xref: /linux/kernel/bpf/verifier.c (revision 02ffd6f89c50ca0bff0c4578949ff99e70451757)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3  * Copyright (c) 2016 Facebook
4  * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
5  */
6 #include <uapi/linux/btf.h>
7 #include <linux/bpf-cgroup.h>
8 #include <linux/kernel.h>
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/bpf.h>
12 #include <linux/btf.h>
13 #include <linux/bpf_verifier.h>
14 #include <linux/filter.h>
15 #include <net/netlink.h>
16 #include <linux/file.h>
17 #include <linux/vmalloc.h>
18 #include <linux/stringify.h>
19 #include <linux/bsearch.h>
20 #include <linux/sort.h>
21 #include <linux/perf_event.h>
22 #include <linux/ctype.h>
23 #include <linux/error-injection.h>
24 #include <linux/bpf_lsm.h>
25 #include <linux/btf_ids.h>
26 #include <linux/poison.h>
27 #include <linux/module.h>
28 #include <linux/cpumask.h>
29 #include <linux/bpf_mem_alloc.h>
30 #include <net/xdp.h>
31 #include <linux/trace_events.h>
32 #include <linux/kallsyms.h>
33 
34 #include "disasm.h"
35 
36 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
37 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
38 	[_id] = & _name ## _verifier_ops,
39 #define BPF_MAP_TYPE(_id, _ops)
40 #define BPF_LINK_TYPE(_id, _name)
41 #include <linux/bpf_types.h>
42 #undef BPF_PROG_TYPE
43 #undef BPF_MAP_TYPE
44 #undef BPF_LINK_TYPE
45 };
46 
47 enum bpf_features {
48 	BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
49 	BPF_FEAT_STREAMS	     = 1,
50 	__MAX_BPF_FEAT,
51 };
52 
53 struct bpf_mem_alloc bpf_global_percpu_ma;
54 static bool bpf_global_percpu_ma_set;
55 
56 /* bpf_check() is a static code analyzer that walks eBPF program
57  * instruction by instruction and updates register/stack state.
58  * All paths of conditional branches are analyzed until 'bpf_exit' insn.
59  *
60  * The first pass is depth-first-search to check that the program is a DAG.
61  * It rejects the following programs:
62  * - larger than BPF_MAXINSNS insns
63  * - if loop is present (detected via back-edge)
64  * - unreachable insns exist (shouldn't be a forest. program = one function)
65  * - out of bounds or malformed jumps
66  * The second pass is all possible path descent from the 1st insn.
67  * Since it's analyzing all paths through the program, the length of the
68  * analysis is limited to 64k insn, which may be hit even if total number of
69  * insn is less then 4K, but there are too many branches that change stack/regs.
70  * Number of 'branches to be analyzed' is limited to 1k
71  *
72  * On entry to each instruction, each register has a type, and the instruction
73  * changes the types of the registers depending on instruction semantics.
74  * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
75  * copied to R1.
76  *
77  * All registers are 64-bit.
78  * R0 - return register
79  * R1-R5 argument passing registers
80  * R6-R9 callee saved registers
81  * R10 - frame pointer read-only
82  *
83  * At the start of BPF program the register R1 contains a pointer to bpf_context
84  * and has type PTR_TO_CTX.
85  *
86  * Verifier tracks arithmetic operations on pointers in case:
87  *    BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
88  *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
89  * 1st insn copies R10 (which has FRAME_PTR) type into R1
90  * and 2nd arithmetic instruction is pattern matched to recognize
91  * that it wants to construct a pointer to some element within stack.
92  * So after 2nd insn, the register R1 has type PTR_TO_STACK
93  * (and -20 constant is saved for further stack bounds checking).
94  * Meaning that this reg is a pointer to stack plus known immediate constant.
95  *
96  * Most of the time the registers have SCALAR_VALUE type, which
97  * means the register has some value, but it's not a valid pointer.
98  * (like pointer plus pointer becomes SCALAR_VALUE type)
99  *
100  * When verifier sees load or store instructions the type of base register
101  * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
102  * four pointer types recognized by check_mem_access() function.
103  *
104  * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
105  * and the range of [ptr, ptr + map's value_size) is accessible.
106  *
107  * registers used to pass values to function calls are checked against
108  * function argument constraints.
109  *
110  * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
111  * It means that the register type passed to this function must be
112  * PTR_TO_STACK and it will be used inside the function as
113  * 'pointer to map element key'
114  *
115  * For example the argument constraints for bpf_map_lookup_elem():
116  *   .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
117  *   .arg1_type = ARG_CONST_MAP_PTR,
118  *   .arg2_type = ARG_PTR_TO_MAP_KEY,
119  *
120  * ret_type says that this function returns 'pointer to map elem value or null'
121  * function expects 1st argument to be a const pointer to 'struct bpf_map' and
122  * 2nd argument should be a pointer to stack, which will be used inside
123  * the helper function as a pointer to map element key.
124  *
125  * On the kernel side the helper function looks like:
126  * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
127  * {
128  *    struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
129  *    void *key = (void *) (unsigned long) r2;
130  *    void *value;
131  *
132  *    here kernel can access 'key' and 'map' pointers safely, knowing that
133  *    [key, key + map->key_size) bytes are valid and were initialized on
134  *    the stack of eBPF program.
135  * }
136  *
137  * Corresponding eBPF program may look like:
138  *    BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),  // after this insn R2 type is FRAME_PTR
139  *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
140  *    BPF_LD_MAP_FD(BPF_REG_1, map_fd),      // after this insn R1 type is CONST_PTR_TO_MAP
141  *    BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
142  * here verifier looks at prototype of map_lookup_elem() and sees:
143  * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
144  * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
145  *
146  * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
147  * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
148  * and were initialized prior to this call.
149  * If it's ok, then verifier allows this BPF_CALL insn and looks at
150  * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
151  * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
152  * returns either pointer to map value or NULL.
153  *
154  * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
155  * insn, the register holding that pointer in the true branch changes state to
156  * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
157  * branch. See check_cond_jmp_op().
158  *
159  * After the call R0 is set to return type of the function and registers R1-R5
160  * are set to NOT_INIT to indicate that they are no longer readable.
161  *
162  * The following reference types represent a potential reference to a kernel
163  * resource which, after first being allocated, must be checked and freed by
164  * the BPF program:
165  * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
166  *
167  * When the verifier sees a helper call return a reference type, it allocates a
168  * pointer id for the reference and stores it in the current function state.
169  * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
170  * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
171  * passes through a NULL-check conditional. For the branch wherein the state is
172  * changed to CONST_IMM, the verifier releases the reference.
173  *
174  * For each helper function that allocates a reference, such as
175  * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
176  * bpf_sk_release(). When a reference type passes into the release function,
177  * the verifier also releases the reference. If any unchecked or unreleased
178  * reference remains at the end of the program, the verifier rejects it.
179  */
180 
181 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
182 struct bpf_verifier_stack_elem {
183 	/* verifier state is 'st'
184 	 * before processing instruction 'insn_idx'
185 	 * and after processing instruction 'prev_insn_idx'
186 	 */
187 	struct bpf_verifier_state st;
188 	int insn_idx;
189 	int prev_insn_idx;
190 	struct bpf_verifier_stack_elem *next;
191 	/* length of verifier log at the time this state was pushed on stack */
192 	u32 log_pos;
193 };
194 
195 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ	8192
196 #define BPF_COMPLEXITY_LIMIT_STATES	64
197 
198 #define BPF_MAP_KEY_POISON	(1ULL << 63)
199 #define BPF_MAP_KEY_SEEN	(1ULL << 62)
200 
201 #define BPF_GLOBAL_PERCPU_MA_MAX_SIZE  512
202 
203 #define BPF_PRIV_STACK_MIN_SIZE		64
204 
205 static int acquire_reference(struct bpf_verifier_env *env, int insn_idx);
206 static int release_reference_nomark(struct bpf_verifier_state *state, int ref_obj_id);
207 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
208 static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
209 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env);
210 static int ref_set_non_owning(struct bpf_verifier_env *env,
211 			      struct bpf_reg_state *reg);
212 static void specialize_kfunc(struct bpf_verifier_env *env,
213 			     u32 func_id, u16 offset, unsigned long *addr);
214 static bool is_trusted_reg(const struct bpf_reg_state *reg);
215 
bpf_map_ptr_poisoned(const struct bpf_insn_aux_data * aux)216 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
217 {
218 	return aux->map_ptr_state.poison;
219 }
220 
bpf_map_ptr_unpriv(const struct bpf_insn_aux_data * aux)221 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
222 {
223 	return aux->map_ptr_state.unpriv;
224 }
225 
bpf_map_ptr_store(struct bpf_insn_aux_data * aux,struct bpf_map * map,bool unpriv,bool poison)226 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
227 			      struct bpf_map *map,
228 			      bool unpriv, bool poison)
229 {
230 	unpriv |= bpf_map_ptr_unpriv(aux);
231 	aux->map_ptr_state.unpriv = unpriv;
232 	aux->map_ptr_state.poison = poison;
233 	aux->map_ptr_state.map_ptr = map;
234 }
235 
bpf_map_key_poisoned(const struct bpf_insn_aux_data * aux)236 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
237 {
238 	return aux->map_key_state & BPF_MAP_KEY_POISON;
239 }
240 
bpf_map_key_unseen(const struct bpf_insn_aux_data * aux)241 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
242 {
243 	return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
244 }
245 
bpf_map_key_immediate(const struct bpf_insn_aux_data * aux)246 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
247 {
248 	return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
249 }
250 
bpf_map_key_store(struct bpf_insn_aux_data * aux,u64 state)251 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
252 {
253 	bool poisoned = bpf_map_key_poisoned(aux);
254 
255 	aux->map_key_state = state | BPF_MAP_KEY_SEEN |
256 			     (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
257 }
258 
bpf_helper_call(const struct bpf_insn * insn)259 static bool bpf_helper_call(const struct bpf_insn *insn)
260 {
261 	return insn->code == (BPF_JMP | BPF_CALL) &&
262 	       insn->src_reg == 0;
263 }
264 
bpf_pseudo_call(const struct bpf_insn * insn)265 static bool bpf_pseudo_call(const struct bpf_insn *insn)
266 {
267 	return insn->code == (BPF_JMP | BPF_CALL) &&
268 	       insn->src_reg == BPF_PSEUDO_CALL;
269 }
270 
bpf_pseudo_kfunc_call(const struct bpf_insn * insn)271 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
272 {
273 	return insn->code == (BPF_JMP | BPF_CALL) &&
274 	       insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
275 }
276 
277 struct bpf_call_arg_meta {
278 	struct bpf_map *map_ptr;
279 	bool raw_mode;
280 	bool pkt_access;
281 	u8 release_regno;
282 	int regno;
283 	int access_size;
284 	int mem_size;
285 	u64 msize_max_value;
286 	int ref_obj_id;
287 	int dynptr_id;
288 	int map_uid;
289 	int func_id;
290 	struct btf *btf;
291 	u32 btf_id;
292 	struct btf *ret_btf;
293 	u32 ret_btf_id;
294 	u32 subprogno;
295 	struct btf_field *kptr_field;
296 	s64 const_map_key;
297 };
298 
299 struct bpf_kfunc_call_arg_meta {
300 	/* In parameters */
301 	struct btf *btf;
302 	u32 func_id;
303 	u32 kfunc_flags;
304 	const struct btf_type *func_proto;
305 	const char *func_name;
306 	/* Out parameters */
307 	u32 ref_obj_id;
308 	u8 release_regno;
309 	bool r0_rdonly;
310 	u32 ret_btf_id;
311 	u64 r0_size;
312 	u32 subprogno;
313 	struct {
314 		u64 value;
315 		bool found;
316 	} arg_constant;
317 
318 	/* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling,
319 	 * generally to pass info about user-defined local kptr types to later
320 	 * verification logic
321 	 *   bpf_obj_drop/bpf_percpu_obj_drop
322 	 *     Record the local kptr type to be drop'd
323 	 *   bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type)
324 	 *     Record the local kptr type to be refcount_incr'd and use
325 	 *     arg_owning_ref to determine whether refcount_acquire should be
326 	 *     fallible
327 	 */
328 	struct btf *arg_btf;
329 	u32 arg_btf_id;
330 	bool arg_owning_ref;
331 	bool arg_prog;
332 
333 	struct {
334 		struct btf_field *field;
335 	} arg_list_head;
336 	struct {
337 		struct btf_field *field;
338 	} arg_rbtree_root;
339 	struct {
340 		enum bpf_dynptr_type type;
341 		u32 id;
342 		u32 ref_obj_id;
343 	} initialized_dynptr;
344 	struct {
345 		u8 spi;
346 		u8 frameno;
347 	} iter;
348 	struct {
349 		struct bpf_map *ptr;
350 		int uid;
351 	} map;
352 	u64 mem_size;
353 };
354 
355 struct btf *btf_vmlinux;
356 
btf_type_name(const struct btf * btf,u32 id)357 static const char *btf_type_name(const struct btf *btf, u32 id)
358 {
359 	return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
360 }
361 
362 static DEFINE_MUTEX(bpf_verifier_lock);
363 static DEFINE_MUTEX(bpf_percpu_ma_lock);
364 
verbose(void * private_data,const char * fmt,...)365 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
366 {
367 	struct bpf_verifier_env *env = private_data;
368 	va_list args;
369 
370 	if (!bpf_verifier_log_needed(&env->log))
371 		return;
372 
373 	va_start(args, fmt);
374 	bpf_verifier_vlog(&env->log, fmt, args);
375 	va_end(args);
376 }
377 
verbose_invalid_scalar(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct bpf_retval_range range,const char * ctx,const char * reg_name)378 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
379 				   struct bpf_reg_state *reg,
380 				   struct bpf_retval_range range, const char *ctx,
381 				   const char *reg_name)
382 {
383 	bool unknown = true;
384 
385 	verbose(env, "%s the register %s has", ctx, reg_name);
386 	if (reg->smin_value > S64_MIN) {
387 		verbose(env, " smin=%lld", reg->smin_value);
388 		unknown = false;
389 	}
390 	if (reg->smax_value < S64_MAX) {
391 		verbose(env, " smax=%lld", reg->smax_value);
392 		unknown = false;
393 	}
394 	if (unknown)
395 		verbose(env, " unknown scalar value");
396 	verbose(env, " should have been in [%d, %d]\n", range.minval, range.maxval);
397 }
398 
reg_not_null(const struct bpf_reg_state * reg)399 static bool reg_not_null(const struct bpf_reg_state *reg)
400 {
401 	enum bpf_reg_type type;
402 
403 	type = reg->type;
404 	if (type_may_be_null(type))
405 		return false;
406 
407 	type = base_type(type);
408 	return type == PTR_TO_SOCKET ||
409 		type == PTR_TO_TCP_SOCK ||
410 		type == PTR_TO_MAP_VALUE ||
411 		type == PTR_TO_MAP_KEY ||
412 		type == PTR_TO_SOCK_COMMON ||
413 		(type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
414 		(type == PTR_TO_MEM && !(reg->type & PTR_UNTRUSTED)) ||
415 		type == CONST_PTR_TO_MAP;
416 }
417 
reg_btf_record(const struct bpf_reg_state * reg)418 static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
419 {
420 	struct btf_record *rec = NULL;
421 	struct btf_struct_meta *meta;
422 
423 	if (reg->type == PTR_TO_MAP_VALUE) {
424 		rec = reg->map_ptr->record;
425 	} else if (type_is_ptr_alloc_obj(reg->type)) {
426 		meta = btf_find_struct_meta(reg->btf, reg->btf_id);
427 		if (meta)
428 			rec = meta->record;
429 	}
430 	return rec;
431 }
432 
subprog_is_global(const struct bpf_verifier_env * env,int subprog)433 static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog)
434 {
435 	struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux;
436 
437 	return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
438 }
439 
subprog_name(const struct bpf_verifier_env * env,int subprog)440 static const char *subprog_name(const struct bpf_verifier_env *env, int subprog)
441 {
442 	struct bpf_func_info *info;
443 
444 	if (!env->prog->aux->func_info)
445 		return "";
446 
447 	info = &env->prog->aux->func_info[subprog];
448 	return btf_type_name(env->prog->aux->btf, info->type_id);
449 }
450 
mark_subprog_exc_cb(struct bpf_verifier_env * env,int subprog)451 static void mark_subprog_exc_cb(struct bpf_verifier_env *env, int subprog)
452 {
453 	struct bpf_subprog_info *info = subprog_info(env, subprog);
454 
455 	info->is_cb = true;
456 	info->is_async_cb = true;
457 	info->is_exception_cb = true;
458 }
459 
subprog_is_exc_cb(struct bpf_verifier_env * env,int subprog)460 static bool subprog_is_exc_cb(struct bpf_verifier_env *env, int subprog)
461 {
462 	return subprog_info(env, subprog)->is_exception_cb;
463 }
464 
reg_may_point_to_spin_lock(const struct bpf_reg_state * reg)465 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
466 {
467 	return btf_record_has_field(reg_btf_record(reg), BPF_SPIN_LOCK | BPF_RES_SPIN_LOCK);
468 }
469 
type_is_rdonly_mem(u32 type)470 static bool type_is_rdonly_mem(u32 type)
471 {
472 	return type & MEM_RDONLY;
473 }
474 
is_acquire_function(enum bpf_func_id func_id,const struct bpf_map * map)475 static bool is_acquire_function(enum bpf_func_id func_id,
476 				const struct bpf_map *map)
477 {
478 	enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
479 
480 	if (func_id == BPF_FUNC_sk_lookup_tcp ||
481 	    func_id == BPF_FUNC_sk_lookup_udp ||
482 	    func_id == BPF_FUNC_skc_lookup_tcp ||
483 	    func_id == BPF_FUNC_ringbuf_reserve ||
484 	    func_id == BPF_FUNC_kptr_xchg)
485 		return true;
486 
487 	if (func_id == BPF_FUNC_map_lookup_elem &&
488 	    (map_type == BPF_MAP_TYPE_SOCKMAP ||
489 	     map_type == BPF_MAP_TYPE_SOCKHASH))
490 		return true;
491 
492 	return false;
493 }
494 
is_ptr_cast_function(enum bpf_func_id func_id)495 static bool is_ptr_cast_function(enum bpf_func_id func_id)
496 {
497 	return func_id == BPF_FUNC_tcp_sock ||
498 		func_id == BPF_FUNC_sk_fullsock ||
499 		func_id == BPF_FUNC_skc_to_tcp_sock ||
500 		func_id == BPF_FUNC_skc_to_tcp6_sock ||
501 		func_id == BPF_FUNC_skc_to_udp6_sock ||
502 		func_id == BPF_FUNC_skc_to_mptcp_sock ||
503 		func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
504 		func_id == BPF_FUNC_skc_to_tcp_request_sock;
505 }
506 
is_dynptr_ref_function(enum bpf_func_id func_id)507 static bool is_dynptr_ref_function(enum bpf_func_id func_id)
508 {
509 	return func_id == BPF_FUNC_dynptr_data;
510 }
511 
512 static bool is_sync_callback_calling_kfunc(u32 btf_id);
513 static bool is_async_callback_calling_kfunc(u32 btf_id);
514 static bool is_callback_calling_kfunc(u32 btf_id);
515 static bool is_bpf_throw_kfunc(struct bpf_insn *insn);
516 
517 static bool is_bpf_wq_set_callback_impl_kfunc(u32 btf_id);
518 
is_sync_callback_calling_function(enum bpf_func_id func_id)519 static bool is_sync_callback_calling_function(enum bpf_func_id func_id)
520 {
521 	return func_id == BPF_FUNC_for_each_map_elem ||
522 	       func_id == BPF_FUNC_find_vma ||
523 	       func_id == BPF_FUNC_loop ||
524 	       func_id == BPF_FUNC_user_ringbuf_drain;
525 }
526 
is_async_callback_calling_function(enum bpf_func_id func_id)527 static bool is_async_callback_calling_function(enum bpf_func_id func_id)
528 {
529 	return func_id == BPF_FUNC_timer_set_callback;
530 }
531 
is_callback_calling_function(enum bpf_func_id func_id)532 static bool is_callback_calling_function(enum bpf_func_id func_id)
533 {
534 	return is_sync_callback_calling_function(func_id) ||
535 	       is_async_callback_calling_function(func_id);
536 }
537 
is_sync_callback_calling_insn(struct bpf_insn * insn)538 static bool is_sync_callback_calling_insn(struct bpf_insn *insn)
539 {
540 	return (bpf_helper_call(insn) && is_sync_callback_calling_function(insn->imm)) ||
541 	       (bpf_pseudo_kfunc_call(insn) && is_sync_callback_calling_kfunc(insn->imm));
542 }
543 
is_async_callback_calling_insn(struct bpf_insn * insn)544 static bool is_async_callback_calling_insn(struct bpf_insn *insn)
545 {
546 	return (bpf_helper_call(insn) && is_async_callback_calling_function(insn->imm)) ||
547 	       (bpf_pseudo_kfunc_call(insn) && is_async_callback_calling_kfunc(insn->imm));
548 }
549 
is_may_goto_insn(struct bpf_insn * insn)550 static bool is_may_goto_insn(struct bpf_insn *insn)
551 {
552 	return insn->code == (BPF_JMP | BPF_JCOND) && insn->src_reg == BPF_MAY_GOTO;
553 }
554 
is_may_goto_insn_at(struct bpf_verifier_env * env,int insn_idx)555 static bool is_may_goto_insn_at(struct bpf_verifier_env *env, int insn_idx)
556 {
557 	return is_may_goto_insn(&env->prog->insnsi[insn_idx]);
558 }
559 
is_storage_get_function(enum bpf_func_id func_id)560 static bool is_storage_get_function(enum bpf_func_id func_id)
561 {
562 	return func_id == BPF_FUNC_sk_storage_get ||
563 	       func_id == BPF_FUNC_inode_storage_get ||
564 	       func_id == BPF_FUNC_task_storage_get ||
565 	       func_id == BPF_FUNC_cgrp_storage_get;
566 }
567 
helper_multiple_ref_obj_use(enum bpf_func_id func_id,const struct bpf_map * map)568 static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
569 					const struct bpf_map *map)
570 {
571 	int ref_obj_uses = 0;
572 
573 	if (is_ptr_cast_function(func_id))
574 		ref_obj_uses++;
575 	if (is_acquire_function(func_id, map))
576 		ref_obj_uses++;
577 	if (is_dynptr_ref_function(func_id))
578 		ref_obj_uses++;
579 
580 	return ref_obj_uses > 1;
581 }
582 
is_cmpxchg_insn(const struct bpf_insn * insn)583 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
584 {
585 	return BPF_CLASS(insn->code) == BPF_STX &&
586 	       BPF_MODE(insn->code) == BPF_ATOMIC &&
587 	       insn->imm == BPF_CMPXCHG;
588 }
589 
is_atomic_load_insn(const struct bpf_insn * insn)590 static bool is_atomic_load_insn(const struct bpf_insn *insn)
591 {
592 	return BPF_CLASS(insn->code) == BPF_STX &&
593 	       BPF_MODE(insn->code) == BPF_ATOMIC &&
594 	       insn->imm == BPF_LOAD_ACQ;
595 }
596 
__get_spi(s32 off)597 static int __get_spi(s32 off)
598 {
599 	return (-off - 1) / BPF_REG_SIZE;
600 }
601 
func(struct bpf_verifier_env * env,const struct bpf_reg_state * reg)602 static struct bpf_func_state *func(struct bpf_verifier_env *env,
603 				   const struct bpf_reg_state *reg)
604 {
605 	struct bpf_verifier_state *cur = env->cur_state;
606 
607 	return cur->frame[reg->frameno];
608 }
609 
is_spi_bounds_valid(struct bpf_func_state * state,int spi,int nr_slots)610 static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
611 {
612        int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
613 
614        /* We need to check that slots between [spi - nr_slots + 1, spi] are
615 	* within [0, allocated_stack).
616 	*
617 	* Please note that the spi grows downwards. For example, a dynptr
618 	* takes the size of two stack slots; the first slot will be at
619 	* spi and the second slot will be at spi - 1.
620 	*/
621        return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
622 }
623 
stack_slot_obj_get_spi(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * obj_kind,int nr_slots)624 static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
625 			          const char *obj_kind, int nr_slots)
626 {
627 	int off, spi;
628 
629 	if (!tnum_is_const(reg->var_off)) {
630 		verbose(env, "%s has to be at a constant offset\n", obj_kind);
631 		return -EINVAL;
632 	}
633 
634 	off = reg->off + reg->var_off.value;
635 	if (off % BPF_REG_SIZE) {
636 		verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
637 		return -EINVAL;
638 	}
639 
640 	spi = __get_spi(off);
641 	if (spi + 1 < nr_slots) {
642 		verbose(env, "cannot pass in %s at an offset=%d\n", obj_kind, off);
643 		return -EINVAL;
644 	}
645 
646 	if (!is_spi_bounds_valid(func(env, reg), spi, nr_slots))
647 		return -ERANGE;
648 	return spi;
649 }
650 
dynptr_get_spi(struct bpf_verifier_env * env,struct bpf_reg_state * reg)651 static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
652 {
653 	return stack_slot_obj_get_spi(env, reg, "dynptr", BPF_DYNPTR_NR_SLOTS);
654 }
655 
iter_get_spi(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int nr_slots)656 static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots)
657 {
658 	return stack_slot_obj_get_spi(env, reg, "iter", nr_slots);
659 }
660 
irq_flag_get_spi(struct bpf_verifier_env * env,struct bpf_reg_state * reg)661 static int irq_flag_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
662 {
663 	return stack_slot_obj_get_spi(env, reg, "irq_flag", 1);
664 }
665 
arg_to_dynptr_type(enum bpf_arg_type arg_type)666 static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
667 {
668 	switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
669 	case DYNPTR_TYPE_LOCAL:
670 		return BPF_DYNPTR_TYPE_LOCAL;
671 	case DYNPTR_TYPE_RINGBUF:
672 		return BPF_DYNPTR_TYPE_RINGBUF;
673 	case DYNPTR_TYPE_SKB:
674 		return BPF_DYNPTR_TYPE_SKB;
675 	case DYNPTR_TYPE_XDP:
676 		return BPF_DYNPTR_TYPE_XDP;
677 	default:
678 		return BPF_DYNPTR_TYPE_INVALID;
679 	}
680 }
681 
get_dynptr_type_flag(enum bpf_dynptr_type type)682 static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type)
683 {
684 	switch (type) {
685 	case BPF_DYNPTR_TYPE_LOCAL:
686 		return DYNPTR_TYPE_LOCAL;
687 	case BPF_DYNPTR_TYPE_RINGBUF:
688 		return DYNPTR_TYPE_RINGBUF;
689 	case BPF_DYNPTR_TYPE_SKB:
690 		return DYNPTR_TYPE_SKB;
691 	case BPF_DYNPTR_TYPE_XDP:
692 		return DYNPTR_TYPE_XDP;
693 	default:
694 		return 0;
695 	}
696 }
697 
dynptr_type_refcounted(enum bpf_dynptr_type type)698 static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
699 {
700 	return type == BPF_DYNPTR_TYPE_RINGBUF;
701 }
702 
703 static void __mark_dynptr_reg(struct bpf_reg_state *reg,
704 			      enum bpf_dynptr_type type,
705 			      bool first_slot, int dynptr_id);
706 
707 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
708 				struct bpf_reg_state *reg);
709 
mark_dynptr_stack_regs(struct bpf_verifier_env * env,struct bpf_reg_state * sreg1,struct bpf_reg_state * sreg2,enum bpf_dynptr_type type)710 static void mark_dynptr_stack_regs(struct bpf_verifier_env *env,
711 				   struct bpf_reg_state *sreg1,
712 				   struct bpf_reg_state *sreg2,
713 				   enum bpf_dynptr_type type)
714 {
715 	int id = ++env->id_gen;
716 
717 	__mark_dynptr_reg(sreg1, type, true, id);
718 	__mark_dynptr_reg(sreg2, type, false, id);
719 }
720 
mark_dynptr_cb_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,enum bpf_dynptr_type type)721 static void mark_dynptr_cb_reg(struct bpf_verifier_env *env,
722 			       struct bpf_reg_state *reg,
723 			       enum bpf_dynptr_type type)
724 {
725 	__mark_dynptr_reg(reg, type, true, ++env->id_gen);
726 }
727 
728 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
729 				        struct bpf_func_state *state, int spi);
730 
mark_stack_slots_dynptr(struct bpf_verifier_env * env,struct bpf_reg_state * reg,enum bpf_arg_type arg_type,int insn_idx,int clone_ref_obj_id)731 static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
732 				   enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id)
733 {
734 	struct bpf_func_state *state = func(env, reg);
735 	enum bpf_dynptr_type type;
736 	int spi, i, err;
737 
738 	spi = dynptr_get_spi(env, reg);
739 	if (spi < 0)
740 		return spi;
741 
742 	/* We cannot assume both spi and spi - 1 belong to the same dynptr,
743 	 * hence we need to call destroy_if_dynptr_stack_slot twice for both,
744 	 * to ensure that for the following example:
745 	 *	[d1][d1][d2][d2]
746 	 * spi    3   2   1   0
747 	 * So marking spi = 2 should lead to destruction of both d1 and d2. In
748 	 * case they do belong to same dynptr, second call won't see slot_type
749 	 * as STACK_DYNPTR and will simply skip destruction.
750 	 */
751 	err = destroy_if_dynptr_stack_slot(env, state, spi);
752 	if (err)
753 		return err;
754 	err = destroy_if_dynptr_stack_slot(env, state, spi - 1);
755 	if (err)
756 		return err;
757 
758 	for (i = 0; i < BPF_REG_SIZE; i++) {
759 		state->stack[spi].slot_type[i] = STACK_DYNPTR;
760 		state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
761 	}
762 
763 	type = arg_to_dynptr_type(arg_type);
764 	if (type == BPF_DYNPTR_TYPE_INVALID)
765 		return -EINVAL;
766 
767 	mark_dynptr_stack_regs(env, &state->stack[spi].spilled_ptr,
768 			       &state->stack[spi - 1].spilled_ptr, type);
769 
770 	if (dynptr_type_refcounted(type)) {
771 		/* The id is used to track proper releasing */
772 		int id;
773 
774 		if (clone_ref_obj_id)
775 			id = clone_ref_obj_id;
776 		else
777 			id = acquire_reference(env, insn_idx);
778 
779 		if (id < 0)
780 			return id;
781 
782 		state->stack[spi].spilled_ptr.ref_obj_id = id;
783 		state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
784 	}
785 
786 	state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
787 	state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
788 
789 	return 0;
790 }
791 
invalidate_dynptr(struct bpf_verifier_env * env,struct bpf_func_state * state,int spi)792 static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi)
793 {
794 	int i;
795 
796 	for (i = 0; i < BPF_REG_SIZE; i++) {
797 		state->stack[spi].slot_type[i] = STACK_INVALID;
798 		state->stack[spi - 1].slot_type[i] = STACK_INVALID;
799 	}
800 
801 	__mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
802 	__mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
803 
804 	/* Why do we need to set REG_LIVE_WRITTEN for STACK_INVALID slot?
805 	 *
806 	 * While we don't allow reading STACK_INVALID, it is still possible to
807 	 * do <8 byte writes marking some but not all slots as STACK_MISC. Then,
808 	 * helpers or insns can do partial read of that part without failing,
809 	 * but check_stack_range_initialized, check_stack_read_var_off, and
810 	 * check_stack_read_fixed_off will do mark_reg_read for all 8-bytes of
811 	 * the slot conservatively. Hence we need to prevent those liveness
812 	 * marking walks.
813 	 *
814 	 * This was not a problem before because STACK_INVALID is only set by
815 	 * default (where the default reg state has its reg->parent as NULL), or
816 	 * in clean_live_states after REG_LIVE_DONE (at which point
817 	 * mark_reg_read won't walk reg->parent chain), but not randomly during
818 	 * verifier state exploration (like we did above). Hence, for our case
819 	 * parentage chain will still be live (i.e. reg->parent may be
820 	 * non-NULL), while earlier reg->parent was NULL, so we need
821 	 * REG_LIVE_WRITTEN to screen off read marker propagation when it is
822 	 * done later on reads or by mark_dynptr_read as well to unnecessary
823 	 * mark registers in verifier state.
824 	 */
825 	state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
826 	state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
827 }
828 
unmark_stack_slots_dynptr(struct bpf_verifier_env * env,struct bpf_reg_state * reg)829 static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
830 {
831 	struct bpf_func_state *state = func(env, reg);
832 	int spi, ref_obj_id, i;
833 
834 	spi = dynptr_get_spi(env, reg);
835 	if (spi < 0)
836 		return spi;
837 
838 	if (!dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
839 		invalidate_dynptr(env, state, spi);
840 		return 0;
841 	}
842 
843 	ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id;
844 
845 	/* If the dynptr has a ref_obj_id, then we need to invalidate
846 	 * two things:
847 	 *
848 	 * 1) Any dynptrs with a matching ref_obj_id (clones)
849 	 * 2) Any slices derived from this dynptr.
850 	 */
851 
852 	/* Invalidate any slices associated with this dynptr */
853 	WARN_ON_ONCE(release_reference(env, ref_obj_id));
854 
855 	/* Invalidate any dynptr clones */
856 	for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
857 		if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
858 			continue;
859 
860 		/* it should always be the case that if the ref obj id
861 		 * matches then the stack slot also belongs to a
862 		 * dynptr
863 		 */
864 		if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
865 			verifier_bug(env, "misconfigured ref_obj_id");
866 			return -EFAULT;
867 		}
868 		if (state->stack[i].spilled_ptr.dynptr.first_slot)
869 			invalidate_dynptr(env, state, i);
870 	}
871 
872 	return 0;
873 }
874 
875 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
876 			       struct bpf_reg_state *reg);
877 
mark_reg_invalid(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)878 static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
879 {
880 	if (!env->allow_ptr_leaks)
881 		__mark_reg_not_init(env, reg);
882 	else
883 		__mark_reg_unknown(env, reg);
884 }
885 
destroy_if_dynptr_stack_slot(struct bpf_verifier_env * env,struct bpf_func_state * state,int spi)886 static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
887 				        struct bpf_func_state *state, int spi)
888 {
889 	struct bpf_func_state *fstate;
890 	struct bpf_reg_state *dreg;
891 	int i, dynptr_id;
892 
893 	/* We always ensure that STACK_DYNPTR is never set partially,
894 	 * hence just checking for slot_type[0] is enough. This is
895 	 * different for STACK_SPILL, where it may be only set for
896 	 * 1 byte, so code has to use is_spilled_reg.
897 	 */
898 	if (state->stack[spi].slot_type[0] != STACK_DYNPTR)
899 		return 0;
900 
901 	/* Reposition spi to first slot */
902 	if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
903 		spi = spi + 1;
904 
905 	if (dynptr_type_refcounted(state->stack[spi].spilled_ptr.dynptr.type)) {
906 		verbose(env, "cannot overwrite referenced dynptr\n");
907 		return -EINVAL;
908 	}
909 
910 	mark_stack_slot_scratched(env, spi);
911 	mark_stack_slot_scratched(env, spi - 1);
912 
913 	/* Writing partially to one dynptr stack slot destroys both. */
914 	for (i = 0; i < BPF_REG_SIZE; i++) {
915 		state->stack[spi].slot_type[i] = STACK_INVALID;
916 		state->stack[spi - 1].slot_type[i] = STACK_INVALID;
917 	}
918 
919 	dynptr_id = state->stack[spi].spilled_ptr.id;
920 	/* Invalidate any slices associated with this dynptr */
921 	bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({
922 		/* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */
923 		if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM)
924 			continue;
925 		if (dreg->dynptr_id == dynptr_id)
926 			mark_reg_invalid(env, dreg);
927 	}));
928 
929 	/* Do not release reference state, we are destroying dynptr on stack,
930 	 * not using some helper to release it. Just reset register.
931 	 */
932 	__mark_reg_not_init(env, &state->stack[spi].spilled_ptr);
933 	__mark_reg_not_init(env, &state->stack[spi - 1].spilled_ptr);
934 
935 	/* Same reason as unmark_stack_slots_dynptr above */
936 	state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
937 	state->stack[spi - 1].spilled_ptr.live |= REG_LIVE_WRITTEN;
938 
939 	return 0;
940 }
941 
is_dynptr_reg_valid_uninit(struct bpf_verifier_env * env,struct bpf_reg_state * reg)942 static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
943 {
944 	int spi;
945 
946 	if (reg->type == CONST_PTR_TO_DYNPTR)
947 		return false;
948 
949 	spi = dynptr_get_spi(env, reg);
950 
951 	/* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an
952 	 * error because this just means the stack state hasn't been updated yet.
953 	 * We will do check_mem_access to check and update stack bounds later.
954 	 */
955 	if (spi < 0 && spi != -ERANGE)
956 		return false;
957 
958 	/* We don't need to check if the stack slots are marked by previous
959 	 * dynptr initializations because we allow overwriting existing unreferenced
960 	 * STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls
961 	 * destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are
962 	 * touching are completely destructed before we reinitialize them for a new
963 	 * one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early
964 	 * instead of delaying it until the end where the user will get "Unreleased
965 	 * reference" error.
966 	 */
967 	return true;
968 }
969 
is_dynptr_reg_valid_init(struct bpf_verifier_env * env,struct bpf_reg_state * reg)970 static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
971 {
972 	struct bpf_func_state *state = func(env, reg);
973 	int i, spi;
974 
975 	/* This already represents first slot of initialized bpf_dynptr.
976 	 *
977 	 * CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
978 	 * check_func_arg_reg_off's logic, so we don't need to check its
979 	 * offset and alignment.
980 	 */
981 	if (reg->type == CONST_PTR_TO_DYNPTR)
982 		return true;
983 
984 	spi = dynptr_get_spi(env, reg);
985 	if (spi < 0)
986 		return false;
987 	if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
988 		return false;
989 
990 	for (i = 0; i < BPF_REG_SIZE; i++) {
991 		if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
992 		    state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
993 			return false;
994 	}
995 
996 	return true;
997 }
998 
is_dynptr_type_expected(struct bpf_verifier_env * env,struct bpf_reg_state * reg,enum bpf_arg_type arg_type)999 static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1000 				    enum bpf_arg_type arg_type)
1001 {
1002 	struct bpf_func_state *state = func(env, reg);
1003 	enum bpf_dynptr_type dynptr_type;
1004 	int spi;
1005 
1006 	/* ARG_PTR_TO_DYNPTR takes any type of dynptr */
1007 	if (arg_type == ARG_PTR_TO_DYNPTR)
1008 		return true;
1009 
1010 	dynptr_type = arg_to_dynptr_type(arg_type);
1011 	if (reg->type == CONST_PTR_TO_DYNPTR) {
1012 		return reg->dynptr.type == dynptr_type;
1013 	} else {
1014 		spi = dynptr_get_spi(env, reg);
1015 		if (spi < 0)
1016 			return false;
1017 		return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
1018 	}
1019 }
1020 
1021 static void __mark_reg_known_zero(struct bpf_reg_state *reg);
1022 
1023 static bool in_rcu_cs(struct bpf_verifier_env *env);
1024 
1025 static bool is_kfunc_rcu_protected(struct bpf_kfunc_call_arg_meta *meta);
1026 
mark_stack_slots_iter(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,struct bpf_reg_state * reg,int insn_idx,struct btf * btf,u32 btf_id,int nr_slots)1027 static int mark_stack_slots_iter(struct bpf_verifier_env *env,
1028 				 struct bpf_kfunc_call_arg_meta *meta,
1029 				 struct bpf_reg_state *reg, int insn_idx,
1030 				 struct btf *btf, u32 btf_id, int nr_slots)
1031 {
1032 	struct bpf_func_state *state = func(env, reg);
1033 	int spi, i, j, id;
1034 
1035 	spi = iter_get_spi(env, reg, nr_slots);
1036 	if (spi < 0)
1037 		return spi;
1038 
1039 	id = acquire_reference(env, insn_idx);
1040 	if (id < 0)
1041 		return id;
1042 
1043 	for (i = 0; i < nr_slots; i++) {
1044 		struct bpf_stack_state *slot = &state->stack[spi - i];
1045 		struct bpf_reg_state *st = &slot->spilled_ptr;
1046 
1047 		__mark_reg_known_zero(st);
1048 		st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1049 		if (is_kfunc_rcu_protected(meta)) {
1050 			if (in_rcu_cs(env))
1051 				st->type |= MEM_RCU;
1052 			else
1053 				st->type |= PTR_UNTRUSTED;
1054 		}
1055 		st->live |= REG_LIVE_WRITTEN;
1056 		st->ref_obj_id = i == 0 ? id : 0;
1057 		st->iter.btf = btf;
1058 		st->iter.btf_id = btf_id;
1059 		st->iter.state = BPF_ITER_STATE_ACTIVE;
1060 		st->iter.depth = 0;
1061 
1062 		for (j = 0; j < BPF_REG_SIZE; j++)
1063 			slot->slot_type[j] = STACK_ITER;
1064 
1065 		mark_stack_slot_scratched(env, spi - i);
1066 	}
1067 
1068 	return 0;
1069 }
1070 
unmark_stack_slots_iter(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int nr_slots)1071 static int unmark_stack_slots_iter(struct bpf_verifier_env *env,
1072 				   struct bpf_reg_state *reg, int nr_slots)
1073 {
1074 	struct bpf_func_state *state = func(env, reg);
1075 	int spi, i, j;
1076 
1077 	spi = iter_get_spi(env, reg, nr_slots);
1078 	if (spi < 0)
1079 		return spi;
1080 
1081 	for (i = 0; i < nr_slots; i++) {
1082 		struct bpf_stack_state *slot = &state->stack[spi - i];
1083 		struct bpf_reg_state *st = &slot->spilled_ptr;
1084 
1085 		if (i == 0)
1086 			WARN_ON_ONCE(release_reference(env, st->ref_obj_id));
1087 
1088 		__mark_reg_not_init(env, st);
1089 
1090 		/* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */
1091 		st->live |= REG_LIVE_WRITTEN;
1092 
1093 		for (j = 0; j < BPF_REG_SIZE; j++)
1094 			slot->slot_type[j] = STACK_INVALID;
1095 
1096 		mark_stack_slot_scratched(env, spi - i);
1097 	}
1098 
1099 	return 0;
1100 }
1101 
is_iter_reg_valid_uninit(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int nr_slots)1102 static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env,
1103 				     struct bpf_reg_state *reg, int nr_slots)
1104 {
1105 	struct bpf_func_state *state = func(env, reg);
1106 	int spi, i, j;
1107 
1108 	/* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1109 	 * will do check_mem_access to check and update stack bounds later, so
1110 	 * return true for that case.
1111 	 */
1112 	spi = iter_get_spi(env, reg, nr_slots);
1113 	if (spi == -ERANGE)
1114 		return true;
1115 	if (spi < 0)
1116 		return false;
1117 
1118 	for (i = 0; i < nr_slots; i++) {
1119 		struct bpf_stack_state *slot = &state->stack[spi - i];
1120 
1121 		for (j = 0; j < BPF_REG_SIZE; j++)
1122 			if (slot->slot_type[j] == STACK_ITER)
1123 				return false;
1124 	}
1125 
1126 	return true;
1127 }
1128 
is_iter_reg_valid_init(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct btf * btf,u32 btf_id,int nr_slots)1129 static int is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1130 				   struct btf *btf, u32 btf_id, int nr_slots)
1131 {
1132 	struct bpf_func_state *state = func(env, reg);
1133 	int spi, i, j;
1134 
1135 	spi = iter_get_spi(env, reg, nr_slots);
1136 	if (spi < 0)
1137 		return -EINVAL;
1138 
1139 	for (i = 0; i < nr_slots; i++) {
1140 		struct bpf_stack_state *slot = &state->stack[spi - i];
1141 		struct bpf_reg_state *st = &slot->spilled_ptr;
1142 
1143 		if (st->type & PTR_UNTRUSTED)
1144 			return -EPROTO;
1145 		/* only main (first) slot has ref_obj_id set */
1146 		if (i == 0 && !st->ref_obj_id)
1147 			return -EINVAL;
1148 		if (i != 0 && st->ref_obj_id)
1149 			return -EINVAL;
1150 		if (st->iter.btf != btf || st->iter.btf_id != btf_id)
1151 			return -EINVAL;
1152 
1153 		for (j = 0; j < BPF_REG_SIZE; j++)
1154 			if (slot->slot_type[j] != STACK_ITER)
1155 				return -EINVAL;
1156 	}
1157 
1158 	return 0;
1159 }
1160 
1161 static int acquire_irq_state(struct bpf_verifier_env *env, int insn_idx);
1162 static int release_irq_state(struct bpf_verifier_state *state, int id);
1163 
mark_stack_slot_irq_flag(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,struct bpf_reg_state * reg,int insn_idx,int kfunc_class)1164 static int mark_stack_slot_irq_flag(struct bpf_verifier_env *env,
1165 				     struct bpf_kfunc_call_arg_meta *meta,
1166 				     struct bpf_reg_state *reg, int insn_idx,
1167 				     int kfunc_class)
1168 {
1169 	struct bpf_func_state *state = func(env, reg);
1170 	struct bpf_stack_state *slot;
1171 	struct bpf_reg_state *st;
1172 	int spi, i, id;
1173 
1174 	spi = irq_flag_get_spi(env, reg);
1175 	if (spi < 0)
1176 		return spi;
1177 
1178 	id = acquire_irq_state(env, insn_idx);
1179 	if (id < 0)
1180 		return id;
1181 
1182 	slot = &state->stack[spi];
1183 	st = &slot->spilled_ptr;
1184 
1185 	__mark_reg_known_zero(st);
1186 	st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1187 	st->live |= REG_LIVE_WRITTEN;
1188 	st->ref_obj_id = id;
1189 	st->irq.kfunc_class = kfunc_class;
1190 
1191 	for (i = 0; i < BPF_REG_SIZE; i++)
1192 		slot->slot_type[i] = STACK_IRQ_FLAG;
1193 
1194 	mark_stack_slot_scratched(env, spi);
1195 	return 0;
1196 }
1197 
unmark_stack_slot_irq_flag(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int kfunc_class)1198 static int unmark_stack_slot_irq_flag(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1199 				      int kfunc_class)
1200 {
1201 	struct bpf_func_state *state = func(env, reg);
1202 	struct bpf_stack_state *slot;
1203 	struct bpf_reg_state *st;
1204 	int spi, i, err;
1205 
1206 	spi = irq_flag_get_spi(env, reg);
1207 	if (spi < 0)
1208 		return spi;
1209 
1210 	slot = &state->stack[spi];
1211 	st = &slot->spilled_ptr;
1212 
1213 	if (st->irq.kfunc_class != kfunc_class) {
1214 		const char *flag_kfunc = st->irq.kfunc_class == IRQ_NATIVE_KFUNC ? "native" : "lock";
1215 		const char *used_kfunc = kfunc_class == IRQ_NATIVE_KFUNC ? "native" : "lock";
1216 
1217 		verbose(env, "irq flag acquired by %s kfuncs cannot be restored with %s kfuncs\n",
1218 			flag_kfunc, used_kfunc);
1219 		return -EINVAL;
1220 	}
1221 
1222 	err = release_irq_state(env->cur_state, st->ref_obj_id);
1223 	WARN_ON_ONCE(err && err != -EACCES);
1224 	if (err) {
1225 		int insn_idx = 0;
1226 
1227 		for (int i = 0; i < env->cur_state->acquired_refs; i++) {
1228 			if (env->cur_state->refs[i].id == env->cur_state->active_irq_id) {
1229 				insn_idx = env->cur_state->refs[i].insn_idx;
1230 				break;
1231 			}
1232 		}
1233 
1234 		verbose(env, "cannot restore irq state out of order, expected id=%d acquired at insn_idx=%d\n",
1235 			env->cur_state->active_irq_id, insn_idx);
1236 		return err;
1237 	}
1238 
1239 	__mark_reg_not_init(env, st);
1240 
1241 	/* see unmark_stack_slots_dynptr() for why we need to set REG_LIVE_WRITTEN */
1242 	st->live |= REG_LIVE_WRITTEN;
1243 
1244 	for (i = 0; i < BPF_REG_SIZE; i++)
1245 		slot->slot_type[i] = STACK_INVALID;
1246 
1247 	mark_stack_slot_scratched(env, spi);
1248 	return 0;
1249 }
1250 
is_irq_flag_reg_valid_uninit(struct bpf_verifier_env * env,struct bpf_reg_state * reg)1251 static bool is_irq_flag_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1252 {
1253 	struct bpf_func_state *state = func(env, reg);
1254 	struct bpf_stack_state *slot;
1255 	int spi, i;
1256 
1257 	/* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1258 	 * will do check_mem_access to check and update stack bounds later, so
1259 	 * return true for that case.
1260 	 */
1261 	spi = irq_flag_get_spi(env, reg);
1262 	if (spi == -ERANGE)
1263 		return true;
1264 	if (spi < 0)
1265 		return false;
1266 
1267 	slot = &state->stack[spi];
1268 
1269 	for (i = 0; i < BPF_REG_SIZE; i++)
1270 		if (slot->slot_type[i] == STACK_IRQ_FLAG)
1271 			return false;
1272 	return true;
1273 }
1274 
is_irq_flag_reg_valid_init(struct bpf_verifier_env * env,struct bpf_reg_state * reg)1275 static int is_irq_flag_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1276 {
1277 	struct bpf_func_state *state = func(env, reg);
1278 	struct bpf_stack_state *slot;
1279 	struct bpf_reg_state *st;
1280 	int spi, i;
1281 
1282 	spi = irq_flag_get_spi(env, reg);
1283 	if (spi < 0)
1284 		return -EINVAL;
1285 
1286 	slot = &state->stack[spi];
1287 	st = &slot->spilled_ptr;
1288 
1289 	if (!st->ref_obj_id)
1290 		return -EINVAL;
1291 
1292 	for (i = 0; i < BPF_REG_SIZE; i++)
1293 		if (slot->slot_type[i] != STACK_IRQ_FLAG)
1294 			return -EINVAL;
1295 	return 0;
1296 }
1297 
1298 /* Check if given stack slot is "special":
1299  *   - spilled register state (STACK_SPILL);
1300  *   - dynptr state (STACK_DYNPTR);
1301  *   - iter state (STACK_ITER).
1302  *   - irq flag state (STACK_IRQ_FLAG)
1303  */
is_stack_slot_special(const struct bpf_stack_state * stack)1304 static bool is_stack_slot_special(const struct bpf_stack_state *stack)
1305 {
1306 	enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1];
1307 
1308 	switch (type) {
1309 	case STACK_SPILL:
1310 	case STACK_DYNPTR:
1311 	case STACK_ITER:
1312 	case STACK_IRQ_FLAG:
1313 		return true;
1314 	case STACK_INVALID:
1315 	case STACK_MISC:
1316 	case STACK_ZERO:
1317 		return false;
1318 	default:
1319 		WARN_ONCE(1, "unknown stack slot type %d\n", type);
1320 		return true;
1321 	}
1322 }
1323 
1324 /* The reg state of a pointer or a bounded scalar was saved when
1325  * it was spilled to the stack.
1326  */
is_spilled_reg(const struct bpf_stack_state * stack)1327 static bool is_spilled_reg(const struct bpf_stack_state *stack)
1328 {
1329 	return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
1330 }
1331 
is_spilled_scalar_reg(const struct bpf_stack_state * stack)1332 static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack)
1333 {
1334 	return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL &&
1335 	       stack->spilled_ptr.type == SCALAR_VALUE;
1336 }
1337 
is_spilled_scalar_reg64(const struct bpf_stack_state * stack)1338 static bool is_spilled_scalar_reg64(const struct bpf_stack_state *stack)
1339 {
1340 	return stack->slot_type[0] == STACK_SPILL &&
1341 	       stack->spilled_ptr.type == SCALAR_VALUE;
1342 }
1343 
1344 /* Mark stack slot as STACK_MISC, unless it is already STACK_INVALID, in which
1345  * case they are equivalent, or it's STACK_ZERO, in which case we preserve
1346  * more precise STACK_ZERO.
1347  * Regardless of allow_ptr_leaks setting (i.e., privileged or unprivileged
1348  * mode), we won't promote STACK_INVALID to STACK_MISC. In privileged case it is
1349  * unnecessary as both are considered equivalent when loading data and pruning,
1350  * in case of unprivileged mode it will be incorrect to allow reads of invalid
1351  * slots.
1352  */
mark_stack_slot_misc(struct bpf_verifier_env * env,u8 * stype)1353 static void mark_stack_slot_misc(struct bpf_verifier_env *env, u8 *stype)
1354 {
1355 	if (*stype == STACK_ZERO)
1356 		return;
1357 	if (*stype == STACK_INVALID)
1358 		return;
1359 	*stype = STACK_MISC;
1360 }
1361 
scrub_spilled_slot(u8 * stype)1362 static void scrub_spilled_slot(u8 *stype)
1363 {
1364 	if (*stype != STACK_INVALID)
1365 		*stype = STACK_MISC;
1366 }
1367 
1368 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1369  * small to hold src. This is different from krealloc since we don't want to preserve
1370  * the contents of dst.
1371  *
1372  * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1373  * not be allocated.
1374  */
copy_array(void * dst,const void * src,size_t n,size_t size,gfp_t flags)1375 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1376 {
1377 	size_t alloc_bytes;
1378 	void *orig = dst;
1379 	size_t bytes;
1380 
1381 	if (ZERO_OR_NULL_PTR(src))
1382 		goto out;
1383 
1384 	if (unlikely(check_mul_overflow(n, size, &bytes)))
1385 		return NULL;
1386 
1387 	alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1388 	dst = krealloc(orig, alloc_bytes, flags);
1389 	if (!dst) {
1390 		kfree(orig);
1391 		return NULL;
1392 	}
1393 
1394 	memcpy(dst, src, bytes);
1395 out:
1396 	return dst ? dst : ZERO_SIZE_PTR;
1397 }
1398 
1399 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
1400  * small to hold new_n items. new items are zeroed out if the array grows.
1401  *
1402  * Contrary to krealloc_array, does not free arr if new_n is zero.
1403  */
realloc_array(void * arr,size_t old_n,size_t new_n,size_t size)1404 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1405 {
1406 	size_t alloc_size;
1407 	void *new_arr;
1408 
1409 	if (!new_n || old_n == new_n)
1410 		goto out;
1411 
1412 	alloc_size = kmalloc_size_roundup(size_mul(new_n, size));
1413 	new_arr = krealloc(arr, alloc_size, GFP_KERNEL_ACCOUNT);
1414 	if (!new_arr) {
1415 		kfree(arr);
1416 		return NULL;
1417 	}
1418 	arr = new_arr;
1419 
1420 	if (new_n > old_n)
1421 		memset(arr + old_n * size, 0, (new_n - old_n) * size);
1422 
1423 out:
1424 	return arr ? arr : ZERO_SIZE_PTR;
1425 }
1426 
copy_reference_state(struct bpf_verifier_state * dst,const struct bpf_verifier_state * src)1427 static int copy_reference_state(struct bpf_verifier_state *dst, const struct bpf_verifier_state *src)
1428 {
1429 	dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
1430 			       sizeof(struct bpf_reference_state), GFP_KERNEL_ACCOUNT);
1431 	if (!dst->refs)
1432 		return -ENOMEM;
1433 
1434 	dst->acquired_refs = src->acquired_refs;
1435 	dst->active_locks = src->active_locks;
1436 	dst->active_preempt_locks = src->active_preempt_locks;
1437 	dst->active_rcu_lock = src->active_rcu_lock;
1438 	dst->active_irq_id = src->active_irq_id;
1439 	dst->active_lock_id = src->active_lock_id;
1440 	dst->active_lock_ptr = src->active_lock_ptr;
1441 	return 0;
1442 }
1443 
copy_stack_state(struct bpf_func_state * dst,const struct bpf_func_state * src)1444 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1445 {
1446 	size_t n = src->allocated_stack / BPF_REG_SIZE;
1447 
1448 	dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
1449 				GFP_KERNEL_ACCOUNT);
1450 	if (!dst->stack)
1451 		return -ENOMEM;
1452 
1453 	dst->allocated_stack = src->allocated_stack;
1454 	return 0;
1455 }
1456 
resize_reference_state(struct bpf_verifier_state * state,size_t n)1457 static int resize_reference_state(struct bpf_verifier_state *state, size_t n)
1458 {
1459 	state->refs = realloc_array(state->refs, state->acquired_refs, n,
1460 				    sizeof(struct bpf_reference_state));
1461 	if (!state->refs)
1462 		return -ENOMEM;
1463 
1464 	state->acquired_refs = n;
1465 	return 0;
1466 }
1467 
1468 /* Possibly update state->allocated_stack to be at least size bytes. Also
1469  * possibly update the function's high-water mark in its bpf_subprog_info.
1470  */
grow_stack_state(struct bpf_verifier_env * env,struct bpf_func_state * state,int size)1471 static int grow_stack_state(struct bpf_verifier_env *env, struct bpf_func_state *state, int size)
1472 {
1473 	size_t old_n = state->allocated_stack / BPF_REG_SIZE, n;
1474 
1475 	/* The stack size is always a multiple of BPF_REG_SIZE. */
1476 	size = round_up(size, BPF_REG_SIZE);
1477 	n = size / BPF_REG_SIZE;
1478 
1479 	if (old_n >= n)
1480 		return 0;
1481 
1482 	state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
1483 	if (!state->stack)
1484 		return -ENOMEM;
1485 
1486 	state->allocated_stack = size;
1487 
1488 	/* update known max for given subprogram */
1489 	if (env->subprog_info[state->subprogno].stack_depth < size)
1490 		env->subprog_info[state->subprogno].stack_depth = size;
1491 
1492 	return 0;
1493 }
1494 
1495 /* Acquire a pointer id from the env and update the state->refs to include
1496  * this new pointer reference.
1497  * On success, returns a valid pointer id to associate with the register
1498  * On failure, returns a negative errno.
1499  */
acquire_reference_state(struct bpf_verifier_env * env,int insn_idx)1500 static struct bpf_reference_state *acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1501 {
1502 	struct bpf_verifier_state *state = env->cur_state;
1503 	int new_ofs = state->acquired_refs;
1504 	int err;
1505 
1506 	err = resize_reference_state(state, state->acquired_refs + 1);
1507 	if (err)
1508 		return NULL;
1509 	state->refs[new_ofs].insn_idx = insn_idx;
1510 
1511 	return &state->refs[new_ofs];
1512 }
1513 
acquire_reference(struct bpf_verifier_env * env,int insn_idx)1514 static int acquire_reference(struct bpf_verifier_env *env, int insn_idx)
1515 {
1516 	struct bpf_reference_state *s;
1517 
1518 	s = acquire_reference_state(env, insn_idx);
1519 	if (!s)
1520 		return -ENOMEM;
1521 	s->type = REF_TYPE_PTR;
1522 	s->id = ++env->id_gen;
1523 	return s->id;
1524 }
1525 
acquire_lock_state(struct bpf_verifier_env * env,int insn_idx,enum ref_state_type type,int id,void * ptr)1526 static int acquire_lock_state(struct bpf_verifier_env *env, int insn_idx, enum ref_state_type type,
1527 			      int id, void *ptr)
1528 {
1529 	struct bpf_verifier_state *state = env->cur_state;
1530 	struct bpf_reference_state *s;
1531 
1532 	s = acquire_reference_state(env, insn_idx);
1533 	if (!s)
1534 		return -ENOMEM;
1535 	s->type = type;
1536 	s->id = id;
1537 	s->ptr = ptr;
1538 
1539 	state->active_locks++;
1540 	state->active_lock_id = id;
1541 	state->active_lock_ptr = ptr;
1542 	return 0;
1543 }
1544 
acquire_irq_state(struct bpf_verifier_env * env,int insn_idx)1545 static int acquire_irq_state(struct bpf_verifier_env *env, int insn_idx)
1546 {
1547 	struct bpf_verifier_state *state = env->cur_state;
1548 	struct bpf_reference_state *s;
1549 
1550 	s = acquire_reference_state(env, insn_idx);
1551 	if (!s)
1552 		return -ENOMEM;
1553 	s->type = REF_TYPE_IRQ;
1554 	s->id = ++env->id_gen;
1555 
1556 	state->active_irq_id = s->id;
1557 	return s->id;
1558 }
1559 
release_reference_state(struct bpf_verifier_state * state,int idx)1560 static void release_reference_state(struct bpf_verifier_state *state, int idx)
1561 {
1562 	int last_idx;
1563 	size_t rem;
1564 
1565 	/* IRQ state requires the relative ordering of elements remaining the
1566 	 * same, since it relies on the refs array to behave as a stack, so that
1567 	 * it can detect out-of-order IRQ restore. Hence use memmove to shift
1568 	 * the array instead of swapping the final element into the deleted idx.
1569 	 */
1570 	last_idx = state->acquired_refs - 1;
1571 	rem = state->acquired_refs - idx - 1;
1572 	if (last_idx && idx != last_idx)
1573 		memmove(&state->refs[idx], &state->refs[idx + 1], sizeof(*state->refs) * rem);
1574 	memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1575 	state->acquired_refs--;
1576 	return;
1577 }
1578 
find_reference_state(struct bpf_verifier_state * state,int ptr_id)1579 static bool find_reference_state(struct bpf_verifier_state *state, int ptr_id)
1580 {
1581 	int i;
1582 
1583 	for (i = 0; i < state->acquired_refs; i++)
1584 		if (state->refs[i].id == ptr_id)
1585 			return true;
1586 
1587 	return false;
1588 }
1589 
release_lock_state(struct bpf_verifier_state * state,int type,int id,void * ptr)1590 static int release_lock_state(struct bpf_verifier_state *state, int type, int id, void *ptr)
1591 {
1592 	void *prev_ptr = NULL;
1593 	u32 prev_id = 0;
1594 	int i;
1595 
1596 	for (i = 0; i < state->acquired_refs; i++) {
1597 		if (state->refs[i].type == type && state->refs[i].id == id &&
1598 		    state->refs[i].ptr == ptr) {
1599 			release_reference_state(state, i);
1600 			state->active_locks--;
1601 			/* Reassign active lock (id, ptr). */
1602 			state->active_lock_id = prev_id;
1603 			state->active_lock_ptr = prev_ptr;
1604 			return 0;
1605 		}
1606 		if (state->refs[i].type & REF_TYPE_LOCK_MASK) {
1607 			prev_id = state->refs[i].id;
1608 			prev_ptr = state->refs[i].ptr;
1609 		}
1610 	}
1611 	return -EINVAL;
1612 }
1613 
release_irq_state(struct bpf_verifier_state * state,int id)1614 static int release_irq_state(struct bpf_verifier_state *state, int id)
1615 {
1616 	u32 prev_id = 0;
1617 	int i;
1618 
1619 	if (id != state->active_irq_id)
1620 		return -EACCES;
1621 
1622 	for (i = 0; i < state->acquired_refs; i++) {
1623 		if (state->refs[i].type != REF_TYPE_IRQ)
1624 			continue;
1625 		if (state->refs[i].id == id) {
1626 			release_reference_state(state, i);
1627 			state->active_irq_id = prev_id;
1628 			return 0;
1629 		} else {
1630 			prev_id = state->refs[i].id;
1631 		}
1632 	}
1633 	return -EINVAL;
1634 }
1635 
find_lock_state(struct bpf_verifier_state * state,enum ref_state_type type,int id,void * ptr)1636 static struct bpf_reference_state *find_lock_state(struct bpf_verifier_state *state, enum ref_state_type type,
1637 						   int id, void *ptr)
1638 {
1639 	int i;
1640 
1641 	for (i = 0; i < state->acquired_refs; i++) {
1642 		struct bpf_reference_state *s = &state->refs[i];
1643 
1644 		if (!(s->type & type))
1645 			continue;
1646 
1647 		if (s->id == id && s->ptr == ptr)
1648 			return s;
1649 	}
1650 	return NULL;
1651 }
1652 
update_peak_states(struct bpf_verifier_env * env)1653 static void update_peak_states(struct bpf_verifier_env *env)
1654 {
1655 	u32 cur_states;
1656 
1657 	cur_states = env->explored_states_size + env->free_list_size + env->num_backedges;
1658 	env->peak_states = max(env->peak_states, cur_states);
1659 }
1660 
free_func_state(struct bpf_func_state * state)1661 static void free_func_state(struct bpf_func_state *state)
1662 {
1663 	if (!state)
1664 		return;
1665 	kfree(state->stack);
1666 	kfree(state);
1667 }
1668 
clear_jmp_history(struct bpf_verifier_state * state)1669 static void clear_jmp_history(struct bpf_verifier_state *state)
1670 {
1671 	kfree(state->jmp_history);
1672 	state->jmp_history = NULL;
1673 	state->jmp_history_cnt = 0;
1674 }
1675 
free_verifier_state(struct bpf_verifier_state * state,bool free_self)1676 static void free_verifier_state(struct bpf_verifier_state *state,
1677 				bool free_self)
1678 {
1679 	int i;
1680 
1681 	for (i = 0; i <= state->curframe; i++) {
1682 		free_func_state(state->frame[i]);
1683 		state->frame[i] = NULL;
1684 	}
1685 	kfree(state->refs);
1686 	clear_jmp_history(state);
1687 	if (free_self)
1688 		kfree(state);
1689 }
1690 
1691 /* struct bpf_verifier_state->parent refers to states
1692  * that are in either of env->{expored_states,free_list}.
1693  * In both cases the state is contained in struct bpf_verifier_state_list.
1694  */
state_parent_as_list(struct bpf_verifier_state * st)1695 static struct bpf_verifier_state_list *state_parent_as_list(struct bpf_verifier_state *st)
1696 {
1697 	if (st->parent)
1698 		return container_of(st->parent, struct bpf_verifier_state_list, state);
1699 	return NULL;
1700 }
1701 
1702 static bool incomplete_read_marks(struct bpf_verifier_env *env,
1703 				  struct bpf_verifier_state *st);
1704 
1705 /* A state can be freed if it is no longer referenced:
1706  * - is in the env->free_list;
1707  * - has no children states;
1708  */
maybe_free_verifier_state(struct bpf_verifier_env * env,struct bpf_verifier_state_list * sl)1709 static void maybe_free_verifier_state(struct bpf_verifier_env *env,
1710 				      struct bpf_verifier_state_list *sl)
1711 {
1712 	if (!sl->in_free_list
1713 	    || sl->state.branches != 0
1714 	    || incomplete_read_marks(env, &sl->state))
1715 		return;
1716 	list_del(&sl->node);
1717 	free_verifier_state(&sl->state, false);
1718 	kfree(sl);
1719 	env->free_list_size--;
1720 }
1721 
1722 /* copy verifier state from src to dst growing dst stack space
1723  * when necessary to accommodate larger src stack
1724  */
copy_func_state(struct bpf_func_state * dst,const struct bpf_func_state * src)1725 static int copy_func_state(struct bpf_func_state *dst,
1726 			   const struct bpf_func_state *src)
1727 {
1728 	memcpy(dst, src, offsetof(struct bpf_func_state, stack));
1729 	return copy_stack_state(dst, src);
1730 }
1731 
copy_verifier_state(struct bpf_verifier_state * dst_state,const struct bpf_verifier_state * src)1732 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1733 			       const struct bpf_verifier_state *src)
1734 {
1735 	struct bpf_func_state *dst;
1736 	int i, err;
1737 
1738 	dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1739 					  src->jmp_history_cnt, sizeof(*dst_state->jmp_history),
1740 					  GFP_KERNEL_ACCOUNT);
1741 	if (!dst_state->jmp_history)
1742 		return -ENOMEM;
1743 	dst_state->jmp_history_cnt = src->jmp_history_cnt;
1744 
1745 	/* if dst has more stack frames then src frame, free them, this is also
1746 	 * necessary in case of exceptional exits using bpf_throw.
1747 	 */
1748 	for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1749 		free_func_state(dst_state->frame[i]);
1750 		dst_state->frame[i] = NULL;
1751 	}
1752 	err = copy_reference_state(dst_state, src);
1753 	if (err)
1754 		return err;
1755 	dst_state->speculative = src->speculative;
1756 	dst_state->in_sleepable = src->in_sleepable;
1757 	dst_state->curframe = src->curframe;
1758 	dst_state->branches = src->branches;
1759 	dst_state->parent = src->parent;
1760 	dst_state->first_insn_idx = src->first_insn_idx;
1761 	dst_state->last_insn_idx = src->last_insn_idx;
1762 	dst_state->dfs_depth = src->dfs_depth;
1763 	dst_state->callback_unroll_depth = src->callback_unroll_depth;
1764 	dst_state->may_goto_depth = src->may_goto_depth;
1765 	dst_state->equal_state = src->equal_state;
1766 	for (i = 0; i <= src->curframe; i++) {
1767 		dst = dst_state->frame[i];
1768 		if (!dst) {
1769 			dst = kzalloc(sizeof(*dst), GFP_KERNEL_ACCOUNT);
1770 			if (!dst)
1771 				return -ENOMEM;
1772 			dst_state->frame[i] = dst;
1773 		}
1774 		err = copy_func_state(dst, src->frame[i]);
1775 		if (err)
1776 			return err;
1777 	}
1778 	return 0;
1779 }
1780 
state_htab_size(struct bpf_verifier_env * env)1781 static u32 state_htab_size(struct bpf_verifier_env *env)
1782 {
1783 	return env->prog->len;
1784 }
1785 
explored_state(struct bpf_verifier_env * env,int idx)1786 static struct list_head *explored_state(struct bpf_verifier_env *env, int idx)
1787 {
1788 	struct bpf_verifier_state *cur = env->cur_state;
1789 	struct bpf_func_state *state = cur->frame[cur->curframe];
1790 
1791 	return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
1792 }
1793 
same_callsites(struct bpf_verifier_state * a,struct bpf_verifier_state * b)1794 static bool same_callsites(struct bpf_verifier_state *a, struct bpf_verifier_state *b)
1795 {
1796 	int fr;
1797 
1798 	if (a->curframe != b->curframe)
1799 		return false;
1800 
1801 	for (fr = a->curframe; fr >= 0; fr--)
1802 		if (a->frame[fr]->callsite != b->frame[fr]->callsite)
1803 			return false;
1804 
1805 	return true;
1806 }
1807 
1808 /* Return IP for a given frame in a call stack */
frame_insn_idx(struct bpf_verifier_state * st,u32 frame)1809 static u32 frame_insn_idx(struct bpf_verifier_state *st, u32 frame)
1810 {
1811 	return frame == st->curframe
1812 	       ? st->insn_idx
1813 	       : st->frame[frame + 1]->callsite;
1814 }
1815 
1816 /* For state @st look for a topmost frame with frame_insn_idx() in some SCC,
1817  * if such frame exists form a corresponding @callchain as an array of
1818  * call sites leading to this frame and SCC id.
1819  * E.g.:
1820  *
1821  *    void foo()  { A: loop {... SCC#1 ...}; }
1822  *    void bar()  { B: loop { C: foo(); ... SCC#2 ... }
1823  *                  D: loop { E: foo(); ... SCC#3 ... } }
1824  *    void main() { F: bar(); }
1825  *
1826  * @callchain at (A) would be either (F,SCC#2) or (F,SCC#3) depending
1827  * on @st frame call sites being (F,C,A) or (F,E,A).
1828  */
compute_scc_callchain(struct bpf_verifier_env * env,struct bpf_verifier_state * st,struct bpf_scc_callchain * callchain)1829 static bool compute_scc_callchain(struct bpf_verifier_env *env,
1830 				  struct bpf_verifier_state *st,
1831 				  struct bpf_scc_callchain *callchain)
1832 {
1833 	u32 i, scc, insn_idx;
1834 
1835 	memset(callchain, 0, sizeof(*callchain));
1836 	for (i = 0; i <= st->curframe; i++) {
1837 		insn_idx = frame_insn_idx(st, i);
1838 		scc = env->insn_aux_data[insn_idx].scc;
1839 		if (scc) {
1840 			callchain->scc = scc;
1841 			break;
1842 		} else if (i < st->curframe) {
1843 			callchain->callsites[i] = insn_idx;
1844 		} else {
1845 			return false;
1846 		}
1847 	}
1848 	return true;
1849 }
1850 
1851 /* Check if bpf_scc_visit instance for @callchain exists. */
scc_visit_lookup(struct bpf_verifier_env * env,struct bpf_scc_callchain * callchain)1852 static struct bpf_scc_visit *scc_visit_lookup(struct bpf_verifier_env *env,
1853 					      struct bpf_scc_callchain *callchain)
1854 {
1855 	struct bpf_scc_info *info = env->scc_info[callchain->scc];
1856 	struct bpf_scc_visit *visits = info->visits;
1857 	u32 i;
1858 
1859 	if (!info)
1860 		return NULL;
1861 	for (i = 0; i < info->num_visits; i++)
1862 		if (memcmp(callchain, &visits[i].callchain, sizeof(*callchain)) == 0)
1863 			return &visits[i];
1864 	return NULL;
1865 }
1866 
1867 /* Allocate a new bpf_scc_visit instance corresponding to @callchain.
1868  * Allocated instances are alive for a duration of the do_check_common()
1869  * call and are freed by free_states().
1870  */
scc_visit_alloc(struct bpf_verifier_env * env,struct bpf_scc_callchain * callchain)1871 static struct bpf_scc_visit *scc_visit_alloc(struct bpf_verifier_env *env,
1872 					     struct bpf_scc_callchain *callchain)
1873 {
1874 	struct bpf_scc_visit *visit;
1875 	struct bpf_scc_info *info;
1876 	u32 scc, num_visits;
1877 	u64 new_sz;
1878 
1879 	scc = callchain->scc;
1880 	info = env->scc_info[scc];
1881 	num_visits = info ? info->num_visits : 0;
1882 	new_sz = sizeof(*info) + sizeof(struct bpf_scc_visit) * (num_visits + 1);
1883 	info = kvrealloc(env->scc_info[scc], new_sz, GFP_KERNEL_ACCOUNT);
1884 	if (!info)
1885 		return NULL;
1886 	env->scc_info[scc] = info;
1887 	info->num_visits = num_visits + 1;
1888 	visit = &info->visits[num_visits];
1889 	memset(visit, 0, sizeof(*visit));
1890 	memcpy(&visit->callchain, callchain, sizeof(*callchain));
1891 	return visit;
1892 }
1893 
1894 /* Form a string '(callsite#1,callsite#2,...,scc)' in env->tmp_str_buf */
format_callchain(struct bpf_verifier_env * env,struct bpf_scc_callchain * callchain)1895 static char *format_callchain(struct bpf_verifier_env *env, struct bpf_scc_callchain *callchain)
1896 {
1897 	char *buf = env->tmp_str_buf;
1898 	int i, delta = 0;
1899 
1900 	delta += snprintf(buf + delta, TMP_STR_BUF_LEN - delta, "(");
1901 	for (i = 0; i < ARRAY_SIZE(callchain->callsites); i++) {
1902 		if (!callchain->callsites[i])
1903 			break;
1904 		delta += snprintf(buf + delta, TMP_STR_BUF_LEN - delta, "%u,",
1905 				  callchain->callsites[i]);
1906 	}
1907 	delta += snprintf(buf + delta, TMP_STR_BUF_LEN - delta, "%u)", callchain->scc);
1908 	return env->tmp_str_buf;
1909 }
1910 
1911 /* If callchain for @st exists (@st is in some SCC), ensure that
1912  * bpf_scc_visit instance for this callchain exists.
1913  * If instance does not exist or is empty, assign visit->entry_state to @st.
1914  */
maybe_enter_scc(struct bpf_verifier_env * env,struct bpf_verifier_state * st)1915 static int maybe_enter_scc(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1916 {
1917 	struct bpf_scc_callchain *callchain = &env->callchain_buf;
1918 	struct bpf_scc_visit *visit;
1919 
1920 	if (!compute_scc_callchain(env, st, callchain))
1921 		return 0;
1922 	visit = scc_visit_lookup(env, callchain);
1923 	visit = visit ?: scc_visit_alloc(env, callchain);
1924 	if (!visit)
1925 		return -ENOMEM;
1926 	if (!visit->entry_state) {
1927 		visit->entry_state = st;
1928 		if (env->log.level & BPF_LOG_LEVEL2)
1929 			verbose(env, "SCC enter %s\n", format_callchain(env, callchain));
1930 	}
1931 	return 0;
1932 }
1933 
1934 static int propagate_backedges(struct bpf_verifier_env *env, struct bpf_scc_visit *visit);
1935 
1936 /* If callchain for @st exists (@st is in some SCC), make it empty:
1937  * - set visit->entry_state to NULL;
1938  * - flush accumulated backedges.
1939  */
maybe_exit_scc(struct bpf_verifier_env * env,struct bpf_verifier_state * st)1940 static int maybe_exit_scc(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1941 {
1942 	struct bpf_scc_callchain *callchain = &env->callchain_buf;
1943 	struct bpf_scc_visit *visit;
1944 
1945 	if (!compute_scc_callchain(env, st, callchain))
1946 		return 0;
1947 	visit = scc_visit_lookup(env, callchain);
1948 	if (!visit) {
1949 		verifier_bug(env, "scc exit: no visit info for call chain %s",
1950 			     format_callchain(env, callchain));
1951 		return -EFAULT;
1952 	}
1953 	if (visit->entry_state != st)
1954 		return 0;
1955 	if (env->log.level & BPF_LOG_LEVEL2)
1956 		verbose(env, "SCC exit %s\n", format_callchain(env, callchain));
1957 	visit->entry_state = NULL;
1958 	env->num_backedges -= visit->num_backedges;
1959 	visit->num_backedges = 0;
1960 	update_peak_states(env);
1961 	return propagate_backedges(env, visit);
1962 }
1963 
1964 /* Lookup an bpf_scc_visit instance corresponding to @st callchain
1965  * and add @backedge to visit->backedges. @st callchain must exist.
1966  */
add_scc_backedge(struct bpf_verifier_env * env,struct bpf_verifier_state * st,struct bpf_scc_backedge * backedge)1967 static int add_scc_backedge(struct bpf_verifier_env *env,
1968 			    struct bpf_verifier_state *st,
1969 			    struct bpf_scc_backedge *backedge)
1970 {
1971 	struct bpf_scc_callchain *callchain = &env->callchain_buf;
1972 	struct bpf_scc_visit *visit;
1973 
1974 	if (!compute_scc_callchain(env, st, callchain)) {
1975 		verifier_bug(env, "add backedge: no SCC in verification path, insn_idx %d",
1976 			     st->insn_idx);
1977 		return -EFAULT;
1978 	}
1979 	visit = scc_visit_lookup(env, callchain);
1980 	if (!visit) {
1981 		verifier_bug(env, "add backedge: no visit info for call chain %s",
1982 			     format_callchain(env, callchain));
1983 		return -EFAULT;
1984 	}
1985 	if (env->log.level & BPF_LOG_LEVEL2)
1986 		verbose(env, "SCC backedge %s\n", format_callchain(env, callchain));
1987 	backedge->next = visit->backedges;
1988 	visit->backedges = backedge;
1989 	visit->num_backedges++;
1990 	env->num_backedges++;
1991 	update_peak_states(env);
1992 	return 0;
1993 }
1994 
1995 /* bpf_reg_state->live marks for registers in a state @st are incomplete,
1996  * if state @st is in some SCC and not all execution paths starting at this
1997  * SCC are fully explored.
1998  */
incomplete_read_marks(struct bpf_verifier_env * env,struct bpf_verifier_state * st)1999 static bool incomplete_read_marks(struct bpf_verifier_env *env,
2000 				  struct bpf_verifier_state *st)
2001 {
2002 	struct bpf_scc_callchain *callchain = &env->callchain_buf;
2003 	struct bpf_scc_visit *visit;
2004 
2005 	if (!compute_scc_callchain(env, st, callchain))
2006 		return false;
2007 	visit = scc_visit_lookup(env, callchain);
2008 	if (!visit)
2009 		return false;
2010 	return !!visit->backedges;
2011 }
2012 
free_backedges(struct bpf_scc_visit * visit)2013 static void free_backedges(struct bpf_scc_visit *visit)
2014 {
2015 	struct bpf_scc_backedge *backedge, *next;
2016 
2017 	for (backedge = visit->backedges; backedge; backedge = next) {
2018 		free_verifier_state(&backedge->state, false);
2019 		next = backedge->next;
2020 		kvfree(backedge);
2021 	}
2022 	visit->backedges = NULL;
2023 }
2024 
update_branch_counts(struct bpf_verifier_env * env,struct bpf_verifier_state * st)2025 static int update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
2026 {
2027 	struct bpf_verifier_state_list *sl = NULL, *parent_sl;
2028 	struct bpf_verifier_state *parent;
2029 	int err;
2030 
2031 	while (st) {
2032 		u32 br = --st->branches;
2033 
2034 		/* verifier_bug_if(br > 1, ...) technically makes sense here,
2035 		 * but see comment in push_stack(), hence:
2036 		 */
2037 		verifier_bug_if((int)br < 0, env, "%s:branches_to_explore=%d", __func__, br);
2038 		if (br)
2039 			break;
2040 		err = maybe_exit_scc(env, st);
2041 		if (err)
2042 			return err;
2043 		parent = st->parent;
2044 		parent_sl = state_parent_as_list(st);
2045 		if (sl)
2046 			maybe_free_verifier_state(env, sl);
2047 		st = parent;
2048 		sl = parent_sl;
2049 	}
2050 	return 0;
2051 }
2052 
pop_stack(struct bpf_verifier_env * env,int * prev_insn_idx,int * insn_idx,bool pop_log)2053 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
2054 		     int *insn_idx, bool pop_log)
2055 {
2056 	struct bpf_verifier_state *cur = env->cur_state;
2057 	struct bpf_verifier_stack_elem *elem, *head = env->head;
2058 	int err;
2059 
2060 	if (env->head == NULL)
2061 		return -ENOENT;
2062 
2063 	if (cur) {
2064 		err = copy_verifier_state(cur, &head->st);
2065 		if (err)
2066 			return err;
2067 	}
2068 	if (pop_log)
2069 		bpf_vlog_reset(&env->log, head->log_pos);
2070 	if (insn_idx)
2071 		*insn_idx = head->insn_idx;
2072 	if (prev_insn_idx)
2073 		*prev_insn_idx = head->prev_insn_idx;
2074 	elem = head->next;
2075 	free_verifier_state(&head->st, false);
2076 	kfree(head);
2077 	env->head = elem;
2078 	env->stack_size--;
2079 	return 0;
2080 }
2081 
error_recoverable_with_nospec(int err)2082 static bool error_recoverable_with_nospec(int err)
2083 {
2084 	/* Should only return true for non-fatal errors that are allowed to
2085 	 * occur during speculative verification. For these we can insert a
2086 	 * nospec and the program might still be accepted. Do not include
2087 	 * something like ENOMEM because it is likely to re-occur for the next
2088 	 * architectural path once it has been recovered-from in all speculative
2089 	 * paths.
2090 	 */
2091 	return err == -EPERM || err == -EACCES || err == -EINVAL;
2092 }
2093 
push_stack(struct bpf_verifier_env * env,int insn_idx,int prev_insn_idx,bool speculative)2094 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
2095 					     int insn_idx, int prev_insn_idx,
2096 					     bool speculative)
2097 {
2098 	struct bpf_verifier_state *cur = env->cur_state;
2099 	struct bpf_verifier_stack_elem *elem;
2100 	int err;
2101 
2102 	elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL_ACCOUNT);
2103 	if (!elem)
2104 		return NULL;
2105 
2106 	elem->insn_idx = insn_idx;
2107 	elem->prev_insn_idx = prev_insn_idx;
2108 	elem->next = env->head;
2109 	elem->log_pos = env->log.end_pos;
2110 	env->head = elem;
2111 	env->stack_size++;
2112 	err = copy_verifier_state(&elem->st, cur);
2113 	if (err)
2114 		return NULL;
2115 	elem->st.speculative |= speculative;
2116 	if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2117 		verbose(env, "The sequence of %d jumps is too complex.\n",
2118 			env->stack_size);
2119 		return NULL;
2120 	}
2121 	if (elem->st.parent) {
2122 		++elem->st.parent->branches;
2123 		/* WARN_ON(branches > 2) technically makes sense here,
2124 		 * but
2125 		 * 1. speculative states will bump 'branches' for non-branch
2126 		 * instructions
2127 		 * 2. is_state_visited() heuristics may decide not to create
2128 		 * a new state for a sequence of branches and all such current
2129 		 * and cloned states will be pointing to a single parent state
2130 		 * which might have large 'branches' count.
2131 		 */
2132 	}
2133 	return &elem->st;
2134 }
2135 
2136 #define CALLER_SAVED_REGS 6
2137 static const int caller_saved[CALLER_SAVED_REGS] = {
2138 	BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
2139 };
2140 
2141 /* This helper doesn't clear reg->id */
___mark_reg_known(struct bpf_reg_state * reg,u64 imm)2142 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
2143 {
2144 	reg->var_off = tnum_const(imm);
2145 	reg->smin_value = (s64)imm;
2146 	reg->smax_value = (s64)imm;
2147 	reg->umin_value = imm;
2148 	reg->umax_value = imm;
2149 
2150 	reg->s32_min_value = (s32)imm;
2151 	reg->s32_max_value = (s32)imm;
2152 	reg->u32_min_value = (u32)imm;
2153 	reg->u32_max_value = (u32)imm;
2154 }
2155 
2156 /* Mark the unknown part of a register (variable offset or scalar value) as
2157  * known to have the value @imm.
2158  */
__mark_reg_known(struct bpf_reg_state * reg,u64 imm)2159 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
2160 {
2161 	/* Clear off and union(map_ptr, range) */
2162 	memset(((u8 *)reg) + sizeof(reg->type), 0,
2163 	       offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
2164 	reg->id = 0;
2165 	reg->ref_obj_id = 0;
2166 	___mark_reg_known(reg, imm);
2167 }
2168 
__mark_reg32_known(struct bpf_reg_state * reg,u64 imm)2169 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
2170 {
2171 	reg->var_off = tnum_const_subreg(reg->var_off, imm);
2172 	reg->s32_min_value = (s32)imm;
2173 	reg->s32_max_value = (s32)imm;
2174 	reg->u32_min_value = (u32)imm;
2175 	reg->u32_max_value = (u32)imm;
2176 }
2177 
2178 /* Mark the 'variable offset' part of a register as zero.  This should be
2179  * used only on registers holding a pointer type.
2180  */
__mark_reg_known_zero(struct bpf_reg_state * reg)2181 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
2182 {
2183 	__mark_reg_known(reg, 0);
2184 }
2185 
__mark_reg_const_zero(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)2186 static void __mark_reg_const_zero(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
2187 {
2188 	__mark_reg_known(reg, 0);
2189 	reg->type = SCALAR_VALUE;
2190 	/* all scalars are assumed imprecise initially (unless unprivileged,
2191 	 * in which case everything is forced to be precise)
2192 	 */
2193 	reg->precise = !env->bpf_capable;
2194 }
2195 
mark_reg_known_zero(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)2196 static void mark_reg_known_zero(struct bpf_verifier_env *env,
2197 				struct bpf_reg_state *regs, u32 regno)
2198 {
2199 	if (WARN_ON(regno >= MAX_BPF_REG)) {
2200 		verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
2201 		/* Something bad happened, let's kill all regs */
2202 		for (regno = 0; regno < MAX_BPF_REG; regno++)
2203 			__mark_reg_not_init(env, regs + regno);
2204 		return;
2205 	}
2206 	__mark_reg_known_zero(regs + regno);
2207 }
2208 
__mark_dynptr_reg(struct bpf_reg_state * reg,enum bpf_dynptr_type type,bool first_slot,int dynptr_id)2209 static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
2210 			      bool first_slot, int dynptr_id)
2211 {
2212 	/* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
2213 	 * callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
2214 	 * set it unconditionally as it is ignored for STACK_DYNPTR anyway.
2215 	 */
2216 	__mark_reg_known_zero(reg);
2217 	reg->type = CONST_PTR_TO_DYNPTR;
2218 	/* Give each dynptr a unique id to uniquely associate slices to it. */
2219 	reg->id = dynptr_id;
2220 	reg->dynptr.type = type;
2221 	reg->dynptr.first_slot = first_slot;
2222 }
2223 
mark_ptr_not_null_reg(struct bpf_reg_state * reg)2224 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
2225 {
2226 	if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
2227 		const struct bpf_map *map = reg->map_ptr;
2228 
2229 		if (map->inner_map_meta) {
2230 			reg->type = CONST_PTR_TO_MAP;
2231 			reg->map_ptr = map->inner_map_meta;
2232 			/* transfer reg's id which is unique for every map_lookup_elem
2233 			 * as UID of the inner map.
2234 			 */
2235 			if (btf_record_has_field(map->inner_map_meta->record, BPF_TIMER))
2236 				reg->map_uid = reg->id;
2237 			if (btf_record_has_field(map->inner_map_meta->record, BPF_WORKQUEUE))
2238 				reg->map_uid = reg->id;
2239 		} else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
2240 			reg->type = PTR_TO_XDP_SOCK;
2241 		} else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
2242 			   map->map_type == BPF_MAP_TYPE_SOCKHASH) {
2243 			reg->type = PTR_TO_SOCKET;
2244 		} else {
2245 			reg->type = PTR_TO_MAP_VALUE;
2246 		}
2247 		return;
2248 	}
2249 
2250 	reg->type &= ~PTR_MAYBE_NULL;
2251 }
2252 
mark_reg_graph_node(struct bpf_reg_state * regs,u32 regno,struct btf_field_graph_root * ds_head)2253 static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
2254 				struct btf_field_graph_root *ds_head)
2255 {
2256 	__mark_reg_known_zero(&regs[regno]);
2257 	regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
2258 	regs[regno].btf = ds_head->btf;
2259 	regs[regno].btf_id = ds_head->value_btf_id;
2260 	regs[regno].off = ds_head->node_offset;
2261 }
2262 
reg_is_pkt_pointer(const struct bpf_reg_state * reg)2263 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
2264 {
2265 	return type_is_pkt_pointer(reg->type);
2266 }
2267 
reg_is_pkt_pointer_any(const struct bpf_reg_state * reg)2268 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2269 {
2270 	return reg_is_pkt_pointer(reg) ||
2271 	       reg->type == PTR_TO_PACKET_END;
2272 }
2273 
reg_is_dynptr_slice_pkt(const struct bpf_reg_state * reg)2274 static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2275 {
2276 	return base_type(reg->type) == PTR_TO_MEM &&
2277 		(reg->type & DYNPTR_TYPE_SKB || reg->type & DYNPTR_TYPE_XDP);
2278 }
2279 
2280 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
reg_is_init_pkt_pointer(const struct bpf_reg_state * reg,enum bpf_reg_type which)2281 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2282 				    enum bpf_reg_type which)
2283 {
2284 	/* The register can already have a range from prior markings.
2285 	 * This is fine as long as it hasn't been advanced from its
2286 	 * origin.
2287 	 */
2288 	return reg->type == which &&
2289 	       reg->id == 0 &&
2290 	       reg->off == 0 &&
2291 	       tnum_equals_const(reg->var_off, 0);
2292 }
2293 
2294 /* Reset the min/max bounds of a register */
__mark_reg_unbounded(struct bpf_reg_state * reg)2295 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2296 {
2297 	reg->smin_value = S64_MIN;
2298 	reg->smax_value = S64_MAX;
2299 	reg->umin_value = 0;
2300 	reg->umax_value = U64_MAX;
2301 
2302 	reg->s32_min_value = S32_MIN;
2303 	reg->s32_max_value = S32_MAX;
2304 	reg->u32_min_value = 0;
2305 	reg->u32_max_value = U32_MAX;
2306 }
2307 
__mark_reg64_unbounded(struct bpf_reg_state * reg)2308 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2309 {
2310 	reg->smin_value = S64_MIN;
2311 	reg->smax_value = S64_MAX;
2312 	reg->umin_value = 0;
2313 	reg->umax_value = U64_MAX;
2314 }
2315 
__mark_reg32_unbounded(struct bpf_reg_state * reg)2316 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2317 {
2318 	reg->s32_min_value = S32_MIN;
2319 	reg->s32_max_value = S32_MAX;
2320 	reg->u32_min_value = 0;
2321 	reg->u32_max_value = U32_MAX;
2322 }
2323 
__update_reg32_bounds(struct bpf_reg_state * reg)2324 static void __update_reg32_bounds(struct bpf_reg_state *reg)
2325 {
2326 	struct tnum var32_off = tnum_subreg(reg->var_off);
2327 
2328 	/* min signed is max(sign bit) | min(other bits) */
2329 	reg->s32_min_value = max_t(s32, reg->s32_min_value,
2330 			var32_off.value | (var32_off.mask & S32_MIN));
2331 	/* max signed is min(sign bit) | max(other bits) */
2332 	reg->s32_max_value = min_t(s32, reg->s32_max_value,
2333 			var32_off.value | (var32_off.mask & S32_MAX));
2334 	reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2335 	reg->u32_max_value = min(reg->u32_max_value,
2336 				 (u32)(var32_off.value | var32_off.mask));
2337 }
2338 
__update_reg64_bounds(struct bpf_reg_state * reg)2339 static void __update_reg64_bounds(struct bpf_reg_state *reg)
2340 {
2341 	/* min signed is max(sign bit) | min(other bits) */
2342 	reg->smin_value = max_t(s64, reg->smin_value,
2343 				reg->var_off.value | (reg->var_off.mask & S64_MIN));
2344 	/* max signed is min(sign bit) | max(other bits) */
2345 	reg->smax_value = min_t(s64, reg->smax_value,
2346 				reg->var_off.value | (reg->var_off.mask & S64_MAX));
2347 	reg->umin_value = max(reg->umin_value, reg->var_off.value);
2348 	reg->umax_value = min(reg->umax_value,
2349 			      reg->var_off.value | reg->var_off.mask);
2350 }
2351 
__update_reg_bounds(struct bpf_reg_state * reg)2352 static void __update_reg_bounds(struct bpf_reg_state *reg)
2353 {
2354 	__update_reg32_bounds(reg);
2355 	__update_reg64_bounds(reg);
2356 }
2357 
2358 /* Uses signed min/max values to inform unsigned, and vice-versa */
__reg32_deduce_bounds(struct bpf_reg_state * reg)2359 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2360 {
2361 	/* If upper 32 bits of u64/s64 range don't change, we can use lower 32
2362 	 * bits to improve our u32/s32 boundaries.
2363 	 *
2364 	 * E.g., the case where we have upper 32 bits as zero ([10, 20] in
2365 	 * u64) is pretty trivial, it's obvious that in u32 we'll also have
2366 	 * [10, 20] range. But this property holds for any 64-bit range as
2367 	 * long as upper 32 bits in that entire range of values stay the same.
2368 	 *
2369 	 * E.g., u64 range [0x10000000A, 0x10000000F] ([4294967306, 4294967311]
2370 	 * in decimal) has the same upper 32 bits throughout all the values in
2371 	 * that range. As such, lower 32 bits form a valid [0xA, 0xF] ([10, 15])
2372 	 * range.
2373 	 *
2374 	 * Note also, that [0xA, 0xF] is a valid range both in u32 and in s32,
2375 	 * following the rules outlined below about u64/s64 correspondence
2376 	 * (which equally applies to u32 vs s32 correspondence). In general it
2377 	 * depends on actual hexadecimal values of 32-bit range. They can form
2378 	 * only valid u32, or only valid s32 ranges in some cases.
2379 	 *
2380 	 * So we use all these insights to derive bounds for subregisters here.
2381 	 */
2382 	if ((reg->umin_value >> 32) == (reg->umax_value >> 32)) {
2383 		/* u64 to u32 casting preserves validity of low 32 bits as
2384 		 * a range, if upper 32 bits are the same
2385 		 */
2386 		reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)reg->umin_value);
2387 		reg->u32_max_value = min_t(u32, reg->u32_max_value, (u32)reg->umax_value);
2388 
2389 		if ((s32)reg->umin_value <= (s32)reg->umax_value) {
2390 			reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->umin_value);
2391 			reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->umax_value);
2392 		}
2393 	}
2394 	if ((reg->smin_value >> 32) == (reg->smax_value >> 32)) {
2395 		/* low 32 bits should form a proper u32 range */
2396 		if ((u32)reg->smin_value <= (u32)reg->smax_value) {
2397 			reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)reg->smin_value);
2398 			reg->u32_max_value = min_t(u32, reg->u32_max_value, (u32)reg->smax_value);
2399 		}
2400 		/* low 32 bits should form a proper s32 range */
2401 		if ((s32)reg->smin_value <= (s32)reg->smax_value) {
2402 			reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->smin_value);
2403 			reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->smax_value);
2404 		}
2405 	}
2406 	/* Special case where upper bits form a small sequence of two
2407 	 * sequential numbers (in 32-bit unsigned space, so 0xffffffff to
2408 	 * 0x00000000 is also valid), while lower bits form a proper s32 range
2409 	 * going from negative numbers to positive numbers. E.g., let's say we
2410 	 * have s64 range [-1, 1] ([0xffffffffffffffff, 0x0000000000000001]).
2411 	 * Possible s64 values are {-1, 0, 1} ({0xffffffffffffffff,
2412 	 * 0x0000000000000000, 0x00000000000001}). Ignoring upper 32 bits,
2413 	 * we still get a valid s32 range [-1, 1] ([0xffffffff, 0x00000001]).
2414 	 * Note that it doesn't have to be 0xffffffff going to 0x00000000 in
2415 	 * upper 32 bits. As a random example, s64 range
2416 	 * [0xfffffff0fffffff0; 0xfffffff100000010], forms a valid s32 range
2417 	 * [-16, 16] ([0xfffffff0; 0x00000010]) in its 32 bit subregister.
2418 	 */
2419 	if ((u32)(reg->umin_value >> 32) + 1 == (u32)(reg->umax_value >> 32) &&
2420 	    (s32)reg->umin_value < 0 && (s32)reg->umax_value >= 0) {
2421 		reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->umin_value);
2422 		reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->umax_value);
2423 	}
2424 	if ((u32)(reg->smin_value >> 32) + 1 == (u32)(reg->smax_value >> 32) &&
2425 	    (s32)reg->smin_value < 0 && (s32)reg->smax_value >= 0) {
2426 		reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->smin_value);
2427 		reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->smax_value);
2428 	}
2429 	/* if u32 range forms a valid s32 range (due to matching sign bit),
2430 	 * try to learn from that
2431 	 */
2432 	if ((s32)reg->u32_min_value <= (s32)reg->u32_max_value) {
2433 		reg->s32_min_value = max_t(s32, reg->s32_min_value, reg->u32_min_value);
2434 		reg->s32_max_value = min_t(s32, reg->s32_max_value, reg->u32_max_value);
2435 	}
2436 	/* If we cannot cross the sign boundary, then signed and unsigned bounds
2437 	 * are the same, so combine.  This works even in the negative case, e.g.
2438 	 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2439 	 */
2440 	if ((u32)reg->s32_min_value <= (u32)reg->s32_max_value) {
2441 		reg->u32_min_value = max_t(u32, reg->s32_min_value, reg->u32_min_value);
2442 		reg->u32_max_value = min_t(u32, reg->s32_max_value, reg->u32_max_value);
2443 	}
2444 }
2445 
__reg64_deduce_bounds(struct bpf_reg_state * reg)2446 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2447 {
2448 	/* If u64 range forms a valid s64 range (due to matching sign bit),
2449 	 * try to learn from that. Let's do a bit of ASCII art to see when
2450 	 * this is happening. Let's take u64 range first:
2451 	 *
2452 	 * 0             0x7fffffffffffffff 0x8000000000000000        U64_MAX
2453 	 * |-------------------------------|--------------------------------|
2454 	 *
2455 	 * Valid u64 range is formed when umin and umax are anywhere in the
2456 	 * range [0, U64_MAX], and umin <= umax. u64 case is simple and
2457 	 * straightforward. Let's see how s64 range maps onto the same range
2458 	 * of values, annotated below the line for comparison:
2459 	 *
2460 	 * 0             0x7fffffffffffffff 0x8000000000000000        U64_MAX
2461 	 * |-------------------------------|--------------------------------|
2462 	 * 0                        S64_MAX S64_MIN                        -1
2463 	 *
2464 	 * So s64 values basically start in the middle and they are logically
2465 	 * contiguous to the right of it, wrapping around from -1 to 0, and
2466 	 * then finishing as S64_MAX (0x7fffffffffffffff) right before
2467 	 * S64_MIN. We can try drawing the continuity of u64 vs s64 values
2468 	 * more visually as mapped to sign-agnostic range of hex values.
2469 	 *
2470 	 *  u64 start                                               u64 end
2471 	 *  _______________________________________________________________
2472 	 * /                                                               \
2473 	 * 0             0x7fffffffffffffff 0x8000000000000000        U64_MAX
2474 	 * |-------------------------------|--------------------------------|
2475 	 * 0                        S64_MAX S64_MIN                        -1
2476 	 *                                / \
2477 	 * >------------------------------   ------------------------------->
2478 	 * s64 continues...        s64 end   s64 start          s64 "midpoint"
2479 	 *
2480 	 * What this means is that, in general, we can't always derive
2481 	 * something new about u64 from any random s64 range, and vice versa.
2482 	 *
2483 	 * But we can do that in two particular cases. One is when entire
2484 	 * u64/s64 range is *entirely* contained within left half of the above
2485 	 * diagram or when it is *entirely* contained in the right half. I.e.:
2486 	 *
2487 	 * |-------------------------------|--------------------------------|
2488 	 *     ^                   ^            ^                 ^
2489 	 *     A                   B            C                 D
2490 	 *
2491 	 * [A, B] and [C, D] are contained entirely in their respective halves
2492 	 * and form valid contiguous ranges as both u64 and s64 values. [A, B]
2493 	 * will be non-negative both as u64 and s64 (and in fact it will be
2494 	 * identical ranges no matter the signedness). [C, D] treated as s64
2495 	 * will be a range of negative values, while in u64 it will be
2496 	 * non-negative range of values larger than 0x8000000000000000.
2497 	 *
2498 	 * Now, any other range here can't be represented in both u64 and s64
2499 	 * simultaneously. E.g., [A, C], [A, D], [B, C], [B, D] are valid
2500 	 * contiguous u64 ranges, but they are discontinuous in s64. [B, C]
2501 	 * in s64 would be properly presented as [S64_MIN, C] and [B, S64_MAX],
2502 	 * for example. Similarly, valid s64 range [D, A] (going from negative
2503 	 * to positive values), would be two separate [D, U64_MAX] and [0, A]
2504 	 * ranges as u64. Currently reg_state can't represent two segments per
2505 	 * numeric domain, so in such situations we can only derive maximal
2506 	 * possible range ([0, U64_MAX] for u64, and [S64_MIN, S64_MAX] for s64).
2507 	 *
2508 	 * So we use these facts to derive umin/umax from smin/smax and vice
2509 	 * versa only if they stay within the same "half". This is equivalent
2510 	 * to checking sign bit: lower half will have sign bit as zero, upper
2511 	 * half have sign bit 1. Below in code we simplify this by just
2512 	 * casting umin/umax as smin/smax and checking if they form valid
2513 	 * range, and vice versa. Those are equivalent checks.
2514 	 */
2515 	if ((s64)reg->umin_value <= (s64)reg->umax_value) {
2516 		reg->smin_value = max_t(s64, reg->smin_value, reg->umin_value);
2517 		reg->smax_value = min_t(s64, reg->smax_value, reg->umax_value);
2518 	}
2519 	/* If we cannot cross the sign boundary, then signed and unsigned bounds
2520 	 * are the same, so combine.  This works even in the negative case, e.g.
2521 	 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2522 	 */
2523 	if ((u64)reg->smin_value <= (u64)reg->smax_value) {
2524 		reg->umin_value = max_t(u64, reg->smin_value, reg->umin_value);
2525 		reg->umax_value = min_t(u64, reg->smax_value, reg->umax_value);
2526 	} else {
2527 		/* If the s64 range crosses the sign boundary, then it's split
2528 		 * between the beginning and end of the U64 domain. In that
2529 		 * case, we can derive new bounds if the u64 range overlaps
2530 		 * with only one end of the s64 range.
2531 		 *
2532 		 * In the following example, the u64 range overlaps only with
2533 		 * positive portion of the s64 range.
2534 		 *
2535 		 * 0                                                   U64_MAX
2536 		 * |  [xxxxxxxxxxxxxx u64 range xxxxxxxxxxxxxx]              |
2537 		 * |----------------------------|----------------------------|
2538 		 * |xxxxx s64 range xxxxxxxxx]                       [xxxxxxx|
2539 		 * 0                     S64_MAX S64_MIN                    -1
2540 		 *
2541 		 * We can thus derive the following new s64 and u64 ranges.
2542 		 *
2543 		 * 0                                                   U64_MAX
2544 		 * |  [xxxxxx u64 range xxxxx]                               |
2545 		 * |----------------------------|----------------------------|
2546 		 * |  [xxxxxx s64 range xxxxx]                               |
2547 		 * 0                     S64_MAX S64_MIN                    -1
2548 		 *
2549 		 * If they overlap in two places, we can't derive anything
2550 		 * because reg_state can't represent two ranges per numeric
2551 		 * domain.
2552 		 *
2553 		 * 0                                                   U64_MAX
2554 		 * |  [xxxxxxxxxxxxxxxxx u64 range xxxxxxxxxxxxxxxxx]        |
2555 		 * |----------------------------|----------------------------|
2556 		 * |xxxxx s64 range xxxxxxxxx]                    [xxxxxxxxxx|
2557 		 * 0                     S64_MAX S64_MIN                    -1
2558 		 *
2559 		 * The first condition below corresponds to the first diagram
2560 		 * above.
2561 		 */
2562 		if (reg->umax_value < (u64)reg->smin_value) {
2563 			reg->smin_value = (s64)reg->umin_value;
2564 			reg->umax_value = min_t(u64, reg->umax_value, reg->smax_value);
2565 		} else if ((u64)reg->smax_value < reg->umin_value) {
2566 			/* This second condition considers the case where the u64 range
2567 			 * overlaps with the negative portion of the s64 range:
2568 			 *
2569 			 * 0                                                   U64_MAX
2570 			 * |              [xxxxxxxxxxxxxx u64 range xxxxxxxxxxxxxx]  |
2571 			 * |----------------------------|----------------------------|
2572 			 * |xxxxxxxxx]                       [xxxxxxxxxxxx s64 range |
2573 			 * 0                     S64_MAX S64_MIN                    -1
2574 			 */
2575 			reg->smax_value = (s64)reg->umax_value;
2576 			reg->umin_value = max_t(u64, reg->umin_value, reg->smin_value);
2577 		}
2578 	}
2579 }
2580 
__reg_deduce_mixed_bounds(struct bpf_reg_state * reg)2581 static void __reg_deduce_mixed_bounds(struct bpf_reg_state *reg)
2582 {
2583 	/* Try to tighten 64-bit bounds from 32-bit knowledge, using 32-bit
2584 	 * values on both sides of 64-bit range in hope to have tighter range.
2585 	 * E.g., if r1 is [0x1'00000000, 0x3'80000000], and we learn from
2586 	 * 32-bit signed > 0 operation that s32 bounds are now [1; 0x7fffffff].
2587 	 * With this, we can substitute 1 as low 32-bits of _low_ 64-bit bound
2588 	 * (0x100000000 -> 0x100000001) and 0x7fffffff as low 32-bits of
2589 	 * _high_ 64-bit bound (0x380000000 -> 0x37fffffff) and arrive at a
2590 	 * better overall bounds for r1 as [0x1'000000001; 0x3'7fffffff].
2591 	 * We just need to make sure that derived bounds we are intersecting
2592 	 * with are well-formed ranges in respective s64 or u64 domain, just
2593 	 * like we do with similar kinds of 32-to-64 or 64-to-32 adjustments.
2594 	 */
2595 	__u64 new_umin, new_umax;
2596 	__s64 new_smin, new_smax;
2597 
2598 	/* u32 -> u64 tightening, it's always well-formed */
2599 	new_umin = (reg->umin_value & ~0xffffffffULL) | reg->u32_min_value;
2600 	new_umax = (reg->umax_value & ~0xffffffffULL) | reg->u32_max_value;
2601 	reg->umin_value = max_t(u64, reg->umin_value, new_umin);
2602 	reg->umax_value = min_t(u64, reg->umax_value, new_umax);
2603 	/* u32 -> s64 tightening, u32 range embedded into s64 preserves range validity */
2604 	new_smin = (reg->smin_value & ~0xffffffffULL) | reg->u32_min_value;
2605 	new_smax = (reg->smax_value & ~0xffffffffULL) | reg->u32_max_value;
2606 	reg->smin_value = max_t(s64, reg->smin_value, new_smin);
2607 	reg->smax_value = min_t(s64, reg->smax_value, new_smax);
2608 
2609 	/* Here we would like to handle a special case after sign extending load,
2610 	 * when upper bits for a 64-bit range are all 1s or all 0s.
2611 	 *
2612 	 * Upper bits are all 1s when register is in a range:
2613 	 *   [0xffff_ffff_0000_0000, 0xffff_ffff_ffff_ffff]
2614 	 * Upper bits are all 0s when register is in a range:
2615 	 *   [0x0000_0000_0000_0000, 0x0000_0000_ffff_ffff]
2616 	 * Together this forms are continuous range:
2617 	 *   [0xffff_ffff_0000_0000, 0x0000_0000_ffff_ffff]
2618 	 *
2619 	 * Now, suppose that register range is in fact tighter:
2620 	 *   [0xffff_ffff_8000_0000, 0x0000_0000_ffff_ffff] (R)
2621 	 * Also suppose that it's 32-bit range is positive,
2622 	 * meaning that lower 32-bits of the full 64-bit register
2623 	 * are in the range:
2624 	 *   [0x0000_0000, 0x7fff_ffff] (W)
2625 	 *
2626 	 * If this happens, then any value in a range:
2627 	 *   [0xffff_ffff_0000_0000, 0xffff_ffff_7fff_ffff]
2628 	 * is smaller than a lowest bound of the range (R):
2629 	 *   0xffff_ffff_8000_0000
2630 	 * which means that upper bits of the full 64-bit register
2631 	 * can't be all 1s, when lower bits are in range (W).
2632 	 *
2633 	 * Note that:
2634 	 *  - 0xffff_ffff_8000_0000 == (s64)S32_MIN
2635 	 *  - 0x0000_0000_7fff_ffff == (s64)S32_MAX
2636 	 * These relations are used in the conditions below.
2637 	 */
2638 	if (reg->s32_min_value >= 0 && reg->smin_value >= S32_MIN && reg->smax_value <= S32_MAX) {
2639 		reg->smin_value = reg->s32_min_value;
2640 		reg->smax_value = reg->s32_max_value;
2641 		reg->umin_value = reg->s32_min_value;
2642 		reg->umax_value = reg->s32_max_value;
2643 		reg->var_off = tnum_intersect(reg->var_off,
2644 					      tnum_range(reg->smin_value, reg->smax_value));
2645 	}
2646 }
2647 
__reg_deduce_bounds(struct bpf_reg_state * reg)2648 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2649 {
2650 	__reg32_deduce_bounds(reg);
2651 	__reg64_deduce_bounds(reg);
2652 	__reg_deduce_mixed_bounds(reg);
2653 }
2654 
2655 /* Attempts to improve var_off based on unsigned min/max information */
__reg_bound_offset(struct bpf_reg_state * reg)2656 static void __reg_bound_offset(struct bpf_reg_state *reg)
2657 {
2658 	struct tnum var64_off = tnum_intersect(reg->var_off,
2659 					       tnum_range(reg->umin_value,
2660 							  reg->umax_value));
2661 	struct tnum var32_off = tnum_intersect(tnum_subreg(var64_off),
2662 					       tnum_range(reg->u32_min_value,
2663 							  reg->u32_max_value));
2664 
2665 	reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
2666 }
2667 
reg_bounds_sync(struct bpf_reg_state * reg)2668 static void reg_bounds_sync(struct bpf_reg_state *reg)
2669 {
2670 	/* We might have learned new bounds from the var_off. */
2671 	__update_reg_bounds(reg);
2672 	/* We might have learned something about the sign bit. */
2673 	__reg_deduce_bounds(reg);
2674 	__reg_deduce_bounds(reg);
2675 	__reg_deduce_bounds(reg);
2676 	/* We might have learned some bits from the bounds. */
2677 	__reg_bound_offset(reg);
2678 	/* Intersecting with the old var_off might have improved our bounds
2679 	 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2680 	 * then new var_off is (0; 0x7f...fc) which improves our umax.
2681 	 */
2682 	__update_reg_bounds(reg);
2683 }
2684 
reg_bounds_sanity_check(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * ctx)2685 static int reg_bounds_sanity_check(struct bpf_verifier_env *env,
2686 				   struct bpf_reg_state *reg, const char *ctx)
2687 {
2688 	const char *msg;
2689 
2690 	if (reg->umin_value > reg->umax_value ||
2691 	    reg->smin_value > reg->smax_value ||
2692 	    reg->u32_min_value > reg->u32_max_value ||
2693 	    reg->s32_min_value > reg->s32_max_value) {
2694 		    msg = "range bounds violation";
2695 		    goto out;
2696 	}
2697 
2698 	if (tnum_is_const(reg->var_off)) {
2699 		u64 uval = reg->var_off.value;
2700 		s64 sval = (s64)uval;
2701 
2702 		if (reg->umin_value != uval || reg->umax_value != uval ||
2703 		    reg->smin_value != sval || reg->smax_value != sval) {
2704 			msg = "const tnum out of sync with range bounds";
2705 			goto out;
2706 		}
2707 	}
2708 
2709 	if (tnum_subreg_is_const(reg->var_off)) {
2710 		u32 uval32 = tnum_subreg(reg->var_off).value;
2711 		s32 sval32 = (s32)uval32;
2712 
2713 		if (reg->u32_min_value != uval32 || reg->u32_max_value != uval32 ||
2714 		    reg->s32_min_value != sval32 || reg->s32_max_value != sval32) {
2715 			msg = "const subreg tnum out of sync with range bounds";
2716 			goto out;
2717 		}
2718 	}
2719 
2720 	return 0;
2721 out:
2722 	verifier_bug(env, "REG INVARIANTS VIOLATION (%s): %s u64=[%#llx, %#llx] "
2723 		     "s64=[%#llx, %#llx] u32=[%#x, %#x] s32=[%#x, %#x] var_off=(%#llx, %#llx)",
2724 		     ctx, msg, reg->umin_value, reg->umax_value,
2725 		     reg->smin_value, reg->smax_value,
2726 		     reg->u32_min_value, reg->u32_max_value,
2727 		     reg->s32_min_value, reg->s32_max_value,
2728 		     reg->var_off.value, reg->var_off.mask);
2729 	if (env->test_reg_invariants)
2730 		return -EFAULT;
2731 	__mark_reg_unbounded(reg);
2732 	return 0;
2733 }
2734 
__reg32_bound_s64(s32 a)2735 static bool __reg32_bound_s64(s32 a)
2736 {
2737 	return a >= 0 && a <= S32_MAX;
2738 }
2739 
__reg_assign_32_into_64(struct bpf_reg_state * reg)2740 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2741 {
2742 	reg->umin_value = reg->u32_min_value;
2743 	reg->umax_value = reg->u32_max_value;
2744 
2745 	/* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2746 	 * be positive otherwise set to worse case bounds and refine later
2747 	 * from tnum.
2748 	 */
2749 	if (__reg32_bound_s64(reg->s32_min_value) &&
2750 	    __reg32_bound_s64(reg->s32_max_value)) {
2751 		reg->smin_value = reg->s32_min_value;
2752 		reg->smax_value = reg->s32_max_value;
2753 	} else {
2754 		reg->smin_value = 0;
2755 		reg->smax_value = U32_MAX;
2756 	}
2757 }
2758 
2759 /* Mark a register as having a completely unknown (scalar) value. */
__mark_reg_unknown_imprecise(struct bpf_reg_state * reg)2760 static void __mark_reg_unknown_imprecise(struct bpf_reg_state *reg)
2761 {
2762 	/*
2763 	 * Clear type, off, and union(map_ptr, range) and
2764 	 * padding between 'type' and union
2765 	 */
2766 	memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2767 	reg->type = SCALAR_VALUE;
2768 	reg->id = 0;
2769 	reg->ref_obj_id = 0;
2770 	reg->var_off = tnum_unknown;
2771 	reg->frameno = 0;
2772 	reg->precise = false;
2773 	__mark_reg_unbounded(reg);
2774 }
2775 
2776 /* Mark a register as having a completely unknown (scalar) value,
2777  * initialize .precise as true when not bpf capable.
2778  */
__mark_reg_unknown(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)2779 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2780 			       struct bpf_reg_state *reg)
2781 {
2782 	__mark_reg_unknown_imprecise(reg);
2783 	reg->precise = !env->bpf_capable;
2784 }
2785 
mark_reg_unknown(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)2786 static void mark_reg_unknown(struct bpf_verifier_env *env,
2787 			     struct bpf_reg_state *regs, u32 regno)
2788 {
2789 	if (WARN_ON(regno >= MAX_BPF_REG)) {
2790 		verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
2791 		/* Something bad happened, let's kill all regs except FP */
2792 		for (regno = 0; regno < BPF_REG_FP; regno++)
2793 			__mark_reg_not_init(env, regs + regno);
2794 		return;
2795 	}
2796 	__mark_reg_unknown(env, regs + regno);
2797 }
2798 
__mark_reg_s32_range(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,s32 s32_min,s32 s32_max)2799 static int __mark_reg_s32_range(struct bpf_verifier_env *env,
2800 				struct bpf_reg_state *regs,
2801 				u32 regno,
2802 				s32 s32_min,
2803 				s32 s32_max)
2804 {
2805 	struct bpf_reg_state *reg = regs + regno;
2806 
2807 	reg->s32_min_value = max_t(s32, reg->s32_min_value, s32_min);
2808 	reg->s32_max_value = min_t(s32, reg->s32_max_value, s32_max);
2809 
2810 	reg->smin_value = max_t(s64, reg->smin_value, s32_min);
2811 	reg->smax_value = min_t(s64, reg->smax_value, s32_max);
2812 
2813 	reg_bounds_sync(reg);
2814 
2815 	return reg_bounds_sanity_check(env, reg, "s32_range");
2816 }
2817 
__mark_reg_not_init(const struct bpf_verifier_env * env,struct bpf_reg_state * reg)2818 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2819 				struct bpf_reg_state *reg)
2820 {
2821 	__mark_reg_unknown(env, reg);
2822 	reg->type = NOT_INIT;
2823 }
2824 
mark_reg_not_init(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno)2825 static void mark_reg_not_init(struct bpf_verifier_env *env,
2826 			      struct bpf_reg_state *regs, u32 regno)
2827 {
2828 	if (WARN_ON(regno >= MAX_BPF_REG)) {
2829 		verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
2830 		/* Something bad happened, let's kill all regs except FP */
2831 		for (regno = 0; regno < BPF_REG_FP; regno++)
2832 			__mark_reg_not_init(env, regs + regno);
2833 		return;
2834 	}
2835 	__mark_reg_not_init(env, regs + regno);
2836 }
2837 
mark_btf_ld_reg(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,enum bpf_reg_type reg_type,struct btf * btf,u32 btf_id,enum bpf_type_flag flag)2838 static int mark_btf_ld_reg(struct bpf_verifier_env *env,
2839 			   struct bpf_reg_state *regs, u32 regno,
2840 			   enum bpf_reg_type reg_type,
2841 			   struct btf *btf, u32 btf_id,
2842 			   enum bpf_type_flag flag)
2843 {
2844 	switch (reg_type) {
2845 	case SCALAR_VALUE:
2846 		mark_reg_unknown(env, regs, regno);
2847 		return 0;
2848 	case PTR_TO_BTF_ID:
2849 		mark_reg_known_zero(env, regs, regno);
2850 		regs[regno].type = PTR_TO_BTF_ID | flag;
2851 		regs[regno].btf = btf;
2852 		regs[regno].btf_id = btf_id;
2853 		if (type_may_be_null(flag))
2854 			regs[regno].id = ++env->id_gen;
2855 		return 0;
2856 	case PTR_TO_MEM:
2857 		mark_reg_known_zero(env, regs, regno);
2858 		regs[regno].type = PTR_TO_MEM | flag;
2859 		regs[regno].mem_size = 0;
2860 		return 0;
2861 	default:
2862 		verifier_bug(env, "unexpected reg_type %d in %s\n", reg_type, __func__);
2863 		return -EFAULT;
2864 	}
2865 }
2866 
2867 #define DEF_NOT_SUBREG	(0)
init_reg_state(struct bpf_verifier_env * env,struct bpf_func_state * state)2868 static void init_reg_state(struct bpf_verifier_env *env,
2869 			   struct bpf_func_state *state)
2870 {
2871 	struct bpf_reg_state *regs = state->regs;
2872 	int i;
2873 
2874 	for (i = 0; i < MAX_BPF_REG; i++) {
2875 		mark_reg_not_init(env, regs, i);
2876 		regs[i].live = REG_LIVE_NONE;
2877 		regs[i].parent = NULL;
2878 		regs[i].subreg_def = DEF_NOT_SUBREG;
2879 	}
2880 
2881 	/* frame pointer */
2882 	regs[BPF_REG_FP].type = PTR_TO_STACK;
2883 	mark_reg_known_zero(env, regs, BPF_REG_FP);
2884 	regs[BPF_REG_FP].frameno = state->frameno;
2885 }
2886 
retval_range(s32 minval,s32 maxval)2887 static struct bpf_retval_range retval_range(s32 minval, s32 maxval)
2888 {
2889 	return (struct bpf_retval_range){ minval, maxval };
2890 }
2891 
2892 #define BPF_MAIN_FUNC (-1)
init_func_state(struct bpf_verifier_env * env,struct bpf_func_state * state,int callsite,int frameno,int subprogno)2893 static void init_func_state(struct bpf_verifier_env *env,
2894 			    struct bpf_func_state *state,
2895 			    int callsite, int frameno, int subprogno)
2896 {
2897 	state->callsite = callsite;
2898 	state->frameno = frameno;
2899 	state->subprogno = subprogno;
2900 	state->callback_ret_range = retval_range(0, 0);
2901 	init_reg_state(env, state);
2902 	mark_verifier_state_scratched(env);
2903 }
2904 
2905 /* Similar to push_stack(), but for async callbacks */
push_async_cb(struct bpf_verifier_env * env,int insn_idx,int prev_insn_idx,int subprog,bool is_sleepable)2906 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2907 						int insn_idx, int prev_insn_idx,
2908 						int subprog, bool is_sleepable)
2909 {
2910 	struct bpf_verifier_stack_elem *elem;
2911 	struct bpf_func_state *frame;
2912 
2913 	elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL_ACCOUNT);
2914 	if (!elem)
2915 		return NULL;
2916 
2917 	elem->insn_idx = insn_idx;
2918 	elem->prev_insn_idx = prev_insn_idx;
2919 	elem->next = env->head;
2920 	elem->log_pos = env->log.end_pos;
2921 	env->head = elem;
2922 	env->stack_size++;
2923 	if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2924 		verbose(env,
2925 			"The sequence of %d jumps is too complex for async cb.\n",
2926 			env->stack_size);
2927 		return NULL;
2928 	}
2929 	/* Unlike push_stack() do not copy_verifier_state().
2930 	 * The caller state doesn't matter.
2931 	 * This is async callback. It starts in a fresh stack.
2932 	 * Initialize it similar to do_check_common().
2933 	 */
2934 	elem->st.branches = 1;
2935 	elem->st.in_sleepable = is_sleepable;
2936 	frame = kzalloc(sizeof(*frame), GFP_KERNEL_ACCOUNT);
2937 	if (!frame)
2938 		return NULL;
2939 	init_func_state(env, frame,
2940 			BPF_MAIN_FUNC /* callsite */,
2941 			0 /* frameno within this callchain */,
2942 			subprog /* subprog number within this prog */);
2943 	elem->st.frame[0] = frame;
2944 	return &elem->st;
2945 }
2946 
2947 
2948 enum reg_arg_type {
2949 	SRC_OP,		/* register is used as source operand */
2950 	DST_OP,		/* register is used as destination operand */
2951 	DST_OP_NO_MARK	/* same as above, check only, don't mark */
2952 };
2953 
cmp_subprogs(const void * a,const void * b)2954 static int cmp_subprogs(const void *a, const void *b)
2955 {
2956 	return ((struct bpf_subprog_info *)a)->start -
2957 	       ((struct bpf_subprog_info *)b)->start;
2958 }
2959 
2960 /* Find subprogram that contains instruction at 'off' */
find_containing_subprog(struct bpf_verifier_env * env,int off)2961 static struct bpf_subprog_info *find_containing_subprog(struct bpf_verifier_env *env, int off)
2962 {
2963 	struct bpf_subprog_info *vals = env->subprog_info;
2964 	int l, r, m;
2965 
2966 	if (off >= env->prog->len || off < 0 || env->subprog_cnt == 0)
2967 		return NULL;
2968 
2969 	l = 0;
2970 	r = env->subprog_cnt - 1;
2971 	while (l < r) {
2972 		m = l + (r - l + 1) / 2;
2973 		if (vals[m].start <= off)
2974 			l = m;
2975 		else
2976 			r = m - 1;
2977 	}
2978 	return &vals[l];
2979 }
2980 
2981 /* Find subprogram that starts exactly at 'off' */
find_subprog(struct bpf_verifier_env * env,int off)2982 static int find_subprog(struct bpf_verifier_env *env, int off)
2983 {
2984 	struct bpf_subprog_info *p;
2985 
2986 	p = find_containing_subprog(env, off);
2987 	if (!p || p->start != off)
2988 		return -ENOENT;
2989 	return p - env->subprog_info;
2990 }
2991 
add_subprog(struct bpf_verifier_env * env,int off)2992 static int add_subprog(struct bpf_verifier_env *env, int off)
2993 {
2994 	int insn_cnt = env->prog->len;
2995 	int ret;
2996 
2997 	if (off >= insn_cnt || off < 0) {
2998 		verbose(env, "call to invalid destination\n");
2999 		return -EINVAL;
3000 	}
3001 	ret = find_subprog(env, off);
3002 	if (ret >= 0)
3003 		return ret;
3004 	if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
3005 		verbose(env, "too many subprograms\n");
3006 		return -E2BIG;
3007 	}
3008 	/* determine subprog starts. The end is one before the next starts */
3009 	env->subprog_info[env->subprog_cnt++].start = off;
3010 	sort(env->subprog_info, env->subprog_cnt,
3011 	     sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
3012 	return env->subprog_cnt - 1;
3013 }
3014 
bpf_find_exception_callback_insn_off(struct bpf_verifier_env * env)3015 static int bpf_find_exception_callback_insn_off(struct bpf_verifier_env *env)
3016 {
3017 	struct bpf_prog_aux *aux = env->prog->aux;
3018 	struct btf *btf = aux->btf;
3019 	const struct btf_type *t;
3020 	u32 main_btf_id, id;
3021 	const char *name;
3022 	int ret, i;
3023 
3024 	/* Non-zero func_info_cnt implies valid btf */
3025 	if (!aux->func_info_cnt)
3026 		return 0;
3027 	main_btf_id = aux->func_info[0].type_id;
3028 
3029 	t = btf_type_by_id(btf, main_btf_id);
3030 	if (!t) {
3031 		verbose(env, "invalid btf id for main subprog in func_info\n");
3032 		return -EINVAL;
3033 	}
3034 
3035 	name = btf_find_decl_tag_value(btf, t, -1, "exception_callback:");
3036 	if (IS_ERR(name)) {
3037 		ret = PTR_ERR(name);
3038 		/* If there is no tag present, there is no exception callback */
3039 		if (ret == -ENOENT)
3040 			ret = 0;
3041 		else if (ret == -EEXIST)
3042 			verbose(env, "multiple exception callback tags for main subprog\n");
3043 		return ret;
3044 	}
3045 
3046 	ret = btf_find_by_name_kind(btf, name, BTF_KIND_FUNC);
3047 	if (ret < 0) {
3048 		verbose(env, "exception callback '%s' could not be found in BTF\n", name);
3049 		return ret;
3050 	}
3051 	id = ret;
3052 	t = btf_type_by_id(btf, id);
3053 	if (btf_func_linkage(t) != BTF_FUNC_GLOBAL) {
3054 		verbose(env, "exception callback '%s' must have global linkage\n", name);
3055 		return -EINVAL;
3056 	}
3057 	ret = 0;
3058 	for (i = 0; i < aux->func_info_cnt; i++) {
3059 		if (aux->func_info[i].type_id != id)
3060 			continue;
3061 		ret = aux->func_info[i].insn_off;
3062 		/* Further func_info and subprog checks will also happen
3063 		 * later, so assume this is the right insn_off for now.
3064 		 */
3065 		if (!ret) {
3066 			verbose(env, "invalid exception callback insn_off in func_info: 0\n");
3067 			ret = -EINVAL;
3068 		}
3069 	}
3070 	if (!ret) {
3071 		verbose(env, "exception callback type id not found in func_info\n");
3072 		ret = -EINVAL;
3073 	}
3074 	return ret;
3075 }
3076 
3077 #define MAX_KFUNC_DESCS 256
3078 #define MAX_KFUNC_BTFS	256
3079 
3080 struct bpf_kfunc_desc {
3081 	struct btf_func_model func_model;
3082 	u32 func_id;
3083 	s32 imm;
3084 	u16 offset;
3085 	unsigned long addr;
3086 };
3087 
3088 struct bpf_kfunc_btf {
3089 	struct btf *btf;
3090 	struct module *module;
3091 	u16 offset;
3092 };
3093 
3094 struct bpf_kfunc_desc_tab {
3095 	/* Sorted by func_id (BTF ID) and offset (fd_array offset) during
3096 	 * verification. JITs do lookups by bpf_insn, where func_id may not be
3097 	 * available, therefore at the end of verification do_misc_fixups()
3098 	 * sorts this by imm and offset.
3099 	 */
3100 	struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
3101 	u32 nr_descs;
3102 };
3103 
3104 struct bpf_kfunc_btf_tab {
3105 	struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
3106 	u32 nr_descs;
3107 };
3108 
kfunc_desc_cmp_by_id_off(const void * a,const void * b)3109 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
3110 {
3111 	const struct bpf_kfunc_desc *d0 = a;
3112 	const struct bpf_kfunc_desc *d1 = b;
3113 
3114 	/* func_id is not greater than BTF_MAX_TYPE */
3115 	return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
3116 }
3117 
kfunc_btf_cmp_by_off(const void * a,const void * b)3118 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
3119 {
3120 	const struct bpf_kfunc_btf *d0 = a;
3121 	const struct bpf_kfunc_btf *d1 = b;
3122 
3123 	return d0->offset - d1->offset;
3124 }
3125 
3126 static const struct bpf_kfunc_desc *
find_kfunc_desc(const struct bpf_prog * prog,u32 func_id,u16 offset)3127 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
3128 {
3129 	struct bpf_kfunc_desc desc = {
3130 		.func_id = func_id,
3131 		.offset = offset,
3132 	};
3133 	struct bpf_kfunc_desc_tab *tab;
3134 
3135 	tab = prog->aux->kfunc_tab;
3136 	return bsearch(&desc, tab->descs, tab->nr_descs,
3137 		       sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
3138 }
3139 
bpf_get_kfunc_addr(const struct bpf_prog * prog,u32 func_id,u16 btf_fd_idx,u8 ** func_addr)3140 int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
3141 		       u16 btf_fd_idx, u8 **func_addr)
3142 {
3143 	const struct bpf_kfunc_desc *desc;
3144 
3145 	desc = find_kfunc_desc(prog, func_id, btf_fd_idx);
3146 	if (!desc)
3147 		return -EFAULT;
3148 
3149 	*func_addr = (u8 *)desc->addr;
3150 	return 0;
3151 }
3152 
__find_kfunc_desc_btf(struct bpf_verifier_env * env,s16 offset)3153 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
3154 					 s16 offset)
3155 {
3156 	struct bpf_kfunc_btf kf_btf = { .offset = offset };
3157 	struct bpf_kfunc_btf_tab *tab;
3158 	struct bpf_kfunc_btf *b;
3159 	struct module *mod;
3160 	struct btf *btf;
3161 	int btf_fd;
3162 
3163 	tab = env->prog->aux->kfunc_btf_tab;
3164 	b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
3165 		    sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
3166 	if (!b) {
3167 		if (tab->nr_descs == MAX_KFUNC_BTFS) {
3168 			verbose(env, "too many different module BTFs\n");
3169 			return ERR_PTR(-E2BIG);
3170 		}
3171 
3172 		if (bpfptr_is_null(env->fd_array)) {
3173 			verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
3174 			return ERR_PTR(-EPROTO);
3175 		}
3176 
3177 		if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
3178 					    offset * sizeof(btf_fd),
3179 					    sizeof(btf_fd)))
3180 			return ERR_PTR(-EFAULT);
3181 
3182 		btf = btf_get_by_fd(btf_fd);
3183 		if (IS_ERR(btf)) {
3184 			verbose(env, "invalid module BTF fd specified\n");
3185 			return btf;
3186 		}
3187 
3188 		if (!btf_is_module(btf)) {
3189 			verbose(env, "BTF fd for kfunc is not a module BTF\n");
3190 			btf_put(btf);
3191 			return ERR_PTR(-EINVAL);
3192 		}
3193 
3194 		mod = btf_try_get_module(btf);
3195 		if (!mod) {
3196 			btf_put(btf);
3197 			return ERR_PTR(-ENXIO);
3198 		}
3199 
3200 		b = &tab->descs[tab->nr_descs++];
3201 		b->btf = btf;
3202 		b->module = mod;
3203 		b->offset = offset;
3204 
3205 		/* sort() reorders entries by value, so b may no longer point
3206 		 * to the right entry after this
3207 		 */
3208 		sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
3209 		     kfunc_btf_cmp_by_off, NULL);
3210 	} else {
3211 		btf = b->btf;
3212 	}
3213 
3214 	return btf;
3215 }
3216 
bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab * tab)3217 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
3218 {
3219 	if (!tab)
3220 		return;
3221 
3222 	while (tab->nr_descs--) {
3223 		module_put(tab->descs[tab->nr_descs].module);
3224 		btf_put(tab->descs[tab->nr_descs].btf);
3225 	}
3226 	kfree(tab);
3227 }
3228 
find_kfunc_desc_btf(struct bpf_verifier_env * env,s16 offset)3229 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
3230 {
3231 	if (offset) {
3232 		if (offset < 0) {
3233 			/* In the future, this can be allowed to increase limit
3234 			 * of fd index into fd_array, interpreted as u16.
3235 			 */
3236 			verbose(env, "negative offset disallowed for kernel module function call\n");
3237 			return ERR_PTR(-EINVAL);
3238 		}
3239 
3240 		return __find_kfunc_desc_btf(env, offset);
3241 	}
3242 	return btf_vmlinux ?: ERR_PTR(-ENOENT);
3243 }
3244 
add_kfunc_call(struct bpf_verifier_env * env,u32 func_id,s16 offset)3245 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
3246 {
3247 	const struct btf_type *func, *func_proto;
3248 	struct bpf_kfunc_btf_tab *btf_tab;
3249 	struct bpf_kfunc_desc_tab *tab;
3250 	struct bpf_prog_aux *prog_aux;
3251 	struct bpf_kfunc_desc *desc;
3252 	const char *func_name;
3253 	struct btf *desc_btf;
3254 	unsigned long call_imm;
3255 	unsigned long addr;
3256 	int err;
3257 
3258 	prog_aux = env->prog->aux;
3259 	tab = prog_aux->kfunc_tab;
3260 	btf_tab = prog_aux->kfunc_btf_tab;
3261 	if (!tab) {
3262 		if (!btf_vmlinux) {
3263 			verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
3264 			return -ENOTSUPP;
3265 		}
3266 
3267 		if (!env->prog->jit_requested) {
3268 			verbose(env, "JIT is required for calling kernel function\n");
3269 			return -ENOTSUPP;
3270 		}
3271 
3272 		if (!bpf_jit_supports_kfunc_call()) {
3273 			verbose(env, "JIT does not support calling kernel function\n");
3274 			return -ENOTSUPP;
3275 		}
3276 
3277 		if (!env->prog->gpl_compatible) {
3278 			verbose(env, "cannot call kernel function from non-GPL compatible program\n");
3279 			return -EINVAL;
3280 		}
3281 
3282 		tab = kzalloc(sizeof(*tab), GFP_KERNEL_ACCOUNT);
3283 		if (!tab)
3284 			return -ENOMEM;
3285 		prog_aux->kfunc_tab = tab;
3286 	}
3287 
3288 	/* func_id == 0 is always invalid, but instead of returning an error, be
3289 	 * conservative and wait until the code elimination pass before returning
3290 	 * error, so that invalid calls that get pruned out can be in BPF programs
3291 	 * loaded from userspace.  It is also required that offset be untouched
3292 	 * for such calls.
3293 	 */
3294 	if (!func_id && !offset)
3295 		return 0;
3296 
3297 	if (!btf_tab && offset) {
3298 		btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL_ACCOUNT);
3299 		if (!btf_tab)
3300 			return -ENOMEM;
3301 		prog_aux->kfunc_btf_tab = btf_tab;
3302 	}
3303 
3304 	desc_btf = find_kfunc_desc_btf(env, offset);
3305 	if (IS_ERR(desc_btf)) {
3306 		verbose(env, "failed to find BTF for kernel function\n");
3307 		return PTR_ERR(desc_btf);
3308 	}
3309 
3310 	if (find_kfunc_desc(env->prog, func_id, offset))
3311 		return 0;
3312 
3313 	if (tab->nr_descs == MAX_KFUNC_DESCS) {
3314 		verbose(env, "too many different kernel function calls\n");
3315 		return -E2BIG;
3316 	}
3317 
3318 	func = btf_type_by_id(desc_btf, func_id);
3319 	if (!func || !btf_type_is_func(func)) {
3320 		verbose(env, "kernel btf_id %u is not a function\n",
3321 			func_id);
3322 		return -EINVAL;
3323 	}
3324 	func_proto = btf_type_by_id(desc_btf, func->type);
3325 	if (!func_proto || !btf_type_is_func_proto(func_proto)) {
3326 		verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
3327 			func_id);
3328 		return -EINVAL;
3329 	}
3330 
3331 	func_name = btf_name_by_offset(desc_btf, func->name_off);
3332 	addr = kallsyms_lookup_name(func_name);
3333 	if (!addr) {
3334 		verbose(env, "cannot find address for kernel function %s\n",
3335 			func_name);
3336 		return -EINVAL;
3337 	}
3338 	specialize_kfunc(env, func_id, offset, &addr);
3339 
3340 	if (bpf_jit_supports_far_kfunc_call()) {
3341 		call_imm = func_id;
3342 	} else {
3343 		call_imm = BPF_CALL_IMM(addr);
3344 		/* Check whether the relative offset overflows desc->imm */
3345 		if ((unsigned long)(s32)call_imm != call_imm) {
3346 			verbose(env, "address of kernel function %s is out of range\n",
3347 				func_name);
3348 			return -EINVAL;
3349 		}
3350 	}
3351 
3352 	if (bpf_dev_bound_kfunc_id(func_id)) {
3353 		err = bpf_dev_bound_kfunc_check(&env->log, prog_aux);
3354 		if (err)
3355 			return err;
3356 	}
3357 
3358 	desc = &tab->descs[tab->nr_descs++];
3359 	desc->func_id = func_id;
3360 	desc->imm = call_imm;
3361 	desc->offset = offset;
3362 	desc->addr = addr;
3363 	err = btf_distill_func_proto(&env->log, desc_btf,
3364 				     func_proto, func_name,
3365 				     &desc->func_model);
3366 	if (!err)
3367 		sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
3368 		     kfunc_desc_cmp_by_id_off, NULL);
3369 	return err;
3370 }
3371 
kfunc_desc_cmp_by_imm_off(const void * a,const void * b)3372 static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
3373 {
3374 	const struct bpf_kfunc_desc *d0 = a;
3375 	const struct bpf_kfunc_desc *d1 = b;
3376 
3377 	if (d0->imm != d1->imm)
3378 		return d0->imm < d1->imm ? -1 : 1;
3379 	if (d0->offset != d1->offset)
3380 		return d0->offset < d1->offset ? -1 : 1;
3381 	return 0;
3382 }
3383 
sort_kfunc_descs_by_imm_off(struct bpf_prog * prog)3384 static void sort_kfunc_descs_by_imm_off(struct bpf_prog *prog)
3385 {
3386 	struct bpf_kfunc_desc_tab *tab;
3387 
3388 	tab = prog->aux->kfunc_tab;
3389 	if (!tab)
3390 		return;
3391 
3392 	sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
3393 	     kfunc_desc_cmp_by_imm_off, NULL);
3394 }
3395 
bpf_prog_has_kfunc_call(const struct bpf_prog * prog)3396 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
3397 {
3398 	return !!prog->aux->kfunc_tab;
3399 }
3400 
3401 const struct btf_func_model *
bpf_jit_find_kfunc_model(const struct bpf_prog * prog,const struct bpf_insn * insn)3402 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
3403 			 const struct bpf_insn *insn)
3404 {
3405 	const struct bpf_kfunc_desc desc = {
3406 		.imm = insn->imm,
3407 		.offset = insn->off,
3408 	};
3409 	const struct bpf_kfunc_desc *res;
3410 	struct bpf_kfunc_desc_tab *tab;
3411 
3412 	tab = prog->aux->kfunc_tab;
3413 	res = bsearch(&desc, tab->descs, tab->nr_descs,
3414 		      sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm_off);
3415 
3416 	return res ? &res->func_model : NULL;
3417 }
3418 
add_kfunc_in_insns(struct bpf_verifier_env * env,struct bpf_insn * insn,int cnt)3419 static int add_kfunc_in_insns(struct bpf_verifier_env *env,
3420 			      struct bpf_insn *insn, int cnt)
3421 {
3422 	int i, ret;
3423 
3424 	for (i = 0; i < cnt; i++, insn++) {
3425 		if (bpf_pseudo_kfunc_call(insn)) {
3426 			ret = add_kfunc_call(env, insn->imm, insn->off);
3427 			if (ret < 0)
3428 				return ret;
3429 		}
3430 	}
3431 	return 0;
3432 }
3433 
add_subprog_and_kfunc(struct bpf_verifier_env * env)3434 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
3435 {
3436 	struct bpf_subprog_info *subprog = env->subprog_info;
3437 	int i, ret, insn_cnt = env->prog->len, ex_cb_insn;
3438 	struct bpf_insn *insn = env->prog->insnsi;
3439 
3440 	/* Add entry function. */
3441 	ret = add_subprog(env, 0);
3442 	if (ret)
3443 		return ret;
3444 
3445 	for (i = 0; i < insn_cnt; i++, insn++) {
3446 		if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
3447 		    !bpf_pseudo_kfunc_call(insn))
3448 			continue;
3449 
3450 		if (!env->bpf_capable) {
3451 			verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
3452 			return -EPERM;
3453 		}
3454 
3455 		if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
3456 			ret = add_subprog(env, i + insn->imm + 1);
3457 		else
3458 			ret = add_kfunc_call(env, insn->imm, insn->off);
3459 
3460 		if (ret < 0)
3461 			return ret;
3462 	}
3463 
3464 	ret = bpf_find_exception_callback_insn_off(env);
3465 	if (ret < 0)
3466 		return ret;
3467 	ex_cb_insn = ret;
3468 
3469 	/* If ex_cb_insn > 0, this means that the main program has a subprog
3470 	 * marked using BTF decl tag to serve as the exception callback.
3471 	 */
3472 	if (ex_cb_insn) {
3473 		ret = add_subprog(env, ex_cb_insn);
3474 		if (ret < 0)
3475 			return ret;
3476 		for (i = 1; i < env->subprog_cnt; i++) {
3477 			if (env->subprog_info[i].start != ex_cb_insn)
3478 				continue;
3479 			env->exception_callback_subprog = i;
3480 			mark_subprog_exc_cb(env, i);
3481 			break;
3482 		}
3483 	}
3484 
3485 	/* Add a fake 'exit' subprog which could simplify subprog iteration
3486 	 * logic. 'subprog_cnt' should not be increased.
3487 	 */
3488 	subprog[env->subprog_cnt].start = insn_cnt;
3489 
3490 	if (env->log.level & BPF_LOG_LEVEL2)
3491 		for (i = 0; i < env->subprog_cnt; i++)
3492 			verbose(env, "func#%d @%d\n", i, subprog[i].start);
3493 
3494 	return 0;
3495 }
3496 
jmp_offset(struct bpf_insn * insn)3497 static int jmp_offset(struct bpf_insn *insn)
3498 {
3499 	u8 code = insn->code;
3500 
3501 	if (code == (BPF_JMP32 | BPF_JA))
3502 		return insn->imm;
3503 	return insn->off;
3504 }
3505 
check_subprogs(struct bpf_verifier_env * env)3506 static int check_subprogs(struct bpf_verifier_env *env)
3507 {
3508 	int i, subprog_start, subprog_end, off, cur_subprog = 0;
3509 	struct bpf_subprog_info *subprog = env->subprog_info;
3510 	struct bpf_insn *insn = env->prog->insnsi;
3511 	int insn_cnt = env->prog->len;
3512 
3513 	/* now check that all jumps are within the same subprog */
3514 	subprog_start = subprog[cur_subprog].start;
3515 	subprog_end = subprog[cur_subprog + 1].start;
3516 	for (i = 0; i < insn_cnt; i++) {
3517 		u8 code = insn[i].code;
3518 
3519 		if (code == (BPF_JMP | BPF_CALL) &&
3520 		    insn[i].src_reg == 0 &&
3521 		    insn[i].imm == BPF_FUNC_tail_call) {
3522 			subprog[cur_subprog].has_tail_call = true;
3523 			subprog[cur_subprog].tail_call_reachable = true;
3524 		}
3525 		if (BPF_CLASS(code) == BPF_LD &&
3526 		    (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
3527 			subprog[cur_subprog].has_ld_abs = true;
3528 		if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
3529 			goto next;
3530 		if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
3531 			goto next;
3532 		off = i + jmp_offset(&insn[i]) + 1;
3533 		if (off < subprog_start || off >= subprog_end) {
3534 			verbose(env, "jump out of range from insn %d to %d\n", i, off);
3535 			return -EINVAL;
3536 		}
3537 next:
3538 		if (i == subprog_end - 1) {
3539 			/* to avoid fall-through from one subprog into another
3540 			 * the last insn of the subprog should be either exit
3541 			 * or unconditional jump back or bpf_throw call
3542 			 */
3543 			if (code != (BPF_JMP | BPF_EXIT) &&
3544 			    code != (BPF_JMP32 | BPF_JA) &&
3545 			    code != (BPF_JMP | BPF_JA)) {
3546 				verbose(env, "last insn is not an exit or jmp\n");
3547 				return -EINVAL;
3548 			}
3549 			subprog_start = subprog_end;
3550 			cur_subprog++;
3551 			if (cur_subprog < env->subprog_cnt)
3552 				subprog_end = subprog[cur_subprog + 1].start;
3553 		}
3554 	}
3555 	return 0;
3556 }
3557 
3558 /* Parentage chain of this register (or stack slot) should take care of all
3559  * issues like callee-saved registers, stack slot allocation time, etc.
3560  */
mark_reg_read(struct bpf_verifier_env * env,const struct bpf_reg_state * state,struct bpf_reg_state * parent,u8 flag)3561 static int mark_reg_read(struct bpf_verifier_env *env,
3562 			 const struct bpf_reg_state *state,
3563 			 struct bpf_reg_state *parent, u8 flag)
3564 {
3565 	bool writes = parent == state->parent; /* Observe write marks */
3566 	int cnt = 0;
3567 
3568 	while (parent) {
3569 		/* if read wasn't screened by an earlier write ... */
3570 		if (writes && state->live & REG_LIVE_WRITTEN)
3571 			break;
3572 		if (verifier_bug_if(parent->live & REG_LIVE_DONE, env,
3573 				    "type %s var_off %lld off %d",
3574 				    reg_type_str(env, parent->type),
3575 				    parent->var_off.value, parent->off))
3576 			return -EFAULT;
3577 		/* The first condition is more likely to be true than the
3578 		 * second, checked it first.
3579 		 */
3580 		if ((parent->live & REG_LIVE_READ) == flag ||
3581 		    parent->live & REG_LIVE_READ64)
3582 			/* The parentage chain never changes and
3583 			 * this parent was already marked as LIVE_READ.
3584 			 * There is no need to keep walking the chain again and
3585 			 * keep re-marking all parents as LIVE_READ.
3586 			 * This case happens when the same register is read
3587 			 * multiple times without writes into it in-between.
3588 			 * Also, if parent has the stronger REG_LIVE_READ64 set,
3589 			 * then no need to set the weak REG_LIVE_READ32.
3590 			 */
3591 			break;
3592 		/* ... then we depend on parent's value */
3593 		parent->live |= flag;
3594 		/* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
3595 		if (flag == REG_LIVE_READ64)
3596 			parent->live &= ~REG_LIVE_READ32;
3597 		state = parent;
3598 		parent = state->parent;
3599 		writes = true;
3600 		cnt++;
3601 	}
3602 
3603 	if (env->longest_mark_read_walk < cnt)
3604 		env->longest_mark_read_walk = cnt;
3605 	return 0;
3606 }
3607 
mark_stack_slot_obj_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int spi,int nr_slots)3608 static int mark_stack_slot_obj_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
3609 				    int spi, int nr_slots)
3610 {
3611 	struct bpf_func_state *state = func(env, reg);
3612 	int err, i;
3613 
3614 	for (i = 0; i < nr_slots; i++) {
3615 		struct bpf_reg_state *st = &state->stack[spi - i].spilled_ptr;
3616 
3617 		err = mark_reg_read(env, st, st->parent, REG_LIVE_READ64);
3618 		if (err)
3619 			return err;
3620 
3621 		mark_stack_slot_scratched(env, spi - i);
3622 	}
3623 	return 0;
3624 }
3625 
mark_dynptr_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg)3626 static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
3627 {
3628 	int spi;
3629 
3630 	/* For CONST_PTR_TO_DYNPTR, it must have already been done by
3631 	 * check_reg_arg in check_helper_call and mark_btf_func_reg_size in
3632 	 * check_kfunc_call.
3633 	 */
3634 	if (reg->type == CONST_PTR_TO_DYNPTR)
3635 		return 0;
3636 	spi = dynptr_get_spi(env, reg);
3637 	if (spi < 0)
3638 		return spi;
3639 	/* Caller ensures dynptr is valid and initialized, which means spi is in
3640 	 * bounds and spi is the first dynptr slot. Simply mark stack slot as
3641 	 * read.
3642 	 */
3643 	return mark_stack_slot_obj_read(env, reg, spi, BPF_DYNPTR_NR_SLOTS);
3644 }
3645 
mark_iter_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int spi,int nr_slots)3646 static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
3647 			  int spi, int nr_slots)
3648 {
3649 	return mark_stack_slot_obj_read(env, reg, spi, nr_slots);
3650 }
3651 
mark_irq_flag_read(struct bpf_verifier_env * env,struct bpf_reg_state * reg)3652 static int mark_irq_flag_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
3653 {
3654 	int spi;
3655 
3656 	spi = irq_flag_get_spi(env, reg);
3657 	if (spi < 0)
3658 		return spi;
3659 	return mark_stack_slot_obj_read(env, reg, spi, 1);
3660 }
3661 
3662 /* This function is supposed to be used by the following 32-bit optimization
3663  * code only. It returns TRUE if the source or destination register operates
3664  * on 64-bit, otherwise return FALSE.
3665  */
is_reg64(struct bpf_verifier_env * env,struct bpf_insn * insn,u32 regno,struct bpf_reg_state * reg,enum reg_arg_type t)3666 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
3667 		     u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
3668 {
3669 	u8 code, class, op;
3670 
3671 	code = insn->code;
3672 	class = BPF_CLASS(code);
3673 	op = BPF_OP(code);
3674 	if (class == BPF_JMP) {
3675 		/* BPF_EXIT for "main" will reach here. Return TRUE
3676 		 * conservatively.
3677 		 */
3678 		if (op == BPF_EXIT)
3679 			return true;
3680 		if (op == BPF_CALL) {
3681 			/* BPF to BPF call will reach here because of marking
3682 			 * caller saved clobber with DST_OP_NO_MARK for which we
3683 			 * don't care the register def because they are anyway
3684 			 * marked as NOT_INIT already.
3685 			 */
3686 			if (insn->src_reg == BPF_PSEUDO_CALL)
3687 				return false;
3688 			/* Helper call will reach here because of arg type
3689 			 * check, conservatively return TRUE.
3690 			 */
3691 			if (t == SRC_OP)
3692 				return true;
3693 
3694 			return false;
3695 		}
3696 	}
3697 
3698 	if (class == BPF_ALU64 && op == BPF_END && (insn->imm == 16 || insn->imm == 32))
3699 		return false;
3700 
3701 	if (class == BPF_ALU64 || class == BPF_JMP ||
3702 	    (class == BPF_ALU && op == BPF_END && insn->imm == 64))
3703 		return true;
3704 
3705 	if (class == BPF_ALU || class == BPF_JMP32)
3706 		return false;
3707 
3708 	if (class == BPF_LDX) {
3709 		if (t != SRC_OP)
3710 			return BPF_SIZE(code) == BPF_DW || BPF_MODE(code) == BPF_MEMSX;
3711 		/* LDX source must be ptr. */
3712 		return true;
3713 	}
3714 
3715 	if (class == BPF_STX) {
3716 		/* BPF_STX (including atomic variants) has one or more source
3717 		 * operands, one of which is a ptr. Check whether the caller is
3718 		 * asking about it.
3719 		 */
3720 		if (t == SRC_OP && reg->type != SCALAR_VALUE)
3721 			return true;
3722 		return BPF_SIZE(code) == BPF_DW;
3723 	}
3724 
3725 	if (class == BPF_LD) {
3726 		u8 mode = BPF_MODE(code);
3727 
3728 		/* LD_IMM64 */
3729 		if (mode == BPF_IMM)
3730 			return true;
3731 
3732 		/* Both LD_IND and LD_ABS return 32-bit data. */
3733 		if (t != SRC_OP)
3734 			return  false;
3735 
3736 		/* Implicit ctx ptr. */
3737 		if (regno == BPF_REG_6)
3738 			return true;
3739 
3740 		/* Explicit source could be any width. */
3741 		return true;
3742 	}
3743 
3744 	if (class == BPF_ST)
3745 		/* The only source register for BPF_ST is a ptr. */
3746 		return true;
3747 
3748 	/* Conservatively return true at default. */
3749 	return true;
3750 }
3751 
3752 /* Return the regno defined by the insn, or -1. */
insn_def_regno(const struct bpf_insn * insn)3753 static int insn_def_regno(const struct bpf_insn *insn)
3754 {
3755 	switch (BPF_CLASS(insn->code)) {
3756 	case BPF_JMP:
3757 	case BPF_JMP32:
3758 	case BPF_ST:
3759 		return -1;
3760 	case BPF_STX:
3761 		if (BPF_MODE(insn->code) == BPF_ATOMIC ||
3762 		    BPF_MODE(insn->code) == BPF_PROBE_ATOMIC) {
3763 			if (insn->imm == BPF_CMPXCHG)
3764 				return BPF_REG_0;
3765 			else if (insn->imm == BPF_LOAD_ACQ)
3766 				return insn->dst_reg;
3767 			else if (insn->imm & BPF_FETCH)
3768 				return insn->src_reg;
3769 		}
3770 		return -1;
3771 	default:
3772 		return insn->dst_reg;
3773 	}
3774 }
3775 
3776 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
insn_has_def32(struct bpf_verifier_env * env,struct bpf_insn * insn)3777 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
3778 {
3779 	int dst_reg = insn_def_regno(insn);
3780 
3781 	if (dst_reg == -1)
3782 		return false;
3783 
3784 	return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
3785 }
3786 
mark_insn_zext(struct bpf_verifier_env * env,struct bpf_reg_state * reg)3787 static void mark_insn_zext(struct bpf_verifier_env *env,
3788 			   struct bpf_reg_state *reg)
3789 {
3790 	s32 def_idx = reg->subreg_def;
3791 
3792 	if (def_idx == DEF_NOT_SUBREG)
3793 		return;
3794 
3795 	env->insn_aux_data[def_idx - 1].zext_dst = true;
3796 	/* The dst will be zero extended, so won't be sub-register anymore. */
3797 	reg->subreg_def = DEF_NOT_SUBREG;
3798 }
3799 
__check_reg_arg(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,enum reg_arg_type t)3800 static int __check_reg_arg(struct bpf_verifier_env *env, struct bpf_reg_state *regs, u32 regno,
3801 			   enum reg_arg_type t)
3802 {
3803 	struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3804 	struct bpf_reg_state *reg;
3805 	bool rw64;
3806 
3807 	if (regno >= MAX_BPF_REG) {
3808 		verbose(env, "R%d is invalid\n", regno);
3809 		return -EINVAL;
3810 	}
3811 
3812 	mark_reg_scratched(env, regno);
3813 
3814 	reg = &regs[regno];
3815 	rw64 = is_reg64(env, insn, regno, reg, t);
3816 	if (t == SRC_OP) {
3817 		/* check whether register used as source operand can be read */
3818 		if (reg->type == NOT_INIT) {
3819 			verbose(env, "R%d !read_ok\n", regno);
3820 			return -EACCES;
3821 		}
3822 		/* We don't need to worry about FP liveness because it's read-only */
3823 		if (regno == BPF_REG_FP)
3824 			return 0;
3825 
3826 		if (rw64)
3827 			mark_insn_zext(env, reg);
3828 
3829 		return mark_reg_read(env, reg, reg->parent,
3830 				     rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
3831 	} else {
3832 		/* check whether register used as dest operand can be written to */
3833 		if (regno == BPF_REG_FP) {
3834 			verbose(env, "frame pointer is read only\n");
3835 			return -EACCES;
3836 		}
3837 		reg->live |= REG_LIVE_WRITTEN;
3838 		reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3839 		if (t == DST_OP)
3840 			mark_reg_unknown(env, regs, regno);
3841 	}
3842 	return 0;
3843 }
3844 
check_reg_arg(struct bpf_verifier_env * env,u32 regno,enum reg_arg_type t)3845 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3846 			 enum reg_arg_type t)
3847 {
3848 	struct bpf_verifier_state *vstate = env->cur_state;
3849 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
3850 
3851 	return __check_reg_arg(env, state->regs, regno, t);
3852 }
3853 
insn_stack_access_flags(int frameno,int spi)3854 static int insn_stack_access_flags(int frameno, int spi)
3855 {
3856 	return INSN_F_STACK_ACCESS | (spi << INSN_F_SPI_SHIFT) | frameno;
3857 }
3858 
insn_stack_access_spi(int insn_flags)3859 static int insn_stack_access_spi(int insn_flags)
3860 {
3861 	return (insn_flags >> INSN_F_SPI_SHIFT) & INSN_F_SPI_MASK;
3862 }
3863 
insn_stack_access_frameno(int insn_flags)3864 static int insn_stack_access_frameno(int insn_flags)
3865 {
3866 	return insn_flags & INSN_F_FRAMENO_MASK;
3867 }
3868 
mark_jmp_point(struct bpf_verifier_env * env,int idx)3869 static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3870 {
3871 	env->insn_aux_data[idx].jmp_point = true;
3872 }
3873 
is_jmp_point(struct bpf_verifier_env * env,int insn_idx)3874 static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3875 {
3876 	return env->insn_aux_data[insn_idx].jmp_point;
3877 }
3878 
3879 #define LR_FRAMENO_BITS	3
3880 #define LR_SPI_BITS	6
3881 #define LR_ENTRY_BITS	(LR_SPI_BITS + LR_FRAMENO_BITS + 1)
3882 #define LR_SIZE_BITS	4
3883 #define LR_FRAMENO_MASK	((1ull << LR_FRAMENO_BITS) - 1)
3884 #define LR_SPI_MASK	((1ull << LR_SPI_BITS)     - 1)
3885 #define LR_SIZE_MASK	((1ull << LR_SIZE_BITS)    - 1)
3886 #define LR_SPI_OFF	LR_FRAMENO_BITS
3887 #define LR_IS_REG_OFF	(LR_SPI_BITS + LR_FRAMENO_BITS)
3888 #define LINKED_REGS_MAX	6
3889 
3890 struct linked_reg {
3891 	u8 frameno;
3892 	union {
3893 		u8 spi;
3894 		u8 regno;
3895 	};
3896 	bool is_reg;
3897 };
3898 
3899 struct linked_regs {
3900 	int cnt;
3901 	struct linked_reg entries[LINKED_REGS_MAX];
3902 };
3903 
linked_regs_push(struct linked_regs * s)3904 static struct linked_reg *linked_regs_push(struct linked_regs *s)
3905 {
3906 	if (s->cnt < LINKED_REGS_MAX)
3907 		return &s->entries[s->cnt++];
3908 
3909 	return NULL;
3910 }
3911 
3912 /* Use u64 as a vector of 6 10-bit values, use first 4-bits to track
3913  * number of elements currently in stack.
3914  * Pack one history entry for linked registers as 10 bits in the following format:
3915  * - 3-bits frameno
3916  * - 6-bits spi_or_reg
3917  * - 1-bit  is_reg
3918  */
linked_regs_pack(struct linked_regs * s)3919 static u64 linked_regs_pack(struct linked_regs *s)
3920 {
3921 	u64 val = 0;
3922 	int i;
3923 
3924 	for (i = 0; i < s->cnt; ++i) {
3925 		struct linked_reg *e = &s->entries[i];
3926 		u64 tmp = 0;
3927 
3928 		tmp |= e->frameno;
3929 		tmp |= e->spi << LR_SPI_OFF;
3930 		tmp |= (e->is_reg ? 1 : 0) << LR_IS_REG_OFF;
3931 
3932 		val <<= LR_ENTRY_BITS;
3933 		val |= tmp;
3934 	}
3935 	val <<= LR_SIZE_BITS;
3936 	val |= s->cnt;
3937 	return val;
3938 }
3939 
linked_regs_unpack(u64 val,struct linked_regs * s)3940 static void linked_regs_unpack(u64 val, struct linked_regs *s)
3941 {
3942 	int i;
3943 
3944 	s->cnt = val & LR_SIZE_MASK;
3945 	val >>= LR_SIZE_BITS;
3946 
3947 	for (i = 0; i < s->cnt; ++i) {
3948 		struct linked_reg *e = &s->entries[i];
3949 
3950 		e->frameno =  val & LR_FRAMENO_MASK;
3951 		e->spi     = (val >> LR_SPI_OFF) & LR_SPI_MASK;
3952 		e->is_reg  = (val >> LR_IS_REG_OFF) & 0x1;
3953 		val >>= LR_ENTRY_BITS;
3954 	}
3955 }
3956 
3957 /* for any branch, call, exit record the history of jmps in the given state */
push_jmp_history(struct bpf_verifier_env * env,struct bpf_verifier_state * cur,int insn_flags,u64 linked_regs)3958 static int push_jmp_history(struct bpf_verifier_env *env, struct bpf_verifier_state *cur,
3959 			    int insn_flags, u64 linked_regs)
3960 {
3961 	u32 cnt = cur->jmp_history_cnt;
3962 	struct bpf_jmp_history_entry *p;
3963 	size_t alloc_size;
3964 
3965 	/* combine instruction flags if we already recorded this instruction */
3966 	if (env->cur_hist_ent) {
3967 		/* atomic instructions push insn_flags twice, for READ and
3968 		 * WRITE sides, but they should agree on stack slot
3969 		 */
3970 		verifier_bug_if((env->cur_hist_ent->flags & insn_flags) &&
3971 				(env->cur_hist_ent->flags & insn_flags) != insn_flags,
3972 				env, "insn history: insn_idx %d cur flags %x new flags %x",
3973 				env->insn_idx, env->cur_hist_ent->flags, insn_flags);
3974 		env->cur_hist_ent->flags |= insn_flags;
3975 		verifier_bug_if(env->cur_hist_ent->linked_regs != 0, env,
3976 				"insn history: insn_idx %d linked_regs: %#llx",
3977 				env->insn_idx, env->cur_hist_ent->linked_regs);
3978 		env->cur_hist_ent->linked_regs = linked_regs;
3979 		return 0;
3980 	}
3981 
3982 	cnt++;
3983 	alloc_size = kmalloc_size_roundup(size_mul(cnt, sizeof(*p)));
3984 	p = krealloc(cur->jmp_history, alloc_size, GFP_KERNEL_ACCOUNT);
3985 	if (!p)
3986 		return -ENOMEM;
3987 	cur->jmp_history = p;
3988 
3989 	p = &cur->jmp_history[cnt - 1];
3990 	p->idx = env->insn_idx;
3991 	p->prev_idx = env->prev_insn_idx;
3992 	p->flags = insn_flags;
3993 	p->linked_regs = linked_regs;
3994 	cur->jmp_history_cnt = cnt;
3995 	env->cur_hist_ent = p;
3996 
3997 	return 0;
3998 }
3999 
get_jmp_hist_entry(struct bpf_verifier_state * st,u32 hist_end,int insn_idx)4000 static struct bpf_jmp_history_entry *get_jmp_hist_entry(struct bpf_verifier_state *st,
4001 						        u32 hist_end, int insn_idx)
4002 {
4003 	if (hist_end > 0 && st->jmp_history[hist_end - 1].idx == insn_idx)
4004 		return &st->jmp_history[hist_end - 1];
4005 	return NULL;
4006 }
4007 
4008 /* Backtrack one insn at a time. If idx is not at the top of recorded
4009  * history then previous instruction came from straight line execution.
4010  * Return -ENOENT if we exhausted all instructions within given state.
4011  *
4012  * It's legal to have a bit of a looping with the same starting and ending
4013  * insn index within the same state, e.g.: 3->4->5->3, so just because current
4014  * instruction index is the same as state's first_idx doesn't mean we are
4015  * done. If there is still some jump history left, we should keep going. We
4016  * need to take into account that we might have a jump history between given
4017  * state's parent and itself, due to checkpointing. In this case, we'll have
4018  * history entry recording a jump from last instruction of parent state and
4019  * first instruction of given state.
4020  */
get_prev_insn_idx(struct bpf_verifier_state * st,int i,u32 * history)4021 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
4022 			     u32 *history)
4023 {
4024 	u32 cnt = *history;
4025 
4026 	if (i == st->first_insn_idx) {
4027 		if (cnt == 0)
4028 			return -ENOENT;
4029 		if (cnt == 1 && st->jmp_history[0].idx == i)
4030 			return -ENOENT;
4031 	}
4032 
4033 	if (cnt && st->jmp_history[cnt - 1].idx == i) {
4034 		i = st->jmp_history[cnt - 1].prev_idx;
4035 		(*history)--;
4036 	} else {
4037 		i--;
4038 	}
4039 	return i;
4040 }
4041 
disasm_kfunc_name(void * data,const struct bpf_insn * insn)4042 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
4043 {
4044 	const struct btf_type *func;
4045 	struct btf *desc_btf;
4046 
4047 	if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
4048 		return NULL;
4049 
4050 	desc_btf = find_kfunc_desc_btf(data, insn->off);
4051 	if (IS_ERR(desc_btf))
4052 		return "<error>";
4053 
4054 	func = btf_type_by_id(desc_btf, insn->imm);
4055 	return btf_name_by_offset(desc_btf, func->name_off);
4056 }
4057 
verbose_insn(struct bpf_verifier_env * env,struct bpf_insn * insn)4058 static void verbose_insn(struct bpf_verifier_env *env, struct bpf_insn *insn)
4059 {
4060 	const struct bpf_insn_cbs cbs = {
4061 		.cb_call	= disasm_kfunc_name,
4062 		.cb_print	= verbose,
4063 		.private_data	= env,
4064 	};
4065 
4066 	print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
4067 }
4068 
bt_init(struct backtrack_state * bt,u32 frame)4069 static inline void bt_init(struct backtrack_state *bt, u32 frame)
4070 {
4071 	bt->frame = frame;
4072 }
4073 
bt_reset(struct backtrack_state * bt)4074 static inline void bt_reset(struct backtrack_state *bt)
4075 {
4076 	struct bpf_verifier_env *env = bt->env;
4077 
4078 	memset(bt, 0, sizeof(*bt));
4079 	bt->env = env;
4080 }
4081 
bt_empty(struct backtrack_state * bt)4082 static inline u32 bt_empty(struct backtrack_state *bt)
4083 {
4084 	u64 mask = 0;
4085 	int i;
4086 
4087 	for (i = 0; i <= bt->frame; i++)
4088 		mask |= bt->reg_masks[i] | bt->stack_masks[i];
4089 
4090 	return mask == 0;
4091 }
4092 
bt_subprog_enter(struct backtrack_state * bt)4093 static inline int bt_subprog_enter(struct backtrack_state *bt)
4094 {
4095 	if (bt->frame == MAX_CALL_FRAMES - 1) {
4096 		verifier_bug(bt->env, "subprog enter from frame %d", bt->frame);
4097 		return -EFAULT;
4098 	}
4099 	bt->frame++;
4100 	return 0;
4101 }
4102 
bt_subprog_exit(struct backtrack_state * bt)4103 static inline int bt_subprog_exit(struct backtrack_state *bt)
4104 {
4105 	if (bt->frame == 0) {
4106 		verifier_bug(bt->env, "subprog exit from frame 0");
4107 		return -EFAULT;
4108 	}
4109 	bt->frame--;
4110 	return 0;
4111 }
4112 
bt_set_frame_reg(struct backtrack_state * bt,u32 frame,u32 reg)4113 static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
4114 {
4115 	bt->reg_masks[frame] |= 1 << reg;
4116 }
4117 
bt_clear_frame_reg(struct backtrack_state * bt,u32 frame,u32 reg)4118 static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
4119 {
4120 	bt->reg_masks[frame] &= ~(1 << reg);
4121 }
4122 
bt_set_reg(struct backtrack_state * bt,u32 reg)4123 static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
4124 {
4125 	bt_set_frame_reg(bt, bt->frame, reg);
4126 }
4127 
bt_clear_reg(struct backtrack_state * bt,u32 reg)4128 static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
4129 {
4130 	bt_clear_frame_reg(bt, bt->frame, reg);
4131 }
4132 
bt_set_frame_slot(struct backtrack_state * bt,u32 frame,u32 slot)4133 static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
4134 {
4135 	bt->stack_masks[frame] |= 1ull << slot;
4136 }
4137 
bt_clear_frame_slot(struct backtrack_state * bt,u32 frame,u32 slot)4138 static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
4139 {
4140 	bt->stack_masks[frame] &= ~(1ull << slot);
4141 }
4142 
bt_frame_reg_mask(struct backtrack_state * bt,u32 frame)4143 static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
4144 {
4145 	return bt->reg_masks[frame];
4146 }
4147 
bt_reg_mask(struct backtrack_state * bt)4148 static inline u32 bt_reg_mask(struct backtrack_state *bt)
4149 {
4150 	return bt->reg_masks[bt->frame];
4151 }
4152 
bt_frame_stack_mask(struct backtrack_state * bt,u32 frame)4153 static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
4154 {
4155 	return bt->stack_masks[frame];
4156 }
4157 
bt_stack_mask(struct backtrack_state * bt)4158 static inline u64 bt_stack_mask(struct backtrack_state *bt)
4159 {
4160 	return bt->stack_masks[bt->frame];
4161 }
4162 
bt_is_reg_set(struct backtrack_state * bt,u32 reg)4163 static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
4164 {
4165 	return bt->reg_masks[bt->frame] & (1 << reg);
4166 }
4167 
bt_is_frame_reg_set(struct backtrack_state * bt,u32 frame,u32 reg)4168 static inline bool bt_is_frame_reg_set(struct backtrack_state *bt, u32 frame, u32 reg)
4169 {
4170 	return bt->reg_masks[frame] & (1 << reg);
4171 }
4172 
bt_is_frame_slot_set(struct backtrack_state * bt,u32 frame,u32 slot)4173 static inline bool bt_is_frame_slot_set(struct backtrack_state *bt, u32 frame, u32 slot)
4174 {
4175 	return bt->stack_masks[frame] & (1ull << slot);
4176 }
4177 
4178 /* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
fmt_reg_mask(char * buf,ssize_t buf_sz,u32 reg_mask)4179 static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
4180 {
4181 	DECLARE_BITMAP(mask, 64);
4182 	bool first = true;
4183 	int i, n;
4184 
4185 	buf[0] = '\0';
4186 
4187 	bitmap_from_u64(mask, reg_mask);
4188 	for_each_set_bit(i, mask, 32) {
4189 		n = snprintf(buf, buf_sz, "%sr%d", first ? "" : ",", i);
4190 		first = false;
4191 		buf += n;
4192 		buf_sz -= n;
4193 		if (buf_sz < 0)
4194 			break;
4195 	}
4196 }
4197 /* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
fmt_stack_mask(char * buf,ssize_t buf_sz,u64 stack_mask)4198 static void fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
4199 {
4200 	DECLARE_BITMAP(mask, 64);
4201 	bool first = true;
4202 	int i, n;
4203 
4204 	buf[0] = '\0';
4205 
4206 	bitmap_from_u64(mask, stack_mask);
4207 	for_each_set_bit(i, mask, 64) {
4208 		n = snprintf(buf, buf_sz, "%s%d", first ? "" : ",", -(i + 1) * 8);
4209 		first = false;
4210 		buf += n;
4211 		buf_sz -= n;
4212 		if (buf_sz < 0)
4213 			break;
4214 	}
4215 }
4216 
4217 /* If any register R in hist->linked_regs is marked as precise in bt,
4218  * do bt_set_frame_{reg,slot}(bt, R) for all registers in hist->linked_regs.
4219  */
bt_sync_linked_regs(struct backtrack_state * bt,struct bpf_jmp_history_entry * hist)4220 static void bt_sync_linked_regs(struct backtrack_state *bt, struct bpf_jmp_history_entry *hist)
4221 {
4222 	struct linked_regs linked_regs;
4223 	bool some_precise = false;
4224 	int i;
4225 
4226 	if (!hist || hist->linked_regs == 0)
4227 		return;
4228 
4229 	linked_regs_unpack(hist->linked_regs, &linked_regs);
4230 	for (i = 0; i < linked_regs.cnt; ++i) {
4231 		struct linked_reg *e = &linked_regs.entries[i];
4232 
4233 		if ((e->is_reg && bt_is_frame_reg_set(bt, e->frameno, e->regno)) ||
4234 		    (!e->is_reg && bt_is_frame_slot_set(bt, e->frameno, e->spi))) {
4235 			some_precise = true;
4236 			break;
4237 		}
4238 	}
4239 
4240 	if (!some_precise)
4241 		return;
4242 
4243 	for (i = 0; i < linked_regs.cnt; ++i) {
4244 		struct linked_reg *e = &linked_regs.entries[i];
4245 
4246 		if (e->is_reg)
4247 			bt_set_frame_reg(bt, e->frameno, e->regno);
4248 		else
4249 			bt_set_frame_slot(bt, e->frameno, e->spi);
4250 	}
4251 }
4252 
4253 static bool calls_callback(struct bpf_verifier_env *env, int insn_idx);
4254 
4255 /* For given verifier state backtrack_insn() is called from the last insn to
4256  * the first insn. Its purpose is to compute a bitmask of registers and
4257  * stack slots that needs precision in the parent verifier state.
4258  *
4259  * @idx is an index of the instruction we are currently processing;
4260  * @subseq_idx is an index of the subsequent instruction that:
4261  *   - *would be* executed next, if jump history is viewed in forward order;
4262  *   - *was* processed previously during backtracking.
4263  */
backtrack_insn(struct bpf_verifier_env * env,int idx,int subseq_idx,struct bpf_jmp_history_entry * hist,struct backtrack_state * bt)4264 static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
4265 			  struct bpf_jmp_history_entry *hist, struct backtrack_state *bt)
4266 {
4267 	struct bpf_insn *insn = env->prog->insnsi + idx;
4268 	u8 class = BPF_CLASS(insn->code);
4269 	u8 opcode = BPF_OP(insn->code);
4270 	u8 mode = BPF_MODE(insn->code);
4271 	u32 dreg = insn->dst_reg;
4272 	u32 sreg = insn->src_reg;
4273 	u32 spi, i, fr;
4274 
4275 	if (insn->code == 0)
4276 		return 0;
4277 	if (env->log.level & BPF_LOG_LEVEL2) {
4278 		fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_reg_mask(bt));
4279 		verbose(env, "mark_precise: frame%d: regs=%s ",
4280 			bt->frame, env->tmp_str_buf);
4281 		fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN, bt_stack_mask(bt));
4282 		verbose(env, "stack=%s before ", env->tmp_str_buf);
4283 		verbose(env, "%d: ", idx);
4284 		verbose_insn(env, insn);
4285 	}
4286 
4287 	/* If there is a history record that some registers gained range at this insn,
4288 	 * propagate precision marks to those registers, so that bt_is_reg_set()
4289 	 * accounts for these registers.
4290 	 */
4291 	bt_sync_linked_regs(bt, hist);
4292 
4293 	if (class == BPF_ALU || class == BPF_ALU64) {
4294 		if (!bt_is_reg_set(bt, dreg))
4295 			return 0;
4296 		if (opcode == BPF_END || opcode == BPF_NEG) {
4297 			/* sreg is reserved and unused
4298 			 * dreg still need precision before this insn
4299 			 */
4300 			return 0;
4301 		} else if (opcode == BPF_MOV) {
4302 			if (BPF_SRC(insn->code) == BPF_X) {
4303 				/* dreg = sreg or dreg = (s8, s16, s32)sreg
4304 				 * dreg needs precision after this insn
4305 				 * sreg needs precision before this insn
4306 				 */
4307 				bt_clear_reg(bt, dreg);
4308 				if (sreg != BPF_REG_FP)
4309 					bt_set_reg(bt, sreg);
4310 			} else {
4311 				/* dreg = K
4312 				 * dreg needs precision after this insn.
4313 				 * Corresponding register is already marked
4314 				 * as precise=true in this verifier state.
4315 				 * No further markings in parent are necessary
4316 				 */
4317 				bt_clear_reg(bt, dreg);
4318 			}
4319 		} else {
4320 			if (BPF_SRC(insn->code) == BPF_X) {
4321 				/* dreg += sreg
4322 				 * both dreg and sreg need precision
4323 				 * before this insn
4324 				 */
4325 				if (sreg != BPF_REG_FP)
4326 					bt_set_reg(bt, sreg);
4327 			} /* else dreg += K
4328 			   * dreg still needs precision before this insn
4329 			   */
4330 		}
4331 	} else if (class == BPF_LDX || is_atomic_load_insn(insn)) {
4332 		if (!bt_is_reg_set(bt, dreg))
4333 			return 0;
4334 		bt_clear_reg(bt, dreg);
4335 
4336 		/* scalars can only be spilled into stack w/o losing precision.
4337 		 * Load from any other memory can be zero extended.
4338 		 * The desire to keep that precision is already indicated
4339 		 * by 'precise' mark in corresponding register of this state.
4340 		 * No further tracking necessary.
4341 		 */
4342 		if (!hist || !(hist->flags & INSN_F_STACK_ACCESS))
4343 			return 0;
4344 		/* dreg = *(u64 *)[fp - off] was a fill from the stack.
4345 		 * that [fp - off] slot contains scalar that needs to be
4346 		 * tracked with precision
4347 		 */
4348 		spi = insn_stack_access_spi(hist->flags);
4349 		fr = insn_stack_access_frameno(hist->flags);
4350 		bt_set_frame_slot(bt, fr, spi);
4351 	} else if (class == BPF_STX || class == BPF_ST) {
4352 		if (bt_is_reg_set(bt, dreg))
4353 			/* stx & st shouldn't be using _scalar_ dst_reg
4354 			 * to access memory. It means backtracking
4355 			 * encountered a case of pointer subtraction.
4356 			 */
4357 			return -ENOTSUPP;
4358 		/* scalars can only be spilled into stack */
4359 		if (!hist || !(hist->flags & INSN_F_STACK_ACCESS))
4360 			return 0;
4361 		spi = insn_stack_access_spi(hist->flags);
4362 		fr = insn_stack_access_frameno(hist->flags);
4363 		if (!bt_is_frame_slot_set(bt, fr, spi))
4364 			return 0;
4365 		bt_clear_frame_slot(bt, fr, spi);
4366 		if (class == BPF_STX)
4367 			bt_set_reg(bt, sreg);
4368 	} else if (class == BPF_JMP || class == BPF_JMP32) {
4369 		if (bpf_pseudo_call(insn)) {
4370 			int subprog_insn_idx, subprog;
4371 
4372 			subprog_insn_idx = idx + insn->imm + 1;
4373 			subprog = find_subprog(env, subprog_insn_idx);
4374 			if (subprog < 0)
4375 				return -EFAULT;
4376 
4377 			if (subprog_is_global(env, subprog)) {
4378 				/* check that jump history doesn't have any
4379 				 * extra instructions from subprog; the next
4380 				 * instruction after call to global subprog
4381 				 * should be literally next instruction in
4382 				 * caller program
4383 				 */
4384 				verifier_bug_if(idx + 1 != subseq_idx, env,
4385 						"extra insn from subprog");
4386 				/* r1-r5 are invalidated after subprog call,
4387 				 * so for global func call it shouldn't be set
4388 				 * anymore
4389 				 */
4390 				if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
4391 					verifier_bug(env, "global subprog unexpected regs %x",
4392 						     bt_reg_mask(bt));
4393 					return -EFAULT;
4394 				}
4395 				/* global subprog always sets R0 */
4396 				bt_clear_reg(bt, BPF_REG_0);
4397 				return 0;
4398 			} else {
4399 				/* static subprog call instruction, which
4400 				 * means that we are exiting current subprog,
4401 				 * so only r1-r5 could be still requested as
4402 				 * precise, r0 and r6-r10 or any stack slot in
4403 				 * the current frame should be zero by now
4404 				 */
4405 				if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
4406 					verifier_bug(env, "static subprog unexpected regs %x",
4407 						     bt_reg_mask(bt));
4408 					return -EFAULT;
4409 				}
4410 				/* we are now tracking register spills correctly,
4411 				 * so any instance of leftover slots is a bug
4412 				 */
4413 				if (bt_stack_mask(bt) != 0) {
4414 					verifier_bug(env,
4415 						     "static subprog leftover stack slots %llx",
4416 						     bt_stack_mask(bt));
4417 					return -EFAULT;
4418 				}
4419 				/* propagate r1-r5 to the caller */
4420 				for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
4421 					if (bt_is_reg_set(bt, i)) {
4422 						bt_clear_reg(bt, i);
4423 						bt_set_frame_reg(bt, bt->frame - 1, i);
4424 					}
4425 				}
4426 				if (bt_subprog_exit(bt))
4427 					return -EFAULT;
4428 				return 0;
4429 			}
4430 		} else if (is_sync_callback_calling_insn(insn) && idx != subseq_idx - 1) {
4431 			/* exit from callback subprog to callback-calling helper or
4432 			 * kfunc call. Use idx/subseq_idx check to discern it from
4433 			 * straight line code backtracking.
4434 			 * Unlike the subprog call handling above, we shouldn't
4435 			 * propagate precision of r1-r5 (if any requested), as they are
4436 			 * not actually arguments passed directly to callback subprogs
4437 			 */
4438 			if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
4439 				verifier_bug(env, "callback unexpected regs %x",
4440 					     bt_reg_mask(bt));
4441 				return -EFAULT;
4442 			}
4443 			if (bt_stack_mask(bt) != 0) {
4444 				verifier_bug(env, "callback leftover stack slots %llx",
4445 					     bt_stack_mask(bt));
4446 				return -EFAULT;
4447 			}
4448 			/* clear r1-r5 in callback subprog's mask */
4449 			for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4450 				bt_clear_reg(bt, i);
4451 			if (bt_subprog_exit(bt))
4452 				return -EFAULT;
4453 			return 0;
4454 		} else if (opcode == BPF_CALL) {
4455 			/* kfunc with imm==0 is invalid and fixup_kfunc_call will
4456 			 * catch this error later. Make backtracking conservative
4457 			 * with ENOTSUPP.
4458 			 */
4459 			if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
4460 				return -ENOTSUPP;
4461 			/* regular helper call sets R0 */
4462 			bt_clear_reg(bt, BPF_REG_0);
4463 			if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
4464 				/* if backtracking was looking for registers R1-R5
4465 				 * they should have been found already.
4466 				 */
4467 				verifier_bug(env, "backtracking call unexpected regs %x",
4468 					     bt_reg_mask(bt));
4469 				return -EFAULT;
4470 			}
4471 		} else if (opcode == BPF_EXIT) {
4472 			bool r0_precise;
4473 
4474 			/* Backtracking to a nested function call, 'idx' is a part of
4475 			 * the inner frame 'subseq_idx' is a part of the outer frame.
4476 			 * In case of a regular function call, instructions giving
4477 			 * precision to registers R1-R5 should have been found already.
4478 			 * In case of a callback, it is ok to have R1-R5 marked for
4479 			 * backtracking, as these registers are set by the function
4480 			 * invoking callback.
4481 			 */
4482 			if (subseq_idx >= 0 && calls_callback(env, subseq_idx))
4483 				for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4484 					bt_clear_reg(bt, i);
4485 			if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
4486 				verifier_bug(env, "backtracking exit unexpected regs %x",
4487 					     bt_reg_mask(bt));
4488 				return -EFAULT;
4489 			}
4490 
4491 			/* BPF_EXIT in subprog or callback always returns
4492 			 * right after the call instruction, so by checking
4493 			 * whether the instruction at subseq_idx-1 is subprog
4494 			 * call or not we can distinguish actual exit from
4495 			 * *subprog* from exit from *callback*. In the former
4496 			 * case, we need to propagate r0 precision, if
4497 			 * necessary. In the former we never do that.
4498 			 */
4499 			r0_precise = subseq_idx - 1 >= 0 &&
4500 				     bpf_pseudo_call(&env->prog->insnsi[subseq_idx - 1]) &&
4501 				     bt_is_reg_set(bt, BPF_REG_0);
4502 
4503 			bt_clear_reg(bt, BPF_REG_0);
4504 			if (bt_subprog_enter(bt))
4505 				return -EFAULT;
4506 
4507 			if (r0_precise)
4508 				bt_set_reg(bt, BPF_REG_0);
4509 			/* r6-r9 and stack slots will stay set in caller frame
4510 			 * bitmasks until we return back from callee(s)
4511 			 */
4512 			return 0;
4513 		} else if (BPF_SRC(insn->code) == BPF_X) {
4514 			if (!bt_is_reg_set(bt, dreg) && !bt_is_reg_set(bt, sreg))
4515 				return 0;
4516 			/* dreg <cond> sreg
4517 			 * Both dreg and sreg need precision before
4518 			 * this insn. If only sreg was marked precise
4519 			 * before it would be equally necessary to
4520 			 * propagate it to dreg.
4521 			 */
4522 			if (!hist || !(hist->flags & INSN_F_SRC_REG_STACK))
4523 				bt_set_reg(bt, sreg);
4524 			if (!hist || !(hist->flags & INSN_F_DST_REG_STACK))
4525 				bt_set_reg(bt, dreg);
4526 		} else if (BPF_SRC(insn->code) == BPF_K) {
4527 			 /* dreg <cond> K
4528 			  * Only dreg still needs precision before
4529 			  * this insn, so for the K-based conditional
4530 			  * there is nothing new to be marked.
4531 			  */
4532 		}
4533 	} else if (class == BPF_LD) {
4534 		if (!bt_is_reg_set(bt, dreg))
4535 			return 0;
4536 		bt_clear_reg(bt, dreg);
4537 		/* It's ld_imm64 or ld_abs or ld_ind.
4538 		 * For ld_imm64 no further tracking of precision
4539 		 * into parent is necessary
4540 		 */
4541 		if (mode == BPF_IND || mode == BPF_ABS)
4542 			/* to be analyzed */
4543 			return -ENOTSUPP;
4544 	}
4545 	/* Propagate precision marks to linked registers, to account for
4546 	 * registers marked as precise in this function.
4547 	 */
4548 	bt_sync_linked_regs(bt, hist);
4549 	return 0;
4550 }
4551 
4552 /* the scalar precision tracking algorithm:
4553  * . at the start all registers have precise=false.
4554  * . scalar ranges are tracked as normal through alu and jmp insns.
4555  * . once precise value of the scalar register is used in:
4556  *   .  ptr + scalar alu
4557  *   . if (scalar cond K|scalar)
4558  *   .  helper_call(.., scalar, ...) where ARG_CONST is expected
4559  *   backtrack through the verifier states and mark all registers and
4560  *   stack slots with spilled constants that these scalar registers
4561  *   should be precise.
4562  * . during state pruning two registers (or spilled stack slots)
4563  *   are equivalent if both are not precise.
4564  *
4565  * Note the verifier cannot simply walk register parentage chain,
4566  * since many different registers and stack slots could have been
4567  * used to compute single precise scalar.
4568  *
4569  * The approach of starting with precise=true for all registers and then
4570  * backtrack to mark a register as not precise when the verifier detects
4571  * that program doesn't care about specific value (e.g., when helper
4572  * takes register as ARG_ANYTHING parameter) is not safe.
4573  *
4574  * It's ok to walk single parentage chain of the verifier states.
4575  * It's possible that this backtracking will go all the way till 1st insn.
4576  * All other branches will be explored for needing precision later.
4577  *
4578  * The backtracking needs to deal with cases like:
4579  *   R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0)
4580  * r9 -= r8
4581  * r5 = r9
4582  * if r5 > 0x79f goto pc+7
4583  *    R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
4584  * r5 += 1
4585  * ...
4586  * call bpf_perf_event_output#25
4587  *   where .arg5_type = ARG_CONST_SIZE_OR_ZERO
4588  *
4589  * and this case:
4590  * r6 = 1
4591  * call foo // uses callee's r6 inside to compute r0
4592  * r0 += r6
4593  * if r0 == 0 goto
4594  *
4595  * to track above reg_mask/stack_mask needs to be independent for each frame.
4596  *
4597  * Also if parent's curframe > frame where backtracking started,
4598  * the verifier need to mark registers in both frames, otherwise callees
4599  * may incorrectly prune callers. This is similar to
4600  * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
4601  *
4602  * For now backtracking falls back into conservative marking.
4603  */
mark_all_scalars_precise(struct bpf_verifier_env * env,struct bpf_verifier_state * st)4604 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
4605 				     struct bpf_verifier_state *st)
4606 {
4607 	struct bpf_func_state *func;
4608 	struct bpf_reg_state *reg;
4609 	int i, j;
4610 
4611 	if (env->log.level & BPF_LOG_LEVEL2) {
4612 		verbose(env, "mark_precise: frame%d: falling back to forcing all scalars precise\n",
4613 			st->curframe);
4614 	}
4615 
4616 	/* big hammer: mark all scalars precise in this path.
4617 	 * pop_stack may still get !precise scalars.
4618 	 * We also skip current state and go straight to first parent state,
4619 	 * because precision markings in current non-checkpointed state are
4620 	 * not needed. See why in the comment in __mark_chain_precision below.
4621 	 */
4622 	for (st = st->parent; st; st = st->parent) {
4623 		for (i = 0; i <= st->curframe; i++) {
4624 			func = st->frame[i];
4625 			for (j = 0; j < BPF_REG_FP; j++) {
4626 				reg = &func->regs[j];
4627 				if (reg->type != SCALAR_VALUE || reg->precise)
4628 					continue;
4629 				reg->precise = true;
4630 				if (env->log.level & BPF_LOG_LEVEL2) {
4631 					verbose(env, "force_precise: frame%d: forcing r%d to be precise\n",
4632 						i, j);
4633 				}
4634 			}
4635 			for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4636 				if (!is_spilled_reg(&func->stack[j]))
4637 					continue;
4638 				reg = &func->stack[j].spilled_ptr;
4639 				if (reg->type != SCALAR_VALUE || reg->precise)
4640 					continue;
4641 				reg->precise = true;
4642 				if (env->log.level & BPF_LOG_LEVEL2) {
4643 					verbose(env, "force_precise: frame%d: forcing fp%d to be precise\n",
4644 						i, -(j + 1) * 8);
4645 				}
4646 			}
4647 		}
4648 	}
4649 }
4650 
mark_all_scalars_imprecise(struct bpf_verifier_env * env,struct bpf_verifier_state * st)4651 static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
4652 {
4653 	struct bpf_func_state *func;
4654 	struct bpf_reg_state *reg;
4655 	int i, j;
4656 
4657 	for (i = 0; i <= st->curframe; i++) {
4658 		func = st->frame[i];
4659 		for (j = 0; j < BPF_REG_FP; j++) {
4660 			reg = &func->regs[j];
4661 			if (reg->type != SCALAR_VALUE)
4662 				continue;
4663 			reg->precise = false;
4664 		}
4665 		for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4666 			if (!is_spilled_reg(&func->stack[j]))
4667 				continue;
4668 			reg = &func->stack[j].spilled_ptr;
4669 			if (reg->type != SCALAR_VALUE)
4670 				continue;
4671 			reg->precise = false;
4672 		}
4673 	}
4674 }
4675 
4676 /*
4677  * __mark_chain_precision() backtracks BPF program instruction sequence and
4678  * chain of verifier states making sure that register *regno* (if regno >= 0)
4679  * and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
4680  * SCALARS, as well as any other registers and slots that contribute to
4681  * a tracked state of given registers/stack slots, depending on specific BPF
4682  * assembly instructions (see backtrack_insns() for exact instruction handling
4683  * logic). This backtracking relies on recorded jmp_history and is able to
4684  * traverse entire chain of parent states. This process ends only when all the
4685  * necessary registers/slots and their transitive dependencies are marked as
4686  * precise.
4687  *
4688  * One important and subtle aspect is that precise marks *do not matter* in
4689  * the currently verified state (current state). It is important to understand
4690  * why this is the case.
4691  *
4692  * First, note that current state is the state that is not yet "checkpointed",
4693  * i.e., it is not yet put into env->explored_states, and it has no children
4694  * states as well. It's ephemeral, and can end up either a) being discarded if
4695  * compatible explored state is found at some point or BPF_EXIT instruction is
4696  * reached or b) checkpointed and put into env->explored_states, branching out
4697  * into one or more children states.
4698  *
4699  * In the former case, precise markings in current state are completely
4700  * ignored by state comparison code (see regsafe() for details). Only
4701  * checkpointed ("old") state precise markings are important, and if old
4702  * state's register/slot is precise, regsafe() assumes current state's
4703  * register/slot as precise and checks value ranges exactly and precisely. If
4704  * states turn out to be compatible, current state's necessary precise
4705  * markings and any required parent states' precise markings are enforced
4706  * after the fact with propagate_precision() logic, after the fact. But it's
4707  * important to realize that in this case, even after marking current state
4708  * registers/slots as precise, we immediately discard current state. So what
4709  * actually matters is any of the precise markings propagated into current
4710  * state's parent states, which are always checkpointed (due to b) case above).
4711  * As such, for scenario a) it doesn't matter if current state has precise
4712  * markings set or not.
4713  *
4714  * Now, for the scenario b), checkpointing and forking into child(ren)
4715  * state(s). Note that before current state gets to checkpointing step, any
4716  * processed instruction always assumes precise SCALAR register/slot
4717  * knowledge: if precise value or range is useful to prune jump branch, BPF
4718  * verifier takes this opportunity enthusiastically. Similarly, when
4719  * register's value is used to calculate offset or memory address, exact
4720  * knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
4721  * what we mentioned above about state comparison ignoring precise markings
4722  * during state comparison, BPF verifier ignores and also assumes precise
4723  * markings *at will* during instruction verification process. But as verifier
4724  * assumes precision, it also propagates any precision dependencies across
4725  * parent states, which are not yet finalized, so can be further restricted
4726  * based on new knowledge gained from restrictions enforced by their children
4727  * states. This is so that once those parent states are finalized, i.e., when
4728  * they have no more active children state, state comparison logic in
4729  * is_state_visited() would enforce strict and precise SCALAR ranges, if
4730  * required for correctness.
4731  *
4732  * To build a bit more intuition, note also that once a state is checkpointed,
4733  * the path we took to get to that state is not important. This is crucial
4734  * property for state pruning. When state is checkpointed and finalized at
4735  * some instruction index, it can be correctly and safely used to "short
4736  * circuit" any *compatible* state that reaches exactly the same instruction
4737  * index. I.e., if we jumped to that instruction from a completely different
4738  * code path than original finalized state was derived from, it doesn't
4739  * matter, current state can be discarded because from that instruction
4740  * forward having a compatible state will ensure we will safely reach the
4741  * exit. States describe preconditions for further exploration, but completely
4742  * forget the history of how we got here.
4743  *
4744  * This also means that even if we needed precise SCALAR range to get to
4745  * finalized state, but from that point forward *that same* SCALAR register is
4746  * never used in a precise context (i.e., it's precise value is not needed for
4747  * correctness), it's correct and safe to mark such register as "imprecise"
4748  * (i.e., precise marking set to false). This is what we rely on when we do
4749  * not set precise marking in current state. If no child state requires
4750  * precision for any given SCALAR register, it's safe to dictate that it can
4751  * be imprecise. If any child state does require this register to be precise,
4752  * we'll mark it precise later retroactively during precise markings
4753  * propagation from child state to parent states.
4754  *
4755  * Skipping precise marking setting in current state is a mild version of
4756  * relying on the above observation. But we can utilize this property even
4757  * more aggressively by proactively forgetting any precise marking in the
4758  * current state (which we inherited from the parent state), right before we
4759  * checkpoint it and branch off into new child state. This is done by
4760  * mark_all_scalars_imprecise() to hopefully get more permissive and generic
4761  * finalized states which help in short circuiting more future states.
4762  */
__mark_chain_precision(struct bpf_verifier_env * env,struct bpf_verifier_state * starting_state,int regno,bool * changed)4763 static int __mark_chain_precision(struct bpf_verifier_env *env,
4764 				  struct bpf_verifier_state *starting_state,
4765 				  int regno,
4766 				  bool *changed)
4767 {
4768 	struct bpf_verifier_state *st = starting_state;
4769 	struct backtrack_state *bt = &env->bt;
4770 	int first_idx = st->first_insn_idx;
4771 	int last_idx = starting_state->insn_idx;
4772 	int subseq_idx = -1;
4773 	struct bpf_func_state *func;
4774 	bool tmp, skip_first = true;
4775 	struct bpf_reg_state *reg;
4776 	int i, fr, err;
4777 
4778 	if (!env->bpf_capable)
4779 		return 0;
4780 
4781 	changed = changed ?: &tmp;
4782 	/* set frame number from which we are starting to backtrack */
4783 	bt_init(bt, starting_state->curframe);
4784 
4785 	/* Do sanity checks against current state of register and/or stack
4786 	 * slot, but don't set precise flag in current state, as precision
4787 	 * tracking in the current state is unnecessary.
4788 	 */
4789 	func = st->frame[bt->frame];
4790 	if (regno >= 0) {
4791 		reg = &func->regs[regno];
4792 		if (reg->type != SCALAR_VALUE) {
4793 			verifier_bug(env, "backtracking misuse");
4794 			return -EFAULT;
4795 		}
4796 		bt_set_reg(bt, regno);
4797 	}
4798 
4799 	if (bt_empty(bt))
4800 		return 0;
4801 
4802 	for (;;) {
4803 		DECLARE_BITMAP(mask, 64);
4804 		u32 history = st->jmp_history_cnt;
4805 		struct bpf_jmp_history_entry *hist;
4806 
4807 		if (env->log.level & BPF_LOG_LEVEL2) {
4808 			verbose(env, "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4809 				bt->frame, last_idx, first_idx, subseq_idx);
4810 		}
4811 
4812 		if (last_idx < 0) {
4813 			/* we are at the entry into subprog, which
4814 			 * is expected for global funcs, but only if
4815 			 * requested precise registers are R1-R5
4816 			 * (which are global func's input arguments)
4817 			 */
4818 			if (st->curframe == 0 &&
4819 			    st->frame[0]->subprogno > 0 &&
4820 			    st->frame[0]->callsite == BPF_MAIN_FUNC &&
4821 			    bt_stack_mask(bt) == 0 &&
4822 			    (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4823 				bitmap_from_u64(mask, bt_reg_mask(bt));
4824 				for_each_set_bit(i, mask, 32) {
4825 					reg = &st->frame[0]->regs[i];
4826 					bt_clear_reg(bt, i);
4827 					if (reg->type == SCALAR_VALUE) {
4828 						reg->precise = true;
4829 						*changed = true;
4830 					}
4831 				}
4832 				return 0;
4833 			}
4834 
4835 			verifier_bug(env, "backtracking func entry subprog %d reg_mask %x stack_mask %llx",
4836 				     st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4837 			return -EFAULT;
4838 		}
4839 
4840 		for (i = last_idx;;) {
4841 			if (skip_first) {
4842 				err = 0;
4843 				skip_first = false;
4844 			} else {
4845 				hist = get_jmp_hist_entry(st, history, i);
4846 				err = backtrack_insn(env, i, subseq_idx, hist, bt);
4847 			}
4848 			if (err == -ENOTSUPP) {
4849 				mark_all_scalars_precise(env, starting_state);
4850 				bt_reset(bt);
4851 				return 0;
4852 			} else if (err) {
4853 				return err;
4854 			}
4855 			if (bt_empty(bt))
4856 				/* Found assignment(s) into tracked register in this state.
4857 				 * Since this state is already marked, just return.
4858 				 * Nothing to be tracked further in the parent state.
4859 				 */
4860 				return 0;
4861 			subseq_idx = i;
4862 			i = get_prev_insn_idx(st, i, &history);
4863 			if (i == -ENOENT)
4864 				break;
4865 			if (i >= env->prog->len) {
4866 				/* This can happen if backtracking reached insn 0
4867 				 * and there are still reg_mask or stack_mask
4868 				 * to backtrack.
4869 				 * It means the backtracking missed the spot where
4870 				 * particular register was initialized with a constant.
4871 				 */
4872 				verifier_bug(env, "backtracking idx %d", i);
4873 				return -EFAULT;
4874 			}
4875 		}
4876 		st = st->parent;
4877 		if (!st)
4878 			break;
4879 
4880 		for (fr = bt->frame; fr >= 0; fr--) {
4881 			func = st->frame[fr];
4882 			bitmap_from_u64(mask, bt_frame_reg_mask(bt, fr));
4883 			for_each_set_bit(i, mask, 32) {
4884 				reg = &func->regs[i];
4885 				if (reg->type != SCALAR_VALUE) {
4886 					bt_clear_frame_reg(bt, fr, i);
4887 					continue;
4888 				}
4889 				if (reg->precise) {
4890 					bt_clear_frame_reg(bt, fr, i);
4891 				} else {
4892 					reg->precise = true;
4893 					*changed = true;
4894 				}
4895 			}
4896 
4897 			bitmap_from_u64(mask, bt_frame_stack_mask(bt, fr));
4898 			for_each_set_bit(i, mask, 64) {
4899 				if (verifier_bug_if(i >= func->allocated_stack / BPF_REG_SIZE,
4900 						    env, "stack slot %d, total slots %d",
4901 						    i, func->allocated_stack / BPF_REG_SIZE))
4902 					return -EFAULT;
4903 
4904 				if (!is_spilled_scalar_reg(&func->stack[i])) {
4905 					bt_clear_frame_slot(bt, fr, i);
4906 					continue;
4907 				}
4908 				reg = &func->stack[i].spilled_ptr;
4909 				if (reg->precise) {
4910 					bt_clear_frame_slot(bt, fr, i);
4911 				} else {
4912 					reg->precise = true;
4913 					*changed = true;
4914 				}
4915 			}
4916 			if (env->log.level & BPF_LOG_LEVEL2) {
4917 				fmt_reg_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4918 					     bt_frame_reg_mask(bt, fr));
4919 				verbose(env, "mark_precise: frame%d: parent state regs=%s ",
4920 					fr, env->tmp_str_buf);
4921 				fmt_stack_mask(env->tmp_str_buf, TMP_STR_BUF_LEN,
4922 					       bt_frame_stack_mask(bt, fr));
4923 				verbose(env, "stack=%s: ", env->tmp_str_buf);
4924 				print_verifier_state(env, st, fr, true);
4925 			}
4926 		}
4927 
4928 		if (bt_empty(bt))
4929 			return 0;
4930 
4931 		subseq_idx = first_idx;
4932 		last_idx = st->last_insn_idx;
4933 		first_idx = st->first_insn_idx;
4934 	}
4935 
4936 	/* if we still have requested precise regs or slots, we missed
4937 	 * something (e.g., stack access through non-r10 register), so
4938 	 * fallback to marking all precise
4939 	 */
4940 	if (!bt_empty(bt)) {
4941 		mark_all_scalars_precise(env, starting_state);
4942 		bt_reset(bt);
4943 	}
4944 
4945 	return 0;
4946 }
4947 
mark_chain_precision(struct bpf_verifier_env * env,int regno)4948 int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4949 {
4950 	return __mark_chain_precision(env, env->cur_state, regno, NULL);
4951 }
4952 
4953 /* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4954  * desired reg and stack masks across all relevant frames
4955  */
mark_chain_precision_batch(struct bpf_verifier_env * env,struct bpf_verifier_state * starting_state)4956 static int mark_chain_precision_batch(struct bpf_verifier_env *env,
4957 				      struct bpf_verifier_state *starting_state)
4958 {
4959 	return __mark_chain_precision(env, starting_state, -1, NULL);
4960 }
4961 
is_spillable_regtype(enum bpf_reg_type type)4962 static bool is_spillable_regtype(enum bpf_reg_type type)
4963 {
4964 	switch (base_type(type)) {
4965 	case PTR_TO_MAP_VALUE:
4966 	case PTR_TO_STACK:
4967 	case PTR_TO_CTX:
4968 	case PTR_TO_PACKET:
4969 	case PTR_TO_PACKET_META:
4970 	case PTR_TO_PACKET_END:
4971 	case PTR_TO_FLOW_KEYS:
4972 	case CONST_PTR_TO_MAP:
4973 	case PTR_TO_SOCKET:
4974 	case PTR_TO_SOCK_COMMON:
4975 	case PTR_TO_TCP_SOCK:
4976 	case PTR_TO_XDP_SOCK:
4977 	case PTR_TO_BTF_ID:
4978 	case PTR_TO_BUF:
4979 	case PTR_TO_MEM:
4980 	case PTR_TO_FUNC:
4981 	case PTR_TO_MAP_KEY:
4982 	case PTR_TO_ARENA:
4983 		return true;
4984 	default:
4985 		return false;
4986 	}
4987 }
4988 
4989 /* Does this register contain a constant zero? */
register_is_null(struct bpf_reg_state * reg)4990 static bool register_is_null(struct bpf_reg_state *reg)
4991 {
4992 	return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
4993 }
4994 
4995 /* check if register is a constant scalar value */
is_reg_const(struct bpf_reg_state * reg,bool subreg32)4996 static bool is_reg_const(struct bpf_reg_state *reg, bool subreg32)
4997 {
4998 	return reg->type == SCALAR_VALUE &&
4999 	       tnum_is_const(subreg32 ? tnum_subreg(reg->var_off) : reg->var_off);
5000 }
5001 
5002 /* assuming is_reg_const() is true, return constant value of a register */
reg_const_value(struct bpf_reg_state * reg,bool subreg32)5003 static u64 reg_const_value(struct bpf_reg_state *reg, bool subreg32)
5004 {
5005 	return subreg32 ? tnum_subreg(reg->var_off).value : reg->var_off.value;
5006 }
5007 
__is_pointer_value(bool allow_ptr_leaks,const struct bpf_reg_state * reg)5008 static bool __is_pointer_value(bool allow_ptr_leaks,
5009 			       const struct bpf_reg_state *reg)
5010 {
5011 	if (allow_ptr_leaks)
5012 		return false;
5013 
5014 	return reg->type != SCALAR_VALUE;
5015 }
5016 
assign_scalar_id_before_mov(struct bpf_verifier_env * env,struct bpf_reg_state * src_reg)5017 static void assign_scalar_id_before_mov(struct bpf_verifier_env *env,
5018 					struct bpf_reg_state *src_reg)
5019 {
5020 	if (src_reg->type != SCALAR_VALUE)
5021 		return;
5022 
5023 	if (src_reg->id & BPF_ADD_CONST) {
5024 		/*
5025 		 * The verifier is processing rX = rY insn and
5026 		 * rY->id has special linked register already.
5027 		 * Cleared it, since multiple rX += const are not supported.
5028 		 */
5029 		src_reg->id = 0;
5030 		src_reg->off = 0;
5031 	}
5032 
5033 	if (!src_reg->id && !tnum_is_const(src_reg->var_off))
5034 		/* Ensure that src_reg has a valid ID that will be copied to
5035 		 * dst_reg and then will be used by sync_linked_regs() to
5036 		 * propagate min/max range.
5037 		 */
5038 		src_reg->id = ++env->id_gen;
5039 }
5040 
5041 /* Copy src state preserving dst->parent and dst->live fields */
copy_register_state(struct bpf_reg_state * dst,const struct bpf_reg_state * src)5042 static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
5043 {
5044 	struct bpf_reg_state *parent = dst->parent;
5045 	enum bpf_reg_liveness live = dst->live;
5046 
5047 	*dst = *src;
5048 	dst->parent = parent;
5049 	dst->live = live;
5050 }
5051 
save_register_state(struct bpf_verifier_env * env,struct bpf_func_state * state,int spi,struct bpf_reg_state * reg,int size)5052 static void save_register_state(struct bpf_verifier_env *env,
5053 				struct bpf_func_state *state,
5054 				int spi, struct bpf_reg_state *reg,
5055 				int size)
5056 {
5057 	int i;
5058 
5059 	copy_register_state(&state->stack[spi].spilled_ptr, reg);
5060 	if (size == BPF_REG_SIZE)
5061 		state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
5062 
5063 	for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
5064 		state->stack[spi].slot_type[i - 1] = STACK_SPILL;
5065 
5066 	/* size < 8 bytes spill */
5067 	for (; i; i--)
5068 		mark_stack_slot_misc(env, &state->stack[spi].slot_type[i - 1]);
5069 }
5070 
is_bpf_st_mem(struct bpf_insn * insn)5071 static bool is_bpf_st_mem(struct bpf_insn *insn)
5072 {
5073 	return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
5074 }
5075 
get_reg_width(struct bpf_reg_state * reg)5076 static int get_reg_width(struct bpf_reg_state *reg)
5077 {
5078 	return fls64(reg->umax_value);
5079 }
5080 
5081 /* See comment for mark_fastcall_pattern_for_call() */
check_fastcall_stack_contract(struct bpf_verifier_env * env,struct bpf_func_state * state,int insn_idx,int off)5082 static void check_fastcall_stack_contract(struct bpf_verifier_env *env,
5083 					  struct bpf_func_state *state, int insn_idx, int off)
5084 {
5085 	struct bpf_subprog_info *subprog = &env->subprog_info[state->subprogno];
5086 	struct bpf_insn_aux_data *aux = env->insn_aux_data;
5087 	int i;
5088 
5089 	if (subprog->fastcall_stack_off <= off || aux[insn_idx].fastcall_pattern)
5090 		return;
5091 	/* access to the region [max_stack_depth .. fastcall_stack_off)
5092 	 * from something that is not a part of the fastcall pattern,
5093 	 * disable fastcall rewrites for current subprogram by setting
5094 	 * fastcall_stack_off to a value smaller than any possible offset.
5095 	 */
5096 	subprog->fastcall_stack_off = S16_MIN;
5097 	/* reset fastcall aux flags within subprogram,
5098 	 * happens at most once per subprogram
5099 	 */
5100 	for (i = subprog->start; i < (subprog + 1)->start; ++i) {
5101 		aux[i].fastcall_spills_num = 0;
5102 		aux[i].fastcall_pattern = 0;
5103 	}
5104 }
5105 
5106 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
5107  * stack boundary and alignment are checked in check_mem_access()
5108  */
check_stack_write_fixed_off(struct bpf_verifier_env * env,struct bpf_func_state * state,int off,int size,int value_regno,int insn_idx)5109 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
5110 				       /* stack frame we're writing to */
5111 				       struct bpf_func_state *state,
5112 				       int off, int size, int value_regno,
5113 				       int insn_idx)
5114 {
5115 	struct bpf_func_state *cur; /* state of the current function */
5116 	int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
5117 	struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5118 	struct bpf_reg_state *reg = NULL;
5119 	int insn_flags = insn_stack_access_flags(state->frameno, spi);
5120 
5121 	/* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
5122 	 * so it's aligned access and [off, off + size) are within stack limits
5123 	 */
5124 	if (!env->allow_ptr_leaks &&
5125 	    is_spilled_reg(&state->stack[spi]) &&
5126 	    !is_spilled_scalar_reg(&state->stack[spi]) &&
5127 	    size != BPF_REG_SIZE) {
5128 		verbose(env, "attempt to corrupt spilled pointer on stack\n");
5129 		return -EACCES;
5130 	}
5131 
5132 	cur = env->cur_state->frame[env->cur_state->curframe];
5133 	if (value_regno >= 0)
5134 		reg = &cur->regs[value_regno];
5135 	if (!env->bypass_spec_v4) {
5136 		bool sanitize = reg && is_spillable_regtype(reg->type);
5137 
5138 		for (i = 0; i < size; i++) {
5139 			u8 type = state->stack[spi].slot_type[i];
5140 
5141 			if (type != STACK_MISC && type != STACK_ZERO) {
5142 				sanitize = true;
5143 				break;
5144 			}
5145 		}
5146 
5147 		if (sanitize)
5148 			env->insn_aux_data[insn_idx].nospec_result = true;
5149 	}
5150 
5151 	err = destroy_if_dynptr_stack_slot(env, state, spi);
5152 	if (err)
5153 		return err;
5154 
5155 	check_fastcall_stack_contract(env, state, insn_idx, off);
5156 	mark_stack_slot_scratched(env, spi);
5157 	if (reg && !(off % BPF_REG_SIZE) && reg->type == SCALAR_VALUE && env->bpf_capable) {
5158 		bool reg_value_fits;
5159 
5160 		reg_value_fits = get_reg_width(reg) <= BITS_PER_BYTE * size;
5161 		/* Make sure that reg had an ID to build a relation on spill. */
5162 		if (reg_value_fits)
5163 			assign_scalar_id_before_mov(env, reg);
5164 		save_register_state(env, state, spi, reg, size);
5165 		/* Break the relation on a narrowing spill. */
5166 		if (!reg_value_fits)
5167 			state->stack[spi].spilled_ptr.id = 0;
5168 	} else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
5169 		   env->bpf_capable) {
5170 		struct bpf_reg_state *tmp_reg = &env->fake_reg[0];
5171 
5172 		memset(tmp_reg, 0, sizeof(*tmp_reg));
5173 		__mark_reg_known(tmp_reg, insn->imm);
5174 		tmp_reg->type = SCALAR_VALUE;
5175 		save_register_state(env, state, spi, tmp_reg, size);
5176 	} else if (reg && is_spillable_regtype(reg->type)) {
5177 		/* register containing pointer is being spilled into stack */
5178 		if (size != BPF_REG_SIZE) {
5179 			verbose_linfo(env, insn_idx, "; ");
5180 			verbose(env, "invalid size of register spill\n");
5181 			return -EACCES;
5182 		}
5183 		if (state != cur && reg->type == PTR_TO_STACK) {
5184 			verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
5185 			return -EINVAL;
5186 		}
5187 		save_register_state(env, state, spi, reg, size);
5188 	} else {
5189 		u8 type = STACK_MISC;
5190 
5191 		/* regular write of data into stack destroys any spilled ptr */
5192 		state->stack[spi].spilled_ptr.type = NOT_INIT;
5193 		/* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
5194 		if (is_stack_slot_special(&state->stack[spi]))
5195 			for (i = 0; i < BPF_REG_SIZE; i++)
5196 				scrub_spilled_slot(&state->stack[spi].slot_type[i]);
5197 
5198 		/* only mark the slot as written if all 8 bytes were written
5199 		 * otherwise read propagation may incorrectly stop too soon
5200 		 * when stack slots are partially written.
5201 		 * This heuristic means that read propagation will be
5202 		 * conservative, since it will add reg_live_read marks
5203 		 * to stack slots all the way to first state when programs
5204 		 * writes+reads less than 8 bytes
5205 		 */
5206 		if (size == BPF_REG_SIZE)
5207 			state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
5208 
5209 		/* when we zero initialize stack slots mark them as such */
5210 		if ((reg && register_is_null(reg)) ||
5211 		    (!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
5212 			/* STACK_ZERO case happened because register spill
5213 			 * wasn't properly aligned at the stack slot boundary,
5214 			 * so it's not a register spill anymore; force
5215 			 * originating register to be precise to make
5216 			 * STACK_ZERO correct for subsequent states
5217 			 */
5218 			err = mark_chain_precision(env, value_regno);
5219 			if (err)
5220 				return err;
5221 			type = STACK_ZERO;
5222 		}
5223 
5224 		/* Mark slots affected by this stack write. */
5225 		for (i = 0; i < size; i++)
5226 			state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] = type;
5227 		insn_flags = 0; /* not a register spill */
5228 	}
5229 
5230 	if (insn_flags)
5231 		return push_jmp_history(env, env->cur_state, insn_flags, 0);
5232 	return 0;
5233 }
5234 
5235 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
5236  * known to contain a variable offset.
5237  * This function checks whether the write is permitted and conservatively
5238  * tracks the effects of the write, considering that each stack slot in the
5239  * dynamic range is potentially written to.
5240  *
5241  * 'off' includes 'regno->off'.
5242  * 'value_regno' can be -1, meaning that an unknown value is being written to
5243  * the stack.
5244  *
5245  * Spilled pointers in range are not marked as written because we don't know
5246  * what's going to be actually written. This means that read propagation for
5247  * future reads cannot be terminated by this write.
5248  *
5249  * For privileged programs, uninitialized stack slots are considered
5250  * initialized by this write (even though we don't know exactly what offsets
5251  * are going to be written to). The idea is that we don't want the verifier to
5252  * reject future reads that access slots written to through variable offsets.
5253  */
check_stack_write_var_off(struct bpf_verifier_env * env,struct bpf_func_state * state,int ptr_regno,int off,int size,int value_regno,int insn_idx)5254 static int check_stack_write_var_off(struct bpf_verifier_env *env,
5255 				     /* func where register points to */
5256 				     struct bpf_func_state *state,
5257 				     int ptr_regno, int off, int size,
5258 				     int value_regno, int insn_idx)
5259 {
5260 	struct bpf_func_state *cur; /* state of the current function */
5261 	int min_off, max_off;
5262 	int i, err;
5263 	struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
5264 	struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5265 	bool writing_zero = false;
5266 	/* set if the fact that we're writing a zero is used to let any
5267 	 * stack slots remain STACK_ZERO
5268 	 */
5269 	bool zero_used = false;
5270 
5271 	cur = env->cur_state->frame[env->cur_state->curframe];
5272 	ptr_reg = &cur->regs[ptr_regno];
5273 	min_off = ptr_reg->smin_value + off;
5274 	max_off = ptr_reg->smax_value + off + size;
5275 	if (value_regno >= 0)
5276 		value_reg = &cur->regs[value_regno];
5277 	if ((value_reg && register_is_null(value_reg)) ||
5278 	    (!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
5279 		writing_zero = true;
5280 
5281 	for (i = min_off; i < max_off; i++) {
5282 		int spi;
5283 
5284 		spi = __get_spi(i);
5285 		err = destroy_if_dynptr_stack_slot(env, state, spi);
5286 		if (err)
5287 			return err;
5288 	}
5289 
5290 	check_fastcall_stack_contract(env, state, insn_idx, min_off);
5291 	/* Variable offset writes destroy any spilled pointers in range. */
5292 	for (i = min_off; i < max_off; i++) {
5293 		u8 new_type, *stype;
5294 		int slot, spi;
5295 
5296 		slot = -i - 1;
5297 		spi = slot / BPF_REG_SIZE;
5298 		stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
5299 		mark_stack_slot_scratched(env, spi);
5300 
5301 		if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
5302 			/* Reject the write if range we may write to has not
5303 			 * been initialized beforehand. If we didn't reject
5304 			 * here, the ptr status would be erased below (even
5305 			 * though not all slots are actually overwritten),
5306 			 * possibly opening the door to leaks.
5307 			 *
5308 			 * We do however catch STACK_INVALID case below, and
5309 			 * only allow reading possibly uninitialized memory
5310 			 * later for CAP_PERFMON, as the write may not happen to
5311 			 * that slot.
5312 			 */
5313 			verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
5314 				insn_idx, i);
5315 			return -EINVAL;
5316 		}
5317 
5318 		/* If writing_zero and the spi slot contains a spill of value 0,
5319 		 * maintain the spill type.
5320 		 */
5321 		if (writing_zero && *stype == STACK_SPILL &&
5322 		    is_spilled_scalar_reg(&state->stack[spi])) {
5323 			struct bpf_reg_state *spill_reg = &state->stack[spi].spilled_ptr;
5324 
5325 			if (tnum_is_const(spill_reg->var_off) && spill_reg->var_off.value == 0) {
5326 				zero_used = true;
5327 				continue;
5328 			}
5329 		}
5330 
5331 		/* Erase all other spilled pointers. */
5332 		state->stack[spi].spilled_ptr.type = NOT_INIT;
5333 
5334 		/* Update the slot type. */
5335 		new_type = STACK_MISC;
5336 		if (writing_zero && *stype == STACK_ZERO) {
5337 			new_type = STACK_ZERO;
5338 			zero_used = true;
5339 		}
5340 		/* If the slot is STACK_INVALID, we check whether it's OK to
5341 		 * pretend that it will be initialized by this write. The slot
5342 		 * might not actually be written to, and so if we mark it as
5343 		 * initialized future reads might leak uninitialized memory.
5344 		 * For privileged programs, we will accept such reads to slots
5345 		 * that may or may not be written because, if we're reject
5346 		 * them, the error would be too confusing.
5347 		 */
5348 		if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
5349 			verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
5350 					insn_idx, i);
5351 			return -EINVAL;
5352 		}
5353 		*stype = new_type;
5354 	}
5355 	if (zero_used) {
5356 		/* backtracking doesn't work for STACK_ZERO yet. */
5357 		err = mark_chain_precision(env, value_regno);
5358 		if (err)
5359 			return err;
5360 	}
5361 	return 0;
5362 }
5363 
5364 /* When register 'dst_regno' is assigned some values from stack[min_off,
5365  * max_off), we set the register's type according to the types of the
5366  * respective stack slots. If all the stack values are known to be zeros, then
5367  * so is the destination reg. Otherwise, the register is considered to be
5368  * SCALAR. This function does not deal with register filling; the caller must
5369  * ensure that all spilled registers in the stack range have been marked as
5370  * read.
5371  */
mark_reg_stack_read(struct bpf_verifier_env * env,struct bpf_func_state * ptr_state,int min_off,int max_off,int dst_regno)5372 static void mark_reg_stack_read(struct bpf_verifier_env *env,
5373 				/* func where src register points to */
5374 				struct bpf_func_state *ptr_state,
5375 				int min_off, int max_off, int dst_regno)
5376 {
5377 	struct bpf_verifier_state *vstate = env->cur_state;
5378 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
5379 	int i, slot, spi;
5380 	u8 *stype;
5381 	int zeros = 0;
5382 
5383 	for (i = min_off; i < max_off; i++) {
5384 		slot = -i - 1;
5385 		spi = slot / BPF_REG_SIZE;
5386 		mark_stack_slot_scratched(env, spi);
5387 		stype = ptr_state->stack[spi].slot_type;
5388 		if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
5389 			break;
5390 		zeros++;
5391 	}
5392 	if (zeros == max_off - min_off) {
5393 		/* Any access_size read into register is zero extended,
5394 		 * so the whole register == const_zero.
5395 		 */
5396 		__mark_reg_const_zero(env, &state->regs[dst_regno]);
5397 	} else {
5398 		/* have read misc data from the stack */
5399 		mark_reg_unknown(env, state->regs, dst_regno);
5400 	}
5401 	state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
5402 }
5403 
5404 /* Read the stack at 'off' and put the results into the register indicated by
5405  * 'dst_regno'. It handles reg filling if the addressed stack slot is a
5406  * spilled reg.
5407  *
5408  * 'dst_regno' can be -1, meaning that the read value is not going to a
5409  * register.
5410  *
5411  * The access is assumed to be within the current stack bounds.
5412  */
check_stack_read_fixed_off(struct bpf_verifier_env * env,struct bpf_func_state * reg_state,int off,int size,int dst_regno)5413 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
5414 				      /* func where src register points to */
5415 				      struct bpf_func_state *reg_state,
5416 				      int off, int size, int dst_regno)
5417 {
5418 	struct bpf_verifier_state *vstate = env->cur_state;
5419 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
5420 	int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
5421 	struct bpf_reg_state *reg;
5422 	u8 *stype, type;
5423 	int insn_flags = insn_stack_access_flags(reg_state->frameno, spi);
5424 
5425 	stype = reg_state->stack[spi].slot_type;
5426 	reg = &reg_state->stack[spi].spilled_ptr;
5427 
5428 	mark_stack_slot_scratched(env, spi);
5429 	check_fastcall_stack_contract(env, state, env->insn_idx, off);
5430 
5431 	if (is_spilled_reg(&reg_state->stack[spi])) {
5432 		u8 spill_size = 1;
5433 
5434 		for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
5435 			spill_size++;
5436 
5437 		if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
5438 			if (reg->type != SCALAR_VALUE) {
5439 				verbose_linfo(env, env->insn_idx, "; ");
5440 				verbose(env, "invalid size of register fill\n");
5441 				return -EACCES;
5442 			}
5443 
5444 			mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
5445 			if (dst_regno < 0)
5446 				return 0;
5447 
5448 			if (size <= spill_size &&
5449 			    bpf_stack_narrow_access_ok(off, size, spill_size)) {
5450 				/* The earlier check_reg_arg() has decided the
5451 				 * subreg_def for this insn.  Save it first.
5452 				 */
5453 				s32 subreg_def = state->regs[dst_regno].subreg_def;
5454 
5455 				copy_register_state(&state->regs[dst_regno], reg);
5456 				state->regs[dst_regno].subreg_def = subreg_def;
5457 
5458 				/* Break the relation on a narrowing fill.
5459 				 * coerce_reg_to_size will adjust the boundaries.
5460 				 */
5461 				if (get_reg_width(reg) > size * BITS_PER_BYTE)
5462 					state->regs[dst_regno].id = 0;
5463 			} else {
5464 				int spill_cnt = 0, zero_cnt = 0;
5465 
5466 				for (i = 0; i < size; i++) {
5467 					type = stype[(slot - i) % BPF_REG_SIZE];
5468 					if (type == STACK_SPILL) {
5469 						spill_cnt++;
5470 						continue;
5471 					}
5472 					if (type == STACK_MISC)
5473 						continue;
5474 					if (type == STACK_ZERO) {
5475 						zero_cnt++;
5476 						continue;
5477 					}
5478 					if (type == STACK_INVALID && env->allow_uninit_stack)
5479 						continue;
5480 					verbose(env, "invalid read from stack off %d+%d size %d\n",
5481 						off, i, size);
5482 					return -EACCES;
5483 				}
5484 
5485 				if (spill_cnt == size &&
5486 				    tnum_is_const(reg->var_off) && reg->var_off.value == 0) {
5487 					__mark_reg_const_zero(env, &state->regs[dst_regno]);
5488 					/* this IS register fill, so keep insn_flags */
5489 				} else if (zero_cnt == size) {
5490 					/* similarly to mark_reg_stack_read(), preserve zeroes */
5491 					__mark_reg_const_zero(env, &state->regs[dst_regno]);
5492 					insn_flags = 0; /* not restoring original register state */
5493 				} else {
5494 					mark_reg_unknown(env, state->regs, dst_regno);
5495 					insn_flags = 0; /* not restoring original register state */
5496 				}
5497 			}
5498 			state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
5499 		} else if (dst_regno >= 0) {
5500 			/* restore register state from stack */
5501 			copy_register_state(&state->regs[dst_regno], reg);
5502 			/* mark reg as written since spilled pointer state likely
5503 			 * has its liveness marks cleared by is_state_visited()
5504 			 * which resets stack/reg liveness for state transitions
5505 			 */
5506 			state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
5507 		} else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
5508 			/* If dst_regno==-1, the caller is asking us whether
5509 			 * it is acceptable to use this value as a SCALAR_VALUE
5510 			 * (e.g. for XADD).
5511 			 * We must not allow unprivileged callers to do that
5512 			 * with spilled pointers.
5513 			 */
5514 			verbose(env, "leaking pointer from stack off %d\n",
5515 				off);
5516 			return -EACCES;
5517 		}
5518 		mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
5519 	} else {
5520 		for (i = 0; i < size; i++) {
5521 			type = stype[(slot - i) % BPF_REG_SIZE];
5522 			if (type == STACK_MISC)
5523 				continue;
5524 			if (type == STACK_ZERO)
5525 				continue;
5526 			if (type == STACK_INVALID && env->allow_uninit_stack)
5527 				continue;
5528 			verbose(env, "invalid read from stack off %d+%d size %d\n",
5529 				off, i, size);
5530 			return -EACCES;
5531 		}
5532 		mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
5533 		if (dst_regno >= 0)
5534 			mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
5535 		insn_flags = 0; /* we are not restoring spilled register */
5536 	}
5537 	if (insn_flags)
5538 		return push_jmp_history(env, env->cur_state, insn_flags, 0);
5539 	return 0;
5540 }
5541 
5542 enum bpf_access_src {
5543 	ACCESS_DIRECT = 1,  /* the access is performed by an instruction */
5544 	ACCESS_HELPER = 2,  /* the access is performed by a helper */
5545 };
5546 
5547 static int check_stack_range_initialized(struct bpf_verifier_env *env,
5548 					 int regno, int off, int access_size,
5549 					 bool zero_size_allowed,
5550 					 enum bpf_access_type type,
5551 					 struct bpf_call_arg_meta *meta);
5552 
reg_state(struct bpf_verifier_env * env,int regno)5553 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
5554 {
5555 	return cur_regs(env) + regno;
5556 }
5557 
5558 /* Read the stack at 'ptr_regno + off' and put the result into the register
5559  * 'dst_regno'.
5560  * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
5561  * but not its variable offset.
5562  * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
5563  *
5564  * As opposed to check_stack_read_fixed_off, this function doesn't deal with
5565  * filling registers (i.e. reads of spilled register cannot be detected when
5566  * the offset is not fixed). We conservatively mark 'dst_regno' as containing
5567  * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
5568  * offset; for a fixed offset check_stack_read_fixed_off should be used
5569  * instead.
5570  */
check_stack_read_var_off(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)5571 static int check_stack_read_var_off(struct bpf_verifier_env *env,
5572 				    int ptr_regno, int off, int size, int dst_regno)
5573 {
5574 	/* The state of the source register. */
5575 	struct bpf_reg_state *reg = reg_state(env, ptr_regno);
5576 	struct bpf_func_state *ptr_state = func(env, reg);
5577 	int err;
5578 	int min_off, max_off;
5579 
5580 	/* Note that we pass a NULL meta, so raw access will not be permitted.
5581 	 */
5582 	err = check_stack_range_initialized(env, ptr_regno, off, size,
5583 					    false, BPF_READ, NULL);
5584 	if (err)
5585 		return err;
5586 
5587 	min_off = reg->smin_value + off;
5588 	max_off = reg->smax_value + off;
5589 	mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
5590 	check_fastcall_stack_contract(env, ptr_state, env->insn_idx, min_off);
5591 	return 0;
5592 }
5593 
5594 /* check_stack_read dispatches to check_stack_read_fixed_off or
5595  * check_stack_read_var_off.
5596  *
5597  * The caller must ensure that the offset falls within the allocated stack
5598  * bounds.
5599  *
5600  * 'dst_regno' is a register which will receive the value from the stack. It
5601  * can be -1, meaning that the read value is not going to a register.
5602  */
check_stack_read(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int dst_regno)5603 static int check_stack_read(struct bpf_verifier_env *env,
5604 			    int ptr_regno, int off, int size,
5605 			    int dst_regno)
5606 {
5607 	struct bpf_reg_state *reg = reg_state(env, ptr_regno);
5608 	struct bpf_func_state *state = func(env, reg);
5609 	int err;
5610 	/* Some accesses are only permitted with a static offset. */
5611 	bool var_off = !tnum_is_const(reg->var_off);
5612 
5613 	/* The offset is required to be static when reads don't go to a
5614 	 * register, in order to not leak pointers (see
5615 	 * check_stack_read_fixed_off).
5616 	 */
5617 	if (dst_regno < 0 && var_off) {
5618 		char tn_buf[48];
5619 
5620 		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5621 		verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
5622 			tn_buf, off, size);
5623 		return -EACCES;
5624 	}
5625 	/* Variable offset is prohibited for unprivileged mode for simplicity
5626 	 * since it requires corresponding support in Spectre masking for stack
5627 	 * ALU. See also retrieve_ptr_limit(). The check in
5628 	 * check_stack_access_for_ptr_arithmetic() called by
5629 	 * adjust_ptr_min_max_vals() prevents users from creating stack pointers
5630 	 * with variable offsets, therefore no check is required here. Further,
5631 	 * just checking it here would be insufficient as speculative stack
5632 	 * writes could still lead to unsafe speculative behaviour.
5633 	 */
5634 	if (!var_off) {
5635 		off += reg->var_off.value;
5636 		err = check_stack_read_fixed_off(env, state, off, size,
5637 						 dst_regno);
5638 	} else {
5639 		/* Variable offset stack reads need more conservative handling
5640 		 * than fixed offset ones. Note that dst_regno >= 0 on this
5641 		 * branch.
5642 		 */
5643 		err = check_stack_read_var_off(env, ptr_regno, off, size,
5644 					       dst_regno);
5645 	}
5646 	return err;
5647 }
5648 
5649 
5650 /* check_stack_write dispatches to check_stack_write_fixed_off or
5651  * check_stack_write_var_off.
5652  *
5653  * 'ptr_regno' is the register used as a pointer into the stack.
5654  * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
5655  * 'value_regno' is the register whose value we're writing to the stack. It can
5656  * be -1, meaning that we're not writing from a register.
5657  *
5658  * The caller must ensure that the offset falls within the maximum stack size.
5659  */
check_stack_write(struct bpf_verifier_env * env,int ptr_regno,int off,int size,int value_regno,int insn_idx)5660 static int check_stack_write(struct bpf_verifier_env *env,
5661 			     int ptr_regno, int off, int size,
5662 			     int value_regno, int insn_idx)
5663 {
5664 	struct bpf_reg_state *reg = reg_state(env, ptr_regno);
5665 	struct bpf_func_state *state = func(env, reg);
5666 	int err;
5667 
5668 	if (tnum_is_const(reg->var_off)) {
5669 		off += reg->var_off.value;
5670 		err = check_stack_write_fixed_off(env, state, off, size,
5671 						  value_regno, insn_idx);
5672 	} else {
5673 		/* Variable offset stack reads need more conservative handling
5674 		 * than fixed offset ones.
5675 		 */
5676 		err = check_stack_write_var_off(env, state,
5677 						ptr_regno, off, size,
5678 						value_regno, insn_idx);
5679 	}
5680 	return err;
5681 }
5682 
check_map_access_type(struct bpf_verifier_env * env,u32 regno,int off,int size,enum bpf_access_type type)5683 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
5684 				 int off, int size, enum bpf_access_type type)
5685 {
5686 	struct bpf_reg_state *regs = cur_regs(env);
5687 	struct bpf_map *map = regs[regno].map_ptr;
5688 	u32 cap = bpf_map_flags_to_cap(map);
5689 
5690 	if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
5691 		verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
5692 			map->value_size, off, size);
5693 		return -EACCES;
5694 	}
5695 
5696 	if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
5697 		verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
5698 			map->value_size, off, size);
5699 		return -EACCES;
5700 	}
5701 
5702 	return 0;
5703 }
5704 
5705 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
__check_mem_access(struct bpf_verifier_env * env,int regno,int off,int size,u32 mem_size,bool zero_size_allowed)5706 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
5707 			      int off, int size, u32 mem_size,
5708 			      bool zero_size_allowed)
5709 {
5710 	bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
5711 	struct bpf_reg_state *reg;
5712 
5713 	if (off >= 0 && size_ok && (u64)off + size <= mem_size)
5714 		return 0;
5715 
5716 	reg = &cur_regs(env)[regno];
5717 	switch (reg->type) {
5718 	case PTR_TO_MAP_KEY:
5719 		verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
5720 			mem_size, off, size);
5721 		break;
5722 	case PTR_TO_MAP_VALUE:
5723 		verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
5724 			mem_size, off, size);
5725 		break;
5726 	case PTR_TO_PACKET:
5727 	case PTR_TO_PACKET_META:
5728 	case PTR_TO_PACKET_END:
5729 		verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
5730 			off, size, regno, reg->id, off, mem_size);
5731 		break;
5732 	case PTR_TO_MEM:
5733 	default:
5734 		verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
5735 			mem_size, off, size);
5736 	}
5737 
5738 	return -EACCES;
5739 }
5740 
5741 /* check read/write into a memory region with possible variable offset */
check_mem_region_access(struct bpf_verifier_env * env,u32 regno,int off,int size,u32 mem_size,bool zero_size_allowed)5742 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
5743 				   int off, int size, u32 mem_size,
5744 				   bool zero_size_allowed)
5745 {
5746 	struct bpf_verifier_state *vstate = env->cur_state;
5747 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
5748 	struct bpf_reg_state *reg = &state->regs[regno];
5749 	int err;
5750 
5751 	/* We may have adjusted the register pointing to memory region, so we
5752 	 * need to try adding each of min_value and max_value to off
5753 	 * to make sure our theoretical access will be safe.
5754 	 *
5755 	 * The minimum value is only important with signed
5756 	 * comparisons where we can't assume the floor of a
5757 	 * value is 0.  If we are using signed variables for our
5758 	 * index'es we need to make sure that whatever we use
5759 	 * will have a set floor within our range.
5760 	 */
5761 	if (reg->smin_value < 0 &&
5762 	    (reg->smin_value == S64_MIN ||
5763 	     (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
5764 	      reg->smin_value + off < 0)) {
5765 		verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5766 			regno);
5767 		return -EACCES;
5768 	}
5769 	err = __check_mem_access(env, regno, reg->smin_value + off, size,
5770 				 mem_size, zero_size_allowed);
5771 	if (err) {
5772 		verbose(env, "R%d min value is outside of the allowed memory range\n",
5773 			regno);
5774 		return err;
5775 	}
5776 
5777 	/* If we haven't set a max value then we need to bail since we can't be
5778 	 * sure we won't do bad things.
5779 	 * If reg->umax_value + off could overflow, treat that as unbounded too.
5780 	 */
5781 	if (reg->umax_value >= BPF_MAX_VAR_OFF) {
5782 		verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
5783 			regno);
5784 		return -EACCES;
5785 	}
5786 	err = __check_mem_access(env, regno, reg->umax_value + off, size,
5787 				 mem_size, zero_size_allowed);
5788 	if (err) {
5789 		verbose(env, "R%d max value is outside of the allowed memory range\n",
5790 			regno);
5791 		return err;
5792 	}
5793 
5794 	return 0;
5795 }
5796 
__check_ptr_off_reg(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,bool fixed_off_ok)5797 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
5798 			       const struct bpf_reg_state *reg, int regno,
5799 			       bool fixed_off_ok)
5800 {
5801 	/* Access to this pointer-typed register or passing it to a helper
5802 	 * is only allowed in its original, unmodified form.
5803 	 */
5804 
5805 	if (reg->off < 0) {
5806 		verbose(env, "negative offset %s ptr R%d off=%d disallowed\n",
5807 			reg_type_str(env, reg->type), regno, reg->off);
5808 		return -EACCES;
5809 	}
5810 
5811 	if (!fixed_off_ok && reg->off) {
5812 		verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
5813 			reg_type_str(env, reg->type), regno, reg->off);
5814 		return -EACCES;
5815 	}
5816 
5817 	if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5818 		char tn_buf[48];
5819 
5820 		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5821 		verbose(env, "variable %s access var_off=%s disallowed\n",
5822 			reg_type_str(env, reg->type), tn_buf);
5823 		return -EACCES;
5824 	}
5825 
5826 	return 0;
5827 }
5828 
check_ptr_off_reg(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno)5829 static int check_ptr_off_reg(struct bpf_verifier_env *env,
5830 		             const struct bpf_reg_state *reg, int regno)
5831 {
5832 	return __check_ptr_off_reg(env, reg, regno, false);
5833 }
5834 
map_kptr_match_type(struct bpf_verifier_env * env,struct btf_field * kptr_field,struct bpf_reg_state * reg,u32 regno)5835 static int map_kptr_match_type(struct bpf_verifier_env *env,
5836 			       struct btf_field *kptr_field,
5837 			       struct bpf_reg_state *reg, u32 regno)
5838 {
5839 	const char *targ_name = btf_type_name(kptr_field->kptr.btf, kptr_field->kptr.btf_id);
5840 	int perm_flags;
5841 	const char *reg_name = "";
5842 
5843 	if (btf_is_kernel(reg->btf)) {
5844 		perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
5845 
5846 		/* Only unreferenced case accepts untrusted pointers */
5847 		if (kptr_field->type == BPF_KPTR_UNREF)
5848 			perm_flags |= PTR_UNTRUSTED;
5849 	} else {
5850 		perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5851 		if (kptr_field->type == BPF_KPTR_PERCPU)
5852 			perm_flags |= MEM_PERCPU;
5853 	}
5854 
5855 	if (base_type(reg->type) != PTR_TO_BTF_ID || (type_flag(reg->type) & ~perm_flags))
5856 		goto bad_type;
5857 
5858 	/* We need to verify reg->type and reg->btf, before accessing reg->btf */
5859 	reg_name = btf_type_name(reg->btf, reg->btf_id);
5860 
5861 	/* For ref_ptr case, release function check should ensure we get one
5862 	 * referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5863 	 * normal store of unreferenced kptr, we must ensure var_off is zero.
5864 	 * Since ref_ptr cannot be accessed directly by BPF insns, checks for
5865 	 * reg->off and reg->ref_obj_id are not needed here.
5866 	 */
5867 	if (__check_ptr_off_reg(env, reg, regno, true))
5868 		return -EACCES;
5869 
5870 	/* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5871 	 * we also need to take into account the reg->off.
5872 	 *
5873 	 * We want to support cases like:
5874 	 *
5875 	 * struct foo {
5876 	 *         struct bar br;
5877 	 *         struct baz bz;
5878 	 * };
5879 	 *
5880 	 * struct foo *v;
5881 	 * v = func();	      // PTR_TO_BTF_ID
5882 	 * val->foo = v;      // reg->off is zero, btf and btf_id match type
5883 	 * val->bar = &v->br; // reg->off is still zero, but we need to retry with
5884 	 *                    // first member type of struct after comparison fails
5885 	 * val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5886 	 *                    // to match type
5887 	 *
5888 	 * In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5889 	 * is zero. We must also ensure that btf_struct_ids_match does not walk
5890 	 * the struct to match type against first member of struct, i.e. reject
5891 	 * second case from above. Hence, when type is BPF_KPTR_REF, we set
5892 	 * strict mode to true for type match.
5893 	 */
5894 	if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5895 				  kptr_field->kptr.btf, kptr_field->kptr.btf_id,
5896 				  kptr_field->type != BPF_KPTR_UNREF))
5897 		goto bad_type;
5898 	return 0;
5899 bad_type:
5900 	verbose(env, "invalid kptr access, R%d type=%s%s ", regno,
5901 		reg_type_str(env, reg->type), reg_name);
5902 	verbose(env, "expected=%s%s", reg_type_str(env, PTR_TO_BTF_ID), targ_name);
5903 	if (kptr_field->type == BPF_KPTR_UNREF)
5904 		verbose(env, " or %s%s\n", reg_type_str(env, PTR_TO_BTF_ID | PTR_UNTRUSTED),
5905 			targ_name);
5906 	else
5907 		verbose(env, "\n");
5908 	return -EINVAL;
5909 }
5910 
in_sleepable(struct bpf_verifier_env * env)5911 static bool in_sleepable(struct bpf_verifier_env *env)
5912 {
5913 	return env->prog->sleepable ||
5914 	       (env->cur_state && env->cur_state->in_sleepable);
5915 }
5916 
5917 /* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5918  * can dereference RCU protected pointers and result is PTR_TRUSTED.
5919  */
in_rcu_cs(struct bpf_verifier_env * env)5920 static bool in_rcu_cs(struct bpf_verifier_env *env)
5921 {
5922 	return env->cur_state->active_rcu_lock ||
5923 	       env->cur_state->active_locks ||
5924 	       !in_sleepable(env);
5925 }
5926 
5927 /* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5928 BTF_SET_START(rcu_protected_types)
5929 #ifdef CONFIG_NET
BTF_ID(struct,prog_test_ref_kfunc)5930 BTF_ID(struct, prog_test_ref_kfunc)
5931 #endif
5932 #ifdef CONFIG_CGROUPS
5933 BTF_ID(struct, cgroup)
5934 #endif
5935 #ifdef CONFIG_BPF_JIT
5936 BTF_ID(struct, bpf_cpumask)
5937 #endif
5938 BTF_ID(struct, task_struct)
5939 #ifdef CONFIG_CRYPTO
5940 BTF_ID(struct, bpf_crypto_ctx)
5941 #endif
5942 BTF_SET_END(rcu_protected_types)
5943 
5944 static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5945 {
5946 	if (!btf_is_kernel(btf))
5947 		return true;
5948 	return btf_id_set_contains(&rcu_protected_types, btf_id);
5949 }
5950 
kptr_pointee_btf_record(struct btf_field * kptr_field)5951 static struct btf_record *kptr_pointee_btf_record(struct btf_field *kptr_field)
5952 {
5953 	struct btf_struct_meta *meta;
5954 
5955 	if (btf_is_kernel(kptr_field->kptr.btf))
5956 		return NULL;
5957 
5958 	meta = btf_find_struct_meta(kptr_field->kptr.btf,
5959 				    kptr_field->kptr.btf_id);
5960 
5961 	return meta ? meta->record : NULL;
5962 }
5963 
rcu_safe_kptr(const struct btf_field * field)5964 static bool rcu_safe_kptr(const struct btf_field *field)
5965 {
5966 	const struct btf_field_kptr *kptr = &field->kptr;
5967 
5968 	return field->type == BPF_KPTR_PERCPU ||
5969 	       (field->type == BPF_KPTR_REF && rcu_protected_object(kptr->btf, kptr->btf_id));
5970 }
5971 
btf_ld_kptr_type(struct bpf_verifier_env * env,struct btf_field * kptr_field)5972 static u32 btf_ld_kptr_type(struct bpf_verifier_env *env, struct btf_field *kptr_field)
5973 {
5974 	struct btf_record *rec;
5975 	u32 ret;
5976 
5977 	ret = PTR_MAYBE_NULL;
5978 	if (rcu_safe_kptr(kptr_field) && in_rcu_cs(env)) {
5979 		ret |= MEM_RCU;
5980 		if (kptr_field->type == BPF_KPTR_PERCPU)
5981 			ret |= MEM_PERCPU;
5982 		else if (!btf_is_kernel(kptr_field->kptr.btf))
5983 			ret |= MEM_ALLOC;
5984 
5985 		rec = kptr_pointee_btf_record(kptr_field);
5986 		if (rec && btf_record_has_field(rec, BPF_GRAPH_NODE))
5987 			ret |= NON_OWN_REF;
5988 	} else {
5989 		ret |= PTR_UNTRUSTED;
5990 	}
5991 
5992 	return ret;
5993 }
5994 
mark_uptr_ld_reg(struct bpf_verifier_env * env,u32 regno,struct btf_field * field)5995 static int mark_uptr_ld_reg(struct bpf_verifier_env *env, u32 regno,
5996 			    struct btf_field *field)
5997 {
5998 	struct bpf_reg_state *reg;
5999 	const struct btf_type *t;
6000 
6001 	t = btf_type_by_id(field->kptr.btf, field->kptr.btf_id);
6002 	mark_reg_known_zero(env, cur_regs(env), regno);
6003 	reg = reg_state(env, regno);
6004 	reg->type = PTR_TO_MEM | PTR_MAYBE_NULL;
6005 	reg->mem_size = t->size;
6006 	reg->id = ++env->id_gen;
6007 
6008 	return 0;
6009 }
6010 
check_map_kptr_access(struct bpf_verifier_env * env,u32 regno,int value_regno,int insn_idx,struct btf_field * kptr_field)6011 static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
6012 				 int value_regno, int insn_idx,
6013 				 struct btf_field *kptr_field)
6014 {
6015 	struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
6016 	int class = BPF_CLASS(insn->code);
6017 	struct bpf_reg_state *val_reg;
6018 	int ret;
6019 
6020 	/* Things we already checked for in check_map_access and caller:
6021 	 *  - Reject cases where variable offset may touch kptr
6022 	 *  - size of access (must be BPF_DW)
6023 	 *  - tnum_is_const(reg->var_off)
6024 	 *  - kptr_field->offset == off + reg->var_off.value
6025 	 */
6026 	/* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
6027 	if (BPF_MODE(insn->code) != BPF_MEM) {
6028 		verbose(env, "kptr in map can only be accessed using BPF_MEM instruction mode\n");
6029 		return -EACCES;
6030 	}
6031 
6032 	/* We only allow loading referenced kptr, since it will be marked as
6033 	 * untrusted, similar to unreferenced kptr.
6034 	 */
6035 	if (class != BPF_LDX &&
6036 	    (kptr_field->type == BPF_KPTR_REF || kptr_field->type == BPF_KPTR_PERCPU)) {
6037 		verbose(env, "store to referenced kptr disallowed\n");
6038 		return -EACCES;
6039 	}
6040 	if (class != BPF_LDX && kptr_field->type == BPF_UPTR) {
6041 		verbose(env, "store to uptr disallowed\n");
6042 		return -EACCES;
6043 	}
6044 
6045 	if (class == BPF_LDX) {
6046 		if (kptr_field->type == BPF_UPTR)
6047 			return mark_uptr_ld_reg(env, value_regno, kptr_field);
6048 
6049 		/* We can simply mark the value_regno receiving the pointer
6050 		 * value from map as PTR_TO_BTF_ID, with the correct type.
6051 		 */
6052 		ret = mark_btf_ld_reg(env, cur_regs(env), value_regno, PTR_TO_BTF_ID,
6053 				      kptr_field->kptr.btf, kptr_field->kptr.btf_id,
6054 				      btf_ld_kptr_type(env, kptr_field));
6055 		if (ret < 0)
6056 			return ret;
6057 	} else if (class == BPF_STX) {
6058 		val_reg = reg_state(env, value_regno);
6059 		if (!register_is_null(val_reg) &&
6060 		    map_kptr_match_type(env, kptr_field, val_reg, value_regno))
6061 			return -EACCES;
6062 	} else if (class == BPF_ST) {
6063 		if (insn->imm) {
6064 			verbose(env, "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
6065 				kptr_field->offset);
6066 			return -EACCES;
6067 		}
6068 	} else {
6069 		verbose(env, "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
6070 		return -EACCES;
6071 	}
6072 	return 0;
6073 }
6074 
6075 /* check read/write into a map element with possible variable offset */
check_map_access(struct bpf_verifier_env * env,u32 regno,int off,int size,bool zero_size_allowed,enum bpf_access_src src)6076 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
6077 			    int off, int size, bool zero_size_allowed,
6078 			    enum bpf_access_src src)
6079 {
6080 	struct bpf_verifier_state *vstate = env->cur_state;
6081 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
6082 	struct bpf_reg_state *reg = &state->regs[regno];
6083 	struct bpf_map *map = reg->map_ptr;
6084 	struct btf_record *rec;
6085 	int err, i;
6086 
6087 	err = check_mem_region_access(env, regno, off, size, map->value_size,
6088 				      zero_size_allowed);
6089 	if (err)
6090 		return err;
6091 
6092 	if (IS_ERR_OR_NULL(map->record))
6093 		return 0;
6094 	rec = map->record;
6095 	for (i = 0; i < rec->cnt; i++) {
6096 		struct btf_field *field = &rec->fields[i];
6097 		u32 p = field->offset;
6098 
6099 		/* If any part of a field  can be touched by load/store, reject
6100 		 * this program. To check that [x1, x2) overlaps with [y1, y2),
6101 		 * it is sufficient to check x1 < y2 && y1 < x2.
6102 		 */
6103 		if (reg->smin_value + off < p + field->size &&
6104 		    p < reg->umax_value + off + size) {
6105 			switch (field->type) {
6106 			case BPF_KPTR_UNREF:
6107 			case BPF_KPTR_REF:
6108 			case BPF_KPTR_PERCPU:
6109 			case BPF_UPTR:
6110 				if (src != ACCESS_DIRECT) {
6111 					verbose(env, "%s cannot be accessed indirectly by helper\n",
6112 						btf_field_type_name(field->type));
6113 					return -EACCES;
6114 				}
6115 				if (!tnum_is_const(reg->var_off)) {
6116 					verbose(env, "%s access cannot have variable offset\n",
6117 						btf_field_type_name(field->type));
6118 					return -EACCES;
6119 				}
6120 				if (p != off + reg->var_off.value) {
6121 					verbose(env, "%s access misaligned expected=%u off=%llu\n",
6122 						btf_field_type_name(field->type),
6123 						p, off + reg->var_off.value);
6124 					return -EACCES;
6125 				}
6126 				if (size != bpf_size_to_bytes(BPF_DW)) {
6127 					verbose(env, "%s access size must be BPF_DW\n",
6128 						btf_field_type_name(field->type));
6129 					return -EACCES;
6130 				}
6131 				break;
6132 			default:
6133 				verbose(env, "%s cannot be accessed directly by load/store\n",
6134 					btf_field_type_name(field->type));
6135 				return -EACCES;
6136 			}
6137 		}
6138 	}
6139 	return 0;
6140 }
6141 
6142 #define MAX_PACKET_OFF 0xffff
6143 
may_access_direct_pkt_data(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_access_type t)6144 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
6145 				       const struct bpf_call_arg_meta *meta,
6146 				       enum bpf_access_type t)
6147 {
6148 	enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
6149 
6150 	switch (prog_type) {
6151 	/* Program types only with direct read access go here! */
6152 	case BPF_PROG_TYPE_LWT_IN:
6153 	case BPF_PROG_TYPE_LWT_OUT:
6154 	case BPF_PROG_TYPE_LWT_SEG6LOCAL:
6155 	case BPF_PROG_TYPE_SK_REUSEPORT:
6156 	case BPF_PROG_TYPE_FLOW_DISSECTOR:
6157 	case BPF_PROG_TYPE_CGROUP_SKB:
6158 		if (t == BPF_WRITE)
6159 			return false;
6160 		fallthrough;
6161 
6162 	/* Program types with direct read + write access go here! */
6163 	case BPF_PROG_TYPE_SCHED_CLS:
6164 	case BPF_PROG_TYPE_SCHED_ACT:
6165 	case BPF_PROG_TYPE_XDP:
6166 	case BPF_PROG_TYPE_LWT_XMIT:
6167 	case BPF_PROG_TYPE_SK_SKB:
6168 	case BPF_PROG_TYPE_SK_MSG:
6169 		if (meta)
6170 			return meta->pkt_access;
6171 
6172 		env->seen_direct_write = true;
6173 		return true;
6174 
6175 	case BPF_PROG_TYPE_CGROUP_SOCKOPT:
6176 		if (t == BPF_WRITE)
6177 			env->seen_direct_write = true;
6178 
6179 		return true;
6180 
6181 	default:
6182 		return false;
6183 	}
6184 }
6185 
check_packet_access(struct bpf_verifier_env * env,u32 regno,int off,int size,bool zero_size_allowed)6186 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
6187 			       int size, bool zero_size_allowed)
6188 {
6189 	struct bpf_reg_state *regs = cur_regs(env);
6190 	struct bpf_reg_state *reg = &regs[regno];
6191 	int err;
6192 
6193 	/* We may have added a variable offset to the packet pointer; but any
6194 	 * reg->range we have comes after that.  We are only checking the fixed
6195 	 * offset.
6196 	 */
6197 
6198 	/* We don't allow negative numbers, because we aren't tracking enough
6199 	 * detail to prove they're safe.
6200 	 */
6201 	if (reg->smin_value < 0) {
6202 		verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
6203 			regno);
6204 		return -EACCES;
6205 	}
6206 
6207 	err = reg->range < 0 ? -EINVAL :
6208 	      __check_mem_access(env, regno, off, size, reg->range,
6209 				 zero_size_allowed);
6210 	if (err) {
6211 		verbose(env, "R%d offset is outside of the packet\n", regno);
6212 		return err;
6213 	}
6214 
6215 	/* __check_mem_access has made sure "off + size - 1" is within u16.
6216 	 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
6217 	 * otherwise find_good_pkt_pointers would have refused to set range info
6218 	 * that __check_mem_access would have rejected this pkt access.
6219 	 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
6220 	 */
6221 	env->prog->aux->max_pkt_offset =
6222 		max_t(u32, env->prog->aux->max_pkt_offset,
6223 		      off + reg->umax_value + size - 1);
6224 
6225 	return err;
6226 }
6227 
6228 /* check access to 'struct bpf_context' fields.  Supports fixed offsets only */
check_ctx_access(struct bpf_verifier_env * env,int insn_idx,int off,int size,enum bpf_access_type t,struct bpf_insn_access_aux * info)6229 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
6230 			    enum bpf_access_type t, struct bpf_insn_access_aux *info)
6231 {
6232 	if (env->ops->is_valid_access &&
6233 	    env->ops->is_valid_access(off, size, t, env->prog, info)) {
6234 		/* A non zero info.ctx_field_size indicates that this field is a
6235 		 * candidate for later verifier transformation to load the whole
6236 		 * field and then apply a mask when accessed with a narrower
6237 		 * access than actual ctx access size. A zero info.ctx_field_size
6238 		 * will only allow for whole field access and rejects any other
6239 		 * type of narrower access.
6240 		 */
6241 		if (base_type(info->reg_type) == PTR_TO_BTF_ID) {
6242 			if (info->ref_obj_id &&
6243 			    !find_reference_state(env->cur_state, info->ref_obj_id)) {
6244 				verbose(env, "invalid bpf_context access off=%d. Reference may already be released\n",
6245 					off);
6246 				return -EACCES;
6247 			}
6248 		} else {
6249 			env->insn_aux_data[insn_idx].ctx_field_size = info->ctx_field_size;
6250 		}
6251 		/* remember the offset of last byte accessed in ctx */
6252 		if (env->prog->aux->max_ctx_offset < off + size)
6253 			env->prog->aux->max_ctx_offset = off + size;
6254 		return 0;
6255 	}
6256 
6257 	verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
6258 	return -EACCES;
6259 }
6260 
check_flow_keys_access(struct bpf_verifier_env * env,int off,int size)6261 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
6262 				  int size)
6263 {
6264 	if (size < 0 || off < 0 ||
6265 	    (u64)off + size > sizeof(struct bpf_flow_keys)) {
6266 		verbose(env, "invalid access to flow keys off=%d size=%d\n",
6267 			off, size);
6268 		return -EACCES;
6269 	}
6270 	return 0;
6271 }
6272 
check_sock_access(struct bpf_verifier_env * env,int insn_idx,u32 regno,int off,int size,enum bpf_access_type t)6273 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
6274 			     u32 regno, int off, int size,
6275 			     enum bpf_access_type t)
6276 {
6277 	struct bpf_reg_state *regs = cur_regs(env);
6278 	struct bpf_reg_state *reg = &regs[regno];
6279 	struct bpf_insn_access_aux info = {};
6280 	bool valid;
6281 
6282 	if (reg->smin_value < 0) {
6283 		verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
6284 			regno);
6285 		return -EACCES;
6286 	}
6287 
6288 	switch (reg->type) {
6289 	case PTR_TO_SOCK_COMMON:
6290 		valid = bpf_sock_common_is_valid_access(off, size, t, &info);
6291 		break;
6292 	case PTR_TO_SOCKET:
6293 		valid = bpf_sock_is_valid_access(off, size, t, &info);
6294 		break;
6295 	case PTR_TO_TCP_SOCK:
6296 		valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
6297 		break;
6298 	case PTR_TO_XDP_SOCK:
6299 		valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
6300 		break;
6301 	default:
6302 		valid = false;
6303 	}
6304 
6305 
6306 	if (valid) {
6307 		env->insn_aux_data[insn_idx].ctx_field_size =
6308 			info.ctx_field_size;
6309 		return 0;
6310 	}
6311 
6312 	verbose(env, "R%d invalid %s access off=%d size=%d\n",
6313 		regno, reg_type_str(env, reg->type), off, size);
6314 
6315 	return -EACCES;
6316 }
6317 
is_pointer_value(struct bpf_verifier_env * env,int regno)6318 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
6319 {
6320 	return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
6321 }
6322 
is_ctx_reg(struct bpf_verifier_env * env,int regno)6323 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
6324 {
6325 	const struct bpf_reg_state *reg = reg_state(env, regno);
6326 
6327 	return reg->type == PTR_TO_CTX;
6328 }
6329 
is_sk_reg(struct bpf_verifier_env * env,int regno)6330 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
6331 {
6332 	const struct bpf_reg_state *reg = reg_state(env, regno);
6333 
6334 	return type_is_sk_pointer(reg->type);
6335 }
6336 
is_pkt_reg(struct bpf_verifier_env * env,int regno)6337 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
6338 {
6339 	const struct bpf_reg_state *reg = reg_state(env, regno);
6340 
6341 	return type_is_pkt_pointer(reg->type);
6342 }
6343 
is_flow_key_reg(struct bpf_verifier_env * env,int regno)6344 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
6345 {
6346 	const struct bpf_reg_state *reg = reg_state(env, regno);
6347 
6348 	/* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
6349 	return reg->type == PTR_TO_FLOW_KEYS;
6350 }
6351 
is_arena_reg(struct bpf_verifier_env * env,int regno)6352 static bool is_arena_reg(struct bpf_verifier_env *env, int regno)
6353 {
6354 	const struct bpf_reg_state *reg = reg_state(env, regno);
6355 
6356 	return reg->type == PTR_TO_ARENA;
6357 }
6358 
6359 /* Return false if @regno contains a pointer whose type isn't supported for
6360  * atomic instruction @insn.
6361  */
atomic_ptr_type_ok(struct bpf_verifier_env * env,int regno,struct bpf_insn * insn)6362 static bool atomic_ptr_type_ok(struct bpf_verifier_env *env, int regno,
6363 			       struct bpf_insn *insn)
6364 {
6365 	if (is_ctx_reg(env, regno))
6366 		return false;
6367 	if (is_pkt_reg(env, regno))
6368 		return false;
6369 	if (is_flow_key_reg(env, regno))
6370 		return false;
6371 	if (is_sk_reg(env, regno))
6372 		return false;
6373 	if (is_arena_reg(env, regno))
6374 		return bpf_jit_supports_insn(insn, true);
6375 
6376 	return true;
6377 }
6378 
6379 static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
6380 #ifdef CONFIG_NET
6381 	[PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
6382 	[PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
6383 	[PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
6384 #endif
6385 	[CONST_PTR_TO_MAP] = btf_bpf_map_id,
6386 };
6387 
is_trusted_reg(const struct bpf_reg_state * reg)6388 static bool is_trusted_reg(const struct bpf_reg_state *reg)
6389 {
6390 	/* A referenced register is always trusted. */
6391 	if (reg->ref_obj_id)
6392 		return true;
6393 
6394 	/* Types listed in the reg2btf_ids are always trusted */
6395 	if (reg2btf_ids[base_type(reg->type)] &&
6396 	    !bpf_type_has_unsafe_modifiers(reg->type))
6397 		return true;
6398 
6399 	/* If a register is not referenced, it is trusted if it has the
6400 	 * MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
6401 	 * other type modifiers may be safe, but we elect to take an opt-in
6402 	 * approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
6403 	 * not.
6404 	 *
6405 	 * Eventually, we should make PTR_TRUSTED the single source of truth
6406 	 * for whether a register is trusted.
6407 	 */
6408 	return type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
6409 	       !bpf_type_has_unsafe_modifiers(reg->type);
6410 }
6411 
is_rcu_reg(const struct bpf_reg_state * reg)6412 static bool is_rcu_reg(const struct bpf_reg_state *reg)
6413 {
6414 	return reg->type & MEM_RCU;
6415 }
6416 
clear_trusted_flags(enum bpf_type_flag * flag)6417 static void clear_trusted_flags(enum bpf_type_flag *flag)
6418 {
6419 	*flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
6420 }
6421 
check_pkt_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict)6422 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
6423 				   const struct bpf_reg_state *reg,
6424 				   int off, int size, bool strict)
6425 {
6426 	struct tnum reg_off;
6427 	int ip_align;
6428 
6429 	/* Byte size accesses are always allowed. */
6430 	if (!strict || size == 1)
6431 		return 0;
6432 
6433 	/* For platforms that do not have a Kconfig enabling
6434 	 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
6435 	 * NET_IP_ALIGN is universally set to '2'.  And on platforms
6436 	 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
6437 	 * to this code only in strict mode where we want to emulate
6438 	 * the NET_IP_ALIGN==2 checking.  Therefore use an
6439 	 * unconditional IP align value of '2'.
6440 	 */
6441 	ip_align = 2;
6442 
6443 	reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
6444 	if (!tnum_is_aligned(reg_off, size)) {
6445 		char tn_buf[48];
6446 
6447 		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6448 		verbose(env,
6449 			"misaligned packet access off %d+%s+%d+%d size %d\n",
6450 			ip_align, tn_buf, reg->off, off, size);
6451 		return -EACCES;
6452 	}
6453 
6454 	return 0;
6455 }
6456 
check_generic_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,const char * pointer_desc,int off,int size,bool strict)6457 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
6458 				       const struct bpf_reg_state *reg,
6459 				       const char *pointer_desc,
6460 				       int off, int size, bool strict)
6461 {
6462 	struct tnum reg_off;
6463 
6464 	/* Byte size accesses are always allowed. */
6465 	if (!strict || size == 1)
6466 		return 0;
6467 
6468 	reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
6469 	if (!tnum_is_aligned(reg_off, size)) {
6470 		char tn_buf[48];
6471 
6472 		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6473 		verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
6474 			pointer_desc, tn_buf, reg->off, off, size);
6475 		return -EACCES;
6476 	}
6477 
6478 	return 0;
6479 }
6480 
check_ptr_alignment(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int off,int size,bool strict_alignment_once)6481 static int check_ptr_alignment(struct bpf_verifier_env *env,
6482 			       const struct bpf_reg_state *reg, int off,
6483 			       int size, bool strict_alignment_once)
6484 {
6485 	bool strict = env->strict_alignment || strict_alignment_once;
6486 	const char *pointer_desc = "";
6487 
6488 	switch (reg->type) {
6489 	case PTR_TO_PACKET:
6490 	case PTR_TO_PACKET_META:
6491 		/* Special case, because of NET_IP_ALIGN. Given metadata sits
6492 		 * right in front, treat it the very same way.
6493 		 */
6494 		return check_pkt_ptr_alignment(env, reg, off, size, strict);
6495 	case PTR_TO_FLOW_KEYS:
6496 		pointer_desc = "flow keys ";
6497 		break;
6498 	case PTR_TO_MAP_KEY:
6499 		pointer_desc = "key ";
6500 		break;
6501 	case PTR_TO_MAP_VALUE:
6502 		pointer_desc = "value ";
6503 		break;
6504 	case PTR_TO_CTX:
6505 		pointer_desc = "context ";
6506 		break;
6507 	case PTR_TO_STACK:
6508 		pointer_desc = "stack ";
6509 		/* The stack spill tracking logic in check_stack_write_fixed_off()
6510 		 * and check_stack_read_fixed_off() relies on stack accesses being
6511 		 * aligned.
6512 		 */
6513 		strict = true;
6514 		break;
6515 	case PTR_TO_SOCKET:
6516 		pointer_desc = "sock ";
6517 		break;
6518 	case PTR_TO_SOCK_COMMON:
6519 		pointer_desc = "sock_common ";
6520 		break;
6521 	case PTR_TO_TCP_SOCK:
6522 		pointer_desc = "tcp_sock ";
6523 		break;
6524 	case PTR_TO_XDP_SOCK:
6525 		pointer_desc = "xdp_sock ";
6526 		break;
6527 	case PTR_TO_ARENA:
6528 		return 0;
6529 	default:
6530 		break;
6531 	}
6532 	return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
6533 					   strict);
6534 }
6535 
bpf_enable_priv_stack(struct bpf_prog * prog)6536 static enum priv_stack_mode bpf_enable_priv_stack(struct bpf_prog *prog)
6537 {
6538 	if (!bpf_jit_supports_private_stack())
6539 		return NO_PRIV_STACK;
6540 
6541 	/* bpf_prog_check_recur() checks all prog types that use bpf trampoline
6542 	 * while kprobe/tp/perf_event/raw_tp don't use trampoline hence checked
6543 	 * explicitly.
6544 	 */
6545 	switch (prog->type) {
6546 	case BPF_PROG_TYPE_KPROBE:
6547 	case BPF_PROG_TYPE_TRACEPOINT:
6548 	case BPF_PROG_TYPE_PERF_EVENT:
6549 	case BPF_PROG_TYPE_RAW_TRACEPOINT:
6550 		return PRIV_STACK_ADAPTIVE;
6551 	case BPF_PROG_TYPE_TRACING:
6552 	case BPF_PROG_TYPE_LSM:
6553 	case BPF_PROG_TYPE_STRUCT_OPS:
6554 		if (prog->aux->priv_stack_requested || bpf_prog_check_recur(prog))
6555 			return PRIV_STACK_ADAPTIVE;
6556 		fallthrough;
6557 	default:
6558 		break;
6559 	}
6560 
6561 	return NO_PRIV_STACK;
6562 }
6563 
round_up_stack_depth(struct bpf_verifier_env * env,int stack_depth)6564 static int round_up_stack_depth(struct bpf_verifier_env *env, int stack_depth)
6565 {
6566 	if (env->prog->jit_requested)
6567 		return round_up(stack_depth, 16);
6568 
6569 	/* round up to 32-bytes, since this is granularity
6570 	 * of interpreter stack size
6571 	 */
6572 	return round_up(max_t(u32, stack_depth, 1), 32);
6573 }
6574 
6575 /* starting from main bpf function walk all instructions of the function
6576  * and recursively walk all callees that given function can call.
6577  * Ignore jump and exit insns.
6578  * Since recursion is prevented by check_cfg() this algorithm
6579  * only needs a local stack of MAX_CALL_FRAMES to remember callsites
6580  */
check_max_stack_depth_subprog(struct bpf_verifier_env * env,int idx,bool priv_stack_supported)6581 static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx,
6582 					 bool priv_stack_supported)
6583 {
6584 	struct bpf_subprog_info *subprog = env->subprog_info;
6585 	struct bpf_insn *insn = env->prog->insnsi;
6586 	int depth = 0, frame = 0, i, subprog_end, subprog_depth;
6587 	bool tail_call_reachable = false;
6588 	int ret_insn[MAX_CALL_FRAMES];
6589 	int ret_prog[MAX_CALL_FRAMES];
6590 	int j;
6591 
6592 	i = subprog[idx].start;
6593 	if (!priv_stack_supported)
6594 		subprog[idx].priv_stack_mode = NO_PRIV_STACK;
6595 process_func:
6596 	/* protect against potential stack overflow that might happen when
6597 	 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
6598 	 * depth for such case down to 256 so that the worst case scenario
6599 	 * would result in 8k stack size (32 which is tailcall limit * 256 =
6600 	 * 8k).
6601 	 *
6602 	 * To get the idea what might happen, see an example:
6603 	 * func1 -> sub rsp, 128
6604 	 *  subfunc1 -> sub rsp, 256
6605 	 *  tailcall1 -> add rsp, 256
6606 	 *   func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
6607 	 *   subfunc2 -> sub rsp, 64
6608 	 *   subfunc22 -> sub rsp, 128
6609 	 *   tailcall2 -> add rsp, 128
6610 	 *    func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
6611 	 *
6612 	 * tailcall will unwind the current stack frame but it will not get rid
6613 	 * of caller's stack as shown on the example above.
6614 	 */
6615 	if (idx && subprog[idx].has_tail_call && depth >= 256) {
6616 		verbose(env,
6617 			"tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
6618 			depth);
6619 		return -EACCES;
6620 	}
6621 
6622 	subprog_depth = round_up_stack_depth(env, subprog[idx].stack_depth);
6623 	if (priv_stack_supported) {
6624 		/* Request private stack support only if the subprog stack
6625 		 * depth is no less than BPF_PRIV_STACK_MIN_SIZE. This is to
6626 		 * avoid jit penalty if the stack usage is small.
6627 		 */
6628 		if (subprog[idx].priv_stack_mode == PRIV_STACK_UNKNOWN &&
6629 		    subprog_depth >= BPF_PRIV_STACK_MIN_SIZE)
6630 			subprog[idx].priv_stack_mode = PRIV_STACK_ADAPTIVE;
6631 	}
6632 
6633 	if (subprog[idx].priv_stack_mode == PRIV_STACK_ADAPTIVE) {
6634 		if (subprog_depth > MAX_BPF_STACK) {
6635 			verbose(env, "stack size of subprog %d is %d. Too large\n",
6636 				idx, subprog_depth);
6637 			return -EACCES;
6638 		}
6639 	} else {
6640 		depth += subprog_depth;
6641 		if (depth > MAX_BPF_STACK) {
6642 			verbose(env, "combined stack size of %d calls is %d. Too large\n",
6643 				frame + 1, depth);
6644 			return -EACCES;
6645 		}
6646 	}
6647 continue_func:
6648 	subprog_end = subprog[idx + 1].start;
6649 	for (; i < subprog_end; i++) {
6650 		int next_insn, sidx;
6651 
6652 		if (bpf_pseudo_kfunc_call(insn + i) && !insn[i].off) {
6653 			bool err = false;
6654 
6655 			if (!is_bpf_throw_kfunc(insn + i))
6656 				continue;
6657 			if (subprog[idx].is_cb)
6658 				err = true;
6659 			for (int c = 0; c < frame && !err; c++) {
6660 				if (subprog[ret_prog[c]].is_cb) {
6661 					err = true;
6662 					break;
6663 				}
6664 			}
6665 			if (!err)
6666 				continue;
6667 			verbose(env,
6668 				"bpf_throw kfunc (insn %d) cannot be called from callback subprog %d\n",
6669 				i, idx);
6670 			return -EINVAL;
6671 		}
6672 
6673 		if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
6674 			continue;
6675 		/* remember insn and function to return to */
6676 		ret_insn[frame] = i + 1;
6677 		ret_prog[frame] = idx;
6678 
6679 		/* find the callee */
6680 		next_insn = i + insn[i].imm + 1;
6681 		sidx = find_subprog(env, next_insn);
6682 		if (verifier_bug_if(sidx < 0, env, "callee not found at insn %d", next_insn))
6683 			return -EFAULT;
6684 		if (subprog[sidx].is_async_cb) {
6685 			if (subprog[sidx].has_tail_call) {
6686 				verifier_bug(env, "subprog has tail_call and async cb");
6687 				return -EFAULT;
6688 			}
6689 			/* async callbacks don't increase bpf prog stack size unless called directly */
6690 			if (!bpf_pseudo_call(insn + i))
6691 				continue;
6692 			if (subprog[sidx].is_exception_cb) {
6693 				verbose(env, "insn %d cannot call exception cb directly", i);
6694 				return -EINVAL;
6695 			}
6696 		}
6697 		i = next_insn;
6698 		idx = sidx;
6699 		if (!priv_stack_supported)
6700 			subprog[idx].priv_stack_mode = NO_PRIV_STACK;
6701 
6702 		if (subprog[idx].has_tail_call)
6703 			tail_call_reachable = true;
6704 
6705 		frame++;
6706 		if (frame >= MAX_CALL_FRAMES) {
6707 			verbose(env, "the call stack of %d frames is too deep !\n",
6708 				frame);
6709 			return -E2BIG;
6710 		}
6711 		goto process_func;
6712 	}
6713 	/* if tail call got detected across bpf2bpf calls then mark each of the
6714 	 * currently present subprog frames as tail call reachable subprogs;
6715 	 * this info will be utilized by JIT so that we will be preserving the
6716 	 * tail call counter throughout bpf2bpf calls combined with tailcalls
6717 	 */
6718 	if (tail_call_reachable)
6719 		for (j = 0; j < frame; j++) {
6720 			if (subprog[ret_prog[j]].is_exception_cb) {
6721 				verbose(env, "cannot tail call within exception cb\n");
6722 				return -EINVAL;
6723 			}
6724 			subprog[ret_prog[j]].tail_call_reachable = true;
6725 		}
6726 	if (subprog[0].tail_call_reachable)
6727 		env->prog->aux->tail_call_reachable = true;
6728 
6729 	/* end of for() loop means the last insn of the 'subprog'
6730 	 * was reached. Doesn't matter whether it was JA or EXIT
6731 	 */
6732 	if (frame == 0)
6733 		return 0;
6734 	if (subprog[idx].priv_stack_mode != PRIV_STACK_ADAPTIVE)
6735 		depth -= round_up_stack_depth(env, subprog[idx].stack_depth);
6736 	frame--;
6737 	i = ret_insn[frame];
6738 	idx = ret_prog[frame];
6739 	goto continue_func;
6740 }
6741 
check_max_stack_depth(struct bpf_verifier_env * env)6742 static int check_max_stack_depth(struct bpf_verifier_env *env)
6743 {
6744 	enum priv_stack_mode priv_stack_mode = PRIV_STACK_UNKNOWN;
6745 	struct bpf_subprog_info *si = env->subprog_info;
6746 	bool priv_stack_supported;
6747 	int ret;
6748 
6749 	for (int i = 0; i < env->subprog_cnt; i++) {
6750 		if (si[i].has_tail_call) {
6751 			priv_stack_mode = NO_PRIV_STACK;
6752 			break;
6753 		}
6754 	}
6755 
6756 	if (priv_stack_mode == PRIV_STACK_UNKNOWN)
6757 		priv_stack_mode = bpf_enable_priv_stack(env->prog);
6758 
6759 	/* All async_cb subprogs use normal kernel stack. If a particular
6760 	 * subprog appears in both main prog and async_cb subtree, that
6761 	 * subprog will use normal kernel stack to avoid potential nesting.
6762 	 * The reverse subprog traversal ensures when main prog subtree is
6763 	 * checked, the subprogs appearing in async_cb subtrees are already
6764 	 * marked as using normal kernel stack, so stack size checking can
6765 	 * be done properly.
6766 	 */
6767 	for (int i = env->subprog_cnt - 1; i >= 0; i--) {
6768 		if (!i || si[i].is_async_cb) {
6769 			priv_stack_supported = !i && priv_stack_mode == PRIV_STACK_ADAPTIVE;
6770 			ret = check_max_stack_depth_subprog(env, i, priv_stack_supported);
6771 			if (ret < 0)
6772 				return ret;
6773 		}
6774 	}
6775 
6776 	for (int i = 0; i < env->subprog_cnt; i++) {
6777 		if (si[i].priv_stack_mode == PRIV_STACK_ADAPTIVE) {
6778 			env->prog->aux->jits_use_priv_stack = true;
6779 			break;
6780 		}
6781 	}
6782 
6783 	return 0;
6784 }
6785 
6786 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
get_callee_stack_depth(struct bpf_verifier_env * env,const struct bpf_insn * insn,int idx)6787 static int get_callee_stack_depth(struct bpf_verifier_env *env,
6788 				  const struct bpf_insn *insn, int idx)
6789 {
6790 	int start = idx + insn->imm + 1, subprog;
6791 
6792 	subprog = find_subprog(env, start);
6793 	if (verifier_bug_if(subprog < 0, env, "get stack depth: no program at insn %d", start))
6794 		return -EFAULT;
6795 	return env->subprog_info[subprog].stack_depth;
6796 }
6797 #endif
6798 
__check_buffer_access(struct bpf_verifier_env * env,const char * buf_info,const struct bpf_reg_state * reg,int regno,int off,int size)6799 static int __check_buffer_access(struct bpf_verifier_env *env,
6800 				 const char *buf_info,
6801 				 const struct bpf_reg_state *reg,
6802 				 int regno, int off, int size)
6803 {
6804 	if (off < 0) {
6805 		verbose(env,
6806 			"R%d invalid %s buffer access: off=%d, size=%d\n",
6807 			regno, buf_info, off, size);
6808 		return -EACCES;
6809 	}
6810 	if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
6811 		char tn_buf[48];
6812 
6813 		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6814 		verbose(env,
6815 			"R%d invalid variable buffer offset: off=%d, var_off=%s\n",
6816 			regno, off, tn_buf);
6817 		return -EACCES;
6818 	}
6819 
6820 	return 0;
6821 }
6822 
check_tp_buffer_access(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,int off,int size)6823 static int check_tp_buffer_access(struct bpf_verifier_env *env,
6824 				  const struct bpf_reg_state *reg,
6825 				  int regno, int off, int size)
6826 {
6827 	int err;
6828 
6829 	err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
6830 	if (err)
6831 		return err;
6832 
6833 	if (off + size > env->prog->aux->max_tp_access)
6834 		env->prog->aux->max_tp_access = off + size;
6835 
6836 	return 0;
6837 }
6838 
check_buffer_access(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,int off,int size,bool zero_size_allowed,u32 * max_access)6839 static int check_buffer_access(struct bpf_verifier_env *env,
6840 			       const struct bpf_reg_state *reg,
6841 			       int regno, int off, int size,
6842 			       bool zero_size_allowed,
6843 			       u32 *max_access)
6844 {
6845 	const char *buf_info = type_is_rdonly_mem(reg->type) ? "rdonly" : "rdwr";
6846 	int err;
6847 
6848 	err = __check_buffer_access(env, buf_info, reg, regno, off, size);
6849 	if (err)
6850 		return err;
6851 
6852 	if (off + size > *max_access)
6853 		*max_access = off + size;
6854 
6855 	return 0;
6856 }
6857 
6858 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
zext_32_to_64(struct bpf_reg_state * reg)6859 static void zext_32_to_64(struct bpf_reg_state *reg)
6860 {
6861 	reg->var_off = tnum_subreg(reg->var_off);
6862 	__reg_assign_32_into_64(reg);
6863 }
6864 
6865 /* truncate register to smaller size (in bytes)
6866  * must be called with size < BPF_REG_SIZE
6867  */
coerce_reg_to_size(struct bpf_reg_state * reg,int size)6868 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
6869 {
6870 	u64 mask;
6871 
6872 	/* clear high bits in bit representation */
6873 	reg->var_off = tnum_cast(reg->var_off, size);
6874 
6875 	/* fix arithmetic bounds */
6876 	mask = ((u64)1 << (size * 8)) - 1;
6877 	if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
6878 		reg->umin_value &= mask;
6879 		reg->umax_value &= mask;
6880 	} else {
6881 		reg->umin_value = 0;
6882 		reg->umax_value = mask;
6883 	}
6884 	reg->smin_value = reg->umin_value;
6885 	reg->smax_value = reg->umax_value;
6886 
6887 	/* If size is smaller than 32bit register the 32bit register
6888 	 * values are also truncated so we push 64-bit bounds into
6889 	 * 32-bit bounds. Above were truncated < 32-bits already.
6890 	 */
6891 	if (size < 4)
6892 		__mark_reg32_unbounded(reg);
6893 
6894 	reg_bounds_sync(reg);
6895 }
6896 
set_sext64_default_val(struct bpf_reg_state * reg,int size)6897 static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
6898 {
6899 	if (size == 1) {
6900 		reg->smin_value = reg->s32_min_value = S8_MIN;
6901 		reg->smax_value = reg->s32_max_value = S8_MAX;
6902 	} else if (size == 2) {
6903 		reg->smin_value = reg->s32_min_value = S16_MIN;
6904 		reg->smax_value = reg->s32_max_value = S16_MAX;
6905 	} else {
6906 		/* size == 4 */
6907 		reg->smin_value = reg->s32_min_value = S32_MIN;
6908 		reg->smax_value = reg->s32_max_value = S32_MAX;
6909 	}
6910 	reg->umin_value = reg->u32_min_value = 0;
6911 	reg->umax_value = U64_MAX;
6912 	reg->u32_max_value = U32_MAX;
6913 	reg->var_off = tnum_unknown;
6914 }
6915 
coerce_reg_to_size_sx(struct bpf_reg_state * reg,int size)6916 static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
6917 {
6918 	s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
6919 	u64 top_smax_value, top_smin_value;
6920 	u64 num_bits = size * 8;
6921 
6922 	if (tnum_is_const(reg->var_off)) {
6923 		u64_cval = reg->var_off.value;
6924 		if (size == 1)
6925 			reg->var_off = tnum_const((s8)u64_cval);
6926 		else if (size == 2)
6927 			reg->var_off = tnum_const((s16)u64_cval);
6928 		else
6929 			/* size == 4 */
6930 			reg->var_off = tnum_const((s32)u64_cval);
6931 
6932 		u64_cval = reg->var_off.value;
6933 		reg->smax_value = reg->smin_value = u64_cval;
6934 		reg->umax_value = reg->umin_value = u64_cval;
6935 		reg->s32_max_value = reg->s32_min_value = u64_cval;
6936 		reg->u32_max_value = reg->u32_min_value = u64_cval;
6937 		return;
6938 	}
6939 
6940 	top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
6941 	top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
6942 
6943 	if (top_smax_value != top_smin_value)
6944 		goto out;
6945 
6946 	/* find the s64_min and s64_min after sign extension */
6947 	if (size == 1) {
6948 		init_s64_max = (s8)reg->smax_value;
6949 		init_s64_min = (s8)reg->smin_value;
6950 	} else if (size == 2) {
6951 		init_s64_max = (s16)reg->smax_value;
6952 		init_s64_min = (s16)reg->smin_value;
6953 	} else {
6954 		init_s64_max = (s32)reg->smax_value;
6955 		init_s64_min = (s32)reg->smin_value;
6956 	}
6957 
6958 	s64_max = max(init_s64_max, init_s64_min);
6959 	s64_min = min(init_s64_max, init_s64_min);
6960 
6961 	/* both of s64_max/s64_min positive or negative */
6962 	if ((s64_max >= 0) == (s64_min >= 0)) {
6963 		reg->s32_min_value = reg->smin_value = s64_min;
6964 		reg->s32_max_value = reg->smax_value = s64_max;
6965 		reg->u32_min_value = reg->umin_value = s64_min;
6966 		reg->u32_max_value = reg->umax_value = s64_max;
6967 		reg->var_off = tnum_range(s64_min, s64_max);
6968 		return;
6969 	}
6970 
6971 out:
6972 	set_sext64_default_val(reg, size);
6973 }
6974 
set_sext32_default_val(struct bpf_reg_state * reg,int size)6975 static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
6976 {
6977 	if (size == 1) {
6978 		reg->s32_min_value = S8_MIN;
6979 		reg->s32_max_value = S8_MAX;
6980 	} else {
6981 		/* size == 2 */
6982 		reg->s32_min_value = S16_MIN;
6983 		reg->s32_max_value = S16_MAX;
6984 	}
6985 	reg->u32_min_value = 0;
6986 	reg->u32_max_value = U32_MAX;
6987 	reg->var_off = tnum_subreg(tnum_unknown);
6988 }
6989 
coerce_subreg_to_size_sx(struct bpf_reg_state * reg,int size)6990 static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
6991 {
6992 	s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
6993 	u32 top_smax_value, top_smin_value;
6994 	u32 num_bits = size * 8;
6995 
6996 	if (tnum_is_const(reg->var_off)) {
6997 		u32_val = reg->var_off.value;
6998 		if (size == 1)
6999 			reg->var_off = tnum_const((s8)u32_val);
7000 		else
7001 			reg->var_off = tnum_const((s16)u32_val);
7002 
7003 		u32_val = reg->var_off.value;
7004 		reg->s32_min_value = reg->s32_max_value = u32_val;
7005 		reg->u32_min_value = reg->u32_max_value = u32_val;
7006 		return;
7007 	}
7008 
7009 	top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
7010 	top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
7011 
7012 	if (top_smax_value != top_smin_value)
7013 		goto out;
7014 
7015 	/* find the s32_min and s32_min after sign extension */
7016 	if (size == 1) {
7017 		init_s32_max = (s8)reg->s32_max_value;
7018 		init_s32_min = (s8)reg->s32_min_value;
7019 	} else {
7020 		/* size == 2 */
7021 		init_s32_max = (s16)reg->s32_max_value;
7022 		init_s32_min = (s16)reg->s32_min_value;
7023 	}
7024 	s32_max = max(init_s32_max, init_s32_min);
7025 	s32_min = min(init_s32_max, init_s32_min);
7026 
7027 	if ((s32_min >= 0) == (s32_max >= 0)) {
7028 		reg->s32_min_value = s32_min;
7029 		reg->s32_max_value = s32_max;
7030 		reg->u32_min_value = (u32)s32_min;
7031 		reg->u32_max_value = (u32)s32_max;
7032 		reg->var_off = tnum_subreg(tnum_range(s32_min, s32_max));
7033 		return;
7034 	}
7035 
7036 out:
7037 	set_sext32_default_val(reg, size);
7038 }
7039 
bpf_map_is_rdonly(const struct bpf_map * map)7040 static bool bpf_map_is_rdonly(const struct bpf_map *map)
7041 {
7042 	/* A map is considered read-only if the following condition are true:
7043 	 *
7044 	 * 1) BPF program side cannot change any of the map content. The
7045 	 *    BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
7046 	 *    and was set at map creation time.
7047 	 * 2) The map value(s) have been initialized from user space by a
7048 	 *    loader and then "frozen", such that no new map update/delete
7049 	 *    operations from syscall side are possible for the rest of
7050 	 *    the map's lifetime from that point onwards.
7051 	 * 3) Any parallel/pending map update/delete operations from syscall
7052 	 *    side have been completed. Only after that point, it's safe to
7053 	 *    assume that map value(s) are immutable.
7054 	 */
7055 	return (map->map_flags & BPF_F_RDONLY_PROG) &&
7056 	       READ_ONCE(map->frozen) &&
7057 	       !bpf_map_write_active(map);
7058 }
7059 
bpf_map_direct_read(struct bpf_map * map,int off,int size,u64 * val,bool is_ldsx)7060 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
7061 			       bool is_ldsx)
7062 {
7063 	void *ptr;
7064 	u64 addr;
7065 	int err;
7066 
7067 	err = map->ops->map_direct_value_addr(map, &addr, off);
7068 	if (err)
7069 		return err;
7070 	ptr = (void *)(long)addr + off;
7071 
7072 	switch (size) {
7073 	case sizeof(u8):
7074 		*val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
7075 		break;
7076 	case sizeof(u16):
7077 		*val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
7078 		break;
7079 	case sizeof(u32):
7080 		*val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
7081 		break;
7082 	case sizeof(u64):
7083 		*val = *(u64 *)ptr;
7084 		break;
7085 	default:
7086 		return -EINVAL;
7087 	}
7088 	return 0;
7089 }
7090 
7091 #define BTF_TYPE_SAFE_RCU(__type)  __PASTE(__type, __safe_rcu)
7092 #define BTF_TYPE_SAFE_RCU_OR_NULL(__type)  __PASTE(__type, __safe_rcu_or_null)
7093 #define BTF_TYPE_SAFE_TRUSTED(__type)  __PASTE(__type, __safe_trusted)
7094 #define BTF_TYPE_SAFE_TRUSTED_OR_NULL(__type)  __PASTE(__type, __safe_trusted_or_null)
7095 
7096 /*
7097  * Allow list few fields as RCU trusted or full trusted.
7098  * This logic doesn't allow mix tagging and will be removed once GCC supports
7099  * btf_type_tag.
7100  */
7101 
7102 /* RCU trusted: these fields are trusted in RCU CS and never NULL */
BTF_TYPE_SAFE_RCU(struct task_struct)7103 BTF_TYPE_SAFE_RCU(struct task_struct) {
7104 	const cpumask_t *cpus_ptr;
7105 	struct css_set __rcu *cgroups;
7106 	struct task_struct __rcu *real_parent;
7107 	struct task_struct *group_leader;
7108 };
7109 
BTF_TYPE_SAFE_RCU(struct cgroup)7110 BTF_TYPE_SAFE_RCU(struct cgroup) {
7111 	/* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
7112 	struct kernfs_node *kn;
7113 };
7114 
BTF_TYPE_SAFE_RCU(struct css_set)7115 BTF_TYPE_SAFE_RCU(struct css_set) {
7116 	struct cgroup *dfl_cgrp;
7117 };
7118 
BTF_TYPE_SAFE_RCU(struct cgroup_subsys_state)7119 BTF_TYPE_SAFE_RCU(struct cgroup_subsys_state) {
7120 	struct cgroup *cgroup;
7121 };
7122 
7123 /* RCU trusted: these fields are trusted in RCU CS and can be NULL */
BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct)7124 BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
7125 	struct file __rcu *exe_file;
7126 };
7127 
7128 /* skb->sk, req->sk are not RCU protected, but we mark them as such
7129  * because bpf prog accessible sockets are SOCK_RCU_FREE.
7130  */
BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff)7131 BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
7132 	struct sock *sk;
7133 };
7134 
BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock)7135 BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
7136 	struct sock *sk;
7137 };
7138 
7139 /* full trusted: these fields are trusted even outside of RCU CS and never NULL */
BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta)7140 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
7141 	struct seq_file *seq;
7142 };
7143 
BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task)7144 BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
7145 	struct bpf_iter_meta *meta;
7146 	struct task_struct *task;
7147 };
7148 
BTF_TYPE_SAFE_TRUSTED(struct linux_binprm)7149 BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
7150 	struct file *file;
7151 };
7152 
BTF_TYPE_SAFE_TRUSTED(struct file)7153 BTF_TYPE_SAFE_TRUSTED(struct file) {
7154 	struct inode *f_inode;
7155 };
7156 
BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct dentry)7157 BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct dentry) {
7158 	struct inode *d_inode;
7159 };
7160 
BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket)7161 BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket) {
7162 	struct sock *sk;
7163 };
7164 
type_is_rcu(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)7165 static bool type_is_rcu(struct bpf_verifier_env *env,
7166 			struct bpf_reg_state *reg,
7167 			const char *field_name, u32 btf_id)
7168 {
7169 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
7170 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
7171 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
7172 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup_subsys_state));
7173 
7174 	return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu");
7175 }
7176 
type_is_rcu_or_null(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)7177 static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
7178 				struct bpf_reg_state *reg,
7179 				const char *field_name, u32 btf_id)
7180 {
7181 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
7182 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
7183 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
7184 
7185 	return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_rcu_or_null");
7186 }
7187 
type_is_trusted(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)7188 static bool type_is_trusted(struct bpf_verifier_env *env,
7189 			    struct bpf_reg_state *reg,
7190 			    const char *field_name, u32 btf_id)
7191 {
7192 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
7193 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
7194 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
7195 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
7196 
7197 	return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id, "__safe_trusted");
7198 }
7199 
type_is_trusted_or_null(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const char * field_name,u32 btf_id)7200 static bool type_is_trusted_or_null(struct bpf_verifier_env *env,
7201 				    struct bpf_reg_state *reg,
7202 				    const char *field_name, u32 btf_id)
7203 {
7204 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket));
7205 	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct dentry));
7206 
7207 	return btf_nested_type_is_trusted(&env->log, reg, field_name, btf_id,
7208 					  "__safe_trusted_or_null");
7209 }
7210 
check_ptr_to_btf_access(struct bpf_verifier_env * env,struct bpf_reg_state * regs,int regno,int off,int size,enum bpf_access_type atype,int value_regno)7211 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
7212 				   struct bpf_reg_state *regs,
7213 				   int regno, int off, int size,
7214 				   enum bpf_access_type atype,
7215 				   int value_regno)
7216 {
7217 	struct bpf_reg_state *reg = regs + regno;
7218 	const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
7219 	const char *tname = btf_name_by_offset(reg->btf, t->name_off);
7220 	const char *field_name = NULL;
7221 	enum bpf_type_flag flag = 0;
7222 	u32 btf_id = 0;
7223 	int ret;
7224 
7225 	if (!env->allow_ptr_leaks) {
7226 		verbose(env,
7227 			"'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
7228 			tname);
7229 		return -EPERM;
7230 	}
7231 	if (!env->prog->gpl_compatible && btf_is_kernel(reg->btf)) {
7232 		verbose(env,
7233 			"Cannot access kernel 'struct %s' from non-GPL compatible program\n",
7234 			tname);
7235 		return -EINVAL;
7236 	}
7237 	if (off < 0) {
7238 		verbose(env,
7239 			"R%d is ptr_%s invalid negative access: off=%d\n",
7240 			regno, tname, off);
7241 		return -EACCES;
7242 	}
7243 	if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
7244 		char tn_buf[48];
7245 
7246 		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7247 		verbose(env,
7248 			"R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
7249 			regno, tname, off, tn_buf);
7250 		return -EACCES;
7251 	}
7252 
7253 	if (reg->type & MEM_USER) {
7254 		verbose(env,
7255 			"R%d is ptr_%s access user memory: off=%d\n",
7256 			regno, tname, off);
7257 		return -EACCES;
7258 	}
7259 
7260 	if (reg->type & MEM_PERCPU) {
7261 		verbose(env,
7262 			"R%d is ptr_%s access percpu memory: off=%d\n",
7263 			regno, tname, off);
7264 		return -EACCES;
7265 	}
7266 
7267 	if (env->ops->btf_struct_access && !type_is_alloc(reg->type) && atype == BPF_WRITE) {
7268 		if (!btf_is_kernel(reg->btf)) {
7269 			verifier_bug(env, "reg->btf must be kernel btf");
7270 			return -EFAULT;
7271 		}
7272 		ret = env->ops->btf_struct_access(&env->log, reg, off, size);
7273 	} else {
7274 		/* Writes are permitted with default btf_struct_access for
7275 		 * program allocated objects (which always have ref_obj_id > 0),
7276 		 * but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
7277 		 */
7278 		if (atype != BPF_READ && !type_is_ptr_alloc_obj(reg->type)) {
7279 			verbose(env, "only read is supported\n");
7280 			return -EACCES;
7281 		}
7282 
7283 		if (type_is_alloc(reg->type) && !type_is_non_owning_ref(reg->type) &&
7284 		    !(reg->type & MEM_RCU) && !reg->ref_obj_id) {
7285 			verifier_bug(env, "ref_obj_id for allocated object must be non-zero");
7286 			return -EFAULT;
7287 		}
7288 
7289 		ret = btf_struct_access(&env->log, reg, off, size, atype, &btf_id, &flag, &field_name);
7290 	}
7291 
7292 	if (ret < 0)
7293 		return ret;
7294 
7295 	if (ret != PTR_TO_BTF_ID) {
7296 		/* just mark; */
7297 
7298 	} else if (type_flag(reg->type) & PTR_UNTRUSTED) {
7299 		/* If this is an untrusted pointer, all pointers formed by walking it
7300 		 * also inherit the untrusted flag.
7301 		 */
7302 		flag = PTR_UNTRUSTED;
7303 
7304 	} else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
7305 		/* By default any pointer obtained from walking a trusted pointer is no
7306 		 * longer trusted, unless the field being accessed has explicitly been
7307 		 * marked as inheriting its parent's state of trust (either full or RCU).
7308 		 * For example:
7309 		 * 'cgroups' pointer is untrusted if task->cgroups dereference
7310 		 * happened in a sleepable program outside of bpf_rcu_read_lock()
7311 		 * section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
7312 		 * Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
7313 		 *
7314 		 * A regular RCU-protected pointer with __rcu tag can also be deemed
7315 		 * trusted if we are in an RCU CS. Such pointer can be NULL.
7316 		 */
7317 		if (type_is_trusted(env, reg, field_name, btf_id)) {
7318 			flag |= PTR_TRUSTED;
7319 		} else if (type_is_trusted_or_null(env, reg, field_name, btf_id)) {
7320 			flag |= PTR_TRUSTED | PTR_MAYBE_NULL;
7321 		} else if (in_rcu_cs(env) && !type_may_be_null(reg->type)) {
7322 			if (type_is_rcu(env, reg, field_name, btf_id)) {
7323 				/* ignore __rcu tag and mark it MEM_RCU */
7324 				flag |= MEM_RCU;
7325 			} else if (flag & MEM_RCU ||
7326 				   type_is_rcu_or_null(env, reg, field_name, btf_id)) {
7327 				/* __rcu tagged pointers can be NULL */
7328 				flag |= MEM_RCU | PTR_MAYBE_NULL;
7329 
7330 				/* We always trust them */
7331 				if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
7332 				    flag & PTR_UNTRUSTED)
7333 					flag &= ~PTR_UNTRUSTED;
7334 			} else if (flag & (MEM_PERCPU | MEM_USER)) {
7335 				/* keep as-is */
7336 			} else {
7337 				/* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
7338 				clear_trusted_flags(&flag);
7339 			}
7340 		} else {
7341 			/*
7342 			 * If not in RCU CS or MEM_RCU pointer can be NULL then
7343 			 * aggressively mark as untrusted otherwise such
7344 			 * pointers will be plain PTR_TO_BTF_ID without flags
7345 			 * and will be allowed to be passed into helpers for
7346 			 * compat reasons.
7347 			 */
7348 			flag = PTR_UNTRUSTED;
7349 		}
7350 	} else {
7351 		/* Old compat. Deprecated */
7352 		clear_trusted_flags(&flag);
7353 	}
7354 
7355 	if (atype == BPF_READ && value_regno >= 0) {
7356 		ret = mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id, flag);
7357 		if (ret < 0)
7358 			return ret;
7359 	}
7360 
7361 	return 0;
7362 }
7363 
check_ptr_to_map_access(struct bpf_verifier_env * env,struct bpf_reg_state * regs,int regno,int off,int size,enum bpf_access_type atype,int value_regno)7364 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
7365 				   struct bpf_reg_state *regs,
7366 				   int regno, int off, int size,
7367 				   enum bpf_access_type atype,
7368 				   int value_regno)
7369 {
7370 	struct bpf_reg_state *reg = regs + regno;
7371 	struct bpf_map *map = reg->map_ptr;
7372 	struct bpf_reg_state map_reg;
7373 	enum bpf_type_flag flag = 0;
7374 	const struct btf_type *t;
7375 	const char *tname;
7376 	u32 btf_id;
7377 	int ret;
7378 
7379 	if (!btf_vmlinux) {
7380 		verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
7381 		return -ENOTSUPP;
7382 	}
7383 
7384 	if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
7385 		verbose(env, "map_ptr access not supported for map type %d\n",
7386 			map->map_type);
7387 		return -ENOTSUPP;
7388 	}
7389 
7390 	t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
7391 	tname = btf_name_by_offset(btf_vmlinux, t->name_off);
7392 
7393 	if (!env->allow_ptr_leaks) {
7394 		verbose(env,
7395 			"'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
7396 			tname);
7397 		return -EPERM;
7398 	}
7399 
7400 	if (off < 0) {
7401 		verbose(env, "R%d is %s invalid negative access: off=%d\n",
7402 			regno, tname, off);
7403 		return -EACCES;
7404 	}
7405 
7406 	if (atype != BPF_READ) {
7407 		verbose(env, "only read from %s is supported\n", tname);
7408 		return -EACCES;
7409 	}
7410 
7411 	/* Simulate access to a PTR_TO_BTF_ID */
7412 	memset(&map_reg, 0, sizeof(map_reg));
7413 	ret = mark_btf_ld_reg(env, &map_reg, 0, PTR_TO_BTF_ID,
7414 			      btf_vmlinux, *map->ops->map_btf_id, 0);
7415 	if (ret < 0)
7416 		return ret;
7417 	ret = btf_struct_access(&env->log, &map_reg, off, size, atype, &btf_id, &flag, NULL);
7418 	if (ret < 0)
7419 		return ret;
7420 
7421 	if (value_regno >= 0) {
7422 		ret = mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id, flag);
7423 		if (ret < 0)
7424 			return ret;
7425 	}
7426 
7427 	return 0;
7428 }
7429 
7430 /* Check that the stack access at the given offset is within bounds. The
7431  * maximum valid offset is -1.
7432  *
7433  * The minimum valid offset is -MAX_BPF_STACK for writes, and
7434  * -state->allocated_stack for reads.
7435  */
check_stack_slot_within_bounds(struct bpf_verifier_env * env,s64 off,struct bpf_func_state * state,enum bpf_access_type t)7436 static int check_stack_slot_within_bounds(struct bpf_verifier_env *env,
7437                                           s64 off,
7438                                           struct bpf_func_state *state,
7439                                           enum bpf_access_type t)
7440 {
7441 	int min_valid_off;
7442 
7443 	if (t == BPF_WRITE || env->allow_uninit_stack)
7444 		min_valid_off = -MAX_BPF_STACK;
7445 	else
7446 		min_valid_off = -state->allocated_stack;
7447 
7448 	if (off < min_valid_off || off > -1)
7449 		return -EACCES;
7450 	return 0;
7451 }
7452 
7453 /* Check that the stack access at 'regno + off' falls within the maximum stack
7454  * bounds.
7455  *
7456  * 'off' includes `regno->offset`, but not its dynamic part (if any).
7457  */
check_stack_access_within_bounds(struct bpf_verifier_env * env,int regno,int off,int access_size,enum bpf_access_type type)7458 static int check_stack_access_within_bounds(
7459 		struct bpf_verifier_env *env,
7460 		int regno, int off, int access_size,
7461 		enum bpf_access_type type)
7462 {
7463 	struct bpf_reg_state *regs = cur_regs(env);
7464 	struct bpf_reg_state *reg = regs + regno;
7465 	struct bpf_func_state *state = func(env, reg);
7466 	s64 min_off, max_off;
7467 	int err;
7468 	char *err_extra;
7469 
7470 	if (type == BPF_READ)
7471 		err_extra = " read from";
7472 	else
7473 		err_extra = " write to";
7474 
7475 	if (tnum_is_const(reg->var_off)) {
7476 		min_off = (s64)reg->var_off.value + off;
7477 		max_off = min_off + access_size;
7478 	} else {
7479 		if (reg->smax_value >= BPF_MAX_VAR_OFF ||
7480 		    reg->smin_value <= -BPF_MAX_VAR_OFF) {
7481 			verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
7482 				err_extra, regno);
7483 			return -EACCES;
7484 		}
7485 		min_off = reg->smin_value + off;
7486 		max_off = reg->smax_value + off + access_size;
7487 	}
7488 
7489 	err = check_stack_slot_within_bounds(env, min_off, state, type);
7490 	if (!err && max_off > 0)
7491 		err = -EINVAL; /* out of stack access into non-negative offsets */
7492 	if (!err && access_size < 0)
7493 		/* access_size should not be negative (or overflow an int); others checks
7494 		 * along the way should have prevented such an access.
7495 		 */
7496 		err = -EFAULT; /* invalid negative access size; integer overflow? */
7497 
7498 	if (err) {
7499 		if (tnum_is_const(reg->var_off)) {
7500 			verbose(env, "invalid%s stack R%d off=%d size=%d\n",
7501 				err_extra, regno, off, access_size);
7502 		} else {
7503 			char tn_buf[48];
7504 
7505 			tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7506 			verbose(env, "invalid variable-offset%s stack R%d var_off=%s off=%d size=%d\n",
7507 				err_extra, regno, tn_buf, off, access_size);
7508 		}
7509 		return err;
7510 	}
7511 
7512 	/* Note that there is no stack access with offset zero, so the needed stack
7513 	 * size is -min_off, not -min_off+1.
7514 	 */
7515 	return grow_stack_state(env, state, -min_off /* size */);
7516 }
7517 
get_func_retval_range(struct bpf_prog * prog,struct bpf_retval_range * range)7518 static bool get_func_retval_range(struct bpf_prog *prog,
7519 				  struct bpf_retval_range *range)
7520 {
7521 	if (prog->type == BPF_PROG_TYPE_LSM &&
7522 		prog->expected_attach_type == BPF_LSM_MAC &&
7523 		!bpf_lsm_get_retval_range(prog, range)) {
7524 		return true;
7525 	}
7526 	return false;
7527 }
7528 
7529 /* check whether memory at (regno + off) is accessible for t = (read | write)
7530  * if t==write, value_regno is a register which value is stored into memory
7531  * if t==read, value_regno is a register which will receive the value from memory
7532  * if t==write && value_regno==-1, some unknown value is stored into memory
7533  * if t==read && value_regno==-1, don't care what we read from memory
7534  */
check_mem_access(struct bpf_verifier_env * env,int insn_idx,u32 regno,int off,int bpf_size,enum bpf_access_type t,int value_regno,bool strict_alignment_once,bool is_ldsx)7535 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
7536 			    int off, int bpf_size, enum bpf_access_type t,
7537 			    int value_regno, bool strict_alignment_once, bool is_ldsx)
7538 {
7539 	struct bpf_reg_state *regs = cur_regs(env);
7540 	struct bpf_reg_state *reg = regs + regno;
7541 	int size, err = 0;
7542 
7543 	size = bpf_size_to_bytes(bpf_size);
7544 	if (size < 0)
7545 		return size;
7546 
7547 	/* alignment checks will add in reg->off themselves */
7548 	err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
7549 	if (err)
7550 		return err;
7551 
7552 	/* for access checks, reg->off is just part of off */
7553 	off += reg->off;
7554 
7555 	if (reg->type == PTR_TO_MAP_KEY) {
7556 		if (t == BPF_WRITE) {
7557 			verbose(env, "write to change key R%d not allowed\n", regno);
7558 			return -EACCES;
7559 		}
7560 
7561 		err = check_mem_region_access(env, regno, off, size,
7562 					      reg->map_ptr->key_size, false);
7563 		if (err)
7564 			return err;
7565 		if (value_regno >= 0)
7566 			mark_reg_unknown(env, regs, value_regno);
7567 	} else if (reg->type == PTR_TO_MAP_VALUE) {
7568 		struct btf_field *kptr_field = NULL;
7569 
7570 		if (t == BPF_WRITE && value_regno >= 0 &&
7571 		    is_pointer_value(env, value_regno)) {
7572 			verbose(env, "R%d leaks addr into map\n", value_regno);
7573 			return -EACCES;
7574 		}
7575 		err = check_map_access_type(env, regno, off, size, t);
7576 		if (err)
7577 			return err;
7578 		err = check_map_access(env, regno, off, size, false, ACCESS_DIRECT);
7579 		if (err)
7580 			return err;
7581 		if (tnum_is_const(reg->var_off))
7582 			kptr_field = btf_record_find(reg->map_ptr->record,
7583 						     off + reg->var_off.value, BPF_KPTR | BPF_UPTR);
7584 		if (kptr_field) {
7585 			err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
7586 		} else if (t == BPF_READ && value_regno >= 0) {
7587 			struct bpf_map *map = reg->map_ptr;
7588 
7589 			/* if map is read-only, track its contents as scalars */
7590 			if (tnum_is_const(reg->var_off) &&
7591 			    bpf_map_is_rdonly(map) &&
7592 			    map->ops->map_direct_value_addr) {
7593 				int map_off = off + reg->var_off.value;
7594 				u64 val = 0;
7595 
7596 				err = bpf_map_direct_read(map, map_off, size,
7597 							  &val, is_ldsx);
7598 				if (err)
7599 					return err;
7600 
7601 				regs[value_regno].type = SCALAR_VALUE;
7602 				__mark_reg_known(&regs[value_regno], val);
7603 			} else {
7604 				mark_reg_unknown(env, regs, value_regno);
7605 			}
7606 		}
7607 	} else if (base_type(reg->type) == PTR_TO_MEM) {
7608 		bool rdonly_mem = type_is_rdonly_mem(reg->type);
7609 		bool rdonly_untrusted = rdonly_mem && (reg->type & PTR_UNTRUSTED);
7610 
7611 		if (type_may_be_null(reg->type)) {
7612 			verbose(env, "R%d invalid mem access '%s'\n", regno,
7613 				reg_type_str(env, reg->type));
7614 			return -EACCES;
7615 		}
7616 
7617 		if (t == BPF_WRITE && rdonly_mem) {
7618 			verbose(env, "R%d cannot write into %s\n",
7619 				regno, reg_type_str(env, reg->type));
7620 			return -EACCES;
7621 		}
7622 
7623 		if (t == BPF_WRITE && value_regno >= 0 &&
7624 		    is_pointer_value(env, value_regno)) {
7625 			verbose(env, "R%d leaks addr into mem\n", value_regno);
7626 			return -EACCES;
7627 		}
7628 
7629 		/*
7630 		 * Accesses to untrusted PTR_TO_MEM are done through probe
7631 		 * instructions, hence no need to check bounds in that case.
7632 		 */
7633 		if (!rdonly_untrusted)
7634 			err = check_mem_region_access(env, regno, off, size,
7635 						      reg->mem_size, false);
7636 		if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
7637 			mark_reg_unknown(env, regs, value_regno);
7638 	} else if (reg->type == PTR_TO_CTX) {
7639 		struct bpf_retval_range range;
7640 		struct bpf_insn_access_aux info = {
7641 			.reg_type = SCALAR_VALUE,
7642 			.is_ldsx = is_ldsx,
7643 			.log = &env->log,
7644 		};
7645 
7646 		if (t == BPF_WRITE && value_regno >= 0 &&
7647 		    is_pointer_value(env, value_regno)) {
7648 			verbose(env, "R%d leaks addr into ctx\n", value_regno);
7649 			return -EACCES;
7650 		}
7651 
7652 		err = check_ptr_off_reg(env, reg, regno);
7653 		if (err < 0)
7654 			return err;
7655 
7656 		err = check_ctx_access(env, insn_idx, off, size, t, &info);
7657 		if (err)
7658 			verbose_linfo(env, insn_idx, "; ");
7659 		if (!err && t == BPF_READ && value_regno >= 0) {
7660 			/* ctx access returns either a scalar, or a
7661 			 * PTR_TO_PACKET[_META,_END]. In the latter
7662 			 * case, we know the offset is zero.
7663 			 */
7664 			if (info.reg_type == SCALAR_VALUE) {
7665 				if (info.is_retval && get_func_retval_range(env->prog, &range)) {
7666 					err = __mark_reg_s32_range(env, regs, value_regno,
7667 								   range.minval, range.maxval);
7668 					if (err)
7669 						return err;
7670 				} else {
7671 					mark_reg_unknown(env, regs, value_regno);
7672 				}
7673 			} else {
7674 				mark_reg_known_zero(env, regs,
7675 						    value_regno);
7676 				if (type_may_be_null(info.reg_type))
7677 					regs[value_regno].id = ++env->id_gen;
7678 				/* A load of ctx field could have different
7679 				 * actual load size with the one encoded in the
7680 				 * insn. When the dst is PTR, it is for sure not
7681 				 * a sub-register.
7682 				 */
7683 				regs[value_regno].subreg_def = DEF_NOT_SUBREG;
7684 				if (base_type(info.reg_type) == PTR_TO_BTF_ID) {
7685 					regs[value_regno].btf = info.btf;
7686 					regs[value_regno].btf_id = info.btf_id;
7687 					regs[value_regno].ref_obj_id = info.ref_obj_id;
7688 				}
7689 			}
7690 			regs[value_regno].type = info.reg_type;
7691 		}
7692 
7693 	} else if (reg->type == PTR_TO_STACK) {
7694 		/* Basic bounds checks. */
7695 		err = check_stack_access_within_bounds(env, regno, off, size, t);
7696 		if (err)
7697 			return err;
7698 
7699 		if (t == BPF_READ)
7700 			err = check_stack_read(env, regno, off, size,
7701 					       value_regno);
7702 		else
7703 			err = check_stack_write(env, regno, off, size,
7704 						value_regno, insn_idx);
7705 	} else if (reg_is_pkt_pointer(reg)) {
7706 		if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
7707 			verbose(env, "cannot write into packet\n");
7708 			return -EACCES;
7709 		}
7710 		if (t == BPF_WRITE && value_regno >= 0 &&
7711 		    is_pointer_value(env, value_regno)) {
7712 			verbose(env, "R%d leaks addr into packet\n",
7713 				value_regno);
7714 			return -EACCES;
7715 		}
7716 		err = check_packet_access(env, regno, off, size, false);
7717 		if (!err && t == BPF_READ && value_regno >= 0)
7718 			mark_reg_unknown(env, regs, value_regno);
7719 	} else if (reg->type == PTR_TO_FLOW_KEYS) {
7720 		if (t == BPF_WRITE && value_regno >= 0 &&
7721 		    is_pointer_value(env, value_regno)) {
7722 			verbose(env, "R%d leaks addr into flow keys\n",
7723 				value_regno);
7724 			return -EACCES;
7725 		}
7726 
7727 		err = check_flow_keys_access(env, off, size);
7728 		if (!err && t == BPF_READ && value_regno >= 0)
7729 			mark_reg_unknown(env, regs, value_regno);
7730 	} else if (type_is_sk_pointer(reg->type)) {
7731 		if (t == BPF_WRITE) {
7732 			verbose(env, "R%d cannot write into %s\n",
7733 				regno, reg_type_str(env, reg->type));
7734 			return -EACCES;
7735 		}
7736 		err = check_sock_access(env, insn_idx, regno, off, size, t);
7737 		if (!err && value_regno >= 0)
7738 			mark_reg_unknown(env, regs, value_regno);
7739 	} else if (reg->type == PTR_TO_TP_BUFFER) {
7740 		err = check_tp_buffer_access(env, reg, regno, off, size);
7741 		if (!err && t == BPF_READ && value_regno >= 0)
7742 			mark_reg_unknown(env, regs, value_regno);
7743 	} else if (base_type(reg->type) == PTR_TO_BTF_ID &&
7744 		   !type_may_be_null(reg->type)) {
7745 		err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
7746 					      value_regno);
7747 	} else if (reg->type == CONST_PTR_TO_MAP) {
7748 		err = check_ptr_to_map_access(env, regs, regno, off, size, t,
7749 					      value_regno);
7750 	} else if (base_type(reg->type) == PTR_TO_BUF) {
7751 		bool rdonly_mem = type_is_rdonly_mem(reg->type);
7752 		u32 *max_access;
7753 
7754 		if (rdonly_mem) {
7755 			if (t == BPF_WRITE) {
7756 				verbose(env, "R%d cannot write into %s\n",
7757 					regno, reg_type_str(env, reg->type));
7758 				return -EACCES;
7759 			}
7760 			max_access = &env->prog->aux->max_rdonly_access;
7761 		} else {
7762 			max_access = &env->prog->aux->max_rdwr_access;
7763 		}
7764 
7765 		err = check_buffer_access(env, reg, regno, off, size, false,
7766 					  max_access);
7767 
7768 		if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
7769 			mark_reg_unknown(env, regs, value_regno);
7770 	} else if (reg->type == PTR_TO_ARENA) {
7771 		if (t == BPF_READ && value_regno >= 0)
7772 			mark_reg_unknown(env, regs, value_regno);
7773 	} else {
7774 		verbose(env, "R%d invalid mem access '%s'\n", regno,
7775 			reg_type_str(env, reg->type));
7776 		return -EACCES;
7777 	}
7778 
7779 	if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
7780 	    regs[value_regno].type == SCALAR_VALUE) {
7781 		if (!is_ldsx)
7782 			/* b/h/w load zero-extends, mark upper bits as known 0 */
7783 			coerce_reg_to_size(&regs[value_regno], size);
7784 		else
7785 			coerce_reg_to_size_sx(&regs[value_regno], size);
7786 	}
7787 	return err;
7788 }
7789 
7790 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
7791 			     bool allow_trust_mismatch);
7792 
check_load_mem(struct bpf_verifier_env * env,struct bpf_insn * insn,bool strict_alignment_once,bool is_ldsx,bool allow_trust_mismatch,const char * ctx)7793 static int check_load_mem(struct bpf_verifier_env *env, struct bpf_insn *insn,
7794 			  bool strict_alignment_once, bool is_ldsx,
7795 			  bool allow_trust_mismatch, const char *ctx)
7796 {
7797 	struct bpf_reg_state *regs = cur_regs(env);
7798 	enum bpf_reg_type src_reg_type;
7799 	int err;
7800 
7801 	/* check src operand */
7802 	err = check_reg_arg(env, insn->src_reg, SRC_OP);
7803 	if (err)
7804 		return err;
7805 
7806 	/* check dst operand */
7807 	err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7808 	if (err)
7809 		return err;
7810 
7811 	src_reg_type = regs[insn->src_reg].type;
7812 
7813 	/* Check if (src_reg + off) is readable. The state of dst_reg will be
7814 	 * updated by this call.
7815 	 */
7816 	err = check_mem_access(env, env->insn_idx, insn->src_reg, insn->off,
7817 			       BPF_SIZE(insn->code), BPF_READ, insn->dst_reg,
7818 			       strict_alignment_once, is_ldsx);
7819 	err = err ?: save_aux_ptr_type(env, src_reg_type,
7820 				       allow_trust_mismatch);
7821 	err = err ?: reg_bounds_sanity_check(env, &regs[insn->dst_reg], ctx);
7822 
7823 	return err;
7824 }
7825 
check_store_reg(struct bpf_verifier_env * env,struct bpf_insn * insn,bool strict_alignment_once)7826 static int check_store_reg(struct bpf_verifier_env *env, struct bpf_insn *insn,
7827 			   bool strict_alignment_once)
7828 {
7829 	struct bpf_reg_state *regs = cur_regs(env);
7830 	enum bpf_reg_type dst_reg_type;
7831 	int err;
7832 
7833 	/* check src1 operand */
7834 	err = check_reg_arg(env, insn->src_reg, SRC_OP);
7835 	if (err)
7836 		return err;
7837 
7838 	/* check src2 operand */
7839 	err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7840 	if (err)
7841 		return err;
7842 
7843 	dst_reg_type = regs[insn->dst_reg].type;
7844 
7845 	/* Check if (dst_reg + off) is writeable. */
7846 	err = check_mem_access(env, env->insn_idx, insn->dst_reg, insn->off,
7847 			       BPF_SIZE(insn->code), BPF_WRITE, insn->src_reg,
7848 			       strict_alignment_once, false);
7849 	err = err ?: save_aux_ptr_type(env, dst_reg_type, false);
7850 
7851 	return err;
7852 }
7853 
check_atomic_rmw(struct bpf_verifier_env * env,struct bpf_insn * insn)7854 static int check_atomic_rmw(struct bpf_verifier_env *env,
7855 			    struct bpf_insn *insn)
7856 {
7857 	int load_reg;
7858 	int err;
7859 
7860 	if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
7861 		verbose(env, "invalid atomic operand size\n");
7862 		return -EINVAL;
7863 	}
7864 
7865 	/* check src1 operand */
7866 	err = check_reg_arg(env, insn->src_reg, SRC_OP);
7867 	if (err)
7868 		return err;
7869 
7870 	/* check src2 operand */
7871 	err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7872 	if (err)
7873 		return err;
7874 
7875 	if (insn->imm == BPF_CMPXCHG) {
7876 		/* Check comparison of R0 with memory location */
7877 		const u32 aux_reg = BPF_REG_0;
7878 
7879 		err = check_reg_arg(env, aux_reg, SRC_OP);
7880 		if (err)
7881 			return err;
7882 
7883 		if (is_pointer_value(env, aux_reg)) {
7884 			verbose(env, "R%d leaks addr into mem\n", aux_reg);
7885 			return -EACCES;
7886 		}
7887 	}
7888 
7889 	if (is_pointer_value(env, insn->src_reg)) {
7890 		verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
7891 		return -EACCES;
7892 	}
7893 
7894 	if (!atomic_ptr_type_ok(env, insn->dst_reg, insn)) {
7895 		verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
7896 			insn->dst_reg,
7897 			reg_type_str(env, reg_state(env, insn->dst_reg)->type));
7898 		return -EACCES;
7899 	}
7900 
7901 	if (insn->imm & BPF_FETCH) {
7902 		if (insn->imm == BPF_CMPXCHG)
7903 			load_reg = BPF_REG_0;
7904 		else
7905 			load_reg = insn->src_reg;
7906 
7907 		/* check and record load of old value */
7908 		err = check_reg_arg(env, load_reg, DST_OP);
7909 		if (err)
7910 			return err;
7911 	} else {
7912 		/* This instruction accesses a memory location but doesn't
7913 		 * actually load it into a register.
7914 		 */
7915 		load_reg = -1;
7916 	}
7917 
7918 	/* Check whether we can read the memory, with second call for fetch
7919 	 * case to simulate the register fill.
7920 	 */
7921 	err = check_mem_access(env, env->insn_idx, insn->dst_reg, insn->off,
7922 			       BPF_SIZE(insn->code), BPF_READ, -1, true, false);
7923 	if (!err && load_reg >= 0)
7924 		err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7925 				       insn->off, BPF_SIZE(insn->code),
7926 				       BPF_READ, load_reg, true, false);
7927 	if (err)
7928 		return err;
7929 
7930 	if (is_arena_reg(env, insn->dst_reg)) {
7931 		err = save_aux_ptr_type(env, PTR_TO_ARENA, false);
7932 		if (err)
7933 			return err;
7934 	}
7935 	/* Check whether we can write into the same memory. */
7936 	err = check_mem_access(env, env->insn_idx, insn->dst_reg, insn->off,
7937 			       BPF_SIZE(insn->code), BPF_WRITE, -1, true, false);
7938 	if (err)
7939 		return err;
7940 	return 0;
7941 }
7942 
check_atomic_load(struct bpf_verifier_env * env,struct bpf_insn * insn)7943 static int check_atomic_load(struct bpf_verifier_env *env,
7944 			     struct bpf_insn *insn)
7945 {
7946 	int err;
7947 
7948 	err = check_load_mem(env, insn, true, false, false, "atomic_load");
7949 	if (err)
7950 		return err;
7951 
7952 	if (!atomic_ptr_type_ok(env, insn->src_reg, insn)) {
7953 		verbose(env, "BPF_ATOMIC loads from R%d %s is not allowed\n",
7954 			insn->src_reg,
7955 			reg_type_str(env, reg_state(env, insn->src_reg)->type));
7956 		return -EACCES;
7957 	}
7958 
7959 	return 0;
7960 }
7961 
check_atomic_store(struct bpf_verifier_env * env,struct bpf_insn * insn)7962 static int check_atomic_store(struct bpf_verifier_env *env,
7963 			      struct bpf_insn *insn)
7964 {
7965 	int err;
7966 
7967 	err = check_store_reg(env, insn, true);
7968 	if (err)
7969 		return err;
7970 
7971 	if (!atomic_ptr_type_ok(env, insn->dst_reg, insn)) {
7972 		verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
7973 			insn->dst_reg,
7974 			reg_type_str(env, reg_state(env, insn->dst_reg)->type));
7975 		return -EACCES;
7976 	}
7977 
7978 	return 0;
7979 }
7980 
check_atomic(struct bpf_verifier_env * env,struct bpf_insn * insn)7981 static int check_atomic(struct bpf_verifier_env *env, struct bpf_insn *insn)
7982 {
7983 	switch (insn->imm) {
7984 	case BPF_ADD:
7985 	case BPF_ADD | BPF_FETCH:
7986 	case BPF_AND:
7987 	case BPF_AND | BPF_FETCH:
7988 	case BPF_OR:
7989 	case BPF_OR | BPF_FETCH:
7990 	case BPF_XOR:
7991 	case BPF_XOR | BPF_FETCH:
7992 	case BPF_XCHG:
7993 	case BPF_CMPXCHG:
7994 		return check_atomic_rmw(env, insn);
7995 	case BPF_LOAD_ACQ:
7996 		if (BPF_SIZE(insn->code) == BPF_DW && BITS_PER_LONG != 64) {
7997 			verbose(env,
7998 				"64-bit load-acquires are only supported on 64-bit arches\n");
7999 			return -EOPNOTSUPP;
8000 		}
8001 		return check_atomic_load(env, insn);
8002 	case BPF_STORE_REL:
8003 		if (BPF_SIZE(insn->code) == BPF_DW && BITS_PER_LONG != 64) {
8004 			verbose(env,
8005 				"64-bit store-releases are only supported on 64-bit arches\n");
8006 			return -EOPNOTSUPP;
8007 		}
8008 		return check_atomic_store(env, insn);
8009 	default:
8010 		verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n",
8011 			insn->imm);
8012 		return -EINVAL;
8013 	}
8014 }
8015 
8016 /* When register 'regno' is used to read the stack (either directly or through
8017  * a helper function) make sure that it's within stack boundary and, depending
8018  * on the access type and privileges, that all elements of the stack are
8019  * initialized.
8020  *
8021  * 'off' includes 'regno->off', but not its dynamic part (if any).
8022  *
8023  * All registers that have been spilled on the stack in the slots within the
8024  * read offsets are marked as read.
8025  */
check_stack_range_initialized(struct bpf_verifier_env * env,int regno,int off,int access_size,bool zero_size_allowed,enum bpf_access_type type,struct bpf_call_arg_meta * meta)8026 static int check_stack_range_initialized(
8027 		struct bpf_verifier_env *env, int regno, int off,
8028 		int access_size, bool zero_size_allowed,
8029 		enum bpf_access_type type, struct bpf_call_arg_meta *meta)
8030 {
8031 	struct bpf_reg_state *reg = reg_state(env, regno);
8032 	struct bpf_func_state *state = func(env, reg);
8033 	int err, min_off, max_off, i, j, slot, spi;
8034 	/* Some accesses can write anything into the stack, others are
8035 	 * read-only.
8036 	 */
8037 	bool clobber = false;
8038 
8039 	if (access_size == 0 && !zero_size_allowed) {
8040 		verbose(env, "invalid zero-sized read\n");
8041 		return -EACCES;
8042 	}
8043 
8044 	if (type == BPF_WRITE)
8045 		clobber = true;
8046 
8047 	err = check_stack_access_within_bounds(env, regno, off, access_size, type);
8048 	if (err)
8049 		return err;
8050 
8051 
8052 	if (tnum_is_const(reg->var_off)) {
8053 		min_off = max_off = reg->var_off.value + off;
8054 	} else {
8055 		/* Variable offset is prohibited for unprivileged mode for
8056 		 * simplicity since it requires corresponding support in
8057 		 * Spectre masking for stack ALU.
8058 		 * See also retrieve_ptr_limit().
8059 		 */
8060 		if (!env->bypass_spec_v1) {
8061 			char tn_buf[48];
8062 
8063 			tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
8064 			verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
8065 				regno, tn_buf);
8066 			return -EACCES;
8067 		}
8068 		/* Only initialized buffer on stack is allowed to be accessed
8069 		 * with variable offset. With uninitialized buffer it's hard to
8070 		 * guarantee that whole memory is marked as initialized on
8071 		 * helper return since specific bounds are unknown what may
8072 		 * cause uninitialized stack leaking.
8073 		 */
8074 		if (meta && meta->raw_mode)
8075 			meta = NULL;
8076 
8077 		min_off = reg->smin_value + off;
8078 		max_off = reg->smax_value + off;
8079 	}
8080 
8081 	if (meta && meta->raw_mode) {
8082 		/* Ensure we won't be overwriting dynptrs when simulating byte
8083 		 * by byte access in check_helper_call using meta.access_size.
8084 		 * This would be a problem if we have a helper in the future
8085 		 * which takes:
8086 		 *
8087 		 *	helper(uninit_mem, len, dynptr)
8088 		 *
8089 		 * Now, uninint_mem may overlap with dynptr pointer. Hence, it
8090 		 * may end up writing to dynptr itself when touching memory from
8091 		 * arg 1. This can be relaxed on a case by case basis for known
8092 		 * safe cases, but reject due to the possibilitiy of aliasing by
8093 		 * default.
8094 		 */
8095 		for (i = min_off; i < max_off + access_size; i++) {
8096 			int stack_off = -i - 1;
8097 
8098 			spi = __get_spi(i);
8099 			/* raw_mode may write past allocated_stack */
8100 			if (state->allocated_stack <= stack_off)
8101 				continue;
8102 			if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
8103 				verbose(env, "potential write to dynptr at off=%d disallowed\n", i);
8104 				return -EACCES;
8105 			}
8106 		}
8107 		meta->access_size = access_size;
8108 		meta->regno = regno;
8109 		return 0;
8110 	}
8111 
8112 	for (i = min_off; i < max_off + access_size; i++) {
8113 		u8 *stype;
8114 
8115 		slot = -i - 1;
8116 		spi = slot / BPF_REG_SIZE;
8117 		if (state->allocated_stack <= slot) {
8118 			verbose(env, "allocated_stack too small\n");
8119 			return -EFAULT;
8120 		}
8121 
8122 		stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
8123 		if (*stype == STACK_MISC)
8124 			goto mark;
8125 		if ((*stype == STACK_ZERO) ||
8126 		    (*stype == STACK_INVALID && env->allow_uninit_stack)) {
8127 			if (clobber) {
8128 				/* helper can write anything into the stack */
8129 				*stype = STACK_MISC;
8130 			}
8131 			goto mark;
8132 		}
8133 
8134 		if (is_spilled_reg(&state->stack[spi]) &&
8135 		    (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
8136 		     env->allow_ptr_leaks)) {
8137 			if (clobber) {
8138 				__mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
8139 				for (j = 0; j < BPF_REG_SIZE; j++)
8140 					scrub_spilled_slot(&state->stack[spi].slot_type[j]);
8141 			}
8142 			goto mark;
8143 		}
8144 
8145 		if (tnum_is_const(reg->var_off)) {
8146 			verbose(env, "invalid read from stack R%d off %d+%d size %d\n",
8147 				regno, min_off, i - min_off, access_size);
8148 		} else {
8149 			char tn_buf[48];
8150 
8151 			tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
8152 			verbose(env, "invalid read from stack R%d var_off %s+%d size %d\n",
8153 				regno, tn_buf, i - min_off, access_size);
8154 		}
8155 		return -EACCES;
8156 mark:
8157 		/* reading any byte out of 8-byte 'spill_slot' will cause
8158 		 * the whole slot to be marked as 'read'
8159 		 */
8160 		mark_reg_read(env, &state->stack[spi].spilled_ptr,
8161 			      state->stack[spi].spilled_ptr.parent,
8162 			      REG_LIVE_READ64);
8163 		/* We do not set REG_LIVE_WRITTEN for stack slot, as we can not
8164 		 * be sure that whether stack slot is written to or not. Hence,
8165 		 * we must still conservatively propagate reads upwards even if
8166 		 * helper may write to the entire memory range.
8167 		 */
8168 	}
8169 	return 0;
8170 }
8171 
check_helper_mem_access(struct bpf_verifier_env * env,int regno,int access_size,enum bpf_access_type access_type,bool zero_size_allowed,struct bpf_call_arg_meta * meta)8172 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
8173 				   int access_size, enum bpf_access_type access_type,
8174 				   bool zero_size_allowed,
8175 				   struct bpf_call_arg_meta *meta)
8176 {
8177 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8178 	u32 *max_access;
8179 
8180 	switch (base_type(reg->type)) {
8181 	case PTR_TO_PACKET:
8182 	case PTR_TO_PACKET_META:
8183 		return check_packet_access(env, regno, reg->off, access_size,
8184 					   zero_size_allowed);
8185 	case PTR_TO_MAP_KEY:
8186 		if (access_type == BPF_WRITE) {
8187 			verbose(env, "R%d cannot write into %s\n", regno,
8188 				reg_type_str(env, reg->type));
8189 			return -EACCES;
8190 		}
8191 		return check_mem_region_access(env, regno, reg->off, access_size,
8192 					       reg->map_ptr->key_size, false);
8193 	case PTR_TO_MAP_VALUE:
8194 		if (check_map_access_type(env, regno, reg->off, access_size, access_type))
8195 			return -EACCES;
8196 		return check_map_access(env, regno, reg->off, access_size,
8197 					zero_size_allowed, ACCESS_HELPER);
8198 	case PTR_TO_MEM:
8199 		if (type_is_rdonly_mem(reg->type)) {
8200 			if (access_type == BPF_WRITE) {
8201 				verbose(env, "R%d cannot write into %s\n", regno,
8202 					reg_type_str(env, reg->type));
8203 				return -EACCES;
8204 			}
8205 		}
8206 		return check_mem_region_access(env, regno, reg->off,
8207 					       access_size, reg->mem_size,
8208 					       zero_size_allowed);
8209 	case PTR_TO_BUF:
8210 		if (type_is_rdonly_mem(reg->type)) {
8211 			if (access_type == BPF_WRITE) {
8212 				verbose(env, "R%d cannot write into %s\n", regno,
8213 					reg_type_str(env, reg->type));
8214 				return -EACCES;
8215 			}
8216 
8217 			max_access = &env->prog->aux->max_rdonly_access;
8218 		} else {
8219 			max_access = &env->prog->aux->max_rdwr_access;
8220 		}
8221 		return check_buffer_access(env, reg, regno, reg->off,
8222 					   access_size, zero_size_allowed,
8223 					   max_access);
8224 	case PTR_TO_STACK:
8225 		return check_stack_range_initialized(
8226 				env,
8227 				regno, reg->off, access_size,
8228 				zero_size_allowed, access_type, meta);
8229 	case PTR_TO_BTF_ID:
8230 		return check_ptr_to_btf_access(env, regs, regno, reg->off,
8231 					       access_size, BPF_READ, -1);
8232 	case PTR_TO_CTX:
8233 		/* in case the function doesn't know how to access the context,
8234 		 * (because we are in a program of type SYSCALL for example), we
8235 		 * can not statically check its size.
8236 		 * Dynamically check it now.
8237 		 */
8238 		if (!env->ops->convert_ctx_access) {
8239 			int offset = access_size - 1;
8240 
8241 			/* Allow zero-byte read from PTR_TO_CTX */
8242 			if (access_size == 0)
8243 				return zero_size_allowed ? 0 : -EACCES;
8244 
8245 			return check_mem_access(env, env->insn_idx, regno, offset, BPF_B,
8246 						access_type, -1, false, false);
8247 		}
8248 
8249 		fallthrough;
8250 	default: /* scalar_value or invalid ptr */
8251 		/* Allow zero-byte read from NULL, regardless of pointer type */
8252 		if (zero_size_allowed && access_size == 0 &&
8253 		    register_is_null(reg))
8254 			return 0;
8255 
8256 		verbose(env, "R%d type=%s ", regno,
8257 			reg_type_str(env, reg->type));
8258 		verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
8259 		return -EACCES;
8260 	}
8261 }
8262 
8263 /* verify arguments to helpers or kfuncs consisting of a pointer and an access
8264  * size.
8265  *
8266  * @regno is the register containing the access size. regno-1 is the register
8267  * containing the pointer.
8268  */
check_mem_size_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,enum bpf_access_type access_type,bool zero_size_allowed,struct bpf_call_arg_meta * meta)8269 static int check_mem_size_reg(struct bpf_verifier_env *env,
8270 			      struct bpf_reg_state *reg, u32 regno,
8271 			      enum bpf_access_type access_type,
8272 			      bool zero_size_allowed,
8273 			      struct bpf_call_arg_meta *meta)
8274 {
8275 	int err;
8276 
8277 	/* This is used to refine r0 return value bounds for helpers
8278 	 * that enforce this value as an upper bound on return values.
8279 	 * See do_refine_retval_range() for helpers that can refine
8280 	 * the return value. C type of helper is u32 so we pull register
8281 	 * bound from umax_value however, if negative verifier errors
8282 	 * out. Only upper bounds can be learned because retval is an
8283 	 * int type and negative retvals are allowed.
8284 	 */
8285 	meta->msize_max_value = reg->umax_value;
8286 
8287 	/* The register is SCALAR_VALUE; the access check happens using
8288 	 * its boundaries. For unprivileged variable accesses, disable
8289 	 * raw mode so that the program is required to initialize all
8290 	 * the memory that the helper could just partially fill up.
8291 	 */
8292 	if (!tnum_is_const(reg->var_off))
8293 		meta = NULL;
8294 
8295 	if (reg->smin_value < 0) {
8296 		verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
8297 			regno);
8298 		return -EACCES;
8299 	}
8300 
8301 	if (reg->umin_value == 0 && !zero_size_allowed) {
8302 		verbose(env, "R%d invalid zero-sized read: u64=[%lld,%lld]\n",
8303 			regno, reg->umin_value, reg->umax_value);
8304 		return -EACCES;
8305 	}
8306 
8307 	if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
8308 		verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
8309 			regno);
8310 		return -EACCES;
8311 	}
8312 	err = check_helper_mem_access(env, regno - 1, reg->umax_value,
8313 				      access_type, zero_size_allowed, meta);
8314 	if (!err)
8315 		err = mark_chain_precision(env, regno);
8316 	return err;
8317 }
8318 
check_mem_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,u32 mem_size)8319 static int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
8320 			 u32 regno, u32 mem_size)
8321 {
8322 	bool may_be_null = type_may_be_null(reg->type);
8323 	struct bpf_reg_state saved_reg;
8324 	int err;
8325 
8326 	if (register_is_null(reg))
8327 		return 0;
8328 
8329 	/* Assuming that the register contains a value check if the memory
8330 	 * access is safe. Temporarily save and restore the register's state as
8331 	 * the conversion shouldn't be visible to a caller.
8332 	 */
8333 	if (may_be_null) {
8334 		saved_reg = *reg;
8335 		mark_ptr_not_null_reg(reg);
8336 	}
8337 
8338 	err = check_helper_mem_access(env, regno, mem_size, BPF_READ, true, NULL);
8339 	err = err ?: check_helper_mem_access(env, regno, mem_size, BPF_WRITE, true, NULL);
8340 
8341 	if (may_be_null)
8342 		*reg = saved_reg;
8343 
8344 	return err;
8345 }
8346 
check_kfunc_mem_size_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno)8347 static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
8348 				    u32 regno)
8349 {
8350 	struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
8351 	bool may_be_null = type_may_be_null(mem_reg->type);
8352 	struct bpf_reg_state saved_reg;
8353 	struct bpf_call_arg_meta meta;
8354 	int err;
8355 
8356 	WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
8357 
8358 	memset(&meta, 0, sizeof(meta));
8359 
8360 	if (may_be_null) {
8361 		saved_reg = *mem_reg;
8362 		mark_ptr_not_null_reg(mem_reg);
8363 	}
8364 
8365 	err = check_mem_size_reg(env, reg, regno, BPF_READ, true, &meta);
8366 	err = err ?: check_mem_size_reg(env, reg, regno, BPF_WRITE, true, &meta);
8367 
8368 	if (may_be_null)
8369 		*mem_reg = saved_reg;
8370 
8371 	return err;
8372 }
8373 
8374 enum {
8375 	PROCESS_SPIN_LOCK = (1 << 0),
8376 	PROCESS_RES_LOCK  = (1 << 1),
8377 	PROCESS_LOCK_IRQ  = (1 << 2),
8378 };
8379 
8380 /* Implementation details:
8381  * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
8382  * bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
8383  * Two bpf_map_lookups (even with the same key) will have different reg->id.
8384  * Two separate bpf_obj_new will also have different reg->id.
8385  * For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
8386  * clears reg->id after value_or_null->value transition, since the verifier only
8387  * cares about the range of access to valid map value pointer and doesn't care
8388  * about actual address of the map element.
8389  * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
8390  * reg->id > 0 after value_or_null->value transition. By doing so
8391  * two bpf_map_lookups will be considered two different pointers that
8392  * point to different bpf_spin_locks. Likewise for pointers to allocated objects
8393  * returned from bpf_obj_new.
8394  * The verifier allows taking only one bpf_spin_lock at a time to avoid
8395  * dead-locks.
8396  * Since only one bpf_spin_lock is allowed the checks are simpler than
8397  * reg_is_refcounted() logic. The verifier needs to remember only
8398  * one spin_lock instead of array of acquired_refs.
8399  * env->cur_state->active_locks remembers which map value element or allocated
8400  * object got locked and clears it after bpf_spin_unlock.
8401  */
process_spin_lock(struct bpf_verifier_env * env,int regno,int flags)8402 static int process_spin_lock(struct bpf_verifier_env *env, int regno, int flags)
8403 {
8404 	bool is_lock = flags & PROCESS_SPIN_LOCK, is_res_lock = flags & PROCESS_RES_LOCK;
8405 	const char *lock_str = is_res_lock ? "bpf_res_spin" : "bpf_spin";
8406 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8407 	struct bpf_verifier_state *cur = env->cur_state;
8408 	bool is_const = tnum_is_const(reg->var_off);
8409 	bool is_irq = flags & PROCESS_LOCK_IRQ;
8410 	u64 val = reg->var_off.value;
8411 	struct bpf_map *map = NULL;
8412 	struct btf *btf = NULL;
8413 	struct btf_record *rec;
8414 	u32 spin_lock_off;
8415 	int err;
8416 
8417 	if (!is_const) {
8418 		verbose(env,
8419 			"R%d doesn't have constant offset. %s_lock has to be at the constant offset\n",
8420 			regno, lock_str);
8421 		return -EINVAL;
8422 	}
8423 	if (reg->type == PTR_TO_MAP_VALUE) {
8424 		map = reg->map_ptr;
8425 		if (!map->btf) {
8426 			verbose(env,
8427 				"map '%s' has to have BTF in order to use %s_lock\n",
8428 				map->name, lock_str);
8429 			return -EINVAL;
8430 		}
8431 	} else {
8432 		btf = reg->btf;
8433 	}
8434 
8435 	rec = reg_btf_record(reg);
8436 	if (!btf_record_has_field(rec, is_res_lock ? BPF_RES_SPIN_LOCK : BPF_SPIN_LOCK)) {
8437 		verbose(env, "%s '%s' has no valid %s_lock\n", map ? "map" : "local",
8438 			map ? map->name : "kptr", lock_str);
8439 		return -EINVAL;
8440 	}
8441 	spin_lock_off = is_res_lock ? rec->res_spin_lock_off : rec->spin_lock_off;
8442 	if (spin_lock_off != val + reg->off) {
8443 		verbose(env, "off %lld doesn't point to 'struct %s_lock' that is at %d\n",
8444 			val + reg->off, lock_str, spin_lock_off);
8445 		return -EINVAL;
8446 	}
8447 	if (is_lock) {
8448 		void *ptr;
8449 		int type;
8450 
8451 		if (map)
8452 			ptr = map;
8453 		else
8454 			ptr = btf;
8455 
8456 		if (!is_res_lock && cur->active_locks) {
8457 			if (find_lock_state(env->cur_state, REF_TYPE_LOCK, 0, NULL)) {
8458 				verbose(env,
8459 					"Locking two bpf_spin_locks are not allowed\n");
8460 				return -EINVAL;
8461 			}
8462 		} else if (is_res_lock && cur->active_locks) {
8463 			if (find_lock_state(env->cur_state, REF_TYPE_RES_LOCK | REF_TYPE_RES_LOCK_IRQ, reg->id, ptr)) {
8464 				verbose(env, "Acquiring the same lock again, AA deadlock detected\n");
8465 				return -EINVAL;
8466 			}
8467 		}
8468 
8469 		if (is_res_lock && is_irq)
8470 			type = REF_TYPE_RES_LOCK_IRQ;
8471 		else if (is_res_lock)
8472 			type = REF_TYPE_RES_LOCK;
8473 		else
8474 			type = REF_TYPE_LOCK;
8475 		err = acquire_lock_state(env, env->insn_idx, type, reg->id, ptr);
8476 		if (err < 0) {
8477 			verbose(env, "Failed to acquire lock state\n");
8478 			return err;
8479 		}
8480 	} else {
8481 		void *ptr;
8482 		int type;
8483 
8484 		if (map)
8485 			ptr = map;
8486 		else
8487 			ptr = btf;
8488 
8489 		if (!cur->active_locks) {
8490 			verbose(env, "%s_unlock without taking a lock\n", lock_str);
8491 			return -EINVAL;
8492 		}
8493 
8494 		if (is_res_lock && is_irq)
8495 			type = REF_TYPE_RES_LOCK_IRQ;
8496 		else if (is_res_lock)
8497 			type = REF_TYPE_RES_LOCK;
8498 		else
8499 			type = REF_TYPE_LOCK;
8500 		if (!find_lock_state(cur, type, reg->id, ptr)) {
8501 			verbose(env, "%s_unlock of different lock\n", lock_str);
8502 			return -EINVAL;
8503 		}
8504 		if (reg->id != cur->active_lock_id || ptr != cur->active_lock_ptr) {
8505 			verbose(env, "%s_unlock cannot be out of order\n", lock_str);
8506 			return -EINVAL;
8507 		}
8508 		if (release_lock_state(cur, type, reg->id, ptr)) {
8509 			verbose(env, "%s_unlock of different lock\n", lock_str);
8510 			return -EINVAL;
8511 		}
8512 
8513 		invalidate_non_owning_refs(env);
8514 	}
8515 	return 0;
8516 }
8517 
process_timer_func(struct bpf_verifier_env * env,int regno,struct bpf_call_arg_meta * meta)8518 static int process_timer_func(struct bpf_verifier_env *env, int regno,
8519 			      struct bpf_call_arg_meta *meta)
8520 {
8521 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8522 	bool is_const = tnum_is_const(reg->var_off);
8523 	struct bpf_map *map = reg->map_ptr;
8524 	u64 val = reg->var_off.value;
8525 
8526 	if (!is_const) {
8527 		verbose(env,
8528 			"R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
8529 			regno);
8530 		return -EINVAL;
8531 	}
8532 	if (!map->btf) {
8533 		verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
8534 			map->name);
8535 		return -EINVAL;
8536 	}
8537 	if (!btf_record_has_field(map->record, BPF_TIMER)) {
8538 		verbose(env, "map '%s' has no valid bpf_timer\n", map->name);
8539 		return -EINVAL;
8540 	}
8541 	if (map->record->timer_off != val + reg->off) {
8542 		verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
8543 			val + reg->off, map->record->timer_off);
8544 		return -EINVAL;
8545 	}
8546 	if (meta->map_ptr) {
8547 		verifier_bug(env, "Two map pointers in a timer helper");
8548 		return -EFAULT;
8549 	}
8550 	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
8551 		verbose(env, "bpf_timer cannot be used for PREEMPT_RT.\n");
8552 		return -EOPNOTSUPP;
8553 	}
8554 	meta->map_uid = reg->map_uid;
8555 	meta->map_ptr = map;
8556 	return 0;
8557 }
8558 
process_wq_func(struct bpf_verifier_env * env,int regno,struct bpf_kfunc_call_arg_meta * meta)8559 static int process_wq_func(struct bpf_verifier_env *env, int regno,
8560 			   struct bpf_kfunc_call_arg_meta *meta)
8561 {
8562 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8563 	struct bpf_map *map = reg->map_ptr;
8564 	u64 val = reg->var_off.value;
8565 
8566 	if (map->record->wq_off != val + reg->off) {
8567 		verbose(env, "off %lld doesn't point to 'struct bpf_wq' that is at %d\n",
8568 			val + reg->off, map->record->wq_off);
8569 		return -EINVAL;
8570 	}
8571 	meta->map.uid = reg->map_uid;
8572 	meta->map.ptr = map;
8573 	return 0;
8574 }
8575 
process_kptr_func(struct bpf_verifier_env * env,int regno,struct bpf_call_arg_meta * meta)8576 static int process_kptr_func(struct bpf_verifier_env *env, int regno,
8577 			     struct bpf_call_arg_meta *meta)
8578 {
8579 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8580 	struct btf_field *kptr_field;
8581 	struct bpf_map *map_ptr;
8582 	struct btf_record *rec;
8583 	u32 kptr_off;
8584 
8585 	if (type_is_ptr_alloc_obj(reg->type)) {
8586 		rec = reg_btf_record(reg);
8587 	} else { /* PTR_TO_MAP_VALUE */
8588 		map_ptr = reg->map_ptr;
8589 		if (!map_ptr->btf) {
8590 			verbose(env, "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
8591 				map_ptr->name);
8592 			return -EINVAL;
8593 		}
8594 		rec = map_ptr->record;
8595 		meta->map_ptr = map_ptr;
8596 	}
8597 
8598 	if (!tnum_is_const(reg->var_off)) {
8599 		verbose(env,
8600 			"R%d doesn't have constant offset. kptr has to be at the constant offset\n",
8601 			regno);
8602 		return -EINVAL;
8603 	}
8604 
8605 	if (!btf_record_has_field(rec, BPF_KPTR)) {
8606 		verbose(env, "R%d has no valid kptr\n", regno);
8607 		return -EINVAL;
8608 	}
8609 
8610 	kptr_off = reg->off + reg->var_off.value;
8611 	kptr_field = btf_record_find(rec, kptr_off, BPF_KPTR);
8612 	if (!kptr_field) {
8613 		verbose(env, "off=%d doesn't point to kptr\n", kptr_off);
8614 		return -EACCES;
8615 	}
8616 	if (kptr_field->type != BPF_KPTR_REF && kptr_field->type != BPF_KPTR_PERCPU) {
8617 		verbose(env, "off=%d kptr isn't referenced kptr\n", kptr_off);
8618 		return -EACCES;
8619 	}
8620 	meta->kptr_field = kptr_field;
8621 	return 0;
8622 }
8623 
8624 /* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
8625  * which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
8626  *
8627  * In both cases we deal with the first 8 bytes, but need to mark the next 8
8628  * bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
8629  * CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
8630  *
8631  * Mutability of bpf_dynptr is at two levels, one is at the level of struct
8632  * bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
8633  * bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
8634  * mutate the view of the dynptr and also possibly destroy it. In the latter
8635  * case, it cannot mutate the bpf_dynptr itself but it can still mutate the
8636  * memory that dynptr points to.
8637  *
8638  * The verifier will keep track both levels of mutation (bpf_dynptr's in
8639  * reg->type and the memory's in reg->dynptr.type), but there is no support for
8640  * readonly dynptr view yet, hence only the first case is tracked and checked.
8641  *
8642  * This is consistent with how C applies the const modifier to a struct object,
8643  * where the pointer itself inside bpf_dynptr becomes const but not what it
8644  * points to.
8645  *
8646  * Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
8647  * type, and declare it as 'const struct bpf_dynptr *' in their prototype.
8648  */
process_dynptr_func(struct bpf_verifier_env * env,int regno,int insn_idx,enum bpf_arg_type arg_type,int clone_ref_obj_id)8649 static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
8650 			       enum bpf_arg_type arg_type, int clone_ref_obj_id)
8651 {
8652 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8653 	int err;
8654 
8655 	if (reg->type != PTR_TO_STACK && reg->type != CONST_PTR_TO_DYNPTR) {
8656 		verbose(env,
8657 			"arg#%d expected pointer to stack or const struct bpf_dynptr\n",
8658 			regno - 1);
8659 		return -EINVAL;
8660 	}
8661 
8662 	/* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
8663 	 * ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
8664 	 */
8665 	if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
8666 		verifier_bug(env, "misconfigured dynptr helper type flags");
8667 		return -EFAULT;
8668 	}
8669 
8670 	/*  MEM_UNINIT - Points to memory that is an appropriate candidate for
8671 	 *		 constructing a mutable bpf_dynptr object.
8672 	 *
8673 	 *		 Currently, this is only possible with PTR_TO_STACK
8674 	 *		 pointing to a region of at least 16 bytes which doesn't
8675 	 *		 contain an existing bpf_dynptr.
8676 	 *
8677 	 *  MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
8678 	 *		 mutated or destroyed. However, the memory it points to
8679 	 *		 may be mutated.
8680 	 *
8681 	 *  None       - Points to a initialized dynptr that can be mutated and
8682 	 *		 destroyed, including mutation of the memory it points
8683 	 *		 to.
8684 	 */
8685 	if (arg_type & MEM_UNINIT) {
8686 		int i;
8687 
8688 		if (!is_dynptr_reg_valid_uninit(env, reg)) {
8689 			verbose(env, "Dynptr has to be an uninitialized dynptr\n");
8690 			return -EINVAL;
8691 		}
8692 
8693 		/* we write BPF_DW bits (8 bytes) at a time */
8694 		for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
8695 			err = check_mem_access(env, insn_idx, regno,
8696 					       i, BPF_DW, BPF_WRITE, -1, false, false);
8697 			if (err)
8698 				return err;
8699 		}
8700 
8701 		err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
8702 	} else /* MEM_RDONLY and None case from above */ {
8703 		/* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
8704 		if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
8705 			verbose(env, "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
8706 			return -EINVAL;
8707 		}
8708 
8709 		if (!is_dynptr_reg_valid_init(env, reg)) {
8710 			verbose(env,
8711 				"Expected an initialized dynptr as arg #%d\n",
8712 				regno - 1);
8713 			return -EINVAL;
8714 		}
8715 
8716 		/* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
8717 		if (!is_dynptr_type_expected(env, reg, arg_type & ~MEM_RDONLY)) {
8718 			verbose(env,
8719 				"Expected a dynptr of type %s as arg #%d\n",
8720 				dynptr_type_str(arg_to_dynptr_type(arg_type)), regno - 1);
8721 			return -EINVAL;
8722 		}
8723 
8724 		err = mark_dynptr_read(env, reg);
8725 	}
8726 	return err;
8727 }
8728 
iter_ref_obj_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg,int spi)8729 static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
8730 {
8731 	struct bpf_func_state *state = func(env, reg);
8732 
8733 	return state->stack[spi].spilled_ptr.ref_obj_id;
8734 }
8735 
is_iter_kfunc(struct bpf_kfunc_call_arg_meta * meta)8736 static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8737 {
8738 	return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
8739 }
8740 
is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta * meta)8741 static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8742 {
8743 	return meta->kfunc_flags & KF_ITER_NEW;
8744 }
8745 
is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta * meta)8746 static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8747 {
8748 	return meta->kfunc_flags & KF_ITER_NEXT;
8749 }
8750 
is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta * meta)8751 static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8752 {
8753 	return meta->kfunc_flags & KF_ITER_DESTROY;
8754 }
8755 
is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta * meta,int arg_idx,const struct btf_param * arg)8756 static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg_idx,
8757 			      const struct btf_param *arg)
8758 {
8759 	/* btf_check_iter_kfuncs() guarantees that first argument of any iter
8760 	 * kfunc is iter state pointer
8761 	 */
8762 	if (is_iter_kfunc(meta))
8763 		return arg_idx == 0;
8764 
8765 	/* iter passed as an argument to a generic kfunc */
8766 	return btf_param_match_suffix(meta->btf, arg, "__iter");
8767 }
8768 
process_iter_arg(struct bpf_verifier_env * env,int regno,int insn_idx,struct bpf_kfunc_call_arg_meta * meta)8769 static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
8770 			    struct bpf_kfunc_call_arg_meta *meta)
8771 {
8772 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8773 	const struct btf_type *t;
8774 	int spi, err, i, nr_slots, btf_id;
8775 
8776 	if (reg->type != PTR_TO_STACK) {
8777 		verbose(env, "arg#%d expected pointer to an iterator on stack\n", regno - 1);
8778 		return -EINVAL;
8779 	}
8780 
8781 	/* For iter_{new,next,destroy} functions, btf_check_iter_kfuncs()
8782 	 * ensures struct convention, so we wouldn't need to do any BTF
8783 	 * validation here. But given iter state can be passed as a parameter
8784 	 * to any kfunc, if arg has "__iter" suffix, we need to be a bit more
8785 	 * conservative here.
8786 	 */
8787 	btf_id = btf_check_iter_arg(meta->btf, meta->func_proto, regno - 1);
8788 	if (btf_id < 0) {
8789 		verbose(env, "expected valid iter pointer as arg #%d\n", regno - 1);
8790 		return -EINVAL;
8791 	}
8792 	t = btf_type_by_id(meta->btf, btf_id);
8793 	nr_slots = t->size / BPF_REG_SIZE;
8794 
8795 	if (is_iter_new_kfunc(meta)) {
8796 		/* bpf_iter_<type>_new() expects pointer to uninit iter state */
8797 		if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
8798 			verbose(env, "expected uninitialized iter_%s as arg #%d\n",
8799 				iter_type_str(meta->btf, btf_id), regno - 1);
8800 			return -EINVAL;
8801 		}
8802 
8803 		for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
8804 			err = check_mem_access(env, insn_idx, regno,
8805 					       i, BPF_DW, BPF_WRITE, -1, false, false);
8806 			if (err)
8807 				return err;
8808 		}
8809 
8810 		err = mark_stack_slots_iter(env, meta, reg, insn_idx, meta->btf, btf_id, nr_slots);
8811 		if (err)
8812 			return err;
8813 	} else {
8814 		/* iter_next() or iter_destroy(), as well as any kfunc
8815 		 * accepting iter argument, expect initialized iter state
8816 		 */
8817 		err = is_iter_reg_valid_init(env, reg, meta->btf, btf_id, nr_slots);
8818 		switch (err) {
8819 		case 0:
8820 			break;
8821 		case -EINVAL:
8822 			verbose(env, "expected an initialized iter_%s as arg #%d\n",
8823 				iter_type_str(meta->btf, btf_id), regno - 1);
8824 			return err;
8825 		case -EPROTO:
8826 			verbose(env, "expected an RCU CS when using %s\n", meta->func_name);
8827 			return err;
8828 		default:
8829 			return err;
8830 		}
8831 
8832 		spi = iter_get_spi(env, reg, nr_slots);
8833 		if (spi < 0)
8834 			return spi;
8835 
8836 		err = mark_iter_read(env, reg, spi, nr_slots);
8837 		if (err)
8838 			return err;
8839 
8840 		/* remember meta->iter info for process_iter_next_call() */
8841 		meta->iter.spi = spi;
8842 		meta->iter.frameno = reg->frameno;
8843 		meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
8844 
8845 		if (is_iter_destroy_kfunc(meta)) {
8846 			err = unmark_stack_slots_iter(env, reg, nr_slots);
8847 			if (err)
8848 				return err;
8849 		}
8850 	}
8851 
8852 	return 0;
8853 }
8854 
8855 /* Look for a previous loop entry at insn_idx: nearest parent state
8856  * stopped at insn_idx with callsites matching those in cur->frame.
8857  */
find_prev_entry(struct bpf_verifier_env * env,struct bpf_verifier_state * cur,int insn_idx)8858 static struct bpf_verifier_state *find_prev_entry(struct bpf_verifier_env *env,
8859 						  struct bpf_verifier_state *cur,
8860 						  int insn_idx)
8861 {
8862 	struct bpf_verifier_state_list *sl;
8863 	struct bpf_verifier_state *st;
8864 	struct list_head *pos, *head;
8865 
8866 	/* Explored states are pushed in stack order, most recent states come first */
8867 	head = explored_state(env, insn_idx);
8868 	list_for_each(pos, head) {
8869 		sl = container_of(pos, struct bpf_verifier_state_list, node);
8870 		/* If st->branches != 0 state is a part of current DFS verification path,
8871 		 * hence cur & st for a loop.
8872 		 */
8873 		st = &sl->state;
8874 		if (st->insn_idx == insn_idx && st->branches && same_callsites(st, cur) &&
8875 		    st->dfs_depth < cur->dfs_depth)
8876 			return st;
8877 	}
8878 
8879 	return NULL;
8880 }
8881 
8882 static void reset_idmap_scratch(struct bpf_verifier_env *env);
8883 static bool regs_exact(const struct bpf_reg_state *rold,
8884 		       const struct bpf_reg_state *rcur,
8885 		       struct bpf_idmap *idmap);
8886 
maybe_widen_reg(struct bpf_verifier_env * env,struct bpf_reg_state * rold,struct bpf_reg_state * rcur,struct bpf_idmap * idmap)8887 static void maybe_widen_reg(struct bpf_verifier_env *env,
8888 			    struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
8889 			    struct bpf_idmap *idmap)
8890 {
8891 	if (rold->type != SCALAR_VALUE)
8892 		return;
8893 	if (rold->type != rcur->type)
8894 		return;
8895 	if (rold->precise || rcur->precise || regs_exact(rold, rcur, idmap))
8896 		return;
8897 	__mark_reg_unknown(env, rcur);
8898 }
8899 
widen_imprecise_scalars(struct bpf_verifier_env * env,struct bpf_verifier_state * old,struct bpf_verifier_state * cur)8900 static int widen_imprecise_scalars(struct bpf_verifier_env *env,
8901 				   struct bpf_verifier_state *old,
8902 				   struct bpf_verifier_state *cur)
8903 {
8904 	struct bpf_func_state *fold, *fcur;
8905 	int i, fr;
8906 
8907 	reset_idmap_scratch(env);
8908 	for (fr = old->curframe; fr >= 0; fr--) {
8909 		fold = old->frame[fr];
8910 		fcur = cur->frame[fr];
8911 
8912 		for (i = 0; i < MAX_BPF_REG; i++)
8913 			maybe_widen_reg(env,
8914 					&fold->regs[i],
8915 					&fcur->regs[i],
8916 					&env->idmap_scratch);
8917 
8918 		for (i = 0; i < fold->allocated_stack / BPF_REG_SIZE; i++) {
8919 			if (!is_spilled_reg(&fold->stack[i]) ||
8920 			    !is_spilled_reg(&fcur->stack[i]))
8921 				continue;
8922 
8923 			maybe_widen_reg(env,
8924 					&fold->stack[i].spilled_ptr,
8925 					&fcur->stack[i].spilled_ptr,
8926 					&env->idmap_scratch);
8927 		}
8928 	}
8929 	return 0;
8930 }
8931 
get_iter_from_state(struct bpf_verifier_state * cur_st,struct bpf_kfunc_call_arg_meta * meta)8932 static struct bpf_reg_state *get_iter_from_state(struct bpf_verifier_state *cur_st,
8933 						 struct bpf_kfunc_call_arg_meta *meta)
8934 {
8935 	int iter_frameno = meta->iter.frameno;
8936 	int iter_spi = meta->iter.spi;
8937 
8938 	return &cur_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
8939 }
8940 
8941 /* process_iter_next_call() is called when verifier gets to iterator's next
8942  * "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
8943  * to it as just "iter_next()" in comments below.
8944  *
8945  * BPF verifier relies on a crucial contract for any iter_next()
8946  * implementation: it should *eventually* return NULL, and once that happens
8947  * it should keep returning NULL. That is, once iterator exhausts elements to
8948  * iterate, it should never reset or spuriously return new elements.
8949  *
8950  * With the assumption of such contract, process_iter_next_call() simulates
8951  * a fork in the verifier state to validate loop logic correctness and safety
8952  * without having to simulate infinite amount of iterations.
8953  *
8954  * In current state, we first assume that iter_next() returned NULL and
8955  * iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
8956  * conditions we should not form an infinite loop and should eventually reach
8957  * exit.
8958  *
8959  * Besides that, we also fork current state and enqueue it for later
8960  * verification. In a forked state we keep iterator state as ACTIVE
8961  * (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
8962  * also bump iteration depth to prevent erroneous infinite loop detection
8963  * later on (see iter_active_depths_differ() comment for details). In this
8964  * state we assume that we'll eventually loop back to another iter_next()
8965  * calls (it could be in exactly same location or in some other instruction,
8966  * it doesn't matter, we don't make any unnecessary assumptions about this,
8967  * everything revolves around iterator state in a stack slot, not which
8968  * instruction is calling iter_next()). When that happens, we either will come
8969  * to iter_next() with equivalent state and can conclude that next iteration
8970  * will proceed in exactly the same way as we just verified, so it's safe to
8971  * assume that loop converges. If not, we'll go on another iteration
8972  * simulation with a different input state, until all possible starting states
8973  * are validated or we reach maximum number of instructions limit.
8974  *
8975  * This way, we will either exhaustively discover all possible input states
8976  * that iterator loop can start with and eventually will converge, or we'll
8977  * effectively regress into bounded loop simulation logic and either reach
8978  * maximum number of instructions if loop is not provably convergent, or there
8979  * is some statically known limit on number of iterations (e.g., if there is
8980  * an explicit `if n > 100 then break;` statement somewhere in the loop).
8981  *
8982  * Iteration convergence logic in is_state_visited() relies on exact
8983  * states comparison, which ignores read and precision marks.
8984  * This is necessary because read and precision marks are not finalized
8985  * while in the loop. Exact comparison might preclude convergence for
8986  * simple programs like below:
8987  *
8988  *     i = 0;
8989  *     while(iter_next(&it))
8990  *       i++;
8991  *
8992  * At each iteration step i++ would produce a new distinct state and
8993  * eventually instruction processing limit would be reached.
8994  *
8995  * To avoid such behavior speculatively forget (widen) range for
8996  * imprecise scalar registers, if those registers were not precise at the
8997  * end of the previous iteration and do not match exactly.
8998  *
8999  * This is a conservative heuristic that allows to verify wide range of programs,
9000  * however it precludes verification of programs that conjure an
9001  * imprecise value on the first loop iteration and use it as precise on a second.
9002  * For example, the following safe program would fail to verify:
9003  *
9004  *     struct bpf_num_iter it;
9005  *     int arr[10];
9006  *     int i = 0, a = 0;
9007  *     bpf_iter_num_new(&it, 0, 10);
9008  *     while (bpf_iter_num_next(&it)) {
9009  *       if (a == 0) {
9010  *         a = 1;
9011  *         i = 7; // Because i changed verifier would forget
9012  *                // it's range on second loop entry.
9013  *       } else {
9014  *         arr[i] = 42; // This would fail to verify.
9015  *       }
9016  *     }
9017  *     bpf_iter_num_destroy(&it);
9018  */
process_iter_next_call(struct bpf_verifier_env * env,int insn_idx,struct bpf_kfunc_call_arg_meta * meta)9019 static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
9020 				  struct bpf_kfunc_call_arg_meta *meta)
9021 {
9022 	struct bpf_verifier_state *cur_st = env->cur_state, *queued_st, *prev_st;
9023 	struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
9024 	struct bpf_reg_state *cur_iter, *queued_iter;
9025 
9026 	BTF_TYPE_EMIT(struct bpf_iter);
9027 
9028 	cur_iter = get_iter_from_state(cur_st, meta);
9029 
9030 	if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
9031 	    cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
9032 		verifier_bug(env, "unexpected iterator state %d (%s)",
9033 			     cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
9034 		return -EFAULT;
9035 	}
9036 
9037 	if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
9038 		/* Because iter_next() call is a checkpoint is_state_visitied()
9039 		 * should guarantee parent state with same call sites and insn_idx.
9040 		 */
9041 		if (!cur_st->parent || cur_st->parent->insn_idx != insn_idx ||
9042 		    !same_callsites(cur_st->parent, cur_st)) {
9043 			verifier_bug(env, "bad parent state for iter next call");
9044 			return -EFAULT;
9045 		}
9046 		/* Note cur_st->parent in the call below, it is necessary to skip
9047 		 * checkpoint created for cur_st by is_state_visited()
9048 		 * right at this instruction.
9049 		 */
9050 		prev_st = find_prev_entry(env, cur_st->parent, insn_idx);
9051 		/* branch out active iter state */
9052 		queued_st = push_stack(env, insn_idx + 1, insn_idx, false);
9053 		if (!queued_st)
9054 			return -ENOMEM;
9055 
9056 		queued_iter = get_iter_from_state(queued_st, meta);
9057 		queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
9058 		queued_iter->iter.depth++;
9059 		if (prev_st)
9060 			widen_imprecise_scalars(env, prev_st, queued_st);
9061 
9062 		queued_fr = queued_st->frame[queued_st->curframe];
9063 		mark_ptr_not_null_reg(&queued_fr->regs[BPF_REG_0]);
9064 	}
9065 
9066 	/* switch to DRAINED state, but keep the depth unchanged */
9067 	/* mark current iter state as drained and assume returned NULL */
9068 	cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
9069 	__mark_reg_const_zero(env, &cur_fr->regs[BPF_REG_0]);
9070 
9071 	return 0;
9072 }
9073 
arg_type_is_mem_size(enum bpf_arg_type type)9074 static bool arg_type_is_mem_size(enum bpf_arg_type type)
9075 {
9076 	return type == ARG_CONST_SIZE ||
9077 	       type == ARG_CONST_SIZE_OR_ZERO;
9078 }
9079 
arg_type_is_raw_mem(enum bpf_arg_type type)9080 static bool arg_type_is_raw_mem(enum bpf_arg_type type)
9081 {
9082 	return base_type(type) == ARG_PTR_TO_MEM &&
9083 	       type & MEM_UNINIT;
9084 }
9085 
arg_type_is_release(enum bpf_arg_type type)9086 static bool arg_type_is_release(enum bpf_arg_type type)
9087 {
9088 	return type & OBJ_RELEASE;
9089 }
9090 
arg_type_is_dynptr(enum bpf_arg_type type)9091 static bool arg_type_is_dynptr(enum bpf_arg_type type)
9092 {
9093 	return base_type(type) == ARG_PTR_TO_DYNPTR;
9094 }
9095 
resolve_map_arg_type(struct bpf_verifier_env * env,const struct bpf_call_arg_meta * meta,enum bpf_arg_type * arg_type)9096 static int resolve_map_arg_type(struct bpf_verifier_env *env,
9097 				 const struct bpf_call_arg_meta *meta,
9098 				 enum bpf_arg_type *arg_type)
9099 {
9100 	if (!meta->map_ptr) {
9101 		/* kernel subsystem misconfigured verifier */
9102 		verifier_bug(env, "invalid map_ptr to access map->type");
9103 		return -EFAULT;
9104 	}
9105 
9106 	switch (meta->map_ptr->map_type) {
9107 	case BPF_MAP_TYPE_SOCKMAP:
9108 	case BPF_MAP_TYPE_SOCKHASH:
9109 		if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
9110 			*arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
9111 		} else {
9112 			verbose(env, "invalid arg_type for sockmap/sockhash\n");
9113 			return -EINVAL;
9114 		}
9115 		break;
9116 	case BPF_MAP_TYPE_BLOOM_FILTER:
9117 		if (meta->func_id == BPF_FUNC_map_peek_elem)
9118 			*arg_type = ARG_PTR_TO_MAP_VALUE;
9119 		break;
9120 	default:
9121 		break;
9122 	}
9123 	return 0;
9124 }
9125 
9126 struct bpf_reg_types {
9127 	const enum bpf_reg_type types[10];
9128 	u32 *btf_id;
9129 };
9130 
9131 static const struct bpf_reg_types sock_types = {
9132 	.types = {
9133 		PTR_TO_SOCK_COMMON,
9134 		PTR_TO_SOCKET,
9135 		PTR_TO_TCP_SOCK,
9136 		PTR_TO_XDP_SOCK,
9137 	},
9138 };
9139 
9140 #ifdef CONFIG_NET
9141 static const struct bpf_reg_types btf_id_sock_common_types = {
9142 	.types = {
9143 		PTR_TO_SOCK_COMMON,
9144 		PTR_TO_SOCKET,
9145 		PTR_TO_TCP_SOCK,
9146 		PTR_TO_XDP_SOCK,
9147 		PTR_TO_BTF_ID,
9148 		PTR_TO_BTF_ID | PTR_TRUSTED,
9149 	},
9150 	.btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
9151 };
9152 #endif
9153 
9154 static const struct bpf_reg_types mem_types = {
9155 	.types = {
9156 		PTR_TO_STACK,
9157 		PTR_TO_PACKET,
9158 		PTR_TO_PACKET_META,
9159 		PTR_TO_MAP_KEY,
9160 		PTR_TO_MAP_VALUE,
9161 		PTR_TO_MEM,
9162 		PTR_TO_MEM | MEM_RINGBUF,
9163 		PTR_TO_BUF,
9164 		PTR_TO_BTF_ID | PTR_TRUSTED,
9165 	},
9166 };
9167 
9168 static const struct bpf_reg_types spin_lock_types = {
9169 	.types = {
9170 		PTR_TO_MAP_VALUE,
9171 		PTR_TO_BTF_ID | MEM_ALLOC,
9172 	}
9173 };
9174 
9175 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
9176 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
9177 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
9178 static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
9179 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
9180 static const struct bpf_reg_types btf_ptr_types = {
9181 	.types = {
9182 		PTR_TO_BTF_ID,
9183 		PTR_TO_BTF_ID | PTR_TRUSTED,
9184 		PTR_TO_BTF_ID | MEM_RCU,
9185 	},
9186 };
9187 static const struct bpf_reg_types percpu_btf_ptr_types = {
9188 	.types = {
9189 		PTR_TO_BTF_ID | MEM_PERCPU,
9190 		PTR_TO_BTF_ID | MEM_PERCPU | MEM_RCU,
9191 		PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
9192 	}
9193 };
9194 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
9195 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
9196 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
9197 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
9198 static const struct bpf_reg_types kptr_xchg_dest_types = {
9199 	.types = {
9200 		PTR_TO_MAP_VALUE,
9201 		PTR_TO_BTF_ID | MEM_ALLOC
9202 	}
9203 };
9204 static const struct bpf_reg_types dynptr_types = {
9205 	.types = {
9206 		PTR_TO_STACK,
9207 		CONST_PTR_TO_DYNPTR,
9208 	}
9209 };
9210 
9211 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
9212 	[ARG_PTR_TO_MAP_KEY]		= &mem_types,
9213 	[ARG_PTR_TO_MAP_VALUE]		= &mem_types,
9214 	[ARG_CONST_SIZE]		= &scalar_types,
9215 	[ARG_CONST_SIZE_OR_ZERO]	= &scalar_types,
9216 	[ARG_CONST_ALLOC_SIZE_OR_ZERO]	= &scalar_types,
9217 	[ARG_CONST_MAP_PTR]		= &const_map_ptr_types,
9218 	[ARG_PTR_TO_CTX]		= &context_types,
9219 	[ARG_PTR_TO_SOCK_COMMON]	= &sock_types,
9220 #ifdef CONFIG_NET
9221 	[ARG_PTR_TO_BTF_ID_SOCK_COMMON]	= &btf_id_sock_common_types,
9222 #endif
9223 	[ARG_PTR_TO_SOCKET]		= &fullsock_types,
9224 	[ARG_PTR_TO_BTF_ID]		= &btf_ptr_types,
9225 	[ARG_PTR_TO_SPIN_LOCK]		= &spin_lock_types,
9226 	[ARG_PTR_TO_MEM]		= &mem_types,
9227 	[ARG_PTR_TO_RINGBUF_MEM]	= &ringbuf_mem_types,
9228 	[ARG_PTR_TO_PERCPU_BTF_ID]	= &percpu_btf_ptr_types,
9229 	[ARG_PTR_TO_FUNC]		= &func_ptr_types,
9230 	[ARG_PTR_TO_STACK]		= &stack_ptr_types,
9231 	[ARG_PTR_TO_CONST_STR]		= &const_str_ptr_types,
9232 	[ARG_PTR_TO_TIMER]		= &timer_types,
9233 	[ARG_KPTR_XCHG_DEST]		= &kptr_xchg_dest_types,
9234 	[ARG_PTR_TO_DYNPTR]		= &dynptr_types,
9235 };
9236 
check_reg_type(struct bpf_verifier_env * env,u32 regno,enum bpf_arg_type arg_type,const u32 * arg_btf_id,struct bpf_call_arg_meta * meta)9237 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
9238 			  enum bpf_arg_type arg_type,
9239 			  const u32 *arg_btf_id,
9240 			  struct bpf_call_arg_meta *meta)
9241 {
9242 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
9243 	enum bpf_reg_type expected, type = reg->type;
9244 	const struct bpf_reg_types *compatible;
9245 	int i, j;
9246 
9247 	compatible = compatible_reg_types[base_type(arg_type)];
9248 	if (!compatible) {
9249 		verifier_bug(env, "unsupported arg type %d", arg_type);
9250 		return -EFAULT;
9251 	}
9252 
9253 	/* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
9254 	 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
9255 	 *
9256 	 * Same for MAYBE_NULL:
9257 	 *
9258 	 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
9259 	 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
9260 	 *
9261 	 * ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
9262 	 *
9263 	 * Therefore we fold these flags depending on the arg_type before comparison.
9264 	 */
9265 	if (arg_type & MEM_RDONLY)
9266 		type &= ~MEM_RDONLY;
9267 	if (arg_type & PTR_MAYBE_NULL)
9268 		type &= ~PTR_MAYBE_NULL;
9269 	if (base_type(arg_type) == ARG_PTR_TO_MEM)
9270 		type &= ~DYNPTR_TYPE_FLAG_MASK;
9271 
9272 	/* Local kptr types are allowed as the source argument of bpf_kptr_xchg */
9273 	if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type) && regno == BPF_REG_2) {
9274 		type &= ~MEM_ALLOC;
9275 		type &= ~MEM_PERCPU;
9276 	}
9277 
9278 	for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
9279 		expected = compatible->types[i];
9280 		if (expected == NOT_INIT)
9281 			break;
9282 
9283 		if (type == expected)
9284 			goto found;
9285 	}
9286 
9287 	verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
9288 	for (j = 0; j + 1 < i; j++)
9289 		verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
9290 	verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
9291 	return -EACCES;
9292 
9293 found:
9294 	if (base_type(reg->type) != PTR_TO_BTF_ID)
9295 		return 0;
9296 
9297 	if (compatible == &mem_types) {
9298 		if (!(arg_type & MEM_RDONLY)) {
9299 			verbose(env,
9300 				"%s() may write into memory pointed by R%d type=%s\n",
9301 				func_id_name(meta->func_id),
9302 				regno, reg_type_str(env, reg->type));
9303 			return -EACCES;
9304 		}
9305 		return 0;
9306 	}
9307 
9308 	switch ((int)reg->type) {
9309 	case PTR_TO_BTF_ID:
9310 	case PTR_TO_BTF_ID | PTR_TRUSTED:
9311 	case PTR_TO_BTF_ID | PTR_TRUSTED | PTR_MAYBE_NULL:
9312 	case PTR_TO_BTF_ID | MEM_RCU:
9313 	case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
9314 	case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
9315 	{
9316 		/* For bpf_sk_release, it needs to match against first member
9317 		 * 'struct sock_common', hence make an exception for it. This
9318 		 * allows bpf_sk_release to work for multiple socket types.
9319 		 */
9320 		bool strict_type_match = arg_type_is_release(arg_type) &&
9321 					 meta->func_id != BPF_FUNC_sk_release;
9322 
9323 		if (type_may_be_null(reg->type) &&
9324 		    (!type_may_be_null(arg_type) || arg_type_is_release(arg_type))) {
9325 			verbose(env, "Possibly NULL pointer passed to helper arg%d\n", regno);
9326 			return -EACCES;
9327 		}
9328 
9329 		if (!arg_btf_id) {
9330 			if (!compatible->btf_id) {
9331 				verifier_bug(env, "missing arg compatible BTF ID");
9332 				return -EFAULT;
9333 			}
9334 			arg_btf_id = compatible->btf_id;
9335 		}
9336 
9337 		if (meta->func_id == BPF_FUNC_kptr_xchg) {
9338 			if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
9339 				return -EACCES;
9340 		} else {
9341 			if (arg_btf_id == BPF_PTR_POISON) {
9342 				verbose(env, "verifier internal error:");
9343 				verbose(env, "R%d has non-overwritten BPF_PTR_POISON type\n",
9344 					regno);
9345 				return -EACCES;
9346 			}
9347 
9348 			if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
9349 						  btf_vmlinux, *arg_btf_id,
9350 						  strict_type_match)) {
9351 				verbose(env, "R%d is of type %s but %s is expected\n",
9352 					regno, btf_type_name(reg->btf, reg->btf_id),
9353 					btf_type_name(btf_vmlinux, *arg_btf_id));
9354 				return -EACCES;
9355 			}
9356 		}
9357 		break;
9358 	}
9359 	case PTR_TO_BTF_ID | MEM_ALLOC:
9360 	case PTR_TO_BTF_ID | MEM_PERCPU | MEM_ALLOC:
9361 		if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
9362 		    meta->func_id != BPF_FUNC_kptr_xchg) {
9363 			verifier_bug(env, "unimplemented handling of MEM_ALLOC");
9364 			return -EFAULT;
9365 		}
9366 		/* Check if local kptr in src arg matches kptr in dst arg */
9367 		if (meta->func_id == BPF_FUNC_kptr_xchg && regno == BPF_REG_2) {
9368 			if (map_kptr_match_type(env, meta->kptr_field, reg, regno))
9369 				return -EACCES;
9370 		}
9371 		break;
9372 	case PTR_TO_BTF_ID | MEM_PERCPU:
9373 	case PTR_TO_BTF_ID | MEM_PERCPU | MEM_RCU:
9374 	case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
9375 		/* Handled by helper specific checks */
9376 		break;
9377 	default:
9378 		verifier_bug(env, "invalid PTR_TO_BTF_ID register for type match");
9379 		return -EFAULT;
9380 	}
9381 	return 0;
9382 }
9383 
9384 static struct btf_field *
reg_find_field_offset(const struct bpf_reg_state * reg,s32 off,u32 fields)9385 reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
9386 {
9387 	struct btf_field *field;
9388 	struct btf_record *rec;
9389 
9390 	rec = reg_btf_record(reg);
9391 	if (!rec)
9392 		return NULL;
9393 
9394 	field = btf_record_find(rec, off, fields);
9395 	if (!field)
9396 		return NULL;
9397 
9398 	return field;
9399 }
9400 
check_func_arg_reg_off(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,int regno,enum bpf_arg_type arg_type)9401 static int check_func_arg_reg_off(struct bpf_verifier_env *env,
9402 				  const struct bpf_reg_state *reg, int regno,
9403 				  enum bpf_arg_type arg_type)
9404 {
9405 	u32 type = reg->type;
9406 
9407 	/* When referenced register is passed to release function, its fixed
9408 	 * offset must be 0.
9409 	 *
9410 	 * We will check arg_type_is_release reg has ref_obj_id when storing
9411 	 * meta->release_regno.
9412 	 */
9413 	if (arg_type_is_release(arg_type)) {
9414 		/* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
9415 		 * may not directly point to the object being released, but to
9416 		 * dynptr pointing to such object, which might be at some offset
9417 		 * on the stack. In that case, we simply to fallback to the
9418 		 * default handling.
9419 		 */
9420 		if (arg_type_is_dynptr(arg_type) && type == PTR_TO_STACK)
9421 			return 0;
9422 
9423 		/* Doing check_ptr_off_reg check for the offset will catch this
9424 		 * because fixed_off_ok is false, but checking here allows us
9425 		 * to give the user a better error message.
9426 		 */
9427 		if (reg->off) {
9428 			verbose(env, "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
9429 				regno);
9430 			return -EINVAL;
9431 		}
9432 		return __check_ptr_off_reg(env, reg, regno, false);
9433 	}
9434 
9435 	switch (type) {
9436 	/* Pointer types where both fixed and variable offset is explicitly allowed: */
9437 	case PTR_TO_STACK:
9438 	case PTR_TO_PACKET:
9439 	case PTR_TO_PACKET_META:
9440 	case PTR_TO_MAP_KEY:
9441 	case PTR_TO_MAP_VALUE:
9442 	case PTR_TO_MEM:
9443 	case PTR_TO_MEM | MEM_RDONLY:
9444 	case PTR_TO_MEM | MEM_RINGBUF:
9445 	case PTR_TO_BUF:
9446 	case PTR_TO_BUF | MEM_RDONLY:
9447 	case PTR_TO_ARENA:
9448 	case SCALAR_VALUE:
9449 		return 0;
9450 	/* All the rest must be rejected, except PTR_TO_BTF_ID which allows
9451 	 * fixed offset.
9452 	 */
9453 	case PTR_TO_BTF_ID:
9454 	case PTR_TO_BTF_ID | MEM_ALLOC:
9455 	case PTR_TO_BTF_ID | PTR_TRUSTED:
9456 	case PTR_TO_BTF_ID | MEM_RCU:
9457 	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
9458 	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
9459 		/* When referenced PTR_TO_BTF_ID is passed to release function,
9460 		 * its fixed offset must be 0. In the other cases, fixed offset
9461 		 * can be non-zero. This was already checked above. So pass
9462 		 * fixed_off_ok as true to allow fixed offset for all other
9463 		 * cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
9464 		 * still need to do checks instead of returning.
9465 		 */
9466 		return __check_ptr_off_reg(env, reg, regno, true);
9467 	default:
9468 		return __check_ptr_off_reg(env, reg, regno, false);
9469 	}
9470 }
9471 
get_dynptr_arg_reg(struct bpf_verifier_env * env,const struct bpf_func_proto * fn,struct bpf_reg_state * regs)9472 static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
9473 						const struct bpf_func_proto *fn,
9474 						struct bpf_reg_state *regs)
9475 {
9476 	struct bpf_reg_state *state = NULL;
9477 	int i;
9478 
9479 	for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
9480 		if (arg_type_is_dynptr(fn->arg_type[i])) {
9481 			if (state) {
9482 				verbose(env, "verifier internal error: multiple dynptr args\n");
9483 				return NULL;
9484 			}
9485 			state = &regs[BPF_REG_1 + i];
9486 		}
9487 
9488 	if (!state)
9489 		verbose(env, "verifier internal error: no dynptr arg found\n");
9490 
9491 	return state;
9492 }
9493 
dynptr_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg)9494 static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
9495 {
9496 	struct bpf_func_state *state = func(env, reg);
9497 	int spi;
9498 
9499 	if (reg->type == CONST_PTR_TO_DYNPTR)
9500 		return reg->id;
9501 	spi = dynptr_get_spi(env, reg);
9502 	if (spi < 0)
9503 		return spi;
9504 	return state->stack[spi].spilled_ptr.id;
9505 }
9506 
dynptr_ref_obj_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg)9507 static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
9508 {
9509 	struct bpf_func_state *state = func(env, reg);
9510 	int spi;
9511 
9512 	if (reg->type == CONST_PTR_TO_DYNPTR)
9513 		return reg->ref_obj_id;
9514 	spi = dynptr_get_spi(env, reg);
9515 	if (spi < 0)
9516 		return spi;
9517 	return state->stack[spi].spilled_ptr.ref_obj_id;
9518 }
9519 
dynptr_get_type(struct bpf_verifier_env * env,struct bpf_reg_state * reg)9520 static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
9521 					    struct bpf_reg_state *reg)
9522 {
9523 	struct bpf_func_state *state = func(env, reg);
9524 	int spi;
9525 
9526 	if (reg->type == CONST_PTR_TO_DYNPTR)
9527 		return reg->dynptr.type;
9528 
9529 	spi = __get_spi(reg->off);
9530 	if (spi < 0) {
9531 		verbose(env, "verifier internal error: invalid spi when querying dynptr type\n");
9532 		return BPF_DYNPTR_TYPE_INVALID;
9533 	}
9534 
9535 	return state->stack[spi].spilled_ptr.dynptr.type;
9536 }
9537 
check_reg_const_str(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno)9538 static int check_reg_const_str(struct bpf_verifier_env *env,
9539 			       struct bpf_reg_state *reg, u32 regno)
9540 {
9541 	struct bpf_map *map = reg->map_ptr;
9542 	int err;
9543 	int map_off;
9544 	u64 map_addr;
9545 	char *str_ptr;
9546 
9547 	if (reg->type != PTR_TO_MAP_VALUE)
9548 		return -EINVAL;
9549 
9550 	if (!bpf_map_is_rdonly(map)) {
9551 		verbose(env, "R%d does not point to a readonly map'\n", regno);
9552 		return -EACCES;
9553 	}
9554 
9555 	if (!tnum_is_const(reg->var_off)) {
9556 		verbose(env, "R%d is not a constant address'\n", regno);
9557 		return -EACCES;
9558 	}
9559 
9560 	if (!map->ops->map_direct_value_addr) {
9561 		verbose(env, "no direct value access support for this map type\n");
9562 		return -EACCES;
9563 	}
9564 
9565 	err = check_map_access(env, regno, reg->off,
9566 			       map->value_size - reg->off, false,
9567 			       ACCESS_HELPER);
9568 	if (err)
9569 		return err;
9570 
9571 	map_off = reg->off + reg->var_off.value;
9572 	err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
9573 	if (err) {
9574 		verbose(env, "direct value access on string failed\n");
9575 		return err;
9576 	}
9577 
9578 	str_ptr = (char *)(long)(map_addr);
9579 	if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
9580 		verbose(env, "string is not zero-terminated\n");
9581 		return -EINVAL;
9582 	}
9583 	return 0;
9584 }
9585 
9586 /* Returns constant key value in `value` if possible, else negative error */
get_constant_map_key(struct bpf_verifier_env * env,struct bpf_reg_state * key,u32 key_size,s64 * value)9587 static int get_constant_map_key(struct bpf_verifier_env *env,
9588 				struct bpf_reg_state *key,
9589 				u32 key_size,
9590 				s64 *value)
9591 {
9592 	struct bpf_func_state *state = func(env, key);
9593 	struct bpf_reg_state *reg;
9594 	int slot, spi, off;
9595 	int spill_size = 0;
9596 	int zero_size = 0;
9597 	int stack_off;
9598 	int i, err;
9599 	u8 *stype;
9600 
9601 	if (!env->bpf_capable)
9602 		return -EOPNOTSUPP;
9603 	if (key->type != PTR_TO_STACK)
9604 		return -EOPNOTSUPP;
9605 	if (!tnum_is_const(key->var_off))
9606 		return -EOPNOTSUPP;
9607 
9608 	stack_off = key->off + key->var_off.value;
9609 	slot = -stack_off - 1;
9610 	spi = slot / BPF_REG_SIZE;
9611 	off = slot % BPF_REG_SIZE;
9612 	stype = state->stack[spi].slot_type;
9613 
9614 	/* First handle precisely tracked STACK_ZERO */
9615 	for (i = off; i >= 0 && stype[i] == STACK_ZERO; i--)
9616 		zero_size++;
9617 	if (zero_size >= key_size) {
9618 		*value = 0;
9619 		return 0;
9620 	}
9621 
9622 	/* Check that stack contains a scalar spill of expected size */
9623 	if (!is_spilled_scalar_reg(&state->stack[spi]))
9624 		return -EOPNOTSUPP;
9625 	for (i = off; i >= 0 && stype[i] == STACK_SPILL; i--)
9626 		spill_size++;
9627 	if (spill_size != key_size)
9628 		return -EOPNOTSUPP;
9629 
9630 	reg = &state->stack[spi].spilled_ptr;
9631 	if (!tnum_is_const(reg->var_off))
9632 		/* Stack value not statically known */
9633 		return -EOPNOTSUPP;
9634 
9635 	/* We are relying on a constant value. So mark as precise
9636 	 * to prevent pruning on it.
9637 	 */
9638 	bt_set_frame_slot(&env->bt, key->frameno, spi);
9639 	err = mark_chain_precision_batch(env, env->cur_state);
9640 	if (err < 0)
9641 		return err;
9642 
9643 	*value = reg->var_off.value;
9644 	return 0;
9645 }
9646 
9647 static bool can_elide_value_nullness(enum bpf_map_type type);
9648 
check_func_arg(struct bpf_verifier_env * env,u32 arg,struct bpf_call_arg_meta * meta,const struct bpf_func_proto * fn,int insn_idx)9649 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
9650 			  struct bpf_call_arg_meta *meta,
9651 			  const struct bpf_func_proto *fn,
9652 			  int insn_idx)
9653 {
9654 	u32 regno = BPF_REG_1 + arg;
9655 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
9656 	enum bpf_arg_type arg_type = fn->arg_type[arg];
9657 	enum bpf_reg_type type = reg->type;
9658 	u32 *arg_btf_id = NULL;
9659 	u32 key_size;
9660 	int err = 0;
9661 
9662 	if (arg_type == ARG_DONTCARE)
9663 		return 0;
9664 
9665 	err = check_reg_arg(env, regno, SRC_OP);
9666 	if (err)
9667 		return err;
9668 
9669 	if (arg_type == ARG_ANYTHING) {
9670 		if (is_pointer_value(env, regno)) {
9671 			verbose(env, "R%d leaks addr into helper function\n",
9672 				regno);
9673 			return -EACCES;
9674 		}
9675 		return 0;
9676 	}
9677 
9678 	if (type_is_pkt_pointer(type) &&
9679 	    !may_access_direct_pkt_data(env, meta, BPF_READ)) {
9680 		verbose(env, "helper access to the packet is not allowed\n");
9681 		return -EACCES;
9682 	}
9683 
9684 	if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE) {
9685 		err = resolve_map_arg_type(env, meta, &arg_type);
9686 		if (err)
9687 			return err;
9688 	}
9689 
9690 	if (register_is_null(reg) && type_may_be_null(arg_type))
9691 		/* A NULL register has a SCALAR_VALUE type, so skip
9692 		 * type checking.
9693 		 */
9694 		goto skip_type_check;
9695 
9696 	/* arg_btf_id and arg_size are in a union. */
9697 	if (base_type(arg_type) == ARG_PTR_TO_BTF_ID ||
9698 	    base_type(arg_type) == ARG_PTR_TO_SPIN_LOCK)
9699 		arg_btf_id = fn->arg_btf_id[arg];
9700 
9701 	err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
9702 	if (err)
9703 		return err;
9704 
9705 	err = check_func_arg_reg_off(env, reg, regno, arg_type);
9706 	if (err)
9707 		return err;
9708 
9709 skip_type_check:
9710 	if (arg_type_is_release(arg_type)) {
9711 		if (arg_type_is_dynptr(arg_type)) {
9712 			struct bpf_func_state *state = func(env, reg);
9713 			int spi;
9714 
9715 			/* Only dynptr created on stack can be released, thus
9716 			 * the get_spi and stack state checks for spilled_ptr
9717 			 * should only be done before process_dynptr_func for
9718 			 * PTR_TO_STACK.
9719 			 */
9720 			if (reg->type == PTR_TO_STACK) {
9721 				spi = dynptr_get_spi(env, reg);
9722 				if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
9723 					verbose(env, "arg %d is an unacquired reference\n", regno);
9724 					return -EINVAL;
9725 				}
9726 			} else {
9727 				verbose(env, "cannot release unowned const bpf_dynptr\n");
9728 				return -EINVAL;
9729 			}
9730 		} else if (!reg->ref_obj_id && !register_is_null(reg)) {
9731 			verbose(env, "R%d must be referenced when passed to release function\n",
9732 				regno);
9733 			return -EINVAL;
9734 		}
9735 		if (meta->release_regno) {
9736 			verifier_bug(env, "more than one release argument");
9737 			return -EFAULT;
9738 		}
9739 		meta->release_regno = regno;
9740 	}
9741 
9742 	if (reg->ref_obj_id && base_type(arg_type) != ARG_KPTR_XCHG_DEST) {
9743 		if (meta->ref_obj_id) {
9744 			verbose(env, "more than one arg with ref_obj_id R%d %u %u",
9745 				regno, reg->ref_obj_id,
9746 				meta->ref_obj_id);
9747 			return -EACCES;
9748 		}
9749 		meta->ref_obj_id = reg->ref_obj_id;
9750 	}
9751 
9752 	switch (base_type(arg_type)) {
9753 	case ARG_CONST_MAP_PTR:
9754 		/* bpf_map_xxx(map_ptr) call: remember that map_ptr */
9755 		if (meta->map_ptr) {
9756 			/* Use map_uid (which is unique id of inner map) to reject:
9757 			 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
9758 			 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
9759 			 * if (inner_map1 && inner_map2) {
9760 			 *     timer = bpf_map_lookup_elem(inner_map1);
9761 			 *     if (timer)
9762 			 *         // mismatch would have been allowed
9763 			 *         bpf_timer_init(timer, inner_map2);
9764 			 * }
9765 			 *
9766 			 * Comparing map_ptr is enough to distinguish normal and outer maps.
9767 			 */
9768 			if (meta->map_ptr != reg->map_ptr ||
9769 			    meta->map_uid != reg->map_uid) {
9770 				verbose(env,
9771 					"timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
9772 					meta->map_uid, reg->map_uid);
9773 				return -EINVAL;
9774 			}
9775 		}
9776 		meta->map_ptr = reg->map_ptr;
9777 		meta->map_uid = reg->map_uid;
9778 		break;
9779 	case ARG_PTR_TO_MAP_KEY:
9780 		/* bpf_map_xxx(..., map_ptr, ..., key) call:
9781 		 * check that [key, key + map->key_size) are within
9782 		 * stack limits and initialized
9783 		 */
9784 		if (!meta->map_ptr) {
9785 			/* in function declaration map_ptr must come before
9786 			 * map_key, so that it's verified and known before
9787 			 * we have to check map_key here. Otherwise it means
9788 			 * that kernel subsystem misconfigured verifier
9789 			 */
9790 			verifier_bug(env, "invalid map_ptr to access map->key");
9791 			return -EFAULT;
9792 		}
9793 		key_size = meta->map_ptr->key_size;
9794 		err = check_helper_mem_access(env, regno, key_size, BPF_READ, false, NULL);
9795 		if (err)
9796 			return err;
9797 		if (can_elide_value_nullness(meta->map_ptr->map_type)) {
9798 			err = get_constant_map_key(env, reg, key_size, &meta->const_map_key);
9799 			if (err < 0) {
9800 				meta->const_map_key = -1;
9801 				if (err == -EOPNOTSUPP)
9802 					err = 0;
9803 				else
9804 					return err;
9805 			}
9806 		}
9807 		break;
9808 	case ARG_PTR_TO_MAP_VALUE:
9809 		if (type_may_be_null(arg_type) && register_is_null(reg))
9810 			return 0;
9811 
9812 		/* bpf_map_xxx(..., map_ptr, ..., value) call:
9813 		 * check [value, value + map->value_size) validity
9814 		 */
9815 		if (!meta->map_ptr) {
9816 			/* kernel subsystem misconfigured verifier */
9817 			verifier_bug(env, "invalid map_ptr to access map->value");
9818 			return -EFAULT;
9819 		}
9820 		meta->raw_mode = arg_type & MEM_UNINIT;
9821 		err = check_helper_mem_access(env, regno, meta->map_ptr->value_size,
9822 					      arg_type & MEM_WRITE ? BPF_WRITE : BPF_READ,
9823 					      false, meta);
9824 		break;
9825 	case ARG_PTR_TO_PERCPU_BTF_ID:
9826 		if (!reg->btf_id) {
9827 			verbose(env, "Helper has invalid btf_id in R%d\n", regno);
9828 			return -EACCES;
9829 		}
9830 		meta->ret_btf = reg->btf;
9831 		meta->ret_btf_id = reg->btf_id;
9832 		break;
9833 	case ARG_PTR_TO_SPIN_LOCK:
9834 		if (in_rbtree_lock_required_cb(env)) {
9835 			verbose(env, "can't spin_{lock,unlock} in rbtree cb\n");
9836 			return -EACCES;
9837 		}
9838 		if (meta->func_id == BPF_FUNC_spin_lock) {
9839 			err = process_spin_lock(env, regno, PROCESS_SPIN_LOCK);
9840 			if (err)
9841 				return err;
9842 		} else if (meta->func_id == BPF_FUNC_spin_unlock) {
9843 			err = process_spin_lock(env, regno, 0);
9844 			if (err)
9845 				return err;
9846 		} else {
9847 			verifier_bug(env, "spin lock arg on unexpected helper");
9848 			return -EFAULT;
9849 		}
9850 		break;
9851 	case ARG_PTR_TO_TIMER:
9852 		err = process_timer_func(env, regno, meta);
9853 		if (err)
9854 			return err;
9855 		break;
9856 	case ARG_PTR_TO_FUNC:
9857 		meta->subprogno = reg->subprogno;
9858 		break;
9859 	case ARG_PTR_TO_MEM:
9860 		/* The access to this pointer is only checked when we hit the
9861 		 * next is_mem_size argument below.
9862 		 */
9863 		meta->raw_mode = arg_type & MEM_UNINIT;
9864 		if (arg_type & MEM_FIXED_SIZE) {
9865 			err = check_helper_mem_access(env, regno, fn->arg_size[arg],
9866 						      arg_type & MEM_WRITE ? BPF_WRITE : BPF_READ,
9867 						      false, meta);
9868 			if (err)
9869 				return err;
9870 			if (arg_type & MEM_ALIGNED)
9871 				err = check_ptr_alignment(env, reg, 0, fn->arg_size[arg], true);
9872 		}
9873 		break;
9874 	case ARG_CONST_SIZE:
9875 		err = check_mem_size_reg(env, reg, regno,
9876 					 fn->arg_type[arg - 1] & MEM_WRITE ?
9877 					 BPF_WRITE : BPF_READ,
9878 					 false, meta);
9879 		break;
9880 	case ARG_CONST_SIZE_OR_ZERO:
9881 		err = check_mem_size_reg(env, reg, regno,
9882 					 fn->arg_type[arg - 1] & MEM_WRITE ?
9883 					 BPF_WRITE : BPF_READ,
9884 					 true, meta);
9885 		break;
9886 	case ARG_PTR_TO_DYNPTR:
9887 		err = process_dynptr_func(env, regno, insn_idx, arg_type, 0);
9888 		if (err)
9889 			return err;
9890 		break;
9891 	case ARG_CONST_ALLOC_SIZE_OR_ZERO:
9892 		if (!tnum_is_const(reg->var_off)) {
9893 			verbose(env, "R%d is not a known constant'\n",
9894 				regno);
9895 			return -EACCES;
9896 		}
9897 		meta->mem_size = reg->var_off.value;
9898 		err = mark_chain_precision(env, regno);
9899 		if (err)
9900 			return err;
9901 		break;
9902 	case ARG_PTR_TO_CONST_STR:
9903 	{
9904 		err = check_reg_const_str(env, reg, regno);
9905 		if (err)
9906 			return err;
9907 		break;
9908 	}
9909 	case ARG_KPTR_XCHG_DEST:
9910 		err = process_kptr_func(env, regno, meta);
9911 		if (err)
9912 			return err;
9913 		break;
9914 	}
9915 
9916 	return err;
9917 }
9918 
may_update_sockmap(struct bpf_verifier_env * env,int func_id)9919 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
9920 {
9921 	enum bpf_attach_type eatype = env->prog->expected_attach_type;
9922 	enum bpf_prog_type type = resolve_prog_type(env->prog);
9923 
9924 	if (func_id != BPF_FUNC_map_update_elem &&
9925 	    func_id != BPF_FUNC_map_delete_elem)
9926 		return false;
9927 
9928 	/* It's not possible to get access to a locked struct sock in these
9929 	 * contexts, so updating is safe.
9930 	 */
9931 	switch (type) {
9932 	case BPF_PROG_TYPE_TRACING:
9933 		if (eatype == BPF_TRACE_ITER)
9934 			return true;
9935 		break;
9936 	case BPF_PROG_TYPE_SOCK_OPS:
9937 		/* map_update allowed only via dedicated helpers with event type checks */
9938 		if (func_id == BPF_FUNC_map_delete_elem)
9939 			return true;
9940 		break;
9941 	case BPF_PROG_TYPE_SOCKET_FILTER:
9942 	case BPF_PROG_TYPE_SCHED_CLS:
9943 	case BPF_PROG_TYPE_SCHED_ACT:
9944 	case BPF_PROG_TYPE_XDP:
9945 	case BPF_PROG_TYPE_SK_REUSEPORT:
9946 	case BPF_PROG_TYPE_FLOW_DISSECTOR:
9947 	case BPF_PROG_TYPE_SK_LOOKUP:
9948 		return true;
9949 	default:
9950 		break;
9951 	}
9952 
9953 	verbose(env, "cannot update sockmap in this context\n");
9954 	return false;
9955 }
9956 
allow_tail_call_in_subprogs(struct bpf_verifier_env * env)9957 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
9958 {
9959 	return env->prog->jit_requested &&
9960 	       bpf_jit_supports_subprog_tailcalls();
9961 }
9962 
check_map_func_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,int func_id)9963 static int check_map_func_compatibility(struct bpf_verifier_env *env,
9964 					struct bpf_map *map, int func_id)
9965 {
9966 	if (!map)
9967 		return 0;
9968 
9969 	/* We need a two way check, first is from map perspective ... */
9970 	switch (map->map_type) {
9971 	case BPF_MAP_TYPE_PROG_ARRAY:
9972 		if (func_id != BPF_FUNC_tail_call)
9973 			goto error;
9974 		break;
9975 	case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
9976 		if (func_id != BPF_FUNC_perf_event_read &&
9977 		    func_id != BPF_FUNC_perf_event_output &&
9978 		    func_id != BPF_FUNC_skb_output &&
9979 		    func_id != BPF_FUNC_perf_event_read_value &&
9980 		    func_id != BPF_FUNC_xdp_output)
9981 			goto error;
9982 		break;
9983 	case BPF_MAP_TYPE_RINGBUF:
9984 		if (func_id != BPF_FUNC_ringbuf_output &&
9985 		    func_id != BPF_FUNC_ringbuf_reserve &&
9986 		    func_id != BPF_FUNC_ringbuf_query &&
9987 		    func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
9988 		    func_id != BPF_FUNC_ringbuf_submit_dynptr &&
9989 		    func_id != BPF_FUNC_ringbuf_discard_dynptr)
9990 			goto error;
9991 		break;
9992 	case BPF_MAP_TYPE_USER_RINGBUF:
9993 		if (func_id != BPF_FUNC_user_ringbuf_drain)
9994 			goto error;
9995 		break;
9996 	case BPF_MAP_TYPE_STACK_TRACE:
9997 		if (func_id != BPF_FUNC_get_stackid)
9998 			goto error;
9999 		break;
10000 	case BPF_MAP_TYPE_CGROUP_ARRAY:
10001 		if (func_id != BPF_FUNC_skb_under_cgroup &&
10002 		    func_id != BPF_FUNC_current_task_under_cgroup)
10003 			goto error;
10004 		break;
10005 	case BPF_MAP_TYPE_CGROUP_STORAGE:
10006 	case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
10007 		if (func_id != BPF_FUNC_get_local_storage)
10008 			goto error;
10009 		break;
10010 	case BPF_MAP_TYPE_DEVMAP:
10011 	case BPF_MAP_TYPE_DEVMAP_HASH:
10012 		if (func_id != BPF_FUNC_redirect_map &&
10013 		    func_id != BPF_FUNC_map_lookup_elem)
10014 			goto error;
10015 		break;
10016 	/* Restrict bpf side of cpumap and xskmap, open when use-cases
10017 	 * appear.
10018 	 */
10019 	case BPF_MAP_TYPE_CPUMAP:
10020 		if (func_id != BPF_FUNC_redirect_map)
10021 			goto error;
10022 		break;
10023 	case BPF_MAP_TYPE_XSKMAP:
10024 		if (func_id != BPF_FUNC_redirect_map &&
10025 		    func_id != BPF_FUNC_map_lookup_elem)
10026 			goto error;
10027 		break;
10028 	case BPF_MAP_TYPE_ARRAY_OF_MAPS:
10029 	case BPF_MAP_TYPE_HASH_OF_MAPS:
10030 		if (func_id != BPF_FUNC_map_lookup_elem)
10031 			goto error;
10032 		break;
10033 	case BPF_MAP_TYPE_SOCKMAP:
10034 		if (func_id != BPF_FUNC_sk_redirect_map &&
10035 		    func_id != BPF_FUNC_sock_map_update &&
10036 		    func_id != BPF_FUNC_msg_redirect_map &&
10037 		    func_id != BPF_FUNC_sk_select_reuseport &&
10038 		    func_id != BPF_FUNC_map_lookup_elem &&
10039 		    !may_update_sockmap(env, func_id))
10040 			goto error;
10041 		break;
10042 	case BPF_MAP_TYPE_SOCKHASH:
10043 		if (func_id != BPF_FUNC_sk_redirect_hash &&
10044 		    func_id != BPF_FUNC_sock_hash_update &&
10045 		    func_id != BPF_FUNC_msg_redirect_hash &&
10046 		    func_id != BPF_FUNC_sk_select_reuseport &&
10047 		    func_id != BPF_FUNC_map_lookup_elem &&
10048 		    !may_update_sockmap(env, func_id))
10049 			goto error;
10050 		break;
10051 	case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
10052 		if (func_id != BPF_FUNC_sk_select_reuseport)
10053 			goto error;
10054 		break;
10055 	case BPF_MAP_TYPE_QUEUE:
10056 	case BPF_MAP_TYPE_STACK:
10057 		if (func_id != BPF_FUNC_map_peek_elem &&
10058 		    func_id != BPF_FUNC_map_pop_elem &&
10059 		    func_id != BPF_FUNC_map_push_elem)
10060 			goto error;
10061 		break;
10062 	case BPF_MAP_TYPE_SK_STORAGE:
10063 		if (func_id != BPF_FUNC_sk_storage_get &&
10064 		    func_id != BPF_FUNC_sk_storage_delete &&
10065 		    func_id != BPF_FUNC_kptr_xchg)
10066 			goto error;
10067 		break;
10068 	case BPF_MAP_TYPE_INODE_STORAGE:
10069 		if (func_id != BPF_FUNC_inode_storage_get &&
10070 		    func_id != BPF_FUNC_inode_storage_delete &&
10071 		    func_id != BPF_FUNC_kptr_xchg)
10072 			goto error;
10073 		break;
10074 	case BPF_MAP_TYPE_TASK_STORAGE:
10075 		if (func_id != BPF_FUNC_task_storage_get &&
10076 		    func_id != BPF_FUNC_task_storage_delete &&
10077 		    func_id != BPF_FUNC_kptr_xchg)
10078 			goto error;
10079 		break;
10080 	case BPF_MAP_TYPE_CGRP_STORAGE:
10081 		if (func_id != BPF_FUNC_cgrp_storage_get &&
10082 		    func_id != BPF_FUNC_cgrp_storage_delete &&
10083 		    func_id != BPF_FUNC_kptr_xchg)
10084 			goto error;
10085 		break;
10086 	case BPF_MAP_TYPE_BLOOM_FILTER:
10087 		if (func_id != BPF_FUNC_map_peek_elem &&
10088 		    func_id != BPF_FUNC_map_push_elem)
10089 			goto error;
10090 		break;
10091 	default:
10092 		break;
10093 	}
10094 
10095 	/* ... and second from the function itself. */
10096 	switch (func_id) {
10097 	case BPF_FUNC_tail_call:
10098 		if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
10099 			goto error;
10100 		if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
10101 			verbose(env, "mixing of tail_calls and bpf-to-bpf calls is not supported\n");
10102 			return -EINVAL;
10103 		}
10104 		break;
10105 	case BPF_FUNC_perf_event_read:
10106 	case BPF_FUNC_perf_event_output:
10107 	case BPF_FUNC_perf_event_read_value:
10108 	case BPF_FUNC_skb_output:
10109 	case BPF_FUNC_xdp_output:
10110 		if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
10111 			goto error;
10112 		break;
10113 	case BPF_FUNC_ringbuf_output:
10114 	case BPF_FUNC_ringbuf_reserve:
10115 	case BPF_FUNC_ringbuf_query:
10116 	case BPF_FUNC_ringbuf_reserve_dynptr:
10117 	case BPF_FUNC_ringbuf_submit_dynptr:
10118 	case BPF_FUNC_ringbuf_discard_dynptr:
10119 		if (map->map_type != BPF_MAP_TYPE_RINGBUF)
10120 			goto error;
10121 		break;
10122 	case BPF_FUNC_user_ringbuf_drain:
10123 		if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
10124 			goto error;
10125 		break;
10126 	case BPF_FUNC_get_stackid:
10127 		if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
10128 			goto error;
10129 		break;
10130 	case BPF_FUNC_current_task_under_cgroup:
10131 	case BPF_FUNC_skb_under_cgroup:
10132 		if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
10133 			goto error;
10134 		break;
10135 	case BPF_FUNC_redirect_map:
10136 		if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
10137 		    map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
10138 		    map->map_type != BPF_MAP_TYPE_CPUMAP &&
10139 		    map->map_type != BPF_MAP_TYPE_XSKMAP)
10140 			goto error;
10141 		break;
10142 	case BPF_FUNC_sk_redirect_map:
10143 	case BPF_FUNC_msg_redirect_map:
10144 	case BPF_FUNC_sock_map_update:
10145 		if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
10146 			goto error;
10147 		break;
10148 	case BPF_FUNC_sk_redirect_hash:
10149 	case BPF_FUNC_msg_redirect_hash:
10150 	case BPF_FUNC_sock_hash_update:
10151 		if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
10152 			goto error;
10153 		break;
10154 	case BPF_FUNC_get_local_storage:
10155 		if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
10156 		    map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
10157 			goto error;
10158 		break;
10159 	case BPF_FUNC_sk_select_reuseport:
10160 		if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
10161 		    map->map_type != BPF_MAP_TYPE_SOCKMAP &&
10162 		    map->map_type != BPF_MAP_TYPE_SOCKHASH)
10163 			goto error;
10164 		break;
10165 	case BPF_FUNC_map_pop_elem:
10166 		if (map->map_type != BPF_MAP_TYPE_QUEUE &&
10167 		    map->map_type != BPF_MAP_TYPE_STACK)
10168 			goto error;
10169 		break;
10170 	case BPF_FUNC_map_peek_elem:
10171 	case BPF_FUNC_map_push_elem:
10172 		if (map->map_type != BPF_MAP_TYPE_QUEUE &&
10173 		    map->map_type != BPF_MAP_TYPE_STACK &&
10174 		    map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
10175 			goto error;
10176 		break;
10177 	case BPF_FUNC_map_lookup_percpu_elem:
10178 		if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
10179 		    map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
10180 		    map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
10181 			goto error;
10182 		break;
10183 	case BPF_FUNC_sk_storage_get:
10184 	case BPF_FUNC_sk_storage_delete:
10185 		if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
10186 			goto error;
10187 		break;
10188 	case BPF_FUNC_inode_storage_get:
10189 	case BPF_FUNC_inode_storage_delete:
10190 		if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
10191 			goto error;
10192 		break;
10193 	case BPF_FUNC_task_storage_get:
10194 	case BPF_FUNC_task_storage_delete:
10195 		if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
10196 			goto error;
10197 		break;
10198 	case BPF_FUNC_cgrp_storage_get:
10199 	case BPF_FUNC_cgrp_storage_delete:
10200 		if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
10201 			goto error;
10202 		break;
10203 	default:
10204 		break;
10205 	}
10206 
10207 	return 0;
10208 error:
10209 	verbose(env, "cannot pass map_type %d into func %s#%d\n",
10210 		map->map_type, func_id_name(func_id), func_id);
10211 	return -EINVAL;
10212 }
10213 
check_raw_mode_ok(const struct bpf_func_proto * fn)10214 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
10215 {
10216 	int count = 0;
10217 
10218 	if (arg_type_is_raw_mem(fn->arg1_type))
10219 		count++;
10220 	if (arg_type_is_raw_mem(fn->arg2_type))
10221 		count++;
10222 	if (arg_type_is_raw_mem(fn->arg3_type))
10223 		count++;
10224 	if (arg_type_is_raw_mem(fn->arg4_type))
10225 		count++;
10226 	if (arg_type_is_raw_mem(fn->arg5_type))
10227 		count++;
10228 
10229 	/* We only support one arg being in raw mode at the moment,
10230 	 * which is sufficient for the helper functions we have
10231 	 * right now.
10232 	 */
10233 	return count <= 1;
10234 }
10235 
check_args_pair_invalid(const struct bpf_func_proto * fn,int arg)10236 static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
10237 {
10238 	bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
10239 	bool has_size = fn->arg_size[arg] != 0;
10240 	bool is_next_size = false;
10241 
10242 	if (arg + 1 < ARRAY_SIZE(fn->arg_type))
10243 		is_next_size = arg_type_is_mem_size(fn->arg_type[arg + 1]);
10244 
10245 	if (base_type(fn->arg_type[arg]) != ARG_PTR_TO_MEM)
10246 		return is_next_size;
10247 
10248 	return has_size == is_next_size || is_next_size == is_fixed;
10249 }
10250 
check_arg_pair_ok(const struct bpf_func_proto * fn)10251 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
10252 {
10253 	/* bpf_xxx(..., buf, len) call will access 'len'
10254 	 * bytes from memory 'buf'. Both arg types need
10255 	 * to be paired, so make sure there's no buggy
10256 	 * helper function specification.
10257 	 */
10258 	if (arg_type_is_mem_size(fn->arg1_type) ||
10259 	    check_args_pair_invalid(fn, 0) ||
10260 	    check_args_pair_invalid(fn, 1) ||
10261 	    check_args_pair_invalid(fn, 2) ||
10262 	    check_args_pair_invalid(fn, 3) ||
10263 	    check_args_pair_invalid(fn, 4))
10264 		return false;
10265 
10266 	return true;
10267 }
10268 
check_btf_id_ok(const struct bpf_func_proto * fn)10269 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
10270 {
10271 	int i;
10272 
10273 	for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
10274 		if (base_type(fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
10275 			return !!fn->arg_btf_id[i];
10276 		if (base_type(fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
10277 			return fn->arg_btf_id[i] == BPF_PTR_POISON;
10278 		if (base_type(fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
10279 		    /* arg_btf_id and arg_size are in a union. */
10280 		    (base_type(fn->arg_type[i]) != ARG_PTR_TO_MEM ||
10281 		     !(fn->arg_type[i] & MEM_FIXED_SIZE)))
10282 			return false;
10283 	}
10284 
10285 	return true;
10286 }
10287 
check_func_proto(const struct bpf_func_proto * fn,int func_id)10288 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
10289 {
10290 	return check_raw_mode_ok(fn) &&
10291 	       check_arg_pair_ok(fn) &&
10292 	       check_btf_id_ok(fn) ? 0 : -EINVAL;
10293 }
10294 
10295 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
10296  * are now invalid, so turn them into unknown SCALAR_VALUE.
10297  *
10298  * This also applies to dynptr slices belonging to skb and xdp dynptrs,
10299  * since these slices point to packet data.
10300  */
clear_all_pkt_pointers(struct bpf_verifier_env * env)10301 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
10302 {
10303 	struct bpf_func_state *state;
10304 	struct bpf_reg_state *reg;
10305 
10306 	bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
10307 		if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
10308 			mark_reg_invalid(env, reg);
10309 	}));
10310 }
10311 
10312 enum {
10313 	AT_PKT_END = -1,
10314 	BEYOND_PKT_END = -2,
10315 };
10316 
mark_pkt_end(struct bpf_verifier_state * vstate,int regn,bool range_open)10317 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
10318 {
10319 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
10320 	struct bpf_reg_state *reg = &state->regs[regn];
10321 
10322 	if (reg->type != PTR_TO_PACKET)
10323 		/* PTR_TO_PACKET_META is not supported yet */
10324 		return;
10325 
10326 	/* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
10327 	 * How far beyond pkt_end it goes is unknown.
10328 	 * if (!range_open) it's the case of pkt >= pkt_end
10329 	 * if (range_open) it's the case of pkt > pkt_end
10330 	 * hence this pointer is at least 1 byte bigger than pkt_end
10331 	 */
10332 	if (range_open)
10333 		reg->range = BEYOND_PKT_END;
10334 	else
10335 		reg->range = AT_PKT_END;
10336 }
10337 
release_reference_nomark(struct bpf_verifier_state * state,int ref_obj_id)10338 static int release_reference_nomark(struct bpf_verifier_state *state, int ref_obj_id)
10339 {
10340 	int i;
10341 
10342 	for (i = 0; i < state->acquired_refs; i++) {
10343 		if (state->refs[i].type != REF_TYPE_PTR)
10344 			continue;
10345 		if (state->refs[i].id == ref_obj_id) {
10346 			release_reference_state(state, i);
10347 			return 0;
10348 		}
10349 	}
10350 	return -EINVAL;
10351 }
10352 
10353 /* The pointer with the specified id has released its reference to kernel
10354  * resources. Identify all copies of the same pointer and clear the reference.
10355  *
10356  * This is the release function corresponding to acquire_reference(). Idempotent.
10357  */
release_reference(struct bpf_verifier_env * env,int ref_obj_id)10358 static int release_reference(struct bpf_verifier_env *env, int ref_obj_id)
10359 {
10360 	struct bpf_verifier_state *vstate = env->cur_state;
10361 	struct bpf_func_state *state;
10362 	struct bpf_reg_state *reg;
10363 	int err;
10364 
10365 	err = release_reference_nomark(vstate, ref_obj_id);
10366 	if (err)
10367 		return err;
10368 
10369 	bpf_for_each_reg_in_vstate(vstate, state, reg, ({
10370 		if (reg->ref_obj_id == ref_obj_id)
10371 			mark_reg_invalid(env, reg);
10372 	}));
10373 
10374 	return 0;
10375 }
10376 
invalidate_non_owning_refs(struct bpf_verifier_env * env)10377 static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
10378 {
10379 	struct bpf_func_state *unused;
10380 	struct bpf_reg_state *reg;
10381 
10382 	bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10383 		if (type_is_non_owning_ref(reg->type))
10384 			mark_reg_invalid(env, reg);
10385 	}));
10386 }
10387 
clear_caller_saved_regs(struct bpf_verifier_env * env,struct bpf_reg_state * regs)10388 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
10389 				    struct bpf_reg_state *regs)
10390 {
10391 	int i;
10392 
10393 	/* after the call registers r0 - r5 were scratched */
10394 	for (i = 0; i < CALLER_SAVED_REGS; i++) {
10395 		mark_reg_not_init(env, regs, caller_saved[i]);
10396 		__check_reg_arg(env, regs, caller_saved[i], DST_OP_NO_MARK);
10397 	}
10398 }
10399 
10400 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
10401 				   struct bpf_func_state *caller,
10402 				   struct bpf_func_state *callee,
10403 				   int insn_idx);
10404 
10405 static int set_callee_state(struct bpf_verifier_env *env,
10406 			    struct bpf_func_state *caller,
10407 			    struct bpf_func_state *callee, int insn_idx);
10408 
setup_func_entry(struct bpf_verifier_env * env,int subprog,int callsite,set_callee_state_fn set_callee_state_cb,struct bpf_verifier_state * state)10409 static int setup_func_entry(struct bpf_verifier_env *env, int subprog, int callsite,
10410 			    set_callee_state_fn set_callee_state_cb,
10411 			    struct bpf_verifier_state *state)
10412 {
10413 	struct bpf_func_state *caller, *callee;
10414 	int err;
10415 
10416 	if (state->curframe + 1 >= MAX_CALL_FRAMES) {
10417 		verbose(env, "the call stack of %d frames is too deep\n",
10418 			state->curframe + 2);
10419 		return -E2BIG;
10420 	}
10421 
10422 	if (state->frame[state->curframe + 1]) {
10423 		verifier_bug(env, "Frame %d already allocated", state->curframe + 1);
10424 		return -EFAULT;
10425 	}
10426 
10427 	caller = state->frame[state->curframe];
10428 	callee = kzalloc(sizeof(*callee), GFP_KERNEL_ACCOUNT);
10429 	if (!callee)
10430 		return -ENOMEM;
10431 	state->frame[state->curframe + 1] = callee;
10432 
10433 	/* callee cannot access r0, r6 - r9 for reading and has to write
10434 	 * into its own stack before reading from it.
10435 	 * callee can read/write into caller's stack
10436 	 */
10437 	init_func_state(env, callee,
10438 			/* remember the callsite, it will be used by bpf_exit */
10439 			callsite,
10440 			state->curframe + 1 /* frameno within this callchain */,
10441 			subprog /* subprog number within this prog */);
10442 	err = set_callee_state_cb(env, caller, callee, callsite);
10443 	if (err)
10444 		goto err_out;
10445 
10446 	/* only increment it after check_reg_arg() finished */
10447 	state->curframe++;
10448 
10449 	return 0;
10450 
10451 err_out:
10452 	free_func_state(callee);
10453 	state->frame[state->curframe + 1] = NULL;
10454 	return err;
10455 }
10456 
btf_check_func_arg_match(struct bpf_verifier_env * env,int subprog,const struct btf * btf,struct bpf_reg_state * regs)10457 static int btf_check_func_arg_match(struct bpf_verifier_env *env, int subprog,
10458 				    const struct btf *btf,
10459 				    struct bpf_reg_state *regs)
10460 {
10461 	struct bpf_subprog_info *sub = subprog_info(env, subprog);
10462 	struct bpf_verifier_log *log = &env->log;
10463 	u32 i;
10464 	int ret;
10465 
10466 	ret = btf_prepare_func_args(env, subprog);
10467 	if (ret)
10468 		return ret;
10469 
10470 	/* check that BTF function arguments match actual types that the
10471 	 * verifier sees.
10472 	 */
10473 	for (i = 0; i < sub->arg_cnt; i++) {
10474 		u32 regno = i + 1;
10475 		struct bpf_reg_state *reg = &regs[regno];
10476 		struct bpf_subprog_arg_info *arg = &sub->args[i];
10477 
10478 		if (arg->arg_type == ARG_ANYTHING) {
10479 			if (reg->type != SCALAR_VALUE) {
10480 				bpf_log(log, "R%d is not a scalar\n", regno);
10481 				return -EINVAL;
10482 			}
10483 		} else if (arg->arg_type & PTR_UNTRUSTED) {
10484 			/*
10485 			 * Anything is allowed for untrusted arguments, as these are
10486 			 * read-only and probe read instructions would protect against
10487 			 * invalid memory access.
10488 			 */
10489 		} else if (arg->arg_type == ARG_PTR_TO_CTX) {
10490 			ret = check_func_arg_reg_off(env, reg, regno, ARG_DONTCARE);
10491 			if (ret < 0)
10492 				return ret;
10493 			/* If function expects ctx type in BTF check that caller
10494 			 * is passing PTR_TO_CTX.
10495 			 */
10496 			if (reg->type != PTR_TO_CTX) {
10497 				bpf_log(log, "arg#%d expects pointer to ctx\n", i);
10498 				return -EINVAL;
10499 			}
10500 		} else if (base_type(arg->arg_type) == ARG_PTR_TO_MEM) {
10501 			ret = check_func_arg_reg_off(env, reg, regno, ARG_DONTCARE);
10502 			if (ret < 0)
10503 				return ret;
10504 			if (check_mem_reg(env, reg, regno, arg->mem_size))
10505 				return -EINVAL;
10506 			if (!(arg->arg_type & PTR_MAYBE_NULL) && (reg->type & PTR_MAYBE_NULL)) {
10507 				bpf_log(log, "arg#%d is expected to be non-NULL\n", i);
10508 				return -EINVAL;
10509 			}
10510 		} else if (base_type(arg->arg_type) == ARG_PTR_TO_ARENA) {
10511 			/*
10512 			 * Can pass any value and the kernel won't crash, but
10513 			 * only PTR_TO_ARENA or SCALAR make sense. Everything
10514 			 * else is a bug in the bpf program. Point it out to
10515 			 * the user at the verification time instead of
10516 			 * run-time debug nightmare.
10517 			 */
10518 			if (reg->type != PTR_TO_ARENA && reg->type != SCALAR_VALUE) {
10519 				bpf_log(log, "R%d is not a pointer to arena or scalar.\n", regno);
10520 				return -EINVAL;
10521 			}
10522 		} else if (arg->arg_type == (ARG_PTR_TO_DYNPTR | MEM_RDONLY)) {
10523 			ret = check_func_arg_reg_off(env, reg, regno, ARG_PTR_TO_DYNPTR);
10524 			if (ret)
10525 				return ret;
10526 
10527 			ret = process_dynptr_func(env, regno, -1, arg->arg_type, 0);
10528 			if (ret)
10529 				return ret;
10530 		} else if (base_type(arg->arg_type) == ARG_PTR_TO_BTF_ID) {
10531 			struct bpf_call_arg_meta meta;
10532 			int err;
10533 
10534 			if (register_is_null(reg) && type_may_be_null(arg->arg_type))
10535 				continue;
10536 
10537 			memset(&meta, 0, sizeof(meta)); /* leave func_id as zero */
10538 			err = check_reg_type(env, regno, arg->arg_type, &arg->btf_id, &meta);
10539 			err = err ?: check_func_arg_reg_off(env, reg, regno, arg->arg_type);
10540 			if (err)
10541 				return err;
10542 		} else {
10543 			verifier_bug(env, "unrecognized arg#%d type %d", i, arg->arg_type);
10544 			return -EFAULT;
10545 		}
10546 	}
10547 
10548 	return 0;
10549 }
10550 
10551 /* Compare BTF of a function call with given bpf_reg_state.
10552  * Returns:
10553  * EFAULT - there is a verifier bug. Abort verification.
10554  * EINVAL - there is a type mismatch or BTF is not available.
10555  * 0 - BTF matches with what bpf_reg_state expects.
10556  * Only PTR_TO_CTX and SCALAR_VALUE states are recognized.
10557  */
btf_check_subprog_call(struct bpf_verifier_env * env,int subprog,struct bpf_reg_state * regs)10558 static int btf_check_subprog_call(struct bpf_verifier_env *env, int subprog,
10559 				  struct bpf_reg_state *regs)
10560 {
10561 	struct bpf_prog *prog = env->prog;
10562 	struct btf *btf = prog->aux->btf;
10563 	u32 btf_id;
10564 	int err;
10565 
10566 	if (!prog->aux->func_info)
10567 		return -EINVAL;
10568 
10569 	btf_id = prog->aux->func_info[subprog].type_id;
10570 	if (!btf_id)
10571 		return -EFAULT;
10572 
10573 	if (prog->aux->func_info_aux[subprog].unreliable)
10574 		return -EINVAL;
10575 
10576 	err = btf_check_func_arg_match(env, subprog, btf, regs);
10577 	/* Compiler optimizations can remove arguments from static functions
10578 	 * or mismatched type can be passed into a global function.
10579 	 * In such cases mark the function as unreliable from BTF point of view.
10580 	 */
10581 	if (err)
10582 		prog->aux->func_info_aux[subprog].unreliable = true;
10583 	return err;
10584 }
10585 
push_callback_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int insn_idx,int subprog,set_callee_state_fn set_callee_state_cb)10586 static int push_callback_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
10587 			      int insn_idx, int subprog,
10588 			      set_callee_state_fn set_callee_state_cb)
10589 {
10590 	struct bpf_verifier_state *state = env->cur_state, *callback_state;
10591 	struct bpf_func_state *caller, *callee;
10592 	int err;
10593 
10594 	caller = state->frame[state->curframe];
10595 	err = btf_check_subprog_call(env, subprog, caller->regs);
10596 	if (err == -EFAULT)
10597 		return err;
10598 
10599 	/* set_callee_state is used for direct subprog calls, but we are
10600 	 * interested in validating only BPF helpers that can call subprogs as
10601 	 * callbacks
10602 	 */
10603 	env->subprog_info[subprog].is_cb = true;
10604 	if (bpf_pseudo_kfunc_call(insn) &&
10605 	    !is_callback_calling_kfunc(insn->imm)) {
10606 		verifier_bug(env, "kfunc %s#%d not marked as callback-calling",
10607 			     func_id_name(insn->imm), insn->imm);
10608 		return -EFAULT;
10609 	} else if (!bpf_pseudo_kfunc_call(insn) &&
10610 		   !is_callback_calling_function(insn->imm)) { /* helper */
10611 		verifier_bug(env, "helper %s#%d not marked as callback-calling",
10612 			     func_id_name(insn->imm), insn->imm);
10613 		return -EFAULT;
10614 	}
10615 
10616 	if (is_async_callback_calling_insn(insn)) {
10617 		struct bpf_verifier_state *async_cb;
10618 
10619 		/* there is no real recursion here. timer and workqueue callbacks are async */
10620 		env->subprog_info[subprog].is_async_cb = true;
10621 		async_cb = push_async_cb(env, env->subprog_info[subprog].start,
10622 					 insn_idx, subprog,
10623 					 is_bpf_wq_set_callback_impl_kfunc(insn->imm));
10624 		if (!async_cb)
10625 			return -EFAULT;
10626 		callee = async_cb->frame[0];
10627 		callee->async_entry_cnt = caller->async_entry_cnt + 1;
10628 
10629 		/* Convert bpf_timer_set_callback() args into timer callback args */
10630 		err = set_callee_state_cb(env, caller, callee, insn_idx);
10631 		if (err)
10632 			return err;
10633 
10634 		return 0;
10635 	}
10636 
10637 	/* for callback functions enqueue entry to callback and
10638 	 * proceed with next instruction within current frame.
10639 	 */
10640 	callback_state = push_stack(env, env->subprog_info[subprog].start, insn_idx, false);
10641 	if (!callback_state)
10642 		return -ENOMEM;
10643 
10644 	err = setup_func_entry(env, subprog, insn_idx, set_callee_state_cb,
10645 			       callback_state);
10646 	if (err)
10647 		return err;
10648 
10649 	callback_state->callback_unroll_depth++;
10650 	callback_state->frame[callback_state->curframe - 1]->callback_depth++;
10651 	caller->callback_depth = 0;
10652 	return 0;
10653 }
10654 
check_func_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)10655 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
10656 			   int *insn_idx)
10657 {
10658 	struct bpf_verifier_state *state = env->cur_state;
10659 	struct bpf_func_state *caller;
10660 	int err, subprog, target_insn;
10661 
10662 	target_insn = *insn_idx + insn->imm + 1;
10663 	subprog = find_subprog(env, target_insn);
10664 	if (verifier_bug_if(subprog < 0, env, "target of func call at insn %d is not a program",
10665 			    target_insn))
10666 		return -EFAULT;
10667 
10668 	caller = state->frame[state->curframe];
10669 	err = btf_check_subprog_call(env, subprog, caller->regs);
10670 	if (err == -EFAULT)
10671 		return err;
10672 	if (subprog_is_global(env, subprog)) {
10673 		const char *sub_name = subprog_name(env, subprog);
10674 
10675 		if (env->cur_state->active_locks) {
10676 			verbose(env, "global function calls are not allowed while holding a lock,\n"
10677 				     "use static function instead\n");
10678 			return -EINVAL;
10679 		}
10680 
10681 		if (env->subprog_info[subprog].might_sleep &&
10682 		    (env->cur_state->active_rcu_lock || env->cur_state->active_preempt_locks ||
10683 		     env->cur_state->active_irq_id || !in_sleepable(env))) {
10684 			verbose(env, "global functions that may sleep are not allowed in non-sleepable context,\n"
10685 				     "i.e., in a RCU/IRQ/preempt-disabled section, or in\n"
10686 				     "a non-sleepable BPF program context\n");
10687 			return -EINVAL;
10688 		}
10689 
10690 		if (err) {
10691 			verbose(env, "Caller passes invalid args into func#%d ('%s')\n",
10692 				subprog, sub_name);
10693 			return err;
10694 		}
10695 
10696 		verbose(env, "Func#%d ('%s') is global and assumed valid.\n",
10697 			subprog, sub_name);
10698 		if (env->subprog_info[subprog].changes_pkt_data)
10699 			clear_all_pkt_pointers(env);
10700 		/* mark global subprog for verifying after main prog */
10701 		subprog_aux(env, subprog)->called = true;
10702 		clear_caller_saved_regs(env, caller->regs);
10703 
10704 		/* All global functions return a 64-bit SCALAR_VALUE */
10705 		mark_reg_unknown(env, caller->regs, BPF_REG_0);
10706 		caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
10707 
10708 		/* continue with next insn after call */
10709 		return 0;
10710 	}
10711 
10712 	/* for regular function entry setup new frame and continue
10713 	 * from that frame.
10714 	 */
10715 	err = setup_func_entry(env, subprog, *insn_idx, set_callee_state, state);
10716 	if (err)
10717 		return err;
10718 
10719 	clear_caller_saved_regs(env, caller->regs);
10720 
10721 	/* and go analyze first insn of the callee */
10722 	*insn_idx = env->subprog_info[subprog].start - 1;
10723 
10724 	if (env->log.level & BPF_LOG_LEVEL) {
10725 		verbose(env, "caller:\n");
10726 		print_verifier_state(env, state, caller->frameno, true);
10727 		verbose(env, "callee:\n");
10728 		print_verifier_state(env, state, state->curframe, true);
10729 	}
10730 
10731 	return 0;
10732 }
10733 
map_set_for_each_callback_args(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee)10734 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
10735 				   struct bpf_func_state *caller,
10736 				   struct bpf_func_state *callee)
10737 {
10738 	/* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
10739 	 *      void *callback_ctx, u64 flags);
10740 	 * callback_fn(struct bpf_map *map, void *key, void *value,
10741 	 *      void *callback_ctx);
10742 	 */
10743 	callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
10744 
10745 	callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
10746 	__mark_reg_known_zero(&callee->regs[BPF_REG_2]);
10747 	callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
10748 
10749 	callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
10750 	__mark_reg_known_zero(&callee->regs[BPF_REG_3]);
10751 	callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
10752 
10753 	/* pointer to stack or null */
10754 	callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
10755 
10756 	/* unused */
10757 	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
10758 	return 0;
10759 }
10760 
set_callee_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)10761 static int set_callee_state(struct bpf_verifier_env *env,
10762 			    struct bpf_func_state *caller,
10763 			    struct bpf_func_state *callee, int insn_idx)
10764 {
10765 	int i;
10766 
10767 	/* copy r1 - r5 args that callee can access.  The copy includes parent
10768 	 * pointers, which connects us up to the liveness chain
10769 	 */
10770 	for (i = BPF_REG_1; i <= BPF_REG_5; i++)
10771 		callee->regs[i] = caller->regs[i];
10772 	return 0;
10773 }
10774 
set_map_elem_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)10775 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
10776 				       struct bpf_func_state *caller,
10777 				       struct bpf_func_state *callee,
10778 				       int insn_idx)
10779 {
10780 	struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
10781 	struct bpf_map *map;
10782 	int err;
10783 
10784 	/* valid map_ptr and poison value does not matter */
10785 	map = insn_aux->map_ptr_state.map_ptr;
10786 	if (!map->ops->map_set_for_each_callback_args ||
10787 	    !map->ops->map_for_each_callback) {
10788 		verbose(env, "callback function not allowed for map\n");
10789 		return -ENOTSUPP;
10790 	}
10791 
10792 	err = map->ops->map_set_for_each_callback_args(env, caller, callee);
10793 	if (err)
10794 		return err;
10795 
10796 	callee->in_callback_fn = true;
10797 	callee->callback_ret_range = retval_range(0, 1);
10798 	return 0;
10799 }
10800 
set_loop_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)10801 static int set_loop_callback_state(struct bpf_verifier_env *env,
10802 				   struct bpf_func_state *caller,
10803 				   struct bpf_func_state *callee,
10804 				   int insn_idx)
10805 {
10806 	/* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
10807 	 *	    u64 flags);
10808 	 * callback_fn(u64 index, void *callback_ctx);
10809 	 */
10810 	callee->regs[BPF_REG_1].type = SCALAR_VALUE;
10811 	callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
10812 
10813 	/* unused */
10814 	__mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
10815 	__mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
10816 	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
10817 
10818 	callee->in_callback_fn = true;
10819 	callee->callback_ret_range = retval_range(0, 1);
10820 	return 0;
10821 }
10822 
set_timer_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)10823 static int set_timer_callback_state(struct bpf_verifier_env *env,
10824 				    struct bpf_func_state *caller,
10825 				    struct bpf_func_state *callee,
10826 				    int insn_idx)
10827 {
10828 	struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
10829 
10830 	/* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
10831 	 * callback_fn(struct bpf_map *map, void *key, void *value);
10832 	 */
10833 	callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
10834 	__mark_reg_known_zero(&callee->regs[BPF_REG_1]);
10835 	callee->regs[BPF_REG_1].map_ptr = map_ptr;
10836 
10837 	callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
10838 	__mark_reg_known_zero(&callee->regs[BPF_REG_2]);
10839 	callee->regs[BPF_REG_2].map_ptr = map_ptr;
10840 
10841 	callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
10842 	__mark_reg_known_zero(&callee->regs[BPF_REG_3]);
10843 	callee->regs[BPF_REG_3].map_ptr = map_ptr;
10844 
10845 	/* unused */
10846 	__mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
10847 	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
10848 	callee->in_async_callback_fn = true;
10849 	callee->callback_ret_range = retval_range(0, 1);
10850 	return 0;
10851 }
10852 
set_find_vma_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)10853 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
10854 				       struct bpf_func_state *caller,
10855 				       struct bpf_func_state *callee,
10856 				       int insn_idx)
10857 {
10858 	/* bpf_find_vma(struct task_struct *task, u64 addr,
10859 	 *               void *callback_fn, void *callback_ctx, u64 flags)
10860 	 * (callback_fn)(struct task_struct *task,
10861 	 *               struct vm_area_struct *vma, void *callback_ctx);
10862 	 */
10863 	callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
10864 
10865 	callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
10866 	__mark_reg_known_zero(&callee->regs[BPF_REG_2]);
10867 	callee->regs[BPF_REG_2].btf =  btf_vmlinux;
10868 	callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA];
10869 
10870 	/* pointer to stack or null */
10871 	callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
10872 
10873 	/* unused */
10874 	__mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
10875 	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
10876 	callee->in_callback_fn = true;
10877 	callee->callback_ret_range = retval_range(0, 1);
10878 	return 0;
10879 }
10880 
set_user_ringbuf_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)10881 static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
10882 					   struct bpf_func_state *caller,
10883 					   struct bpf_func_state *callee,
10884 					   int insn_idx)
10885 {
10886 	/* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
10887 	 *			  callback_ctx, u64 flags);
10888 	 * callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
10889 	 */
10890 	__mark_reg_not_init(env, &callee->regs[BPF_REG_0]);
10891 	mark_dynptr_cb_reg(env, &callee->regs[BPF_REG_1], BPF_DYNPTR_TYPE_LOCAL);
10892 	callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
10893 
10894 	/* unused */
10895 	__mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
10896 	__mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
10897 	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
10898 
10899 	callee->in_callback_fn = true;
10900 	callee->callback_ret_range = retval_range(0, 1);
10901 	return 0;
10902 }
10903 
set_rbtree_add_callback_state(struct bpf_verifier_env * env,struct bpf_func_state * caller,struct bpf_func_state * callee,int insn_idx)10904 static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
10905 					 struct bpf_func_state *caller,
10906 					 struct bpf_func_state *callee,
10907 					 int insn_idx)
10908 {
10909 	/* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
10910 	 *                     bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
10911 	 *
10912 	 * 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
10913 	 * that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
10914 	 * by this point, so look at 'root'
10915 	 */
10916 	struct btf_field *field;
10917 
10918 	field = reg_find_field_offset(&caller->regs[BPF_REG_1], caller->regs[BPF_REG_1].off,
10919 				      BPF_RB_ROOT);
10920 	if (!field || !field->graph_root.value_btf_id)
10921 		return -EFAULT;
10922 
10923 	mark_reg_graph_node(callee->regs, BPF_REG_1, &field->graph_root);
10924 	ref_set_non_owning(env, &callee->regs[BPF_REG_1]);
10925 	mark_reg_graph_node(callee->regs, BPF_REG_2, &field->graph_root);
10926 	ref_set_non_owning(env, &callee->regs[BPF_REG_2]);
10927 
10928 	__mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
10929 	__mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
10930 	__mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
10931 	callee->in_callback_fn = true;
10932 	callee->callback_ret_range = retval_range(0, 1);
10933 	return 0;
10934 }
10935 
10936 static bool is_rbtree_lock_required_kfunc(u32 btf_id);
10937 
10938 /* Are we currently verifying the callback for a rbtree helper that must
10939  * be called with lock held? If so, no need to complain about unreleased
10940  * lock
10941  */
in_rbtree_lock_required_cb(struct bpf_verifier_env * env)10942 static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
10943 {
10944 	struct bpf_verifier_state *state = env->cur_state;
10945 	struct bpf_insn *insn = env->prog->insnsi;
10946 	struct bpf_func_state *callee;
10947 	int kfunc_btf_id;
10948 
10949 	if (!state->curframe)
10950 		return false;
10951 
10952 	callee = state->frame[state->curframe];
10953 
10954 	if (!callee->in_callback_fn)
10955 		return false;
10956 
10957 	kfunc_btf_id = insn[callee->callsite].imm;
10958 	return is_rbtree_lock_required_kfunc(kfunc_btf_id);
10959 }
10960 
retval_range_within(struct bpf_retval_range range,const struct bpf_reg_state * reg,bool return_32bit)10961 static bool retval_range_within(struct bpf_retval_range range, const struct bpf_reg_state *reg,
10962 				bool return_32bit)
10963 {
10964 	if (return_32bit)
10965 		return range.minval <= reg->s32_min_value && reg->s32_max_value <= range.maxval;
10966 	else
10967 		return range.minval <= reg->smin_value && reg->smax_value <= range.maxval;
10968 }
10969 
prepare_func_exit(struct bpf_verifier_env * env,int * insn_idx)10970 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
10971 {
10972 	struct bpf_verifier_state *state = env->cur_state, *prev_st;
10973 	struct bpf_func_state *caller, *callee;
10974 	struct bpf_reg_state *r0;
10975 	bool in_callback_fn;
10976 	int err;
10977 
10978 	callee = state->frame[state->curframe];
10979 	r0 = &callee->regs[BPF_REG_0];
10980 	if (r0->type == PTR_TO_STACK) {
10981 		/* technically it's ok to return caller's stack pointer
10982 		 * (or caller's caller's pointer) back to the caller,
10983 		 * since these pointers are valid. Only current stack
10984 		 * pointer will be invalid as soon as function exits,
10985 		 * but let's be conservative
10986 		 */
10987 		verbose(env, "cannot return stack pointer to the caller\n");
10988 		return -EINVAL;
10989 	}
10990 
10991 	caller = state->frame[state->curframe - 1];
10992 	if (callee->in_callback_fn) {
10993 		if (r0->type != SCALAR_VALUE) {
10994 			verbose(env, "R0 not a scalar value\n");
10995 			return -EACCES;
10996 		}
10997 
10998 		/* we are going to rely on register's precise value */
10999 		err = mark_reg_read(env, r0, r0->parent, REG_LIVE_READ64);
11000 		err = err ?: mark_chain_precision(env, BPF_REG_0);
11001 		if (err)
11002 			return err;
11003 
11004 		/* enforce R0 return value range, and bpf_callback_t returns 64bit */
11005 		if (!retval_range_within(callee->callback_ret_range, r0, false)) {
11006 			verbose_invalid_scalar(env, r0, callee->callback_ret_range,
11007 					       "At callback return", "R0");
11008 			return -EINVAL;
11009 		}
11010 		if (!calls_callback(env, callee->callsite)) {
11011 			verifier_bug(env, "in callback at %d, callsite %d !calls_callback",
11012 				     *insn_idx, callee->callsite);
11013 			return -EFAULT;
11014 		}
11015 	} else {
11016 		/* return to the caller whatever r0 had in the callee */
11017 		caller->regs[BPF_REG_0] = *r0;
11018 	}
11019 
11020 	/* for callbacks like bpf_loop or bpf_for_each_map_elem go back to callsite,
11021 	 * there function call logic would reschedule callback visit. If iteration
11022 	 * converges is_state_visited() would prune that visit eventually.
11023 	 */
11024 	in_callback_fn = callee->in_callback_fn;
11025 	if (in_callback_fn)
11026 		*insn_idx = callee->callsite;
11027 	else
11028 		*insn_idx = callee->callsite + 1;
11029 
11030 	if (env->log.level & BPF_LOG_LEVEL) {
11031 		verbose(env, "returning from callee:\n");
11032 		print_verifier_state(env, state, callee->frameno, true);
11033 		verbose(env, "to caller at %d:\n", *insn_idx);
11034 		print_verifier_state(env, state, caller->frameno, true);
11035 	}
11036 	/* clear everything in the callee. In case of exceptional exits using
11037 	 * bpf_throw, this will be done by copy_verifier_state for extra frames. */
11038 	free_func_state(callee);
11039 	state->frame[state->curframe--] = NULL;
11040 
11041 	/* for callbacks widen imprecise scalars to make programs like below verify:
11042 	 *
11043 	 *   struct ctx { int i; }
11044 	 *   void cb(int idx, struct ctx *ctx) { ctx->i++; ... }
11045 	 *   ...
11046 	 *   struct ctx = { .i = 0; }
11047 	 *   bpf_loop(100, cb, &ctx, 0);
11048 	 *
11049 	 * This is similar to what is done in process_iter_next_call() for open
11050 	 * coded iterators.
11051 	 */
11052 	prev_st = in_callback_fn ? find_prev_entry(env, state, *insn_idx) : NULL;
11053 	if (prev_st) {
11054 		err = widen_imprecise_scalars(env, prev_st, state);
11055 		if (err)
11056 			return err;
11057 	}
11058 	return 0;
11059 }
11060 
do_refine_retval_range(struct bpf_verifier_env * env,struct bpf_reg_state * regs,int ret_type,int func_id,struct bpf_call_arg_meta * meta)11061 static int do_refine_retval_range(struct bpf_verifier_env *env,
11062 				  struct bpf_reg_state *regs, int ret_type,
11063 				  int func_id,
11064 				  struct bpf_call_arg_meta *meta)
11065 {
11066 	struct bpf_reg_state *ret_reg = &regs[BPF_REG_0];
11067 
11068 	if (ret_type != RET_INTEGER)
11069 		return 0;
11070 
11071 	switch (func_id) {
11072 	case BPF_FUNC_get_stack:
11073 	case BPF_FUNC_get_task_stack:
11074 	case BPF_FUNC_probe_read_str:
11075 	case BPF_FUNC_probe_read_kernel_str:
11076 	case BPF_FUNC_probe_read_user_str:
11077 		ret_reg->smax_value = meta->msize_max_value;
11078 		ret_reg->s32_max_value = meta->msize_max_value;
11079 		ret_reg->smin_value = -MAX_ERRNO;
11080 		ret_reg->s32_min_value = -MAX_ERRNO;
11081 		reg_bounds_sync(ret_reg);
11082 		break;
11083 	case BPF_FUNC_get_smp_processor_id:
11084 		ret_reg->umax_value = nr_cpu_ids - 1;
11085 		ret_reg->u32_max_value = nr_cpu_ids - 1;
11086 		ret_reg->smax_value = nr_cpu_ids - 1;
11087 		ret_reg->s32_max_value = nr_cpu_ids - 1;
11088 		ret_reg->umin_value = 0;
11089 		ret_reg->u32_min_value = 0;
11090 		ret_reg->smin_value = 0;
11091 		ret_reg->s32_min_value = 0;
11092 		reg_bounds_sync(ret_reg);
11093 		break;
11094 	}
11095 
11096 	return reg_bounds_sanity_check(env, ret_reg, "retval");
11097 }
11098 
11099 static int
record_func_map(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)11100 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
11101 		int func_id, int insn_idx)
11102 {
11103 	struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
11104 	struct bpf_map *map = meta->map_ptr;
11105 
11106 	if (func_id != BPF_FUNC_tail_call &&
11107 	    func_id != BPF_FUNC_map_lookup_elem &&
11108 	    func_id != BPF_FUNC_map_update_elem &&
11109 	    func_id != BPF_FUNC_map_delete_elem &&
11110 	    func_id != BPF_FUNC_map_push_elem &&
11111 	    func_id != BPF_FUNC_map_pop_elem &&
11112 	    func_id != BPF_FUNC_map_peek_elem &&
11113 	    func_id != BPF_FUNC_for_each_map_elem &&
11114 	    func_id != BPF_FUNC_redirect_map &&
11115 	    func_id != BPF_FUNC_map_lookup_percpu_elem)
11116 		return 0;
11117 
11118 	if (map == NULL) {
11119 		verifier_bug(env, "expected map for helper call");
11120 		return -EFAULT;
11121 	}
11122 
11123 	/* In case of read-only, some additional restrictions
11124 	 * need to be applied in order to prevent altering the
11125 	 * state of the map from program side.
11126 	 */
11127 	if ((map->map_flags & BPF_F_RDONLY_PROG) &&
11128 	    (func_id == BPF_FUNC_map_delete_elem ||
11129 	     func_id == BPF_FUNC_map_update_elem ||
11130 	     func_id == BPF_FUNC_map_push_elem ||
11131 	     func_id == BPF_FUNC_map_pop_elem)) {
11132 		verbose(env, "write into map forbidden\n");
11133 		return -EACCES;
11134 	}
11135 
11136 	if (!aux->map_ptr_state.map_ptr)
11137 		bpf_map_ptr_store(aux, meta->map_ptr,
11138 				  !meta->map_ptr->bypass_spec_v1, false);
11139 	else if (aux->map_ptr_state.map_ptr != meta->map_ptr)
11140 		bpf_map_ptr_store(aux, meta->map_ptr,
11141 				  !meta->map_ptr->bypass_spec_v1, true);
11142 	return 0;
11143 }
11144 
11145 static int
record_func_key(struct bpf_verifier_env * env,struct bpf_call_arg_meta * meta,int func_id,int insn_idx)11146 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
11147 		int func_id, int insn_idx)
11148 {
11149 	struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
11150 	struct bpf_reg_state *regs = cur_regs(env), *reg;
11151 	struct bpf_map *map = meta->map_ptr;
11152 	u64 val, max;
11153 	int err;
11154 
11155 	if (func_id != BPF_FUNC_tail_call)
11156 		return 0;
11157 	if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
11158 		verbose(env, "expected prog array map for tail call");
11159 		return -EINVAL;
11160 	}
11161 
11162 	reg = &regs[BPF_REG_3];
11163 	val = reg->var_off.value;
11164 	max = map->max_entries;
11165 
11166 	if (!(is_reg_const(reg, false) && val < max)) {
11167 		bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
11168 		return 0;
11169 	}
11170 
11171 	err = mark_chain_precision(env, BPF_REG_3);
11172 	if (err)
11173 		return err;
11174 	if (bpf_map_key_unseen(aux))
11175 		bpf_map_key_store(aux, val);
11176 	else if (!bpf_map_key_poisoned(aux) &&
11177 		  bpf_map_key_immediate(aux) != val)
11178 		bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
11179 	return 0;
11180 }
11181 
check_reference_leak(struct bpf_verifier_env * env,bool exception_exit)11182 static int check_reference_leak(struct bpf_verifier_env *env, bool exception_exit)
11183 {
11184 	struct bpf_verifier_state *state = env->cur_state;
11185 	enum bpf_prog_type type = resolve_prog_type(env->prog);
11186 	struct bpf_reg_state *reg = reg_state(env, BPF_REG_0);
11187 	bool refs_lingering = false;
11188 	int i;
11189 
11190 	if (!exception_exit && cur_func(env)->frameno)
11191 		return 0;
11192 
11193 	for (i = 0; i < state->acquired_refs; i++) {
11194 		if (state->refs[i].type != REF_TYPE_PTR)
11195 			continue;
11196 		/* Allow struct_ops programs to return a referenced kptr back to
11197 		 * kernel. Type checks are performed later in check_return_code.
11198 		 */
11199 		if (type == BPF_PROG_TYPE_STRUCT_OPS && !exception_exit &&
11200 		    reg->ref_obj_id == state->refs[i].id)
11201 			continue;
11202 		verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
11203 			state->refs[i].id, state->refs[i].insn_idx);
11204 		refs_lingering = true;
11205 	}
11206 	return refs_lingering ? -EINVAL : 0;
11207 }
11208 
check_resource_leak(struct bpf_verifier_env * env,bool exception_exit,bool check_lock,const char * prefix)11209 static int check_resource_leak(struct bpf_verifier_env *env, bool exception_exit, bool check_lock, const char *prefix)
11210 {
11211 	int err;
11212 
11213 	if (check_lock && env->cur_state->active_locks) {
11214 		verbose(env, "%s cannot be used inside bpf_spin_lock-ed region\n", prefix);
11215 		return -EINVAL;
11216 	}
11217 
11218 	err = check_reference_leak(env, exception_exit);
11219 	if (err) {
11220 		verbose(env, "%s would lead to reference leak\n", prefix);
11221 		return err;
11222 	}
11223 
11224 	if (check_lock && env->cur_state->active_irq_id) {
11225 		verbose(env, "%s cannot be used inside bpf_local_irq_save-ed region\n", prefix);
11226 		return -EINVAL;
11227 	}
11228 
11229 	if (check_lock && env->cur_state->active_rcu_lock) {
11230 		verbose(env, "%s cannot be used inside bpf_rcu_read_lock-ed region\n", prefix);
11231 		return -EINVAL;
11232 	}
11233 
11234 	if (check_lock && env->cur_state->active_preempt_locks) {
11235 		verbose(env, "%s cannot be used inside bpf_preempt_disable-ed region\n", prefix);
11236 		return -EINVAL;
11237 	}
11238 
11239 	return 0;
11240 }
11241 
check_bpf_snprintf_call(struct bpf_verifier_env * env,struct bpf_reg_state * regs)11242 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
11243 				   struct bpf_reg_state *regs)
11244 {
11245 	struct bpf_reg_state *fmt_reg = &regs[BPF_REG_3];
11246 	struct bpf_reg_state *data_len_reg = &regs[BPF_REG_5];
11247 	struct bpf_map *fmt_map = fmt_reg->map_ptr;
11248 	struct bpf_bprintf_data data = {};
11249 	int err, fmt_map_off, num_args;
11250 	u64 fmt_addr;
11251 	char *fmt;
11252 
11253 	/* data must be an array of u64 */
11254 	if (data_len_reg->var_off.value % 8)
11255 		return -EINVAL;
11256 	num_args = data_len_reg->var_off.value / 8;
11257 
11258 	/* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
11259 	 * and map_direct_value_addr is set.
11260 	 */
11261 	fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
11262 	err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
11263 						  fmt_map_off);
11264 	if (err) {
11265 		verbose(env, "failed to retrieve map value address\n");
11266 		return -EFAULT;
11267 	}
11268 	fmt = (char *)(long)fmt_addr + fmt_map_off;
11269 
11270 	/* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
11271 	 * can focus on validating the format specifiers.
11272 	 */
11273 	err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, &data);
11274 	if (err < 0)
11275 		verbose(env, "Invalid format string\n");
11276 
11277 	return err;
11278 }
11279 
check_get_func_ip(struct bpf_verifier_env * env)11280 static int check_get_func_ip(struct bpf_verifier_env *env)
11281 {
11282 	enum bpf_prog_type type = resolve_prog_type(env->prog);
11283 	int func_id = BPF_FUNC_get_func_ip;
11284 
11285 	if (type == BPF_PROG_TYPE_TRACING) {
11286 		if (!bpf_prog_has_trampoline(env->prog)) {
11287 			verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
11288 				func_id_name(func_id), func_id);
11289 			return -ENOTSUPP;
11290 		}
11291 		return 0;
11292 	} else if (type == BPF_PROG_TYPE_KPROBE) {
11293 		return 0;
11294 	}
11295 
11296 	verbose(env, "func %s#%d not supported for program type %d\n",
11297 		func_id_name(func_id), func_id, type);
11298 	return -ENOTSUPP;
11299 }
11300 
cur_aux(const struct bpf_verifier_env * env)11301 static struct bpf_insn_aux_data *cur_aux(const struct bpf_verifier_env *env)
11302 {
11303 	return &env->insn_aux_data[env->insn_idx];
11304 }
11305 
loop_flag_is_zero(struct bpf_verifier_env * env)11306 static bool loop_flag_is_zero(struct bpf_verifier_env *env)
11307 {
11308 	struct bpf_reg_state *regs = cur_regs(env);
11309 	struct bpf_reg_state *reg = &regs[BPF_REG_4];
11310 	bool reg_is_null = register_is_null(reg);
11311 
11312 	if (reg_is_null)
11313 		mark_chain_precision(env, BPF_REG_4);
11314 
11315 	return reg_is_null;
11316 }
11317 
update_loop_inline_state(struct bpf_verifier_env * env,u32 subprogno)11318 static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
11319 {
11320 	struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
11321 
11322 	if (!state->initialized) {
11323 		state->initialized = 1;
11324 		state->fit_for_inline = loop_flag_is_zero(env);
11325 		state->callback_subprogno = subprogno;
11326 		return;
11327 	}
11328 
11329 	if (!state->fit_for_inline)
11330 		return;
11331 
11332 	state->fit_for_inline = (loop_flag_is_zero(env) &&
11333 				 state->callback_subprogno == subprogno);
11334 }
11335 
11336 /* Returns whether or not the given map type can potentially elide
11337  * lookup return value nullness check. This is possible if the key
11338  * is statically known.
11339  */
can_elide_value_nullness(enum bpf_map_type type)11340 static bool can_elide_value_nullness(enum bpf_map_type type)
11341 {
11342 	switch (type) {
11343 	case BPF_MAP_TYPE_ARRAY:
11344 	case BPF_MAP_TYPE_PERCPU_ARRAY:
11345 		return true;
11346 	default:
11347 		return false;
11348 	}
11349 }
11350 
get_helper_proto(struct bpf_verifier_env * env,int func_id,const struct bpf_func_proto ** ptr)11351 static int get_helper_proto(struct bpf_verifier_env *env, int func_id,
11352 			    const struct bpf_func_proto **ptr)
11353 {
11354 	if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID)
11355 		return -ERANGE;
11356 
11357 	if (!env->ops->get_func_proto)
11358 		return -EINVAL;
11359 
11360 	*ptr = env->ops->get_func_proto(func_id, env->prog);
11361 	return *ptr && (*ptr)->func ? 0 : -EINVAL;
11362 }
11363 
check_helper_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx_p)11364 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11365 			     int *insn_idx_p)
11366 {
11367 	enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
11368 	bool returns_cpu_specific_alloc_ptr = false;
11369 	const struct bpf_func_proto *fn = NULL;
11370 	enum bpf_return_type ret_type;
11371 	enum bpf_type_flag ret_flag;
11372 	struct bpf_reg_state *regs;
11373 	struct bpf_call_arg_meta meta;
11374 	int insn_idx = *insn_idx_p;
11375 	bool changes_data;
11376 	int i, err, func_id;
11377 
11378 	/* find function prototype */
11379 	func_id = insn->imm;
11380 	err = get_helper_proto(env, insn->imm, &fn);
11381 	if (err == -ERANGE) {
11382 		verbose(env, "invalid func %s#%d\n", func_id_name(func_id), func_id);
11383 		return -EINVAL;
11384 	}
11385 
11386 	if (err) {
11387 		verbose(env, "program of this type cannot use helper %s#%d\n",
11388 			func_id_name(func_id), func_id);
11389 		return err;
11390 	}
11391 
11392 	/* eBPF programs must be GPL compatible to use GPL-ed functions */
11393 	if (!env->prog->gpl_compatible && fn->gpl_only) {
11394 		verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
11395 		return -EINVAL;
11396 	}
11397 
11398 	if (fn->allowed && !fn->allowed(env->prog)) {
11399 		verbose(env, "helper call is not allowed in probe\n");
11400 		return -EINVAL;
11401 	}
11402 
11403 	if (!in_sleepable(env) && fn->might_sleep) {
11404 		verbose(env, "helper call might sleep in a non-sleepable prog\n");
11405 		return -EINVAL;
11406 	}
11407 
11408 	/* With LD_ABS/IND some JITs save/restore skb from r1. */
11409 	changes_data = bpf_helper_changes_pkt_data(func_id);
11410 	if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
11411 		verifier_bug(env, "func %s#%d: r1 != ctx", func_id_name(func_id), func_id);
11412 		return -EFAULT;
11413 	}
11414 
11415 	memset(&meta, 0, sizeof(meta));
11416 	meta.pkt_access = fn->pkt_access;
11417 
11418 	err = check_func_proto(fn, func_id);
11419 	if (err) {
11420 		verifier_bug(env, "incorrect func proto %s#%d", func_id_name(func_id), func_id);
11421 		return err;
11422 	}
11423 
11424 	if (env->cur_state->active_rcu_lock) {
11425 		if (fn->might_sleep) {
11426 			verbose(env, "sleepable helper %s#%d in rcu_read_lock region\n",
11427 				func_id_name(func_id), func_id);
11428 			return -EINVAL;
11429 		}
11430 
11431 		if (in_sleepable(env) && is_storage_get_function(func_id))
11432 			env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
11433 	}
11434 
11435 	if (env->cur_state->active_preempt_locks) {
11436 		if (fn->might_sleep) {
11437 			verbose(env, "sleepable helper %s#%d in non-preemptible region\n",
11438 				func_id_name(func_id), func_id);
11439 			return -EINVAL;
11440 		}
11441 
11442 		if (in_sleepable(env) && is_storage_get_function(func_id))
11443 			env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
11444 	}
11445 
11446 	if (env->cur_state->active_irq_id) {
11447 		if (fn->might_sleep) {
11448 			verbose(env, "sleepable helper %s#%d in IRQ-disabled region\n",
11449 				func_id_name(func_id), func_id);
11450 			return -EINVAL;
11451 		}
11452 
11453 		if (in_sleepable(env) && is_storage_get_function(func_id))
11454 			env->insn_aux_data[insn_idx].storage_get_func_atomic = true;
11455 	}
11456 
11457 	meta.func_id = func_id;
11458 	/* check args */
11459 	for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
11460 		err = check_func_arg(env, i, &meta, fn, insn_idx);
11461 		if (err)
11462 			return err;
11463 	}
11464 
11465 	err = record_func_map(env, &meta, func_id, insn_idx);
11466 	if (err)
11467 		return err;
11468 
11469 	err = record_func_key(env, &meta, func_id, insn_idx);
11470 	if (err)
11471 		return err;
11472 
11473 	/* Mark slots with STACK_MISC in case of raw mode, stack offset
11474 	 * is inferred from register state.
11475 	 */
11476 	for (i = 0; i < meta.access_size; i++) {
11477 		err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
11478 				       BPF_WRITE, -1, false, false);
11479 		if (err)
11480 			return err;
11481 	}
11482 
11483 	regs = cur_regs(env);
11484 
11485 	if (meta.release_regno) {
11486 		err = -EINVAL;
11487 		/* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
11488 		 * be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
11489 		 * is safe to do directly.
11490 		 */
11491 		if (arg_type_is_dynptr(fn->arg_type[meta.release_regno - BPF_REG_1])) {
11492 			if (regs[meta.release_regno].type == CONST_PTR_TO_DYNPTR) {
11493 				verifier_bug(env, "CONST_PTR_TO_DYNPTR cannot be released");
11494 				return -EFAULT;
11495 			}
11496 			err = unmark_stack_slots_dynptr(env, &regs[meta.release_regno]);
11497 		} else if (func_id == BPF_FUNC_kptr_xchg && meta.ref_obj_id) {
11498 			u32 ref_obj_id = meta.ref_obj_id;
11499 			bool in_rcu = in_rcu_cs(env);
11500 			struct bpf_func_state *state;
11501 			struct bpf_reg_state *reg;
11502 
11503 			err = release_reference_nomark(env->cur_state, ref_obj_id);
11504 			if (!err) {
11505 				bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11506 					if (reg->ref_obj_id == ref_obj_id) {
11507 						if (in_rcu && (reg->type & MEM_ALLOC) && (reg->type & MEM_PERCPU)) {
11508 							reg->ref_obj_id = 0;
11509 							reg->type &= ~MEM_ALLOC;
11510 							reg->type |= MEM_RCU;
11511 						} else {
11512 							mark_reg_invalid(env, reg);
11513 						}
11514 					}
11515 				}));
11516 			}
11517 		} else if (meta.ref_obj_id) {
11518 			err = release_reference(env, meta.ref_obj_id);
11519 		} else if (register_is_null(&regs[meta.release_regno])) {
11520 			/* meta.ref_obj_id can only be 0 if register that is meant to be
11521 			 * released is NULL, which must be > R0.
11522 			 */
11523 			err = 0;
11524 		}
11525 		if (err) {
11526 			verbose(env, "func %s#%d reference has not been acquired before\n",
11527 				func_id_name(func_id), func_id);
11528 			return err;
11529 		}
11530 	}
11531 
11532 	switch (func_id) {
11533 	case BPF_FUNC_tail_call:
11534 		err = check_resource_leak(env, false, true, "tail_call");
11535 		if (err)
11536 			return err;
11537 		break;
11538 	case BPF_FUNC_get_local_storage:
11539 		/* check that flags argument in get_local_storage(map, flags) is 0,
11540 		 * this is required because get_local_storage() can't return an error.
11541 		 */
11542 		if (!register_is_null(&regs[BPF_REG_2])) {
11543 			verbose(env, "get_local_storage() doesn't support non-zero flags\n");
11544 			return -EINVAL;
11545 		}
11546 		break;
11547 	case BPF_FUNC_for_each_map_elem:
11548 		err = push_callback_call(env, insn, insn_idx, meta.subprogno,
11549 					 set_map_elem_callback_state);
11550 		break;
11551 	case BPF_FUNC_timer_set_callback:
11552 		err = push_callback_call(env, insn, insn_idx, meta.subprogno,
11553 					 set_timer_callback_state);
11554 		break;
11555 	case BPF_FUNC_find_vma:
11556 		err = push_callback_call(env, insn, insn_idx, meta.subprogno,
11557 					 set_find_vma_callback_state);
11558 		break;
11559 	case BPF_FUNC_snprintf:
11560 		err = check_bpf_snprintf_call(env, regs);
11561 		break;
11562 	case BPF_FUNC_loop:
11563 		update_loop_inline_state(env, meta.subprogno);
11564 		/* Verifier relies on R1 value to determine if bpf_loop() iteration
11565 		 * is finished, thus mark it precise.
11566 		 */
11567 		err = mark_chain_precision(env, BPF_REG_1);
11568 		if (err)
11569 			return err;
11570 		if (cur_func(env)->callback_depth < regs[BPF_REG_1].umax_value) {
11571 			err = push_callback_call(env, insn, insn_idx, meta.subprogno,
11572 						 set_loop_callback_state);
11573 		} else {
11574 			cur_func(env)->callback_depth = 0;
11575 			if (env->log.level & BPF_LOG_LEVEL2)
11576 				verbose(env, "frame%d bpf_loop iteration limit reached\n",
11577 					env->cur_state->curframe);
11578 		}
11579 		break;
11580 	case BPF_FUNC_dynptr_from_mem:
11581 		if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
11582 			verbose(env, "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
11583 				reg_type_str(env, regs[BPF_REG_1].type));
11584 			return -EACCES;
11585 		}
11586 		break;
11587 	case BPF_FUNC_set_retval:
11588 		if (prog_type == BPF_PROG_TYPE_LSM &&
11589 		    env->prog->expected_attach_type == BPF_LSM_CGROUP) {
11590 			if (!env->prog->aux->attach_func_proto->type) {
11591 				/* Make sure programs that attach to void
11592 				 * hooks don't try to modify return value.
11593 				 */
11594 				verbose(env, "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
11595 				return -EINVAL;
11596 			}
11597 		}
11598 		break;
11599 	case BPF_FUNC_dynptr_data:
11600 	{
11601 		struct bpf_reg_state *reg;
11602 		int id, ref_obj_id;
11603 
11604 		reg = get_dynptr_arg_reg(env, fn, regs);
11605 		if (!reg)
11606 			return -EFAULT;
11607 
11608 
11609 		if (meta.dynptr_id) {
11610 			verifier_bug(env, "meta.dynptr_id already set");
11611 			return -EFAULT;
11612 		}
11613 		if (meta.ref_obj_id) {
11614 			verifier_bug(env, "meta.ref_obj_id already set");
11615 			return -EFAULT;
11616 		}
11617 
11618 		id = dynptr_id(env, reg);
11619 		if (id < 0) {
11620 			verifier_bug(env, "failed to obtain dynptr id");
11621 			return id;
11622 		}
11623 
11624 		ref_obj_id = dynptr_ref_obj_id(env, reg);
11625 		if (ref_obj_id < 0) {
11626 			verifier_bug(env, "failed to obtain dynptr ref_obj_id");
11627 			return ref_obj_id;
11628 		}
11629 
11630 		meta.dynptr_id = id;
11631 		meta.ref_obj_id = ref_obj_id;
11632 
11633 		break;
11634 	}
11635 	case BPF_FUNC_dynptr_write:
11636 	{
11637 		enum bpf_dynptr_type dynptr_type;
11638 		struct bpf_reg_state *reg;
11639 
11640 		reg = get_dynptr_arg_reg(env, fn, regs);
11641 		if (!reg)
11642 			return -EFAULT;
11643 
11644 		dynptr_type = dynptr_get_type(env, reg);
11645 		if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
11646 			return -EFAULT;
11647 
11648 		if (dynptr_type == BPF_DYNPTR_TYPE_SKB)
11649 			/* this will trigger clear_all_pkt_pointers(), which will
11650 			 * invalidate all dynptr slices associated with the skb
11651 			 */
11652 			changes_data = true;
11653 
11654 		break;
11655 	}
11656 	case BPF_FUNC_per_cpu_ptr:
11657 	case BPF_FUNC_this_cpu_ptr:
11658 	{
11659 		struct bpf_reg_state *reg = &regs[BPF_REG_1];
11660 		const struct btf_type *type;
11661 
11662 		if (reg->type & MEM_RCU) {
11663 			type = btf_type_by_id(reg->btf, reg->btf_id);
11664 			if (!type || !btf_type_is_struct(type)) {
11665 				verbose(env, "Helper has invalid btf/btf_id in R1\n");
11666 				return -EFAULT;
11667 			}
11668 			returns_cpu_specific_alloc_ptr = true;
11669 			env->insn_aux_data[insn_idx].call_with_percpu_alloc_ptr = true;
11670 		}
11671 		break;
11672 	}
11673 	case BPF_FUNC_user_ringbuf_drain:
11674 		err = push_callback_call(env, insn, insn_idx, meta.subprogno,
11675 					 set_user_ringbuf_callback_state);
11676 		break;
11677 	}
11678 
11679 	if (err)
11680 		return err;
11681 
11682 	/* reset caller saved regs */
11683 	for (i = 0; i < CALLER_SAVED_REGS; i++) {
11684 		mark_reg_not_init(env, regs, caller_saved[i]);
11685 		check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
11686 	}
11687 
11688 	/* helper call returns 64-bit value. */
11689 	regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
11690 
11691 	/* update return register (already marked as written above) */
11692 	ret_type = fn->ret_type;
11693 	ret_flag = type_flag(ret_type);
11694 
11695 	switch (base_type(ret_type)) {
11696 	case RET_INTEGER:
11697 		/* sets type to SCALAR_VALUE */
11698 		mark_reg_unknown(env, regs, BPF_REG_0);
11699 		break;
11700 	case RET_VOID:
11701 		regs[BPF_REG_0].type = NOT_INIT;
11702 		break;
11703 	case RET_PTR_TO_MAP_VALUE:
11704 		/* There is no offset yet applied, variable or fixed */
11705 		mark_reg_known_zero(env, regs, BPF_REG_0);
11706 		/* remember map_ptr, so that check_map_access()
11707 		 * can check 'value_size' boundary of memory access
11708 		 * to map element returned from bpf_map_lookup_elem()
11709 		 */
11710 		if (meta.map_ptr == NULL) {
11711 			verifier_bug(env, "unexpected null map_ptr");
11712 			return -EFAULT;
11713 		}
11714 
11715 		if (func_id == BPF_FUNC_map_lookup_elem &&
11716 		    can_elide_value_nullness(meta.map_ptr->map_type) &&
11717 		    meta.const_map_key >= 0 &&
11718 		    meta.const_map_key < meta.map_ptr->max_entries)
11719 			ret_flag &= ~PTR_MAYBE_NULL;
11720 
11721 		regs[BPF_REG_0].map_ptr = meta.map_ptr;
11722 		regs[BPF_REG_0].map_uid = meta.map_uid;
11723 		regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
11724 		if (!type_may_be_null(ret_flag) &&
11725 		    btf_record_has_field(meta.map_ptr->record, BPF_SPIN_LOCK | BPF_RES_SPIN_LOCK)) {
11726 			regs[BPF_REG_0].id = ++env->id_gen;
11727 		}
11728 		break;
11729 	case RET_PTR_TO_SOCKET:
11730 		mark_reg_known_zero(env, regs, BPF_REG_0);
11731 		regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
11732 		break;
11733 	case RET_PTR_TO_SOCK_COMMON:
11734 		mark_reg_known_zero(env, regs, BPF_REG_0);
11735 		regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
11736 		break;
11737 	case RET_PTR_TO_TCP_SOCK:
11738 		mark_reg_known_zero(env, regs, BPF_REG_0);
11739 		regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
11740 		break;
11741 	case RET_PTR_TO_MEM:
11742 		mark_reg_known_zero(env, regs, BPF_REG_0);
11743 		regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
11744 		regs[BPF_REG_0].mem_size = meta.mem_size;
11745 		break;
11746 	case RET_PTR_TO_MEM_OR_BTF_ID:
11747 	{
11748 		const struct btf_type *t;
11749 
11750 		mark_reg_known_zero(env, regs, BPF_REG_0);
11751 		t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
11752 		if (!btf_type_is_struct(t)) {
11753 			u32 tsize;
11754 			const struct btf_type *ret;
11755 			const char *tname;
11756 
11757 			/* resolve the type size of ksym. */
11758 			ret = btf_resolve_size(meta.ret_btf, t, &tsize);
11759 			if (IS_ERR(ret)) {
11760 				tname = btf_name_by_offset(meta.ret_btf, t->name_off);
11761 				verbose(env, "unable to resolve the size of type '%s': %ld\n",
11762 					tname, PTR_ERR(ret));
11763 				return -EINVAL;
11764 			}
11765 			regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
11766 			regs[BPF_REG_0].mem_size = tsize;
11767 		} else {
11768 			if (returns_cpu_specific_alloc_ptr) {
11769 				regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC | MEM_RCU;
11770 			} else {
11771 				/* MEM_RDONLY may be carried from ret_flag, but it
11772 				 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
11773 				 * it will confuse the check of PTR_TO_BTF_ID in
11774 				 * check_mem_access().
11775 				 */
11776 				ret_flag &= ~MEM_RDONLY;
11777 				regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
11778 			}
11779 
11780 			regs[BPF_REG_0].btf = meta.ret_btf;
11781 			regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11782 		}
11783 		break;
11784 	}
11785 	case RET_PTR_TO_BTF_ID:
11786 	{
11787 		struct btf *ret_btf;
11788 		int ret_btf_id;
11789 
11790 		mark_reg_known_zero(env, regs, BPF_REG_0);
11791 		regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
11792 		if (func_id == BPF_FUNC_kptr_xchg) {
11793 			ret_btf = meta.kptr_field->kptr.btf;
11794 			ret_btf_id = meta.kptr_field->kptr.btf_id;
11795 			if (!btf_is_kernel(ret_btf)) {
11796 				regs[BPF_REG_0].type |= MEM_ALLOC;
11797 				if (meta.kptr_field->type == BPF_KPTR_PERCPU)
11798 					regs[BPF_REG_0].type |= MEM_PERCPU;
11799 			}
11800 		} else {
11801 			if (fn->ret_btf_id == BPF_PTR_POISON) {
11802 				verifier_bug(env, "func %s has non-overwritten BPF_PTR_POISON return type",
11803 					     func_id_name(func_id));
11804 				return -EFAULT;
11805 			}
11806 			ret_btf = btf_vmlinux;
11807 			ret_btf_id = *fn->ret_btf_id;
11808 		}
11809 		if (ret_btf_id == 0) {
11810 			verbose(env, "invalid return type %u of func %s#%d\n",
11811 				base_type(ret_type), func_id_name(func_id),
11812 				func_id);
11813 			return -EINVAL;
11814 		}
11815 		regs[BPF_REG_0].btf = ret_btf;
11816 		regs[BPF_REG_0].btf_id = ret_btf_id;
11817 		break;
11818 	}
11819 	default:
11820 		verbose(env, "unknown return type %u of func %s#%d\n",
11821 			base_type(ret_type), func_id_name(func_id), func_id);
11822 		return -EINVAL;
11823 	}
11824 
11825 	if (type_may_be_null(regs[BPF_REG_0].type))
11826 		regs[BPF_REG_0].id = ++env->id_gen;
11827 
11828 	if (helper_multiple_ref_obj_use(func_id, meta.map_ptr)) {
11829 		verifier_bug(env, "func %s#%d sets ref_obj_id more than once",
11830 			     func_id_name(func_id), func_id);
11831 		return -EFAULT;
11832 	}
11833 
11834 	if (is_dynptr_ref_function(func_id))
11835 		regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
11836 
11837 	if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
11838 		/* For release_reference() */
11839 		regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11840 	} else if (is_acquire_function(func_id, meta.map_ptr)) {
11841 		int id = acquire_reference(env, insn_idx);
11842 
11843 		if (id < 0)
11844 			return id;
11845 		/* For mark_ptr_or_null_reg() */
11846 		regs[BPF_REG_0].id = id;
11847 		/* For release_reference() */
11848 		regs[BPF_REG_0].ref_obj_id = id;
11849 	}
11850 
11851 	err = do_refine_retval_range(env, regs, fn->ret_type, func_id, &meta);
11852 	if (err)
11853 		return err;
11854 
11855 	err = check_map_func_compatibility(env, meta.map_ptr, func_id);
11856 	if (err)
11857 		return err;
11858 
11859 	if ((func_id == BPF_FUNC_get_stack ||
11860 	     func_id == BPF_FUNC_get_task_stack) &&
11861 	    !env->prog->has_callchain_buf) {
11862 		const char *err_str;
11863 
11864 #ifdef CONFIG_PERF_EVENTS
11865 		err = get_callchain_buffers(sysctl_perf_event_max_stack);
11866 		err_str = "cannot get callchain buffer for func %s#%d\n";
11867 #else
11868 		err = -ENOTSUPP;
11869 		err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
11870 #endif
11871 		if (err) {
11872 			verbose(env, err_str, func_id_name(func_id), func_id);
11873 			return err;
11874 		}
11875 
11876 		env->prog->has_callchain_buf = true;
11877 	}
11878 
11879 	if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
11880 		env->prog->call_get_stack = true;
11881 
11882 	if (func_id == BPF_FUNC_get_func_ip) {
11883 		if (check_get_func_ip(env))
11884 			return -ENOTSUPP;
11885 		env->prog->call_get_func_ip = true;
11886 	}
11887 
11888 	if (changes_data)
11889 		clear_all_pkt_pointers(env);
11890 	return 0;
11891 }
11892 
11893 /* mark_btf_func_reg_size() is used when the reg size is determined by
11894  * the BTF func_proto's return value size and argument.
11895  */
__mark_btf_func_reg_size(struct bpf_verifier_env * env,struct bpf_reg_state * regs,u32 regno,size_t reg_size)11896 static void __mark_btf_func_reg_size(struct bpf_verifier_env *env, struct bpf_reg_state *regs,
11897 				     u32 regno, size_t reg_size)
11898 {
11899 	struct bpf_reg_state *reg = &regs[regno];
11900 
11901 	if (regno == BPF_REG_0) {
11902 		/* Function return value */
11903 		reg->live |= REG_LIVE_WRITTEN;
11904 		reg->subreg_def = reg_size == sizeof(u64) ?
11905 			DEF_NOT_SUBREG : env->insn_idx + 1;
11906 	} else {
11907 		/* Function argument */
11908 		if (reg_size == sizeof(u64)) {
11909 			mark_insn_zext(env, reg);
11910 			mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
11911 		} else {
11912 			mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
11913 		}
11914 	}
11915 }
11916 
mark_btf_func_reg_size(struct bpf_verifier_env * env,u32 regno,size_t reg_size)11917 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
11918 				   size_t reg_size)
11919 {
11920 	return __mark_btf_func_reg_size(env, cur_regs(env), regno, reg_size);
11921 }
11922 
is_kfunc_acquire(struct bpf_kfunc_call_arg_meta * meta)11923 static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
11924 {
11925 	return meta->kfunc_flags & KF_ACQUIRE;
11926 }
11927 
is_kfunc_release(struct bpf_kfunc_call_arg_meta * meta)11928 static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
11929 {
11930 	return meta->kfunc_flags & KF_RELEASE;
11931 }
11932 
is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta * meta)11933 static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
11934 {
11935 	return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
11936 }
11937 
is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta * meta)11938 static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
11939 {
11940 	return meta->kfunc_flags & KF_SLEEPABLE;
11941 }
11942 
is_kfunc_destructive(struct bpf_kfunc_call_arg_meta * meta)11943 static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
11944 {
11945 	return meta->kfunc_flags & KF_DESTRUCTIVE;
11946 }
11947 
is_kfunc_rcu(struct bpf_kfunc_call_arg_meta * meta)11948 static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
11949 {
11950 	return meta->kfunc_flags & KF_RCU;
11951 }
11952 
is_kfunc_rcu_protected(struct bpf_kfunc_call_arg_meta * meta)11953 static bool is_kfunc_rcu_protected(struct bpf_kfunc_call_arg_meta *meta)
11954 {
11955 	return meta->kfunc_flags & KF_RCU_PROTECTED;
11956 }
11957 
is_kfunc_arg_mem_size(const struct btf * btf,const struct btf_param * arg,const struct bpf_reg_state * reg)11958 static bool is_kfunc_arg_mem_size(const struct btf *btf,
11959 				  const struct btf_param *arg,
11960 				  const struct bpf_reg_state *reg)
11961 {
11962 	const struct btf_type *t;
11963 
11964 	t = btf_type_skip_modifiers(btf, arg->type, NULL);
11965 	if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
11966 		return false;
11967 
11968 	return btf_param_match_suffix(btf, arg, "__sz");
11969 }
11970 
is_kfunc_arg_const_mem_size(const struct btf * btf,const struct btf_param * arg,const struct bpf_reg_state * reg)11971 static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
11972 					const struct btf_param *arg,
11973 					const struct bpf_reg_state *reg)
11974 {
11975 	const struct btf_type *t;
11976 
11977 	t = btf_type_skip_modifiers(btf, arg->type, NULL);
11978 	if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
11979 		return false;
11980 
11981 	return btf_param_match_suffix(btf, arg, "__szk");
11982 }
11983 
is_kfunc_arg_optional(const struct btf * btf,const struct btf_param * arg)11984 static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
11985 {
11986 	return btf_param_match_suffix(btf, arg, "__opt");
11987 }
11988 
is_kfunc_arg_constant(const struct btf * btf,const struct btf_param * arg)11989 static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
11990 {
11991 	return btf_param_match_suffix(btf, arg, "__k");
11992 }
11993 
is_kfunc_arg_ignore(const struct btf * btf,const struct btf_param * arg)11994 static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
11995 {
11996 	return btf_param_match_suffix(btf, arg, "__ign");
11997 }
11998 
is_kfunc_arg_map(const struct btf * btf,const struct btf_param * arg)11999 static bool is_kfunc_arg_map(const struct btf *btf, const struct btf_param *arg)
12000 {
12001 	return btf_param_match_suffix(btf, arg, "__map");
12002 }
12003 
is_kfunc_arg_alloc_obj(const struct btf * btf,const struct btf_param * arg)12004 static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
12005 {
12006 	return btf_param_match_suffix(btf, arg, "__alloc");
12007 }
12008 
is_kfunc_arg_uninit(const struct btf * btf,const struct btf_param * arg)12009 static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
12010 {
12011 	return btf_param_match_suffix(btf, arg, "__uninit");
12012 }
12013 
is_kfunc_arg_refcounted_kptr(const struct btf * btf,const struct btf_param * arg)12014 static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
12015 {
12016 	return btf_param_match_suffix(btf, arg, "__refcounted_kptr");
12017 }
12018 
is_kfunc_arg_nullable(const struct btf * btf,const struct btf_param * arg)12019 static bool is_kfunc_arg_nullable(const struct btf *btf, const struct btf_param *arg)
12020 {
12021 	return btf_param_match_suffix(btf, arg, "__nullable");
12022 }
12023 
is_kfunc_arg_const_str(const struct btf * btf,const struct btf_param * arg)12024 static bool is_kfunc_arg_const_str(const struct btf *btf, const struct btf_param *arg)
12025 {
12026 	return btf_param_match_suffix(btf, arg, "__str");
12027 }
12028 
is_kfunc_arg_irq_flag(const struct btf * btf,const struct btf_param * arg)12029 static bool is_kfunc_arg_irq_flag(const struct btf *btf, const struct btf_param *arg)
12030 {
12031 	return btf_param_match_suffix(btf, arg, "__irq_flag");
12032 }
12033 
is_kfunc_arg_prog(const struct btf * btf,const struct btf_param * arg)12034 static bool is_kfunc_arg_prog(const struct btf *btf, const struct btf_param *arg)
12035 {
12036 	return btf_param_match_suffix(btf, arg, "__prog");
12037 }
12038 
is_kfunc_arg_scalar_with_name(const struct btf * btf,const struct btf_param * arg,const char * name)12039 static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
12040 					  const struct btf_param *arg,
12041 					  const char *name)
12042 {
12043 	int len, target_len = strlen(name);
12044 	const char *param_name;
12045 
12046 	param_name = btf_name_by_offset(btf, arg->name_off);
12047 	if (str_is_empty(param_name))
12048 		return false;
12049 	len = strlen(param_name);
12050 	if (len != target_len)
12051 		return false;
12052 	if (strcmp(param_name, name))
12053 		return false;
12054 
12055 	return true;
12056 }
12057 
12058 enum {
12059 	KF_ARG_DYNPTR_ID,
12060 	KF_ARG_LIST_HEAD_ID,
12061 	KF_ARG_LIST_NODE_ID,
12062 	KF_ARG_RB_ROOT_ID,
12063 	KF_ARG_RB_NODE_ID,
12064 	KF_ARG_WORKQUEUE_ID,
12065 	KF_ARG_RES_SPIN_LOCK_ID,
12066 };
12067 
12068 BTF_ID_LIST(kf_arg_btf_ids)
BTF_ID(struct,bpf_dynptr)12069 BTF_ID(struct, bpf_dynptr)
12070 BTF_ID(struct, bpf_list_head)
12071 BTF_ID(struct, bpf_list_node)
12072 BTF_ID(struct, bpf_rb_root)
12073 BTF_ID(struct, bpf_rb_node)
12074 BTF_ID(struct, bpf_wq)
12075 BTF_ID(struct, bpf_res_spin_lock)
12076 
12077 static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
12078 				    const struct btf_param *arg, int type)
12079 {
12080 	const struct btf_type *t;
12081 	u32 res_id;
12082 
12083 	t = btf_type_skip_modifiers(btf, arg->type, NULL);
12084 	if (!t)
12085 		return false;
12086 	if (!btf_type_is_ptr(t))
12087 		return false;
12088 	t = btf_type_skip_modifiers(btf, t->type, &res_id);
12089 	if (!t)
12090 		return false;
12091 	return btf_types_are_same(btf, res_id, btf_vmlinux, kf_arg_btf_ids[type]);
12092 }
12093 
is_kfunc_arg_dynptr(const struct btf * btf,const struct btf_param * arg)12094 static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
12095 {
12096 	return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_DYNPTR_ID);
12097 }
12098 
is_kfunc_arg_list_head(const struct btf * btf,const struct btf_param * arg)12099 static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
12100 {
12101 	return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_HEAD_ID);
12102 }
12103 
is_kfunc_arg_list_node(const struct btf * btf,const struct btf_param * arg)12104 static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
12105 {
12106 	return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_LIST_NODE_ID);
12107 }
12108 
is_kfunc_arg_rbtree_root(const struct btf * btf,const struct btf_param * arg)12109 static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
12110 {
12111 	return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_ROOT_ID);
12112 }
12113 
is_kfunc_arg_rbtree_node(const struct btf * btf,const struct btf_param * arg)12114 static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
12115 {
12116 	return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RB_NODE_ID);
12117 }
12118 
is_kfunc_arg_wq(const struct btf * btf,const struct btf_param * arg)12119 static bool is_kfunc_arg_wq(const struct btf *btf, const struct btf_param *arg)
12120 {
12121 	return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_WORKQUEUE_ID);
12122 }
12123 
is_kfunc_arg_res_spin_lock(const struct btf * btf,const struct btf_param * arg)12124 static bool is_kfunc_arg_res_spin_lock(const struct btf *btf, const struct btf_param *arg)
12125 {
12126 	return __is_kfunc_ptr_arg_type(btf, arg, KF_ARG_RES_SPIN_LOCK_ID);
12127 }
12128 
is_rbtree_node_type(const struct btf_type * t)12129 static bool is_rbtree_node_type(const struct btf_type *t)
12130 {
12131 	return t == btf_type_by_id(btf_vmlinux, kf_arg_btf_ids[KF_ARG_RB_NODE_ID]);
12132 }
12133 
is_list_node_type(const struct btf_type * t)12134 static bool is_list_node_type(const struct btf_type *t)
12135 {
12136 	return t == btf_type_by_id(btf_vmlinux, kf_arg_btf_ids[KF_ARG_LIST_NODE_ID]);
12137 }
12138 
is_kfunc_arg_callback(struct bpf_verifier_env * env,const struct btf * btf,const struct btf_param * arg)12139 static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
12140 				  const struct btf_param *arg)
12141 {
12142 	const struct btf_type *t;
12143 
12144 	t = btf_type_resolve_func_ptr(btf, arg->type, NULL);
12145 	if (!t)
12146 		return false;
12147 
12148 	return true;
12149 }
12150 
12151 /* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
__btf_type_is_scalar_struct(struct bpf_verifier_env * env,const struct btf * btf,const struct btf_type * t,int rec)12152 static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
12153 					const struct btf *btf,
12154 					const struct btf_type *t, int rec)
12155 {
12156 	const struct btf_type *member_type;
12157 	const struct btf_member *member;
12158 	u32 i;
12159 
12160 	if (!btf_type_is_struct(t))
12161 		return false;
12162 
12163 	for_each_member(i, t, member) {
12164 		const struct btf_array *array;
12165 
12166 		member_type = btf_type_skip_modifiers(btf, member->type, NULL);
12167 		if (btf_type_is_struct(member_type)) {
12168 			if (rec >= 3) {
12169 				verbose(env, "max struct nesting depth exceeded\n");
12170 				return false;
12171 			}
12172 			if (!__btf_type_is_scalar_struct(env, btf, member_type, rec + 1))
12173 				return false;
12174 			continue;
12175 		}
12176 		if (btf_type_is_array(member_type)) {
12177 			array = btf_array(member_type);
12178 			if (!array->nelems)
12179 				return false;
12180 			member_type = btf_type_skip_modifiers(btf, array->type, NULL);
12181 			if (!btf_type_is_scalar(member_type))
12182 				return false;
12183 			continue;
12184 		}
12185 		if (!btf_type_is_scalar(member_type))
12186 			return false;
12187 	}
12188 	return true;
12189 }
12190 
12191 enum kfunc_ptr_arg_type {
12192 	KF_ARG_PTR_TO_CTX,
12193 	KF_ARG_PTR_TO_ALLOC_BTF_ID,    /* Allocated object */
12194 	KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
12195 	KF_ARG_PTR_TO_DYNPTR,
12196 	KF_ARG_PTR_TO_ITER,
12197 	KF_ARG_PTR_TO_LIST_HEAD,
12198 	KF_ARG_PTR_TO_LIST_NODE,
12199 	KF_ARG_PTR_TO_BTF_ID,	       /* Also covers reg2btf_ids conversions */
12200 	KF_ARG_PTR_TO_MEM,
12201 	KF_ARG_PTR_TO_MEM_SIZE,	       /* Size derived from next argument, skip it */
12202 	KF_ARG_PTR_TO_CALLBACK,
12203 	KF_ARG_PTR_TO_RB_ROOT,
12204 	KF_ARG_PTR_TO_RB_NODE,
12205 	KF_ARG_PTR_TO_NULL,
12206 	KF_ARG_PTR_TO_CONST_STR,
12207 	KF_ARG_PTR_TO_MAP,
12208 	KF_ARG_PTR_TO_WORKQUEUE,
12209 	KF_ARG_PTR_TO_IRQ_FLAG,
12210 	KF_ARG_PTR_TO_RES_SPIN_LOCK,
12211 };
12212 
12213 enum special_kfunc_type {
12214 	KF_bpf_obj_new_impl,
12215 	KF_bpf_obj_drop_impl,
12216 	KF_bpf_refcount_acquire_impl,
12217 	KF_bpf_list_push_front_impl,
12218 	KF_bpf_list_push_back_impl,
12219 	KF_bpf_list_pop_front,
12220 	KF_bpf_list_pop_back,
12221 	KF_bpf_list_front,
12222 	KF_bpf_list_back,
12223 	KF_bpf_cast_to_kern_ctx,
12224 	KF_bpf_rdonly_cast,
12225 	KF_bpf_rcu_read_lock,
12226 	KF_bpf_rcu_read_unlock,
12227 	KF_bpf_rbtree_remove,
12228 	KF_bpf_rbtree_add_impl,
12229 	KF_bpf_rbtree_first,
12230 	KF_bpf_rbtree_root,
12231 	KF_bpf_rbtree_left,
12232 	KF_bpf_rbtree_right,
12233 	KF_bpf_dynptr_from_skb,
12234 	KF_bpf_dynptr_from_xdp,
12235 	KF_bpf_dynptr_slice,
12236 	KF_bpf_dynptr_slice_rdwr,
12237 	KF_bpf_dynptr_clone,
12238 	KF_bpf_percpu_obj_new_impl,
12239 	KF_bpf_percpu_obj_drop_impl,
12240 	KF_bpf_throw,
12241 	KF_bpf_wq_set_callback_impl,
12242 	KF_bpf_preempt_disable,
12243 	KF_bpf_preempt_enable,
12244 	KF_bpf_iter_css_task_new,
12245 	KF_bpf_session_cookie,
12246 	KF_bpf_get_kmem_cache,
12247 	KF_bpf_local_irq_save,
12248 	KF_bpf_local_irq_restore,
12249 	KF_bpf_iter_num_new,
12250 	KF_bpf_iter_num_next,
12251 	KF_bpf_iter_num_destroy,
12252 	KF_bpf_set_dentry_xattr,
12253 	KF_bpf_remove_dentry_xattr,
12254 	KF_bpf_res_spin_lock,
12255 	KF_bpf_res_spin_unlock,
12256 	KF_bpf_res_spin_lock_irqsave,
12257 	KF_bpf_res_spin_unlock_irqrestore,
12258 	KF___bpf_trap,
12259 };
12260 
12261 BTF_ID_LIST(special_kfunc_list)
BTF_ID(func,bpf_obj_new_impl)12262 BTF_ID(func, bpf_obj_new_impl)
12263 BTF_ID(func, bpf_obj_drop_impl)
12264 BTF_ID(func, bpf_refcount_acquire_impl)
12265 BTF_ID(func, bpf_list_push_front_impl)
12266 BTF_ID(func, bpf_list_push_back_impl)
12267 BTF_ID(func, bpf_list_pop_front)
12268 BTF_ID(func, bpf_list_pop_back)
12269 BTF_ID(func, bpf_list_front)
12270 BTF_ID(func, bpf_list_back)
12271 BTF_ID(func, bpf_cast_to_kern_ctx)
12272 BTF_ID(func, bpf_rdonly_cast)
12273 BTF_ID(func, bpf_rcu_read_lock)
12274 BTF_ID(func, bpf_rcu_read_unlock)
12275 BTF_ID(func, bpf_rbtree_remove)
12276 BTF_ID(func, bpf_rbtree_add_impl)
12277 BTF_ID(func, bpf_rbtree_first)
12278 BTF_ID(func, bpf_rbtree_root)
12279 BTF_ID(func, bpf_rbtree_left)
12280 BTF_ID(func, bpf_rbtree_right)
12281 #ifdef CONFIG_NET
12282 BTF_ID(func, bpf_dynptr_from_skb)
12283 BTF_ID(func, bpf_dynptr_from_xdp)
12284 #else
12285 BTF_ID_UNUSED
12286 BTF_ID_UNUSED
12287 #endif
12288 BTF_ID(func, bpf_dynptr_slice)
12289 BTF_ID(func, bpf_dynptr_slice_rdwr)
12290 BTF_ID(func, bpf_dynptr_clone)
12291 BTF_ID(func, bpf_percpu_obj_new_impl)
12292 BTF_ID(func, bpf_percpu_obj_drop_impl)
12293 BTF_ID(func, bpf_throw)
12294 BTF_ID(func, bpf_wq_set_callback_impl)
12295 BTF_ID(func, bpf_preempt_disable)
12296 BTF_ID(func, bpf_preempt_enable)
12297 #ifdef CONFIG_CGROUPS
12298 BTF_ID(func, bpf_iter_css_task_new)
12299 #else
12300 BTF_ID_UNUSED
12301 #endif
12302 #ifdef CONFIG_BPF_EVENTS
12303 BTF_ID(func, bpf_session_cookie)
12304 #else
12305 BTF_ID_UNUSED
12306 #endif
12307 BTF_ID(func, bpf_get_kmem_cache)
12308 BTF_ID(func, bpf_local_irq_save)
12309 BTF_ID(func, bpf_local_irq_restore)
12310 BTF_ID(func, bpf_iter_num_new)
12311 BTF_ID(func, bpf_iter_num_next)
12312 BTF_ID(func, bpf_iter_num_destroy)
12313 #ifdef CONFIG_BPF_LSM
12314 BTF_ID(func, bpf_set_dentry_xattr)
12315 BTF_ID(func, bpf_remove_dentry_xattr)
12316 #else
12317 BTF_ID_UNUSED
12318 BTF_ID_UNUSED
12319 #endif
12320 BTF_ID(func, bpf_res_spin_lock)
12321 BTF_ID(func, bpf_res_spin_unlock)
12322 BTF_ID(func, bpf_res_spin_lock_irqsave)
12323 BTF_ID(func, bpf_res_spin_unlock_irqrestore)
12324 BTF_ID(func, __bpf_trap)
12325 
12326 static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
12327 {
12328 	if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
12329 	    meta->arg_owning_ref) {
12330 		return false;
12331 	}
12332 
12333 	return meta->kfunc_flags & KF_RET_NULL;
12334 }
12335 
is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta * meta)12336 static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
12337 {
12338 	return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
12339 }
12340 
is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta * meta)12341 static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
12342 {
12343 	return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
12344 }
12345 
is_kfunc_bpf_preempt_disable(struct bpf_kfunc_call_arg_meta * meta)12346 static bool is_kfunc_bpf_preempt_disable(struct bpf_kfunc_call_arg_meta *meta)
12347 {
12348 	return meta->func_id == special_kfunc_list[KF_bpf_preempt_disable];
12349 }
12350 
is_kfunc_bpf_preempt_enable(struct bpf_kfunc_call_arg_meta * meta)12351 static bool is_kfunc_bpf_preempt_enable(struct bpf_kfunc_call_arg_meta *meta)
12352 {
12353 	return meta->func_id == special_kfunc_list[KF_bpf_preempt_enable];
12354 }
12355 
12356 static enum kfunc_ptr_arg_type
get_kfunc_ptr_arg_type(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,const struct btf_type * t,const struct btf_type * ref_t,const char * ref_tname,const struct btf_param * args,int argno,int nargs)12357 get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
12358 		       struct bpf_kfunc_call_arg_meta *meta,
12359 		       const struct btf_type *t, const struct btf_type *ref_t,
12360 		       const char *ref_tname, const struct btf_param *args,
12361 		       int argno, int nargs)
12362 {
12363 	u32 regno = argno + 1;
12364 	struct bpf_reg_state *regs = cur_regs(env);
12365 	struct bpf_reg_state *reg = &regs[regno];
12366 	bool arg_mem_size = false;
12367 
12368 	if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
12369 		return KF_ARG_PTR_TO_CTX;
12370 
12371 	/* In this function, we verify the kfunc's BTF as per the argument type,
12372 	 * leaving the rest of the verification with respect to the register
12373 	 * type to our caller. When a set of conditions hold in the BTF type of
12374 	 * arguments, we resolve it to a known kfunc_ptr_arg_type.
12375 	 */
12376 	if (btf_is_prog_ctx_type(&env->log, meta->btf, t, resolve_prog_type(env->prog), argno))
12377 		return KF_ARG_PTR_TO_CTX;
12378 
12379 	if (is_kfunc_arg_nullable(meta->btf, &args[argno]) && register_is_null(reg))
12380 		return KF_ARG_PTR_TO_NULL;
12381 
12382 	if (is_kfunc_arg_alloc_obj(meta->btf, &args[argno]))
12383 		return KF_ARG_PTR_TO_ALLOC_BTF_ID;
12384 
12385 	if (is_kfunc_arg_refcounted_kptr(meta->btf, &args[argno]))
12386 		return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
12387 
12388 	if (is_kfunc_arg_dynptr(meta->btf, &args[argno]))
12389 		return KF_ARG_PTR_TO_DYNPTR;
12390 
12391 	if (is_kfunc_arg_iter(meta, argno, &args[argno]))
12392 		return KF_ARG_PTR_TO_ITER;
12393 
12394 	if (is_kfunc_arg_list_head(meta->btf, &args[argno]))
12395 		return KF_ARG_PTR_TO_LIST_HEAD;
12396 
12397 	if (is_kfunc_arg_list_node(meta->btf, &args[argno]))
12398 		return KF_ARG_PTR_TO_LIST_NODE;
12399 
12400 	if (is_kfunc_arg_rbtree_root(meta->btf, &args[argno]))
12401 		return KF_ARG_PTR_TO_RB_ROOT;
12402 
12403 	if (is_kfunc_arg_rbtree_node(meta->btf, &args[argno]))
12404 		return KF_ARG_PTR_TO_RB_NODE;
12405 
12406 	if (is_kfunc_arg_const_str(meta->btf, &args[argno]))
12407 		return KF_ARG_PTR_TO_CONST_STR;
12408 
12409 	if (is_kfunc_arg_map(meta->btf, &args[argno]))
12410 		return KF_ARG_PTR_TO_MAP;
12411 
12412 	if (is_kfunc_arg_wq(meta->btf, &args[argno]))
12413 		return KF_ARG_PTR_TO_WORKQUEUE;
12414 
12415 	if (is_kfunc_arg_irq_flag(meta->btf, &args[argno]))
12416 		return KF_ARG_PTR_TO_IRQ_FLAG;
12417 
12418 	if (is_kfunc_arg_res_spin_lock(meta->btf, &args[argno]))
12419 		return KF_ARG_PTR_TO_RES_SPIN_LOCK;
12420 
12421 	if ((base_type(reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(reg->type)])) {
12422 		if (!btf_type_is_struct(ref_t)) {
12423 			verbose(env, "kernel function %s args#%d pointer type %s %s is not supported\n",
12424 				meta->func_name, argno, btf_type_str(ref_t), ref_tname);
12425 			return -EINVAL;
12426 		}
12427 		return KF_ARG_PTR_TO_BTF_ID;
12428 	}
12429 
12430 	if (is_kfunc_arg_callback(env, meta->btf, &args[argno]))
12431 		return KF_ARG_PTR_TO_CALLBACK;
12432 
12433 	if (argno + 1 < nargs &&
12434 	    (is_kfunc_arg_mem_size(meta->btf, &args[argno + 1], &regs[regno + 1]) ||
12435 	     is_kfunc_arg_const_mem_size(meta->btf, &args[argno + 1], &regs[regno + 1])))
12436 		arg_mem_size = true;
12437 
12438 	/* This is the catch all argument type of register types supported by
12439 	 * check_helper_mem_access. However, we only allow when argument type is
12440 	 * pointer to scalar, or struct composed (recursively) of scalars. When
12441 	 * arg_mem_size is true, the pointer can be void *.
12442 	 */
12443 	if (!btf_type_is_scalar(ref_t) && !__btf_type_is_scalar_struct(env, meta->btf, ref_t, 0) &&
12444 	    (arg_mem_size ? !btf_type_is_void(ref_t) : 1)) {
12445 		verbose(env, "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
12446 			argno, btf_type_str(ref_t), ref_tname, arg_mem_size ? "void, " : "");
12447 		return -EINVAL;
12448 	}
12449 	return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
12450 }
12451 
process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env * env,struct bpf_reg_state * reg,const struct btf_type * ref_t,const char * ref_tname,u32 ref_id,struct bpf_kfunc_call_arg_meta * meta,int argno)12452 static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
12453 					struct bpf_reg_state *reg,
12454 					const struct btf_type *ref_t,
12455 					const char *ref_tname, u32 ref_id,
12456 					struct bpf_kfunc_call_arg_meta *meta,
12457 					int argno)
12458 {
12459 	const struct btf_type *reg_ref_t;
12460 	bool strict_type_match = false;
12461 	const struct btf *reg_btf;
12462 	const char *reg_ref_tname;
12463 	bool taking_projection;
12464 	bool struct_same;
12465 	u32 reg_ref_id;
12466 
12467 	if (base_type(reg->type) == PTR_TO_BTF_ID) {
12468 		reg_btf = reg->btf;
12469 		reg_ref_id = reg->btf_id;
12470 	} else {
12471 		reg_btf = btf_vmlinux;
12472 		reg_ref_id = *reg2btf_ids[base_type(reg->type)];
12473 	}
12474 
12475 	/* Enforce strict type matching for calls to kfuncs that are acquiring
12476 	 * or releasing a reference, or are no-cast aliases. We do _not_
12477 	 * enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
12478 	 * as we want to enable BPF programs to pass types that are bitwise
12479 	 * equivalent without forcing them to explicitly cast with something
12480 	 * like bpf_cast_to_kern_ctx().
12481 	 *
12482 	 * For example, say we had a type like the following:
12483 	 *
12484 	 * struct bpf_cpumask {
12485 	 *	cpumask_t cpumask;
12486 	 *	refcount_t usage;
12487 	 * };
12488 	 *
12489 	 * Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
12490 	 * to a struct cpumask, so it would be safe to pass a struct
12491 	 * bpf_cpumask * to a kfunc expecting a struct cpumask *.
12492 	 *
12493 	 * The philosophy here is similar to how we allow scalars of different
12494 	 * types to be passed to kfuncs as long as the size is the same. The
12495 	 * only difference here is that we're simply allowing
12496 	 * btf_struct_ids_match() to walk the struct at the 0th offset, and
12497 	 * resolve types.
12498 	 */
12499 	if ((is_kfunc_release(meta) && reg->ref_obj_id) ||
12500 	    btf_type_ids_nocast_alias(&env->log, reg_btf, reg_ref_id, meta->btf, ref_id))
12501 		strict_type_match = true;
12502 
12503 	WARN_ON_ONCE(is_kfunc_release(meta) &&
12504 		     (reg->off || !tnum_is_const(reg->var_off) ||
12505 		      reg->var_off.value));
12506 
12507 	reg_ref_t = btf_type_skip_modifiers(reg_btf, reg_ref_id, &reg_ref_id);
12508 	reg_ref_tname = btf_name_by_offset(reg_btf, reg_ref_t->name_off);
12509 	struct_same = btf_struct_ids_match(&env->log, reg_btf, reg_ref_id, reg->off, meta->btf, ref_id, strict_type_match);
12510 	/* If kfunc is accepting a projection type (ie. __sk_buff), it cannot
12511 	 * actually use it -- it must cast to the underlying type. So we allow
12512 	 * caller to pass in the underlying type.
12513 	 */
12514 	taking_projection = btf_is_projection_of(ref_tname, reg_ref_tname);
12515 	if (!taking_projection && !struct_same) {
12516 		verbose(env, "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
12517 			meta->func_name, argno, btf_type_str(ref_t), ref_tname, argno + 1,
12518 			btf_type_str(reg_ref_t), reg_ref_tname);
12519 		return -EINVAL;
12520 	}
12521 	return 0;
12522 }
12523 
process_irq_flag(struct bpf_verifier_env * env,int regno,struct bpf_kfunc_call_arg_meta * meta)12524 static int process_irq_flag(struct bpf_verifier_env *env, int regno,
12525 			     struct bpf_kfunc_call_arg_meta *meta)
12526 {
12527 	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
12528 	int err, kfunc_class = IRQ_NATIVE_KFUNC;
12529 	bool irq_save;
12530 
12531 	if (meta->func_id == special_kfunc_list[KF_bpf_local_irq_save] ||
12532 	    meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave]) {
12533 		irq_save = true;
12534 		if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave])
12535 			kfunc_class = IRQ_LOCK_KFUNC;
12536 	} else if (meta->func_id == special_kfunc_list[KF_bpf_local_irq_restore] ||
12537 		   meta->func_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore]) {
12538 		irq_save = false;
12539 		if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore])
12540 			kfunc_class = IRQ_LOCK_KFUNC;
12541 	} else {
12542 		verifier_bug(env, "unknown irq flags kfunc");
12543 		return -EFAULT;
12544 	}
12545 
12546 	if (irq_save) {
12547 		if (!is_irq_flag_reg_valid_uninit(env, reg)) {
12548 			verbose(env, "expected uninitialized irq flag as arg#%d\n", regno - 1);
12549 			return -EINVAL;
12550 		}
12551 
12552 		err = check_mem_access(env, env->insn_idx, regno, 0, BPF_DW, BPF_WRITE, -1, false, false);
12553 		if (err)
12554 			return err;
12555 
12556 		err = mark_stack_slot_irq_flag(env, meta, reg, env->insn_idx, kfunc_class);
12557 		if (err)
12558 			return err;
12559 	} else {
12560 		err = is_irq_flag_reg_valid_init(env, reg);
12561 		if (err) {
12562 			verbose(env, "expected an initialized irq flag as arg#%d\n", regno - 1);
12563 			return err;
12564 		}
12565 
12566 		err = mark_irq_flag_read(env, reg);
12567 		if (err)
12568 			return err;
12569 
12570 		err = unmark_stack_slot_irq_flag(env, reg, kfunc_class);
12571 		if (err)
12572 			return err;
12573 	}
12574 	return 0;
12575 }
12576 
12577 
ref_set_non_owning(struct bpf_verifier_env * env,struct bpf_reg_state * reg)12578 static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
12579 {
12580 	struct btf_record *rec = reg_btf_record(reg);
12581 
12582 	if (!env->cur_state->active_locks) {
12583 		verifier_bug(env, "%s w/o active lock", __func__);
12584 		return -EFAULT;
12585 	}
12586 
12587 	if (type_flag(reg->type) & NON_OWN_REF) {
12588 		verifier_bug(env, "NON_OWN_REF already set");
12589 		return -EFAULT;
12590 	}
12591 
12592 	reg->type |= NON_OWN_REF;
12593 	if (rec->refcount_off >= 0)
12594 		reg->type |= MEM_RCU;
12595 
12596 	return 0;
12597 }
12598 
ref_convert_owning_non_owning(struct bpf_verifier_env * env,u32 ref_obj_id)12599 static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
12600 {
12601 	struct bpf_verifier_state *state = env->cur_state;
12602 	struct bpf_func_state *unused;
12603 	struct bpf_reg_state *reg;
12604 	int i;
12605 
12606 	if (!ref_obj_id) {
12607 		verifier_bug(env, "ref_obj_id is zero for owning -> non-owning conversion");
12608 		return -EFAULT;
12609 	}
12610 
12611 	for (i = 0; i < state->acquired_refs; i++) {
12612 		if (state->refs[i].id != ref_obj_id)
12613 			continue;
12614 
12615 		/* Clear ref_obj_id here so release_reference doesn't clobber
12616 		 * the whole reg
12617 		 */
12618 		bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
12619 			if (reg->ref_obj_id == ref_obj_id) {
12620 				reg->ref_obj_id = 0;
12621 				ref_set_non_owning(env, reg);
12622 			}
12623 		}));
12624 		return 0;
12625 	}
12626 
12627 	verifier_bug(env, "ref state missing for ref_obj_id");
12628 	return -EFAULT;
12629 }
12630 
12631 /* Implementation details:
12632  *
12633  * Each register points to some region of memory, which we define as an
12634  * allocation. Each allocation may embed a bpf_spin_lock which protects any
12635  * special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
12636  * allocation. The lock and the data it protects are colocated in the same
12637  * memory region.
12638  *
12639  * Hence, everytime a register holds a pointer value pointing to such
12640  * allocation, the verifier preserves a unique reg->id for it.
12641  *
12642  * The verifier remembers the lock 'ptr' and the lock 'id' whenever
12643  * bpf_spin_lock is called.
12644  *
12645  * To enable this, lock state in the verifier captures two values:
12646  *	active_lock.ptr = Register's type specific pointer
12647  *	active_lock.id  = A unique ID for each register pointer value
12648  *
12649  * Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
12650  * supported register types.
12651  *
12652  * The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
12653  * allocated objects is the reg->btf pointer.
12654  *
12655  * The active_lock.id is non-unique for maps supporting direct_value_addr, as we
12656  * can establish the provenance of the map value statically for each distinct
12657  * lookup into such maps. They always contain a single map value hence unique
12658  * IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
12659  *
12660  * So, in case of global variables, they use array maps with max_entries = 1,
12661  * hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
12662  * into the same map value as max_entries is 1, as described above).
12663  *
12664  * In case of inner map lookups, the inner map pointer has same map_ptr as the
12665  * outer map pointer (in verifier context), but each lookup into an inner map
12666  * assigns a fresh reg->id to the lookup, so while lookups into distinct inner
12667  * maps from the same outer map share the same map_ptr as active_lock.ptr, they
12668  * will get different reg->id assigned to each lookup, hence different
12669  * active_lock.id.
12670  *
12671  * In case of allocated objects, active_lock.ptr is the reg->btf, and the
12672  * reg->id is a unique ID preserved after the NULL pointer check on the pointer
12673  * returned from bpf_obj_new. Each allocation receives a new reg->id.
12674  */
check_reg_allocation_locked(struct bpf_verifier_env * env,struct bpf_reg_state * reg)12675 static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
12676 {
12677 	struct bpf_reference_state *s;
12678 	void *ptr;
12679 	u32 id;
12680 
12681 	switch ((int)reg->type) {
12682 	case PTR_TO_MAP_VALUE:
12683 		ptr = reg->map_ptr;
12684 		break;
12685 	case PTR_TO_BTF_ID | MEM_ALLOC:
12686 		ptr = reg->btf;
12687 		break;
12688 	default:
12689 		verifier_bug(env, "unknown reg type for lock check");
12690 		return -EFAULT;
12691 	}
12692 	id = reg->id;
12693 
12694 	if (!env->cur_state->active_locks)
12695 		return -EINVAL;
12696 	s = find_lock_state(env->cur_state, REF_TYPE_LOCK_MASK, id, ptr);
12697 	if (!s) {
12698 		verbose(env, "held lock and object are not in the same allocation\n");
12699 		return -EINVAL;
12700 	}
12701 	return 0;
12702 }
12703 
is_bpf_list_api_kfunc(u32 btf_id)12704 static bool is_bpf_list_api_kfunc(u32 btf_id)
12705 {
12706 	return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
12707 	       btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
12708 	       btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
12709 	       btf_id == special_kfunc_list[KF_bpf_list_pop_back] ||
12710 	       btf_id == special_kfunc_list[KF_bpf_list_front] ||
12711 	       btf_id == special_kfunc_list[KF_bpf_list_back];
12712 }
12713 
is_bpf_rbtree_api_kfunc(u32 btf_id)12714 static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
12715 {
12716 	return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
12717 	       btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
12718 	       btf_id == special_kfunc_list[KF_bpf_rbtree_first] ||
12719 	       btf_id == special_kfunc_list[KF_bpf_rbtree_root] ||
12720 	       btf_id == special_kfunc_list[KF_bpf_rbtree_left] ||
12721 	       btf_id == special_kfunc_list[KF_bpf_rbtree_right];
12722 }
12723 
is_bpf_iter_num_api_kfunc(u32 btf_id)12724 static bool is_bpf_iter_num_api_kfunc(u32 btf_id)
12725 {
12726 	return btf_id == special_kfunc_list[KF_bpf_iter_num_new] ||
12727 	       btf_id == special_kfunc_list[KF_bpf_iter_num_next] ||
12728 	       btf_id == special_kfunc_list[KF_bpf_iter_num_destroy];
12729 }
12730 
is_bpf_graph_api_kfunc(u32 btf_id)12731 static bool is_bpf_graph_api_kfunc(u32 btf_id)
12732 {
12733 	return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
12734 	       btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
12735 }
12736 
is_bpf_res_spin_lock_kfunc(u32 btf_id)12737 static bool is_bpf_res_spin_lock_kfunc(u32 btf_id)
12738 {
12739 	return btf_id == special_kfunc_list[KF_bpf_res_spin_lock] ||
12740 	       btf_id == special_kfunc_list[KF_bpf_res_spin_unlock] ||
12741 	       btf_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave] ||
12742 	       btf_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore];
12743 }
12744 
kfunc_spin_allowed(u32 btf_id)12745 static bool kfunc_spin_allowed(u32 btf_id)
12746 {
12747 	return is_bpf_graph_api_kfunc(btf_id) || is_bpf_iter_num_api_kfunc(btf_id) ||
12748 	       is_bpf_res_spin_lock_kfunc(btf_id);
12749 }
12750 
is_sync_callback_calling_kfunc(u32 btf_id)12751 static bool is_sync_callback_calling_kfunc(u32 btf_id)
12752 {
12753 	return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
12754 }
12755 
is_async_callback_calling_kfunc(u32 btf_id)12756 static bool is_async_callback_calling_kfunc(u32 btf_id)
12757 {
12758 	return btf_id == special_kfunc_list[KF_bpf_wq_set_callback_impl];
12759 }
12760 
is_bpf_throw_kfunc(struct bpf_insn * insn)12761 static bool is_bpf_throw_kfunc(struct bpf_insn *insn)
12762 {
12763 	return bpf_pseudo_kfunc_call(insn) && insn->off == 0 &&
12764 	       insn->imm == special_kfunc_list[KF_bpf_throw];
12765 }
12766 
is_bpf_wq_set_callback_impl_kfunc(u32 btf_id)12767 static bool is_bpf_wq_set_callback_impl_kfunc(u32 btf_id)
12768 {
12769 	return btf_id == special_kfunc_list[KF_bpf_wq_set_callback_impl];
12770 }
12771 
is_callback_calling_kfunc(u32 btf_id)12772 static bool is_callback_calling_kfunc(u32 btf_id)
12773 {
12774 	return is_sync_callback_calling_kfunc(btf_id) ||
12775 	       is_async_callback_calling_kfunc(btf_id);
12776 }
12777 
is_rbtree_lock_required_kfunc(u32 btf_id)12778 static bool is_rbtree_lock_required_kfunc(u32 btf_id)
12779 {
12780 	return is_bpf_rbtree_api_kfunc(btf_id);
12781 }
12782 
check_kfunc_is_graph_root_api(struct bpf_verifier_env * env,enum btf_field_type head_field_type,u32 kfunc_btf_id)12783 static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
12784 					  enum btf_field_type head_field_type,
12785 					  u32 kfunc_btf_id)
12786 {
12787 	bool ret;
12788 
12789 	switch (head_field_type) {
12790 	case BPF_LIST_HEAD:
12791 		ret = is_bpf_list_api_kfunc(kfunc_btf_id);
12792 		break;
12793 	case BPF_RB_ROOT:
12794 		ret = is_bpf_rbtree_api_kfunc(kfunc_btf_id);
12795 		break;
12796 	default:
12797 		verbose(env, "verifier internal error: unexpected graph root argument type %s\n",
12798 			btf_field_type_name(head_field_type));
12799 		return false;
12800 	}
12801 
12802 	if (!ret)
12803 		verbose(env, "verifier internal error: %s head arg for unknown kfunc\n",
12804 			btf_field_type_name(head_field_type));
12805 	return ret;
12806 }
12807 
check_kfunc_is_graph_node_api(struct bpf_verifier_env * env,enum btf_field_type node_field_type,u32 kfunc_btf_id)12808 static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
12809 					  enum btf_field_type node_field_type,
12810 					  u32 kfunc_btf_id)
12811 {
12812 	bool ret;
12813 
12814 	switch (node_field_type) {
12815 	case BPF_LIST_NODE:
12816 		ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
12817 		       kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
12818 		break;
12819 	case BPF_RB_NODE:
12820 		ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
12821 		       kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
12822 		       kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_left] ||
12823 		       kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_right]);
12824 		break;
12825 	default:
12826 		verbose(env, "verifier internal error: unexpected graph node argument type %s\n",
12827 			btf_field_type_name(node_field_type));
12828 		return false;
12829 	}
12830 
12831 	if (!ret)
12832 		verbose(env, "verifier internal error: %s node arg for unknown kfunc\n",
12833 			btf_field_type_name(node_field_type));
12834 	return ret;
12835 }
12836 
12837 static int
__process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta,enum btf_field_type head_field_type,struct btf_field ** head_field)12838 __process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
12839 				   struct bpf_reg_state *reg, u32 regno,
12840 				   struct bpf_kfunc_call_arg_meta *meta,
12841 				   enum btf_field_type head_field_type,
12842 				   struct btf_field **head_field)
12843 {
12844 	const char *head_type_name;
12845 	struct btf_field *field;
12846 	struct btf_record *rec;
12847 	u32 head_off;
12848 
12849 	if (meta->btf != btf_vmlinux) {
12850 		verifier_bug(env, "unexpected btf mismatch in kfunc call");
12851 		return -EFAULT;
12852 	}
12853 
12854 	if (!check_kfunc_is_graph_root_api(env, head_field_type, meta->func_id))
12855 		return -EFAULT;
12856 
12857 	head_type_name = btf_field_type_name(head_field_type);
12858 	if (!tnum_is_const(reg->var_off)) {
12859 		verbose(env,
12860 			"R%d doesn't have constant offset. %s has to be at the constant offset\n",
12861 			regno, head_type_name);
12862 		return -EINVAL;
12863 	}
12864 
12865 	rec = reg_btf_record(reg);
12866 	head_off = reg->off + reg->var_off.value;
12867 	field = btf_record_find(rec, head_off, head_field_type);
12868 	if (!field) {
12869 		verbose(env, "%s not found at offset=%u\n", head_type_name, head_off);
12870 		return -EINVAL;
12871 	}
12872 
12873 	/* All functions require bpf_list_head to be protected using a bpf_spin_lock */
12874 	if (check_reg_allocation_locked(env, reg)) {
12875 		verbose(env, "bpf_spin_lock at off=%d must be held for %s\n",
12876 			rec->spin_lock_off, head_type_name);
12877 		return -EINVAL;
12878 	}
12879 
12880 	if (*head_field) {
12881 		verifier_bug(env, "repeating %s arg", head_type_name);
12882 		return -EFAULT;
12883 	}
12884 	*head_field = field;
12885 	return 0;
12886 }
12887 
process_kf_arg_ptr_to_list_head(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)12888 static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
12889 					   struct bpf_reg_state *reg, u32 regno,
12890 					   struct bpf_kfunc_call_arg_meta *meta)
12891 {
12892 	return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_LIST_HEAD,
12893 							  &meta->arg_list_head.field);
12894 }
12895 
process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)12896 static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
12897 					     struct bpf_reg_state *reg, u32 regno,
12898 					     struct bpf_kfunc_call_arg_meta *meta)
12899 {
12900 	return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, BPF_RB_ROOT,
12901 							  &meta->arg_rbtree_root.field);
12902 }
12903 
12904 static int
__process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta,enum btf_field_type head_field_type,enum btf_field_type node_field_type,struct btf_field ** node_field)12905 __process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
12906 				   struct bpf_reg_state *reg, u32 regno,
12907 				   struct bpf_kfunc_call_arg_meta *meta,
12908 				   enum btf_field_type head_field_type,
12909 				   enum btf_field_type node_field_type,
12910 				   struct btf_field **node_field)
12911 {
12912 	const char *node_type_name;
12913 	const struct btf_type *et, *t;
12914 	struct btf_field *field;
12915 	u32 node_off;
12916 
12917 	if (meta->btf != btf_vmlinux) {
12918 		verifier_bug(env, "unexpected btf mismatch in kfunc call");
12919 		return -EFAULT;
12920 	}
12921 
12922 	if (!check_kfunc_is_graph_node_api(env, node_field_type, meta->func_id))
12923 		return -EFAULT;
12924 
12925 	node_type_name = btf_field_type_name(node_field_type);
12926 	if (!tnum_is_const(reg->var_off)) {
12927 		verbose(env,
12928 			"R%d doesn't have constant offset. %s has to be at the constant offset\n",
12929 			regno, node_type_name);
12930 		return -EINVAL;
12931 	}
12932 
12933 	node_off = reg->off + reg->var_off.value;
12934 	field = reg_find_field_offset(reg, node_off, node_field_type);
12935 	if (!field) {
12936 		verbose(env, "%s not found at offset=%u\n", node_type_name, node_off);
12937 		return -EINVAL;
12938 	}
12939 
12940 	field = *node_field;
12941 
12942 	et = btf_type_by_id(field->graph_root.btf, field->graph_root.value_btf_id);
12943 	t = btf_type_by_id(reg->btf, reg->btf_id);
12944 	if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, 0, field->graph_root.btf,
12945 				  field->graph_root.value_btf_id, true)) {
12946 		verbose(env, "operation on %s expects arg#1 %s at offset=%d "
12947 			"in struct %s, but arg is at offset=%d in struct %s\n",
12948 			btf_field_type_name(head_field_type),
12949 			btf_field_type_name(node_field_type),
12950 			field->graph_root.node_offset,
12951 			btf_name_by_offset(field->graph_root.btf, et->name_off),
12952 			node_off, btf_name_by_offset(reg->btf, t->name_off));
12953 		return -EINVAL;
12954 	}
12955 	meta->arg_btf = reg->btf;
12956 	meta->arg_btf_id = reg->btf_id;
12957 
12958 	if (node_off != field->graph_root.node_offset) {
12959 		verbose(env, "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
12960 			node_off, btf_field_type_name(node_field_type),
12961 			field->graph_root.node_offset,
12962 			btf_name_by_offset(field->graph_root.btf, et->name_off));
12963 		return -EINVAL;
12964 	}
12965 
12966 	return 0;
12967 }
12968 
process_kf_arg_ptr_to_list_node(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)12969 static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
12970 					   struct bpf_reg_state *reg, u32 regno,
12971 					   struct bpf_kfunc_call_arg_meta *meta)
12972 {
12973 	return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
12974 						  BPF_LIST_HEAD, BPF_LIST_NODE,
12975 						  &meta->arg_list_head.field);
12976 }
12977 
process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env * env,struct bpf_reg_state * reg,u32 regno,struct bpf_kfunc_call_arg_meta * meta)12978 static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
12979 					     struct bpf_reg_state *reg, u32 regno,
12980 					     struct bpf_kfunc_call_arg_meta *meta)
12981 {
12982 	return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
12983 						  BPF_RB_ROOT, BPF_RB_NODE,
12984 						  &meta->arg_rbtree_root.field);
12985 }
12986 
12987 /*
12988  * css_task iter allowlist is needed to avoid dead locking on css_set_lock.
12989  * LSM hooks and iters (both sleepable and non-sleepable) are safe.
12990  * Any sleepable progs are also safe since bpf_check_attach_target() enforce
12991  * them can only be attached to some specific hook points.
12992  */
check_css_task_iter_allowlist(struct bpf_verifier_env * env)12993 static bool check_css_task_iter_allowlist(struct bpf_verifier_env *env)
12994 {
12995 	enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
12996 
12997 	switch (prog_type) {
12998 	case BPF_PROG_TYPE_LSM:
12999 		return true;
13000 	case BPF_PROG_TYPE_TRACING:
13001 		if (env->prog->expected_attach_type == BPF_TRACE_ITER)
13002 			return true;
13003 		fallthrough;
13004 	default:
13005 		return in_sleepable(env);
13006 	}
13007 }
13008 
check_kfunc_args(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,int insn_idx)13009 static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
13010 			    int insn_idx)
13011 {
13012 	const char *func_name = meta->func_name, *ref_tname;
13013 	const struct btf *btf = meta->btf;
13014 	const struct btf_param *args;
13015 	struct btf_record *rec;
13016 	u32 i, nargs;
13017 	int ret;
13018 
13019 	args = (const struct btf_param *)(meta->func_proto + 1);
13020 	nargs = btf_type_vlen(meta->func_proto);
13021 	if (nargs > MAX_BPF_FUNC_REG_ARGS) {
13022 		verbose(env, "Function %s has %d > %d args\n", func_name, nargs,
13023 			MAX_BPF_FUNC_REG_ARGS);
13024 		return -EINVAL;
13025 	}
13026 
13027 	/* Check that BTF function arguments match actual types that the
13028 	 * verifier sees.
13029 	 */
13030 	for (i = 0; i < nargs; i++) {
13031 		struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[i + 1];
13032 		const struct btf_type *t, *ref_t, *resolve_ret;
13033 		enum bpf_arg_type arg_type = ARG_DONTCARE;
13034 		u32 regno = i + 1, ref_id, type_size;
13035 		bool is_ret_buf_sz = false;
13036 		int kf_arg_type;
13037 
13038 		t = btf_type_skip_modifiers(btf, args[i].type, NULL);
13039 
13040 		if (is_kfunc_arg_ignore(btf, &args[i]))
13041 			continue;
13042 
13043 		if (is_kfunc_arg_prog(btf, &args[i])) {
13044 			/* Used to reject repeated use of __prog. */
13045 			if (meta->arg_prog) {
13046 				verifier_bug(env, "Only 1 prog->aux argument supported per-kfunc");
13047 				return -EFAULT;
13048 			}
13049 			meta->arg_prog = true;
13050 			cur_aux(env)->arg_prog = regno;
13051 			continue;
13052 		}
13053 
13054 		if (btf_type_is_scalar(t)) {
13055 			if (reg->type != SCALAR_VALUE) {
13056 				verbose(env, "R%d is not a scalar\n", regno);
13057 				return -EINVAL;
13058 			}
13059 
13060 			if (is_kfunc_arg_constant(meta->btf, &args[i])) {
13061 				if (meta->arg_constant.found) {
13062 					verifier_bug(env, "only one constant argument permitted");
13063 					return -EFAULT;
13064 				}
13065 				if (!tnum_is_const(reg->var_off)) {
13066 					verbose(env, "R%d must be a known constant\n", regno);
13067 					return -EINVAL;
13068 				}
13069 				ret = mark_chain_precision(env, regno);
13070 				if (ret < 0)
13071 					return ret;
13072 				meta->arg_constant.found = true;
13073 				meta->arg_constant.value = reg->var_off.value;
13074 			} else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdonly_buf_size")) {
13075 				meta->r0_rdonly = true;
13076 				is_ret_buf_sz = true;
13077 			} else if (is_kfunc_arg_scalar_with_name(btf, &args[i], "rdwr_buf_size")) {
13078 				is_ret_buf_sz = true;
13079 			}
13080 
13081 			if (is_ret_buf_sz) {
13082 				if (meta->r0_size) {
13083 					verbose(env, "2 or more rdonly/rdwr_buf_size parameters for kfunc");
13084 					return -EINVAL;
13085 				}
13086 
13087 				if (!tnum_is_const(reg->var_off)) {
13088 					verbose(env, "R%d is not a const\n", regno);
13089 					return -EINVAL;
13090 				}
13091 
13092 				meta->r0_size = reg->var_off.value;
13093 				ret = mark_chain_precision(env, regno);
13094 				if (ret)
13095 					return ret;
13096 			}
13097 			continue;
13098 		}
13099 
13100 		if (!btf_type_is_ptr(t)) {
13101 			verbose(env, "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
13102 			return -EINVAL;
13103 		}
13104 
13105 		if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
13106 		    (register_is_null(reg) || type_may_be_null(reg->type)) &&
13107 			!is_kfunc_arg_nullable(meta->btf, &args[i])) {
13108 			verbose(env, "Possibly NULL pointer passed to trusted arg%d\n", i);
13109 			return -EACCES;
13110 		}
13111 
13112 		if (reg->ref_obj_id) {
13113 			if (is_kfunc_release(meta) && meta->ref_obj_id) {
13114 				verifier_bug(env, "more than one arg with ref_obj_id R%d %u %u",
13115 					     regno, reg->ref_obj_id,
13116 					     meta->ref_obj_id);
13117 				return -EFAULT;
13118 			}
13119 			meta->ref_obj_id = reg->ref_obj_id;
13120 			if (is_kfunc_release(meta))
13121 				meta->release_regno = regno;
13122 		}
13123 
13124 		ref_t = btf_type_skip_modifiers(btf, t->type, &ref_id);
13125 		ref_tname = btf_name_by_offset(btf, ref_t->name_off);
13126 
13127 		kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, i, nargs);
13128 		if (kf_arg_type < 0)
13129 			return kf_arg_type;
13130 
13131 		switch (kf_arg_type) {
13132 		case KF_ARG_PTR_TO_NULL:
13133 			continue;
13134 		case KF_ARG_PTR_TO_MAP:
13135 			if (!reg->map_ptr) {
13136 				verbose(env, "pointer in R%d isn't map pointer\n", regno);
13137 				return -EINVAL;
13138 			}
13139 			if (meta->map.ptr && reg->map_ptr->record->wq_off >= 0) {
13140 				/* Use map_uid (which is unique id of inner map) to reject:
13141 				 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
13142 				 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
13143 				 * if (inner_map1 && inner_map2) {
13144 				 *     wq = bpf_map_lookup_elem(inner_map1);
13145 				 *     if (wq)
13146 				 *         // mismatch would have been allowed
13147 				 *         bpf_wq_init(wq, inner_map2);
13148 				 * }
13149 				 *
13150 				 * Comparing map_ptr is enough to distinguish normal and outer maps.
13151 				 */
13152 				if (meta->map.ptr != reg->map_ptr ||
13153 				    meta->map.uid != reg->map_uid) {
13154 					verbose(env,
13155 						"workqueue pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
13156 						meta->map.uid, reg->map_uid);
13157 					return -EINVAL;
13158 				}
13159 			}
13160 			meta->map.ptr = reg->map_ptr;
13161 			meta->map.uid = reg->map_uid;
13162 			fallthrough;
13163 		case KF_ARG_PTR_TO_ALLOC_BTF_ID:
13164 		case KF_ARG_PTR_TO_BTF_ID:
13165 			if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
13166 				break;
13167 
13168 			if (!is_trusted_reg(reg)) {
13169 				if (!is_kfunc_rcu(meta)) {
13170 					verbose(env, "R%d must be referenced or trusted\n", regno);
13171 					return -EINVAL;
13172 				}
13173 				if (!is_rcu_reg(reg)) {
13174 					verbose(env, "R%d must be a rcu pointer\n", regno);
13175 					return -EINVAL;
13176 				}
13177 			}
13178 			fallthrough;
13179 		case KF_ARG_PTR_TO_CTX:
13180 		case KF_ARG_PTR_TO_DYNPTR:
13181 		case KF_ARG_PTR_TO_ITER:
13182 		case KF_ARG_PTR_TO_LIST_HEAD:
13183 		case KF_ARG_PTR_TO_LIST_NODE:
13184 		case KF_ARG_PTR_TO_RB_ROOT:
13185 		case KF_ARG_PTR_TO_RB_NODE:
13186 		case KF_ARG_PTR_TO_MEM:
13187 		case KF_ARG_PTR_TO_MEM_SIZE:
13188 		case KF_ARG_PTR_TO_CALLBACK:
13189 		case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
13190 		case KF_ARG_PTR_TO_CONST_STR:
13191 		case KF_ARG_PTR_TO_WORKQUEUE:
13192 		case KF_ARG_PTR_TO_IRQ_FLAG:
13193 		case KF_ARG_PTR_TO_RES_SPIN_LOCK:
13194 			break;
13195 		default:
13196 			verifier_bug(env, "unknown kfunc arg type %d", kf_arg_type);
13197 			return -EFAULT;
13198 		}
13199 
13200 		if (is_kfunc_release(meta) && reg->ref_obj_id)
13201 			arg_type |= OBJ_RELEASE;
13202 		ret = check_func_arg_reg_off(env, reg, regno, arg_type);
13203 		if (ret < 0)
13204 			return ret;
13205 
13206 		switch (kf_arg_type) {
13207 		case KF_ARG_PTR_TO_CTX:
13208 			if (reg->type != PTR_TO_CTX) {
13209 				verbose(env, "arg#%d expected pointer to ctx, but got %s\n",
13210 					i, reg_type_str(env, reg->type));
13211 				return -EINVAL;
13212 			}
13213 
13214 			if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
13215 				ret = get_kern_ctx_btf_id(&env->log, resolve_prog_type(env->prog));
13216 				if (ret < 0)
13217 					return -EINVAL;
13218 				meta->ret_btf_id  = ret;
13219 			}
13220 			break;
13221 		case KF_ARG_PTR_TO_ALLOC_BTF_ID:
13222 			if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC)) {
13223 				if (meta->func_id != special_kfunc_list[KF_bpf_obj_drop_impl]) {
13224 					verbose(env, "arg#%d expected for bpf_obj_drop_impl()\n", i);
13225 					return -EINVAL;
13226 				}
13227 			} else if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC | MEM_PERCPU)) {
13228 				if (meta->func_id != special_kfunc_list[KF_bpf_percpu_obj_drop_impl]) {
13229 					verbose(env, "arg#%d expected for bpf_percpu_obj_drop_impl()\n", i);
13230 					return -EINVAL;
13231 				}
13232 			} else {
13233 				verbose(env, "arg#%d expected pointer to allocated object\n", i);
13234 				return -EINVAL;
13235 			}
13236 			if (!reg->ref_obj_id) {
13237 				verbose(env, "allocated object must be referenced\n");
13238 				return -EINVAL;
13239 			}
13240 			if (meta->btf == btf_vmlinux) {
13241 				meta->arg_btf = reg->btf;
13242 				meta->arg_btf_id = reg->btf_id;
13243 			}
13244 			break;
13245 		case KF_ARG_PTR_TO_DYNPTR:
13246 		{
13247 			enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
13248 			int clone_ref_obj_id = 0;
13249 
13250 			if (reg->type == CONST_PTR_TO_DYNPTR)
13251 				dynptr_arg_type |= MEM_RDONLY;
13252 
13253 			if (is_kfunc_arg_uninit(btf, &args[i]))
13254 				dynptr_arg_type |= MEM_UNINIT;
13255 
13256 			if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
13257 				dynptr_arg_type |= DYNPTR_TYPE_SKB;
13258 			} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
13259 				dynptr_arg_type |= DYNPTR_TYPE_XDP;
13260 			} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
13261 				   (dynptr_arg_type & MEM_UNINIT)) {
13262 				enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
13263 
13264 				if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
13265 					verifier_bug(env, "no dynptr type for parent of clone");
13266 					return -EFAULT;
13267 				}
13268 
13269 				dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(parent_type);
13270 				clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
13271 				if (dynptr_type_refcounted(parent_type) && !clone_ref_obj_id) {
13272 					verifier_bug(env, "missing ref obj id for parent of clone");
13273 					return -EFAULT;
13274 				}
13275 			}
13276 
13277 			ret = process_dynptr_func(env, regno, insn_idx, dynptr_arg_type, clone_ref_obj_id);
13278 			if (ret < 0)
13279 				return ret;
13280 
13281 			if (!(dynptr_arg_type & MEM_UNINIT)) {
13282 				int id = dynptr_id(env, reg);
13283 
13284 				if (id < 0) {
13285 					verifier_bug(env, "failed to obtain dynptr id");
13286 					return id;
13287 				}
13288 				meta->initialized_dynptr.id = id;
13289 				meta->initialized_dynptr.type = dynptr_get_type(env, reg);
13290 				meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
13291 			}
13292 
13293 			break;
13294 		}
13295 		case KF_ARG_PTR_TO_ITER:
13296 			if (meta->func_id == special_kfunc_list[KF_bpf_iter_css_task_new]) {
13297 				if (!check_css_task_iter_allowlist(env)) {
13298 					verbose(env, "css_task_iter is only allowed in bpf_lsm, bpf_iter and sleepable progs\n");
13299 					return -EINVAL;
13300 				}
13301 			}
13302 			ret = process_iter_arg(env, regno, insn_idx, meta);
13303 			if (ret < 0)
13304 				return ret;
13305 			break;
13306 		case KF_ARG_PTR_TO_LIST_HEAD:
13307 			if (reg->type != PTR_TO_MAP_VALUE &&
13308 			    reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13309 				verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
13310 				return -EINVAL;
13311 			}
13312 			if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
13313 				verbose(env, "allocated object must be referenced\n");
13314 				return -EINVAL;
13315 			}
13316 			ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
13317 			if (ret < 0)
13318 				return ret;
13319 			break;
13320 		case KF_ARG_PTR_TO_RB_ROOT:
13321 			if (reg->type != PTR_TO_MAP_VALUE &&
13322 			    reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13323 				verbose(env, "arg#%d expected pointer to map value or allocated object\n", i);
13324 				return -EINVAL;
13325 			}
13326 			if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
13327 				verbose(env, "allocated object must be referenced\n");
13328 				return -EINVAL;
13329 			}
13330 			ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
13331 			if (ret < 0)
13332 				return ret;
13333 			break;
13334 		case KF_ARG_PTR_TO_LIST_NODE:
13335 			if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13336 				verbose(env, "arg#%d expected pointer to allocated object\n", i);
13337 				return -EINVAL;
13338 			}
13339 			if (!reg->ref_obj_id) {
13340 				verbose(env, "allocated object must be referenced\n");
13341 				return -EINVAL;
13342 			}
13343 			ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
13344 			if (ret < 0)
13345 				return ret;
13346 			break;
13347 		case KF_ARG_PTR_TO_RB_NODE:
13348 			if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
13349 				if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13350 					verbose(env, "arg#%d expected pointer to allocated object\n", i);
13351 					return -EINVAL;
13352 				}
13353 				if (!reg->ref_obj_id) {
13354 					verbose(env, "allocated object must be referenced\n");
13355 					return -EINVAL;
13356 				}
13357 			} else {
13358 				if (!type_is_non_owning_ref(reg->type) && !reg->ref_obj_id) {
13359 					verbose(env, "%s can only take non-owning or refcounted bpf_rb_node pointer\n", func_name);
13360 					return -EINVAL;
13361 				}
13362 				if (in_rbtree_lock_required_cb(env)) {
13363 					verbose(env, "%s not allowed in rbtree cb\n", func_name);
13364 					return -EINVAL;
13365 				}
13366 			}
13367 
13368 			ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
13369 			if (ret < 0)
13370 				return ret;
13371 			break;
13372 		case KF_ARG_PTR_TO_MAP:
13373 			/* If argument has '__map' suffix expect 'struct bpf_map *' */
13374 			ref_id = *reg2btf_ids[CONST_PTR_TO_MAP];
13375 			ref_t = btf_type_by_id(btf_vmlinux, ref_id);
13376 			ref_tname = btf_name_by_offset(btf, ref_t->name_off);
13377 			fallthrough;
13378 		case KF_ARG_PTR_TO_BTF_ID:
13379 			/* Only base_type is checked, further checks are done here */
13380 			if ((base_type(reg->type) != PTR_TO_BTF_ID ||
13381 			     (bpf_type_has_unsafe_modifiers(reg->type) && !is_rcu_reg(reg))) &&
13382 			    !reg2btf_ids[base_type(reg->type)]) {
13383 				verbose(env, "arg#%d is %s ", i, reg_type_str(env, reg->type));
13384 				verbose(env, "expected %s or socket\n",
13385 					reg_type_str(env, base_type(reg->type) |
13386 							  (type_flag(reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
13387 				return -EINVAL;
13388 			}
13389 			ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, i);
13390 			if (ret < 0)
13391 				return ret;
13392 			break;
13393 		case KF_ARG_PTR_TO_MEM:
13394 			resolve_ret = btf_resolve_size(btf, ref_t, &type_size);
13395 			if (IS_ERR(resolve_ret)) {
13396 				verbose(env, "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
13397 					i, btf_type_str(ref_t), ref_tname, PTR_ERR(resolve_ret));
13398 				return -EINVAL;
13399 			}
13400 			ret = check_mem_reg(env, reg, regno, type_size);
13401 			if (ret < 0)
13402 				return ret;
13403 			break;
13404 		case KF_ARG_PTR_TO_MEM_SIZE:
13405 		{
13406 			struct bpf_reg_state *buff_reg = &regs[regno];
13407 			const struct btf_param *buff_arg = &args[i];
13408 			struct bpf_reg_state *size_reg = &regs[regno + 1];
13409 			const struct btf_param *size_arg = &args[i + 1];
13410 
13411 			if (!register_is_null(buff_reg) || !is_kfunc_arg_optional(meta->btf, buff_arg)) {
13412 				ret = check_kfunc_mem_size_reg(env, size_reg, regno + 1);
13413 				if (ret < 0) {
13414 					verbose(env, "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
13415 					return ret;
13416 				}
13417 			}
13418 
13419 			if (is_kfunc_arg_const_mem_size(meta->btf, size_arg, size_reg)) {
13420 				if (meta->arg_constant.found) {
13421 					verifier_bug(env, "only one constant argument permitted");
13422 					return -EFAULT;
13423 				}
13424 				if (!tnum_is_const(size_reg->var_off)) {
13425 					verbose(env, "R%d must be a known constant\n", regno + 1);
13426 					return -EINVAL;
13427 				}
13428 				meta->arg_constant.found = true;
13429 				meta->arg_constant.value = size_reg->var_off.value;
13430 			}
13431 
13432 			/* Skip next '__sz' or '__szk' argument */
13433 			i++;
13434 			break;
13435 		}
13436 		case KF_ARG_PTR_TO_CALLBACK:
13437 			if (reg->type != PTR_TO_FUNC) {
13438 				verbose(env, "arg%d expected pointer to func\n", i);
13439 				return -EINVAL;
13440 			}
13441 			meta->subprogno = reg->subprogno;
13442 			break;
13443 		case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
13444 			if (!type_is_ptr_alloc_obj(reg->type)) {
13445 				verbose(env, "arg#%d is neither owning or non-owning ref\n", i);
13446 				return -EINVAL;
13447 			}
13448 			if (!type_is_non_owning_ref(reg->type))
13449 				meta->arg_owning_ref = true;
13450 
13451 			rec = reg_btf_record(reg);
13452 			if (!rec) {
13453 				verifier_bug(env, "Couldn't find btf_record");
13454 				return -EFAULT;
13455 			}
13456 
13457 			if (rec->refcount_off < 0) {
13458 				verbose(env, "arg#%d doesn't point to a type with bpf_refcount field\n", i);
13459 				return -EINVAL;
13460 			}
13461 
13462 			meta->arg_btf = reg->btf;
13463 			meta->arg_btf_id = reg->btf_id;
13464 			break;
13465 		case KF_ARG_PTR_TO_CONST_STR:
13466 			if (reg->type != PTR_TO_MAP_VALUE) {
13467 				verbose(env, "arg#%d doesn't point to a const string\n", i);
13468 				return -EINVAL;
13469 			}
13470 			ret = check_reg_const_str(env, reg, regno);
13471 			if (ret)
13472 				return ret;
13473 			break;
13474 		case KF_ARG_PTR_TO_WORKQUEUE:
13475 			if (reg->type != PTR_TO_MAP_VALUE) {
13476 				verbose(env, "arg#%d doesn't point to a map value\n", i);
13477 				return -EINVAL;
13478 			}
13479 			ret = process_wq_func(env, regno, meta);
13480 			if (ret < 0)
13481 				return ret;
13482 			break;
13483 		case KF_ARG_PTR_TO_IRQ_FLAG:
13484 			if (reg->type != PTR_TO_STACK) {
13485 				verbose(env, "arg#%d doesn't point to an irq flag on stack\n", i);
13486 				return -EINVAL;
13487 			}
13488 			ret = process_irq_flag(env, regno, meta);
13489 			if (ret < 0)
13490 				return ret;
13491 			break;
13492 		case KF_ARG_PTR_TO_RES_SPIN_LOCK:
13493 		{
13494 			int flags = PROCESS_RES_LOCK;
13495 
13496 			if (reg->type != PTR_TO_MAP_VALUE && reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13497 				verbose(env, "arg#%d doesn't point to map value or allocated object\n", i);
13498 				return -EINVAL;
13499 			}
13500 
13501 			if (!is_bpf_res_spin_lock_kfunc(meta->func_id))
13502 				return -EFAULT;
13503 			if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock] ||
13504 			    meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave])
13505 				flags |= PROCESS_SPIN_LOCK;
13506 			if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave] ||
13507 			    meta->func_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore])
13508 				flags |= PROCESS_LOCK_IRQ;
13509 			ret = process_spin_lock(env, regno, flags);
13510 			if (ret < 0)
13511 				return ret;
13512 			break;
13513 		}
13514 		}
13515 	}
13516 
13517 	if (is_kfunc_release(meta) && !meta->release_regno) {
13518 		verbose(env, "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
13519 			func_name);
13520 		return -EINVAL;
13521 	}
13522 
13523 	return 0;
13524 }
13525 
fetch_kfunc_meta(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_kfunc_call_arg_meta * meta,const char ** kfunc_name)13526 static int fetch_kfunc_meta(struct bpf_verifier_env *env,
13527 			    struct bpf_insn *insn,
13528 			    struct bpf_kfunc_call_arg_meta *meta,
13529 			    const char **kfunc_name)
13530 {
13531 	const struct btf_type *func, *func_proto;
13532 	u32 func_id, *kfunc_flags;
13533 	const char *func_name;
13534 	struct btf *desc_btf;
13535 
13536 	if (kfunc_name)
13537 		*kfunc_name = NULL;
13538 
13539 	if (!insn->imm)
13540 		return -EINVAL;
13541 
13542 	desc_btf = find_kfunc_desc_btf(env, insn->off);
13543 	if (IS_ERR(desc_btf))
13544 		return PTR_ERR(desc_btf);
13545 
13546 	func_id = insn->imm;
13547 	func = btf_type_by_id(desc_btf, func_id);
13548 	func_name = btf_name_by_offset(desc_btf, func->name_off);
13549 	if (kfunc_name)
13550 		*kfunc_name = func_name;
13551 	func_proto = btf_type_by_id(desc_btf, func->type);
13552 
13553 	kfunc_flags = btf_kfunc_id_set_contains(desc_btf, func_id, env->prog);
13554 	if (!kfunc_flags) {
13555 		return -EACCES;
13556 	}
13557 
13558 	memset(meta, 0, sizeof(*meta));
13559 	meta->btf = desc_btf;
13560 	meta->func_id = func_id;
13561 	meta->kfunc_flags = *kfunc_flags;
13562 	meta->func_proto = func_proto;
13563 	meta->func_name = func_name;
13564 
13565 	return 0;
13566 }
13567 
13568 /* check special kfuncs and return:
13569  *  1  - not fall-through to 'else' branch, continue verification
13570  *  0  - fall-through to 'else' branch
13571  * < 0 - not fall-through to 'else' branch, return error
13572  */
check_special_kfunc(struct bpf_verifier_env * env,struct bpf_kfunc_call_arg_meta * meta,struct bpf_reg_state * regs,struct bpf_insn_aux_data * insn_aux,const struct btf_type * ptr_type,struct btf * desc_btf)13573 static int check_special_kfunc(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
13574 			       struct bpf_reg_state *regs, struct bpf_insn_aux_data *insn_aux,
13575 			       const struct btf_type *ptr_type, struct btf *desc_btf)
13576 {
13577 	const struct btf_type *ret_t;
13578 	int err = 0;
13579 
13580 	if (meta->btf != btf_vmlinux)
13581 		return 0;
13582 
13583 	if (meta->func_id == special_kfunc_list[KF_bpf_obj_new_impl] ||
13584 	    meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
13585 		struct btf_struct_meta *struct_meta;
13586 		struct btf *ret_btf;
13587 		u32 ret_btf_id;
13588 
13589 		if (meta->func_id == special_kfunc_list[KF_bpf_obj_new_impl] && !bpf_global_ma_set)
13590 			return -ENOMEM;
13591 
13592 		if (((u64)(u32)meta->arg_constant.value) != meta->arg_constant.value) {
13593 			verbose(env, "local type ID argument must be in range [0, U32_MAX]\n");
13594 			return -EINVAL;
13595 		}
13596 
13597 		ret_btf = env->prog->aux->btf;
13598 		ret_btf_id = meta->arg_constant.value;
13599 
13600 		/* This may be NULL due to user not supplying a BTF */
13601 		if (!ret_btf) {
13602 			verbose(env, "bpf_obj_new/bpf_percpu_obj_new requires prog BTF\n");
13603 			return -EINVAL;
13604 		}
13605 
13606 		ret_t = btf_type_by_id(ret_btf, ret_btf_id);
13607 		if (!ret_t || !__btf_type_is_struct(ret_t)) {
13608 			verbose(env, "bpf_obj_new/bpf_percpu_obj_new type ID argument must be of a struct\n");
13609 			return -EINVAL;
13610 		}
13611 
13612 		if (meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
13613 			if (ret_t->size > BPF_GLOBAL_PERCPU_MA_MAX_SIZE) {
13614 				verbose(env, "bpf_percpu_obj_new type size (%d) is greater than %d\n",
13615 					ret_t->size, BPF_GLOBAL_PERCPU_MA_MAX_SIZE);
13616 				return -EINVAL;
13617 			}
13618 
13619 			if (!bpf_global_percpu_ma_set) {
13620 				mutex_lock(&bpf_percpu_ma_lock);
13621 				if (!bpf_global_percpu_ma_set) {
13622 					/* Charge memory allocated with bpf_global_percpu_ma to
13623 					 * root memcg. The obj_cgroup for root memcg is NULL.
13624 					 */
13625 					err = bpf_mem_alloc_percpu_init(&bpf_global_percpu_ma, NULL);
13626 					if (!err)
13627 						bpf_global_percpu_ma_set = true;
13628 				}
13629 				mutex_unlock(&bpf_percpu_ma_lock);
13630 				if (err)
13631 					return err;
13632 			}
13633 
13634 			mutex_lock(&bpf_percpu_ma_lock);
13635 			err = bpf_mem_alloc_percpu_unit_init(&bpf_global_percpu_ma, ret_t->size);
13636 			mutex_unlock(&bpf_percpu_ma_lock);
13637 			if (err)
13638 				return err;
13639 		}
13640 
13641 		struct_meta = btf_find_struct_meta(ret_btf, ret_btf_id);
13642 		if (meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
13643 			if (!__btf_type_is_scalar_struct(env, ret_btf, ret_t, 0)) {
13644 				verbose(env, "bpf_percpu_obj_new type ID argument must be of a struct of scalars\n");
13645 				return -EINVAL;
13646 			}
13647 
13648 			if (struct_meta) {
13649 				verbose(env, "bpf_percpu_obj_new type ID argument must not contain special fields\n");
13650 				return -EINVAL;
13651 			}
13652 		}
13653 
13654 		mark_reg_known_zero(env, regs, BPF_REG_0);
13655 		regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
13656 		regs[BPF_REG_0].btf = ret_btf;
13657 		regs[BPF_REG_0].btf_id = ret_btf_id;
13658 		if (meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl])
13659 			regs[BPF_REG_0].type |= MEM_PERCPU;
13660 
13661 		insn_aux->obj_new_size = ret_t->size;
13662 		insn_aux->kptr_struct_meta = struct_meta;
13663 	} else if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
13664 		mark_reg_known_zero(env, regs, BPF_REG_0);
13665 		regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
13666 		regs[BPF_REG_0].btf = meta->arg_btf;
13667 		regs[BPF_REG_0].btf_id = meta->arg_btf_id;
13668 
13669 		insn_aux->kptr_struct_meta =
13670 			btf_find_struct_meta(meta->arg_btf,
13671 					     meta->arg_btf_id);
13672 	} else if (is_list_node_type(ptr_type)) {
13673 		struct btf_field *field = meta->arg_list_head.field;
13674 
13675 		mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
13676 	} else if (is_rbtree_node_type(ptr_type)) {
13677 		struct btf_field *field = meta->arg_rbtree_root.field;
13678 
13679 		mark_reg_graph_node(regs, BPF_REG_0, &field->graph_root);
13680 	} else if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
13681 		mark_reg_known_zero(env, regs, BPF_REG_0);
13682 		regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
13683 		regs[BPF_REG_0].btf = desc_btf;
13684 		regs[BPF_REG_0].btf_id = meta->ret_btf_id;
13685 	} else if (meta->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
13686 		ret_t = btf_type_by_id(desc_btf, meta->arg_constant.value);
13687 		if (!ret_t) {
13688 			verbose(env, "Unknown type ID %lld passed to kfunc bpf_rdonly_cast\n",
13689 				meta->arg_constant.value);
13690 			return -EINVAL;
13691 		} else if (btf_type_is_struct(ret_t)) {
13692 			mark_reg_known_zero(env, regs, BPF_REG_0);
13693 			regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
13694 			regs[BPF_REG_0].btf = desc_btf;
13695 			regs[BPF_REG_0].btf_id = meta->arg_constant.value;
13696 		} else if (btf_type_is_void(ret_t)) {
13697 			mark_reg_known_zero(env, regs, BPF_REG_0);
13698 			regs[BPF_REG_0].type = PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED;
13699 			regs[BPF_REG_0].mem_size = 0;
13700 		} else {
13701 			verbose(env,
13702 				"kfunc bpf_rdonly_cast type ID argument must be of a struct or void\n");
13703 			return -EINVAL;
13704 		}
13705 	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
13706 		   meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
13707 		enum bpf_type_flag type_flag = get_dynptr_type_flag(meta->initialized_dynptr.type);
13708 
13709 		mark_reg_known_zero(env, regs, BPF_REG_0);
13710 
13711 		if (!meta->arg_constant.found) {
13712 			verifier_bug(env, "bpf_dynptr_slice(_rdwr) no constant size");
13713 			return -EFAULT;
13714 		}
13715 
13716 		regs[BPF_REG_0].mem_size = meta->arg_constant.value;
13717 
13718 		/* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
13719 		regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
13720 
13721 		if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
13722 			regs[BPF_REG_0].type |= MEM_RDONLY;
13723 		} else {
13724 			/* this will set env->seen_direct_write to true */
13725 			if (!may_access_direct_pkt_data(env, NULL, BPF_WRITE)) {
13726 				verbose(env, "the prog does not allow writes to packet data\n");
13727 				return -EINVAL;
13728 			}
13729 		}
13730 
13731 		if (!meta->initialized_dynptr.id) {
13732 			verifier_bug(env, "no dynptr id");
13733 			return -EFAULT;
13734 		}
13735 		regs[BPF_REG_0].dynptr_id = meta->initialized_dynptr.id;
13736 
13737 		/* we don't need to set BPF_REG_0's ref obj id
13738 		 * because packet slices are not refcounted (see
13739 		 * dynptr_type_refcounted)
13740 		 */
13741 	} else {
13742 		return 0;
13743 	}
13744 
13745 	return 1;
13746 }
13747 
13748 static int check_return_code(struct bpf_verifier_env *env, int regno, const char *reg_name);
13749 
check_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx_p)13750 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
13751 			    int *insn_idx_p)
13752 {
13753 	bool sleepable, rcu_lock, rcu_unlock, preempt_disable, preempt_enable;
13754 	u32 i, nargs, ptr_type_id, release_ref_obj_id;
13755 	struct bpf_reg_state *regs = cur_regs(env);
13756 	const char *func_name, *ptr_type_name;
13757 	const struct btf_type *t, *ptr_type;
13758 	struct bpf_kfunc_call_arg_meta meta;
13759 	struct bpf_insn_aux_data *insn_aux;
13760 	int err, insn_idx = *insn_idx_p;
13761 	const struct btf_param *args;
13762 	struct btf *desc_btf;
13763 
13764 	/* skip for now, but return error when we find this in fixup_kfunc_call */
13765 	if (!insn->imm)
13766 		return 0;
13767 
13768 	err = fetch_kfunc_meta(env, insn, &meta, &func_name);
13769 	if (err == -EACCES && func_name)
13770 		verbose(env, "calling kernel function %s is not allowed\n", func_name);
13771 	if (err)
13772 		return err;
13773 	desc_btf = meta.btf;
13774 	insn_aux = &env->insn_aux_data[insn_idx];
13775 
13776 	insn_aux->is_iter_next = is_iter_next_kfunc(&meta);
13777 
13778 	if (!insn->off &&
13779 	    (insn->imm == special_kfunc_list[KF_bpf_res_spin_lock] ||
13780 	     insn->imm == special_kfunc_list[KF_bpf_res_spin_lock_irqsave])) {
13781 		struct bpf_verifier_state *branch;
13782 		struct bpf_reg_state *regs;
13783 
13784 		branch = push_stack(env, env->insn_idx + 1, env->insn_idx, false);
13785 		if (!branch) {
13786 			verbose(env, "failed to push state for failed lock acquisition\n");
13787 			return -ENOMEM;
13788 		}
13789 
13790 		regs = branch->frame[branch->curframe]->regs;
13791 
13792 		/* Clear r0-r5 registers in forked state */
13793 		for (i = 0; i < CALLER_SAVED_REGS; i++)
13794 			mark_reg_not_init(env, regs, caller_saved[i]);
13795 
13796 		mark_reg_unknown(env, regs, BPF_REG_0);
13797 		err = __mark_reg_s32_range(env, regs, BPF_REG_0, -MAX_ERRNO, -1);
13798 		if (err) {
13799 			verbose(env, "failed to mark s32 range for retval in forked state for lock\n");
13800 			return err;
13801 		}
13802 		__mark_btf_func_reg_size(env, regs, BPF_REG_0, sizeof(u32));
13803 	} else if (!insn->off && insn->imm == special_kfunc_list[KF___bpf_trap]) {
13804 		verbose(env, "unexpected __bpf_trap() due to uninitialized variable?\n");
13805 		return -EFAULT;
13806 	}
13807 
13808 	if (is_kfunc_destructive(&meta) && !capable(CAP_SYS_BOOT)) {
13809 		verbose(env, "destructive kfunc calls require CAP_SYS_BOOT capability\n");
13810 		return -EACCES;
13811 	}
13812 
13813 	sleepable = is_kfunc_sleepable(&meta);
13814 	if (sleepable && !in_sleepable(env)) {
13815 		verbose(env, "program must be sleepable to call sleepable kfunc %s\n", func_name);
13816 		return -EACCES;
13817 	}
13818 
13819 	/* Check the arguments */
13820 	err = check_kfunc_args(env, &meta, insn_idx);
13821 	if (err < 0)
13822 		return err;
13823 
13824 	if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
13825 		err = push_callback_call(env, insn, insn_idx, meta.subprogno,
13826 					 set_rbtree_add_callback_state);
13827 		if (err) {
13828 			verbose(env, "kfunc %s#%d failed callback verification\n",
13829 				func_name, meta.func_id);
13830 			return err;
13831 		}
13832 	}
13833 
13834 	if (meta.func_id == special_kfunc_list[KF_bpf_session_cookie]) {
13835 		meta.r0_size = sizeof(u64);
13836 		meta.r0_rdonly = false;
13837 	}
13838 
13839 	if (is_bpf_wq_set_callback_impl_kfunc(meta.func_id)) {
13840 		err = push_callback_call(env, insn, insn_idx, meta.subprogno,
13841 					 set_timer_callback_state);
13842 		if (err) {
13843 			verbose(env, "kfunc %s#%d failed callback verification\n",
13844 				func_name, meta.func_id);
13845 			return err;
13846 		}
13847 	}
13848 
13849 	rcu_lock = is_kfunc_bpf_rcu_read_lock(&meta);
13850 	rcu_unlock = is_kfunc_bpf_rcu_read_unlock(&meta);
13851 
13852 	preempt_disable = is_kfunc_bpf_preempt_disable(&meta);
13853 	preempt_enable = is_kfunc_bpf_preempt_enable(&meta);
13854 
13855 	if (env->cur_state->active_rcu_lock) {
13856 		struct bpf_func_state *state;
13857 		struct bpf_reg_state *reg;
13858 		u32 clear_mask = (1 << STACK_SPILL) | (1 << STACK_ITER);
13859 
13860 		if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
13861 			verbose(env, "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
13862 			return -EACCES;
13863 		}
13864 
13865 		if (rcu_lock) {
13866 			verbose(env, "nested rcu read lock (kernel function %s)\n", func_name);
13867 			return -EINVAL;
13868 		} else if (rcu_unlock) {
13869 			bpf_for_each_reg_in_vstate_mask(env->cur_state, state, reg, clear_mask, ({
13870 				if (reg->type & MEM_RCU) {
13871 					reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
13872 					reg->type |= PTR_UNTRUSTED;
13873 				}
13874 			}));
13875 			env->cur_state->active_rcu_lock = false;
13876 		} else if (sleepable) {
13877 			verbose(env, "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
13878 			return -EACCES;
13879 		}
13880 	} else if (rcu_lock) {
13881 		env->cur_state->active_rcu_lock = true;
13882 	} else if (rcu_unlock) {
13883 		verbose(env, "unmatched rcu read unlock (kernel function %s)\n", func_name);
13884 		return -EINVAL;
13885 	}
13886 
13887 	if (env->cur_state->active_preempt_locks) {
13888 		if (preempt_disable) {
13889 			env->cur_state->active_preempt_locks++;
13890 		} else if (preempt_enable) {
13891 			env->cur_state->active_preempt_locks--;
13892 		} else if (sleepable) {
13893 			verbose(env, "kernel func %s is sleepable within non-preemptible region\n", func_name);
13894 			return -EACCES;
13895 		}
13896 	} else if (preempt_disable) {
13897 		env->cur_state->active_preempt_locks++;
13898 	} else if (preempt_enable) {
13899 		verbose(env, "unmatched attempt to enable preemption (kernel function %s)\n", func_name);
13900 		return -EINVAL;
13901 	}
13902 
13903 	if (env->cur_state->active_irq_id && sleepable) {
13904 		verbose(env, "kernel func %s is sleepable within IRQ-disabled region\n", func_name);
13905 		return -EACCES;
13906 	}
13907 
13908 	/* In case of release function, we get register number of refcounted
13909 	 * PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
13910 	 */
13911 	if (meta.release_regno) {
13912 		err = release_reference(env, regs[meta.release_regno].ref_obj_id);
13913 		if (err) {
13914 			verbose(env, "kfunc %s#%d reference has not been acquired before\n",
13915 				func_name, meta.func_id);
13916 			return err;
13917 		}
13918 	}
13919 
13920 	if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
13921 	    meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
13922 	    meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
13923 		release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
13924 		insn_aux->insert_off = regs[BPF_REG_2].off;
13925 		insn_aux->kptr_struct_meta = btf_find_struct_meta(meta.arg_btf, meta.arg_btf_id);
13926 		err = ref_convert_owning_non_owning(env, release_ref_obj_id);
13927 		if (err) {
13928 			verbose(env, "kfunc %s#%d conversion of owning ref to non-owning failed\n",
13929 				func_name, meta.func_id);
13930 			return err;
13931 		}
13932 
13933 		err = release_reference(env, release_ref_obj_id);
13934 		if (err) {
13935 			verbose(env, "kfunc %s#%d reference has not been acquired before\n",
13936 				func_name, meta.func_id);
13937 			return err;
13938 		}
13939 	}
13940 
13941 	if (meta.func_id == special_kfunc_list[KF_bpf_throw]) {
13942 		if (!bpf_jit_supports_exceptions()) {
13943 			verbose(env, "JIT does not support calling kfunc %s#%d\n",
13944 				func_name, meta.func_id);
13945 			return -ENOTSUPP;
13946 		}
13947 		env->seen_exception = true;
13948 
13949 		/* In the case of the default callback, the cookie value passed
13950 		 * to bpf_throw becomes the return value of the program.
13951 		 */
13952 		if (!env->exception_callback_subprog) {
13953 			err = check_return_code(env, BPF_REG_1, "R1");
13954 			if (err < 0)
13955 				return err;
13956 		}
13957 	}
13958 
13959 	for (i = 0; i < CALLER_SAVED_REGS; i++)
13960 		mark_reg_not_init(env, regs, caller_saved[i]);
13961 
13962 	/* Check return type */
13963 	t = btf_type_skip_modifiers(desc_btf, meta.func_proto->type, NULL);
13964 
13965 	if (is_kfunc_acquire(&meta) && !btf_type_is_struct_ptr(meta.btf, t)) {
13966 		/* Only exception is bpf_obj_new_impl */
13967 		if (meta.btf != btf_vmlinux ||
13968 		    (meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
13969 		     meta.func_id != special_kfunc_list[KF_bpf_percpu_obj_new_impl] &&
13970 		     meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
13971 			verbose(env, "acquire kernel function does not return PTR_TO_BTF_ID\n");
13972 			return -EINVAL;
13973 		}
13974 	}
13975 
13976 	if (btf_type_is_scalar(t)) {
13977 		mark_reg_unknown(env, regs, BPF_REG_0);
13978 		if (meta.btf == btf_vmlinux && (meta.func_id == special_kfunc_list[KF_bpf_res_spin_lock] ||
13979 		    meta.func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave]))
13980 			__mark_reg_const_zero(env, &regs[BPF_REG_0]);
13981 		mark_btf_func_reg_size(env, BPF_REG_0, t->size);
13982 	} else if (btf_type_is_ptr(t)) {
13983 		ptr_type = btf_type_skip_modifiers(desc_btf, t->type, &ptr_type_id);
13984 		err = check_special_kfunc(env, &meta, regs, insn_aux, ptr_type, desc_btf);
13985 		if (err) {
13986 			if (err < 0)
13987 				return err;
13988 		} else if (btf_type_is_void(ptr_type)) {
13989 			/* kfunc returning 'void *' is equivalent to returning scalar */
13990 			mark_reg_unknown(env, regs, BPF_REG_0);
13991 		} else if (!__btf_type_is_struct(ptr_type)) {
13992 			if (!meta.r0_size) {
13993 				__u32 sz;
13994 
13995 				if (!IS_ERR(btf_resolve_size(desc_btf, ptr_type, &sz))) {
13996 					meta.r0_size = sz;
13997 					meta.r0_rdonly = true;
13998 				}
13999 			}
14000 			if (!meta.r0_size) {
14001 				ptr_type_name = btf_name_by_offset(desc_btf,
14002 								   ptr_type->name_off);
14003 				verbose(env,
14004 					"kernel function %s returns pointer type %s %s is not supported\n",
14005 					func_name,
14006 					btf_type_str(ptr_type),
14007 					ptr_type_name);
14008 				return -EINVAL;
14009 			}
14010 
14011 			mark_reg_known_zero(env, regs, BPF_REG_0);
14012 			regs[BPF_REG_0].type = PTR_TO_MEM;
14013 			regs[BPF_REG_0].mem_size = meta.r0_size;
14014 
14015 			if (meta.r0_rdonly)
14016 				regs[BPF_REG_0].type |= MEM_RDONLY;
14017 
14018 			/* Ensures we don't access the memory after a release_reference() */
14019 			if (meta.ref_obj_id)
14020 				regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
14021 		} else {
14022 			mark_reg_known_zero(env, regs, BPF_REG_0);
14023 			regs[BPF_REG_0].btf = desc_btf;
14024 			regs[BPF_REG_0].type = PTR_TO_BTF_ID;
14025 			regs[BPF_REG_0].btf_id = ptr_type_id;
14026 
14027 			if (meta.func_id == special_kfunc_list[KF_bpf_get_kmem_cache])
14028 				regs[BPF_REG_0].type |= PTR_UNTRUSTED;
14029 
14030 			if (is_iter_next_kfunc(&meta)) {
14031 				struct bpf_reg_state *cur_iter;
14032 
14033 				cur_iter = get_iter_from_state(env->cur_state, &meta);
14034 
14035 				if (cur_iter->type & MEM_RCU) /* KF_RCU_PROTECTED */
14036 					regs[BPF_REG_0].type |= MEM_RCU;
14037 				else
14038 					regs[BPF_REG_0].type |= PTR_TRUSTED;
14039 			}
14040 		}
14041 
14042 		if (is_kfunc_ret_null(&meta)) {
14043 			regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
14044 			/* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
14045 			regs[BPF_REG_0].id = ++env->id_gen;
14046 		}
14047 		mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
14048 		if (is_kfunc_acquire(&meta)) {
14049 			int id = acquire_reference(env, insn_idx);
14050 
14051 			if (id < 0)
14052 				return id;
14053 			if (is_kfunc_ret_null(&meta))
14054 				regs[BPF_REG_0].id = id;
14055 			regs[BPF_REG_0].ref_obj_id = id;
14056 		} else if (is_rbtree_node_type(ptr_type) || is_list_node_type(ptr_type)) {
14057 			ref_set_non_owning(env, &regs[BPF_REG_0]);
14058 		}
14059 
14060 		if (reg_may_point_to_spin_lock(&regs[BPF_REG_0]) && !regs[BPF_REG_0].id)
14061 			regs[BPF_REG_0].id = ++env->id_gen;
14062 	} else if (btf_type_is_void(t)) {
14063 		if (meta.btf == btf_vmlinux) {
14064 			if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
14065 			    meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_drop_impl]) {
14066 				insn_aux->kptr_struct_meta =
14067 					btf_find_struct_meta(meta.arg_btf,
14068 							     meta.arg_btf_id);
14069 			}
14070 		}
14071 	}
14072 
14073 	nargs = btf_type_vlen(meta.func_proto);
14074 	args = (const struct btf_param *)(meta.func_proto + 1);
14075 	for (i = 0; i < nargs; i++) {
14076 		u32 regno = i + 1;
14077 
14078 		t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
14079 		if (btf_type_is_ptr(t))
14080 			mark_btf_func_reg_size(env, regno, sizeof(void *));
14081 		else
14082 			/* scalar. ensured by btf_check_kfunc_arg_match() */
14083 			mark_btf_func_reg_size(env, regno, t->size);
14084 	}
14085 
14086 	if (is_iter_next_kfunc(&meta)) {
14087 		err = process_iter_next_call(env, insn_idx, &meta);
14088 		if (err)
14089 			return err;
14090 	}
14091 
14092 	return 0;
14093 }
14094 
check_reg_sane_offset(struct bpf_verifier_env * env,const struct bpf_reg_state * reg,enum bpf_reg_type type)14095 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
14096 				  const struct bpf_reg_state *reg,
14097 				  enum bpf_reg_type type)
14098 {
14099 	bool known = tnum_is_const(reg->var_off);
14100 	s64 val = reg->var_off.value;
14101 	s64 smin = reg->smin_value;
14102 
14103 	if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
14104 		verbose(env, "math between %s pointer and %lld is not allowed\n",
14105 			reg_type_str(env, type), val);
14106 		return false;
14107 	}
14108 
14109 	if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
14110 		verbose(env, "%s pointer offset %d is not allowed\n",
14111 			reg_type_str(env, type), reg->off);
14112 		return false;
14113 	}
14114 
14115 	if (smin == S64_MIN) {
14116 		verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
14117 			reg_type_str(env, type));
14118 		return false;
14119 	}
14120 
14121 	if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
14122 		verbose(env, "value %lld makes %s pointer be out of bounds\n",
14123 			smin, reg_type_str(env, type));
14124 		return false;
14125 	}
14126 
14127 	return true;
14128 }
14129 
14130 enum {
14131 	REASON_BOUNDS	= -1,
14132 	REASON_TYPE	= -2,
14133 	REASON_PATHS	= -3,
14134 	REASON_LIMIT	= -4,
14135 	REASON_STACK	= -5,
14136 };
14137 
retrieve_ptr_limit(const struct bpf_reg_state * ptr_reg,u32 * alu_limit,bool mask_to_left)14138 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
14139 			      u32 *alu_limit, bool mask_to_left)
14140 {
14141 	u32 max = 0, ptr_limit = 0;
14142 
14143 	switch (ptr_reg->type) {
14144 	case PTR_TO_STACK:
14145 		/* Offset 0 is out-of-bounds, but acceptable start for the
14146 		 * left direction, see BPF_REG_FP. Also, unknown scalar
14147 		 * offset where we would need to deal with min/max bounds is
14148 		 * currently prohibited for unprivileged.
14149 		 */
14150 		max = MAX_BPF_STACK + mask_to_left;
14151 		ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
14152 		break;
14153 	case PTR_TO_MAP_VALUE:
14154 		max = ptr_reg->map_ptr->value_size;
14155 		ptr_limit = (mask_to_left ?
14156 			     ptr_reg->smin_value :
14157 			     ptr_reg->umax_value) + ptr_reg->off;
14158 		break;
14159 	default:
14160 		return REASON_TYPE;
14161 	}
14162 
14163 	if (ptr_limit >= max)
14164 		return REASON_LIMIT;
14165 	*alu_limit = ptr_limit;
14166 	return 0;
14167 }
14168 
can_skip_alu_sanitation(const struct bpf_verifier_env * env,const struct bpf_insn * insn)14169 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
14170 				    const struct bpf_insn *insn)
14171 {
14172 	return env->bypass_spec_v1 ||
14173 		BPF_SRC(insn->code) == BPF_K ||
14174 		cur_aux(env)->nospec;
14175 }
14176 
update_alu_sanitation_state(struct bpf_insn_aux_data * aux,u32 alu_state,u32 alu_limit)14177 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
14178 				       u32 alu_state, u32 alu_limit)
14179 {
14180 	/* If we arrived here from different branches with different
14181 	 * state or limits to sanitize, then this won't work.
14182 	 */
14183 	if (aux->alu_state &&
14184 	    (aux->alu_state != alu_state ||
14185 	     aux->alu_limit != alu_limit))
14186 		return REASON_PATHS;
14187 
14188 	/* Corresponding fixup done in do_misc_fixups(). */
14189 	aux->alu_state = alu_state;
14190 	aux->alu_limit = alu_limit;
14191 	return 0;
14192 }
14193 
sanitize_val_alu(struct bpf_verifier_env * env,struct bpf_insn * insn)14194 static int sanitize_val_alu(struct bpf_verifier_env *env,
14195 			    struct bpf_insn *insn)
14196 {
14197 	struct bpf_insn_aux_data *aux = cur_aux(env);
14198 
14199 	if (can_skip_alu_sanitation(env, insn))
14200 		return 0;
14201 
14202 	return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
14203 }
14204 
sanitize_needed(u8 opcode)14205 static bool sanitize_needed(u8 opcode)
14206 {
14207 	return opcode == BPF_ADD || opcode == BPF_SUB;
14208 }
14209 
14210 struct bpf_sanitize_info {
14211 	struct bpf_insn_aux_data aux;
14212 	bool mask_to_left;
14213 };
14214 
14215 static struct bpf_verifier_state *
sanitize_speculative_path(struct bpf_verifier_env * env,const struct bpf_insn * insn,u32 next_idx,u32 curr_idx)14216 sanitize_speculative_path(struct bpf_verifier_env *env,
14217 			  const struct bpf_insn *insn,
14218 			  u32 next_idx, u32 curr_idx)
14219 {
14220 	struct bpf_verifier_state *branch;
14221 	struct bpf_reg_state *regs;
14222 
14223 	branch = push_stack(env, next_idx, curr_idx, true);
14224 	if (branch && insn) {
14225 		regs = branch->frame[branch->curframe]->regs;
14226 		if (BPF_SRC(insn->code) == BPF_K) {
14227 			mark_reg_unknown(env, regs, insn->dst_reg);
14228 		} else if (BPF_SRC(insn->code) == BPF_X) {
14229 			mark_reg_unknown(env, regs, insn->dst_reg);
14230 			mark_reg_unknown(env, regs, insn->src_reg);
14231 		}
14232 	}
14233 	return branch;
14234 }
14235 
sanitize_ptr_alu(struct bpf_verifier_env * env,struct bpf_insn * insn,const struct bpf_reg_state * ptr_reg,const struct bpf_reg_state * off_reg,struct bpf_reg_state * dst_reg,struct bpf_sanitize_info * info,const bool commit_window)14236 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
14237 			    struct bpf_insn *insn,
14238 			    const struct bpf_reg_state *ptr_reg,
14239 			    const struct bpf_reg_state *off_reg,
14240 			    struct bpf_reg_state *dst_reg,
14241 			    struct bpf_sanitize_info *info,
14242 			    const bool commit_window)
14243 {
14244 	struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
14245 	struct bpf_verifier_state *vstate = env->cur_state;
14246 	bool off_is_imm = tnum_is_const(off_reg->var_off);
14247 	bool off_is_neg = off_reg->smin_value < 0;
14248 	bool ptr_is_dst_reg = ptr_reg == dst_reg;
14249 	u8 opcode = BPF_OP(insn->code);
14250 	u32 alu_state, alu_limit;
14251 	struct bpf_reg_state tmp;
14252 	bool ret;
14253 	int err;
14254 
14255 	if (can_skip_alu_sanitation(env, insn))
14256 		return 0;
14257 
14258 	/* We already marked aux for masking from non-speculative
14259 	 * paths, thus we got here in the first place. We only care
14260 	 * to explore bad access from here.
14261 	 */
14262 	if (vstate->speculative)
14263 		goto do_sim;
14264 
14265 	if (!commit_window) {
14266 		if (!tnum_is_const(off_reg->var_off) &&
14267 		    (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
14268 			return REASON_BOUNDS;
14269 
14270 		info->mask_to_left = (opcode == BPF_ADD &&  off_is_neg) ||
14271 				     (opcode == BPF_SUB && !off_is_neg);
14272 	}
14273 
14274 	err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
14275 	if (err < 0)
14276 		return err;
14277 
14278 	if (commit_window) {
14279 		/* In commit phase we narrow the masking window based on
14280 		 * the observed pointer move after the simulated operation.
14281 		 */
14282 		alu_state = info->aux.alu_state;
14283 		alu_limit = abs(info->aux.alu_limit - alu_limit);
14284 	} else {
14285 		alu_state  = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
14286 		alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
14287 		alu_state |= ptr_is_dst_reg ?
14288 			     BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
14289 
14290 		/* Limit pruning on unknown scalars to enable deep search for
14291 		 * potential masking differences from other program paths.
14292 		 */
14293 		if (!off_is_imm)
14294 			env->explore_alu_limits = true;
14295 	}
14296 
14297 	err = update_alu_sanitation_state(aux, alu_state, alu_limit);
14298 	if (err < 0)
14299 		return err;
14300 do_sim:
14301 	/* If we're in commit phase, we're done here given we already
14302 	 * pushed the truncated dst_reg into the speculative verification
14303 	 * stack.
14304 	 *
14305 	 * Also, when register is a known constant, we rewrite register-based
14306 	 * operation to immediate-based, and thus do not need masking (and as
14307 	 * a consequence, do not need to simulate the zero-truncation either).
14308 	 */
14309 	if (commit_window || off_is_imm)
14310 		return 0;
14311 
14312 	/* Simulate and find potential out-of-bounds access under
14313 	 * speculative execution from truncation as a result of
14314 	 * masking when off was not within expected range. If off
14315 	 * sits in dst, then we temporarily need to move ptr there
14316 	 * to simulate dst (== 0) +/-= ptr. Needed, for example,
14317 	 * for cases where we use K-based arithmetic in one direction
14318 	 * and truncated reg-based in the other in order to explore
14319 	 * bad access.
14320 	 */
14321 	if (!ptr_is_dst_reg) {
14322 		tmp = *dst_reg;
14323 		copy_register_state(dst_reg, ptr_reg);
14324 	}
14325 	ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
14326 					env->insn_idx);
14327 	if (!ptr_is_dst_reg && ret)
14328 		*dst_reg = tmp;
14329 	return !ret ? REASON_STACK : 0;
14330 }
14331 
sanitize_mark_insn_seen(struct bpf_verifier_env * env)14332 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
14333 {
14334 	struct bpf_verifier_state *vstate = env->cur_state;
14335 
14336 	/* If we simulate paths under speculation, we don't update the
14337 	 * insn as 'seen' such that when we verify unreachable paths in
14338 	 * the non-speculative domain, sanitize_dead_code() can still
14339 	 * rewrite/sanitize them.
14340 	 */
14341 	if (!vstate->speculative)
14342 		env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
14343 }
14344 
sanitize_err(struct bpf_verifier_env * env,const struct bpf_insn * insn,int reason,const struct bpf_reg_state * off_reg,const struct bpf_reg_state * dst_reg)14345 static int sanitize_err(struct bpf_verifier_env *env,
14346 			const struct bpf_insn *insn, int reason,
14347 			const struct bpf_reg_state *off_reg,
14348 			const struct bpf_reg_state *dst_reg)
14349 {
14350 	static const char *err = "pointer arithmetic with it prohibited for !root";
14351 	const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
14352 	u32 dst = insn->dst_reg, src = insn->src_reg;
14353 
14354 	switch (reason) {
14355 	case REASON_BOUNDS:
14356 		verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
14357 			off_reg == dst_reg ? dst : src, err);
14358 		break;
14359 	case REASON_TYPE:
14360 		verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
14361 			off_reg == dst_reg ? src : dst, err);
14362 		break;
14363 	case REASON_PATHS:
14364 		verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
14365 			dst, op, err);
14366 		break;
14367 	case REASON_LIMIT:
14368 		verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
14369 			dst, op, err);
14370 		break;
14371 	case REASON_STACK:
14372 		verbose(env, "R%d could not be pushed for speculative verification, %s\n",
14373 			dst, err);
14374 		return -ENOMEM;
14375 	default:
14376 		verifier_bug(env, "unknown reason (%d)", reason);
14377 		break;
14378 	}
14379 
14380 	return -EACCES;
14381 }
14382 
14383 /* check that stack access falls within stack limits and that 'reg' doesn't
14384  * have a variable offset.
14385  *
14386  * Variable offset is prohibited for unprivileged mode for simplicity since it
14387  * requires corresponding support in Spectre masking for stack ALU.  See also
14388  * retrieve_ptr_limit().
14389  *
14390  *
14391  * 'off' includes 'reg->off'.
14392  */
check_stack_access_for_ptr_arithmetic(struct bpf_verifier_env * env,int regno,const struct bpf_reg_state * reg,int off)14393 static int check_stack_access_for_ptr_arithmetic(
14394 				struct bpf_verifier_env *env,
14395 				int regno,
14396 				const struct bpf_reg_state *reg,
14397 				int off)
14398 {
14399 	if (!tnum_is_const(reg->var_off)) {
14400 		char tn_buf[48];
14401 
14402 		tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
14403 		verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
14404 			regno, tn_buf, off);
14405 		return -EACCES;
14406 	}
14407 
14408 	if (off >= 0 || off < -MAX_BPF_STACK) {
14409 		verbose(env, "R%d stack pointer arithmetic goes out of range, "
14410 			"prohibited for !root; off=%d\n", regno, off);
14411 		return -EACCES;
14412 	}
14413 
14414 	return 0;
14415 }
14416 
sanitize_check_bounds(struct bpf_verifier_env * env,const struct bpf_insn * insn,const struct bpf_reg_state * dst_reg)14417 static int sanitize_check_bounds(struct bpf_verifier_env *env,
14418 				 const struct bpf_insn *insn,
14419 				 const struct bpf_reg_state *dst_reg)
14420 {
14421 	u32 dst = insn->dst_reg;
14422 
14423 	/* For unprivileged we require that resulting offset must be in bounds
14424 	 * in order to be able to sanitize access later on.
14425 	 */
14426 	if (env->bypass_spec_v1)
14427 		return 0;
14428 
14429 	switch (dst_reg->type) {
14430 	case PTR_TO_STACK:
14431 		if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
14432 					dst_reg->off + dst_reg->var_off.value))
14433 			return -EACCES;
14434 		break;
14435 	case PTR_TO_MAP_VALUE:
14436 		if (check_map_access(env, dst, dst_reg->off, 1, false, ACCESS_HELPER)) {
14437 			verbose(env, "R%d pointer arithmetic of map value goes out of range, "
14438 				"prohibited for !root\n", dst);
14439 			return -EACCES;
14440 		}
14441 		break;
14442 	default:
14443 		return -EOPNOTSUPP;
14444 	}
14445 
14446 	return 0;
14447 }
14448 
14449 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
14450  * Caller should also handle BPF_MOV case separately.
14451  * If we return -EACCES, caller may want to try again treating pointer as a
14452  * scalar.  So we only emit a diagnostic if !env->allow_ptr_leaks.
14453  */
adjust_ptr_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn,const struct bpf_reg_state * ptr_reg,const struct bpf_reg_state * off_reg)14454 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
14455 				   struct bpf_insn *insn,
14456 				   const struct bpf_reg_state *ptr_reg,
14457 				   const struct bpf_reg_state *off_reg)
14458 {
14459 	struct bpf_verifier_state *vstate = env->cur_state;
14460 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
14461 	struct bpf_reg_state *regs = state->regs, *dst_reg;
14462 	bool known = tnum_is_const(off_reg->var_off);
14463 	s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
14464 	    smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
14465 	u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
14466 	    umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
14467 	struct bpf_sanitize_info info = {};
14468 	u8 opcode = BPF_OP(insn->code);
14469 	u32 dst = insn->dst_reg;
14470 	int ret, bounds_ret;
14471 
14472 	dst_reg = &regs[dst];
14473 
14474 	if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
14475 	    smin_val > smax_val || umin_val > umax_val) {
14476 		/* Taint dst register if offset had invalid bounds derived from
14477 		 * e.g. dead branches.
14478 		 */
14479 		__mark_reg_unknown(env, dst_reg);
14480 		return 0;
14481 	}
14482 
14483 	if (BPF_CLASS(insn->code) != BPF_ALU64) {
14484 		/* 32-bit ALU ops on pointers produce (meaningless) scalars */
14485 		if (opcode == BPF_SUB && env->allow_ptr_leaks) {
14486 			__mark_reg_unknown(env, dst_reg);
14487 			return 0;
14488 		}
14489 
14490 		verbose(env,
14491 			"R%d 32-bit pointer arithmetic prohibited\n",
14492 			dst);
14493 		return -EACCES;
14494 	}
14495 
14496 	if (ptr_reg->type & PTR_MAYBE_NULL) {
14497 		verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
14498 			dst, reg_type_str(env, ptr_reg->type));
14499 		return -EACCES;
14500 	}
14501 
14502 	/*
14503 	 * Accesses to untrusted PTR_TO_MEM are done through probe
14504 	 * instructions, hence no need to track offsets.
14505 	 */
14506 	if (base_type(ptr_reg->type) == PTR_TO_MEM && (ptr_reg->type & PTR_UNTRUSTED))
14507 		return 0;
14508 
14509 	switch (base_type(ptr_reg->type)) {
14510 	case PTR_TO_CTX:
14511 	case PTR_TO_MAP_VALUE:
14512 	case PTR_TO_MAP_KEY:
14513 	case PTR_TO_STACK:
14514 	case PTR_TO_PACKET_META:
14515 	case PTR_TO_PACKET:
14516 	case PTR_TO_TP_BUFFER:
14517 	case PTR_TO_BTF_ID:
14518 	case PTR_TO_MEM:
14519 	case PTR_TO_BUF:
14520 	case PTR_TO_FUNC:
14521 	case CONST_PTR_TO_DYNPTR:
14522 		break;
14523 	case PTR_TO_FLOW_KEYS:
14524 		if (known)
14525 			break;
14526 		fallthrough;
14527 	case CONST_PTR_TO_MAP:
14528 		/* smin_val represents the known value */
14529 		if (known && smin_val == 0 && opcode == BPF_ADD)
14530 			break;
14531 		fallthrough;
14532 	default:
14533 		verbose(env, "R%d pointer arithmetic on %s prohibited\n",
14534 			dst, reg_type_str(env, ptr_reg->type));
14535 		return -EACCES;
14536 	}
14537 
14538 	/* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
14539 	 * The id may be overwritten later if we create a new variable offset.
14540 	 */
14541 	dst_reg->type = ptr_reg->type;
14542 	dst_reg->id = ptr_reg->id;
14543 
14544 	if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
14545 	    !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
14546 		return -EINVAL;
14547 
14548 	/* pointer types do not carry 32-bit bounds at the moment. */
14549 	__mark_reg32_unbounded(dst_reg);
14550 
14551 	if (sanitize_needed(opcode)) {
14552 		ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
14553 				       &info, false);
14554 		if (ret < 0)
14555 			return sanitize_err(env, insn, ret, off_reg, dst_reg);
14556 	}
14557 
14558 	switch (opcode) {
14559 	case BPF_ADD:
14560 		/* We can take a fixed offset as long as it doesn't overflow
14561 		 * the s32 'off' field
14562 		 */
14563 		if (known && (ptr_reg->off + smin_val ==
14564 			      (s64)(s32)(ptr_reg->off + smin_val))) {
14565 			/* pointer += K.  Accumulate it into fixed offset */
14566 			dst_reg->smin_value = smin_ptr;
14567 			dst_reg->smax_value = smax_ptr;
14568 			dst_reg->umin_value = umin_ptr;
14569 			dst_reg->umax_value = umax_ptr;
14570 			dst_reg->var_off = ptr_reg->var_off;
14571 			dst_reg->off = ptr_reg->off + smin_val;
14572 			dst_reg->raw = ptr_reg->raw;
14573 			break;
14574 		}
14575 		/* A new variable offset is created.  Note that off_reg->off
14576 		 * == 0, since it's a scalar.
14577 		 * dst_reg gets the pointer type and since some positive
14578 		 * integer value was added to the pointer, give it a new 'id'
14579 		 * if it's a PTR_TO_PACKET.
14580 		 * this creates a new 'base' pointer, off_reg (variable) gets
14581 		 * added into the variable offset, and we copy the fixed offset
14582 		 * from ptr_reg.
14583 		 */
14584 		if (check_add_overflow(smin_ptr, smin_val, &dst_reg->smin_value) ||
14585 		    check_add_overflow(smax_ptr, smax_val, &dst_reg->smax_value)) {
14586 			dst_reg->smin_value = S64_MIN;
14587 			dst_reg->smax_value = S64_MAX;
14588 		}
14589 		if (check_add_overflow(umin_ptr, umin_val, &dst_reg->umin_value) ||
14590 		    check_add_overflow(umax_ptr, umax_val, &dst_reg->umax_value)) {
14591 			dst_reg->umin_value = 0;
14592 			dst_reg->umax_value = U64_MAX;
14593 		}
14594 		dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
14595 		dst_reg->off = ptr_reg->off;
14596 		dst_reg->raw = ptr_reg->raw;
14597 		if (reg_is_pkt_pointer(ptr_reg)) {
14598 			dst_reg->id = ++env->id_gen;
14599 			/* something was added to pkt_ptr, set range to zero */
14600 			memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
14601 		}
14602 		break;
14603 	case BPF_SUB:
14604 		if (dst_reg == off_reg) {
14605 			/* scalar -= pointer.  Creates an unknown scalar */
14606 			verbose(env, "R%d tried to subtract pointer from scalar\n",
14607 				dst);
14608 			return -EACCES;
14609 		}
14610 		/* We don't allow subtraction from FP, because (according to
14611 		 * test_verifier.c test "invalid fp arithmetic", JITs might not
14612 		 * be able to deal with it.
14613 		 */
14614 		if (ptr_reg->type == PTR_TO_STACK) {
14615 			verbose(env, "R%d subtraction from stack pointer prohibited\n",
14616 				dst);
14617 			return -EACCES;
14618 		}
14619 		if (known && (ptr_reg->off - smin_val ==
14620 			      (s64)(s32)(ptr_reg->off - smin_val))) {
14621 			/* pointer -= K.  Subtract it from fixed offset */
14622 			dst_reg->smin_value = smin_ptr;
14623 			dst_reg->smax_value = smax_ptr;
14624 			dst_reg->umin_value = umin_ptr;
14625 			dst_reg->umax_value = umax_ptr;
14626 			dst_reg->var_off = ptr_reg->var_off;
14627 			dst_reg->id = ptr_reg->id;
14628 			dst_reg->off = ptr_reg->off - smin_val;
14629 			dst_reg->raw = ptr_reg->raw;
14630 			break;
14631 		}
14632 		/* A new variable offset is created.  If the subtrahend is known
14633 		 * nonnegative, then any reg->range we had before is still good.
14634 		 */
14635 		if (check_sub_overflow(smin_ptr, smax_val, &dst_reg->smin_value) ||
14636 		    check_sub_overflow(smax_ptr, smin_val, &dst_reg->smax_value)) {
14637 			/* Overflow possible, we know nothing */
14638 			dst_reg->smin_value = S64_MIN;
14639 			dst_reg->smax_value = S64_MAX;
14640 		}
14641 		if (umin_ptr < umax_val) {
14642 			/* Overflow possible, we know nothing */
14643 			dst_reg->umin_value = 0;
14644 			dst_reg->umax_value = U64_MAX;
14645 		} else {
14646 			/* Cannot overflow (as long as bounds are consistent) */
14647 			dst_reg->umin_value = umin_ptr - umax_val;
14648 			dst_reg->umax_value = umax_ptr - umin_val;
14649 		}
14650 		dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
14651 		dst_reg->off = ptr_reg->off;
14652 		dst_reg->raw = ptr_reg->raw;
14653 		if (reg_is_pkt_pointer(ptr_reg)) {
14654 			dst_reg->id = ++env->id_gen;
14655 			/* something was added to pkt_ptr, set range to zero */
14656 			if (smin_val < 0)
14657 				memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
14658 		}
14659 		break;
14660 	case BPF_AND:
14661 	case BPF_OR:
14662 	case BPF_XOR:
14663 		/* bitwise ops on pointers are troublesome, prohibit. */
14664 		verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
14665 			dst, bpf_alu_string[opcode >> 4]);
14666 		return -EACCES;
14667 	default:
14668 		/* other operators (e.g. MUL,LSH) produce non-pointer results */
14669 		verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
14670 			dst, bpf_alu_string[opcode >> 4]);
14671 		return -EACCES;
14672 	}
14673 
14674 	if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
14675 		return -EINVAL;
14676 	reg_bounds_sync(dst_reg);
14677 	bounds_ret = sanitize_check_bounds(env, insn, dst_reg);
14678 	if (bounds_ret == -EACCES)
14679 		return bounds_ret;
14680 	if (sanitize_needed(opcode)) {
14681 		ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
14682 				       &info, true);
14683 		if (verifier_bug_if(!can_skip_alu_sanitation(env, insn)
14684 				    && !env->cur_state->speculative
14685 				    && bounds_ret
14686 				    && !ret,
14687 				    env, "Pointer type unsupported by sanitize_check_bounds() not rejected by retrieve_ptr_limit() as required")) {
14688 			return -EFAULT;
14689 		}
14690 		if (ret < 0)
14691 			return sanitize_err(env, insn, ret, off_reg, dst_reg);
14692 	}
14693 
14694 	return 0;
14695 }
14696 
scalar32_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14697 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
14698 				 struct bpf_reg_state *src_reg)
14699 {
14700 	s32 *dst_smin = &dst_reg->s32_min_value;
14701 	s32 *dst_smax = &dst_reg->s32_max_value;
14702 	u32 *dst_umin = &dst_reg->u32_min_value;
14703 	u32 *dst_umax = &dst_reg->u32_max_value;
14704 	u32 umin_val = src_reg->u32_min_value;
14705 	u32 umax_val = src_reg->u32_max_value;
14706 	bool min_overflow, max_overflow;
14707 
14708 	if (check_add_overflow(*dst_smin, src_reg->s32_min_value, dst_smin) ||
14709 	    check_add_overflow(*dst_smax, src_reg->s32_max_value, dst_smax)) {
14710 		*dst_smin = S32_MIN;
14711 		*dst_smax = S32_MAX;
14712 	}
14713 
14714 	/* If either all additions overflow or no additions overflow, then
14715 	 * it is okay to set: dst_umin = dst_umin + src_umin, dst_umax =
14716 	 * dst_umax + src_umax. Otherwise (some additions overflow), set
14717 	 * the output bounds to unbounded.
14718 	 */
14719 	min_overflow = check_add_overflow(*dst_umin, umin_val, dst_umin);
14720 	max_overflow = check_add_overflow(*dst_umax, umax_val, dst_umax);
14721 
14722 	if (!min_overflow && max_overflow) {
14723 		*dst_umin = 0;
14724 		*dst_umax = U32_MAX;
14725 	}
14726 }
14727 
scalar_min_max_add(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14728 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
14729 			       struct bpf_reg_state *src_reg)
14730 {
14731 	s64 *dst_smin = &dst_reg->smin_value;
14732 	s64 *dst_smax = &dst_reg->smax_value;
14733 	u64 *dst_umin = &dst_reg->umin_value;
14734 	u64 *dst_umax = &dst_reg->umax_value;
14735 	u64 umin_val = src_reg->umin_value;
14736 	u64 umax_val = src_reg->umax_value;
14737 	bool min_overflow, max_overflow;
14738 
14739 	if (check_add_overflow(*dst_smin, src_reg->smin_value, dst_smin) ||
14740 	    check_add_overflow(*dst_smax, src_reg->smax_value, dst_smax)) {
14741 		*dst_smin = S64_MIN;
14742 		*dst_smax = S64_MAX;
14743 	}
14744 
14745 	/* If either all additions overflow or no additions overflow, then
14746 	 * it is okay to set: dst_umin = dst_umin + src_umin, dst_umax =
14747 	 * dst_umax + src_umax. Otherwise (some additions overflow), set
14748 	 * the output bounds to unbounded.
14749 	 */
14750 	min_overflow = check_add_overflow(*dst_umin, umin_val, dst_umin);
14751 	max_overflow = check_add_overflow(*dst_umax, umax_val, dst_umax);
14752 
14753 	if (!min_overflow && max_overflow) {
14754 		*dst_umin = 0;
14755 		*dst_umax = U64_MAX;
14756 	}
14757 }
14758 
scalar32_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14759 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
14760 				 struct bpf_reg_state *src_reg)
14761 {
14762 	s32 *dst_smin = &dst_reg->s32_min_value;
14763 	s32 *dst_smax = &dst_reg->s32_max_value;
14764 	u32 *dst_umin = &dst_reg->u32_min_value;
14765 	u32 *dst_umax = &dst_reg->u32_max_value;
14766 	u32 umin_val = src_reg->u32_min_value;
14767 	u32 umax_val = src_reg->u32_max_value;
14768 	bool min_underflow, max_underflow;
14769 
14770 	if (check_sub_overflow(*dst_smin, src_reg->s32_max_value, dst_smin) ||
14771 	    check_sub_overflow(*dst_smax, src_reg->s32_min_value, dst_smax)) {
14772 		/* Overflow possible, we know nothing */
14773 		*dst_smin = S32_MIN;
14774 		*dst_smax = S32_MAX;
14775 	}
14776 
14777 	/* If either all subtractions underflow or no subtractions
14778 	 * underflow, it is okay to set: dst_umin = dst_umin - src_umax,
14779 	 * dst_umax = dst_umax - src_umin. Otherwise (some subtractions
14780 	 * underflow), set the output bounds to unbounded.
14781 	 */
14782 	min_underflow = check_sub_overflow(*dst_umin, umax_val, dst_umin);
14783 	max_underflow = check_sub_overflow(*dst_umax, umin_val, dst_umax);
14784 
14785 	if (min_underflow && !max_underflow) {
14786 		*dst_umin = 0;
14787 		*dst_umax = U32_MAX;
14788 	}
14789 }
14790 
scalar_min_max_sub(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14791 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
14792 			       struct bpf_reg_state *src_reg)
14793 {
14794 	s64 *dst_smin = &dst_reg->smin_value;
14795 	s64 *dst_smax = &dst_reg->smax_value;
14796 	u64 *dst_umin = &dst_reg->umin_value;
14797 	u64 *dst_umax = &dst_reg->umax_value;
14798 	u64 umin_val = src_reg->umin_value;
14799 	u64 umax_val = src_reg->umax_value;
14800 	bool min_underflow, max_underflow;
14801 
14802 	if (check_sub_overflow(*dst_smin, src_reg->smax_value, dst_smin) ||
14803 	    check_sub_overflow(*dst_smax, src_reg->smin_value, dst_smax)) {
14804 		/* Overflow possible, we know nothing */
14805 		*dst_smin = S64_MIN;
14806 		*dst_smax = S64_MAX;
14807 	}
14808 
14809 	/* If either all subtractions underflow or no subtractions
14810 	 * underflow, it is okay to set: dst_umin = dst_umin - src_umax,
14811 	 * dst_umax = dst_umax - src_umin. Otherwise (some subtractions
14812 	 * underflow), set the output bounds to unbounded.
14813 	 */
14814 	min_underflow = check_sub_overflow(*dst_umin, umax_val, dst_umin);
14815 	max_underflow = check_sub_overflow(*dst_umax, umin_val, dst_umax);
14816 
14817 	if (min_underflow && !max_underflow) {
14818 		*dst_umin = 0;
14819 		*dst_umax = U64_MAX;
14820 	}
14821 }
14822 
scalar32_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14823 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
14824 				 struct bpf_reg_state *src_reg)
14825 {
14826 	s32 *dst_smin = &dst_reg->s32_min_value;
14827 	s32 *dst_smax = &dst_reg->s32_max_value;
14828 	u32 *dst_umin = &dst_reg->u32_min_value;
14829 	u32 *dst_umax = &dst_reg->u32_max_value;
14830 	s32 tmp_prod[4];
14831 
14832 	if (check_mul_overflow(*dst_umax, src_reg->u32_max_value, dst_umax) ||
14833 	    check_mul_overflow(*dst_umin, src_reg->u32_min_value, dst_umin)) {
14834 		/* Overflow possible, we know nothing */
14835 		*dst_umin = 0;
14836 		*dst_umax = U32_MAX;
14837 	}
14838 	if (check_mul_overflow(*dst_smin, src_reg->s32_min_value, &tmp_prod[0]) ||
14839 	    check_mul_overflow(*dst_smin, src_reg->s32_max_value, &tmp_prod[1]) ||
14840 	    check_mul_overflow(*dst_smax, src_reg->s32_min_value, &tmp_prod[2]) ||
14841 	    check_mul_overflow(*dst_smax, src_reg->s32_max_value, &tmp_prod[3])) {
14842 		/* Overflow possible, we know nothing */
14843 		*dst_smin = S32_MIN;
14844 		*dst_smax = S32_MAX;
14845 	} else {
14846 		*dst_smin = min_array(tmp_prod, 4);
14847 		*dst_smax = max_array(tmp_prod, 4);
14848 	}
14849 }
14850 
scalar_min_max_mul(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14851 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
14852 			       struct bpf_reg_state *src_reg)
14853 {
14854 	s64 *dst_smin = &dst_reg->smin_value;
14855 	s64 *dst_smax = &dst_reg->smax_value;
14856 	u64 *dst_umin = &dst_reg->umin_value;
14857 	u64 *dst_umax = &dst_reg->umax_value;
14858 	s64 tmp_prod[4];
14859 
14860 	if (check_mul_overflow(*dst_umax, src_reg->umax_value, dst_umax) ||
14861 	    check_mul_overflow(*dst_umin, src_reg->umin_value, dst_umin)) {
14862 		/* Overflow possible, we know nothing */
14863 		*dst_umin = 0;
14864 		*dst_umax = U64_MAX;
14865 	}
14866 	if (check_mul_overflow(*dst_smin, src_reg->smin_value, &tmp_prod[0]) ||
14867 	    check_mul_overflow(*dst_smin, src_reg->smax_value, &tmp_prod[1]) ||
14868 	    check_mul_overflow(*dst_smax, src_reg->smin_value, &tmp_prod[2]) ||
14869 	    check_mul_overflow(*dst_smax, src_reg->smax_value, &tmp_prod[3])) {
14870 		/* Overflow possible, we know nothing */
14871 		*dst_smin = S64_MIN;
14872 		*dst_smax = S64_MAX;
14873 	} else {
14874 		*dst_smin = min_array(tmp_prod, 4);
14875 		*dst_smax = max_array(tmp_prod, 4);
14876 	}
14877 }
14878 
scalar32_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14879 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
14880 				 struct bpf_reg_state *src_reg)
14881 {
14882 	bool src_known = tnum_subreg_is_const(src_reg->var_off);
14883 	bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
14884 	struct tnum var32_off = tnum_subreg(dst_reg->var_off);
14885 	u32 umax_val = src_reg->u32_max_value;
14886 
14887 	if (src_known && dst_known) {
14888 		__mark_reg32_known(dst_reg, var32_off.value);
14889 		return;
14890 	}
14891 
14892 	/* We get our minimum from the var_off, since that's inherently
14893 	 * bitwise.  Our maximum is the minimum of the operands' maxima.
14894 	 */
14895 	dst_reg->u32_min_value = var32_off.value;
14896 	dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
14897 
14898 	/* Safe to set s32 bounds by casting u32 result into s32 when u32
14899 	 * doesn't cross sign boundary. Otherwise set s32 bounds to unbounded.
14900 	 */
14901 	if ((s32)dst_reg->u32_min_value <= (s32)dst_reg->u32_max_value) {
14902 		dst_reg->s32_min_value = dst_reg->u32_min_value;
14903 		dst_reg->s32_max_value = dst_reg->u32_max_value;
14904 	} else {
14905 		dst_reg->s32_min_value = S32_MIN;
14906 		dst_reg->s32_max_value = S32_MAX;
14907 	}
14908 }
14909 
scalar_min_max_and(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14910 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
14911 			       struct bpf_reg_state *src_reg)
14912 {
14913 	bool src_known = tnum_is_const(src_reg->var_off);
14914 	bool dst_known = tnum_is_const(dst_reg->var_off);
14915 	u64 umax_val = src_reg->umax_value;
14916 
14917 	if (src_known && dst_known) {
14918 		__mark_reg_known(dst_reg, dst_reg->var_off.value);
14919 		return;
14920 	}
14921 
14922 	/* We get our minimum from the var_off, since that's inherently
14923 	 * bitwise.  Our maximum is the minimum of the operands' maxima.
14924 	 */
14925 	dst_reg->umin_value = dst_reg->var_off.value;
14926 	dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
14927 
14928 	/* Safe to set s64 bounds by casting u64 result into s64 when u64
14929 	 * doesn't cross sign boundary. Otherwise set s64 bounds to unbounded.
14930 	 */
14931 	if ((s64)dst_reg->umin_value <= (s64)dst_reg->umax_value) {
14932 		dst_reg->smin_value = dst_reg->umin_value;
14933 		dst_reg->smax_value = dst_reg->umax_value;
14934 	} else {
14935 		dst_reg->smin_value = S64_MIN;
14936 		dst_reg->smax_value = S64_MAX;
14937 	}
14938 	/* We may learn something more from the var_off */
14939 	__update_reg_bounds(dst_reg);
14940 }
14941 
scalar32_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14942 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
14943 				struct bpf_reg_state *src_reg)
14944 {
14945 	bool src_known = tnum_subreg_is_const(src_reg->var_off);
14946 	bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
14947 	struct tnum var32_off = tnum_subreg(dst_reg->var_off);
14948 	u32 umin_val = src_reg->u32_min_value;
14949 
14950 	if (src_known && dst_known) {
14951 		__mark_reg32_known(dst_reg, var32_off.value);
14952 		return;
14953 	}
14954 
14955 	/* We get our maximum from the var_off, and our minimum is the
14956 	 * maximum of the operands' minima
14957 	 */
14958 	dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
14959 	dst_reg->u32_max_value = var32_off.value | var32_off.mask;
14960 
14961 	/* Safe to set s32 bounds by casting u32 result into s32 when u32
14962 	 * doesn't cross sign boundary. Otherwise set s32 bounds to unbounded.
14963 	 */
14964 	if ((s32)dst_reg->u32_min_value <= (s32)dst_reg->u32_max_value) {
14965 		dst_reg->s32_min_value = dst_reg->u32_min_value;
14966 		dst_reg->s32_max_value = dst_reg->u32_max_value;
14967 	} else {
14968 		dst_reg->s32_min_value = S32_MIN;
14969 		dst_reg->s32_max_value = S32_MAX;
14970 	}
14971 }
14972 
scalar_min_max_or(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)14973 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
14974 			      struct bpf_reg_state *src_reg)
14975 {
14976 	bool src_known = tnum_is_const(src_reg->var_off);
14977 	bool dst_known = tnum_is_const(dst_reg->var_off);
14978 	u64 umin_val = src_reg->umin_value;
14979 
14980 	if (src_known && dst_known) {
14981 		__mark_reg_known(dst_reg, dst_reg->var_off.value);
14982 		return;
14983 	}
14984 
14985 	/* We get our maximum from the var_off, and our minimum is the
14986 	 * maximum of the operands' minima
14987 	 */
14988 	dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
14989 	dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
14990 
14991 	/* Safe to set s64 bounds by casting u64 result into s64 when u64
14992 	 * doesn't cross sign boundary. Otherwise set s64 bounds to unbounded.
14993 	 */
14994 	if ((s64)dst_reg->umin_value <= (s64)dst_reg->umax_value) {
14995 		dst_reg->smin_value = dst_reg->umin_value;
14996 		dst_reg->smax_value = dst_reg->umax_value;
14997 	} else {
14998 		dst_reg->smin_value = S64_MIN;
14999 		dst_reg->smax_value = S64_MAX;
15000 	}
15001 	/* We may learn something more from the var_off */
15002 	__update_reg_bounds(dst_reg);
15003 }
15004 
scalar32_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15005 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
15006 				 struct bpf_reg_state *src_reg)
15007 {
15008 	bool src_known = tnum_subreg_is_const(src_reg->var_off);
15009 	bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
15010 	struct tnum var32_off = tnum_subreg(dst_reg->var_off);
15011 
15012 	if (src_known && dst_known) {
15013 		__mark_reg32_known(dst_reg, var32_off.value);
15014 		return;
15015 	}
15016 
15017 	/* We get both minimum and maximum from the var32_off. */
15018 	dst_reg->u32_min_value = var32_off.value;
15019 	dst_reg->u32_max_value = var32_off.value | var32_off.mask;
15020 
15021 	/* Safe to set s32 bounds by casting u32 result into s32 when u32
15022 	 * doesn't cross sign boundary. Otherwise set s32 bounds to unbounded.
15023 	 */
15024 	if ((s32)dst_reg->u32_min_value <= (s32)dst_reg->u32_max_value) {
15025 		dst_reg->s32_min_value = dst_reg->u32_min_value;
15026 		dst_reg->s32_max_value = dst_reg->u32_max_value;
15027 	} else {
15028 		dst_reg->s32_min_value = S32_MIN;
15029 		dst_reg->s32_max_value = S32_MAX;
15030 	}
15031 }
15032 
scalar_min_max_xor(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15033 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
15034 			       struct bpf_reg_state *src_reg)
15035 {
15036 	bool src_known = tnum_is_const(src_reg->var_off);
15037 	bool dst_known = tnum_is_const(dst_reg->var_off);
15038 
15039 	if (src_known && dst_known) {
15040 		/* dst_reg->var_off.value has been updated earlier */
15041 		__mark_reg_known(dst_reg, dst_reg->var_off.value);
15042 		return;
15043 	}
15044 
15045 	/* We get both minimum and maximum from the var_off. */
15046 	dst_reg->umin_value = dst_reg->var_off.value;
15047 	dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
15048 
15049 	/* Safe to set s64 bounds by casting u64 result into s64 when u64
15050 	 * doesn't cross sign boundary. Otherwise set s64 bounds to unbounded.
15051 	 */
15052 	if ((s64)dst_reg->umin_value <= (s64)dst_reg->umax_value) {
15053 		dst_reg->smin_value = dst_reg->umin_value;
15054 		dst_reg->smax_value = dst_reg->umax_value;
15055 	} else {
15056 		dst_reg->smin_value = S64_MIN;
15057 		dst_reg->smax_value = S64_MAX;
15058 	}
15059 
15060 	__update_reg_bounds(dst_reg);
15061 }
15062 
__scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)15063 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
15064 				   u64 umin_val, u64 umax_val)
15065 {
15066 	/* We lose all sign bit information (except what we can pick
15067 	 * up from var_off)
15068 	 */
15069 	dst_reg->s32_min_value = S32_MIN;
15070 	dst_reg->s32_max_value = S32_MAX;
15071 	/* If we might shift our top bit out, then we know nothing */
15072 	if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
15073 		dst_reg->u32_min_value = 0;
15074 		dst_reg->u32_max_value = U32_MAX;
15075 	} else {
15076 		dst_reg->u32_min_value <<= umin_val;
15077 		dst_reg->u32_max_value <<= umax_val;
15078 	}
15079 }
15080 
scalar32_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15081 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
15082 				 struct bpf_reg_state *src_reg)
15083 {
15084 	u32 umax_val = src_reg->u32_max_value;
15085 	u32 umin_val = src_reg->u32_min_value;
15086 	/* u32 alu operation will zext upper bits */
15087 	struct tnum subreg = tnum_subreg(dst_reg->var_off);
15088 
15089 	__scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
15090 	dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
15091 	/* Not required but being careful mark reg64 bounds as unknown so
15092 	 * that we are forced to pick them up from tnum and zext later and
15093 	 * if some path skips this step we are still safe.
15094 	 */
15095 	__mark_reg64_unbounded(dst_reg);
15096 	__update_reg32_bounds(dst_reg);
15097 }
15098 
__scalar64_min_max_lsh(struct bpf_reg_state * dst_reg,u64 umin_val,u64 umax_val)15099 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
15100 				   u64 umin_val, u64 umax_val)
15101 {
15102 	/* Special case <<32 because it is a common compiler pattern to sign
15103 	 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
15104 	 * positive we know this shift will also be positive so we can track
15105 	 * bounds correctly. Otherwise we lose all sign bit information except
15106 	 * what we can pick up from var_off. Perhaps we can generalize this
15107 	 * later to shifts of any length.
15108 	 */
15109 	if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
15110 		dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
15111 	else
15112 		dst_reg->smax_value = S64_MAX;
15113 
15114 	if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
15115 		dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
15116 	else
15117 		dst_reg->smin_value = S64_MIN;
15118 
15119 	/* If we might shift our top bit out, then we know nothing */
15120 	if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
15121 		dst_reg->umin_value = 0;
15122 		dst_reg->umax_value = U64_MAX;
15123 	} else {
15124 		dst_reg->umin_value <<= umin_val;
15125 		dst_reg->umax_value <<= umax_val;
15126 	}
15127 }
15128 
scalar_min_max_lsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15129 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
15130 			       struct bpf_reg_state *src_reg)
15131 {
15132 	u64 umax_val = src_reg->umax_value;
15133 	u64 umin_val = src_reg->umin_value;
15134 
15135 	/* scalar64 calc uses 32bit unshifted bounds so must be called first */
15136 	__scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
15137 	__scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
15138 
15139 	dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
15140 	/* We may learn something more from the var_off */
15141 	__update_reg_bounds(dst_reg);
15142 }
15143 
scalar32_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15144 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
15145 				 struct bpf_reg_state *src_reg)
15146 {
15147 	struct tnum subreg = tnum_subreg(dst_reg->var_off);
15148 	u32 umax_val = src_reg->u32_max_value;
15149 	u32 umin_val = src_reg->u32_min_value;
15150 
15151 	/* BPF_RSH is an unsigned shift.  If the value in dst_reg might
15152 	 * be negative, then either:
15153 	 * 1) src_reg might be zero, so the sign bit of the result is
15154 	 *    unknown, so we lose our signed bounds
15155 	 * 2) it's known negative, thus the unsigned bounds capture the
15156 	 *    signed bounds
15157 	 * 3) the signed bounds cross zero, so they tell us nothing
15158 	 *    about the result
15159 	 * If the value in dst_reg is known nonnegative, then again the
15160 	 * unsigned bounds capture the signed bounds.
15161 	 * Thus, in all cases it suffices to blow away our signed bounds
15162 	 * and rely on inferring new ones from the unsigned bounds and
15163 	 * var_off of the result.
15164 	 */
15165 	dst_reg->s32_min_value = S32_MIN;
15166 	dst_reg->s32_max_value = S32_MAX;
15167 
15168 	dst_reg->var_off = tnum_rshift(subreg, umin_val);
15169 	dst_reg->u32_min_value >>= umax_val;
15170 	dst_reg->u32_max_value >>= umin_val;
15171 
15172 	__mark_reg64_unbounded(dst_reg);
15173 	__update_reg32_bounds(dst_reg);
15174 }
15175 
scalar_min_max_rsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15176 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
15177 			       struct bpf_reg_state *src_reg)
15178 {
15179 	u64 umax_val = src_reg->umax_value;
15180 	u64 umin_val = src_reg->umin_value;
15181 
15182 	/* BPF_RSH is an unsigned shift.  If the value in dst_reg might
15183 	 * be negative, then either:
15184 	 * 1) src_reg might be zero, so the sign bit of the result is
15185 	 *    unknown, so we lose our signed bounds
15186 	 * 2) it's known negative, thus the unsigned bounds capture the
15187 	 *    signed bounds
15188 	 * 3) the signed bounds cross zero, so they tell us nothing
15189 	 *    about the result
15190 	 * If the value in dst_reg is known nonnegative, then again the
15191 	 * unsigned bounds capture the signed bounds.
15192 	 * Thus, in all cases it suffices to blow away our signed bounds
15193 	 * and rely on inferring new ones from the unsigned bounds and
15194 	 * var_off of the result.
15195 	 */
15196 	dst_reg->smin_value = S64_MIN;
15197 	dst_reg->smax_value = S64_MAX;
15198 	dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
15199 	dst_reg->umin_value >>= umax_val;
15200 	dst_reg->umax_value >>= umin_val;
15201 
15202 	/* Its not easy to operate on alu32 bounds here because it depends
15203 	 * on bits being shifted in. Take easy way out and mark unbounded
15204 	 * so we can recalculate later from tnum.
15205 	 */
15206 	__mark_reg32_unbounded(dst_reg);
15207 	__update_reg_bounds(dst_reg);
15208 }
15209 
scalar32_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15210 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
15211 				  struct bpf_reg_state *src_reg)
15212 {
15213 	u64 umin_val = src_reg->u32_min_value;
15214 
15215 	/* Upon reaching here, src_known is true and
15216 	 * umax_val is equal to umin_val.
15217 	 */
15218 	dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
15219 	dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
15220 
15221 	dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
15222 
15223 	/* blow away the dst_reg umin_value/umax_value and rely on
15224 	 * dst_reg var_off to refine the result.
15225 	 */
15226 	dst_reg->u32_min_value = 0;
15227 	dst_reg->u32_max_value = U32_MAX;
15228 
15229 	__mark_reg64_unbounded(dst_reg);
15230 	__update_reg32_bounds(dst_reg);
15231 }
15232 
scalar_min_max_arsh(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg)15233 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
15234 				struct bpf_reg_state *src_reg)
15235 {
15236 	u64 umin_val = src_reg->umin_value;
15237 
15238 	/* Upon reaching here, src_known is true and umax_val is equal
15239 	 * to umin_val.
15240 	 */
15241 	dst_reg->smin_value >>= umin_val;
15242 	dst_reg->smax_value >>= umin_val;
15243 
15244 	dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
15245 
15246 	/* blow away the dst_reg umin_value/umax_value and rely on
15247 	 * dst_reg var_off to refine the result.
15248 	 */
15249 	dst_reg->umin_value = 0;
15250 	dst_reg->umax_value = U64_MAX;
15251 
15252 	/* Its not easy to operate on alu32 bounds here because it depends
15253 	 * on bits being shifted in from upper 32-bits. Take easy way out
15254 	 * and mark unbounded so we can recalculate later from tnum.
15255 	 */
15256 	__mark_reg32_unbounded(dst_reg);
15257 	__update_reg_bounds(dst_reg);
15258 }
15259 
is_safe_to_compute_dst_reg_range(struct bpf_insn * insn,const struct bpf_reg_state * src_reg)15260 static bool is_safe_to_compute_dst_reg_range(struct bpf_insn *insn,
15261 					     const struct bpf_reg_state *src_reg)
15262 {
15263 	bool src_is_const = false;
15264 	u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
15265 
15266 	if (insn_bitness == 32) {
15267 		if (tnum_subreg_is_const(src_reg->var_off)
15268 		    && src_reg->s32_min_value == src_reg->s32_max_value
15269 		    && src_reg->u32_min_value == src_reg->u32_max_value)
15270 			src_is_const = true;
15271 	} else {
15272 		if (tnum_is_const(src_reg->var_off)
15273 		    && src_reg->smin_value == src_reg->smax_value
15274 		    && src_reg->umin_value == src_reg->umax_value)
15275 			src_is_const = true;
15276 	}
15277 
15278 	switch (BPF_OP(insn->code)) {
15279 	case BPF_ADD:
15280 	case BPF_SUB:
15281 	case BPF_NEG:
15282 	case BPF_AND:
15283 	case BPF_XOR:
15284 	case BPF_OR:
15285 	case BPF_MUL:
15286 		return true;
15287 
15288 	/* Shift operators range is only computable if shift dimension operand
15289 	 * is a constant. Shifts greater than 31 or 63 are undefined. This
15290 	 * includes shifts by a negative number.
15291 	 */
15292 	case BPF_LSH:
15293 	case BPF_RSH:
15294 	case BPF_ARSH:
15295 		return (src_is_const && src_reg->umax_value < insn_bitness);
15296 	default:
15297 		return false;
15298 	}
15299 }
15300 
15301 /* WARNING: This function does calculations on 64-bit values, but the actual
15302  * execution may occur on 32-bit values. Therefore, things like bitshifts
15303  * need extra checks in the 32-bit case.
15304  */
adjust_scalar_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_reg_state * dst_reg,struct bpf_reg_state src_reg)15305 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
15306 				      struct bpf_insn *insn,
15307 				      struct bpf_reg_state *dst_reg,
15308 				      struct bpf_reg_state src_reg)
15309 {
15310 	u8 opcode = BPF_OP(insn->code);
15311 	bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
15312 	int ret;
15313 
15314 	if (!is_safe_to_compute_dst_reg_range(insn, &src_reg)) {
15315 		__mark_reg_unknown(env, dst_reg);
15316 		return 0;
15317 	}
15318 
15319 	if (sanitize_needed(opcode)) {
15320 		ret = sanitize_val_alu(env, insn);
15321 		if (ret < 0)
15322 			return sanitize_err(env, insn, ret, NULL, NULL);
15323 	}
15324 
15325 	/* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
15326 	 * There are two classes of instructions: The first class we track both
15327 	 * alu32 and alu64 sign/unsigned bounds independently this provides the
15328 	 * greatest amount of precision when alu operations are mixed with jmp32
15329 	 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
15330 	 * and BPF_OR. This is possible because these ops have fairly easy to
15331 	 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
15332 	 * See alu32 verifier tests for examples. The second class of
15333 	 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
15334 	 * with regards to tracking sign/unsigned bounds because the bits may
15335 	 * cross subreg boundaries in the alu64 case. When this happens we mark
15336 	 * the reg unbounded in the subreg bound space and use the resulting
15337 	 * tnum to calculate an approximation of the sign/unsigned bounds.
15338 	 */
15339 	switch (opcode) {
15340 	case BPF_ADD:
15341 		scalar32_min_max_add(dst_reg, &src_reg);
15342 		scalar_min_max_add(dst_reg, &src_reg);
15343 		dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
15344 		break;
15345 	case BPF_SUB:
15346 		scalar32_min_max_sub(dst_reg, &src_reg);
15347 		scalar_min_max_sub(dst_reg, &src_reg);
15348 		dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
15349 		break;
15350 	case BPF_NEG:
15351 		env->fake_reg[0] = *dst_reg;
15352 		__mark_reg_known(dst_reg, 0);
15353 		scalar32_min_max_sub(dst_reg, &env->fake_reg[0]);
15354 		scalar_min_max_sub(dst_reg, &env->fake_reg[0]);
15355 		dst_reg->var_off = tnum_neg(env->fake_reg[0].var_off);
15356 		break;
15357 	case BPF_MUL:
15358 		dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
15359 		scalar32_min_max_mul(dst_reg, &src_reg);
15360 		scalar_min_max_mul(dst_reg, &src_reg);
15361 		break;
15362 	case BPF_AND:
15363 		dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
15364 		scalar32_min_max_and(dst_reg, &src_reg);
15365 		scalar_min_max_and(dst_reg, &src_reg);
15366 		break;
15367 	case BPF_OR:
15368 		dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
15369 		scalar32_min_max_or(dst_reg, &src_reg);
15370 		scalar_min_max_or(dst_reg, &src_reg);
15371 		break;
15372 	case BPF_XOR:
15373 		dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
15374 		scalar32_min_max_xor(dst_reg, &src_reg);
15375 		scalar_min_max_xor(dst_reg, &src_reg);
15376 		break;
15377 	case BPF_LSH:
15378 		if (alu32)
15379 			scalar32_min_max_lsh(dst_reg, &src_reg);
15380 		else
15381 			scalar_min_max_lsh(dst_reg, &src_reg);
15382 		break;
15383 	case BPF_RSH:
15384 		if (alu32)
15385 			scalar32_min_max_rsh(dst_reg, &src_reg);
15386 		else
15387 			scalar_min_max_rsh(dst_reg, &src_reg);
15388 		break;
15389 	case BPF_ARSH:
15390 		if (alu32)
15391 			scalar32_min_max_arsh(dst_reg, &src_reg);
15392 		else
15393 			scalar_min_max_arsh(dst_reg, &src_reg);
15394 		break;
15395 	default:
15396 		break;
15397 	}
15398 
15399 	/* ALU32 ops are zero extended into 64bit register */
15400 	if (alu32)
15401 		zext_32_to_64(dst_reg);
15402 	reg_bounds_sync(dst_reg);
15403 	return 0;
15404 }
15405 
15406 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
15407  * and var_off.
15408  */
adjust_reg_min_max_vals(struct bpf_verifier_env * env,struct bpf_insn * insn)15409 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
15410 				   struct bpf_insn *insn)
15411 {
15412 	struct bpf_verifier_state *vstate = env->cur_state;
15413 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
15414 	struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
15415 	struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
15416 	bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
15417 	u8 opcode = BPF_OP(insn->code);
15418 	int err;
15419 
15420 	dst_reg = &regs[insn->dst_reg];
15421 	src_reg = NULL;
15422 
15423 	if (dst_reg->type == PTR_TO_ARENA) {
15424 		struct bpf_insn_aux_data *aux = cur_aux(env);
15425 
15426 		if (BPF_CLASS(insn->code) == BPF_ALU64)
15427 			/*
15428 			 * 32-bit operations zero upper bits automatically.
15429 			 * 64-bit operations need to be converted to 32.
15430 			 */
15431 			aux->needs_zext = true;
15432 
15433 		/* Any arithmetic operations are allowed on arena pointers */
15434 		return 0;
15435 	}
15436 
15437 	if (dst_reg->type != SCALAR_VALUE)
15438 		ptr_reg = dst_reg;
15439 
15440 	if (BPF_SRC(insn->code) == BPF_X) {
15441 		src_reg = &regs[insn->src_reg];
15442 		if (src_reg->type != SCALAR_VALUE) {
15443 			if (dst_reg->type != SCALAR_VALUE) {
15444 				/* Combining two pointers by any ALU op yields
15445 				 * an arbitrary scalar. Disallow all math except
15446 				 * pointer subtraction
15447 				 */
15448 				if (opcode == BPF_SUB && env->allow_ptr_leaks) {
15449 					mark_reg_unknown(env, regs, insn->dst_reg);
15450 					return 0;
15451 				}
15452 				verbose(env, "R%d pointer %s pointer prohibited\n",
15453 					insn->dst_reg,
15454 					bpf_alu_string[opcode >> 4]);
15455 				return -EACCES;
15456 			} else {
15457 				/* scalar += pointer
15458 				 * This is legal, but we have to reverse our
15459 				 * src/dest handling in computing the range
15460 				 */
15461 				err = mark_chain_precision(env, insn->dst_reg);
15462 				if (err)
15463 					return err;
15464 				return adjust_ptr_min_max_vals(env, insn,
15465 							       src_reg, dst_reg);
15466 			}
15467 		} else if (ptr_reg) {
15468 			/* pointer += scalar */
15469 			err = mark_chain_precision(env, insn->src_reg);
15470 			if (err)
15471 				return err;
15472 			return adjust_ptr_min_max_vals(env, insn,
15473 						       dst_reg, src_reg);
15474 		} else if (dst_reg->precise) {
15475 			/* if dst_reg is precise, src_reg should be precise as well */
15476 			err = mark_chain_precision(env, insn->src_reg);
15477 			if (err)
15478 				return err;
15479 		}
15480 	} else {
15481 		/* Pretend the src is a reg with a known value, since we only
15482 		 * need to be able to read from this state.
15483 		 */
15484 		off_reg.type = SCALAR_VALUE;
15485 		__mark_reg_known(&off_reg, insn->imm);
15486 		src_reg = &off_reg;
15487 		if (ptr_reg) /* pointer += K */
15488 			return adjust_ptr_min_max_vals(env, insn,
15489 						       ptr_reg, src_reg);
15490 	}
15491 
15492 	/* Got here implies adding two SCALAR_VALUEs */
15493 	if (WARN_ON_ONCE(ptr_reg)) {
15494 		print_verifier_state(env, vstate, vstate->curframe, true);
15495 		verbose(env, "verifier internal error: unexpected ptr_reg\n");
15496 		return -EFAULT;
15497 	}
15498 	if (WARN_ON(!src_reg)) {
15499 		print_verifier_state(env, vstate, vstate->curframe, true);
15500 		verbose(env, "verifier internal error: no src_reg\n");
15501 		return -EFAULT;
15502 	}
15503 	err = adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
15504 	if (err)
15505 		return err;
15506 	/*
15507 	 * Compilers can generate the code
15508 	 * r1 = r2
15509 	 * r1 += 0x1
15510 	 * if r2 < 1000 goto ...
15511 	 * use r1 in memory access
15512 	 * So for 64-bit alu remember constant delta between r2 and r1 and
15513 	 * update r1 after 'if' condition.
15514 	 */
15515 	if (env->bpf_capable &&
15516 	    BPF_OP(insn->code) == BPF_ADD && !alu32 &&
15517 	    dst_reg->id && is_reg_const(src_reg, false)) {
15518 		u64 val = reg_const_value(src_reg, false);
15519 
15520 		if ((dst_reg->id & BPF_ADD_CONST) ||
15521 		    /* prevent overflow in sync_linked_regs() later */
15522 		    val > (u32)S32_MAX) {
15523 			/*
15524 			 * If the register already went through rX += val
15525 			 * we cannot accumulate another val into rx->off.
15526 			 */
15527 			dst_reg->off = 0;
15528 			dst_reg->id = 0;
15529 		} else {
15530 			dst_reg->id |= BPF_ADD_CONST;
15531 			dst_reg->off = val;
15532 		}
15533 	} else {
15534 		/*
15535 		 * Make sure ID is cleared otherwise dst_reg min/max could be
15536 		 * incorrectly propagated into other registers by sync_linked_regs()
15537 		 */
15538 		dst_reg->id = 0;
15539 	}
15540 	return 0;
15541 }
15542 
15543 /* check validity of 32-bit and 64-bit arithmetic operations */
check_alu_op(struct bpf_verifier_env * env,struct bpf_insn * insn)15544 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
15545 {
15546 	struct bpf_reg_state *regs = cur_regs(env);
15547 	u8 opcode = BPF_OP(insn->code);
15548 	int err;
15549 
15550 	if (opcode == BPF_END || opcode == BPF_NEG) {
15551 		if (opcode == BPF_NEG) {
15552 			if (BPF_SRC(insn->code) != BPF_K ||
15553 			    insn->src_reg != BPF_REG_0 ||
15554 			    insn->off != 0 || insn->imm != 0) {
15555 				verbose(env, "BPF_NEG uses reserved fields\n");
15556 				return -EINVAL;
15557 			}
15558 		} else {
15559 			if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
15560 			    (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
15561 			    (BPF_CLASS(insn->code) == BPF_ALU64 &&
15562 			     BPF_SRC(insn->code) != BPF_TO_LE)) {
15563 				verbose(env, "BPF_END uses reserved fields\n");
15564 				return -EINVAL;
15565 			}
15566 		}
15567 
15568 		/* check src operand */
15569 		err = check_reg_arg(env, insn->dst_reg, SRC_OP);
15570 		if (err)
15571 			return err;
15572 
15573 		if (is_pointer_value(env, insn->dst_reg)) {
15574 			verbose(env, "R%d pointer arithmetic prohibited\n",
15575 				insn->dst_reg);
15576 			return -EACCES;
15577 		}
15578 
15579 		/* check dest operand */
15580 		if (opcode == BPF_NEG) {
15581 			err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
15582 			err = err ?: adjust_scalar_min_max_vals(env, insn,
15583 							 &regs[insn->dst_reg],
15584 							 regs[insn->dst_reg]);
15585 		} else {
15586 			err = check_reg_arg(env, insn->dst_reg, DST_OP);
15587 		}
15588 		if (err)
15589 			return err;
15590 
15591 	} else if (opcode == BPF_MOV) {
15592 
15593 		if (BPF_SRC(insn->code) == BPF_X) {
15594 			if (BPF_CLASS(insn->code) == BPF_ALU) {
15595 				if ((insn->off != 0 && insn->off != 8 && insn->off != 16) ||
15596 				    insn->imm) {
15597 					verbose(env, "BPF_MOV uses reserved fields\n");
15598 					return -EINVAL;
15599 				}
15600 			} else if (insn->off == BPF_ADDR_SPACE_CAST) {
15601 				if (insn->imm != 1 && insn->imm != 1u << 16) {
15602 					verbose(env, "addr_space_cast insn can only convert between address space 1 and 0\n");
15603 					return -EINVAL;
15604 				}
15605 				if (!env->prog->aux->arena) {
15606 					verbose(env, "addr_space_cast insn can only be used in a program that has an associated arena\n");
15607 					return -EINVAL;
15608 				}
15609 			} else {
15610 				if ((insn->off != 0 && insn->off != 8 && insn->off != 16 &&
15611 				     insn->off != 32) || insn->imm) {
15612 					verbose(env, "BPF_MOV uses reserved fields\n");
15613 					return -EINVAL;
15614 				}
15615 			}
15616 
15617 			/* check src operand */
15618 			err = check_reg_arg(env, insn->src_reg, SRC_OP);
15619 			if (err)
15620 				return err;
15621 		} else {
15622 			if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
15623 				verbose(env, "BPF_MOV uses reserved fields\n");
15624 				return -EINVAL;
15625 			}
15626 		}
15627 
15628 		/* check dest operand, mark as required later */
15629 		err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
15630 		if (err)
15631 			return err;
15632 
15633 		if (BPF_SRC(insn->code) == BPF_X) {
15634 			struct bpf_reg_state *src_reg = regs + insn->src_reg;
15635 			struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
15636 
15637 			if (BPF_CLASS(insn->code) == BPF_ALU64) {
15638 				if (insn->imm) {
15639 					/* off == BPF_ADDR_SPACE_CAST */
15640 					mark_reg_unknown(env, regs, insn->dst_reg);
15641 					if (insn->imm == 1) { /* cast from as(1) to as(0) */
15642 						dst_reg->type = PTR_TO_ARENA;
15643 						/* PTR_TO_ARENA is 32-bit */
15644 						dst_reg->subreg_def = env->insn_idx + 1;
15645 					}
15646 				} else if (insn->off == 0) {
15647 					/* case: R1 = R2
15648 					 * copy register state to dest reg
15649 					 */
15650 					assign_scalar_id_before_mov(env, src_reg);
15651 					copy_register_state(dst_reg, src_reg);
15652 					dst_reg->live |= REG_LIVE_WRITTEN;
15653 					dst_reg->subreg_def = DEF_NOT_SUBREG;
15654 				} else {
15655 					/* case: R1 = (s8, s16 s32)R2 */
15656 					if (is_pointer_value(env, insn->src_reg)) {
15657 						verbose(env,
15658 							"R%d sign-extension part of pointer\n",
15659 							insn->src_reg);
15660 						return -EACCES;
15661 					} else if (src_reg->type == SCALAR_VALUE) {
15662 						bool no_sext;
15663 
15664 						no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
15665 						if (no_sext)
15666 							assign_scalar_id_before_mov(env, src_reg);
15667 						copy_register_state(dst_reg, src_reg);
15668 						if (!no_sext)
15669 							dst_reg->id = 0;
15670 						coerce_reg_to_size_sx(dst_reg, insn->off >> 3);
15671 						dst_reg->live |= REG_LIVE_WRITTEN;
15672 						dst_reg->subreg_def = DEF_NOT_SUBREG;
15673 					} else {
15674 						mark_reg_unknown(env, regs, insn->dst_reg);
15675 					}
15676 				}
15677 			} else {
15678 				/* R1 = (u32) R2 */
15679 				if (is_pointer_value(env, insn->src_reg)) {
15680 					verbose(env,
15681 						"R%d partial copy of pointer\n",
15682 						insn->src_reg);
15683 					return -EACCES;
15684 				} else if (src_reg->type == SCALAR_VALUE) {
15685 					if (insn->off == 0) {
15686 						bool is_src_reg_u32 = get_reg_width(src_reg) <= 32;
15687 
15688 						if (is_src_reg_u32)
15689 							assign_scalar_id_before_mov(env, src_reg);
15690 						copy_register_state(dst_reg, src_reg);
15691 						/* Make sure ID is cleared if src_reg is not in u32
15692 						 * range otherwise dst_reg min/max could be incorrectly
15693 						 * propagated into src_reg by sync_linked_regs()
15694 						 */
15695 						if (!is_src_reg_u32)
15696 							dst_reg->id = 0;
15697 						dst_reg->live |= REG_LIVE_WRITTEN;
15698 						dst_reg->subreg_def = env->insn_idx + 1;
15699 					} else {
15700 						/* case: W1 = (s8, s16)W2 */
15701 						bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
15702 
15703 						if (no_sext)
15704 							assign_scalar_id_before_mov(env, src_reg);
15705 						copy_register_state(dst_reg, src_reg);
15706 						if (!no_sext)
15707 							dst_reg->id = 0;
15708 						dst_reg->live |= REG_LIVE_WRITTEN;
15709 						dst_reg->subreg_def = env->insn_idx + 1;
15710 						coerce_subreg_to_size_sx(dst_reg, insn->off >> 3);
15711 					}
15712 				} else {
15713 					mark_reg_unknown(env, regs,
15714 							 insn->dst_reg);
15715 				}
15716 				zext_32_to_64(dst_reg);
15717 				reg_bounds_sync(dst_reg);
15718 			}
15719 		} else {
15720 			/* case: R = imm
15721 			 * remember the value we stored into this reg
15722 			 */
15723 			/* clear any state __mark_reg_known doesn't set */
15724 			mark_reg_unknown(env, regs, insn->dst_reg);
15725 			regs[insn->dst_reg].type = SCALAR_VALUE;
15726 			if (BPF_CLASS(insn->code) == BPF_ALU64) {
15727 				__mark_reg_known(regs + insn->dst_reg,
15728 						 insn->imm);
15729 			} else {
15730 				__mark_reg_known(regs + insn->dst_reg,
15731 						 (u32)insn->imm);
15732 			}
15733 		}
15734 
15735 	} else if (opcode > BPF_END) {
15736 		verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
15737 		return -EINVAL;
15738 
15739 	} else {	/* all other ALU ops: and, sub, xor, add, ... */
15740 
15741 		if (BPF_SRC(insn->code) == BPF_X) {
15742 			if (insn->imm != 0 || insn->off > 1 ||
15743 			    (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
15744 				verbose(env, "BPF_ALU uses reserved fields\n");
15745 				return -EINVAL;
15746 			}
15747 			/* check src1 operand */
15748 			err = check_reg_arg(env, insn->src_reg, SRC_OP);
15749 			if (err)
15750 				return err;
15751 		} else {
15752 			if (insn->src_reg != BPF_REG_0 || insn->off > 1 ||
15753 			    (insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
15754 				verbose(env, "BPF_ALU uses reserved fields\n");
15755 				return -EINVAL;
15756 			}
15757 		}
15758 
15759 		/* check src2 operand */
15760 		err = check_reg_arg(env, insn->dst_reg, SRC_OP);
15761 		if (err)
15762 			return err;
15763 
15764 		if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
15765 		    BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
15766 			verbose(env, "div by zero\n");
15767 			return -EINVAL;
15768 		}
15769 
15770 		if ((opcode == BPF_LSH || opcode == BPF_RSH ||
15771 		     opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
15772 			int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
15773 
15774 			if (insn->imm < 0 || insn->imm >= size) {
15775 				verbose(env, "invalid shift %d\n", insn->imm);
15776 				return -EINVAL;
15777 			}
15778 		}
15779 
15780 		/* check dest operand */
15781 		err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
15782 		err = err ?: adjust_reg_min_max_vals(env, insn);
15783 		if (err)
15784 			return err;
15785 	}
15786 
15787 	return reg_bounds_sanity_check(env, &regs[insn->dst_reg], "alu");
15788 }
15789 
find_good_pkt_pointers(struct bpf_verifier_state * vstate,struct bpf_reg_state * dst_reg,enum bpf_reg_type type,bool range_right_open)15790 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
15791 				   struct bpf_reg_state *dst_reg,
15792 				   enum bpf_reg_type type,
15793 				   bool range_right_open)
15794 {
15795 	struct bpf_func_state *state;
15796 	struct bpf_reg_state *reg;
15797 	int new_range;
15798 
15799 	if (dst_reg->off < 0 ||
15800 	    (dst_reg->off == 0 && range_right_open))
15801 		/* This doesn't give us any range */
15802 		return;
15803 
15804 	if (dst_reg->umax_value > MAX_PACKET_OFF ||
15805 	    dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
15806 		/* Risk of overflow.  For instance, ptr + (1<<63) may be less
15807 		 * than pkt_end, but that's because it's also less than pkt.
15808 		 */
15809 		return;
15810 
15811 	new_range = dst_reg->off;
15812 	if (range_right_open)
15813 		new_range++;
15814 
15815 	/* Examples for register markings:
15816 	 *
15817 	 * pkt_data in dst register:
15818 	 *
15819 	 *   r2 = r3;
15820 	 *   r2 += 8;
15821 	 *   if (r2 > pkt_end) goto <handle exception>
15822 	 *   <access okay>
15823 	 *
15824 	 *   r2 = r3;
15825 	 *   r2 += 8;
15826 	 *   if (r2 < pkt_end) goto <access okay>
15827 	 *   <handle exception>
15828 	 *
15829 	 *   Where:
15830 	 *     r2 == dst_reg, pkt_end == src_reg
15831 	 *     r2=pkt(id=n,off=8,r=0)
15832 	 *     r3=pkt(id=n,off=0,r=0)
15833 	 *
15834 	 * pkt_data in src register:
15835 	 *
15836 	 *   r2 = r3;
15837 	 *   r2 += 8;
15838 	 *   if (pkt_end >= r2) goto <access okay>
15839 	 *   <handle exception>
15840 	 *
15841 	 *   r2 = r3;
15842 	 *   r2 += 8;
15843 	 *   if (pkt_end <= r2) goto <handle exception>
15844 	 *   <access okay>
15845 	 *
15846 	 *   Where:
15847 	 *     pkt_end == dst_reg, r2 == src_reg
15848 	 *     r2=pkt(id=n,off=8,r=0)
15849 	 *     r3=pkt(id=n,off=0,r=0)
15850 	 *
15851 	 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
15852 	 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
15853 	 * and [r3, r3 + 8-1) respectively is safe to access depending on
15854 	 * the check.
15855 	 */
15856 
15857 	/* If our ids match, then we must have the same max_value.  And we
15858 	 * don't care about the other reg's fixed offset, since if it's too big
15859 	 * the range won't allow anything.
15860 	 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
15861 	 */
15862 	bpf_for_each_reg_in_vstate(vstate, state, reg, ({
15863 		if (reg->type == type && reg->id == dst_reg->id)
15864 			/* keep the maximum range already checked */
15865 			reg->range = max(reg->range, new_range);
15866 	}));
15867 }
15868 
15869 /*
15870  * <reg1> <op> <reg2>, currently assuming reg2 is a constant
15871  */
is_scalar_branch_taken(struct bpf_reg_state * reg1,struct bpf_reg_state * reg2,u8 opcode,bool is_jmp32)15872 static int is_scalar_branch_taken(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
15873 				  u8 opcode, bool is_jmp32)
15874 {
15875 	struct tnum t1 = is_jmp32 ? tnum_subreg(reg1->var_off) : reg1->var_off;
15876 	struct tnum t2 = is_jmp32 ? tnum_subreg(reg2->var_off) : reg2->var_off;
15877 	u64 umin1 = is_jmp32 ? (u64)reg1->u32_min_value : reg1->umin_value;
15878 	u64 umax1 = is_jmp32 ? (u64)reg1->u32_max_value : reg1->umax_value;
15879 	s64 smin1 = is_jmp32 ? (s64)reg1->s32_min_value : reg1->smin_value;
15880 	s64 smax1 = is_jmp32 ? (s64)reg1->s32_max_value : reg1->smax_value;
15881 	u64 umin2 = is_jmp32 ? (u64)reg2->u32_min_value : reg2->umin_value;
15882 	u64 umax2 = is_jmp32 ? (u64)reg2->u32_max_value : reg2->umax_value;
15883 	s64 smin2 = is_jmp32 ? (s64)reg2->s32_min_value : reg2->smin_value;
15884 	s64 smax2 = is_jmp32 ? (s64)reg2->s32_max_value : reg2->smax_value;
15885 
15886 	switch (opcode) {
15887 	case BPF_JEQ:
15888 		/* constants, umin/umax and smin/smax checks would be
15889 		 * redundant in this case because they all should match
15890 		 */
15891 		if (tnum_is_const(t1) && tnum_is_const(t2))
15892 			return t1.value == t2.value;
15893 		/* non-overlapping ranges */
15894 		if (umin1 > umax2 || umax1 < umin2)
15895 			return 0;
15896 		if (smin1 > smax2 || smax1 < smin2)
15897 			return 0;
15898 		if (!is_jmp32) {
15899 			/* if 64-bit ranges are inconclusive, see if we can
15900 			 * utilize 32-bit subrange knowledge to eliminate
15901 			 * branches that can't be taken a priori
15902 			 */
15903 			if (reg1->u32_min_value > reg2->u32_max_value ||
15904 			    reg1->u32_max_value < reg2->u32_min_value)
15905 				return 0;
15906 			if (reg1->s32_min_value > reg2->s32_max_value ||
15907 			    reg1->s32_max_value < reg2->s32_min_value)
15908 				return 0;
15909 		}
15910 		break;
15911 	case BPF_JNE:
15912 		/* constants, umin/umax and smin/smax checks would be
15913 		 * redundant in this case because they all should match
15914 		 */
15915 		if (tnum_is_const(t1) && tnum_is_const(t2))
15916 			return t1.value != t2.value;
15917 		/* non-overlapping ranges */
15918 		if (umin1 > umax2 || umax1 < umin2)
15919 			return 1;
15920 		if (smin1 > smax2 || smax1 < smin2)
15921 			return 1;
15922 		if (!is_jmp32) {
15923 			/* if 64-bit ranges are inconclusive, see if we can
15924 			 * utilize 32-bit subrange knowledge to eliminate
15925 			 * branches that can't be taken a priori
15926 			 */
15927 			if (reg1->u32_min_value > reg2->u32_max_value ||
15928 			    reg1->u32_max_value < reg2->u32_min_value)
15929 				return 1;
15930 			if (reg1->s32_min_value > reg2->s32_max_value ||
15931 			    reg1->s32_max_value < reg2->s32_min_value)
15932 				return 1;
15933 		}
15934 		break;
15935 	case BPF_JSET:
15936 		if (!is_reg_const(reg2, is_jmp32)) {
15937 			swap(reg1, reg2);
15938 			swap(t1, t2);
15939 		}
15940 		if (!is_reg_const(reg2, is_jmp32))
15941 			return -1;
15942 		if ((~t1.mask & t1.value) & t2.value)
15943 			return 1;
15944 		if (!((t1.mask | t1.value) & t2.value))
15945 			return 0;
15946 		break;
15947 	case BPF_JGT:
15948 		if (umin1 > umax2)
15949 			return 1;
15950 		else if (umax1 <= umin2)
15951 			return 0;
15952 		break;
15953 	case BPF_JSGT:
15954 		if (smin1 > smax2)
15955 			return 1;
15956 		else if (smax1 <= smin2)
15957 			return 0;
15958 		break;
15959 	case BPF_JLT:
15960 		if (umax1 < umin2)
15961 			return 1;
15962 		else if (umin1 >= umax2)
15963 			return 0;
15964 		break;
15965 	case BPF_JSLT:
15966 		if (smax1 < smin2)
15967 			return 1;
15968 		else if (smin1 >= smax2)
15969 			return 0;
15970 		break;
15971 	case BPF_JGE:
15972 		if (umin1 >= umax2)
15973 			return 1;
15974 		else if (umax1 < umin2)
15975 			return 0;
15976 		break;
15977 	case BPF_JSGE:
15978 		if (smin1 >= smax2)
15979 			return 1;
15980 		else if (smax1 < smin2)
15981 			return 0;
15982 		break;
15983 	case BPF_JLE:
15984 		if (umax1 <= umin2)
15985 			return 1;
15986 		else if (umin1 > umax2)
15987 			return 0;
15988 		break;
15989 	case BPF_JSLE:
15990 		if (smax1 <= smin2)
15991 			return 1;
15992 		else if (smin1 > smax2)
15993 			return 0;
15994 		break;
15995 	}
15996 
15997 	return -1;
15998 }
15999 
flip_opcode(u32 opcode)16000 static int flip_opcode(u32 opcode)
16001 {
16002 	/* How can we transform "a <op> b" into "b <op> a"? */
16003 	static const u8 opcode_flip[16] = {
16004 		/* these stay the same */
16005 		[BPF_JEQ  >> 4] = BPF_JEQ,
16006 		[BPF_JNE  >> 4] = BPF_JNE,
16007 		[BPF_JSET >> 4] = BPF_JSET,
16008 		/* these swap "lesser" and "greater" (L and G in the opcodes) */
16009 		[BPF_JGE  >> 4] = BPF_JLE,
16010 		[BPF_JGT  >> 4] = BPF_JLT,
16011 		[BPF_JLE  >> 4] = BPF_JGE,
16012 		[BPF_JLT  >> 4] = BPF_JGT,
16013 		[BPF_JSGE >> 4] = BPF_JSLE,
16014 		[BPF_JSGT >> 4] = BPF_JSLT,
16015 		[BPF_JSLE >> 4] = BPF_JSGE,
16016 		[BPF_JSLT >> 4] = BPF_JSGT
16017 	};
16018 	return opcode_flip[opcode >> 4];
16019 }
16020 
is_pkt_ptr_branch_taken(struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg,u8 opcode)16021 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
16022 				   struct bpf_reg_state *src_reg,
16023 				   u8 opcode)
16024 {
16025 	struct bpf_reg_state *pkt;
16026 
16027 	if (src_reg->type == PTR_TO_PACKET_END) {
16028 		pkt = dst_reg;
16029 	} else if (dst_reg->type == PTR_TO_PACKET_END) {
16030 		pkt = src_reg;
16031 		opcode = flip_opcode(opcode);
16032 	} else {
16033 		return -1;
16034 	}
16035 
16036 	if (pkt->range >= 0)
16037 		return -1;
16038 
16039 	switch (opcode) {
16040 	case BPF_JLE:
16041 		/* pkt <= pkt_end */
16042 		fallthrough;
16043 	case BPF_JGT:
16044 		/* pkt > pkt_end */
16045 		if (pkt->range == BEYOND_PKT_END)
16046 			/* pkt has at last one extra byte beyond pkt_end */
16047 			return opcode == BPF_JGT;
16048 		break;
16049 	case BPF_JLT:
16050 		/* pkt < pkt_end */
16051 		fallthrough;
16052 	case BPF_JGE:
16053 		/* pkt >= pkt_end */
16054 		if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
16055 			return opcode == BPF_JGE;
16056 		break;
16057 	}
16058 	return -1;
16059 }
16060 
16061 /* compute branch direction of the expression "if (<reg1> opcode <reg2>) goto target;"
16062  * and return:
16063  *  1 - branch will be taken and "goto target" will be executed
16064  *  0 - branch will not be taken and fall-through to next insn
16065  * -1 - unknown. Example: "if (reg1 < 5)" is unknown when register value
16066  *      range [0,10]
16067  */
is_branch_taken(struct bpf_reg_state * reg1,struct bpf_reg_state * reg2,u8 opcode,bool is_jmp32)16068 static int is_branch_taken(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
16069 			   u8 opcode, bool is_jmp32)
16070 {
16071 	if (reg_is_pkt_pointer_any(reg1) && reg_is_pkt_pointer_any(reg2) && !is_jmp32)
16072 		return is_pkt_ptr_branch_taken(reg1, reg2, opcode);
16073 
16074 	if (__is_pointer_value(false, reg1) || __is_pointer_value(false, reg2)) {
16075 		u64 val;
16076 
16077 		/* arrange that reg2 is a scalar, and reg1 is a pointer */
16078 		if (!is_reg_const(reg2, is_jmp32)) {
16079 			opcode = flip_opcode(opcode);
16080 			swap(reg1, reg2);
16081 		}
16082 		/* and ensure that reg2 is a constant */
16083 		if (!is_reg_const(reg2, is_jmp32))
16084 			return -1;
16085 
16086 		if (!reg_not_null(reg1))
16087 			return -1;
16088 
16089 		/* If pointer is valid tests against zero will fail so we can
16090 		 * use this to direct branch taken.
16091 		 */
16092 		val = reg_const_value(reg2, is_jmp32);
16093 		if (val != 0)
16094 			return -1;
16095 
16096 		switch (opcode) {
16097 		case BPF_JEQ:
16098 			return 0;
16099 		case BPF_JNE:
16100 			return 1;
16101 		default:
16102 			return -1;
16103 		}
16104 	}
16105 
16106 	/* now deal with two scalars, but not necessarily constants */
16107 	return is_scalar_branch_taken(reg1, reg2, opcode, is_jmp32);
16108 }
16109 
16110 /* Opcode that corresponds to a *false* branch condition.
16111  * E.g., if r1 < r2, then reverse (false) condition is r1 >= r2
16112  */
rev_opcode(u8 opcode)16113 static u8 rev_opcode(u8 opcode)
16114 {
16115 	switch (opcode) {
16116 	case BPF_JEQ:		return BPF_JNE;
16117 	case BPF_JNE:		return BPF_JEQ;
16118 	/* JSET doesn't have it's reverse opcode in BPF, so add
16119 	 * BPF_X flag to denote the reverse of that operation
16120 	 */
16121 	case BPF_JSET:		return BPF_JSET | BPF_X;
16122 	case BPF_JSET | BPF_X:	return BPF_JSET;
16123 	case BPF_JGE:		return BPF_JLT;
16124 	case BPF_JGT:		return BPF_JLE;
16125 	case BPF_JLE:		return BPF_JGT;
16126 	case BPF_JLT:		return BPF_JGE;
16127 	case BPF_JSGE:		return BPF_JSLT;
16128 	case BPF_JSGT:		return BPF_JSLE;
16129 	case BPF_JSLE:		return BPF_JSGT;
16130 	case BPF_JSLT:		return BPF_JSGE;
16131 	default:		return 0;
16132 	}
16133 }
16134 
16135 /* Refine range knowledge for <reg1> <op> <reg>2 conditional operation. */
regs_refine_cond_op(struct bpf_reg_state * reg1,struct bpf_reg_state * reg2,u8 opcode,bool is_jmp32)16136 static void regs_refine_cond_op(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
16137 				u8 opcode, bool is_jmp32)
16138 {
16139 	struct tnum t;
16140 	u64 val;
16141 
16142 	/* In case of GE/GT/SGE/JST, reuse LE/LT/SLE/SLT logic from below */
16143 	switch (opcode) {
16144 	case BPF_JGE:
16145 	case BPF_JGT:
16146 	case BPF_JSGE:
16147 	case BPF_JSGT:
16148 		opcode = flip_opcode(opcode);
16149 		swap(reg1, reg2);
16150 		break;
16151 	default:
16152 		break;
16153 	}
16154 
16155 	switch (opcode) {
16156 	case BPF_JEQ:
16157 		if (is_jmp32) {
16158 			reg1->u32_min_value = max(reg1->u32_min_value, reg2->u32_min_value);
16159 			reg1->u32_max_value = min(reg1->u32_max_value, reg2->u32_max_value);
16160 			reg1->s32_min_value = max(reg1->s32_min_value, reg2->s32_min_value);
16161 			reg1->s32_max_value = min(reg1->s32_max_value, reg2->s32_max_value);
16162 			reg2->u32_min_value = reg1->u32_min_value;
16163 			reg2->u32_max_value = reg1->u32_max_value;
16164 			reg2->s32_min_value = reg1->s32_min_value;
16165 			reg2->s32_max_value = reg1->s32_max_value;
16166 
16167 			t = tnum_intersect(tnum_subreg(reg1->var_off), tnum_subreg(reg2->var_off));
16168 			reg1->var_off = tnum_with_subreg(reg1->var_off, t);
16169 			reg2->var_off = tnum_with_subreg(reg2->var_off, t);
16170 		} else {
16171 			reg1->umin_value = max(reg1->umin_value, reg2->umin_value);
16172 			reg1->umax_value = min(reg1->umax_value, reg2->umax_value);
16173 			reg1->smin_value = max(reg1->smin_value, reg2->smin_value);
16174 			reg1->smax_value = min(reg1->smax_value, reg2->smax_value);
16175 			reg2->umin_value = reg1->umin_value;
16176 			reg2->umax_value = reg1->umax_value;
16177 			reg2->smin_value = reg1->smin_value;
16178 			reg2->smax_value = reg1->smax_value;
16179 
16180 			reg1->var_off = tnum_intersect(reg1->var_off, reg2->var_off);
16181 			reg2->var_off = reg1->var_off;
16182 		}
16183 		break;
16184 	case BPF_JNE:
16185 		if (!is_reg_const(reg2, is_jmp32))
16186 			swap(reg1, reg2);
16187 		if (!is_reg_const(reg2, is_jmp32))
16188 			break;
16189 
16190 		/* try to recompute the bound of reg1 if reg2 is a const and
16191 		 * is exactly the edge of reg1.
16192 		 */
16193 		val = reg_const_value(reg2, is_jmp32);
16194 		if (is_jmp32) {
16195 			/* u32_min_value is not equal to 0xffffffff at this point,
16196 			 * because otherwise u32_max_value is 0xffffffff as well,
16197 			 * in such a case both reg1 and reg2 would be constants,
16198 			 * jump would be predicted and reg_set_min_max() won't
16199 			 * be called.
16200 			 *
16201 			 * Same reasoning works for all {u,s}{min,max}{32,64} cases
16202 			 * below.
16203 			 */
16204 			if (reg1->u32_min_value == (u32)val)
16205 				reg1->u32_min_value++;
16206 			if (reg1->u32_max_value == (u32)val)
16207 				reg1->u32_max_value--;
16208 			if (reg1->s32_min_value == (s32)val)
16209 				reg1->s32_min_value++;
16210 			if (reg1->s32_max_value == (s32)val)
16211 				reg1->s32_max_value--;
16212 		} else {
16213 			if (reg1->umin_value == (u64)val)
16214 				reg1->umin_value++;
16215 			if (reg1->umax_value == (u64)val)
16216 				reg1->umax_value--;
16217 			if (reg1->smin_value == (s64)val)
16218 				reg1->smin_value++;
16219 			if (reg1->smax_value == (s64)val)
16220 				reg1->smax_value--;
16221 		}
16222 		break;
16223 	case BPF_JSET:
16224 		if (!is_reg_const(reg2, is_jmp32))
16225 			swap(reg1, reg2);
16226 		if (!is_reg_const(reg2, is_jmp32))
16227 			break;
16228 		val = reg_const_value(reg2, is_jmp32);
16229 		/* BPF_JSET (i.e., TRUE branch, *not* BPF_JSET | BPF_X)
16230 		 * requires single bit to learn something useful. E.g., if we
16231 		 * know that `r1 & 0x3` is true, then which bits (0, 1, or both)
16232 		 * are actually set? We can learn something definite only if
16233 		 * it's a single-bit value to begin with.
16234 		 *
16235 		 * BPF_JSET | BPF_X (i.e., negation of BPF_JSET) doesn't have
16236 		 * this restriction. I.e., !(r1 & 0x3) means neither bit 0 nor
16237 		 * bit 1 is set, which we can readily use in adjustments.
16238 		 */
16239 		if (!is_power_of_2(val))
16240 			break;
16241 		if (is_jmp32) {
16242 			t = tnum_or(tnum_subreg(reg1->var_off), tnum_const(val));
16243 			reg1->var_off = tnum_with_subreg(reg1->var_off, t);
16244 		} else {
16245 			reg1->var_off = tnum_or(reg1->var_off, tnum_const(val));
16246 		}
16247 		break;
16248 	case BPF_JSET | BPF_X: /* reverse of BPF_JSET, see rev_opcode() */
16249 		if (!is_reg_const(reg2, is_jmp32))
16250 			swap(reg1, reg2);
16251 		if (!is_reg_const(reg2, is_jmp32))
16252 			break;
16253 		val = reg_const_value(reg2, is_jmp32);
16254 		/* Forget the ranges before narrowing tnums, to avoid invariant
16255 		 * violations if we're on a dead branch.
16256 		 */
16257 		__mark_reg_unbounded(reg1);
16258 		if (is_jmp32) {
16259 			t = tnum_and(tnum_subreg(reg1->var_off), tnum_const(~val));
16260 			reg1->var_off = tnum_with_subreg(reg1->var_off, t);
16261 		} else {
16262 			reg1->var_off = tnum_and(reg1->var_off, tnum_const(~val));
16263 		}
16264 		break;
16265 	case BPF_JLE:
16266 		if (is_jmp32) {
16267 			reg1->u32_max_value = min(reg1->u32_max_value, reg2->u32_max_value);
16268 			reg2->u32_min_value = max(reg1->u32_min_value, reg2->u32_min_value);
16269 		} else {
16270 			reg1->umax_value = min(reg1->umax_value, reg2->umax_value);
16271 			reg2->umin_value = max(reg1->umin_value, reg2->umin_value);
16272 		}
16273 		break;
16274 	case BPF_JLT:
16275 		if (is_jmp32) {
16276 			reg1->u32_max_value = min(reg1->u32_max_value, reg2->u32_max_value - 1);
16277 			reg2->u32_min_value = max(reg1->u32_min_value + 1, reg2->u32_min_value);
16278 		} else {
16279 			reg1->umax_value = min(reg1->umax_value, reg2->umax_value - 1);
16280 			reg2->umin_value = max(reg1->umin_value + 1, reg2->umin_value);
16281 		}
16282 		break;
16283 	case BPF_JSLE:
16284 		if (is_jmp32) {
16285 			reg1->s32_max_value = min(reg1->s32_max_value, reg2->s32_max_value);
16286 			reg2->s32_min_value = max(reg1->s32_min_value, reg2->s32_min_value);
16287 		} else {
16288 			reg1->smax_value = min(reg1->smax_value, reg2->smax_value);
16289 			reg2->smin_value = max(reg1->smin_value, reg2->smin_value);
16290 		}
16291 		break;
16292 	case BPF_JSLT:
16293 		if (is_jmp32) {
16294 			reg1->s32_max_value = min(reg1->s32_max_value, reg2->s32_max_value - 1);
16295 			reg2->s32_min_value = max(reg1->s32_min_value + 1, reg2->s32_min_value);
16296 		} else {
16297 			reg1->smax_value = min(reg1->smax_value, reg2->smax_value - 1);
16298 			reg2->smin_value = max(reg1->smin_value + 1, reg2->smin_value);
16299 		}
16300 		break;
16301 	default:
16302 		return;
16303 	}
16304 }
16305 
16306 /* Adjusts the register min/max values in the case that the dst_reg and
16307  * src_reg are both SCALAR_VALUE registers (or we are simply doing a BPF_K
16308  * check, in which case we have a fake SCALAR_VALUE representing insn->imm).
16309  * Technically we can do similar adjustments for pointers to the same object,
16310  * but we don't support that right now.
16311  */
reg_set_min_max(struct bpf_verifier_env * env,struct bpf_reg_state * true_reg1,struct bpf_reg_state * true_reg2,struct bpf_reg_state * false_reg1,struct bpf_reg_state * false_reg2,u8 opcode,bool is_jmp32)16312 static int reg_set_min_max(struct bpf_verifier_env *env,
16313 			   struct bpf_reg_state *true_reg1,
16314 			   struct bpf_reg_state *true_reg2,
16315 			   struct bpf_reg_state *false_reg1,
16316 			   struct bpf_reg_state *false_reg2,
16317 			   u8 opcode, bool is_jmp32)
16318 {
16319 	int err;
16320 
16321 	/* If either register is a pointer, we can't learn anything about its
16322 	 * variable offset from the compare (unless they were a pointer into
16323 	 * the same object, but we don't bother with that).
16324 	 */
16325 	if (false_reg1->type != SCALAR_VALUE || false_reg2->type != SCALAR_VALUE)
16326 		return 0;
16327 
16328 	/* fallthrough (FALSE) branch */
16329 	regs_refine_cond_op(false_reg1, false_reg2, rev_opcode(opcode), is_jmp32);
16330 	reg_bounds_sync(false_reg1);
16331 	reg_bounds_sync(false_reg2);
16332 
16333 	/* jump (TRUE) branch */
16334 	regs_refine_cond_op(true_reg1, true_reg2, opcode, is_jmp32);
16335 	reg_bounds_sync(true_reg1);
16336 	reg_bounds_sync(true_reg2);
16337 
16338 	err = reg_bounds_sanity_check(env, true_reg1, "true_reg1");
16339 	err = err ?: reg_bounds_sanity_check(env, true_reg2, "true_reg2");
16340 	err = err ?: reg_bounds_sanity_check(env, false_reg1, "false_reg1");
16341 	err = err ?: reg_bounds_sanity_check(env, false_reg2, "false_reg2");
16342 	return err;
16343 }
16344 
mark_ptr_or_null_reg(struct bpf_func_state * state,struct bpf_reg_state * reg,u32 id,bool is_null)16345 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
16346 				 struct bpf_reg_state *reg, u32 id,
16347 				 bool is_null)
16348 {
16349 	if (type_may_be_null(reg->type) && reg->id == id &&
16350 	    (is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
16351 		/* Old offset (both fixed and variable parts) should have been
16352 		 * known-zero, because we don't allow pointer arithmetic on
16353 		 * pointers that might be NULL. If we see this happening, don't
16354 		 * convert the register.
16355 		 *
16356 		 * But in some cases, some helpers that return local kptrs
16357 		 * advance offset for the returned pointer. In those cases, it
16358 		 * is fine to expect to see reg->off.
16359 		 */
16360 		if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
16361 			return;
16362 		if (!(type_is_ptr_alloc_obj(reg->type) || type_is_non_owning_ref(reg->type)) &&
16363 		    WARN_ON_ONCE(reg->off))
16364 			return;
16365 
16366 		if (is_null) {
16367 			reg->type = SCALAR_VALUE;
16368 			/* We don't need id and ref_obj_id from this point
16369 			 * onwards anymore, thus we should better reset it,
16370 			 * so that state pruning has chances to take effect.
16371 			 */
16372 			reg->id = 0;
16373 			reg->ref_obj_id = 0;
16374 
16375 			return;
16376 		}
16377 
16378 		mark_ptr_not_null_reg(reg);
16379 
16380 		if (!reg_may_point_to_spin_lock(reg)) {
16381 			/* For not-NULL ptr, reg->ref_obj_id will be reset
16382 			 * in release_reference().
16383 			 *
16384 			 * reg->id is still used by spin_lock ptr. Other
16385 			 * than spin_lock ptr type, reg->id can be reset.
16386 			 */
16387 			reg->id = 0;
16388 		}
16389 	}
16390 }
16391 
16392 /* The logic is similar to find_good_pkt_pointers(), both could eventually
16393  * be folded together at some point.
16394  */
mark_ptr_or_null_regs(struct bpf_verifier_state * vstate,u32 regno,bool is_null)16395 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
16396 				  bool is_null)
16397 {
16398 	struct bpf_func_state *state = vstate->frame[vstate->curframe];
16399 	struct bpf_reg_state *regs = state->regs, *reg;
16400 	u32 ref_obj_id = regs[regno].ref_obj_id;
16401 	u32 id = regs[regno].id;
16402 
16403 	if (ref_obj_id && ref_obj_id == id && is_null)
16404 		/* regs[regno] is in the " == NULL" branch.
16405 		 * No one could have freed the reference state before
16406 		 * doing the NULL check.
16407 		 */
16408 		WARN_ON_ONCE(release_reference_nomark(vstate, id));
16409 
16410 	bpf_for_each_reg_in_vstate(vstate, state, reg, ({
16411 		mark_ptr_or_null_reg(state, reg, id, is_null);
16412 	}));
16413 }
16414 
try_match_pkt_pointers(const struct bpf_insn * insn,struct bpf_reg_state * dst_reg,struct bpf_reg_state * src_reg,struct bpf_verifier_state * this_branch,struct bpf_verifier_state * other_branch)16415 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
16416 				   struct bpf_reg_state *dst_reg,
16417 				   struct bpf_reg_state *src_reg,
16418 				   struct bpf_verifier_state *this_branch,
16419 				   struct bpf_verifier_state *other_branch)
16420 {
16421 	if (BPF_SRC(insn->code) != BPF_X)
16422 		return false;
16423 
16424 	/* Pointers are always 64-bit. */
16425 	if (BPF_CLASS(insn->code) == BPF_JMP32)
16426 		return false;
16427 
16428 	switch (BPF_OP(insn->code)) {
16429 	case BPF_JGT:
16430 		if ((dst_reg->type == PTR_TO_PACKET &&
16431 		     src_reg->type == PTR_TO_PACKET_END) ||
16432 		    (dst_reg->type == PTR_TO_PACKET_META &&
16433 		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
16434 			/* pkt_data' > pkt_end, pkt_meta' > pkt_data */
16435 			find_good_pkt_pointers(this_branch, dst_reg,
16436 					       dst_reg->type, false);
16437 			mark_pkt_end(other_branch, insn->dst_reg, true);
16438 		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16439 			    src_reg->type == PTR_TO_PACKET) ||
16440 			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
16441 			    src_reg->type == PTR_TO_PACKET_META)) {
16442 			/* pkt_end > pkt_data', pkt_data > pkt_meta' */
16443 			find_good_pkt_pointers(other_branch, src_reg,
16444 					       src_reg->type, true);
16445 			mark_pkt_end(this_branch, insn->src_reg, false);
16446 		} else {
16447 			return false;
16448 		}
16449 		break;
16450 	case BPF_JLT:
16451 		if ((dst_reg->type == PTR_TO_PACKET &&
16452 		     src_reg->type == PTR_TO_PACKET_END) ||
16453 		    (dst_reg->type == PTR_TO_PACKET_META &&
16454 		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
16455 			/* pkt_data' < pkt_end, pkt_meta' < pkt_data */
16456 			find_good_pkt_pointers(other_branch, dst_reg,
16457 					       dst_reg->type, true);
16458 			mark_pkt_end(this_branch, insn->dst_reg, false);
16459 		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16460 			    src_reg->type == PTR_TO_PACKET) ||
16461 			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
16462 			    src_reg->type == PTR_TO_PACKET_META)) {
16463 			/* pkt_end < pkt_data', pkt_data > pkt_meta' */
16464 			find_good_pkt_pointers(this_branch, src_reg,
16465 					       src_reg->type, false);
16466 			mark_pkt_end(other_branch, insn->src_reg, true);
16467 		} else {
16468 			return false;
16469 		}
16470 		break;
16471 	case BPF_JGE:
16472 		if ((dst_reg->type == PTR_TO_PACKET &&
16473 		     src_reg->type == PTR_TO_PACKET_END) ||
16474 		    (dst_reg->type == PTR_TO_PACKET_META &&
16475 		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
16476 			/* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
16477 			find_good_pkt_pointers(this_branch, dst_reg,
16478 					       dst_reg->type, true);
16479 			mark_pkt_end(other_branch, insn->dst_reg, false);
16480 		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16481 			    src_reg->type == PTR_TO_PACKET) ||
16482 			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
16483 			    src_reg->type == PTR_TO_PACKET_META)) {
16484 			/* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
16485 			find_good_pkt_pointers(other_branch, src_reg,
16486 					       src_reg->type, false);
16487 			mark_pkt_end(this_branch, insn->src_reg, true);
16488 		} else {
16489 			return false;
16490 		}
16491 		break;
16492 	case BPF_JLE:
16493 		if ((dst_reg->type == PTR_TO_PACKET &&
16494 		     src_reg->type == PTR_TO_PACKET_END) ||
16495 		    (dst_reg->type == PTR_TO_PACKET_META &&
16496 		     reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
16497 			/* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
16498 			find_good_pkt_pointers(other_branch, dst_reg,
16499 					       dst_reg->type, false);
16500 			mark_pkt_end(this_branch, insn->dst_reg, true);
16501 		} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16502 			    src_reg->type == PTR_TO_PACKET) ||
16503 			   (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
16504 			    src_reg->type == PTR_TO_PACKET_META)) {
16505 			/* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
16506 			find_good_pkt_pointers(this_branch, src_reg,
16507 					       src_reg->type, true);
16508 			mark_pkt_end(other_branch, insn->src_reg, false);
16509 		} else {
16510 			return false;
16511 		}
16512 		break;
16513 	default:
16514 		return false;
16515 	}
16516 
16517 	return true;
16518 }
16519 
__collect_linked_regs(struct linked_regs * reg_set,struct bpf_reg_state * reg,u32 id,u32 frameno,u32 spi_or_reg,bool is_reg)16520 static void __collect_linked_regs(struct linked_regs *reg_set, struct bpf_reg_state *reg,
16521 				  u32 id, u32 frameno, u32 spi_or_reg, bool is_reg)
16522 {
16523 	struct linked_reg *e;
16524 
16525 	if (reg->type != SCALAR_VALUE || (reg->id & ~BPF_ADD_CONST) != id)
16526 		return;
16527 
16528 	e = linked_regs_push(reg_set);
16529 	if (e) {
16530 		e->frameno = frameno;
16531 		e->is_reg = is_reg;
16532 		e->regno = spi_or_reg;
16533 	} else {
16534 		reg->id = 0;
16535 	}
16536 }
16537 
16538 /* For all R being scalar registers or spilled scalar registers
16539  * in verifier state, save R in linked_regs if R->id == id.
16540  * If there are too many Rs sharing same id, reset id for leftover Rs.
16541  */
collect_linked_regs(struct bpf_verifier_state * vstate,u32 id,struct linked_regs * linked_regs)16542 static void collect_linked_regs(struct bpf_verifier_state *vstate, u32 id,
16543 				struct linked_regs *linked_regs)
16544 {
16545 	struct bpf_func_state *func;
16546 	struct bpf_reg_state *reg;
16547 	int i, j;
16548 
16549 	id = id & ~BPF_ADD_CONST;
16550 	for (i = vstate->curframe; i >= 0; i--) {
16551 		func = vstate->frame[i];
16552 		for (j = 0; j < BPF_REG_FP; j++) {
16553 			reg = &func->regs[j];
16554 			__collect_linked_regs(linked_regs, reg, id, i, j, true);
16555 		}
16556 		for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
16557 			if (!is_spilled_reg(&func->stack[j]))
16558 				continue;
16559 			reg = &func->stack[j].spilled_ptr;
16560 			__collect_linked_regs(linked_regs, reg, id, i, j, false);
16561 		}
16562 	}
16563 }
16564 
16565 /* For all R in linked_regs, copy known_reg range into R
16566  * if R->id == known_reg->id.
16567  */
sync_linked_regs(struct bpf_verifier_state * vstate,struct bpf_reg_state * known_reg,struct linked_regs * linked_regs)16568 static void sync_linked_regs(struct bpf_verifier_state *vstate, struct bpf_reg_state *known_reg,
16569 			     struct linked_regs *linked_regs)
16570 {
16571 	struct bpf_reg_state fake_reg;
16572 	struct bpf_reg_state *reg;
16573 	struct linked_reg *e;
16574 	int i;
16575 
16576 	for (i = 0; i < linked_regs->cnt; ++i) {
16577 		e = &linked_regs->entries[i];
16578 		reg = e->is_reg ? &vstate->frame[e->frameno]->regs[e->regno]
16579 				: &vstate->frame[e->frameno]->stack[e->spi].spilled_ptr;
16580 		if (reg->type != SCALAR_VALUE || reg == known_reg)
16581 			continue;
16582 		if ((reg->id & ~BPF_ADD_CONST) != (known_reg->id & ~BPF_ADD_CONST))
16583 			continue;
16584 		if ((!(reg->id & BPF_ADD_CONST) && !(known_reg->id & BPF_ADD_CONST)) ||
16585 		    reg->off == known_reg->off) {
16586 			s32 saved_subreg_def = reg->subreg_def;
16587 
16588 			copy_register_state(reg, known_reg);
16589 			reg->subreg_def = saved_subreg_def;
16590 		} else {
16591 			s32 saved_subreg_def = reg->subreg_def;
16592 			s32 saved_off = reg->off;
16593 
16594 			fake_reg.type = SCALAR_VALUE;
16595 			__mark_reg_known(&fake_reg, (s32)reg->off - (s32)known_reg->off);
16596 
16597 			/* reg = known_reg; reg += delta */
16598 			copy_register_state(reg, known_reg);
16599 			/*
16600 			 * Must preserve off, id and add_const flag,
16601 			 * otherwise another sync_linked_regs() will be incorrect.
16602 			 */
16603 			reg->off = saved_off;
16604 			reg->subreg_def = saved_subreg_def;
16605 
16606 			scalar32_min_max_add(reg, &fake_reg);
16607 			scalar_min_max_add(reg, &fake_reg);
16608 			reg->var_off = tnum_add(reg->var_off, fake_reg.var_off);
16609 		}
16610 	}
16611 }
16612 
check_cond_jmp_op(struct bpf_verifier_env * env,struct bpf_insn * insn,int * insn_idx)16613 static int check_cond_jmp_op(struct bpf_verifier_env *env,
16614 			     struct bpf_insn *insn, int *insn_idx)
16615 {
16616 	struct bpf_verifier_state *this_branch = env->cur_state;
16617 	struct bpf_verifier_state *other_branch;
16618 	struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
16619 	struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
16620 	struct bpf_reg_state *eq_branch_regs;
16621 	struct linked_regs linked_regs = {};
16622 	u8 opcode = BPF_OP(insn->code);
16623 	int insn_flags = 0;
16624 	bool is_jmp32;
16625 	int pred = -1;
16626 	int err;
16627 
16628 	/* Only conditional jumps are expected to reach here. */
16629 	if (opcode == BPF_JA || opcode > BPF_JCOND) {
16630 		verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
16631 		return -EINVAL;
16632 	}
16633 
16634 	if (opcode == BPF_JCOND) {
16635 		struct bpf_verifier_state *cur_st = env->cur_state, *queued_st, *prev_st;
16636 		int idx = *insn_idx;
16637 
16638 		if (insn->code != (BPF_JMP | BPF_JCOND) ||
16639 		    insn->src_reg != BPF_MAY_GOTO ||
16640 		    insn->dst_reg || insn->imm) {
16641 			verbose(env, "invalid may_goto imm %d\n", insn->imm);
16642 			return -EINVAL;
16643 		}
16644 		prev_st = find_prev_entry(env, cur_st->parent, idx);
16645 
16646 		/* branch out 'fallthrough' insn as a new state to explore */
16647 		queued_st = push_stack(env, idx + 1, idx, false);
16648 		if (!queued_st)
16649 			return -ENOMEM;
16650 
16651 		queued_st->may_goto_depth++;
16652 		if (prev_st)
16653 			widen_imprecise_scalars(env, prev_st, queued_st);
16654 		*insn_idx += insn->off;
16655 		return 0;
16656 	}
16657 
16658 	/* check src2 operand */
16659 	err = check_reg_arg(env, insn->dst_reg, SRC_OP);
16660 	if (err)
16661 		return err;
16662 
16663 	dst_reg = &regs[insn->dst_reg];
16664 	if (BPF_SRC(insn->code) == BPF_X) {
16665 		if (insn->imm != 0) {
16666 			verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
16667 			return -EINVAL;
16668 		}
16669 
16670 		/* check src1 operand */
16671 		err = check_reg_arg(env, insn->src_reg, SRC_OP);
16672 		if (err)
16673 			return err;
16674 
16675 		src_reg = &regs[insn->src_reg];
16676 		if (!(reg_is_pkt_pointer_any(dst_reg) && reg_is_pkt_pointer_any(src_reg)) &&
16677 		    is_pointer_value(env, insn->src_reg)) {
16678 			verbose(env, "R%d pointer comparison prohibited\n",
16679 				insn->src_reg);
16680 			return -EACCES;
16681 		}
16682 
16683 		if (src_reg->type == PTR_TO_STACK)
16684 			insn_flags |= INSN_F_SRC_REG_STACK;
16685 		if (dst_reg->type == PTR_TO_STACK)
16686 			insn_flags |= INSN_F_DST_REG_STACK;
16687 	} else {
16688 		if (insn->src_reg != BPF_REG_0) {
16689 			verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
16690 			return -EINVAL;
16691 		}
16692 		src_reg = &env->fake_reg[0];
16693 		memset(src_reg, 0, sizeof(*src_reg));
16694 		src_reg->type = SCALAR_VALUE;
16695 		__mark_reg_known(src_reg, insn->imm);
16696 
16697 		if (dst_reg->type == PTR_TO_STACK)
16698 			insn_flags |= INSN_F_DST_REG_STACK;
16699 	}
16700 
16701 	if (insn_flags) {
16702 		err = push_jmp_history(env, this_branch, insn_flags, 0);
16703 		if (err)
16704 			return err;
16705 	}
16706 
16707 	is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
16708 	pred = is_branch_taken(dst_reg, src_reg, opcode, is_jmp32);
16709 	if (pred >= 0) {
16710 		/* If we get here with a dst_reg pointer type it is because
16711 		 * above is_branch_taken() special cased the 0 comparison.
16712 		 */
16713 		if (!__is_pointer_value(false, dst_reg))
16714 			err = mark_chain_precision(env, insn->dst_reg);
16715 		if (BPF_SRC(insn->code) == BPF_X && !err &&
16716 		    !__is_pointer_value(false, src_reg))
16717 			err = mark_chain_precision(env, insn->src_reg);
16718 		if (err)
16719 			return err;
16720 	}
16721 
16722 	if (pred == 1) {
16723 		/* Only follow the goto, ignore fall-through. If needed, push
16724 		 * the fall-through branch for simulation under speculative
16725 		 * execution.
16726 		 */
16727 		if (!env->bypass_spec_v1 &&
16728 		    !sanitize_speculative_path(env, insn, *insn_idx + 1,
16729 					       *insn_idx))
16730 			return -EFAULT;
16731 		if (env->log.level & BPF_LOG_LEVEL)
16732 			print_insn_state(env, this_branch, this_branch->curframe);
16733 		*insn_idx += insn->off;
16734 		return 0;
16735 	} else if (pred == 0) {
16736 		/* Only follow the fall-through branch, since that's where the
16737 		 * program will go. If needed, push the goto branch for
16738 		 * simulation under speculative execution.
16739 		 */
16740 		if (!env->bypass_spec_v1 &&
16741 		    !sanitize_speculative_path(env, insn,
16742 					       *insn_idx + insn->off + 1,
16743 					       *insn_idx))
16744 			return -EFAULT;
16745 		if (env->log.level & BPF_LOG_LEVEL)
16746 			print_insn_state(env, this_branch, this_branch->curframe);
16747 		return 0;
16748 	}
16749 
16750 	/* Push scalar registers sharing same ID to jump history,
16751 	 * do this before creating 'other_branch', so that both
16752 	 * 'this_branch' and 'other_branch' share this history
16753 	 * if parent state is created.
16754 	 */
16755 	if (BPF_SRC(insn->code) == BPF_X && src_reg->type == SCALAR_VALUE && src_reg->id)
16756 		collect_linked_regs(this_branch, src_reg->id, &linked_regs);
16757 	if (dst_reg->type == SCALAR_VALUE && dst_reg->id)
16758 		collect_linked_regs(this_branch, dst_reg->id, &linked_regs);
16759 	if (linked_regs.cnt > 1) {
16760 		err = push_jmp_history(env, this_branch, 0, linked_regs_pack(&linked_regs));
16761 		if (err)
16762 			return err;
16763 	}
16764 
16765 	other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
16766 				  false);
16767 	if (!other_branch)
16768 		return -EFAULT;
16769 	other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
16770 
16771 	if (BPF_SRC(insn->code) == BPF_X) {
16772 		err = reg_set_min_max(env,
16773 				      &other_branch_regs[insn->dst_reg],
16774 				      &other_branch_regs[insn->src_reg],
16775 				      dst_reg, src_reg, opcode, is_jmp32);
16776 	} else /* BPF_SRC(insn->code) == BPF_K */ {
16777 		/* reg_set_min_max() can mangle the fake_reg. Make a copy
16778 		 * so that these are two different memory locations. The
16779 		 * src_reg is not used beyond here in context of K.
16780 		 */
16781 		memcpy(&env->fake_reg[1], &env->fake_reg[0],
16782 		       sizeof(env->fake_reg[0]));
16783 		err = reg_set_min_max(env,
16784 				      &other_branch_regs[insn->dst_reg],
16785 				      &env->fake_reg[0],
16786 				      dst_reg, &env->fake_reg[1],
16787 				      opcode, is_jmp32);
16788 	}
16789 	if (err)
16790 		return err;
16791 
16792 	if (BPF_SRC(insn->code) == BPF_X &&
16793 	    src_reg->type == SCALAR_VALUE && src_reg->id &&
16794 	    !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
16795 		sync_linked_regs(this_branch, src_reg, &linked_regs);
16796 		sync_linked_regs(other_branch, &other_branch_regs[insn->src_reg], &linked_regs);
16797 	}
16798 	if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
16799 	    !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
16800 		sync_linked_regs(this_branch, dst_reg, &linked_regs);
16801 		sync_linked_regs(other_branch, &other_branch_regs[insn->dst_reg], &linked_regs);
16802 	}
16803 
16804 	/* if one pointer register is compared to another pointer
16805 	 * register check if PTR_MAYBE_NULL could be lifted.
16806 	 * E.g. register A - maybe null
16807 	 *      register B - not null
16808 	 * for JNE A, B, ... - A is not null in the false branch;
16809 	 * for JEQ A, B, ... - A is not null in the true branch.
16810 	 *
16811 	 * Since PTR_TO_BTF_ID points to a kernel struct that does
16812 	 * not need to be null checked by the BPF program, i.e.,
16813 	 * could be null even without PTR_MAYBE_NULL marking, so
16814 	 * only propagate nullness when neither reg is that type.
16815 	 */
16816 	if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
16817 	    __is_pointer_value(false, src_reg) && __is_pointer_value(false, dst_reg) &&
16818 	    type_may_be_null(src_reg->type) != type_may_be_null(dst_reg->type) &&
16819 	    base_type(src_reg->type) != PTR_TO_BTF_ID &&
16820 	    base_type(dst_reg->type) != PTR_TO_BTF_ID) {
16821 		eq_branch_regs = NULL;
16822 		switch (opcode) {
16823 		case BPF_JEQ:
16824 			eq_branch_regs = other_branch_regs;
16825 			break;
16826 		case BPF_JNE:
16827 			eq_branch_regs = regs;
16828 			break;
16829 		default:
16830 			/* do nothing */
16831 			break;
16832 		}
16833 		if (eq_branch_regs) {
16834 			if (type_may_be_null(src_reg->type))
16835 				mark_ptr_not_null_reg(&eq_branch_regs[insn->src_reg]);
16836 			else
16837 				mark_ptr_not_null_reg(&eq_branch_regs[insn->dst_reg]);
16838 		}
16839 	}
16840 
16841 	/* detect if R == 0 where R is returned from bpf_map_lookup_elem().
16842 	 * NOTE: these optimizations below are related with pointer comparison
16843 	 *       which will never be JMP32.
16844 	 */
16845 	if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
16846 	    insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
16847 	    type_may_be_null(dst_reg->type)) {
16848 		/* Mark all identical registers in each branch as either
16849 		 * safe or unknown depending R == 0 or R != 0 conditional.
16850 		 */
16851 		mark_ptr_or_null_regs(this_branch, insn->dst_reg,
16852 				      opcode == BPF_JNE);
16853 		mark_ptr_or_null_regs(other_branch, insn->dst_reg,
16854 				      opcode == BPF_JEQ);
16855 	} else if (!try_match_pkt_pointers(insn, dst_reg, &regs[insn->src_reg],
16856 					   this_branch, other_branch) &&
16857 		   is_pointer_value(env, insn->dst_reg)) {
16858 		verbose(env, "R%d pointer comparison prohibited\n",
16859 			insn->dst_reg);
16860 		return -EACCES;
16861 	}
16862 	if (env->log.level & BPF_LOG_LEVEL)
16863 		print_insn_state(env, this_branch, this_branch->curframe);
16864 	return 0;
16865 }
16866 
16867 /* verify BPF_LD_IMM64 instruction */
check_ld_imm(struct bpf_verifier_env * env,struct bpf_insn * insn)16868 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
16869 {
16870 	struct bpf_insn_aux_data *aux = cur_aux(env);
16871 	struct bpf_reg_state *regs = cur_regs(env);
16872 	struct bpf_reg_state *dst_reg;
16873 	struct bpf_map *map;
16874 	int err;
16875 
16876 	if (BPF_SIZE(insn->code) != BPF_DW) {
16877 		verbose(env, "invalid BPF_LD_IMM insn\n");
16878 		return -EINVAL;
16879 	}
16880 	if (insn->off != 0) {
16881 		verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
16882 		return -EINVAL;
16883 	}
16884 
16885 	err = check_reg_arg(env, insn->dst_reg, DST_OP);
16886 	if (err)
16887 		return err;
16888 
16889 	dst_reg = &regs[insn->dst_reg];
16890 	if (insn->src_reg == 0) {
16891 		u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
16892 
16893 		dst_reg->type = SCALAR_VALUE;
16894 		__mark_reg_known(&regs[insn->dst_reg], imm);
16895 		return 0;
16896 	}
16897 
16898 	/* All special src_reg cases are listed below. From this point onwards
16899 	 * we either succeed and assign a corresponding dst_reg->type after
16900 	 * zeroing the offset, or fail and reject the program.
16901 	 */
16902 	mark_reg_known_zero(env, regs, insn->dst_reg);
16903 
16904 	if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
16905 		dst_reg->type = aux->btf_var.reg_type;
16906 		switch (base_type(dst_reg->type)) {
16907 		case PTR_TO_MEM:
16908 			dst_reg->mem_size = aux->btf_var.mem_size;
16909 			break;
16910 		case PTR_TO_BTF_ID:
16911 			dst_reg->btf = aux->btf_var.btf;
16912 			dst_reg->btf_id = aux->btf_var.btf_id;
16913 			break;
16914 		default:
16915 			verifier_bug(env, "pseudo btf id: unexpected dst reg type");
16916 			return -EFAULT;
16917 		}
16918 		return 0;
16919 	}
16920 
16921 	if (insn->src_reg == BPF_PSEUDO_FUNC) {
16922 		struct bpf_prog_aux *aux = env->prog->aux;
16923 		u32 subprogno = find_subprog(env,
16924 					     env->insn_idx + insn->imm + 1);
16925 
16926 		if (!aux->func_info) {
16927 			verbose(env, "missing btf func_info\n");
16928 			return -EINVAL;
16929 		}
16930 		if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
16931 			verbose(env, "callback function not static\n");
16932 			return -EINVAL;
16933 		}
16934 
16935 		dst_reg->type = PTR_TO_FUNC;
16936 		dst_reg->subprogno = subprogno;
16937 		return 0;
16938 	}
16939 
16940 	map = env->used_maps[aux->map_index];
16941 	dst_reg->map_ptr = map;
16942 
16943 	if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
16944 	    insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
16945 		if (map->map_type == BPF_MAP_TYPE_ARENA) {
16946 			__mark_reg_unknown(env, dst_reg);
16947 			return 0;
16948 		}
16949 		dst_reg->type = PTR_TO_MAP_VALUE;
16950 		dst_reg->off = aux->map_off;
16951 		WARN_ON_ONCE(map->max_entries != 1);
16952 		/* We want reg->id to be same (0) as map_value is not distinct */
16953 	} else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
16954 		   insn->src_reg == BPF_PSEUDO_MAP_IDX) {
16955 		dst_reg->type = CONST_PTR_TO_MAP;
16956 	} else {
16957 		verifier_bug(env, "unexpected src reg value for ldimm64");
16958 		return -EFAULT;
16959 	}
16960 
16961 	return 0;
16962 }
16963 
may_access_skb(enum bpf_prog_type type)16964 static bool may_access_skb(enum bpf_prog_type type)
16965 {
16966 	switch (type) {
16967 	case BPF_PROG_TYPE_SOCKET_FILTER:
16968 	case BPF_PROG_TYPE_SCHED_CLS:
16969 	case BPF_PROG_TYPE_SCHED_ACT:
16970 		return true;
16971 	default:
16972 		return false;
16973 	}
16974 }
16975 
16976 /* verify safety of LD_ABS|LD_IND instructions:
16977  * - they can only appear in the programs where ctx == skb
16978  * - since they are wrappers of function calls, they scratch R1-R5 registers,
16979  *   preserve R6-R9, and store return value into R0
16980  *
16981  * Implicit input:
16982  *   ctx == skb == R6 == CTX
16983  *
16984  * Explicit input:
16985  *   SRC == any register
16986  *   IMM == 32-bit immediate
16987  *
16988  * Output:
16989  *   R0 - 8/16/32-bit skb data converted to cpu endianness
16990  */
check_ld_abs(struct bpf_verifier_env * env,struct bpf_insn * insn)16991 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
16992 {
16993 	struct bpf_reg_state *regs = cur_regs(env);
16994 	static const int ctx_reg = BPF_REG_6;
16995 	u8 mode = BPF_MODE(insn->code);
16996 	int i, err;
16997 
16998 	if (!may_access_skb(resolve_prog_type(env->prog))) {
16999 		verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
17000 		return -EINVAL;
17001 	}
17002 
17003 	if (!env->ops->gen_ld_abs) {
17004 		verifier_bug(env, "gen_ld_abs is null");
17005 		return -EFAULT;
17006 	}
17007 
17008 	if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
17009 	    BPF_SIZE(insn->code) == BPF_DW ||
17010 	    (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
17011 		verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
17012 		return -EINVAL;
17013 	}
17014 
17015 	/* check whether implicit source operand (register R6) is readable */
17016 	err = check_reg_arg(env, ctx_reg, SRC_OP);
17017 	if (err)
17018 		return err;
17019 
17020 	/* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
17021 	 * gen_ld_abs() may terminate the program at runtime, leading to
17022 	 * reference leak.
17023 	 */
17024 	err = check_resource_leak(env, false, true, "BPF_LD_[ABS|IND]");
17025 	if (err)
17026 		return err;
17027 
17028 	if (regs[ctx_reg].type != PTR_TO_CTX) {
17029 		verbose(env,
17030 			"at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
17031 		return -EINVAL;
17032 	}
17033 
17034 	if (mode == BPF_IND) {
17035 		/* check explicit source operand */
17036 		err = check_reg_arg(env, insn->src_reg, SRC_OP);
17037 		if (err)
17038 			return err;
17039 	}
17040 
17041 	err = check_ptr_off_reg(env, &regs[ctx_reg], ctx_reg);
17042 	if (err < 0)
17043 		return err;
17044 
17045 	/* reset caller saved regs to unreadable */
17046 	for (i = 0; i < CALLER_SAVED_REGS; i++) {
17047 		mark_reg_not_init(env, regs, caller_saved[i]);
17048 		check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
17049 	}
17050 
17051 	/* mark destination R0 register as readable, since it contains
17052 	 * the value fetched from the packet.
17053 	 * Already marked as written above.
17054 	 */
17055 	mark_reg_unknown(env, regs, BPF_REG_0);
17056 	/* ld_abs load up to 32-bit skb data. */
17057 	regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
17058 	return 0;
17059 }
17060 
check_return_code(struct bpf_verifier_env * env,int regno,const char * reg_name)17061 static int check_return_code(struct bpf_verifier_env *env, int regno, const char *reg_name)
17062 {
17063 	const char *exit_ctx = "At program exit";
17064 	struct tnum enforce_attach_type_range = tnum_unknown;
17065 	const struct bpf_prog *prog = env->prog;
17066 	struct bpf_reg_state *reg = reg_state(env, regno);
17067 	struct bpf_retval_range range = retval_range(0, 1);
17068 	enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
17069 	int err;
17070 	struct bpf_func_state *frame = env->cur_state->frame[0];
17071 	const bool is_subprog = frame->subprogno;
17072 	bool return_32bit = false;
17073 	const struct btf_type *reg_type, *ret_type = NULL;
17074 
17075 	/* LSM and struct_ops func-ptr's return type could be "void" */
17076 	if (!is_subprog || frame->in_exception_callback_fn) {
17077 		switch (prog_type) {
17078 		case BPF_PROG_TYPE_LSM:
17079 			if (prog->expected_attach_type == BPF_LSM_CGROUP)
17080 				/* See below, can be 0 or 0-1 depending on hook. */
17081 				break;
17082 			if (!prog->aux->attach_func_proto->type)
17083 				return 0;
17084 			break;
17085 		case BPF_PROG_TYPE_STRUCT_OPS:
17086 			if (!prog->aux->attach_func_proto->type)
17087 				return 0;
17088 
17089 			if (frame->in_exception_callback_fn)
17090 				break;
17091 
17092 			/* Allow a struct_ops program to return a referenced kptr if it
17093 			 * matches the operator's return type and is in its unmodified
17094 			 * form. A scalar zero (i.e., a null pointer) is also allowed.
17095 			 */
17096 			reg_type = reg->btf ? btf_type_by_id(reg->btf, reg->btf_id) : NULL;
17097 			ret_type = btf_type_resolve_ptr(prog->aux->attach_btf,
17098 							prog->aux->attach_func_proto->type,
17099 							NULL);
17100 			if (ret_type && ret_type == reg_type && reg->ref_obj_id)
17101 				return __check_ptr_off_reg(env, reg, regno, false);
17102 			break;
17103 		default:
17104 			break;
17105 		}
17106 	}
17107 
17108 	/* eBPF calling convention is such that R0 is used
17109 	 * to return the value from eBPF program.
17110 	 * Make sure that it's readable at this time
17111 	 * of bpf_exit, which means that program wrote
17112 	 * something into it earlier
17113 	 */
17114 	err = check_reg_arg(env, regno, SRC_OP);
17115 	if (err)
17116 		return err;
17117 
17118 	if (is_pointer_value(env, regno)) {
17119 		verbose(env, "R%d leaks addr as return value\n", regno);
17120 		return -EACCES;
17121 	}
17122 
17123 	if (frame->in_async_callback_fn) {
17124 		/* enforce return zero from async callbacks like timer */
17125 		exit_ctx = "At async callback return";
17126 		range = retval_range(0, 0);
17127 		goto enforce_retval;
17128 	}
17129 
17130 	if (is_subprog && !frame->in_exception_callback_fn) {
17131 		if (reg->type != SCALAR_VALUE) {
17132 			verbose(env, "At subprogram exit the register R%d is not a scalar value (%s)\n",
17133 				regno, reg_type_str(env, reg->type));
17134 			return -EINVAL;
17135 		}
17136 		return 0;
17137 	}
17138 
17139 	switch (prog_type) {
17140 	case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
17141 		if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
17142 		    env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
17143 		    env->prog->expected_attach_type == BPF_CGROUP_UNIX_RECVMSG ||
17144 		    env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
17145 		    env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
17146 		    env->prog->expected_attach_type == BPF_CGROUP_UNIX_GETPEERNAME ||
17147 		    env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
17148 		    env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME ||
17149 		    env->prog->expected_attach_type == BPF_CGROUP_UNIX_GETSOCKNAME)
17150 			range = retval_range(1, 1);
17151 		if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
17152 		    env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
17153 			range = retval_range(0, 3);
17154 		break;
17155 	case BPF_PROG_TYPE_CGROUP_SKB:
17156 		if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
17157 			range = retval_range(0, 3);
17158 			enforce_attach_type_range = tnum_range(2, 3);
17159 		}
17160 		break;
17161 	case BPF_PROG_TYPE_CGROUP_SOCK:
17162 	case BPF_PROG_TYPE_SOCK_OPS:
17163 	case BPF_PROG_TYPE_CGROUP_DEVICE:
17164 	case BPF_PROG_TYPE_CGROUP_SYSCTL:
17165 	case BPF_PROG_TYPE_CGROUP_SOCKOPT:
17166 		break;
17167 	case BPF_PROG_TYPE_RAW_TRACEPOINT:
17168 		if (!env->prog->aux->attach_btf_id)
17169 			return 0;
17170 		range = retval_range(0, 0);
17171 		break;
17172 	case BPF_PROG_TYPE_TRACING:
17173 		switch (env->prog->expected_attach_type) {
17174 		case BPF_TRACE_FENTRY:
17175 		case BPF_TRACE_FEXIT:
17176 			range = retval_range(0, 0);
17177 			break;
17178 		case BPF_TRACE_RAW_TP:
17179 		case BPF_MODIFY_RETURN:
17180 			return 0;
17181 		case BPF_TRACE_ITER:
17182 			break;
17183 		default:
17184 			return -ENOTSUPP;
17185 		}
17186 		break;
17187 	case BPF_PROG_TYPE_KPROBE:
17188 		switch (env->prog->expected_attach_type) {
17189 		case BPF_TRACE_KPROBE_SESSION:
17190 		case BPF_TRACE_UPROBE_SESSION:
17191 			range = retval_range(0, 1);
17192 			break;
17193 		default:
17194 			return 0;
17195 		}
17196 		break;
17197 	case BPF_PROG_TYPE_SK_LOOKUP:
17198 		range = retval_range(SK_DROP, SK_PASS);
17199 		break;
17200 
17201 	case BPF_PROG_TYPE_LSM:
17202 		if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
17203 			/* no range found, any return value is allowed */
17204 			if (!get_func_retval_range(env->prog, &range))
17205 				return 0;
17206 			/* no restricted range, any return value is allowed */
17207 			if (range.minval == S32_MIN && range.maxval == S32_MAX)
17208 				return 0;
17209 			return_32bit = true;
17210 		} else if (!env->prog->aux->attach_func_proto->type) {
17211 			/* Make sure programs that attach to void
17212 			 * hooks don't try to modify return value.
17213 			 */
17214 			range = retval_range(1, 1);
17215 		}
17216 		break;
17217 
17218 	case BPF_PROG_TYPE_NETFILTER:
17219 		range = retval_range(NF_DROP, NF_ACCEPT);
17220 		break;
17221 	case BPF_PROG_TYPE_STRUCT_OPS:
17222 		if (!ret_type)
17223 			return 0;
17224 		range = retval_range(0, 0);
17225 		break;
17226 	case BPF_PROG_TYPE_EXT:
17227 		/* freplace program can return anything as its return value
17228 		 * depends on the to-be-replaced kernel func or bpf program.
17229 		 */
17230 	default:
17231 		return 0;
17232 	}
17233 
17234 enforce_retval:
17235 	if (reg->type != SCALAR_VALUE) {
17236 		verbose(env, "%s the register R%d is not a known value (%s)\n",
17237 			exit_ctx, regno, reg_type_str(env, reg->type));
17238 		return -EINVAL;
17239 	}
17240 
17241 	err = mark_chain_precision(env, regno);
17242 	if (err)
17243 		return err;
17244 
17245 	if (!retval_range_within(range, reg, return_32bit)) {
17246 		verbose_invalid_scalar(env, reg, range, exit_ctx, reg_name);
17247 		if (!is_subprog &&
17248 		    prog->expected_attach_type == BPF_LSM_CGROUP &&
17249 		    prog_type == BPF_PROG_TYPE_LSM &&
17250 		    !prog->aux->attach_func_proto->type)
17251 			verbose(env, "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
17252 		return -EINVAL;
17253 	}
17254 
17255 	if (!tnum_is_unknown(enforce_attach_type_range) &&
17256 	    tnum_in(enforce_attach_type_range, reg->var_off))
17257 		env->prog->enforce_expected_attach_type = 1;
17258 	return 0;
17259 }
17260 
mark_subprog_changes_pkt_data(struct bpf_verifier_env * env,int off)17261 static void mark_subprog_changes_pkt_data(struct bpf_verifier_env *env, int off)
17262 {
17263 	struct bpf_subprog_info *subprog;
17264 
17265 	subprog = find_containing_subprog(env, off);
17266 	subprog->changes_pkt_data = true;
17267 }
17268 
mark_subprog_might_sleep(struct bpf_verifier_env * env,int off)17269 static void mark_subprog_might_sleep(struct bpf_verifier_env *env, int off)
17270 {
17271 	struct bpf_subprog_info *subprog;
17272 
17273 	subprog = find_containing_subprog(env, off);
17274 	subprog->might_sleep = true;
17275 }
17276 
17277 /* 't' is an index of a call-site.
17278  * 'w' is a callee entry point.
17279  * Eventually this function would be called when env->cfg.insn_state[w] == EXPLORED.
17280  * Rely on DFS traversal order and absence of recursive calls to guarantee that
17281  * callee's change_pkt_data marks would be correct at that moment.
17282  */
merge_callee_effects(struct bpf_verifier_env * env,int t,int w)17283 static void merge_callee_effects(struct bpf_verifier_env *env, int t, int w)
17284 {
17285 	struct bpf_subprog_info *caller, *callee;
17286 
17287 	caller = find_containing_subprog(env, t);
17288 	callee = find_containing_subprog(env, w);
17289 	caller->changes_pkt_data |= callee->changes_pkt_data;
17290 	caller->might_sleep |= callee->might_sleep;
17291 }
17292 
17293 /* non-recursive DFS pseudo code
17294  * 1  procedure DFS-iterative(G,v):
17295  * 2      label v as discovered
17296  * 3      let S be a stack
17297  * 4      S.push(v)
17298  * 5      while S is not empty
17299  * 6            t <- S.peek()
17300  * 7            if t is what we're looking for:
17301  * 8                return t
17302  * 9            for all edges e in G.adjacentEdges(t) do
17303  * 10               if edge e is already labelled
17304  * 11                   continue with the next edge
17305  * 12               w <- G.adjacentVertex(t,e)
17306  * 13               if vertex w is not discovered and not explored
17307  * 14                   label e as tree-edge
17308  * 15                   label w as discovered
17309  * 16                   S.push(w)
17310  * 17                   continue at 5
17311  * 18               else if vertex w is discovered
17312  * 19                   label e as back-edge
17313  * 20               else
17314  * 21                   // vertex w is explored
17315  * 22                   label e as forward- or cross-edge
17316  * 23           label t as explored
17317  * 24           S.pop()
17318  *
17319  * convention:
17320  * 0x10 - discovered
17321  * 0x11 - discovered and fall-through edge labelled
17322  * 0x12 - discovered and fall-through and branch edges labelled
17323  * 0x20 - explored
17324  */
17325 
17326 enum {
17327 	DISCOVERED = 0x10,
17328 	EXPLORED = 0x20,
17329 	FALLTHROUGH = 1,
17330 	BRANCH = 2,
17331 };
17332 
mark_prune_point(struct bpf_verifier_env * env,int idx)17333 static void mark_prune_point(struct bpf_verifier_env *env, int idx)
17334 {
17335 	env->insn_aux_data[idx].prune_point = true;
17336 }
17337 
is_prune_point(struct bpf_verifier_env * env,int insn_idx)17338 static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
17339 {
17340 	return env->insn_aux_data[insn_idx].prune_point;
17341 }
17342 
mark_force_checkpoint(struct bpf_verifier_env * env,int idx)17343 static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
17344 {
17345 	env->insn_aux_data[idx].force_checkpoint = true;
17346 }
17347 
is_force_checkpoint(struct bpf_verifier_env * env,int insn_idx)17348 static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
17349 {
17350 	return env->insn_aux_data[insn_idx].force_checkpoint;
17351 }
17352 
mark_calls_callback(struct bpf_verifier_env * env,int idx)17353 static void mark_calls_callback(struct bpf_verifier_env *env, int idx)
17354 {
17355 	env->insn_aux_data[idx].calls_callback = true;
17356 }
17357 
calls_callback(struct bpf_verifier_env * env,int insn_idx)17358 static bool calls_callback(struct bpf_verifier_env *env, int insn_idx)
17359 {
17360 	return env->insn_aux_data[insn_idx].calls_callback;
17361 }
17362 
17363 enum {
17364 	DONE_EXPLORING = 0,
17365 	KEEP_EXPLORING = 1,
17366 };
17367 
17368 /* t, w, e - match pseudo-code above:
17369  * t - index of current instruction
17370  * w - next instruction
17371  * e - edge
17372  */
push_insn(int t,int w,int e,struct bpf_verifier_env * env)17373 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
17374 {
17375 	int *insn_stack = env->cfg.insn_stack;
17376 	int *insn_state = env->cfg.insn_state;
17377 
17378 	if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
17379 		return DONE_EXPLORING;
17380 
17381 	if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
17382 		return DONE_EXPLORING;
17383 
17384 	if (w < 0 || w >= env->prog->len) {
17385 		verbose_linfo(env, t, "%d: ", t);
17386 		verbose(env, "jump out of range from insn %d to %d\n", t, w);
17387 		return -EINVAL;
17388 	}
17389 
17390 	if (e == BRANCH) {
17391 		/* mark branch target for state pruning */
17392 		mark_prune_point(env, w);
17393 		mark_jmp_point(env, w);
17394 	}
17395 
17396 	if (insn_state[w] == 0) {
17397 		/* tree-edge */
17398 		insn_state[t] = DISCOVERED | e;
17399 		insn_state[w] = DISCOVERED;
17400 		if (env->cfg.cur_stack >= env->prog->len)
17401 			return -E2BIG;
17402 		insn_stack[env->cfg.cur_stack++] = w;
17403 		return KEEP_EXPLORING;
17404 	} else if ((insn_state[w] & 0xF0) == DISCOVERED) {
17405 		if (env->bpf_capable)
17406 			return DONE_EXPLORING;
17407 		verbose_linfo(env, t, "%d: ", t);
17408 		verbose_linfo(env, w, "%d: ", w);
17409 		verbose(env, "back-edge from insn %d to %d\n", t, w);
17410 		return -EINVAL;
17411 	} else if (insn_state[w] == EXPLORED) {
17412 		/* forward- or cross-edge */
17413 		insn_state[t] = DISCOVERED | e;
17414 	} else {
17415 		verifier_bug(env, "insn state internal bug");
17416 		return -EFAULT;
17417 	}
17418 	return DONE_EXPLORING;
17419 }
17420 
visit_func_call_insn(int t,struct bpf_insn * insns,struct bpf_verifier_env * env,bool visit_callee)17421 static int visit_func_call_insn(int t, struct bpf_insn *insns,
17422 				struct bpf_verifier_env *env,
17423 				bool visit_callee)
17424 {
17425 	int ret, insn_sz;
17426 	int w;
17427 
17428 	insn_sz = bpf_is_ldimm64(&insns[t]) ? 2 : 1;
17429 	ret = push_insn(t, t + insn_sz, FALLTHROUGH, env);
17430 	if (ret)
17431 		return ret;
17432 
17433 	mark_prune_point(env, t + insn_sz);
17434 	/* when we exit from subprog, we need to record non-linear history */
17435 	mark_jmp_point(env, t + insn_sz);
17436 
17437 	if (visit_callee) {
17438 		w = t + insns[t].imm + 1;
17439 		mark_prune_point(env, t);
17440 		merge_callee_effects(env, t, w);
17441 		ret = push_insn(t, w, BRANCH, env);
17442 	}
17443 	return ret;
17444 }
17445 
17446 /* Bitmask with 1s for all caller saved registers */
17447 #define ALL_CALLER_SAVED_REGS ((1u << CALLER_SAVED_REGS) - 1)
17448 
17449 /* True if do_misc_fixups() replaces calls to helper number 'imm',
17450  * replacement patch is presumed to follow bpf_fastcall contract
17451  * (see mark_fastcall_pattern_for_call() below).
17452  */
verifier_inlines_helper_call(struct bpf_verifier_env * env,s32 imm)17453 static bool verifier_inlines_helper_call(struct bpf_verifier_env *env, s32 imm)
17454 {
17455 	switch (imm) {
17456 #ifdef CONFIG_X86_64
17457 	case BPF_FUNC_get_smp_processor_id:
17458 		return env->prog->jit_requested && bpf_jit_supports_percpu_insn();
17459 #endif
17460 	default:
17461 		return false;
17462 	}
17463 }
17464 
17465 struct call_summary {
17466 	u8 num_params;
17467 	bool is_void;
17468 	bool fastcall;
17469 };
17470 
17471 /* If @call is a kfunc or helper call, fills @cs and returns true,
17472  * otherwise returns false.
17473  */
get_call_summary(struct bpf_verifier_env * env,struct bpf_insn * call,struct call_summary * cs)17474 static bool get_call_summary(struct bpf_verifier_env *env, struct bpf_insn *call,
17475 			     struct call_summary *cs)
17476 {
17477 	struct bpf_kfunc_call_arg_meta meta;
17478 	const struct bpf_func_proto *fn;
17479 	int i;
17480 
17481 	if (bpf_helper_call(call)) {
17482 
17483 		if (get_helper_proto(env, call->imm, &fn) < 0)
17484 			/* error would be reported later */
17485 			return false;
17486 		cs->fastcall = fn->allow_fastcall &&
17487 			       (verifier_inlines_helper_call(env, call->imm) ||
17488 				bpf_jit_inlines_helper_call(call->imm));
17489 		cs->is_void = fn->ret_type == RET_VOID;
17490 		cs->num_params = 0;
17491 		for (i = 0; i < ARRAY_SIZE(fn->arg_type); ++i) {
17492 			if (fn->arg_type[i] == ARG_DONTCARE)
17493 				break;
17494 			cs->num_params++;
17495 		}
17496 		return true;
17497 	}
17498 
17499 	if (bpf_pseudo_kfunc_call(call)) {
17500 		int err;
17501 
17502 		err = fetch_kfunc_meta(env, call, &meta, NULL);
17503 		if (err < 0)
17504 			/* error would be reported later */
17505 			return false;
17506 		cs->num_params = btf_type_vlen(meta.func_proto);
17507 		cs->fastcall = meta.kfunc_flags & KF_FASTCALL;
17508 		cs->is_void = btf_type_is_void(btf_type_by_id(meta.btf, meta.func_proto->type));
17509 		return true;
17510 	}
17511 
17512 	return false;
17513 }
17514 
17515 /* LLVM define a bpf_fastcall function attribute.
17516  * This attribute means that function scratches only some of
17517  * the caller saved registers defined by ABI.
17518  * For BPF the set of such registers could be defined as follows:
17519  * - R0 is scratched only if function is non-void;
17520  * - R1-R5 are scratched only if corresponding parameter type is defined
17521  *   in the function prototype.
17522  *
17523  * The contract between kernel and clang allows to simultaneously use
17524  * such functions and maintain backwards compatibility with old
17525  * kernels that don't understand bpf_fastcall calls:
17526  *
17527  * - for bpf_fastcall calls clang allocates registers as-if relevant r0-r5
17528  *   registers are not scratched by the call;
17529  *
17530  * - as a post-processing step, clang visits each bpf_fastcall call and adds
17531  *   spill/fill for every live r0-r5;
17532  *
17533  * - stack offsets used for the spill/fill are allocated as lowest
17534  *   stack offsets in whole function and are not used for any other
17535  *   purposes;
17536  *
17537  * - when kernel loads a program, it looks for such patterns
17538  *   (bpf_fastcall function surrounded by spills/fills) and checks if
17539  *   spill/fill stack offsets are used exclusively in fastcall patterns;
17540  *
17541  * - if so, and if verifier or current JIT inlines the call to the
17542  *   bpf_fastcall function (e.g. a helper call), kernel removes unnecessary
17543  *   spill/fill pairs;
17544  *
17545  * - when old kernel loads a program, presence of spill/fill pairs
17546  *   keeps BPF program valid, albeit slightly less efficient.
17547  *
17548  * For example:
17549  *
17550  *   r1 = 1;
17551  *   r2 = 2;
17552  *   *(u64 *)(r10 - 8)  = r1;            r1 = 1;
17553  *   *(u64 *)(r10 - 16) = r2;            r2 = 2;
17554  *   call %[to_be_inlined]         -->   call %[to_be_inlined]
17555  *   r2 = *(u64 *)(r10 - 16);            r0 = r1;
17556  *   r1 = *(u64 *)(r10 - 8);             r0 += r2;
17557  *   r0 = r1;                            exit;
17558  *   r0 += r2;
17559  *   exit;
17560  *
17561  * The purpose of mark_fastcall_pattern_for_call is to:
17562  * - look for such patterns;
17563  * - mark spill and fill instructions in env->insn_aux_data[*].fastcall_pattern;
17564  * - mark set env->insn_aux_data[*].fastcall_spills_num for call instruction;
17565  * - update env->subprog_info[*]->fastcall_stack_off to find an offset
17566  *   at which bpf_fastcall spill/fill stack slots start;
17567  * - update env->subprog_info[*]->keep_fastcall_stack.
17568  *
17569  * The .fastcall_pattern and .fastcall_stack_off are used by
17570  * check_fastcall_stack_contract() to check if every stack access to
17571  * fastcall spill/fill stack slot originates from spill/fill
17572  * instructions, members of fastcall patterns.
17573  *
17574  * If such condition holds true for a subprogram, fastcall patterns could
17575  * be rewritten by remove_fastcall_spills_fills().
17576  * Otherwise bpf_fastcall patterns are not changed in the subprogram
17577  * (code, presumably, generated by an older clang version).
17578  *
17579  * For example, it is *not* safe to remove spill/fill below:
17580  *
17581  *   r1 = 1;
17582  *   *(u64 *)(r10 - 8)  = r1;            r1 = 1;
17583  *   call %[to_be_inlined]         -->   call %[to_be_inlined]
17584  *   r1 = *(u64 *)(r10 - 8);             r0 = *(u64 *)(r10 - 8);  <---- wrong !!!
17585  *   r0 = *(u64 *)(r10 - 8);             r0 += r1;
17586  *   r0 += r1;                           exit;
17587  *   exit;
17588  */
mark_fastcall_pattern_for_call(struct bpf_verifier_env * env,struct bpf_subprog_info * subprog,int insn_idx,s16 lowest_off)17589 static void mark_fastcall_pattern_for_call(struct bpf_verifier_env *env,
17590 					   struct bpf_subprog_info *subprog,
17591 					   int insn_idx, s16 lowest_off)
17592 {
17593 	struct bpf_insn *insns = env->prog->insnsi, *stx, *ldx;
17594 	struct bpf_insn *call = &env->prog->insnsi[insn_idx];
17595 	u32 clobbered_regs_mask;
17596 	struct call_summary cs;
17597 	u32 expected_regs_mask;
17598 	s16 off;
17599 	int i;
17600 
17601 	if (!get_call_summary(env, call, &cs))
17602 		return;
17603 
17604 	/* A bitmask specifying which caller saved registers are clobbered
17605 	 * by a call to a helper/kfunc *as if* this helper/kfunc follows
17606 	 * bpf_fastcall contract:
17607 	 * - includes R0 if function is non-void;
17608 	 * - includes R1-R5 if corresponding parameter has is described
17609 	 *   in the function prototype.
17610 	 */
17611 	clobbered_regs_mask = GENMASK(cs.num_params, cs.is_void ? 1 : 0);
17612 	/* e.g. if helper call clobbers r{0,1}, expect r{2,3,4,5} in the pattern */
17613 	expected_regs_mask = ~clobbered_regs_mask & ALL_CALLER_SAVED_REGS;
17614 
17615 	/* match pairs of form:
17616 	 *
17617 	 * *(u64 *)(r10 - Y) = rX   (where Y % 8 == 0)
17618 	 * ...
17619 	 * call %[to_be_inlined]
17620 	 * ...
17621 	 * rX = *(u64 *)(r10 - Y)
17622 	 */
17623 	for (i = 1, off = lowest_off; i <= ARRAY_SIZE(caller_saved); ++i, off += BPF_REG_SIZE) {
17624 		if (insn_idx - i < 0 || insn_idx + i >= env->prog->len)
17625 			break;
17626 		stx = &insns[insn_idx - i];
17627 		ldx = &insns[insn_idx + i];
17628 		/* must be a stack spill/fill pair */
17629 		if (stx->code != (BPF_STX | BPF_MEM | BPF_DW) ||
17630 		    ldx->code != (BPF_LDX | BPF_MEM | BPF_DW) ||
17631 		    stx->dst_reg != BPF_REG_10 ||
17632 		    ldx->src_reg != BPF_REG_10)
17633 			break;
17634 		/* must be a spill/fill for the same reg */
17635 		if (stx->src_reg != ldx->dst_reg)
17636 			break;
17637 		/* must be one of the previously unseen registers */
17638 		if ((BIT(stx->src_reg) & expected_regs_mask) == 0)
17639 			break;
17640 		/* must be a spill/fill for the same expected offset,
17641 		 * no need to check offset alignment, BPF_DW stack access
17642 		 * is always 8-byte aligned.
17643 		 */
17644 		if (stx->off != off || ldx->off != off)
17645 			break;
17646 		expected_regs_mask &= ~BIT(stx->src_reg);
17647 		env->insn_aux_data[insn_idx - i].fastcall_pattern = 1;
17648 		env->insn_aux_data[insn_idx + i].fastcall_pattern = 1;
17649 	}
17650 	if (i == 1)
17651 		return;
17652 
17653 	/* Conditionally set 'fastcall_spills_num' to allow forward
17654 	 * compatibility when more helper functions are marked as
17655 	 * bpf_fastcall at compile time than current kernel supports, e.g:
17656 	 *
17657 	 *   1: *(u64 *)(r10 - 8) = r1
17658 	 *   2: call A                  ;; assume A is bpf_fastcall for current kernel
17659 	 *   3: r1 = *(u64 *)(r10 - 8)
17660 	 *   4: *(u64 *)(r10 - 8) = r1
17661 	 *   5: call B                  ;; assume B is not bpf_fastcall for current kernel
17662 	 *   6: r1 = *(u64 *)(r10 - 8)
17663 	 *
17664 	 * There is no need to block bpf_fastcall rewrite for such program.
17665 	 * Set 'fastcall_pattern' for both calls to keep check_fastcall_stack_contract() happy,
17666 	 * don't set 'fastcall_spills_num' for call B so that remove_fastcall_spills_fills()
17667 	 * does not remove spill/fill pair {4,6}.
17668 	 */
17669 	if (cs.fastcall)
17670 		env->insn_aux_data[insn_idx].fastcall_spills_num = i - 1;
17671 	else
17672 		subprog->keep_fastcall_stack = 1;
17673 	subprog->fastcall_stack_off = min(subprog->fastcall_stack_off, off);
17674 }
17675 
mark_fastcall_patterns(struct bpf_verifier_env * env)17676 static int mark_fastcall_patterns(struct bpf_verifier_env *env)
17677 {
17678 	struct bpf_subprog_info *subprog = env->subprog_info;
17679 	struct bpf_insn *insn;
17680 	s16 lowest_off;
17681 	int s, i;
17682 
17683 	for (s = 0; s < env->subprog_cnt; ++s, ++subprog) {
17684 		/* find lowest stack spill offset used in this subprog */
17685 		lowest_off = 0;
17686 		for (i = subprog->start; i < (subprog + 1)->start; ++i) {
17687 			insn = env->prog->insnsi + i;
17688 			if (insn->code != (BPF_STX | BPF_MEM | BPF_DW) ||
17689 			    insn->dst_reg != BPF_REG_10)
17690 				continue;
17691 			lowest_off = min(lowest_off, insn->off);
17692 		}
17693 		/* use this offset to find fastcall patterns */
17694 		for (i = subprog->start; i < (subprog + 1)->start; ++i) {
17695 			insn = env->prog->insnsi + i;
17696 			if (insn->code != (BPF_JMP | BPF_CALL))
17697 				continue;
17698 			mark_fastcall_pattern_for_call(env, subprog, i, lowest_off);
17699 		}
17700 	}
17701 	return 0;
17702 }
17703 
17704 /* Visits the instruction at index t and returns one of the following:
17705  *  < 0 - an error occurred
17706  *  DONE_EXPLORING - the instruction was fully explored
17707  *  KEEP_EXPLORING - there is still work to be done before it is fully explored
17708  */
visit_insn(int t,struct bpf_verifier_env * env)17709 static int visit_insn(int t, struct bpf_verifier_env *env)
17710 {
17711 	struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
17712 	int ret, off, insn_sz;
17713 
17714 	if (bpf_pseudo_func(insn))
17715 		return visit_func_call_insn(t, insns, env, true);
17716 
17717 	/* All non-branch instructions have a single fall-through edge. */
17718 	if (BPF_CLASS(insn->code) != BPF_JMP &&
17719 	    BPF_CLASS(insn->code) != BPF_JMP32) {
17720 		insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
17721 		return push_insn(t, t + insn_sz, FALLTHROUGH, env);
17722 	}
17723 
17724 	switch (BPF_OP(insn->code)) {
17725 	case BPF_EXIT:
17726 		return DONE_EXPLORING;
17727 
17728 	case BPF_CALL:
17729 		if (is_async_callback_calling_insn(insn))
17730 			/* Mark this call insn as a prune point to trigger
17731 			 * is_state_visited() check before call itself is
17732 			 * processed by __check_func_call(). Otherwise new
17733 			 * async state will be pushed for further exploration.
17734 			 */
17735 			mark_prune_point(env, t);
17736 		/* For functions that invoke callbacks it is not known how many times
17737 		 * callback would be called. Verifier models callback calling functions
17738 		 * by repeatedly visiting callback bodies and returning to origin call
17739 		 * instruction.
17740 		 * In order to stop such iteration verifier needs to identify when a
17741 		 * state identical some state from a previous iteration is reached.
17742 		 * Check below forces creation of checkpoint before callback calling
17743 		 * instruction to allow search for such identical states.
17744 		 */
17745 		if (is_sync_callback_calling_insn(insn)) {
17746 			mark_calls_callback(env, t);
17747 			mark_force_checkpoint(env, t);
17748 			mark_prune_point(env, t);
17749 			mark_jmp_point(env, t);
17750 		}
17751 		if (bpf_helper_call(insn)) {
17752 			const struct bpf_func_proto *fp;
17753 
17754 			ret = get_helper_proto(env, insn->imm, &fp);
17755 			/* If called in a non-sleepable context program will be
17756 			 * rejected anyway, so we should end up with precise
17757 			 * sleepable marks on subprogs, except for dead code
17758 			 * elimination.
17759 			 */
17760 			if (ret == 0 && fp->might_sleep)
17761 				mark_subprog_might_sleep(env, t);
17762 			if (bpf_helper_changes_pkt_data(insn->imm))
17763 				mark_subprog_changes_pkt_data(env, t);
17764 		} else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
17765 			struct bpf_kfunc_call_arg_meta meta;
17766 
17767 			ret = fetch_kfunc_meta(env, insn, &meta, NULL);
17768 			if (ret == 0 && is_iter_next_kfunc(&meta)) {
17769 				mark_prune_point(env, t);
17770 				/* Checking and saving state checkpoints at iter_next() call
17771 				 * is crucial for fast convergence of open-coded iterator loop
17772 				 * logic, so we need to force it. If we don't do that,
17773 				 * is_state_visited() might skip saving a checkpoint, causing
17774 				 * unnecessarily long sequence of not checkpointed
17775 				 * instructions and jumps, leading to exhaustion of jump
17776 				 * history buffer, and potentially other undesired outcomes.
17777 				 * It is expected that with correct open-coded iterators
17778 				 * convergence will happen quickly, so we don't run a risk of
17779 				 * exhausting memory.
17780 				 */
17781 				mark_force_checkpoint(env, t);
17782 			}
17783 			/* Same as helpers, if called in a non-sleepable context
17784 			 * program will be rejected anyway, so we should end up
17785 			 * with precise sleepable marks on subprogs, except for
17786 			 * dead code elimination.
17787 			 */
17788 			if (ret == 0 && is_kfunc_sleepable(&meta))
17789 				mark_subprog_might_sleep(env, t);
17790 		}
17791 		return visit_func_call_insn(t, insns, env, insn->src_reg == BPF_PSEUDO_CALL);
17792 
17793 	case BPF_JA:
17794 		if (BPF_SRC(insn->code) != BPF_K)
17795 			return -EINVAL;
17796 
17797 		if (BPF_CLASS(insn->code) == BPF_JMP)
17798 			off = insn->off;
17799 		else
17800 			off = insn->imm;
17801 
17802 		/* unconditional jump with single edge */
17803 		ret = push_insn(t, t + off + 1, FALLTHROUGH, env);
17804 		if (ret)
17805 			return ret;
17806 
17807 		mark_prune_point(env, t + off + 1);
17808 		mark_jmp_point(env, t + off + 1);
17809 
17810 		return ret;
17811 
17812 	default:
17813 		/* conditional jump with two edges */
17814 		mark_prune_point(env, t);
17815 		if (is_may_goto_insn(insn))
17816 			mark_force_checkpoint(env, t);
17817 
17818 		ret = push_insn(t, t + 1, FALLTHROUGH, env);
17819 		if (ret)
17820 			return ret;
17821 
17822 		return push_insn(t, t + insn->off + 1, BRANCH, env);
17823 	}
17824 }
17825 
17826 /* non-recursive depth-first-search to detect loops in BPF program
17827  * loop == back-edge in directed graph
17828  */
check_cfg(struct bpf_verifier_env * env)17829 static int check_cfg(struct bpf_verifier_env *env)
17830 {
17831 	int insn_cnt = env->prog->len;
17832 	int *insn_stack, *insn_state, *insn_postorder;
17833 	int ex_insn_beg, i, ret = 0;
17834 
17835 	insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
17836 	if (!insn_state)
17837 		return -ENOMEM;
17838 
17839 	insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
17840 	if (!insn_stack) {
17841 		kvfree(insn_state);
17842 		return -ENOMEM;
17843 	}
17844 
17845 	insn_postorder = env->cfg.insn_postorder =
17846 		kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
17847 	if (!insn_postorder) {
17848 		kvfree(insn_state);
17849 		kvfree(insn_stack);
17850 		return -ENOMEM;
17851 	}
17852 
17853 	ex_insn_beg = env->exception_callback_subprog
17854 		      ? env->subprog_info[env->exception_callback_subprog].start
17855 		      : 0;
17856 
17857 	insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
17858 	insn_stack[0] = 0; /* 0 is the first instruction */
17859 	env->cfg.cur_stack = 1;
17860 
17861 walk_cfg:
17862 	while (env->cfg.cur_stack > 0) {
17863 		int t = insn_stack[env->cfg.cur_stack - 1];
17864 
17865 		ret = visit_insn(t, env);
17866 		switch (ret) {
17867 		case DONE_EXPLORING:
17868 			insn_state[t] = EXPLORED;
17869 			env->cfg.cur_stack--;
17870 			insn_postorder[env->cfg.cur_postorder++] = t;
17871 			break;
17872 		case KEEP_EXPLORING:
17873 			break;
17874 		default:
17875 			if (ret > 0) {
17876 				verifier_bug(env, "visit_insn internal bug");
17877 				ret = -EFAULT;
17878 			}
17879 			goto err_free;
17880 		}
17881 	}
17882 
17883 	if (env->cfg.cur_stack < 0) {
17884 		verifier_bug(env, "pop stack internal bug");
17885 		ret = -EFAULT;
17886 		goto err_free;
17887 	}
17888 
17889 	if (ex_insn_beg && insn_state[ex_insn_beg] != EXPLORED) {
17890 		insn_state[ex_insn_beg] = DISCOVERED;
17891 		insn_stack[0] = ex_insn_beg;
17892 		env->cfg.cur_stack = 1;
17893 		goto walk_cfg;
17894 	}
17895 
17896 	for (i = 0; i < insn_cnt; i++) {
17897 		struct bpf_insn *insn = &env->prog->insnsi[i];
17898 
17899 		if (insn_state[i] != EXPLORED) {
17900 			verbose(env, "unreachable insn %d\n", i);
17901 			ret = -EINVAL;
17902 			goto err_free;
17903 		}
17904 		if (bpf_is_ldimm64(insn)) {
17905 			if (insn_state[i + 1] != 0) {
17906 				verbose(env, "jump into the middle of ldimm64 insn %d\n", i);
17907 				ret = -EINVAL;
17908 				goto err_free;
17909 			}
17910 			i++; /* skip second half of ldimm64 */
17911 		}
17912 	}
17913 	ret = 0; /* cfg looks good */
17914 	env->prog->aux->changes_pkt_data = env->subprog_info[0].changes_pkt_data;
17915 	env->prog->aux->might_sleep = env->subprog_info[0].might_sleep;
17916 
17917 err_free:
17918 	kvfree(insn_state);
17919 	kvfree(insn_stack);
17920 	env->cfg.insn_state = env->cfg.insn_stack = NULL;
17921 	return ret;
17922 }
17923 
check_abnormal_return(struct bpf_verifier_env * env)17924 static int check_abnormal_return(struct bpf_verifier_env *env)
17925 {
17926 	int i;
17927 
17928 	for (i = 1; i < env->subprog_cnt; i++) {
17929 		if (env->subprog_info[i].has_ld_abs) {
17930 			verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
17931 			return -EINVAL;
17932 		}
17933 		if (env->subprog_info[i].has_tail_call) {
17934 			verbose(env, "tail_call is not allowed in subprogs without BTF\n");
17935 			return -EINVAL;
17936 		}
17937 	}
17938 	return 0;
17939 }
17940 
17941 /* The minimum supported BTF func info size */
17942 #define MIN_BPF_FUNCINFO_SIZE	8
17943 #define MAX_FUNCINFO_REC_SIZE	252
17944 
check_btf_func_early(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)17945 static int check_btf_func_early(struct bpf_verifier_env *env,
17946 				const union bpf_attr *attr,
17947 				bpfptr_t uattr)
17948 {
17949 	u32 krec_size = sizeof(struct bpf_func_info);
17950 	const struct btf_type *type, *func_proto;
17951 	u32 i, nfuncs, urec_size, min_size;
17952 	struct bpf_func_info *krecord;
17953 	struct bpf_prog *prog;
17954 	const struct btf *btf;
17955 	u32 prev_offset = 0;
17956 	bpfptr_t urecord;
17957 	int ret = -ENOMEM;
17958 
17959 	nfuncs = attr->func_info_cnt;
17960 	if (!nfuncs) {
17961 		if (check_abnormal_return(env))
17962 			return -EINVAL;
17963 		return 0;
17964 	}
17965 
17966 	urec_size = attr->func_info_rec_size;
17967 	if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
17968 	    urec_size > MAX_FUNCINFO_REC_SIZE ||
17969 	    urec_size % sizeof(u32)) {
17970 		verbose(env, "invalid func info rec size %u\n", urec_size);
17971 		return -EINVAL;
17972 	}
17973 
17974 	prog = env->prog;
17975 	btf = prog->aux->btf;
17976 
17977 	urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
17978 	min_size = min_t(u32, krec_size, urec_size);
17979 
17980 	krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
17981 	if (!krecord)
17982 		return -ENOMEM;
17983 
17984 	for (i = 0; i < nfuncs; i++) {
17985 		ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
17986 		if (ret) {
17987 			if (ret == -E2BIG) {
17988 				verbose(env, "nonzero tailing record in func info");
17989 				/* set the size kernel expects so loader can zero
17990 				 * out the rest of the record.
17991 				 */
17992 				if (copy_to_bpfptr_offset(uattr,
17993 							  offsetof(union bpf_attr, func_info_rec_size),
17994 							  &min_size, sizeof(min_size)))
17995 					ret = -EFAULT;
17996 			}
17997 			goto err_free;
17998 		}
17999 
18000 		if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
18001 			ret = -EFAULT;
18002 			goto err_free;
18003 		}
18004 
18005 		/* check insn_off */
18006 		ret = -EINVAL;
18007 		if (i == 0) {
18008 			if (krecord[i].insn_off) {
18009 				verbose(env,
18010 					"nonzero insn_off %u for the first func info record",
18011 					krecord[i].insn_off);
18012 				goto err_free;
18013 			}
18014 		} else if (krecord[i].insn_off <= prev_offset) {
18015 			verbose(env,
18016 				"same or smaller insn offset (%u) than previous func info record (%u)",
18017 				krecord[i].insn_off, prev_offset);
18018 			goto err_free;
18019 		}
18020 
18021 		/* check type_id */
18022 		type = btf_type_by_id(btf, krecord[i].type_id);
18023 		if (!type || !btf_type_is_func(type)) {
18024 			verbose(env, "invalid type id %d in func info",
18025 				krecord[i].type_id);
18026 			goto err_free;
18027 		}
18028 
18029 		func_proto = btf_type_by_id(btf, type->type);
18030 		if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
18031 			/* btf_func_check() already verified it during BTF load */
18032 			goto err_free;
18033 
18034 		prev_offset = krecord[i].insn_off;
18035 		bpfptr_add(&urecord, urec_size);
18036 	}
18037 
18038 	prog->aux->func_info = krecord;
18039 	prog->aux->func_info_cnt = nfuncs;
18040 	return 0;
18041 
18042 err_free:
18043 	kvfree(krecord);
18044 	return ret;
18045 }
18046 
check_btf_func(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)18047 static int check_btf_func(struct bpf_verifier_env *env,
18048 			  const union bpf_attr *attr,
18049 			  bpfptr_t uattr)
18050 {
18051 	const struct btf_type *type, *func_proto, *ret_type;
18052 	u32 i, nfuncs, urec_size;
18053 	struct bpf_func_info *krecord;
18054 	struct bpf_func_info_aux *info_aux = NULL;
18055 	struct bpf_prog *prog;
18056 	const struct btf *btf;
18057 	bpfptr_t urecord;
18058 	bool scalar_return;
18059 	int ret = -ENOMEM;
18060 
18061 	nfuncs = attr->func_info_cnt;
18062 	if (!nfuncs) {
18063 		if (check_abnormal_return(env))
18064 			return -EINVAL;
18065 		return 0;
18066 	}
18067 	if (nfuncs != env->subprog_cnt) {
18068 		verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
18069 		return -EINVAL;
18070 	}
18071 
18072 	urec_size = attr->func_info_rec_size;
18073 
18074 	prog = env->prog;
18075 	btf = prog->aux->btf;
18076 
18077 	urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
18078 
18079 	krecord = prog->aux->func_info;
18080 	info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
18081 	if (!info_aux)
18082 		return -ENOMEM;
18083 
18084 	for (i = 0; i < nfuncs; i++) {
18085 		/* check insn_off */
18086 		ret = -EINVAL;
18087 
18088 		if (env->subprog_info[i].start != krecord[i].insn_off) {
18089 			verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
18090 			goto err_free;
18091 		}
18092 
18093 		/* Already checked type_id */
18094 		type = btf_type_by_id(btf, krecord[i].type_id);
18095 		info_aux[i].linkage = BTF_INFO_VLEN(type->info);
18096 		/* Already checked func_proto */
18097 		func_proto = btf_type_by_id(btf, type->type);
18098 
18099 		ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
18100 		scalar_return =
18101 			btf_type_is_small_int(ret_type) || btf_is_any_enum(ret_type);
18102 		if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
18103 			verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
18104 			goto err_free;
18105 		}
18106 		if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
18107 			verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
18108 			goto err_free;
18109 		}
18110 
18111 		bpfptr_add(&urecord, urec_size);
18112 	}
18113 
18114 	prog->aux->func_info_aux = info_aux;
18115 	return 0;
18116 
18117 err_free:
18118 	kfree(info_aux);
18119 	return ret;
18120 }
18121 
adjust_btf_func(struct bpf_verifier_env * env)18122 static void adjust_btf_func(struct bpf_verifier_env *env)
18123 {
18124 	struct bpf_prog_aux *aux = env->prog->aux;
18125 	int i;
18126 
18127 	if (!aux->func_info)
18128 		return;
18129 
18130 	/* func_info is not available for hidden subprogs */
18131 	for (i = 0; i < env->subprog_cnt - env->hidden_subprog_cnt; i++)
18132 		aux->func_info[i].insn_off = env->subprog_info[i].start;
18133 }
18134 
18135 #define MIN_BPF_LINEINFO_SIZE	offsetofend(struct bpf_line_info, line_col)
18136 #define MAX_LINEINFO_REC_SIZE	MAX_FUNCINFO_REC_SIZE
18137 
check_btf_line(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)18138 static int check_btf_line(struct bpf_verifier_env *env,
18139 			  const union bpf_attr *attr,
18140 			  bpfptr_t uattr)
18141 {
18142 	u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
18143 	struct bpf_subprog_info *sub;
18144 	struct bpf_line_info *linfo;
18145 	struct bpf_prog *prog;
18146 	const struct btf *btf;
18147 	bpfptr_t ulinfo;
18148 	int err;
18149 
18150 	nr_linfo = attr->line_info_cnt;
18151 	if (!nr_linfo)
18152 		return 0;
18153 	if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
18154 		return -EINVAL;
18155 
18156 	rec_size = attr->line_info_rec_size;
18157 	if (rec_size < MIN_BPF_LINEINFO_SIZE ||
18158 	    rec_size > MAX_LINEINFO_REC_SIZE ||
18159 	    rec_size & (sizeof(u32) - 1))
18160 		return -EINVAL;
18161 
18162 	/* Need to zero it in case the userspace may
18163 	 * pass in a smaller bpf_line_info object.
18164 	 */
18165 	linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
18166 			 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
18167 	if (!linfo)
18168 		return -ENOMEM;
18169 
18170 	prog = env->prog;
18171 	btf = prog->aux->btf;
18172 
18173 	s = 0;
18174 	sub = env->subprog_info;
18175 	ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
18176 	expected_size = sizeof(struct bpf_line_info);
18177 	ncopy = min_t(u32, expected_size, rec_size);
18178 	for (i = 0; i < nr_linfo; i++) {
18179 		err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
18180 		if (err) {
18181 			if (err == -E2BIG) {
18182 				verbose(env, "nonzero tailing record in line_info");
18183 				if (copy_to_bpfptr_offset(uattr,
18184 							  offsetof(union bpf_attr, line_info_rec_size),
18185 							  &expected_size, sizeof(expected_size)))
18186 					err = -EFAULT;
18187 			}
18188 			goto err_free;
18189 		}
18190 
18191 		if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
18192 			err = -EFAULT;
18193 			goto err_free;
18194 		}
18195 
18196 		/*
18197 		 * Check insn_off to ensure
18198 		 * 1) strictly increasing AND
18199 		 * 2) bounded by prog->len
18200 		 *
18201 		 * The linfo[0].insn_off == 0 check logically falls into
18202 		 * the later "missing bpf_line_info for func..." case
18203 		 * because the first linfo[0].insn_off must be the
18204 		 * first sub also and the first sub must have
18205 		 * subprog_info[0].start == 0.
18206 		 */
18207 		if ((i && linfo[i].insn_off <= prev_offset) ||
18208 		    linfo[i].insn_off >= prog->len) {
18209 			verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
18210 				i, linfo[i].insn_off, prev_offset,
18211 				prog->len);
18212 			err = -EINVAL;
18213 			goto err_free;
18214 		}
18215 
18216 		if (!prog->insnsi[linfo[i].insn_off].code) {
18217 			verbose(env,
18218 				"Invalid insn code at line_info[%u].insn_off\n",
18219 				i);
18220 			err = -EINVAL;
18221 			goto err_free;
18222 		}
18223 
18224 		if (!btf_name_by_offset(btf, linfo[i].line_off) ||
18225 		    !btf_name_by_offset(btf, linfo[i].file_name_off)) {
18226 			verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
18227 			err = -EINVAL;
18228 			goto err_free;
18229 		}
18230 
18231 		if (s != env->subprog_cnt) {
18232 			if (linfo[i].insn_off == sub[s].start) {
18233 				sub[s].linfo_idx = i;
18234 				s++;
18235 			} else if (sub[s].start < linfo[i].insn_off) {
18236 				verbose(env, "missing bpf_line_info for func#%u\n", s);
18237 				err = -EINVAL;
18238 				goto err_free;
18239 			}
18240 		}
18241 
18242 		prev_offset = linfo[i].insn_off;
18243 		bpfptr_add(&ulinfo, rec_size);
18244 	}
18245 
18246 	if (s != env->subprog_cnt) {
18247 		verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
18248 			env->subprog_cnt - s, s);
18249 		err = -EINVAL;
18250 		goto err_free;
18251 	}
18252 
18253 	prog->aux->linfo = linfo;
18254 	prog->aux->nr_linfo = nr_linfo;
18255 
18256 	return 0;
18257 
18258 err_free:
18259 	kvfree(linfo);
18260 	return err;
18261 }
18262 
18263 #define MIN_CORE_RELO_SIZE	sizeof(struct bpf_core_relo)
18264 #define MAX_CORE_RELO_SIZE	MAX_FUNCINFO_REC_SIZE
18265 
check_core_relo(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)18266 static int check_core_relo(struct bpf_verifier_env *env,
18267 			   const union bpf_attr *attr,
18268 			   bpfptr_t uattr)
18269 {
18270 	u32 i, nr_core_relo, ncopy, expected_size, rec_size;
18271 	struct bpf_core_relo core_relo = {};
18272 	struct bpf_prog *prog = env->prog;
18273 	const struct btf *btf = prog->aux->btf;
18274 	struct bpf_core_ctx ctx = {
18275 		.log = &env->log,
18276 		.btf = btf,
18277 	};
18278 	bpfptr_t u_core_relo;
18279 	int err;
18280 
18281 	nr_core_relo = attr->core_relo_cnt;
18282 	if (!nr_core_relo)
18283 		return 0;
18284 	if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
18285 		return -EINVAL;
18286 
18287 	rec_size = attr->core_relo_rec_size;
18288 	if (rec_size < MIN_CORE_RELO_SIZE ||
18289 	    rec_size > MAX_CORE_RELO_SIZE ||
18290 	    rec_size % sizeof(u32))
18291 		return -EINVAL;
18292 
18293 	u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
18294 	expected_size = sizeof(struct bpf_core_relo);
18295 	ncopy = min_t(u32, expected_size, rec_size);
18296 
18297 	/* Unlike func_info and line_info, copy and apply each CO-RE
18298 	 * relocation record one at a time.
18299 	 */
18300 	for (i = 0; i < nr_core_relo; i++) {
18301 		/* future proofing when sizeof(bpf_core_relo) changes */
18302 		err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
18303 		if (err) {
18304 			if (err == -E2BIG) {
18305 				verbose(env, "nonzero tailing record in core_relo");
18306 				if (copy_to_bpfptr_offset(uattr,
18307 							  offsetof(union bpf_attr, core_relo_rec_size),
18308 							  &expected_size, sizeof(expected_size)))
18309 					err = -EFAULT;
18310 			}
18311 			break;
18312 		}
18313 
18314 		if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
18315 			err = -EFAULT;
18316 			break;
18317 		}
18318 
18319 		if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
18320 			verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
18321 				i, core_relo.insn_off, prog->len);
18322 			err = -EINVAL;
18323 			break;
18324 		}
18325 
18326 		err = bpf_core_apply(&ctx, &core_relo, i,
18327 				     &prog->insnsi[core_relo.insn_off / 8]);
18328 		if (err)
18329 			break;
18330 		bpfptr_add(&u_core_relo, rec_size);
18331 	}
18332 	return err;
18333 }
18334 
check_btf_info_early(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)18335 static int check_btf_info_early(struct bpf_verifier_env *env,
18336 				const union bpf_attr *attr,
18337 				bpfptr_t uattr)
18338 {
18339 	struct btf *btf;
18340 	int err;
18341 
18342 	if (!attr->func_info_cnt && !attr->line_info_cnt) {
18343 		if (check_abnormal_return(env))
18344 			return -EINVAL;
18345 		return 0;
18346 	}
18347 
18348 	btf = btf_get_by_fd(attr->prog_btf_fd);
18349 	if (IS_ERR(btf))
18350 		return PTR_ERR(btf);
18351 	if (btf_is_kernel(btf)) {
18352 		btf_put(btf);
18353 		return -EACCES;
18354 	}
18355 	env->prog->aux->btf = btf;
18356 
18357 	err = check_btf_func_early(env, attr, uattr);
18358 	if (err)
18359 		return err;
18360 	return 0;
18361 }
18362 
check_btf_info(struct bpf_verifier_env * env,const union bpf_attr * attr,bpfptr_t uattr)18363 static int check_btf_info(struct bpf_verifier_env *env,
18364 			  const union bpf_attr *attr,
18365 			  bpfptr_t uattr)
18366 {
18367 	int err;
18368 
18369 	if (!attr->func_info_cnt && !attr->line_info_cnt) {
18370 		if (check_abnormal_return(env))
18371 			return -EINVAL;
18372 		return 0;
18373 	}
18374 
18375 	err = check_btf_func(env, attr, uattr);
18376 	if (err)
18377 		return err;
18378 
18379 	err = check_btf_line(env, attr, uattr);
18380 	if (err)
18381 		return err;
18382 
18383 	err = check_core_relo(env, attr, uattr);
18384 	if (err)
18385 		return err;
18386 
18387 	return 0;
18388 }
18389 
18390 /* check %cur's range satisfies %old's */
range_within(const struct bpf_reg_state * old,const struct bpf_reg_state * cur)18391 static bool range_within(const struct bpf_reg_state *old,
18392 			 const struct bpf_reg_state *cur)
18393 {
18394 	return old->umin_value <= cur->umin_value &&
18395 	       old->umax_value >= cur->umax_value &&
18396 	       old->smin_value <= cur->smin_value &&
18397 	       old->smax_value >= cur->smax_value &&
18398 	       old->u32_min_value <= cur->u32_min_value &&
18399 	       old->u32_max_value >= cur->u32_max_value &&
18400 	       old->s32_min_value <= cur->s32_min_value &&
18401 	       old->s32_max_value >= cur->s32_max_value;
18402 }
18403 
18404 /* If in the old state two registers had the same id, then they need to have
18405  * the same id in the new state as well.  But that id could be different from
18406  * the old state, so we need to track the mapping from old to new ids.
18407  * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
18408  * regs with old id 5 must also have new id 9 for the new state to be safe.  But
18409  * regs with a different old id could still have new id 9, we don't care about
18410  * that.
18411  * So we look through our idmap to see if this old id has been seen before.  If
18412  * so, we require the new id to match; otherwise, we add the id pair to the map.
18413  */
check_ids(u32 old_id,u32 cur_id,struct bpf_idmap * idmap)18414 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
18415 {
18416 	struct bpf_id_pair *map = idmap->map;
18417 	unsigned int i;
18418 
18419 	/* either both IDs should be set or both should be zero */
18420 	if (!!old_id != !!cur_id)
18421 		return false;
18422 
18423 	if (old_id == 0) /* cur_id == 0 as well */
18424 		return true;
18425 
18426 	for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
18427 		if (!map[i].old) {
18428 			/* Reached an empty slot; haven't seen this id before */
18429 			map[i].old = old_id;
18430 			map[i].cur = cur_id;
18431 			return true;
18432 		}
18433 		if (map[i].old == old_id)
18434 			return map[i].cur == cur_id;
18435 		if (map[i].cur == cur_id)
18436 			return false;
18437 	}
18438 	/* We ran out of idmap slots, which should be impossible */
18439 	WARN_ON_ONCE(1);
18440 	return false;
18441 }
18442 
18443 /* Similar to check_ids(), but allocate a unique temporary ID
18444  * for 'old_id' or 'cur_id' of zero.
18445  * This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
18446  */
check_scalar_ids(u32 old_id,u32 cur_id,struct bpf_idmap * idmap)18447 static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
18448 {
18449 	old_id = old_id ? old_id : ++idmap->tmp_id_gen;
18450 	cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
18451 
18452 	return check_ids(old_id, cur_id, idmap);
18453 }
18454 
clean_func_state(struct bpf_verifier_env * env,struct bpf_func_state * st)18455 static void clean_func_state(struct bpf_verifier_env *env,
18456 			     struct bpf_func_state *st)
18457 {
18458 	enum bpf_reg_liveness live;
18459 	int i, j;
18460 
18461 	for (i = 0; i < BPF_REG_FP; i++) {
18462 		live = st->regs[i].live;
18463 		/* liveness must not touch this register anymore */
18464 		st->regs[i].live |= REG_LIVE_DONE;
18465 		if (!(live & REG_LIVE_READ))
18466 			/* since the register is unused, clear its state
18467 			 * to make further comparison simpler
18468 			 */
18469 			__mark_reg_not_init(env, &st->regs[i]);
18470 	}
18471 
18472 	for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
18473 		live = st->stack[i].spilled_ptr.live;
18474 		/* liveness must not touch this stack slot anymore */
18475 		st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
18476 		if (!(live & REG_LIVE_READ)) {
18477 			__mark_reg_not_init(env, &st->stack[i].spilled_ptr);
18478 			for (j = 0; j < BPF_REG_SIZE; j++)
18479 				st->stack[i].slot_type[j] = STACK_INVALID;
18480 		}
18481 	}
18482 }
18483 
clean_verifier_state(struct bpf_verifier_env * env,struct bpf_verifier_state * st)18484 static void clean_verifier_state(struct bpf_verifier_env *env,
18485 				 struct bpf_verifier_state *st)
18486 {
18487 	int i;
18488 
18489 	for (i = 0; i <= st->curframe; i++)
18490 		clean_func_state(env, st->frame[i]);
18491 }
18492 
18493 /* the parentage chains form a tree.
18494  * the verifier states are added to state lists at given insn and
18495  * pushed into state stack for future exploration.
18496  * when the verifier reaches bpf_exit insn some of the verifier states
18497  * stored in the state lists have their final liveness state already,
18498  * but a lot of states will get revised from liveness point of view when
18499  * the verifier explores other branches.
18500  * Example:
18501  * 1: r0 = 1
18502  * 2: if r1 == 100 goto pc+1
18503  * 3: r0 = 2
18504  * 4: exit
18505  * when the verifier reaches exit insn the register r0 in the state list of
18506  * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
18507  * of insn 2 and goes exploring further. At the insn 4 it will walk the
18508  * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
18509  *
18510  * Since the verifier pushes the branch states as it sees them while exploring
18511  * the program the condition of walking the branch instruction for the second
18512  * time means that all states below this branch were already explored and
18513  * their final liveness marks are already propagated.
18514  * Hence when the verifier completes the search of state list in is_state_visited()
18515  * we can call this clean_live_states() function to mark all liveness states
18516  * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
18517  * will not be used.
18518  * This function also clears the registers and stack for states that !READ
18519  * to simplify state merging.
18520  *
18521  * Important note here that walking the same branch instruction in the callee
18522  * doesn't meant that the states are DONE. The verifier has to compare
18523  * the callsites
18524  */
clean_live_states(struct bpf_verifier_env * env,int insn,struct bpf_verifier_state * cur)18525 static void clean_live_states(struct bpf_verifier_env *env, int insn,
18526 			      struct bpf_verifier_state *cur)
18527 {
18528 	struct bpf_verifier_state_list *sl;
18529 	struct list_head *pos, *head;
18530 
18531 	head = explored_state(env, insn);
18532 	list_for_each(pos, head) {
18533 		sl = container_of(pos, struct bpf_verifier_state_list, node);
18534 		if (sl->state.branches)
18535 			continue;
18536 		if (sl->state.insn_idx != insn ||
18537 		    !same_callsites(&sl->state, cur))
18538 			continue;
18539 		if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE)
18540 			/* all regs in this state in all frames were already marked */
18541 			continue;
18542 		if (incomplete_read_marks(env, &sl->state))
18543 			continue;
18544 		clean_verifier_state(env, &sl->state);
18545 	}
18546 }
18547 
regs_exact(const struct bpf_reg_state * rold,const struct bpf_reg_state * rcur,struct bpf_idmap * idmap)18548 static bool regs_exact(const struct bpf_reg_state *rold,
18549 		       const struct bpf_reg_state *rcur,
18550 		       struct bpf_idmap *idmap)
18551 {
18552 	return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
18553 	       check_ids(rold->id, rcur->id, idmap) &&
18554 	       check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
18555 }
18556 
18557 enum exact_level {
18558 	NOT_EXACT,
18559 	EXACT,
18560 	RANGE_WITHIN
18561 };
18562 
18563 /* Returns true if (rold safe implies rcur safe) */
regsafe(struct bpf_verifier_env * env,struct bpf_reg_state * rold,struct bpf_reg_state * rcur,struct bpf_idmap * idmap,enum exact_level exact)18564 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
18565 		    struct bpf_reg_state *rcur, struct bpf_idmap *idmap,
18566 		    enum exact_level exact)
18567 {
18568 	if (exact == EXACT)
18569 		return regs_exact(rold, rcur, idmap);
18570 
18571 	if (!(rold->live & REG_LIVE_READ) && exact == NOT_EXACT)
18572 		/* explored state didn't use this */
18573 		return true;
18574 	if (rold->type == NOT_INIT) {
18575 		if (exact == NOT_EXACT || rcur->type == NOT_INIT)
18576 			/* explored state can't have used this */
18577 			return true;
18578 	}
18579 
18580 	/* Enforce that register types have to match exactly, including their
18581 	 * modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
18582 	 * rule.
18583 	 *
18584 	 * One can make a point that using a pointer register as unbounded
18585 	 * SCALAR would be technically acceptable, but this could lead to
18586 	 * pointer leaks because scalars are allowed to leak while pointers
18587 	 * are not. We could make this safe in special cases if root is
18588 	 * calling us, but it's probably not worth the hassle.
18589 	 *
18590 	 * Also, register types that are *not* MAYBE_NULL could technically be
18591 	 * safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
18592 	 * is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
18593 	 * to the same map).
18594 	 * However, if the old MAYBE_NULL register then got NULL checked,
18595 	 * doing so could have affected others with the same id, and we can't
18596 	 * check for that because we lost the id when we converted to
18597 	 * a non-MAYBE_NULL variant.
18598 	 * So, as a general rule we don't allow mixing MAYBE_NULL and
18599 	 * non-MAYBE_NULL registers as well.
18600 	 */
18601 	if (rold->type != rcur->type)
18602 		return false;
18603 
18604 	switch (base_type(rold->type)) {
18605 	case SCALAR_VALUE:
18606 		if (env->explore_alu_limits) {
18607 			/* explore_alu_limits disables tnum_in() and range_within()
18608 			 * logic and requires everything to be strict
18609 			 */
18610 			return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
18611 			       check_scalar_ids(rold->id, rcur->id, idmap);
18612 		}
18613 		if (!rold->precise && exact == NOT_EXACT)
18614 			return true;
18615 		if ((rold->id & BPF_ADD_CONST) != (rcur->id & BPF_ADD_CONST))
18616 			return false;
18617 		if ((rold->id & BPF_ADD_CONST) && (rold->off != rcur->off))
18618 			return false;
18619 		/* Why check_ids() for scalar registers?
18620 		 *
18621 		 * Consider the following BPF code:
18622 		 *   1: r6 = ... unbound scalar, ID=a ...
18623 		 *   2: r7 = ... unbound scalar, ID=b ...
18624 		 *   3: if (r6 > r7) goto +1
18625 		 *   4: r6 = r7
18626 		 *   5: if (r6 > X) goto ...
18627 		 *   6: ... memory operation using r7 ...
18628 		 *
18629 		 * First verification path is [1-6]:
18630 		 * - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
18631 		 * - at (5) r6 would be marked <= X, sync_linked_regs() would also mark
18632 		 *   r7 <= X, because r6 and r7 share same id.
18633 		 * Next verification path is [1-4, 6].
18634 		 *
18635 		 * Instruction (6) would be reached in two states:
18636 		 *   I.  r6{.id=b}, r7{.id=b} via path 1-6;
18637 		 *   II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
18638 		 *
18639 		 * Use check_ids() to distinguish these states.
18640 		 * ---
18641 		 * Also verify that new value satisfies old value range knowledge.
18642 		 */
18643 		return range_within(rold, rcur) &&
18644 		       tnum_in(rold->var_off, rcur->var_off) &&
18645 		       check_scalar_ids(rold->id, rcur->id, idmap);
18646 	case PTR_TO_MAP_KEY:
18647 	case PTR_TO_MAP_VALUE:
18648 	case PTR_TO_MEM:
18649 	case PTR_TO_BUF:
18650 	case PTR_TO_TP_BUFFER:
18651 		/* If the new min/max/var_off satisfy the old ones and
18652 		 * everything else matches, we are OK.
18653 		 */
18654 		return memcmp(rold, rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
18655 		       range_within(rold, rcur) &&
18656 		       tnum_in(rold->var_off, rcur->var_off) &&
18657 		       check_ids(rold->id, rcur->id, idmap) &&
18658 		       check_ids(rold->ref_obj_id, rcur->ref_obj_id, idmap);
18659 	case PTR_TO_PACKET_META:
18660 	case PTR_TO_PACKET:
18661 		/* We must have at least as much range as the old ptr
18662 		 * did, so that any accesses which were safe before are
18663 		 * still safe.  This is true even if old range < old off,
18664 		 * since someone could have accessed through (ptr - k), or
18665 		 * even done ptr -= k in a register, to get a safe access.
18666 		 */
18667 		if (rold->range > rcur->range)
18668 			return false;
18669 		/* If the offsets don't match, we can't trust our alignment;
18670 		 * nor can we be sure that we won't fall out of range.
18671 		 */
18672 		if (rold->off != rcur->off)
18673 			return false;
18674 		/* id relations must be preserved */
18675 		if (!check_ids(rold->id, rcur->id, idmap))
18676 			return false;
18677 		/* new val must satisfy old val knowledge */
18678 		return range_within(rold, rcur) &&
18679 		       tnum_in(rold->var_off, rcur->var_off);
18680 	case PTR_TO_STACK:
18681 		/* two stack pointers are equal only if they're pointing to
18682 		 * the same stack frame, since fp-8 in foo != fp-8 in bar
18683 		 */
18684 		return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
18685 	case PTR_TO_ARENA:
18686 		return true;
18687 	default:
18688 		return regs_exact(rold, rcur, idmap);
18689 	}
18690 }
18691 
18692 static struct bpf_reg_state unbound_reg;
18693 
unbound_reg_init(void)18694 static __init int unbound_reg_init(void)
18695 {
18696 	__mark_reg_unknown_imprecise(&unbound_reg);
18697 	unbound_reg.live |= REG_LIVE_READ;
18698 	return 0;
18699 }
18700 late_initcall(unbound_reg_init);
18701 
is_stack_all_misc(struct bpf_verifier_env * env,struct bpf_stack_state * stack)18702 static bool is_stack_all_misc(struct bpf_verifier_env *env,
18703 			      struct bpf_stack_state *stack)
18704 {
18705 	u32 i;
18706 
18707 	for (i = 0; i < ARRAY_SIZE(stack->slot_type); ++i) {
18708 		if ((stack->slot_type[i] == STACK_MISC) ||
18709 		    (stack->slot_type[i] == STACK_INVALID && env->allow_uninit_stack))
18710 			continue;
18711 		return false;
18712 	}
18713 
18714 	return true;
18715 }
18716 
scalar_reg_for_stack(struct bpf_verifier_env * env,struct bpf_stack_state * stack)18717 static struct bpf_reg_state *scalar_reg_for_stack(struct bpf_verifier_env *env,
18718 						  struct bpf_stack_state *stack)
18719 {
18720 	if (is_spilled_scalar_reg64(stack))
18721 		return &stack->spilled_ptr;
18722 
18723 	if (is_stack_all_misc(env, stack))
18724 		return &unbound_reg;
18725 
18726 	return NULL;
18727 }
18728 
stacksafe(struct bpf_verifier_env * env,struct bpf_func_state * old,struct bpf_func_state * cur,struct bpf_idmap * idmap,enum exact_level exact)18729 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
18730 		      struct bpf_func_state *cur, struct bpf_idmap *idmap,
18731 		      enum exact_level exact)
18732 {
18733 	int i, spi;
18734 
18735 	/* walk slots of the explored stack and ignore any additional
18736 	 * slots in the current stack, since explored(safe) state
18737 	 * didn't use them
18738 	 */
18739 	for (i = 0; i < old->allocated_stack; i++) {
18740 		struct bpf_reg_state *old_reg, *cur_reg;
18741 
18742 		spi = i / BPF_REG_SIZE;
18743 
18744 		if (exact != NOT_EXACT &&
18745 		    (i >= cur->allocated_stack ||
18746 		     old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
18747 		     cur->stack[spi].slot_type[i % BPF_REG_SIZE]))
18748 			return false;
18749 
18750 		if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)
18751 		    && exact == NOT_EXACT) {
18752 			i += BPF_REG_SIZE - 1;
18753 			/* explored state didn't use this */
18754 			continue;
18755 		}
18756 
18757 		if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
18758 			continue;
18759 
18760 		if (env->allow_uninit_stack &&
18761 		    old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
18762 			continue;
18763 
18764 		/* explored stack has more populated slots than current stack
18765 		 * and these slots were used
18766 		 */
18767 		if (i >= cur->allocated_stack)
18768 			return false;
18769 
18770 		/* 64-bit scalar spill vs all slots MISC and vice versa.
18771 		 * Load from all slots MISC produces unbound scalar.
18772 		 * Construct a fake register for such stack and call
18773 		 * regsafe() to ensure scalar ids are compared.
18774 		 */
18775 		old_reg = scalar_reg_for_stack(env, &old->stack[spi]);
18776 		cur_reg = scalar_reg_for_stack(env, &cur->stack[spi]);
18777 		if (old_reg && cur_reg) {
18778 			if (!regsafe(env, old_reg, cur_reg, idmap, exact))
18779 				return false;
18780 			i += BPF_REG_SIZE - 1;
18781 			continue;
18782 		}
18783 
18784 		/* if old state was safe with misc data in the stack
18785 		 * it will be safe with zero-initialized stack.
18786 		 * The opposite is not true
18787 		 */
18788 		if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
18789 		    cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
18790 			continue;
18791 		if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
18792 		    cur->stack[spi].slot_type[i % BPF_REG_SIZE])
18793 			/* Ex: old explored (safe) state has STACK_SPILL in
18794 			 * this stack slot, but current has STACK_MISC ->
18795 			 * this verifier states are not equivalent,
18796 			 * return false to continue verification of this path
18797 			 */
18798 			return false;
18799 		if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
18800 			continue;
18801 		/* Both old and cur are having same slot_type */
18802 		switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
18803 		case STACK_SPILL:
18804 			/* when explored and current stack slot are both storing
18805 			 * spilled registers, check that stored pointers types
18806 			 * are the same as well.
18807 			 * Ex: explored safe path could have stored
18808 			 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
18809 			 * but current path has stored:
18810 			 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
18811 			 * such verifier states are not equivalent.
18812 			 * return false to continue verification of this path
18813 			 */
18814 			if (!regsafe(env, &old->stack[spi].spilled_ptr,
18815 				     &cur->stack[spi].spilled_ptr, idmap, exact))
18816 				return false;
18817 			break;
18818 		case STACK_DYNPTR:
18819 			old_reg = &old->stack[spi].spilled_ptr;
18820 			cur_reg = &cur->stack[spi].spilled_ptr;
18821 			if (old_reg->dynptr.type != cur_reg->dynptr.type ||
18822 			    old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
18823 			    !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
18824 				return false;
18825 			break;
18826 		case STACK_ITER:
18827 			old_reg = &old->stack[spi].spilled_ptr;
18828 			cur_reg = &cur->stack[spi].spilled_ptr;
18829 			/* iter.depth is not compared between states as it
18830 			 * doesn't matter for correctness and would otherwise
18831 			 * prevent convergence; we maintain it only to prevent
18832 			 * infinite loop check triggering, see
18833 			 * iter_active_depths_differ()
18834 			 */
18835 			if (old_reg->iter.btf != cur_reg->iter.btf ||
18836 			    old_reg->iter.btf_id != cur_reg->iter.btf_id ||
18837 			    old_reg->iter.state != cur_reg->iter.state ||
18838 			    /* ignore {old_reg,cur_reg}->iter.depth, see above */
18839 			    !check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap))
18840 				return false;
18841 			break;
18842 		case STACK_IRQ_FLAG:
18843 			old_reg = &old->stack[spi].spilled_ptr;
18844 			cur_reg = &cur->stack[spi].spilled_ptr;
18845 			if (!check_ids(old_reg->ref_obj_id, cur_reg->ref_obj_id, idmap) ||
18846 			    old_reg->irq.kfunc_class != cur_reg->irq.kfunc_class)
18847 				return false;
18848 			break;
18849 		case STACK_MISC:
18850 		case STACK_ZERO:
18851 		case STACK_INVALID:
18852 			continue;
18853 		/* Ensure that new unhandled slot types return false by default */
18854 		default:
18855 			return false;
18856 		}
18857 	}
18858 	return true;
18859 }
18860 
refsafe(struct bpf_verifier_state * old,struct bpf_verifier_state * cur,struct bpf_idmap * idmap)18861 static bool refsafe(struct bpf_verifier_state *old, struct bpf_verifier_state *cur,
18862 		    struct bpf_idmap *idmap)
18863 {
18864 	int i;
18865 
18866 	if (old->acquired_refs != cur->acquired_refs)
18867 		return false;
18868 
18869 	if (old->active_locks != cur->active_locks)
18870 		return false;
18871 
18872 	if (old->active_preempt_locks != cur->active_preempt_locks)
18873 		return false;
18874 
18875 	if (old->active_rcu_lock != cur->active_rcu_lock)
18876 		return false;
18877 
18878 	if (!check_ids(old->active_irq_id, cur->active_irq_id, idmap))
18879 		return false;
18880 
18881 	if (!check_ids(old->active_lock_id, cur->active_lock_id, idmap) ||
18882 	    old->active_lock_ptr != cur->active_lock_ptr)
18883 		return false;
18884 
18885 	for (i = 0; i < old->acquired_refs; i++) {
18886 		if (!check_ids(old->refs[i].id, cur->refs[i].id, idmap) ||
18887 		    old->refs[i].type != cur->refs[i].type)
18888 			return false;
18889 		switch (old->refs[i].type) {
18890 		case REF_TYPE_PTR:
18891 		case REF_TYPE_IRQ:
18892 			break;
18893 		case REF_TYPE_LOCK:
18894 		case REF_TYPE_RES_LOCK:
18895 		case REF_TYPE_RES_LOCK_IRQ:
18896 			if (old->refs[i].ptr != cur->refs[i].ptr)
18897 				return false;
18898 			break;
18899 		default:
18900 			WARN_ONCE(1, "Unhandled enum type for reference state: %d\n", old->refs[i].type);
18901 			return false;
18902 		}
18903 	}
18904 
18905 	return true;
18906 }
18907 
18908 /* compare two verifier states
18909  *
18910  * all states stored in state_list are known to be valid, since
18911  * verifier reached 'bpf_exit' instruction through them
18912  *
18913  * this function is called when verifier exploring different branches of
18914  * execution popped from the state stack. If it sees an old state that has
18915  * more strict register state and more strict stack state then this execution
18916  * branch doesn't need to be explored further, since verifier already
18917  * concluded that more strict state leads to valid finish.
18918  *
18919  * Therefore two states are equivalent if register state is more conservative
18920  * and explored stack state is more conservative than the current one.
18921  * Example:
18922  *       explored                   current
18923  * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
18924  * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
18925  *
18926  * In other words if current stack state (one being explored) has more
18927  * valid slots than old one that already passed validation, it means
18928  * the verifier can stop exploring and conclude that current state is valid too
18929  *
18930  * Similarly with registers. If explored state has register type as invalid
18931  * whereas register type in current state is meaningful, it means that
18932  * the current state will reach 'bpf_exit' instruction safely
18933  */
func_states_equal(struct bpf_verifier_env * env,struct bpf_func_state * old,struct bpf_func_state * cur,u32 insn_idx,enum exact_level exact)18934 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
18935 			      struct bpf_func_state *cur, u32 insn_idx, enum exact_level exact)
18936 {
18937 	u16 live_regs = env->insn_aux_data[insn_idx].live_regs_before;
18938 	u16 i;
18939 
18940 	if (old->callback_depth > cur->callback_depth)
18941 		return false;
18942 
18943 	for (i = 0; i < MAX_BPF_REG; i++)
18944 		if (((1 << i) & live_regs) &&
18945 		    !regsafe(env, &old->regs[i], &cur->regs[i],
18946 			     &env->idmap_scratch, exact))
18947 			return false;
18948 
18949 	if (!stacksafe(env, old, cur, &env->idmap_scratch, exact))
18950 		return false;
18951 
18952 	return true;
18953 }
18954 
reset_idmap_scratch(struct bpf_verifier_env * env)18955 static void reset_idmap_scratch(struct bpf_verifier_env *env)
18956 {
18957 	env->idmap_scratch.tmp_id_gen = env->id_gen;
18958 	memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
18959 }
18960 
states_equal(struct bpf_verifier_env * env,struct bpf_verifier_state * old,struct bpf_verifier_state * cur,enum exact_level exact)18961 static bool states_equal(struct bpf_verifier_env *env,
18962 			 struct bpf_verifier_state *old,
18963 			 struct bpf_verifier_state *cur,
18964 			 enum exact_level exact)
18965 {
18966 	u32 insn_idx;
18967 	int i;
18968 
18969 	if (old->curframe != cur->curframe)
18970 		return false;
18971 
18972 	reset_idmap_scratch(env);
18973 
18974 	/* Verification state from speculative execution simulation
18975 	 * must never prune a non-speculative execution one.
18976 	 */
18977 	if (old->speculative && !cur->speculative)
18978 		return false;
18979 
18980 	if (old->in_sleepable != cur->in_sleepable)
18981 		return false;
18982 
18983 	if (!refsafe(old, cur, &env->idmap_scratch))
18984 		return false;
18985 
18986 	/* for states to be equal callsites have to be the same
18987 	 * and all frame states need to be equivalent
18988 	 */
18989 	for (i = 0; i <= old->curframe; i++) {
18990 		insn_idx = frame_insn_idx(old, i);
18991 		if (old->frame[i]->callsite != cur->frame[i]->callsite)
18992 			return false;
18993 		if (!func_states_equal(env, old->frame[i], cur->frame[i], insn_idx, exact))
18994 			return false;
18995 	}
18996 	return true;
18997 }
18998 
18999 /* Return 0 if no propagation happened. Return negative error code if error
19000  * happened. Otherwise, return the propagated bit.
19001  */
propagate_liveness_reg(struct bpf_verifier_env * env,struct bpf_reg_state * reg,struct bpf_reg_state * parent_reg)19002 static int propagate_liveness_reg(struct bpf_verifier_env *env,
19003 				  struct bpf_reg_state *reg,
19004 				  struct bpf_reg_state *parent_reg)
19005 {
19006 	u8 parent_flag = parent_reg->live & REG_LIVE_READ;
19007 	u8 flag = reg->live & REG_LIVE_READ;
19008 	int err;
19009 
19010 	/* When comes here, read flags of PARENT_REG or REG could be any of
19011 	 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
19012 	 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
19013 	 */
19014 	if (parent_flag == REG_LIVE_READ64 ||
19015 	    /* Or if there is no read flag from REG. */
19016 	    !flag ||
19017 	    /* Or if the read flag from REG is the same as PARENT_REG. */
19018 	    parent_flag == flag)
19019 		return 0;
19020 
19021 	err = mark_reg_read(env, reg, parent_reg, flag);
19022 	if (err)
19023 		return err;
19024 
19025 	return flag;
19026 }
19027 
19028 /* A write screens off any subsequent reads; but write marks come from the
19029  * straight-line code between a state and its parent.  When we arrive at an
19030  * equivalent state (jump target or such) we didn't arrive by the straight-line
19031  * code, so read marks in the state must propagate to the parent regardless
19032  * of the state's write marks. That's what 'parent == state->parent' comparison
19033  * in mark_reg_read() is for.
19034  */
propagate_liveness(struct bpf_verifier_env * env,const struct bpf_verifier_state * vstate,struct bpf_verifier_state * vparent,bool * changed)19035 static int propagate_liveness(struct bpf_verifier_env *env,
19036 			      const struct bpf_verifier_state *vstate,
19037 			      struct bpf_verifier_state *vparent,
19038 			      bool *changed)
19039 {
19040 	struct bpf_reg_state *state_reg, *parent_reg;
19041 	struct bpf_func_state *state, *parent;
19042 	int i, frame, err = 0;
19043 	bool tmp = false;
19044 
19045 	changed = changed ?: &tmp;
19046 	if (vparent->curframe != vstate->curframe) {
19047 		WARN(1, "propagate_live: parent frame %d current frame %d\n",
19048 		     vparent->curframe, vstate->curframe);
19049 		return -EFAULT;
19050 	}
19051 	/* Propagate read liveness of registers... */
19052 	BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
19053 	for (frame = 0; frame <= vstate->curframe; frame++) {
19054 		parent = vparent->frame[frame];
19055 		state = vstate->frame[frame];
19056 		parent_reg = parent->regs;
19057 		state_reg = state->regs;
19058 		/* We don't need to worry about FP liveness, it's read-only */
19059 		for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
19060 			err = propagate_liveness_reg(env, &state_reg[i],
19061 						     &parent_reg[i]);
19062 			if (err < 0)
19063 				return err;
19064 			*changed |= err > 0;
19065 			if (err == REG_LIVE_READ64)
19066 				mark_insn_zext(env, &parent_reg[i]);
19067 		}
19068 
19069 		/* Propagate stack slots. */
19070 		for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
19071 			    i < parent->allocated_stack / BPF_REG_SIZE; i++) {
19072 			parent_reg = &parent->stack[i].spilled_ptr;
19073 			state_reg = &state->stack[i].spilled_ptr;
19074 			err = propagate_liveness_reg(env, state_reg,
19075 						     parent_reg);
19076 			*changed |= err > 0;
19077 			if (err < 0)
19078 				return err;
19079 		}
19080 	}
19081 	return 0;
19082 }
19083 
19084 /* find precise scalars in the previous equivalent state and
19085  * propagate them into the current state
19086  */
propagate_precision(struct bpf_verifier_env * env,const struct bpf_verifier_state * old,struct bpf_verifier_state * cur,bool * changed)19087 static int propagate_precision(struct bpf_verifier_env *env,
19088 			       const struct bpf_verifier_state *old,
19089 			       struct bpf_verifier_state *cur,
19090 			       bool *changed)
19091 {
19092 	struct bpf_reg_state *state_reg;
19093 	struct bpf_func_state *state;
19094 	int i, err = 0, fr;
19095 	bool first;
19096 
19097 	for (fr = old->curframe; fr >= 0; fr--) {
19098 		state = old->frame[fr];
19099 		state_reg = state->regs;
19100 		first = true;
19101 		for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
19102 			if (state_reg->type != SCALAR_VALUE ||
19103 			    !state_reg->precise ||
19104 			    !(state_reg->live & REG_LIVE_READ))
19105 				continue;
19106 			if (env->log.level & BPF_LOG_LEVEL2) {
19107 				if (first)
19108 					verbose(env, "frame %d: propagating r%d", fr, i);
19109 				else
19110 					verbose(env, ",r%d", i);
19111 			}
19112 			bt_set_frame_reg(&env->bt, fr, i);
19113 			first = false;
19114 		}
19115 
19116 		for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
19117 			if (!is_spilled_reg(&state->stack[i]))
19118 				continue;
19119 			state_reg = &state->stack[i].spilled_ptr;
19120 			if (state_reg->type != SCALAR_VALUE ||
19121 			    !state_reg->precise ||
19122 			    !(state_reg->live & REG_LIVE_READ))
19123 				continue;
19124 			if (env->log.level & BPF_LOG_LEVEL2) {
19125 				if (first)
19126 					verbose(env, "frame %d: propagating fp%d",
19127 						fr, (-i - 1) * BPF_REG_SIZE);
19128 				else
19129 					verbose(env, ",fp%d", (-i - 1) * BPF_REG_SIZE);
19130 			}
19131 			bt_set_frame_slot(&env->bt, fr, i);
19132 			first = false;
19133 		}
19134 		if (!first)
19135 			verbose(env, "\n");
19136 	}
19137 
19138 	err = __mark_chain_precision(env, cur, -1, changed);
19139 	if (err < 0)
19140 		return err;
19141 
19142 	return 0;
19143 }
19144 
19145 #define MAX_BACKEDGE_ITERS 64
19146 
19147 /* Propagate read and precision marks from visit->backedges[*].state->equal_state
19148  * to corresponding parent states of visit->backedges[*].state until fixed point is reached,
19149  * then free visit->backedges.
19150  * After execution of this function incomplete_read_marks() will return false
19151  * for all states corresponding to @visit->callchain.
19152  */
propagate_backedges(struct bpf_verifier_env * env,struct bpf_scc_visit * visit)19153 static int propagate_backedges(struct bpf_verifier_env *env, struct bpf_scc_visit *visit)
19154 {
19155 	struct bpf_scc_backedge *backedge;
19156 	struct bpf_verifier_state *st;
19157 	bool changed;
19158 	int i, err;
19159 
19160 	i = 0;
19161 	do {
19162 		if (i++ > MAX_BACKEDGE_ITERS) {
19163 			if (env->log.level & BPF_LOG_LEVEL2)
19164 				verbose(env, "%s: too many iterations\n", __func__);
19165 			for (backedge = visit->backedges; backedge; backedge = backedge->next)
19166 				mark_all_scalars_precise(env, &backedge->state);
19167 			break;
19168 		}
19169 		changed = false;
19170 		for (backedge = visit->backedges; backedge; backedge = backedge->next) {
19171 			st = &backedge->state;
19172 			err = propagate_liveness(env, st->equal_state, st, &changed);
19173 			if (err)
19174 				return err;
19175 			err = propagate_precision(env, st->equal_state, st, &changed);
19176 			if (err)
19177 				return err;
19178 		}
19179 	} while (changed);
19180 
19181 	free_backedges(visit);
19182 	return 0;
19183 }
19184 
states_maybe_looping(struct bpf_verifier_state * old,struct bpf_verifier_state * cur)19185 static bool states_maybe_looping(struct bpf_verifier_state *old,
19186 				 struct bpf_verifier_state *cur)
19187 {
19188 	struct bpf_func_state *fold, *fcur;
19189 	int i, fr = cur->curframe;
19190 
19191 	if (old->curframe != fr)
19192 		return false;
19193 
19194 	fold = old->frame[fr];
19195 	fcur = cur->frame[fr];
19196 	for (i = 0; i < MAX_BPF_REG; i++)
19197 		if (memcmp(&fold->regs[i], &fcur->regs[i],
19198 			   offsetof(struct bpf_reg_state, parent)))
19199 			return false;
19200 	return true;
19201 }
19202 
is_iter_next_insn(struct bpf_verifier_env * env,int insn_idx)19203 static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
19204 {
19205 	return env->insn_aux_data[insn_idx].is_iter_next;
19206 }
19207 
19208 /* is_state_visited() handles iter_next() (see process_iter_next_call() for
19209  * terminology) calls specially: as opposed to bounded BPF loops, it *expects*
19210  * states to match, which otherwise would look like an infinite loop. So while
19211  * iter_next() calls are taken care of, we still need to be careful and
19212  * prevent erroneous and too eager declaration of "infinite loop", when
19213  * iterators are involved.
19214  *
19215  * Here's a situation in pseudo-BPF assembly form:
19216  *
19217  *   0: again:                          ; set up iter_next() call args
19218  *   1:   r1 = &it                      ; <CHECKPOINT HERE>
19219  *   2:   call bpf_iter_num_next        ; this is iter_next() call
19220  *   3:   if r0 == 0 goto done
19221  *   4:   ... something useful here ...
19222  *   5:   goto again                    ; another iteration
19223  *   6: done:
19224  *   7:   r1 = &it
19225  *   8:   call bpf_iter_num_destroy     ; clean up iter state
19226  *   9:   exit
19227  *
19228  * This is a typical loop. Let's assume that we have a prune point at 1:,
19229  * before we get to `call bpf_iter_num_next` (e.g., because of that `goto
19230  * again`, assuming other heuristics don't get in a way).
19231  *
19232  * When we first time come to 1:, let's say we have some state X. We proceed
19233  * to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
19234  * Now we come back to validate that forked ACTIVE state. We proceed through
19235  * 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
19236  * are converging. But the problem is that we don't know that yet, as this
19237  * convergence has to happen at iter_next() call site only. So if nothing is
19238  * done, at 1: verifier will use bounded loop logic and declare infinite
19239  * looping (and would be *technically* correct, if not for iterator's
19240  * "eventual sticky NULL" contract, see process_iter_next_call()). But we
19241  * don't want that. So what we do in process_iter_next_call() when we go on
19242  * another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
19243  * a different iteration. So when we suspect an infinite loop, we additionally
19244  * check if any of the *ACTIVE* iterator states depths differ. If yes, we
19245  * pretend we are not looping and wait for next iter_next() call.
19246  *
19247  * This only applies to ACTIVE state. In DRAINED state we don't expect to
19248  * loop, because that would actually mean infinite loop, as DRAINED state is
19249  * "sticky", and so we'll keep returning into the same instruction with the
19250  * same state (at least in one of possible code paths).
19251  *
19252  * This approach allows to keep infinite loop heuristic even in the face of
19253  * active iterator. E.g., C snippet below is and will be detected as
19254  * infinitely looping:
19255  *
19256  *   struct bpf_iter_num it;
19257  *   int *p, x;
19258  *
19259  *   bpf_iter_num_new(&it, 0, 10);
19260  *   while ((p = bpf_iter_num_next(&t))) {
19261  *       x = p;
19262  *       while (x--) {} // <<-- infinite loop here
19263  *   }
19264  *
19265  */
iter_active_depths_differ(struct bpf_verifier_state * old,struct bpf_verifier_state * cur)19266 static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
19267 {
19268 	struct bpf_reg_state *slot, *cur_slot;
19269 	struct bpf_func_state *state;
19270 	int i, fr;
19271 
19272 	for (fr = old->curframe; fr >= 0; fr--) {
19273 		state = old->frame[fr];
19274 		for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
19275 			if (state->stack[i].slot_type[0] != STACK_ITER)
19276 				continue;
19277 
19278 			slot = &state->stack[i].spilled_ptr;
19279 			if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
19280 				continue;
19281 
19282 			cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
19283 			if (cur_slot->iter.depth != slot->iter.depth)
19284 				return true;
19285 		}
19286 	}
19287 	return false;
19288 }
19289 
is_state_visited(struct bpf_verifier_env * env,int insn_idx)19290 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
19291 {
19292 	struct bpf_verifier_state_list *new_sl;
19293 	struct bpf_verifier_state_list *sl;
19294 	struct bpf_verifier_state *cur = env->cur_state, *new;
19295 	bool force_new_state, add_new_state, loop;
19296 	int i, j, n, err, states_cnt = 0;
19297 	struct list_head *pos, *tmp, *head;
19298 
19299 	force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx) ||
19300 			  /* Avoid accumulating infinitely long jmp history */
19301 			  cur->jmp_history_cnt > 40;
19302 
19303 	/* bpf progs typically have pruning point every 4 instructions
19304 	 * http://vger.kernel.org/bpfconf2019.html#session-1
19305 	 * Do not add new state for future pruning if the verifier hasn't seen
19306 	 * at least 2 jumps and at least 8 instructions.
19307 	 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
19308 	 * In tests that amounts to up to 50% reduction into total verifier
19309 	 * memory consumption and 20% verifier time speedup.
19310 	 */
19311 	add_new_state = force_new_state;
19312 	if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
19313 	    env->insn_processed - env->prev_insn_processed >= 8)
19314 		add_new_state = true;
19315 
19316 	clean_live_states(env, insn_idx, cur);
19317 
19318 	loop = false;
19319 	head = explored_state(env, insn_idx);
19320 	list_for_each_safe(pos, tmp, head) {
19321 		sl = container_of(pos, struct bpf_verifier_state_list, node);
19322 		states_cnt++;
19323 		if (sl->state.insn_idx != insn_idx)
19324 			continue;
19325 
19326 		if (sl->state.branches) {
19327 			struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
19328 
19329 			if (frame->in_async_callback_fn &&
19330 			    frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
19331 				/* Different async_entry_cnt means that the verifier is
19332 				 * processing another entry into async callback.
19333 				 * Seeing the same state is not an indication of infinite
19334 				 * loop or infinite recursion.
19335 				 * But finding the same state doesn't mean that it's safe
19336 				 * to stop processing the current state. The previous state
19337 				 * hasn't yet reached bpf_exit, since state.branches > 0.
19338 				 * Checking in_async_callback_fn alone is not enough either.
19339 				 * Since the verifier still needs to catch infinite loops
19340 				 * inside async callbacks.
19341 				 */
19342 				goto skip_inf_loop_check;
19343 			}
19344 			/* BPF open-coded iterators loop detection is special.
19345 			 * states_maybe_looping() logic is too simplistic in detecting
19346 			 * states that *might* be equivalent, because it doesn't know
19347 			 * about ID remapping, so don't even perform it.
19348 			 * See process_iter_next_call() and iter_active_depths_differ()
19349 			 * for overview of the logic. When current and one of parent
19350 			 * states are detected as equivalent, it's a good thing: we prove
19351 			 * convergence and can stop simulating further iterations.
19352 			 * It's safe to assume that iterator loop will finish, taking into
19353 			 * account iter_next() contract of eventually returning
19354 			 * sticky NULL result.
19355 			 *
19356 			 * Note, that states have to be compared exactly in this case because
19357 			 * read and precision marks might not be finalized inside the loop.
19358 			 * E.g. as in the program below:
19359 			 *
19360 			 *     1. r7 = -16
19361 			 *     2. r6 = bpf_get_prandom_u32()
19362 			 *     3. while (bpf_iter_num_next(&fp[-8])) {
19363 			 *     4.   if (r6 != 42) {
19364 			 *     5.     r7 = -32
19365 			 *     6.     r6 = bpf_get_prandom_u32()
19366 			 *     7.     continue
19367 			 *     8.   }
19368 			 *     9.   r0 = r10
19369 			 *    10.   r0 += r7
19370 			 *    11.   r8 = *(u64 *)(r0 + 0)
19371 			 *    12.   r6 = bpf_get_prandom_u32()
19372 			 *    13. }
19373 			 *
19374 			 * Here verifier would first visit path 1-3, create a checkpoint at 3
19375 			 * with r7=-16, continue to 4-7,3. Existing checkpoint at 3 does
19376 			 * not have read or precision mark for r7 yet, thus inexact states
19377 			 * comparison would discard current state with r7=-32
19378 			 * => unsafe memory access at 11 would not be caught.
19379 			 */
19380 			if (is_iter_next_insn(env, insn_idx)) {
19381 				if (states_equal(env, &sl->state, cur, RANGE_WITHIN)) {
19382 					struct bpf_func_state *cur_frame;
19383 					struct bpf_reg_state *iter_state, *iter_reg;
19384 					int spi;
19385 
19386 					cur_frame = cur->frame[cur->curframe];
19387 					/* btf_check_iter_kfuncs() enforces that
19388 					 * iter state pointer is always the first arg
19389 					 */
19390 					iter_reg = &cur_frame->regs[BPF_REG_1];
19391 					/* current state is valid due to states_equal(),
19392 					 * so we can assume valid iter and reg state,
19393 					 * no need for extra (re-)validations
19394 					 */
19395 					spi = __get_spi(iter_reg->off + iter_reg->var_off.value);
19396 					iter_state = &func(env, iter_reg)->stack[spi].spilled_ptr;
19397 					if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE) {
19398 						loop = true;
19399 						goto hit;
19400 					}
19401 				}
19402 				goto skip_inf_loop_check;
19403 			}
19404 			if (is_may_goto_insn_at(env, insn_idx)) {
19405 				if (sl->state.may_goto_depth != cur->may_goto_depth &&
19406 				    states_equal(env, &sl->state, cur, RANGE_WITHIN)) {
19407 					loop = true;
19408 					goto hit;
19409 				}
19410 			}
19411 			if (calls_callback(env, insn_idx)) {
19412 				if (states_equal(env, &sl->state, cur, RANGE_WITHIN))
19413 					goto hit;
19414 				goto skip_inf_loop_check;
19415 			}
19416 			/* attempt to detect infinite loop to avoid unnecessary doomed work */
19417 			if (states_maybe_looping(&sl->state, cur) &&
19418 			    states_equal(env, &sl->state, cur, EXACT) &&
19419 			    !iter_active_depths_differ(&sl->state, cur) &&
19420 			    sl->state.may_goto_depth == cur->may_goto_depth &&
19421 			    sl->state.callback_unroll_depth == cur->callback_unroll_depth) {
19422 				verbose_linfo(env, insn_idx, "; ");
19423 				verbose(env, "infinite loop detected at insn %d\n", insn_idx);
19424 				verbose(env, "cur state:");
19425 				print_verifier_state(env, cur, cur->curframe, true);
19426 				verbose(env, "old state:");
19427 				print_verifier_state(env, &sl->state, cur->curframe, true);
19428 				return -EINVAL;
19429 			}
19430 			/* if the verifier is processing a loop, avoid adding new state
19431 			 * too often, since different loop iterations have distinct
19432 			 * states and may not help future pruning.
19433 			 * This threshold shouldn't be too low to make sure that
19434 			 * a loop with large bound will be rejected quickly.
19435 			 * The most abusive loop will be:
19436 			 * r1 += 1
19437 			 * if r1 < 1000000 goto pc-2
19438 			 * 1M insn_procssed limit / 100 == 10k peak states.
19439 			 * This threshold shouldn't be too high either, since states
19440 			 * at the end of the loop are likely to be useful in pruning.
19441 			 */
19442 skip_inf_loop_check:
19443 			if (!force_new_state &&
19444 			    env->jmps_processed - env->prev_jmps_processed < 20 &&
19445 			    env->insn_processed - env->prev_insn_processed < 100)
19446 				add_new_state = false;
19447 			goto miss;
19448 		}
19449 		/* See comments for mark_all_regs_read_and_precise() */
19450 		loop = incomplete_read_marks(env, &sl->state);
19451 		if (states_equal(env, &sl->state, cur, loop ? RANGE_WITHIN : NOT_EXACT)) {
19452 hit:
19453 			sl->hit_cnt++;
19454 			/* reached equivalent register/stack state,
19455 			 * prune the search.
19456 			 * Registers read by the continuation are read by us.
19457 			 * If we have any write marks in env->cur_state, they
19458 			 * will prevent corresponding reads in the continuation
19459 			 * from reaching our parent (an explored_state).  Our
19460 			 * own state will get the read marks recorded, but
19461 			 * they'll be immediately forgotten as we're pruning
19462 			 * this state and will pop a new one.
19463 			 */
19464 			err = propagate_liveness(env, &sl->state, cur, NULL);
19465 
19466 			/* if previous state reached the exit with precision and
19467 			 * current state is equivalent to it (except precision marks)
19468 			 * the precision needs to be propagated back in
19469 			 * the current state.
19470 			 */
19471 			if (is_jmp_point(env, env->insn_idx))
19472 				err = err ? : push_jmp_history(env, cur, 0, 0);
19473 			err = err ? : propagate_precision(env, &sl->state, cur, NULL);
19474 			if (err)
19475 				return err;
19476 			/* When processing iterator based loops above propagate_liveness and
19477 			 * propagate_precision calls are not sufficient to transfer all relevant
19478 			 * read and precision marks. E.g. consider the following case:
19479 			 *
19480 			 *  .-> A --.  Assume the states are visited in the order A, B, C.
19481 			 *  |   |   |  Assume that state B reaches a state equivalent to state A.
19482 			 *  |   v   v  At this point, state C is not processed yet, so state A
19483 			 *  '-- B   C  has not received any read or precision marks from C.
19484 			 *             Thus, marks propagated from A to B are incomplete.
19485 			 *
19486 			 * The verifier mitigates this by performing the following steps:
19487 			 *
19488 			 * - Prior to the main verification pass, strongly connected components
19489 			 *   (SCCs) are computed over the program's control flow graph,
19490 			 *   intraprocedurally.
19491 			 *
19492 			 * - During the main verification pass, `maybe_enter_scc()` checks
19493 			 *   whether the current verifier state is entering an SCC. If so, an
19494 			 *   instance of a `bpf_scc_visit` object is created, and the state
19495 			 *   entering the SCC is recorded as the entry state.
19496 			 *
19497 			 * - This instance is associated not with the SCC itself, but with a
19498 			 *   `bpf_scc_callchain`: a tuple consisting of the call sites leading to
19499 			 *   the SCC and the SCC id. See `compute_scc_callchain()`.
19500 			 *
19501 			 * - When a verification path encounters a `states_equal(...,
19502 			 *   RANGE_WITHIN)` condition, there exists a call chain describing the
19503 			 *   current state and a corresponding `bpf_scc_visit` instance. A copy
19504 			 *   of the current state is created and added to
19505 			 *   `bpf_scc_visit->backedges`.
19506 			 *
19507 			 * - When a verification path terminates, `maybe_exit_scc()` is called
19508 			 *   from `update_branch_counts()`. For states with `branches == 0`, it
19509 			 *   checks whether the state is the entry state of any `bpf_scc_visit`
19510 			 *   instance. If it is, this indicates that all paths originating from
19511 			 *   this SCC visit have been explored. `propagate_backedges()` is then
19512 			 *   called, which propagates read and precision marks through the
19513 			 *   backedges until a fixed point is reached.
19514 			 *   (In the earlier example, this would propagate marks from A to B,
19515 			 *    from C to A, and then again from A to B.)
19516 			 *
19517 			 * A note on callchains
19518 			 * --------------------
19519 			 *
19520 			 * Consider the following example:
19521 			 *
19522 			 *     void foo() { loop { ... SCC#1 ... } }
19523 			 *     void main() {
19524 			 *       A: foo();
19525 			 *       B: ...
19526 			 *       C: foo();
19527 			 *     }
19528 			 *
19529 			 * Here, there are two distinct callchains leading to SCC#1:
19530 			 * - (A, SCC#1)
19531 			 * - (C, SCC#1)
19532 			 *
19533 			 * Each callchain identifies a separate `bpf_scc_visit` instance that
19534 			 * accumulates backedge states. The `propagate_{liveness,precision}()`
19535 			 * functions traverse the parent state of each backedge state, which
19536 			 * means these parent states must remain valid (i.e., not freed) while
19537 			 * the corresponding `bpf_scc_visit` instance exists.
19538 			 *
19539 			 * Associating `bpf_scc_visit` instances directly with SCCs instead of
19540 			 * callchains would break this invariant:
19541 			 * - States explored during `C: foo()` would contribute backedges to
19542 			 *   SCC#1, but SCC#1 would only be exited once the exploration of
19543 			 *   `A: foo()` completes.
19544 			 * - By that time, the states explored between `A: foo()` and `C: foo()`
19545 			 *   (i.e., `B: ...`) may have already been freed, causing the parent
19546 			 *   links for states from `C: foo()` to become invalid.
19547 			 */
19548 			if (loop) {
19549 				struct bpf_scc_backedge *backedge;
19550 
19551 				backedge = kzalloc(sizeof(*backedge), GFP_KERNEL_ACCOUNT);
19552 				if (!backedge)
19553 					return -ENOMEM;
19554 				err = copy_verifier_state(&backedge->state, cur);
19555 				backedge->state.equal_state = &sl->state;
19556 				backedge->state.insn_idx = insn_idx;
19557 				err = err ?: add_scc_backedge(env, &sl->state, backedge);
19558 				if (err) {
19559 					free_verifier_state(&backedge->state, false);
19560 					kvfree(backedge);
19561 					return err;
19562 				}
19563 			}
19564 			return 1;
19565 		}
19566 miss:
19567 		/* when new state is not going to be added do not increase miss count.
19568 		 * Otherwise several loop iterations will remove the state
19569 		 * recorded earlier. The goal of these heuristics is to have
19570 		 * states from some iterations of the loop (some in the beginning
19571 		 * and some at the end) to help pruning.
19572 		 */
19573 		if (add_new_state)
19574 			sl->miss_cnt++;
19575 		/* heuristic to determine whether this state is beneficial
19576 		 * to keep checking from state equivalence point of view.
19577 		 * Higher numbers increase max_states_per_insn and verification time,
19578 		 * but do not meaningfully decrease insn_processed.
19579 		 * 'n' controls how many times state could miss before eviction.
19580 		 * Use bigger 'n' for checkpoints because evicting checkpoint states
19581 		 * too early would hinder iterator convergence.
19582 		 */
19583 		n = is_force_checkpoint(env, insn_idx) && sl->state.branches > 0 ? 64 : 3;
19584 		if (sl->miss_cnt > sl->hit_cnt * n + n) {
19585 			/* the state is unlikely to be useful. Remove it to
19586 			 * speed up verification
19587 			 */
19588 			sl->in_free_list = true;
19589 			list_del(&sl->node);
19590 			list_add(&sl->node, &env->free_list);
19591 			env->free_list_size++;
19592 			env->explored_states_size--;
19593 			maybe_free_verifier_state(env, sl);
19594 		}
19595 	}
19596 
19597 	if (env->max_states_per_insn < states_cnt)
19598 		env->max_states_per_insn = states_cnt;
19599 
19600 	if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
19601 		return 0;
19602 
19603 	if (!add_new_state)
19604 		return 0;
19605 
19606 	/* There were no equivalent states, remember the current one.
19607 	 * Technically the current state is not proven to be safe yet,
19608 	 * but it will either reach outer most bpf_exit (which means it's safe)
19609 	 * or it will be rejected. When there are no loops the verifier won't be
19610 	 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
19611 	 * again on the way to bpf_exit.
19612 	 * When looping the sl->state.branches will be > 0 and this state
19613 	 * will not be considered for equivalence until branches == 0.
19614 	 */
19615 	new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL_ACCOUNT);
19616 	if (!new_sl)
19617 		return -ENOMEM;
19618 	env->total_states++;
19619 	env->explored_states_size++;
19620 	update_peak_states(env);
19621 	env->prev_jmps_processed = env->jmps_processed;
19622 	env->prev_insn_processed = env->insn_processed;
19623 
19624 	/* forget precise markings we inherited, see __mark_chain_precision */
19625 	if (env->bpf_capable)
19626 		mark_all_scalars_imprecise(env, cur);
19627 
19628 	/* add new state to the head of linked list */
19629 	new = &new_sl->state;
19630 	err = copy_verifier_state(new, cur);
19631 	if (err) {
19632 		free_verifier_state(new, false);
19633 		kfree(new_sl);
19634 		return err;
19635 	}
19636 	new->insn_idx = insn_idx;
19637 	verifier_bug_if(new->branches != 1, env,
19638 			"%s:branches_to_explore=%d insn %d",
19639 			__func__, new->branches, insn_idx);
19640 	err = maybe_enter_scc(env, new);
19641 	if (err) {
19642 		free_verifier_state(new, false);
19643 		kvfree(new_sl);
19644 		return err;
19645 	}
19646 
19647 	cur->parent = new;
19648 	cur->first_insn_idx = insn_idx;
19649 	cur->dfs_depth = new->dfs_depth + 1;
19650 	clear_jmp_history(cur);
19651 	list_add(&new_sl->node, head);
19652 
19653 	/* connect new state to parentage chain. Current frame needs all
19654 	 * registers connected. Only r6 - r9 of the callers are alive (pushed
19655 	 * to the stack implicitly by JITs) so in callers' frames connect just
19656 	 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
19657 	 * the state of the call instruction (with WRITTEN set), and r0 comes
19658 	 * from callee with its full parentage chain, anyway.
19659 	 */
19660 	/* clear write marks in current state: the writes we did are not writes
19661 	 * our child did, so they don't screen off its reads from us.
19662 	 * (There are no read marks in current state, because reads always mark
19663 	 * their parent and current state never has children yet.  Only
19664 	 * explored_states can get read marks.)
19665 	 */
19666 	for (j = 0; j <= cur->curframe; j++) {
19667 		for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
19668 			cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
19669 		for (i = 0; i < BPF_REG_FP; i++)
19670 			cur->frame[j]->regs[i].live = REG_LIVE_NONE;
19671 	}
19672 
19673 	/* all stack frames are accessible from callee, clear them all */
19674 	for (j = 0; j <= cur->curframe; j++) {
19675 		struct bpf_func_state *frame = cur->frame[j];
19676 		struct bpf_func_state *newframe = new->frame[j];
19677 
19678 		for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
19679 			frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
19680 			frame->stack[i].spilled_ptr.parent =
19681 						&newframe->stack[i].spilled_ptr;
19682 		}
19683 	}
19684 	return 0;
19685 }
19686 
19687 /* Return true if it's OK to have the same insn return a different type. */
reg_type_mismatch_ok(enum bpf_reg_type type)19688 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
19689 {
19690 	switch (base_type(type)) {
19691 	case PTR_TO_CTX:
19692 	case PTR_TO_SOCKET:
19693 	case PTR_TO_SOCK_COMMON:
19694 	case PTR_TO_TCP_SOCK:
19695 	case PTR_TO_XDP_SOCK:
19696 	case PTR_TO_BTF_ID:
19697 	case PTR_TO_ARENA:
19698 		return false;
19699 	default:
19700 		return true;
19701 	}
19702 }
19703 
19704 /* If an instruction was previously used with particular pointer types, then we
19705  * need to be careful to avoid cases such as the below, where it may be ok
19706  * for one branch accessing the pointer, but not ok for the other branch:
19707  *
19708  * R1 = sock_ptr
19709  * goto X;
19710  * ...
19711  * R1 = some_other_valid_ptr;
19712  * goto X;
19713  * ...
19714  * R2 = *(u32 *)(R1 + 0);
19715  */
reg_type_mismatch(enum bpf_reg_type src,enum bpf_reg_type prev)19716 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
19717 {
19718 	return src != prev && (!reg_type_mismatch_ok(src) ||
19719 			       !reg_type_mismatch_ok(prev));
19720 }
19721 
is_ptr_to_mem_or_btf_id(enum bpf_reg_type type)19722 static bool is_ptr_to_mem_or_btf_id(enum bpf_reg_type type)
19723 {
19724 	switch (base_type(type)) {
19725 	case PTR_TO_MEM:
19726 	case PTR_TO_BTF_ID:
19727 		return true;
19728 	default:
19729 		return false;
19730 	}
19731 }
19732 
is_ptr_to_mem(enum bpf_reg_type type)19733 static bool is_ptr_to_mem(enum bpf_reg_type type)
19734 {
19735 	return base_type(type) == PTR_TO_MEM;
19736 }
19737 
save_aux_ptr_type(struct bpf_verifier_env * env,enum bpf_reg_type type,bool allow_trust_mismatch)19738 static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
19739 			     bool allow_trust_mismatch)
19740 {
19741 	enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
19742 	enum bpf_reg_type merged_type;
19743 
19744 	if (*prev_type == NOT_INIT) {
19745 		/* Saw a valid insn
19746 		 * dst_reg = *(u32 *)(src_reg + off)
19747 		 * save type to validate intersecting paths
19748 		 */
19749 		*prev_type = type;
19750 	} else if (reg_type_mismatch(type, *prev_type)) {
19751 		/* Abuser program is trying to use the same insn
19752 		 * dst_reg = *(u32*) (src_reg + off)
19753 		 * with different pointer types:
19754 		 * src_reg == ctx in one branch and
19755 		 * src_reg == stack|map in some other branch.
19756 		 * Reject it.
19757 		 */
19758 		if (allow_trust_mismatch &&
19759 		    is_ptr_to_mem_or_btf_id(type) &&
19760 		    is_ptr_to_mem_or_btf_id(*prev_type)) {
19761 			/*
19762 			 * Have to support a use case when one path through
19763 			 * the program yields TRUSTED pointer while another
19764 			 * is UNTRUSTED. Fallback to UNTRUSTED to generate
19765 			 * BPF_PROBE_MEM/BPF_PROBE_MEMSX.
19766 			 * Same behavior of MEM_RDONLY flag.
19767 			 */
19768 			if (is_ptr_to_mem(type) || is_ptr_to_mem(*prev_type))
19769 				merged_type = PTR_TO_MEM;
19770 			else
19771 				merged_type = PTR_TO_BTF_ID;
19772 			if ((type & PTR_UNTRUSTED) || (*prev_type & PTR_UNTRUSTED))
19773 				merged_type |= PTR_UNTRUSTED;
19774 			if ((type & MEM_RDONLY) || (*prev_type & MEM_RDONLY))
19775 				merged_type |= MEM_RDONLY;
19776 			*prev_type = merged_type;
19777 		} else {
19778 			verbose(env, "same insn cannot be used with different pointers\n");
19779 			return -EINVAL;
19780 		}
19781 	}
19782 
19783 	return 0;
19784 }
19785 
19786 enum {
19787 	PROCESS_BPF_EXIT = 1
19788 };
19789 
process_bpf_exit_full(struct bpf_verifier_env * env,bool * do_print_state,bool exception_exit)19790 static int process_bpf_exit_full(struct bpf_verifier_env *env,
19791 				 bool *do_print_state,
19792 				 bool exception_exit)
19793 {
19794 	/* We must do check_reference_leak here before
19795 	 * prepare_func_exit to handle the case when
19796 	 * state->curframe > 0, it may be a callback function,
19797 	 * for which reference_state must match caller reference
19798 	 * state when it exits.
19799 	 */
19800 	int err = check_resource_leak(env, exception_exit,
19801 				      !env->cur_state->curframe,
19802 				      "BPF_EXIT instruction in main prog");
19803 	if (err)
19804 		return err;
19805 
19806 	/* The side effect of the prepare_func_exit which is
19807 	 * being skipped is that it frees bpf_func_state.
19808 	 * Typically, process_bpf_exit will only be hit with
19809 	 * outermost exit. copy_verifier_state in pop_stack will
19810 	 * handle freeing of any extra bpf_func_state left over
19811 	 * from not processing all nested function exits. We
19812 	 * also skip return code checks as they are not needed
19813 	 * for exceptional exits.
19814 	 */
19815 	if (exception_exit)
19816 		return PROCESS_BPF_EXIT;
19817 
19818 	if (env->cur_state->curframe) {
19819 		/* exit from nested function */
19820 		err = prepare_func_exit(env, &env->insn_idx);
19821 		if (err)
19822 			return err;
19823 		*do_print_state = true;
19824 		return 0;
19825 	}
19826 
19827 	err = check_return_code(env, BPF_REG_0, "R0");
19828 	if (err)
19829 		return err;
19830 	return PROCESS_BPF_EXIT;
19831 }
19832 
do_check_insn(struct bpf_verifier_env * env,bool * do_print_state)19833 static int do_check_insn(struct bpf_verifier_env *env, bool *do_print_state)
19834 {
19835 	int err;
19836 	struct bpf_insn *insn = &env->prog->insnsi[env->insn_idx];
19837 	u8 class = BPF_CLASS(insn->code);
19838 
19839 	if (class == BPF_ALU || class == BPF_ALU64) {
19840 		err = check_alu_op(env, insn);
19841 		if (err)
19842 			return err;
19843 
19844 	} else if (class == BPF_LDX) {
19845 		bool is_ldsx = BPF_MODE(insn->code) == BPF_MEMSX;
19846 
19847 		/* Check for reserved fields is already done in
19848 		 * resolve_pseudo_ldimm64().
19849 		 */
19850 		err = check_load_mem(env, insn, false, is_ldsx, true, "ldx");
19851 		if (err)
19852 			return err;
19853 	} else if (class == BPF_STX) {
19854 		if (BPF_MODE(insn->code) == BPF_ATOMIC) {
19855 			err = check_atomic(env, insn);
19856 			if (err)
19857 				return err;
19858 			env->insn_idx++;
19859 			return 0;
19860 		}
19861 
19862 		if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
19863 			verbose(env, "BPF_STX uses reserved fields\n");
19864 			return -EINVAL;
19865 		}
19866 
19867 		err = check_store_reg(env, insn, false);
19868 		if (err)
19869 			return err;
19870 	} else if (class == BPF_ST) {
19871 		enum bpf_reg_type dst_reg_type;
19872 
19873 		if (BPF_MODE(insn->code) != BPF_MEM ||
19874 		    insn->src_reg != BPF_REG_0) {
19875 			verbose(env, "BPF_ST uses reserved fields\n");
19876 			return -EINVAL;
19877 		}
19878 		/* check src operand */
19879 		err = check_reg_arg(env, insn->dst_reg, SRC_OP);
19880 		if (err)
19881 			return err;
19882 
19883 		dst_reg_type = cur_regs(env)[insn->dst_reg].type;
19884 
19885 		/* check that memory (dst_reg + off) is writeable */
19886 		err = check_mem_access(env, env->insn_idx, insn->dst_reg,
19887 				       insn->off, BPF_SIZE(insn->code),
19888 				       BPF_WRITE, -1, false, false);
19889 		if (err)
19890 			return err;
19891 
19892 		err = save_aux_ptr_type(env, dst_reg_type, false);
19893 		if (err)
19894 			return err;
19895 	} else if (class == BPF_JMP || class == BPF_JMP32) {
19896 		u8 opcode = BPF_OP(insn->code);
19897 
19898 		env->jmps_processed++;
19899 		if (opcode == BPF_CALL) {
19900 			if (BPF_SRC(insn->code) != BPF_K ||
19901 			    (insn->src_reg != BPF_PSEUDO_KFUNC_CALL &&
19902 			     insn->off != 0) ||
19903 			    (insn->src_reg != BPF_REG_0 &&
19904 			     insn->src_reg != BPF_PSEUDO_CALL &&
19905 			     insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
19906 			    insn->dst_reg != BPF_REG_0 || class == BPF_JMP32) {
19907 				verbose(env, "BPF_CALL uses reserved fields\n");
19908 				return -EINVAL;
19909 			}
19910 
19911 			if (env->cur_state->active_locks) {
19912 				if ((insn->src_reg == BPF_REG_0 &&
19913 				     insn->imm != BPF_FUNC_spin_unlock) ||
19914 				    (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
19915 				     (insn->off != 0 || !kfunc_spin_allowed(insn->imm)))) {
19916 					verbose(env,
19917 						"function calls are not allowed while holding a lock\n");
19918 					return -EINVAL;
19919 				}
19920 			}
19921 			if (insn->src_reg == BPF_PSEUDO_CALL) {
19922 				err = check_func_call(env, insn, &env->insn_idx);
19923 			} else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
19924 				err = check_kfunc_call(env, insn, &env->insn_idx);
19925 				if (!err && is_bpf_throw_kfunc(insn))
19926 					return process_bpf_exit_full(env, do_print_state, true);
19927 			} else {
19928 				err = check_helper_call(env, insn, &env->insn_idx);
19929 			}
19930 			if (err)
19931 				return err;
19932 
19933 			mark_reg_scratched(env, BPF_REG_0);
19934 		} else if (opcode == BPF_JA) {
19935 			if (BPF_SRC(insn->code) != BPF_K ||
19936 			    insn->src_reg != BPF_REG_0 ||
19937 			    insn->dst_reg != BPF_REG_0 ||
19938 			    (class == BPF_JMP && insn->imm != 0) ||
19939 			    (class == BPF_JMP32 && insn->off != 0)) {
19940 				verbose(env, "BPF_JA uses reserved fields\n");
19941 				return -EINVAL;
19942 			}
19943 
19944 			if (class == BPF_JMP)
19945 				env->insn_idx += insn->off + 1;
19946 			else
19947 				env->insn_idx += insn->imm + 1;
19948 			return 0;
19949 		} else if (opcode == BPF_EXIT) {
19950 			if (BPF_SRC(insn->code) != BPF_K ||
19951 			    insn->imm != 0 ||
19952 			    insn->src_reg != BPF_REG_0 ||
19953 			    insn->dst_reg != BPF_REG_0 ||
19954 			    class == BPF_JMP32) {
19955 				verbose(env, "BPF_EXIT uses reserved fields\n");
19956 				return -EINVAL;
19957 			}
19958 			return process_bpf_exit_full(env, do_print_state, false);
19959 		} else {
19960 			err = check_cond_jmp_op(env, insn, &env->insn_idx);
19961 			if (err)
19962 				return err;
19963 		}
19964 	} else if (class == BPF_LD) {
19965 		u8 mode = BPF_MODE(insn->code);
19966 
19967 		if (mode == BPF_ABS || mode == BPF_IND) {
19968 			err = check_ld_abs(env, insn);
19969 			if (err)
19970 				return err;
19971 
19972 		} else if (mode == BPF_IMM) {
19973 			err = check_ld_imm(env, insn);
19974 			if (err)
19975 				return err;
19976 
19977 			env->insn_idx++;
19978 			sanitize_mark_insn_seen(env);
19979 		} else {
19980 			verbose(env, "invalid BPF_LD mode\n");
19981 			return -EINVAL;
19982 		}
19983 	} else {
19984 		verbose(env, "unknown insn class %d\n", class);
19985 		return -EINVAL;
19986 	}
19987 
19988 	env->insn_idx++;
19989 	return 0;
19990 }
19991 
do_check(struct bpf_verifier_env * env)19992 static int do_check(struct bpf_verifier_env *env)
19993 {
19994 	bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
19995 	struct bpf_verifier_state *state = env->cur_state;
19996 	struct bpf_insn *insns = env->prog->insnsi;
19997 	int insn_cnt = env->prog->len;
19998 	bool do_print_state = false;
19999 	int prev_insn_idx = -1;
20000 
20001 	for (;;) {
20002 		struct bpf_insn *insn;
20003 		struct bpf_insn_aux_data *insn_aux;
20004 		int err;
20005 
20006 		/* reset current history entry on each new instruction */
20007 		env->cur_hist_ent = NULL;
20008 
20009 		env->prev_insn_idx = prev_insn_idx;
20010 		if (env->insn_idx >= insn_cnt) {
20011 			verbose(env, "invalid insn idx %d insn_cnt %d\n",
20012 				env->insn_idx, insn_cnt);
20013 			return -EFAULT;
20014 		}
20015 
20016 		insn = &insns[env->insn_idx];
20017 		insn_aux = &env->insn_aux_data[env->insn_idx];
20018 
20019 		if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
20020 			verbose(env,
20021 				"BPF program is too large. Processed %d insn\n",
20022 				env->insn_processed);
20023 			return -E2BIG;
20024 		}
20025 
20026 		state->last_insn_idx = env->prev_insn_idx;
20027 		state->insn_idx = env->insn_idx;
20028 
20029 		if (is_prune_point(env, env->insn_idx)) {
20030 			err = is_state_visited(env, env->insn_idx);
20031 			if (err < 0)
20032 				return err;
20033 			if (err == 1) {
20034 				/* found equivalent state, can prune the search */
20035 				if (env->log.level & BPF_LOG_LEVEL) {
20036 					if (do_print_state)
20037 						verbose(env, "\nfrom %d to %d%s: safe\n",
20038 							env->prev_insn_idx, env->insn_idx,
20039 							env->cur_state->speculative ?
20040 							" (speculative execution)" : "");
20041 					else
20042 						verbose(env, "%d: safe\n", env->insn_idx);
20043 				}
20044 				goto process_bpf_exit;
20045 			}
20046 		}
20047 
20048 		if (is_jmp_point(env, env->insn_idx)) {
20049 			err = push_jmp_history(env, state, 0, 0);
20050 			if (err)
20051 				return err;
20052 		}
20053 
20054 		if (signal_pending(current))
20055 			return -EAGAIN;
20056 
20057 		if (need_resched())
20058 			cond_resched();
20059 
20060 		if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
20061 			verbose(env, "\nfrom %d to %d%s:",
20062 				env->prev_insn_idx, env->insn_idx,
20063 				env->cur_state->speculative ?
20064 				" (speculative execution)" : "");
20065 			print_verifier_state(env, state, state->curframe, true);
20066 			do_print_state = false;
20067 		}
20068 
20069 		if (env->log.level & BPF_LOG_LEVEL) {
20070 			if (verifier_state_scratched(env))
20071 				print_insn_state(env, state, state->curframe);
20072 
20073 			verbose_linfo(env, env->insn_idx, "; ");
20074 			env->prev_log_pos = env->log.end_pos;
20075 			verbose(env, "%d: ", env->insn_idx);
20076 			verbose_insn(env, insn);
20077 			env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
20078 			env->prev_log_pos = env->log.end_pos;
20079 		}
20080 
20081 		if (bpf_prog_is_offloaded(env->prog->aux)) {
20082 			err = bpf_prog_offload_verify_insn(env, env->insn_idx,
20083 							   env->prev_insn_idx);
20084 			if (err)
20085 				return err;
20086 		}
20087 
20088 		sanitize_mark_insn_seen(env);
20089 		prev_insn_idx = env->insn_idx;
20090 
20091 		/* Reduce verification complexity by stopping speculative path
20092 		 * verification when a nospec is encountered.
20093 		 */
20094 		if (state->speculative && insn_aux->nospec)
20095 			goto process_bpf_exit;
20096 
20097 		err = do_check_insn(env, &do_print_state);
20098 		if (error_recoverable_with_nospec(err) && state->speculative) {
20099 			/* Prevent this speculative path from ever reaching the
20100 			 * insn that would have been unsafe to execute.
20101 			 */
20102 			insn_aux->nospec = true;
20103 			/* If it was an ADD/SUB insn, potentially remove any
20104 			 * markings for alu sanitization.
20105 			 */
20106 			insn_aux->alu_state = 0;
20107 			goto process_bpf_exit;
20108 		} else if (err < 0) {
20109 			return err;
20110 		} else if (err == PROCESS_BPF_EXIT) {
20111 			goto process_bpf_exit;
20112 		}
20113 		WARN_ON_ONCE(err);
20114 
20115 		if (state->speculative && insn_aux->nospec_result) {
20116 			/* If we are on a path that performed a jump-op, this
20117 			 * may skip a nospec patched-in after the jump. This can
20118 			 * currently never happen because nospec_result is only
20119 			 * used for the write-ops
20120 			 * `*(size*)(dst_reg+off)=src_reg|imm32` which must
20121 			 * never skip the following insn. Still, add a warning
20122 			 * to document this in case nospec_result is used
20123 			 * elsewhere in the future.
20124 			 *
20125 			 * All non-branch instructions have a single
20126 			 * fall-through edge. For these, nospec_result should
20127 			 * already work.
20128 			 */
20129 			if (verifier_bug_if(BPF_CLASS(insn->code) == BPF_JMP ||
20130 					    BPF_CLASS(insn->code) == BPF_JMP32, env,
20131 					    "speculation barrier after jump instruction may not have the desired effect"))
20132 				return -EFAULT;
20133 process_bpf_exit:
20134 			mark_verifier_state_scratched(env);
20135 			err = update_branch_counts(env, env->cur_state);
20136 			if (err)
20137 				return err;
20138 			err = pop_stack(env, &prev_insn_idx, &env->insn_idx,
20139 					pop_log);
20140 			if (err < 0) {
20141 				if (err != -ENOENT)
20142 					return err;
20143 				break;
20144 			} else {
20145 				do_print_state = true;
20146 				continue;
20147 			}
20148 		}
20149 	}
20150 
20151 	return 0;
20152 }
20153 
find_btf_percpu_datasec(struct btf * btf)20154 static int find_btf_percpu_datasec(struct btf *btf)
20155 {
20156 	const struct btf_type *t;
20157 	const char *tname;
20158 	int i, n;
20159 
20160 	/*
20161 	 * Both vmlinux and module each have their own ".data..percpu"
20162 	 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
20163 	 * types to look at only module's own BTF types.
20164 	 */
20165 	n = btf_nr_types(btf);
20166 	if (btf_is_module(btf))
20167 		i = btf_nr_types(btf_vmlinux);
20168 	else
20169 		i = 1;
20170 
20171 	for(; i < n; i++) {
20172 		t = btf_type_by_id(btf, i);
20173 		if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
20174 			continue;
20175 
20176 		tname = btf_name_by_offset(btf, t->name_off);
20177 		if (!strcmp(tname, ".data..percpu"))
20178 			return i;
20179 	}
20180 
20181 	return -ENOENT;
20182 }
20183 
20184 /*
20185  * Add btf to the used_btfs array and return the index. (If the btf was
20186  * already added, then just return the index.) Upon successful insertion
20187  * increase btf refcnt, and, if present, also refcount the corresponding
20188  * kernel module.
20189  */
__add_used_btf(struct bpf_verifier_env * env,struct btf * btf)20190 static int __add_used_btf(struct bpf_verifier_env *env, struct btf *btf)
20191 {
20192 	struct btf_mod_pair *btf_mod;
20193 	int i;
20194 
20195 	/* check whether we recorded this BTF (and maybe module) already */
20196 	for (i = 0; i < env->used_btf_cnt; i++)
20197 		if (env->used_btfs[i].btf == btf)
20198 			return i;
20199 
20200 	if (env->used_btf_cnt >= MAX_USED_BTFS)
20201 		return -E2BIG;
20202 
20203 	btf_get(btf);
20204 
20205 	btf_mod = &env->used_btfs[env->used_btf_cnt];
20206 	btf_mod->btf = btf;
20207 	btf_mod->module = NULL;
20208 
20209 	/* if we reference variables from kernel module, bump its refcount */
20210 	if (btf_is_module(btf)) {
20211 		btf_mod->module = btf_try_get_module(btf);
20212 		if (!btf_mod->module) {
20213 			btf_put(btf);
20214 			return -ENXIO;
20215 		}
20216 	}
20217 
20218 	return env->used_btf_cnt++;
20219 }
20220 
20221 /* replace pseudo btf_id with kernel symbol address */
__check_pseudo_btf_id(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_insn_aux_data * aux,struct btf * btf)20222 static int __check_pseudo_btf_id(struct bpf_verifier_env *env,
20223 				 struct bpf_insn *insn,
20224 				 struct bpf_insn_aux_data *aux,
20225 				 struct btf *btf)
20226 {
20227 	const struct btf_var_secinfo *vsi;
20228 	const struct btf_type *datasec;
20229 	const struct btf_type *t;
20230 	const char *sym_name;
20231 	bool percpu = false;
20232 	u32 type, id = insn->imm;
20233 	s32 datasec_id;
20234 	u64 addr;
20235 	int i;
20236 
20237 	t = btf_type_by_id(btf, id);
20238 	if (!t) {
20239 		verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
20240 		return -ENOENT;
20241 	}
20242 
20243 	if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
20244 		verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
20245 		return -EINVAL;
20246 	}
20247 
20248 	sym_name = btf_name_by_offset(btf, t->name_off);
20249 	addr = kallsyms_lookup_name(sym_name);
20250 	if (!addr) {
20251 		verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
20252 			sym_name);
20253 		return -ENOENT;
20254 	}
20255 	insn[0].imm = (u32)addr;
20256 	insn[1].imm = addr >> 32;
20257 
20258 	if (btf_type_is_func(t)) {
20259 		aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
20260 		aux->btf_var.mem_size = 0;
20261 		return 0;
20262 	}
20263 
20264 	datasec_id = find_btf_percpu_datasec(btf);
20265 	if (datasec_id > 0) {
20266 		datasec = btf_type_by_id(btf, datasec_id);
20267 		for_each_vsi(i, datasec, vsi) {
20268 			if (vsi->type == id) {
20269 				percpu = true;
20270 				break;
20271 			}
20272 		}
20273 	}
20274 
20275 	type = t->type;
20276 	t = btf_type_skip_modifiers(btf, type, NULL);
20277 	if (percpu) {
20278 		aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
20279 		aux->btf_var.btf = btf;
20280 		aux->btf_var.btf_id = type;
20281 	} else if (!btf_type_is_struct(t)) {
20282 		const struct btf_type *ret;
20283 		const char *tname;
20284 		u32 tsize;
20285 
20286 		/* resolve the type size of ksym. */
20287 		ret = btf_resolve_size(btf, t, &tsize);
20288 		if (IS_ERR(ret)) {
20289 			tname = btf_name_by_offset(btf, t->name_off);
20290 			verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
20291 				tname, PTR_ERR(ret));
20292 			return -EINVAL;
20293 		}
20294 		aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
20295 		aux->btf_var.mem_size = tsize;
20296 	} else {
20297 		aux->btf_var.reg_type = PTR_TO_BTF_ID;
20298 		aux->btf_var.btf = btf;
20299 		aux->btf_var.btf_id = type;
20300 	}
20301 
20302 	return 0;
20303 }
20304 
check_pseudo_btf_id(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_insn_aux_data * aux)20305 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
20306 			       struct bpf_insn *insn,
20307 			       struct bpf_insn_aux_data *aux)
20308 {
20309 	struct btf *btf;
20310 	int btf_fd;
20311 	int err;
20312 
20313 	btf_fd = insn[1].imm;
20314 	if (btf_fd) {
20315 		CLASS(fd, f)(btf_fd);
20316 
20317 		btf = __btf_get_by_fd(f);
20318 		if (IS_ERR(btf)) {
20319 			verbose(env, "invalid module BTF object FD specified.\n");
20320 			return -EINVAL;
20321 		}
20322 	} else {
20323 		if (!btf_vmlinux) {
20324 			verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
20325 			return -EINVAL;
20326 		}
20327 		btf = btf_vmlinux;
20328 	}
20329 
20330 	err = __check_pseudo_btf_id(env, insn, aux, btf);
20331 	if (err)
20332 		return err;
20333 
20334 	err = __add_used_btf(env, btf);
20335 	if (err < 0)
20336 		return err;
20337 	return 0;
20338 }
20339 
is_tracing_prog_type(enum bpf_prog_type type)20340 static bool is_tracing_prog_type(enum bpf_prog_type type)
20341 {
20342 	switch (type) {
20343 	case BPF_PROG_TYPE_KPROBE:
20344 	case BPF_PROG_TYPE_TRACEPOINT:
20345 	case BPF_PROG_TYPE_PERF_EVENT:
20346 	case BPF_PROG_TYPE_RAW_TRACEPOINT:
20347 	case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
20348 		return true;
20349 	default:
20350 		return false;
20351 	}
20352 }
20353 
bpf_map_is_cgroup_storage(struct bpf_map * map)20354 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
20355 {
20356 	return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
20357 		map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
20358 }
20359 
check_map_prog_compatibility(struct bpf_verifier_env * env,struct bpf_map * map,struct bpf_prog * prog)20360 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
20361 					struct bpf_map *map,
20362 					struct bpf_prog *prog)
20363 
20364 {
20365 	enum bpf_prog_type prog_type = resolve_prog_type(prog);
20366 
20367 	if (btf_record_has_field(map->record, BPF_LIST_HEAD) ||
20368 	    btf_record_has_field(map->record, BPF_RB_ROOT)) {
20369 		if (is_tracing_prog_type(prog_type)) {
20370 			verbose(env, "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
20371 			return -EINVAL;
20372 		}
20373 	}
20374 
20375 	if (btf_record_has_field(map->record, BPF_SPIN_LOCK | BPF_RES_SPIN_LOCK)) {
20376 		if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
20377 			verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
20378 			return -EINVAL;
20379 		}
20380 
20381 		if (is_tracing_prog_type(prog_type)) {
20382 			verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
20383 			return -EINVAL;
20384 		}
20385 	}
20386 
20387 	if (btf_record_has_field(map->record, BPF_TIMER)) {
20388 		if (is_tracing_prog_type(prog_type)) {
20389 			verbose(env, "tracing progs cannot use bpf_timer yet\n");
20390 			return -EINVAL;
20391 		}
20392 	}
20393 
20394 	if (btf_record_has_field(map->record, BPF_WORKQUEUE)) {
20395 		if (is_tracing_prog_type(prog_type)) {
20396 			verbose(env, "tracing progs cannot use bpf_wq yet\n");
20397 			return -EINVAL;
20398 		}
20399 	}
20400 
20401 	if ((bpf_prog_is_offloaded(prog->aux) || bpf_map_is_offloaded(map)) &&
20402 	    !bpf_offload_prog_map_match(prog, map)) {
20403 		verbose(env, "offload device mismatch between prog and map\n");
20404 		return -EINVAL;
20405 	}
20406 
20407 	if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
20408 		verbose(env, "bpf_struct_ops map cannot be used in prog\n");
20409 		return -EINVAL;
20410 	}
20411 
20412 	if (prog->sleepable)
20413 		switch (map->map_type) {
20414 		case BPF_MAP_TYPE_HASH:
20415 		case BPF_MAP_TYPE_LRU_HASH:
20416 		case BPF_MAP_TYPE_ARRAY:
20417 		case BPF_MAP_TYPE_PERCPU_HASH:
20418 		case BPF_MAP_TYPE_PERCPU_ARRAY:
20419 		case BPF_MAP_TYPE_LRU_PERCPU_HASH:
20420 		case BPF_MAP_TYPE_ARRAY_OF_MAPS:
20421 		case BPF_MAP_TYPE_HASH_OF_MAPS:
20422 		case BPF_MAP_TYPE_RINGBUF:
20423 		case BPF_MAP_TYPE_USER_RINGBUF:
20424 		case BPF_MAP_TYPE_INODE_STORAGE:
20425 		case BPF_MAP_TYPE_SK_STORAGE:
20426 		case BPF_MAP_TYPE_TASK_STORAGE:
20427 		case BPF_MAP_TYPE_CGRP_STORAGE:
20428 		case BPF_MAP_TYPE_QUEUE:
20429 		case BPF_MAP_TYPE_STACK:
20430 		case BPF_MAP_TYPE_ARENA:
20431 			break;
20432 		default:
20433 			verbose(env,
20434 				"Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
20435 			return -EINVAL;
20436 		}
20437 
20438 	if (bpf_map_is_cgroup_storage(map) &&
20439 	    bpf_cgroup_storage_assign(env->prog->aux, map)) {
20440 		verbose(env, "only one cgroup storage of each type is allowed\n");
20441 		return -EBUSY;
20442 	}
20443 
20444 	if (map->map_type == BPF_MAP_TYPE_ARENA) {
20445 		if (env->prog->aux->arena) {
20446 			verbose(env, "Only one arena per program\n");
20447 			return -EBUSY;
20448 		}
20449 		if (!env->allow_ptr_leaks || !env->bpf_capable) {
20450 			verbose(env, "CAP_BPF and CAP_PERFMON are required to use arena\n");
20451 			return -EPERM;
20452 		}
20453 		if (!env->prog->jit_requested) {
20454 			verbose(env, "JIT is required to use arena\n");
20455 			return -EOPNOTSUPP;
20456 		}
20457 		if (!bpf_jit_supports_arena()) {
20458 			verbose(env, "JIT doesn't support arena\n");
20459 			return -EOPNOTSUPP;
20460 		}
20461 		env->prog->aux->arena = (void *)map;
20462 		if (!bpf_arena_get_user_vm_start(env->prog->aux->arena)) {
20463 			verbose(env, "arena's user address must be set via map_extra or mmap()\n");
20464 			return -EINVAL;
20465 		}
20466 	}
20467 
20468 	return 0;
20469 }
20470 
__add_used_map(struct bpf_verifier_env * env,struct bpf_map * map)20471 static int __add_used_map(struct bpf_verifier_env *env, struct bpf_map *map)
20472 {
20473 	int i, err;
20474 
20475 	/* check whether we recorded this map already */
20476 	for (i = 0; i < env->used_map_cnt; i++)
20477 		if (env->used_maps[i] == map)
20478 			return i;
20479 
20480 	if (env->used_map_cnt >= MAX_USED_MAPS) {
20481 		verbose(env, "The total number of maps per program has reached the limit of %u\n",
20482 			MAX_USED_MAPS);
20483 		return -E2BIG;
20484 	}
20485 
20486 	err = check_map_prog_compatibility(env, map, env->prog);
20487 	if (err)
20488 		return err;
20489 
20490 	if (env->prog->sleepable)
20491 		atomic64_inc(&map->sleepable_refcnt);
20492 
20493 	/* hold the map. If the program is rejected by verifier,
20494 	 * the map will be released by release_maps() or it
20495 	 * will be used by the valid program until it's unloaded
20496 	 * and all maps are released in bpf_free_used_maps()
20497 	 */
20498 	bpf_map_inc(map);
20499 
20500 	env->used_maps[env->used_map_cnt++] = map;
20501 
20502 	return env->used_map_cnt - 1;
20503 }
20504 
20505 /* Add map behind fd to used maps list, if it's not already there, and return
20506  * its index.
20507  * Returns <0 on error, or >= 0 index, on success.
20508  */
add_used_map(struct bpf_verifier_env * env,int fd)20509 static int add_used_map(struct bpf_verifier_env *env, int fd)
20510 {
20511 	struct bpf_map *map;
20512 	CLASS(fd, f)(fd);
20513 
20514 	map = __bpf_map_get(f);
20515 	if (IS_ERR(map)) {
20516 		verbose(env, "fd %d is not pointing to valid bpf_map\n", fd);
20517 		return PTR_ERR(map);
20518 	}
20519 
20520 	return __add_used_map(env, map);
20521 }
20522 
20523 /* find and rewrite pseudo imm in ld_imm64 instructions:
20524  *
20525  * 1. if it accesses map FD, replace it with actual map pointer.
20526  * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
20527  *
20528  * NOTE: btf_vmlinux is required for converting pseudo btf_id.
20529  */
resolve_pseudo_ldimm64(struct bpf_verifier_env * env)20530 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
20531 {
20532 	struct bpf_insn *insn = env->prog->insnsi;
20533 	int insn_cnt = env->prog->len;
20534 	int i, err;
20535 
20536 	err = bpf_prog_calc_tag(env->prog);
20537 	if (err)
20538 		return err;
20539 
20540 	for (i = 0; i < insn_cnt; i++, insn++) {
20541 		if (BPF_CLASS(insn->code) == BPF_LDX &&
20542 		    ((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
20543 		    insn->imm != 0)) {
20544 			verbose(env, "BPF_LDX uses reserved fields\n");
20545 			return -EINVAL;
20546 		}
20547 
20548 		if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
20549 			struct bpf_insn_aux_data *aux;
20550 			struct bpf_map *map;
20551 			int map_idx;
20552 			u64 addr;
20553 			u32 fd;
20554 
20555 			if (i == insn_cnt - 1 || insn[1].code != 0 ||
20556 			    insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
20557 			    insn[1].off != 0) {
20558 				verbose(env, "invalid bpf_ld_imm64 insn\n");
20559 				return -EINVAL;
20560 			}
20561 
20562 			if (insn[0].src_reg == 0)
20563 				/* valid generic load 64-bit imm */
20564 				goto next_insn;
20565 
20566 			if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
20567 				aux = &env->insn_aux_data[i];
20568 				err = check_pseudo_btf_id(env, insn, aux);
20569 				if (err)
20570 					return err;
20571 				goto next_insn;
20572 			}
20573 
20574 			if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
20575 				aux = &env->insn_aux_data[i];
20576 				aux->ptr_type = PTR_TO_FUNC;
20577 				goto next_insn;
20578 			}
20579 
20580 			/* In final convert_pseudo_ld_imm64() step, this is
20581 			 * converted into regular 64-bit imm load insn.
20582 			 */
20583 			switch (insn[0].src_reg) {
20584 			case BPF_PSEUDO_MAP_VALUE:
20585 			case BPF_PSEUDO_MAP_IDX_VALUE:
20586 				break;
20587 			case BPF_PSEUDO_MAP_FD:
20588 			case BPF_PSEUDO_MAP_IDX:
20589 				if (insn[1].imm == 0)
20590 					break;
20591 				fallthrough;
20592 			default:
20593 				verbose(env, "unrecognized bpf_ld_imm64 insn\n");
20594 				return -EINVAL;
20595 			}
20596 
20597 			switch (insn[0].src_reg) {
20598 			case BPF_PSEUDO_MAP_IDX_VALUE:
20599 			case BPF_PSEUDO_MAP_IDX:
20600 				if (bpfptr_is_null(env->fd_array)) {
20601 					verbose(env, "fd_idx without fd_array is invalid\n");
20602 					return -EPROTO;
20603 				}
20604 				if (copy_from_bpfptr_offset(&fd, env->fd_array,
20605 							    insn[0].imm * sizeof(fd),
20606 							    sizeof(fd)))
20607 					return -EFAULT;
20608 				break;
20609 			default:
20610 				fd = insn[0].imm;
20611 				break;
20612 			}
20613 
20614 			map_idx = add_used_map(env, fd);
20615 			if (map_idx < 0)
20616 				return map_idx;
20617 			map = env->used_maps[map_idx];
20618 
20619 			aux = &env->insn_aux_data[i];
20620 			aux->map_index = map_idx;
20621 
20622 			if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
20623 			    insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
20624 				addr = (unsigned long)map;
20625 			} else {
20626 				u32 off = insn[1].imm;
20627 
20628 				if (off >= BPF_MAX_VAR_OFF) {
20629 					verbose(env, "direct value offset of %u is not allowed\n", off);
20630 					return -EINVAL;
20631 				}
20632 
20633 				if (!map->ops->map_direct_value_addr) {
20634 					verbose(env, "no direct value access support for this map type\n");
20635 					return -EINVAL;
20636 				}
20637 
20638 				err = map->ops->map_direct_value_addr(map, &addr, off);
20639 				if (err) {
20640 					verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
20641 						map->value_size, off);
20642 					return err;
20643 				}
20644 
20645 				aux->map_off = off;
20646 				addr += off;
20647 			}
20648 
20649 			insn[0].imm = (u32)addr;
20650 			insn[1].imm = addr >> 32;
20651 
20652 next_insn:
20653 			insn++;
20654 			i++;
20655 			continue;
20656 		}
20657 
20658 		/* Basic sanity check before we invest more work here. */
20659 		if (!bpf_opcode_in_insntable(insn->code)) {
20660 			verbose(env, "unknown opcode %02x\n", insn->code);
20661 			return -EINVAL;
20662 		}
20663 	}
20664 
20665 	/* now all pseudo BPF_LD_IMM64 instructions load valid
20666 	 * 'struct bpf_map *' into a register instead of user map_fd.
20667 	 * These pointers will be used later by verifier to validate map access.
20668 	 */
20669 	return 0;
20670 }
20671 
20672 /* drop refcnt of maps used by the rejected program */
release_maps(struct bpf_verifier_env * env)20673 static void release_maps(struct bpf_verifier_env *env)
20674 {
20675 	__bpf_free_used_maps(env->prog->aux, env->used_maps,
20676 			     env->used_map_cnt);
20677 }
20678 
20679 /* drop refcnt of maps used by the rejected program */
release_btfs(struct bpf_verifier_env * env)20680 static void release_btfs(struct bpf_verifier_env *env)
20681 {
20682 	__bpf_free_used_btfs(env->used_btfs, env->used_btf_cnt);
20683 }
20684 
20685 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
convert_pseudo_ld_imm64(struct bpf_verifier_env * env)20686 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
20687 {
20688 	struct bpf_insn *insn = env->prog->insnsi;
20689 	int insn_cnt = env->prog->len;
20690 	int i;
20691 
20692 	for (i = 0; i < insn_cnt; i++, insn++) {
20693 		if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
20694 			continue;
20695 		if (insn->src_reg == BPF_PSEUDO_FUNC)
20696 			continue;
20697 		insn->src_reg = 0;
20698 	}
20699 }
20700 
20701 /* single env->prog->insni[off] instruction was replaced with the range
20702  * insni[off, off + cnt).  Adjust corresponding insn_aux_data by copying
20703  * [0, off) and [off, end) to new locations, so the patched range stays zero
20704  */
adjust_insn_aux_data(struct bpf_verifier_env * env,struct bpf_insn_aux_data * new_data,struct bpf_prog * new_prog,u32 off,u32 cnt)20705 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
20706 				 struct bpf_insn_aux_data *new_data,
20707 				 struct bpf_prog *new_prog, u32 off, u32 cnt)
20708 {
20709 	struct bpf_insn_aux_data *old_data = env->insn_aux_data;
20710 	struct bpf_insn *insn = new_prog->insnsi;
20711 	u32 old_seen = old_data[off].seen;
20712 	u32 prog_len;
20713 	int i;
20714 
20715 	/* aux info at OFF always needs adjustment, no matter fast path
20716 	 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
20717 	 * original insn at old prog.
20718 	 */
20719 	old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
20720 
20721 	if (cnt == 1)
20722 		return;
20723 	prog_len = new_prog->len;
20724 
20725 	memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
20726 	memcpy(new_data + off + cnt - 1, old_data + off,
20727 	       sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
20728 	for (i = off; i < off + cnt - 1; i++) {
20729 		/* Expand insni[off]'s seen count to the patched range. */
20730 		new_data[i].seen = old_seen;
20731 		new_data[i].zext_dst = insn_has_def32(env, insn + i);
20732 	}
20733 	env->insn_aux_data = new_data;
20734 	vfree(old_data);
20735 }
20736 
adjust_subprog_starts(struct bpf_verifier_env * env,u32 off,u32 len)20737 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
20738 {
20739 	int i;
20740 
20741 	if (len == 1)
20742 		return;
20743 	/* NOTE: fake 'exit' subprog should be updated as well. */
20744 	for (i = 0; i <= env->subprog_cnt; i++) {
20745 		if (env->subprog_info[i].start <= off)
20746 			continue;
20747 		env->subprog_info[i].start += len - 1;
20748 	}
20749 }
20750 
adjust_poke_descs(struct bpf_prog * prog,u32 off,u32 len)20751 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
20752 {
20753 	struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
20754 	int i, sz = prog->aux->size_poke_tab;
20755 	struct bpf_jit_poke_descriptor *desc;
20756 
20757 	for (i = 0; i < sz; i++) {
20758 		desc = &tab[i];
20759 		if (desc->insn_idx <= off)
20760 			continue;
20761 		desc->insn_idx += len - 1;
20762 	}
20763 }
20764 
bpf_patch_insn_data(struct bpf_verifier_env * env,u32 off,const struct bpf_insn * patch,u32 len)20765 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
20766 					    const struct bpf_insn *patch, u32 len)
20767 {
20768 	struct bpf_prog *new_prog;
20769 	struct bpf_insn_aux_data *new_data = NULL;
20770 
20771 	if (len > 1) {
20772 		new_data = vzalloc(array_size(env->prog->len + len - 1,
20773 					      sizeof(struct bpf_insn_aux_data)));
20774 		if (!new_data)
20775 			return NULL;
20776 	}
20777 
20778 	new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
20779 	if (IS_ERR(new_prog)) {
20780 		if (PTR_ERR(new_prog) == -ERANGE)
20781 			verbose(env,
20782 				"insn %d cannot be patched due to 16-bit range\n",
20783 				env->insn_aux_data[off].orig_idx);
20784 		vfree(new_data);
20785 		return NULL;
20786 	}
20787 	adjust_insn_aux_data(env, new_data, new_prog, off, len);
20788 	adjust_subprog_starts(env, off, len);
20789 	adjust_poke_descs(new_prog, off, len);
20790 	return new_prog;
20791 }
20792 
20793 /*
20794  * For all jmp insns in a given 'prog' that point to 'tgt_idx' insn adjust the
20795  * jump offset by 'delta'.
20796  */
adjust_jmp_off(struct bpf_prog * prog,u32 tgt_idx,u32 delta)20797 static int adjust_jmp_off(struct bpf_prog *prog, u32 tgt_idx, u32 delta)
20798 {
20799 	struct bpf_insn *insn = prog->insnsi;
20800 	u32 insn_cnt = prog->len, i;
20801 	s32 imm;
20802 	s16 off;
20803 
20804 	for (i = 0; i < insn_cnt; i++, insn++) {
20805 		u8 code = insn->code;
20806 
20807 		if (tgt_idx <= i && i < tgt_idx + delta)
20808 			continue;
20809 
20810 		if ((BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32) ||
20811 		    BPF_OP(code) == BPF_CALL || BPF_OP(code) == BPF_EXIT)
20812 			continue;
20813 
20814 		if (insn->code == (BPF_JMP32 | BPF_JA)) {
20815 			if (i + 1 + insn->imm != tgt_idx)
20816 				continue;
20817 			if (check_add_overflow(insn->imm, delta, &imm))
20818 				return -ERANGE;
20819 			insn->imm = imm;
20820 		} else {
20821 			if (i + 1 + insn->off != tgt_idx)
20822 				continue;
20823 			if (check_add_overflow(insn->off, delta, &off))
20824 				return -ERANGE;
20825 			insn->off = off;
20826 		}
20827 	}
20828 	return 0;
20829 }
20830 
adjust_subprog_starts_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)20831 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
20832 					      u32 off, u32 cnt)
20833 {
20834 	int i, j;
20835 
20836 	/* find first prog starting at or after off (first to remove) */
20837 	for (i = 0; i < env->subprog_cnt; i++)
20838 		if (env->subprog_info[i].start >= off)
20839 			break;
20840 	/* find first prog starting at or after off + cnt (first to stay) */
20841 	for (j = i; j < env->subprog_cnt; j++)
20842 		if (env->subprog_info[j].start >= off + cnt)
20843 			break;
20844 	/* if j doesn't start exactly at off + cnt, we are just removing
20845 	 * the front of previous prog
20846 	 */
20847 	if (env->subprog_info[j].start != off + cnt)
20848 		j--;
20849 
20850 	if (j > i) {
20851 		struct bpf_prog_aux *aux = env->prog->aux;
20852 		int move;
20853 
20854 		/* move fake 'exit' subprog as well */
20855 		move = env->subprog_cnt + 1 - j;
20856 
20857 		memmove(env->subprog_info + i,
20858 			env->subprog_info + j,
20859 			sizeof(*env->subprog_info) * move);
20860 		env->subprog_cnt -= j - i;
20861 
20862 		/* remove func_info */
20863 		if (aux->func_info) {
20864 			move = aux->func_info_cnt - j;
20865 
20866 			memmove(aux->func_info + i,
20867 				aux->func_info + j,
20868 				sizeof(*aux->func_info) * move);
20869 			aux->func_info_cnt -= j - i;
20870 			/* func_info->insn_off is set after all code rewrites,
20871 			 * in adjust_btf_func() - no need to adjust
20872 			 */
20873 		}
20874 	} else {
20875 		/* convert i from "first prog to remove" to "first to adjust" */
20876 		if (env->subprog_info[i].start == off)
20877 			i++;
20878 	}
20879 
20880 	/* update fake 'exit' subprog as well */
20881 	for (; i <= env->subprog_cnt; i++)
20882 		env->subprog_info[i].start -= cnt;
20883 
20884 	return 0;
20885 }
20886 
bpf_adj_linfo_after_remove(struct bpf_verifier_env * env,u32 off,u32 cnt)20887 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
20888 				      u32 cnt)
20889 {
20890 	struct bpf_prog *prog = env->prog;
20891 	u32 i, l_off, l_cnt, nr_linfo;
20892 	struct bpf_line_info *linfo;
20893 
20894 	nr_linfo = prog->aux->nr_linfo;
20895 	if (!nr_linfo)
20896 		return 0;
20897 
20898 	linfo = prog->aux->linfo;
20899 
20900 	/* find first line info to remove, count lines to be removed */
20901 	for (i = 0; i < nr_linfo; i++)
20902 		if (linfo[i].insn_off >= off)
20903 			break;
20904 
20905 	l_off = i;
20906 	l_cnt = 0;
20907 	for (; i < nr_linfo; i++)
20908 		if (linfo[i].insn_off < off + cnt)
20909 			l_cnt++;
20910 		else
20911 			break;
20912 
20913 	/* First live insn doesn't match first live linfo, it needs to "inherit"
20914 	 * last removed linfo.  prog is already modified, so prog->len == off
20915 	 * means no live instructions after (tail of the program was removed).
20916 	 */
20917 	if (prog->len != off && l_cnt &&
20918 	    (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
20919 		l_cnt--;
20920 		linfo[--i].insn_off = off + cnt;
20921 	}
20922 
20923 	/* remove the line info which refer to the removed instructions */
20924 	if (l_cnt) {
20925 		memmove(linfo + l_off, linfo + i,
20926 			sizeof(*linfo) * (nr_linfo - i));
20927 
20928 		prog->aux->nr_linfo -= l_cnt;
20929 		nr_linfo = prog->aux->nr_linfo;
20930 	}
20931 
20932 	/* pull all linfo[i].insn_off >= off + cnt in by cnt */
20933 	for (i = l_off; i < nr_linfo; i++)
20934 		linfo[i].insn_off -= cnt;
20935 
20936 	/* fix up all subprogs (incl. 'exit') which start >= off */
20937 	for (i = 0; i <= env->subprog_cnt; i++)
20938 		if (env->subprog_info[i].linfo_idx > l_off) {
20939 			/* program may have started in the removed region but
20940 			 * may not be fully removed
20941 			 */
20942 			if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
20943 				env->subprog_info[i].linfo_idx -= l_cnt;
20944 			else
20945 				env->subprog_info[i].linfo_idx = l_off;
20946 		}
20947 
20948 	return 0;
20949 }
20950 
verifier_remove_insns(struct bpf_verifier_env * env,u32 off,u32 cnt)20951 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
20952 {
20953 	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
20954 	unsigned int orig_prog_len = env->prog->len;
20955 	int err;
20956 
20957 	if (bpf_prog_is_offloaded(env->prog->aux))
20958 		bpf_prog_offload_remove_insns(env, off, cnt);
20959 
20960 	err = bpf_remove_insns(env->prog, off, cnt);
20961 	if (err)
20962 		return err;
20963 
20964 	err = adjust_subprog_starts_after_remove(env, off, cnt);
20965 	if (err)
20966 		return err;
20967 
20968 	err = bpf_adj_linfo_after_remove(env, off, cnt);
20969 	if (err)
20970 		return err;
20971 
20972 	memmove(aux_data + off,	aux_data + off + cnt,
20973 		sizeof(*aux_data) * (orig_prog_len - off - cnt));
20974 
20975 	return 0;
20976 }
20977 
20978 /* The verifier does more data flow analysis than llvm and will not
20979  * explore branches that are dead at run time. Malicious programs can
20980  * have dead code too. Therefore replace all dead at-run-time code
20981  * with 'ja -1'.
20982  *
20983  * Just nops are not optimal, e.g. if they would sit at the end of the
20984  * program and through another bug we would manage to jump there, then
20985  * we'd execute beyond program memory otherwise. Returning exception
20986  * code also wouldn't work since we can have subprogs where the dead
20987  * code could be located.
20988  */
sanitize_dead_code(struct bpf_verifier_env * env)20989 static void sanitize_dead_code(struct bpf_verifier_env *env)
20990 {
20991 	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
20992 	struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
20993 	struct bpf_insn *insn = env->prog->insnsi;
20994 	const int insn_cnt = env->prog->len;
20995 	int i;
20996 
20997 	for (i = 0; i < insn_cnt; i++) {
20998 		if (aux_data[i].seen)
20999 			continue;
21000 		memcpy(insn + i, &trap, sizeof(trap));
21001 		aux_data[i].zext_dst = false;
21002 	}
21003 }
21004 
insn_is_cond_jump(u8 code)21005 static bool insn_is_cond_jump(u8 code)
21006 {
21007 	u8 op;
21008 
21009 	op = BPF_OP(code);
21010 	if (BPF_CLASS(code) == BPF_JMP32)
21011 		return op != BPF_JA;
21012 
21013 	if (BPF_CLASS(code) != BPF_JMP)
21014 		return false;
21015 
21016 	return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
21017 }
21018 
opt_hard_wire_dead_code_branches(struct bpf_verifier_env * env)21019 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
21020 {
21021 	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
21022 	struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
21023 	struct bpf_insn *insn = env->prog->insnsi;
21024 	const int insn_cnt = env->prog->len;
21025 	int i;
21026 
21027 	for (i = 0; i < insn_cnt; i++, insn++) {
21028 		if (!insn_is_cond_jump(insn->code))
21029 			continue;
21030 
21031 		if (!aux_data[i + 1].seen)
21032 			ja.off = insn->off;
21033 		else if (!aux_data[i + 1 + insn->off].seen)
21034 			ja.off = 0;
21035 		else
21036 			continue;
21037 
21038 		if (bpf_prog_is_offloaded(env->prog->aux))
21039 			bpf_prog_offload_replace_insn(env, i, &ja);
21040 
21041 		memcpy(insn, &ja, sizeof(ja));
21042 	}
21043 }
21044 
opt_remove_dead_code(struct bpf_verifier_env * env)21045 static int opt_remove_dead_code(struct bpf_verifier_env *env)
21046 {
21047 	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
21048 	int insn_cnt = env->prog->len;
21049 	int i, err;
21050 
21051 	for (i = 0; i < insn_cnt; i++) {
21052 		int j;
21053 
21054 		j = 0;
21055 		while (i + j < insn_cnt && !aux_data[i + j].seen)
21056 			j++;
21057 		if (!j)
21058 			continue;
21059 
21060 		err = verifier_remove_insns(env, i, j);
21061 		if (err)
21062 			return err;
21063 		insn_cnt = env->prog->len;
21064 	}
21065 
21066 	return 0;
21067 }
21068 
21069 static const struct bpf_insn NOP = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
21070 static const struct bpf_insn MAY_GOTO_0 = BPF_RAW_INSN(BPF_JMP | BPF_JCOND, 0, 0, 0, 0);
21071 
opt_remove_nops(struct bpf_verifier_env * env)21072 static int opt_remove_nops(struct bpf_verifier_env *env)
21073 {
21074 	struct bpf_insn *insn = env->prog->insnsi;
21075 	int insn_cnt = env->prog->len;
21076 	bool is_may_goto_0, is_ja;
21077 	int i, err;
21078 
21079 	for (i = 0; i < insn_cnt; i++) {
21080 		is_may_goto_0 = !memcmp(&insn[i], &MAY_GOTO_0, sizeof(MAY_GOTO_0));
21081 		is_ja = !memcmp(&insn[i], &NOP, sizeof(NOP));
21082 
21083 		if (!is_may_goto_0 && !is_ja)
21084 			continue;
21085 
21086 		err = verifier_remove_insns(env, i, 1);
21087 		if (err)
21088 			return err;
21089 		insn_cnt--;
21090 		/* Go back one insn to catch may_goto +1; may_goto +0 sequence */
21091 		i -= (is_may_goto_0 && i > 0) ? 2 : 1;
21092 	}
21093 
21094 	return 0;
21095 }
21096 
opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env * env,const union bpf_attr * attr)21097 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
21098 					 const union bpf_attr *attr)
21099 {
21100 	struct bpf_insn *patch;
21101 	/* use env->insn_buf as two independent buffers */
21102 	struct bpf_insn *zext_patch = env->insn_buf;
21103 	struct bpf_insn *rnd_hi32_patch = &env->insn_buf[2];
21104 	struct bpf_insn_aux_data *aux = env->insn_aux_data;
21105 	int i, patch_len, delta = 0, len = env->prog->len;
21106 	struct bpf_insn *insns = env->prog->insnsi;
21107 	struct bpf_prog *new_prog;
21108 	bool rnd_hi32;
21109 
21110 	rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
21111 	zext_patch[1] = BPF_ZEXT_REG(0);
21112 	rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
21113 	rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
21114 	rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
21115 	for (i = 0; i < len; i++) {
21116 		int adj_idx = i + delta;
21117 		struct bpf_insn insn;
21118 		int load_reg;
21119 
21120 		insn = insns[adj_idx];
21121 		load_reg = insn_def_regno(&insn);
21122 		if (!aux[adj_idx].zext_dst) {
21123 			u8 code, class;
21124 			u32 imm_rnd;
21125 
21126 			if (!rnd_hi32)
21127 				continue;
21128 
21129 			code = insn.code;
21130 			class = BPF_CLASS(code);
21131 			if (load_reg == -1)
21132 				continue;
21133 
21134 			/* NOTE: arg "reg" (the fourth one) is only used for
21135 			 *       BPF_STX + SRC_OP, so it is safe to pass NULL
21136 			 *       here.
21137 			 */
21138 			if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
21139 				if (class == BPF_LD &&
21140 				    BPF_MODE(code) == BPF_IMM)
21141 					i++;
21142 				continue;
21143 			}
21144 
21145 			/* ctx load could be transformed into wider load. */
21146 			if (class == BPF_LDX &&
21147 			    aux[adj_idx].ptr_type == PTR_TO_CTX)
21148 				continue;
21149 
21150 			imm_rnd = get_random_u32();
21151 			rnd_hi32_patch[0] = insn;
21152 			rnd_hi32_patch[1].imm = imm_rnd;
21153 			rnd_hi32_patch[3].dst_reg = load_reg;
21154 			patch = rnd_hi32_patch;
21155 			patch_len = 4;
21156 			goto apply_patch_buffer;
21157 		}
21158 
21159 		/* Add in an zero-extend instruction if a) the JIT has requested
21160 		 * it or b) it's a CMPXCHG.
21161 		 *
21162 		 * The latter is because: BPF_CMPXCHG always loads a value into
21163 		 * R0, therefore always zero-extends. However some archs'
21164 		 * equivalent instruction only does this load when the
21165 		 * comparison is successful. This detail of CMPXCHG is
21166 		 * orthogonal to the general zero-extension behaviour of the
21167 		 * CPU, so it's treated independently of bpf_jit_needs_zext.
21168 		 */
21169 		if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
21170 			continue;
21171 
21172 		/* Zero-extension is done by the caller. */
21173 		if (bpf_pseudo_kfunc_call(&insn))
21174 			continue;
21175 
21176 		if (verifier_bug_if(load_reg == -1, env,
21177 				    "zext_dst is set, but no reg is defined"))
21178 			return -EFAULT;
21179 
21180 		zext_patch[0] = insn;
21181 		zext_patch[1].dst_reg = load_reg;
21182 		zext_patch[1].src_reg = load_reg;
21183 		patch = zext_patch;
21184 		patch_len = 2;
21185 apply_patch_buffer:
21186 		new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
21187 		if (!new_prog)
21188 			return -ENOMEM;
21189 		env->prog = new_prog;
21190 		insns = new_prog->insnsi;
21191 		aux = env->insn_aux_data;
21192 		delta += patch_len - 1;
21193 	}
21194 
21195 	return 0;
21196 }
21197 
21198 /* convert load instructions that access fields of a context type into a
21199  * sequence of instructions that access fields of the underlying structure:
21200  *     struct __sk_buff    -> struct sk_buff
21201  *     struct bpf_sock_ops -> struct sock
21202  */
convert_ctx_accesses(struct bpf_verifier_env * env)21203 static int convert_ctx_accesses(struct bpf_verifier_env *env)
21204 {
21205 	struct bpf_subprog_info *subprogs = env->subprog_info;
21206 	const struct bpf_verifier_ops *ops = env->ops;
21207 	int i, cnt, size, ctx_field_size, ret, delta = 0, epilogue_cnt = 0;
21208 	const int insn_cnt = env->prog->len;
21209 	struct bpf_insn *epilogue_buf = env->epilogue_buf;
21210 	struct bpf_insn *insn_buf = env->insn_buf;
21211 	struct bpf_insn *insn;
21212 	u32 target_size, size_default, off;
21213 	struct bpf_prog *new_prog;
21214 	enum bpf_access_type type;
21215 	bool is_narrower_load;
21216 	int epilogue_idx = 0;
21217 
21218 	if (ops->gen_epilogue) {
21219 		epilogue_cnt = ops->gen_epilogue(epilogue_buf, env->prog,
21220 						 -(subprogs[0].stack_depth + 8));
21221 		if (epilogue_cnt >= INSN_BUF_SIZE) {
21222 			verifier_bug(env, "epilogue is too long");
21223 			return -EFAULT;
21224 		} else if (epilogue_cnt) {
21225 			/* Save the ARG_PTR_TO_CTX for the epilogue to use */
21226 			cnt = 0;
21227 			subprogs[0].stack_depth += 8;
21228 			insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_FP, BPF_REG_1,
21229 						      -subprogs[0].stack_depth);
21230 			insn_buf[cnt++] = env->prog->insnsi[0];
21231 			new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
21232 			if (!new_prog)
21233 				return -ENOMEM;
21234 			env->prog = new_prog;
21235 			delta += cnt - 1;
21236 
21237 			ret = add_kfunc_in_insns(env, epilogue_buf, epilogue_cnt - 1);
21238 			if (ret < 0)
21239 				return ret;
21240 		}
21241 	}
21242 
21243 	if (ops->gen_prologue || env->seen_direct_write) {
21244 		if (!ops->gen_prologue) {
21245 			verifier_bug(env, "gen_prologue is null");
21246 			return -EFAULT;
21247 		}
21248 		cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
21249 					env->prog);
21250 		if (cnt >= INSN_BUF_SIZE) {
21251 			verifier_bug(env, "prologue is too long");
21252 			return -EFAULT;
21253 		} else if (cnt) {
21254 			new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
21255 			if (!new_prog)
21256 				return -ENOMEM;
21257 
21258 			env->prog = new_prog;
21259 			delta += cnt - 1;
21260 
21261 			ret = add_kfunc_in_insns(env, insn_buf, cnt - 1);
21262 			if (ret < 0)
21263 				return ret;
21264 		}
21265 	}
21266 
21267 	if (delta)
21268 		WARN_ON(adjust_jmp_off(env->prog, 0, delta));
21269 
21270 	if (bpf_prog_is_offloaded(env->prog->aux))
21271 		return 0;
21272 
21273 	insn = env->prog->insnsi + delta;
21274 
21275 	for (i = 0; i < insn_cnt; i++, insn++) {
21276 		bpf_convert_ctx_access_t convert_ctx_access;
21277 		u8 mode;
21278 
21279 		if (env->insn_aux_data[i + delta].nospec) {
21280 			WARN_ON_ONCE(env->insn_aux_data[i + delta].alu_state);
21281 			struct bpf_insn *patch = insn_buf;
21282 
21283 			*patch++ = BPF_ST_NOSPEC();
21284 			*patch++ = *insn;
21285 			cnt = patch - insn_buf;
21286 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
21287 			if (!new_prog)
21288 				return -ENOMEM;
21289 
21290 			delta    += cnt - 1;
21291 			env->prog = new_prog;
21292 			insn      = new_prog->insnsi + i + delta;
21293 			/* This can not be easily merged with the
21294 			 * nospec_result-case, because an insn may require a
21295 			 * nospec before and after itself. Therefore also do not
21296 			 * 'continue' here but potentially apply further
21297 			 * patching to insn. *insn should equal patch[1] now.
21298 			 */
21299 		}
21300 
21301 		if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
21302 		    insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
21303 		    insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
21304 		    insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
21305 		    insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
21306 		    insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
21307 		    insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
21308 			type = BPF_READ;
21309 		} else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
21310 			   insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
21311 			   insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
21312 			   insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
21313 			   insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
21314 			   insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
21315 			   insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
21316 			   insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
21317 			type = BPF_WRITE;
21318 		} else if ((insn->code == (BPF_STX | BPF_ATOMIC | BPF_B) ||
21319 			    insn->code == (BPF_STX | BPF_ATOMIC | BPF_H) ||
21320 			    insn->code == (BPF_STX | BPF_ATOMIC | BPF_W) ||
21321 			    insn->code == (BPF_STX | BPF_ATOMIC | BPF_DW)) &&
21322 			   env->insn_aux_data[i + delta].ptr_type == PTR_TO_ARENA) {
21323 			insn->code = BPF_STX | BPF_PROBE_ATOMIC | BPF_SIZE(insn->code);
21324 			env->prog->aux->num_exentries++;
21325 			continue;
21326 		} else if (insn->code == (BPF_JMP | BPF_EXIT) &&
21327 			   epilogue_cnt &&
21328 			   i + delta < subprogs[1].start) {
21329 			/* Generate epilogue for the main prog */
21330 			if (epilogue_idx) {
21331 				/* jump back to the earlier generated epilogue */
21332 				insn_buf[0] = BPF_JMP32_A(epilogue_idx - i - delta - 1);
21333 				cnt = 1;
21334 			} else {
21335 				memcpy(insn_buf, epilogue_buf,
21336 				       epilogue_cnt * sizeof(*epilogue_buf));
21337 				cnt = epilogue_cnt;
21338 				/* epilogue_idx cannot be 0. It must have at
21339 				 * least one ctx ptr saving insn before the
21340 				 * epilogue.
21341 				 */
21342 				epilogue_idx = i + delta;
21343 			}
21344 			goto patch_insn_buf;
21345 		} else {
21346 			continue;
21347 		}
21348 
21349 		if (type == BPF_WRITE &&
21350 		    env->insn_aux_data[i + delta].nospec_result) {
21351 			/* nospec_result is only used to mitigate Spectre v4 and
21352 			 * to limit verification-time for Spectre v1.
21353 			 */
21354 			struct bpf_insn *patch = insn_buf;
21355 
21356 			*patch++ = *insn;
21357 			*patch++ = BPF_ST_NOSPEC();
21358 			cnt = patch - insn_buf;
21359 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
21360 			if (!new_prog)
21361 				return -ENOMEM;
21362 
21363 			delta    += cnt - 1;
21364 			env->prog = new_prog;
21365 			insn      = new_prog->insnsi + i + delta;
21366 			continue;
21367 		}
21368 
21369 		switch ((int)env->insn_aux_data[i + delta].ptr_type) {
21370 		case PTR_TO_CTX:
21371 			if (!ops->convert_ctx_access)
21372 				continue;
21373 			convert_ctx_access = ops->convert_ctx_access;
21374 			break;
21375 		case PTR_TO_SOCKET:
21376 		case PTR_TO_SOCK_COMMON:
21377 			convert_ctx_access = bpf_sock_convert_ctx_access;
21378 			break;
21379 		case PTR_TO_TCP_SOCK:
21380 			convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
21381 			break;
21382 		case PTR_TO_XDP_SOCK:
21383 			convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
21384 			break;
21385 		case PTR_TO_BTF_ID:
21386 		case PTR_TO_BTF_ID | PTR_UNTRUSTED:
21387 		/* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
21388 		 * PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
21389 		 * be said once it is marked PTR_UNTRUSTED, hence we must handle
21390 		 * any faults for loads into such types. BPF_WRITE is disallowed
21391 		 * for this case.
21392 		 */
21393 		case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
21394 		case PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED:
21395 			if (type == BPF_READ) {
21396 				if (BPF_MODE(insn->code) == BPF_MEM)
21397 					insn->code = BPF_LDX | BPF_PROBE_MEM |
21398 						     BPF_SIZE((insn)->code);
21399 				else
21400 					insn->code = BPF_LDX | BPF_PROBE_MEMSX |
21401 						     BPF_SIZE((insn)->code);
21402 				env->prog->aux->num_exentries++;
21403 			}
21404 			continue;
21405 		case PTR_TO_ARENA:
21406 			if (BPF_MODE(insn->code) == BPF_MEMSX) {
21407 				verbose(env, "sign extending loads from arena are not supported yet\n");
21408 				return -EOPNOTSUPP;
21409 			}
21410 			insn->code = BPF_CLASS(insn->code) | BPF_PROBE_MEM32 | BPF_SIZE(insn->code);
21411 			env->prog->aux->num_exentries++;
21412 			continue;
21413 		default:
21414 			continue;
21415 		}
21416 
21417 		ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
21418 		size = BPF_LDST_BYTES(insn);
21419 		mode = BPF_MODE(insn->code);
21420 
21421 		/* If the read access is a narrower load of the field,
21422 		 * convert to a 4/8-byte load, to minimum program type specific
21423 		 * convert_ctx_access changes. If conversion is successful,
21424 		 * we will apply proper mask to the result.
21425 		 */
21426 		is_narrower_load = size < ctx_field_size;
21427 		size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
21428 		off = insn->off;
21429 		if (is_narrower_load) {
21430 			u8 size_code;
21431 
21432 			if (type == BPF_WRITE) {
21433 				verifier_bug(env, "narrow ctx access misconfigured");
21434 				return -EFAULT;
21435 			}
21436 
21437 			size_code = BPF_H;
21438 			if (ctx_field_size == 4)
21439 				size_code = BPF_W;
21440 			else if (ctx_field_size == 8)
21441 				size_code = BPF_DW;
21442 
21443 			insn->off = off & ~(size_default - 1);
21444 			insn->code = BPF_LDX | BPF_MEM | size_code;
21445 		}
21446 
21447 		target_size = 0;
21448 		cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
21449 					 &target_size);
21450 		if (cnt == 0 || cnt >= INSN_BUF_SIZE ||
21451 		    (ctx_field_size && !target_size)) {
21452 			verifier_bug(env, "error during ctx access conversion (%d)", cnt);
21453 			return -EFAULT;
21454 		}
21455 
21456 		if (is_narrower_load && size < target_size) {
21457 			u8 shift = bpf_ctx_narrow_access_offset(
21458 				off, size, size_default) * 8;
21459 			if (shift && cnt + 1 >= INSN_BUF_SIZE) {
21460 				verifier_bug(env, "narrow ctx load misconfigured");
21461 				return -EFAULT;
21462 			}
21463 			if (ctx_field_size <= 4) {
21464 				if (shift)
21465 					insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
21466 									insn->dst_reg,
21467 									shift);
21468 				insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
21469 								(1 << size * 8) - 1);
21470 			} else {
21471 				if (shift)
21472 					insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
21473 									insn->dst_reg,
21474 									shift);
21475 				insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
21476 								(1ULL << size * 8) - 1);
21477 			}
21478 		}
21479 		if (mode == BPF_MEMSX)
21480 			insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
21481 						       insn->dst_reg, insn->dst_reg,
21482 						       size * 8, 0);
21483 
21484 patch_insn_buf:
21485 		new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
21486 		if (!new_prog)
21487 			return -ENOMEM;
21488 
21489 		delta += cnt - 1;
21490 
21491 		/* keep walking new program and skip insns we just inserted */
21492 		env->prog = new_prog;
21493 		insn      = new_prog->insnsi + i + delta;
21494 	}
21495 
21496 	return 0;
21497 }
21498 
jit_subprogs(struct bpf_verifier_env * env)21499 static int jit_subprogs(struct bpf_verifier_env *env)
21500 {
21501 	struct bpf_prog *prog = env->prog, **func, *tmp;
21502 	int i, j, subprog_start, subprog_end = 0, len, subprog;
21503 	struct bpf_map *map_ptr;
21504 	struct bpf_insn *insn;
21505 	void *old_bpf_func;
21506 	int err, num_exentries;
21507 
21508 	if (env->subprog_cnt <= 1)
21509 		return 0;
21510 
21511 	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
21512 		if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
21513 			continue;
21514 
21515 		/* Upon error here we cannot fall back to interpreter but
21516 		 * need a hard reject of the program. Thus -EFAULT is
21517 		 * propagated in any case.
21518 		 */
21519 		subprog = find_subprog(env, i + insn->imm + 1);
21520 		if (verifier_bug_if(subprog < 0, env, "No program to jit at insn %d",
21521 				    i + insn->imm + 1))
21522 			return -EFAULT;
21523 		/* temporarily remember subprog id inside insn instead of
21524 		 * aux_data, since next loop will split up all insns into funcs
21525 		 */
21526 		insn->off = subprog;
21527 		/* remember original imm in case JIT fails and fallback
21528 		 * to interpreter will be needed
21529 		 */
21530 		env->insn_aux_data[i].call_imm = insn->imm;
21531 		/* point imm to __bpf_call_base+1 from JITs point of view */
21532 		insn->imm = 1;
21533 		if (bpf_pseudo_func(insn)) {
21534 #if defined(MODULES_VADDR)
21535 			u64 addr = MODULES_VADDR;
21536 #else
21537 			u64 addr = VMALLOC_START;
21538 #endif
21539 			/* jit (e.g. x86_64) may emit fewer instructions
21540 			 * if it learns a u32 imm is the same as a u64 imm.
21541 			 * Set close enough to possible prog address.
21542 			 */
21543 			insn[0].imm = (u32)addr;
21544 			insn[1].imm = addr >> 32;
21545 		}
21546 	}
21547 
21548 	err = bpf_prog_alloc_jited_linfo(prog);
21549 	if (err)
21550 		goto out_undo_insn;
21551 
21552 	err = -ENOMEM;
21553 	func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
21554 	if (!func)
21555 		goto out_undo_insn;
21556 
21557 	for (i = 0; i < env->subprog_cnt; i++) {
21558 		subprog_start = subprog_end;
21559 		subprog_end = env->subprog_info[i + 1].start;
21560 
21561 		len = subprog_end - subprog_start;
21562 		/* bpf_prog_run() doesn't call subprogs directly,
21563 		 * hence main prog stats include the runtime of subprogs.
21564 		 * subprogs don't have IDs and not reachable via prog_get_next_id
21565 		 * func[i]->stats will never be accessed and stays NULL
21566 		 */
21567 		func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
21568 		if (!func[i])
21569 			goto out_free;
21570 		memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
21571 		       len * sizeof(struct bpf_insn));
21572 		func[i]->type = prog->type;
21573 		func[i]->len = len;
21574 		if (bpf_prog_calc_tag(func[i]))
21575 			goto out_free;
21576 		func[i]->is_func = 1;
21577 		func[i]->sleepable = prog->sleepable;
21578 		func[i]->aux->func_idx = i;
21579 		/* Below members will be freed only at prog->aux */
21580 		func[i]->aux->btf = prog->aux->btf;
21581 		func[i]->aux->func_info = prog->aux->func_info;
21582 		func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
21583 		func[i]->aux->poke_tab = prog->aux->poke_tab;
21584 		func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
21585 
21586 		for (j = 0; j < prog->aux->size_poke_tab; j++) {
21587 			struct bpf_jit_poke_descriptor *poke;
21588 
21589 			poke = &prog->aux->poke_tab[j];
21590 			if (poke->insn_idx < subprog_end &&
21591 			    poke->insn_idx >= subprog_start)
21592 				poke->aux = func[i]->aux;
21593 		}
21594 
21595 		func[i]->aux->name[0] = 'F';
21596 		func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
21597 		if (env->subprog_info[i].priv_stack_mode == PRIV_STACK_ADAPTIVE)
21598 			func[i]->aux->jits_use_priv_stack = true;
21599 
21600 		func[i]->jit_requested = 1;
21601 		func[i]->blinding_requested = prog->blinding_requested;
21602 		func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
21603 		func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
21604 		func[i]->aux->linfo = prog->aux->linfo;
21605 		func[i]->aux->nr_linfo = prog->aux->nr_linfo;
21606 		func[i]->aux->jited_linfo = prog->aux->jited_linfo;
21607 		func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
21608 		func[i]->aux->arena = prog->aux->arena;
21609 		num_exentries = 0;
21610 		insn = func[i]->insnsi;
21611 		for (j = 0; j < func[i]->len; j++, insn++) {
21612 			if (BPF_CLASS(insn->code) == BPF_LDX &&
21613 			    (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
21614 			     BPF_MODE(insn->code) == BPF_PROBE_MEM32 ||
21615 			     BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
21616 				num_exentries++;
21617 			if ((BPF_CLASS(insn->code) == BPF_STX ||
21618 			     BPF_CLASS(insn->code) == BPF_ST) &&
21619 			     BPF_MODE(insn->code) == BPF_PROBE_MEM32)
21620 				num_exentries++;
21621 			if (BPF_CLASS(insn->code) == BPF_STX &&
21622 			     BPF_MODE(insn->code) == BPF_PROBE_ATOMIC)
21623 				num_exentries++;
21624 		}
21625 		func[i]->aux->num_exentries = num_exentries;
21626 		func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
21627 		func[i]->aux->exception_cb = env->subprog_info[i].is_exception_cb;
21628 		func[i]->aux->changes_pkt_data = env->subprog_info[i].changes_pkt_data;
21629 		func[i]->aux->might_sleep = env->subprog_info[i].might_sleep;
21630 		if (!i)
21631 			func[i]->aux->exception_boundary = env->seen_exception;
21632 		func[i] = bpf_int_jit_compile(func[i]);
21633 		if (!func[i]->jited) {
21634 			err = -ENOTSUPP;
21635 			goto out_free;
21636 		}
21637 		cond_resched();
21638 	}
21639 
21640 	/* at this point all bpf functions were successfully JITed
21641 	 * now populate all bpf_calls with correct addresses and
21642 	 * run last pass of JIT
21643 	 */
21644 	for (i = 0; i < env->subprog_cnt; i++) {
21645 		insn = func[i]->insnsi;
21646 		for (j = 0; j < func[i]->len; j++, insn++) {
21647 			if (bpf_pseudo_func(insn)) {
21648 				subprog = insn->off;
21649 				insn[0].imm = (u32)(long)func[subprog]->bpf_func;
21650 				insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
21651 				continue;
21652 			}
21653 			if (!bpf_pseudo_call(insn))
21654 				continue;
21655 			subprog = insn->off;
21656 			insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
21657 		}
21658 
21659 		/* we use the aux data to keep a list of the start addresses
21660 		 * of the JITed images for each function in the program
21661 		 *
21662 		 * for some architectures, such as powerpc64, the imm field
21663 		 * might not be large enough to hold the offset of the start
21664 		 * address of the callee's JITed image from __bpf_call_base
21665 		 *
21666 		 * in such cases, we can lookup the start address of a callee
21667 		 * by using its subprog id, available from the off field of
21668 		 * the call instruction, as an index for this list
21669 		 */
21670 		func[i]->aux->func = func;
21671 		func[i]->aux->func_cnt = env->subprog_cnt - env->hidden_subprog_cnt;
21672 		func[i]->aux->real_func_cnt = env->subprog_cnt;
21673 	}
21674 	for (i = 0; i < env->subprog_cnt; i++) {
21675 		old_bpf_func = func[i]->bpf_func;
21676 		tmp = bpf_int_jit_compile(func[i]);
21677 		if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
21678 			verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
21679 			err = -ENOTSUPP;
21680 			goto out_free;
21681 		}
21682 		cond_resched();
21683 	}
21684 
21685 	/* finally lock prog and jit images for all functions and
21686 	 * populate kallsysm. Begin at the first subprogram, since
21687 	 * bpf_prog_load will add the kallsyms for the main program.
21688 	 */
21689 	for (i = 1; i < env->subprog_cnt; i++) {
21690 		err = bpf_prog_lock_ro(func[i]);
21691 		if (err)
21692 			goto out_free;
21693 	}
21694 
21695 	for (i = 1; i < env->subprog_cnt; i++)
21696 		bpf_prog_kallsyms_add(func[i]);
21697 
21698 	/* Last step: make now unused interpreter insns from main
21699 	 * prog consistent for later dump requests, so they can
21700 	 * later look the same as if they were interpreted only.
21701 	 */
21702 	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
21703 		if (bpf_pseudo_func(insn)) {
21704 			insn[0].imm = env->insn_aux_data[i].call_imm;
21705 			insn[1].imm = insn->off;
21706 			insn->off = 0;
21707 			continue;
21708 		}
21709 		if (!bpf_pseudo_call(insn))
21710 			continue;
21711 		insn->off = env->insn_aux_data[i].call_imm;
21712 		subprog = find_subprog(env, i + insn->off + 1);
21713 		insn->imm = subprog;
21714 	}
21715 
21716 	prog->jited = 1;
21717 	prog->bpf_func = func[0]->bpf_func;
21718 	prog->jited_len = func[0]->jited_len;
21719 	prog->aux->extable = func[0]->aux->extable;
21720 	prog->aux->num_exentries = func[0]->aux->num_exentries;
21721 	prog->aux->func = func;
21722 	prog->aux->func_cnt = env->subprog_cnt - env->hidden_subprog_cnt;
21723 	prog->aux->real_func_cnt = env->subprog_cnt;
21724 	prog->aux->bpf_exception_cb = (void *)func[env->exception_callback_subprog]->bpf_func;
21725 	prog->aux->exception_boundary = func[0]->aux->exception_boundary;
21726 	bpf_prog_jit_attempt_done(prog);
21727 	return 0;
21728 out_free:
21729 	/* We failed JIT'ing, so at this point we need to unregister poke
21730 	 * descriptors from subprogs, so that kernel is not attempting to
21731 	 * patch it anymore as we're freeing the subprog JIT memory.
21732 	 */
21733 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
21734 		map_ptr = prog->aux->poke_tab[i].tail_call.map;
21735 		map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
21736 	}
21737 	/* At this point we're guaranteed that poke descriptors are not
21738 	 * live anymore. We can just unlink its descriptor table as it's
21739 	 * released with the main prog.
21740 	 */
21741 	for (i = 0; i < env->subprog_cnt; i++) {
21742 		if (!func[i])
21743 			continue;
21744 		func[i]->aux->poke_tab = NULL;
21745 		bpf_jit_free(func[i]);
21746 	}
21747 	kfree(func);
21748 out_undo_insn:
21749 	/* cleanup main prog to be interpreted */
21750 	prog->jit_requested = 0;
21751 	prog->blinding_requested = 0;
21752 	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
21753 		if (!bpf_pseudo_call(insn))
21754 			continue;
21755 		insn->off = 0;
21756 		insn->imm = env->insn_aux_data[i].call_imm;
21757 	}
21758 	bpf_prog_jit_attempt_done(prog);
21759 	return err;
21760 }
21761 
fixup_call_args(struct bpf_verifier_env * env)21762 static int fixup_call_args(struct bpf_verifier_env *env)
21763 {
21764 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
21765 	struct bpf_prog *prog = env->prog;
21766 	struct bpf_insn *insn = prog->insnsi;
21767 	bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
21768 	int i, depth;
21769 #endif
21770 	int err = 0;
21771 
21772 	if (env->prog->jit_requested &&
21773 	    !bpf_prog_is_offloaded(env->prog->aux)) {
21774 		err = jit_subprogs(env);
21775 		if (err == 0)
21776 			return 0;
21777 		if (err == -EFAULT)
21778 			return err;
21779 	}
21780 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
21781 	if (has_kfunc_call) {
21782 		verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
21783 		return -EINVAL;
21784 	}
21785 	if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
21786 		/* When JIT fails the progs with bpf2bpf calls and tail_calls
21787 		 * have to be rejected, since interpreter doesn't support them yet.
21788 		 */
21789 		verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
21790 		return -EINVAL;
21791 	}
21792 	for (i = 0; i < prog->len; i++, insn++) {
21793 		if (bpf_pseudo_func(insn)) {
21794 			/* When JIT fails the progs with callback calls
21795 			 * have to be rejected, since interpreter doesn't support them yet.
21796 			 */
21797 			verbose(env, "callbacks are not allowed in non-JITed programs\n");
21798 			return -EINVAL;
21799 		}
21800 
21801 		if (!bpf_pseudo_call(insn))
21802 			continue;
21803 		depth = get_callee_stack_depth(env, insn, i);
21804 		if (depth < 0)
21805 			return depth;
21806 		bpf_patch_call_args(insn, depth);
21807 	}
21808 	err = 0;
21809 #endif
21810 	return err;
21811 }
21812 
21813 /* replace a generic kfunc with a specialized version if necessary */
specialize_kfunc(struct bpf_verifier_env * env,u32 func_id,u16 offset,unsigned long * addr)21814 static void specialize_kfunc(struct bpf_verifier_env *env,
21815 			     u32 func_id, u16 offset, unsigned long *addr)
21816 {
21817 	struct bpf_prog *prog = env->prog;
21818 	bool seen_direct_write;
21819 	void *xdp_kfunc;
21820 	bool is_rdonly;
21821 
21822 	if (bpf_dev_bound_kfunc_id(func_id)) {
21823 		xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
21824 		if (xdp_kfunc) {
21825 			*addr = (unsigned long)xdp_kfunc;
21826 			return;
21827 		}
21828 		/* fallback to default kfunc when not supported by netdev */
21829 	}
21830 
21831 	if (offset)
21832 		return;
21833 
21834 	if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
21835 		seen_direct_write = env->seen_direct_write;
21836 		is_rdonly = !may_access_direct_pkt_data(env, NULL, BPF_WRITE);
21837 
21838 		if (is_rdonly)
21839 			*addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
21840 
21841 		/* restore env->seen_direct_write to its original value, since
21842 		 * may_access_direct_pkt_data mutates it
21843 		 */
21844 		env->seen_direct_write = seen_direct_write;
21845 	}
21846 
21847 	if (func_id == special_kfunc_list[KF_bpf_set_dentry_xattr] &&
21848 	    bpf_lsm_has_d_inode_locked(prog))
21849 		*addr = (unsigned long)bpf_set_dentry_xattr_locked;
21850 
21851 	if (func_id == special_kfunc_list[KF_bpf_remove_dentry_xattr] &&
21852 	    bpf_lsm_has_d_inode_locked(prog))
21853 		*addr = (unsigned long)bpf_remove_dentry_xattr_locked;
21854 }
21855 
__fixup_collection_insert_kfunc(struct bpf_insn_aux_data * insn_aux,u16 struct_meta_reg,u16 node_offset_reg,struct bpf_insn * insn,struct bpf_insn * insn_buf,int * cnt)21856 static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
21857 					    u16 struct_meta_reg,
21858 					    u16 node_offset_reg,
21859 					    struct bpf_insn *insn,
21860 					    struct bpf_insn *insn_buf,
21861 					    int *cnt)
21862 {
21863 	struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
21864 	struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
21865 
21866 	insn_buf[0] = addr[0];
21867 	insn_buf[1] = addr[1];
21868 	insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
21869 	insn_buf[3] = *insn;
21870 	*cnt = 4;
21871 }
21872 
fixup_kfunc_call(struct bpf_verifier_env * env,struct bpf_insn * insn,struct bpf_insn * insn_buf,int insn_idx,int * cnt)21873 static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
21874 			    struct bpf_insn *insn_buf, int insn_idx, int *cnt)
21875 {
21876 	const struct bpf_kfunc_desc *desc;
21877 
21878 	if (!insn->imm) {
21879 		verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
21880 		return -EINVAL;
21881 	}
21882 
21883 	*cnt = 0;
21884 
21885 	/* insn->imm has the btf func_id. Replace it with an offset relative to
21886 	 * __bpf_call_base, unless the JIT needs to call functions that are
21887 	 * further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
21888 	 */
21889 	desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
21890 	if (!desc) {
21891 		verifier_bug(env, "kernel function descriptor not found for func_id %u",
21892 			     insn->imm);
21893 		return -EFAULT;
21894 	}
21895 
21896 	if (!bpf_jit_supports_far_kfunc_call())
21897 		insn->imm = BPF_CALL_IMM(desc->addr);
21898 	if (insn->off)
21899 		return 0;
21900 	if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl] ||
21901 	    desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
21902 		struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
21903 		struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
21904 		u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
21905 
21906 		if (desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl] && kptr_struct_meta) {
21907 			verifier_bug(env, "NULL kptr_struct_meta expected at insn_idx %d",
21908 				     insn_idx);
21909 			return -EFAULT;
21910 		}
21911 
21912 		insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
21913 		insn_buf[1] = addr[0];
21914 		insn_buf[2] = addr[1];
21915 		insn_buf[3] = *insn;
21916 		*cnt = 4;
21917 	} else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
21918 		   desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_drop_impl] ||
21919 		   desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
21920 		struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
21921 		struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
21922 
21923 		if (desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_drop_impl] && kptr_struct_meta) {
21924 			verifier_bug(env, "NULL kptr_struct_meta expected at insn_idx %d",
21925 				     insn_idx);
21926 			return -EFAULT;
21927 		}
21928 
21929 		if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
21930 		    !kptr_struct_meta) {
21931 			verifier_bug(env, "kptr_struct_meta expected at insn_idx %d",
21932 				     insn_idx);
21933 			return -EFAULT;
21934 		}
21935 
21936 		insn_buf[0] = addr[0];
21937 		insn_buf[1] = addr[1];
21938 		insn_buf[2] = *insn;
21939 		*cnt = 3;
21940 	} else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
21941 		   desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
21942 		   desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
21943 		struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
21944 		int struct_meta_reg = BPF_REG_3;
21945 		int node_offset_reg = BPF_REG_4;
21946 
21947 		/* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
21948 		if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
21949 			struct_meta_reg = BPF_REG_4;
21950 			node_offset_reg = BPF_REG_5;
21951 		}
21952 
21953 		if (!kptr_struct_meta) {
21954 			verifier_bug(env, "kptr_struct_meta expected at insn_idx %d",
21955 				     insn_idx);
21956 			return -EFAULT;
21957 		}
21958 
21959 		__fixup_collection_insert_kfunc(&env->insn_aux_data[insn_idx], struct_meta_reg,
21960 						node_offset_reg, insn, insn_buf, cnt);
21961 	} else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
21962 		   desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
21963 		insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
21964 		*cnt = 1;
21965 	}
21966 
21967 	if (env->insn_aux_data[insn_idx].arg_prog) {
21968 		u32 regno = env->insn_aux_data[insn_idx].arg_prog;
21969 		struct bpf_insn ld_addrs[2] = { BPF_LD_IMM64(regno, (long)env->prog->aux) };
21970 		int idx = *cnt;
21971 
21972 		insn_buf[idx++] = ld_addrs[0];
21973 		insn_buf[idx++] = ld_addrs[1];
21974 		insn_buf[idx++] = *insn;
21975 		*cnt = idx;
21976 	}
21977 	return 0;
21978 }
21979 
21980 /* The function requires that first instruction in 'patch' is insnsi[prog->len - 1] */
add_hidden_subprog(struct bpf_verifier_env * env,struct bpf_insn * patch,int len)21981 static int add_hidden_subprog(struct bpf_verifier_env *env, struct bpf_insn *patch, int len)
21982 {
21983 	struct bpf_subprog_info *info = env->subprog_info;
21984 	int cnt = env->subprog_cnt;
21985 	struct bpf_prog *prog;
21986 
21987 	/* We only reserve one slot for hidden subprogs in subprog_info. */
21988 	if (env->hidden_subprog_cnt) {
21989 		verifier_bug(env, "only one hidden subprog supported");
21990 		return -EFAULT;
21991 	}
21992 	/* We're not patching any existing instruction, just appending the new
21993 	 * ones for the hidden subprog. Hence all of the adjustment operations
21994 	 * in bpf_patch_insn_data are no-ops.
21995 	 */
21996 	prog = bpf_patch_insn_data(env, env->prog->len - 1, patch, len);
21997 	if (!prog)
21998 		return -ENOMEM;
21999 	env->prog = prog;
22000 	info[cnt + 1].start = info[cnt].start;
22001 	info[cnt].start = prog->len - len + 1;
22002 	env->subprog_cnt++;
22003 	env->hidden_subprog_cnt++;
22004 	return 0;
22005 }
22006 
22007 /* Do various post-verification rewrites in a single program pass.
22008  * These rewrites simplify JIT and interpreter implementations.
22009  */
do_misc_fixups(struct bpf_verifier_env * env)22010 static int do_misc_fixups(struct bpf_verifier_env *env)
22011 {
22012 	struct bpf_prog *prog = env->prog;
22013 	enum bpf_attach_type eatype = prog->expected_attach_type;
22014 	enum bpf_prog_type prog_type = resolve_prog_type(prog);
22015 	struct bpf_insn *insn = prog->insnsi;
22016 	const struct bpf_func_proto *fn;
22017 	const int insn_cnt = prog->len;
22018 	const struct bpf_map_ops *ops;
22019 	struct bpf_insn_aux_data *aux;
22020 	struct bpf_insn *insn_buf = env->insn_buf;
22021 	struct bpf_prog *new_prog;
22022 	struct bpf_map *map_ptr;
22023 	int i, ret, cnt, delta = 0, cur_subprog = 0;
22024 	struct bpf_subprog_info *subprogs = env->subprog_info;
22025 	u16 stack_depth = subprogs[cur_subprog].stack_depth;
22026 	u16 stack_depth_extra = 0;
22027 
22028 	if (env->seen_exception && !env->exception_callback_subprog) {
22029 		struct bpf_insn *patch = insn_buf;
22030 
22031 		*patch++ = env->prog->insnsi[insn_cnt - 1];
22032 		*patch++ = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
22033 		*patch++ = BPF_EXIT_INSN();
22034 		ret = add_hidden_subprog(env, insn_buf, patch - insn_buf);
22035 		if (ret < 0)
22036 			return ret;
22037 		prog = env->prog;
22038 		insn = prog->insnsi;
22039 
22040 		env->exception_callback_subprog = env->subprog_cnt - 1;
22041 		/* Don't update insn_cnt, as add_hidden_subprog always appends insns */
22042 		mark_subprog_exc_cb(env, env->exception_callback_subprog);
22043 	}
22044 
22045 	for (i = 0; i < insn_cnt;) {
22046 		if (insn->code == (BPF_ALU64 | BPF_MOV | BPF_X) && insn->imm) {
22047 			if ((insn->off == BPF_ADDR_SPACE_CAST && insn->imm == 1) ||
22048 			    (((struct bpf_map *)env->prog->aux->arena)->map_flags & BPF_F_NO_USER_CONV)) {
22049 				/* convert to 32-bit mov that clears upper 32-bit */
22050 				insn->code = BPF_ALU | BPF_MOV | BPF_X;
22051 				/* clear off and imm, so it's a normal 'wX = wY' from JIT pov */
22052 				insn->off = 0;
22053 				insn->imm = 0;
22054 			} /* cast from as(0) to as(1) should be handled by JIT */
22055 			goto next_insn;
22056 		}
22057 
22058 		if (env->insn_aux_data[i + delta].needs_zext)
22059 			/* Convert BPF_CLASS(insn->code) == BPF_ALU64 to 32-bit ALU */
22060 			insn->code = BPF_ALU | BPF_OP(insn->code) | BPF_SRC(insn->code);
22061 
22062 		/* Make sdiv/smod divide-by-minus-one exceptions impossible. */
22063 		if ((insn->code == (BPF_ALU64 | BPF_MOD | BPF_K) ||
22064 		     insn->code == (BPF_ALU64 | BPF_DIV | BPF_K) ||
22065 		     insn->code == (BPF_ALU | BPF_MOD | BPF_K) ||
22066 		     insn->code == (BPF_ALU | BPF_DIV | BPF_K)) &&
22067 		    insn->off == 1 && insn->imm == -1) {
22068 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
22069 			bool isdiv = BPF_OP(insn->code) == BPF_DIV;
22070 			struct bpf_insn *patch = insn_buf;
22071 
22072 			if (isdiv)
22073 				*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22074 							BPF_NEG | BPF_K, insn->dst_reg,
22075 							0, 0, 0);
22076 			else
22077 				*patch++ = BPF_MOV32_IMM(insn->dst_reg, 0);
22078 
22079 			cnt = patch - insn_buf;
22080 
22081 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22082 			if (!new_prog)
22083 				return -ENOMEM;
22084 
22085 			delta    += cnt - 1;
22086 			env->prog = prog = new_prog;
22087 			insn      = new_prog->insnsi + i + delta;
22088 			goto next_insn;
22089 		}
22090 
22091 		/* Make divide-by-zero and divide-by-minus-one exceptions impossible. */
22092 		if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
22093 		    insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
22094 		    insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
22095 		    insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
22096 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
22097 			bool isdiv = BPF_OP(insn->code) == BPF_DIV;
22098 			bool is_sdiv = isdiv && insn->off == 1;
22099 			bool is_smod = !isdiv && insn->off == 1;
22100 			struct bpf_insn *patch = insn_buf;
22101 
22102 			if (is_sdiv) {
22103 				/* [R,W]x sdiv 0 -> 0
22104 				 * LLONG_MIN sdiv -1 -> LLONG_MIN
22105 				 * INT_MIN sdiv -1 -> INT_MIN
22106 				 */
22107 				*patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg);
22108 				*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22109 							BPF_ADD | BPF_K, BPF_REG_AX,
22110 							0, 0, 1);
22111 				*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22112 							BPF_JGT | BPF_K, BPF_REG_AX,
22113 							0, 4, 1);
22114 				*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22115 							BPF_JEQ | BPF_K, BPF_REG_AX,
22116 							0, 1, 0);
22117 				*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22118 							BPF_MOV | BPF_K, insn->dst_reg,
22119 							0, 0, 0);
22120 				/* BPF_NEG(LLONG_MIN) == -LLONG_MIN == LLONG_MIN */
22121 				*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22122 							BPF_NEG | BPF_K, insn->dst_reg,
22123 							0, 0, 0);
22124 				*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22125 				*patch++ = *insn;
22126 				cnt = patch - insn_buf;
22127 			} else if (is_smod) {
22128 				/* [R,W]x mod 0 -> [R,W]x */
22129 				/* [R,W]x mod -1 -> 0 */
22130 				*patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg);
22131 				*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22132 							BPF_ADD | BPF_K, BPF_REG_AX,
22133 							0, 0, 1);
22134 				*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22135 							BPF_JGT | BPF_K, BPF_REG_AX,
22136 							0, 3, 1);
22137 				*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22138 							BPF_JEQ | BPF_K, BPF_REG_AX,
22139 							0, 3 + (is64 ? 0 : 1), 1);
22140 				*patch++ = BPF_MOV32_IMM(insn->dst_reg, 0);
22141 				*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22142 				*patch++ = *insn;
22143 
22144 				if (!is64) {
22145 					*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22146 					*patch++ = BPF_MOV32_REG(insn->dst_reg, insn->dst_reg);
22147 				}
22148 				cnt = patch - insn_buf;
22149 			} else if (isdiv) {
22150 				/* [R,W]x div 0 -> 0 */
22151 				*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22152 							BPF_JNE | BPF_K, insn->src_reg,
22153 							0, 2, 0);
22154 				*patch++ = BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg);
22155 				*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22156 				*patch++ = *insn;
22157 				cnt = patch - insn_buf;
22158 			} else {
22159 				/* [R,W]x mod 0 -> [R,W]x */
22160 				*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22161 							BPF_JEQ | BPF_K, insn->src_reg,
22162 							0, 1 + (is64 ? 0 : 1), 0);
22163 				*patch++ = *insn;
22164 
22165 				if (!is64) {
22166 					*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22167 					*patch++ = BPF_MOV32_REG(insn->dst_reg, insn->dst_reg);
22168 				}
22169 				cnt = patch - insn_buf;
22170 			}
22171 
22172 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22173 			if (!new_prog)
22174 				return -ENOMEM;
22175 
22176 			delta    += cnt - 1;
22177 			env->prog = prog = new_prog;
22178 			insn      = new_prog->insnsi + i + delta;
22179 			goto next_insn;
22180 		}
22181 
22182 		/* Make it impossible to de-reference a userspace address */
22183 		if (BPF_CLASS(insn->code) == BPF_LDX &&
22184 		    (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
22185 		     BPF_MODE(insn->code) == BPF_PROBE_MEMSX)) {
22186 			struct bpf_insn *patch = insn_buf;
22187 			u64 uaddress_limit = bpf_arch_uaddress_limit();
22188 
22189 			if (!uaddress_limit)
22190 				goto next_insn;
22191 
22192 			*patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg);
22193 			if (insn->off)
22194 				*patch++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_AX, insn->off);
22195 			*patch++ = BPF_ALU64_IMM(BPF_RSH, BPF_REG_AX, 32);
22196 			*patch++ = BPF_JMP_IMM(BPF_JLE, BPF_REG_AX, uaddress_limit >> 32, 2);
22197 			*patch++ = *insn;
22198 			*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22199 			*patch++ = BPF_MOV64_IMM(insn->dst_reg, 0);
22200 
22201 			cnt = patch - insn_buf;
22202 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22203 			if (!new_prog)
22204 				return -ENOMEM;
22205 
22206 			delta    += cnt - 1;
22207 			env->prog = prog = new_prog;
22208 			insn      = new_prog->insnsi + i + delta;
22209 			goto next_insn;
22210 		}
22211 
22212 		/* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
22213 		if (BPF_CLASS(insn->code) == BPF_LD &&
22214 		    (BPF_MODE(insn->code) == BPF_ABS ||
22215 		     BPF_MODE(insn->code) == BPF_IND)) {
22216 			cnt = env->ops->gen_ld_abs(insn, insn_buf);
22217 			if (cnt == 0 || cnt >= INSN_BUF_SIZE) {
22218 				verifier_bug(env, "%d insns generated for ld_abs", cnt);
22219 				return -EFAULT;
22220 			}
22221 
22222 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22223 			if (!new_prog)
22224 				return -ENOMEM;
22225 
22226 			delta    += cnt - 1;
22227 			env->prog = prog = new_prog;
22228 			insn      = new_prog->insnsi + i + delta;
22229 			goto next_insn;
22230 		}
22231 
22232 		/* Rewrite pointer arithmetic to mitigate speculation attacks. */
22233 		if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
22234 		    insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
22235 			const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
22236 			const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
22237 			struct bpf_insn *patch = insn_buf;
22238 			bool issrc, isneg, isimm;
22239 			u32 off_reg;
22240 
22241 			aux = &env->insn_aux_data[i + delta];
22242 			if (!aux->alu_state ||
22243 			    aux->alu_state == BPF_ALU_NON_POINTER)
22244 				goto next_insn;
22245 
22246 			isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
22247 			issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
22248 				BPF_ALU_SANITIZE_SRC;
22249 			isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
22250 
22251 			off_reg = issrc ? insn->src_reg : insn->dst_reg;
22252 			if (isimm) {
22253 				*patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
22254 			} else {
22255 				if (isneg)
22256 					*patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
22257 				*patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
22258 				*patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
22259 				*patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
22260 				*patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
22261 				*patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
22262 				*patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
22263 			}
22264 			if (!issrc)
22265 				*patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
22266 			insn->src_reg = BPF_REG_AX;
22267 			if (isneg)
22268 				insn->code = insn->code == code_add ?
22269 					     code_sub : code_add;
22270 			*patch++ = *insn;
22271 			if (issrc && isneg && !isimm)
22272 				*patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
22273 			cnt = patch - insn_buf;
22274 
22275 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22276 			if (!new_prog)
22277 				return -ENOMEM;
22278 
22279 			delta    += cnt - 1;
22280 			env->prog = prog = new_prog;
22281 			insn      = new_prog->insnsi + i + delta;
22282 			goto next_insn;
22283 		}
22284 
22285 		if (is_may_goto_insn(insn) && bpf_jit_supports_timed_may_goto()) {
22286 			int stack_off_cnt = -stack_depth - 16;
22287 
22288 			/*
22289 			 * Two 8 byte slots, depth-16 stores the count, and
22290 			 * depth-8 stores the start timestamp of the loop.
22291 			 *
22292 			 * The starting value of count is BPF_MAX_TIMED_LOOPS
22293 			 * (0xffff).  Every iteration loads it and subs it by 1,
22294 			 * until the value becomes 0 in AX (thus, 1 in stack),
22295 			 * after which we call arch_bpf_timed_may_goto, which
22296 			 * either sets AX to 0xffff to keep looping, or to 0
22297 			 * upon timeout. AX is then stored into the stack. In
22298 			 * the next iteration, we either see 0 and break out, or
22299 			 * continue iterating until the next time value is 0
22300 			 * after subtraction, rinse and repeat.
22301 			 */
22302 			stack_depth_extra = 16;
22303 			insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_AX, BPF_REG_10, stack_off_cnt);
22304 			if (insn->off >= 0)
22305 				insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off + 5);
22306 			else
22307 				insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off - 1);
22308 			insn_buf[2] = BPF_ALU64_IMM(BPF_SUB, BPF_REG_AX, 1);
22309 			insn_buf[3] = BPF_JMP_IMM(BPF_JNE, BPF_REG_AX, 0, 2);
22310 			/*
22311 			 * AX is used as an argument to pass in stack_off_cnt
22312 			 * (to add to r10/fp), and also as the return value of
22313 			 * the call to arch_bpf_timed_may_goto.
22314 			 */
22315 			insn_buf[4] = BPF_MOV64_IMM(BPF_REG_AX, stack_off_cnt);
22316 			insn_buf[5] = BPF_EMIT_CALL(arch_bpf_timed_may_goto);
22317 			insn_buf[6] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_AX, stack_off_cnt);
22318 			cnt = 7;
22319 
22320 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22321 			if (!new_prog)
22322 				return -ENOMEM;
22323 
22324 			delta += cnt - 1;
22325 			env->prog = prog = new_prog;
22326 			insn = new_prog->insnsi + i + delta;
22327 			goto next_insn;
22328 		} else if (is_may_goto_insn(insn)) {
22329 			int stack_off = -stack_depth - 8;
22330 
22331 			stack_depth_extra = 8;
22332 			insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_AX, BPF_REG_10, stack_off);
22333 			if (insn->off >= 0)
22334 				insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off + 2);
22335 			else
22336 				insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off - 1);
22337 			insn_buf[2] = BPF_ALU64_IMM(BPF_SUB, BPF_REG_AX, 1);
22338 			insn_buf[3] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_AX, stack_off);
22339 			cnt = 4;
22340 
22341 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22342 			if (!new_prog)
22343 				return -ENOMEM;
22344 
22345 			delta += cnt - 1;
22346 			env->prog = prog = new_prog;
22347 			insn = new_prog->insnsi + i + delta;
22348 			goto next_insn;
22349 		}
22350 
22351 		if (insn->code != (BPF_JMP | BPF_CALL))
22352 			goto next_insn;
22353 		if (insn->src_reg == BPF_PSEUDO_CALL)
22354 			goto next_insn;
22355 		if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
22356 			ret = fixup_kfunc_call(env, insn, insn_buf, i + delta, &cnt);
22357 			if (ret)
22358 				return ret;
22359 			if (cnt == 0)
22360 				goto next_insn;
22361 
22362 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22363 			if (!new_prog)
22364 				return -ENOMEM;
22365 
22366 			delta	 += cnt - 1;
22367 			env->prog = prog = new_prog;
22368 			insn	  = new_prog->insnsi + i + delta;
22369 			goto next_insn;
22370 		}
22371 
22372 		/* Skip inlining the helper call if the JIT does it. */
22373 		if (bpf_jit_inlines_helper_call(insn->imm))
22374 			goto next_insn;
22375 
22376 		if (insn->imm == BPF_FUNC_get_route_realm)
22377 			prog->dst_needed = 1;
22378 		if (insn->imm == BPF_FUNC_get_prandom_u32)
22379 			bpf_user_rnd_init_once();
22380 		if (insn->imm == BPF_FUNC_override_return)
22381 			prog->kprobe_override = 1;
22382 		if (insn->imm == BPF_FUNC_tail_call) {
22383 			/* If we tail call into other programs, we
22384 			 * cannot make any assumptions since they can
22385 			 * be replaced dynamically during runtime in
22386 			 * the program array.
22387 			 */
22388 			prog->cb_access = 1;
22389 			if (!allow_tail_call_in_subprogs(env))
22390 				prog->aux->stack_depth = MAX_BPF_STACK;
22391 			prog->aux->max_pkt_offset = MAX_PACKET_OFF;
22392 
22393 			/* mark bpf_tail_call as different opcode to avoid
22394 			 * conditional branch in the interpreter for every normal
22395 			 * call and to prevent accidental JITing by JIT compiler
22396 			 * that doesn't support bpf_tail_call yet
22397 			 */
22398 			insn->imm = 0;
22399 			insn->code = BPF_JMP | BPF_TAIL_CALL;
22400 
22401 			aux = &env->insn_aux_data[i + delta];
22402 			if (env->bpf_capable && !prog->blinding_requested &&
22403 			    prog->jit_requested &&
22404 			    !bpf_map_key_poisoned(aux) &&
22405 			    !bpf_map_ptr_poisoned(aux) &&
22406 			    !bpf_map_ptr_unpriv(aux)) {
22407 				struct bpf_jit_poke_descriptor desc = {
22408 					.reason = BPF_POKE_REASON_TAIL_CALL,
22409 					.tail_call.map = aux->map_ptr_state.map_ptr,
22410 					.tail_call.key = bpf_map_key_immediate(aux),
22411 					.insn_idx = i + delta,
22412 				};
22413 
22414 				ret = bpf_jit_add_poke_descriptor(prog, &desc);
22415 				if (ret < 0) {
22416 					verbose(env, "adding tail call poke descriptor failed\n");
22417 					return ret;
22418 				}
22419 
22420 				insn->imm = ret + 1;
22421 				goto next_insn;
22422 			}
22423 
22424 			if (!bpf_map_ptr_unpriv(aux))
22425 				goto next_insn;
22426 
22427 			/* instead of changing every JIT dealing with tail_call
22428 			 * emit two extra insns:
22429 			 * if (index >= max_entries) goto out;
22430 			 * index &= array->index_mask;
22431 			 * to avoid out-of-bounds cpu speculation
22432 			 */
22433 			if (bpf_map_ptr_poisoned(aux)) {
22434 				verbose(env, "tail_call abusing map_ptr\n");
22435 				return -EINVAL;
22436 			}
22437 
22438 			map_ptr = aux->map_ptr_state.map_ptr;
22439 			insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
22440 						  map_ptr->max_entries, 2);
22441 			insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
22442 						    container_of(map_ptr,
22443 								 struct bpf_array,
22444 								 map)->index_mask);
22445 			insn_buf[2] = *insn;
22446 			cnt = 3;
22447 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22448 			if (!new_prog)
22449 				return -ENOMEM;
22450 
22451 			delta    += cnt - 1;
22452 			env->prog = prog = new_prog;
22453 			insn      = new_prog->insnsi + i + delta;
22454 			goto next_insn;
22455 		}
22456 
22457 		if (insn->imm == BPF_FUNC_timer_set_callback) {
22458 			/* The verifier will process callback_fn as many times as necessary
22459 			 * with different maps and the register states prepared by
22460 			 * set_timer_callback_state will be accurate.
22461 			 *
22462 			 * The following use case is valid:
22463 			 *   map1 is shared by prog1, prog2, prog3.
22464 			 *   prog1 calls bpf_timer_init for some map1 elements
22465 			 *   prog2 calls bpf_timer_set_callback for some map1 elements.
22466 			 *     Those that were not bpf_timer_init-ed will return -EINVAL.
22467 			 *   prog3 calls bpf_timer_start for some map1 elements.
22468 			 *     Those that were not both bpf_timer_init-ed and
22469 			 *     bpf_timer_set_callback-ed will return -EINVAL.
22470 			 */
22471 			struct bpf_insn ld_addrs[2] = {
22472 				BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
22473 			};
22474 
22475 			insn_buf[0] = ld_addrs[0];
22476 			insn_buf[1] = ld_addrs[1];
22477 			insn_buf[2] = *insn;
22478 			cnt = 3;
22479 
22480 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22481 			if (!new_prog)
22482 				return -ENOMEM;
22483 
22484 			delta    += cnt - 1;
22485 			env->prog = prog = new_prog;
22486 			insn      = new_prog->insnsi + i + delta;
22487 			goto patch_call_imm;
22488 		}
22489 
22490 		if (is_storage_get_function(insn->imm)) {
22491 			if (!in_sleepable(env) ||
22492 			    env->insn_aux_data[i + delta].storage_get_func_atomic)
22493 				insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
22494 			else
22495 				insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
22496 			insn_buf[1] = *insn;
22497 			cnt = 2;
22498 
22499 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22500 			if (!new_prog)
22501 				return -ENOMEM;
22502 
22503 			delta += cnt - 1;
22504 			env->prog = prog = new_prog;
22505 			insn = new_prog->insnsi + i + delta;
22506 			goto patch_call_imm;
22507 		}
22508 
22509 		/* bpf_per_cpu_ptr() and bpf_this_cpu_ptr() */
22510 		if (env->insn_aux_data[i + delta].call_with_percpu_alloc_ptr) {
22511 			/* patch with 'r1 = *(u64 *)(r1 + 0)' since for percpu data,
22512 			 * bpf_mem_alloc() returns a ptr to the percpu data ptr.
22513 			 */
22514 			insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_1, BPF_REG_1, 0);
22515 			insn_buf[1] = *insn;
22516 			cnt = 2;
22517 
22518 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22519 			if (!new_prog)
22520 				return -ENOMEM;
22521 
22522 			delta += cnt - 1;
22523 			env->prog = prog = new_prog;
22524 			insn = new_prog->insnsi + i + delta;
22525 			goto patch_call_imm;
22526 		}
22527 
22528 		/* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
22529 		 * and other inlining handlers are currently limited to 64 bit
22530 		 * only.
22531 		 */
22532 		if (prog->jit_requested && BITS_PER_LONG == 64 &&
22533 		    (insn->imm == BPF_FUNC_map_lookup_elem ||
22534 		     insn->imm == BPF_FUNC_map_update_elem ||
22535 		     insn->imm == BPF_FUNC_map_delete_elem ||
22536 		     insn->imm == BPF_FUNC_map_push_elem   ||
22537 		     insn->imm == BPF_FUNC_map_pop_elem    ||
22538 		     insn->imm == BPF_FUNC_map_peek_elem   ||
22539 		     insn->imm == BPF_FUNC_redirect_map    ||
22540 		     insn->imm == BPF_FUNC_for_each_map_elem ||
22541 		     insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
22542 			aux = &env->insn_aux_data[i + delta];
22543 			if (bpf_map_ptr_poisoned(aux))
22544 				goto patch_call_imm;
22545 
22546 			map_ptr = aux->map_ptr_state.map_ptr;
22547 			ops = map_ptr->ops;
22548 			if (insn->imm == BPF_FUNC_map_lookup_elem &&
22549 			    ops->map_gen_lookup) {
22550 				cnt = ops->map_gen_lookup(map_ptr, insn_buf);
22551 				if (cnt == -EOPNOTSUPP)
22552 					goto patch_map_ops_generic;
22553 				if (cnt <= 0 || cnt >= INSN_BUF_SIZE) {
22554 					verifier_bug(env, "%d insns generated for map lookup", cnt);
22555 					return -EFAULT;
22556 				}
22557 
22558 				new_prog = bpf_patch_insn_data(env, i + delta,
22559 							       insn_buf, cnt);
22560 				if (!new_prog)
22561 					return -ENOMEM;
22562 
22563 				delta    += cnt - 1;
22564 				env->prog = prog = new_prog;
22565 				insn      = new_prog->insnsi + i + delta;
22566 				goto next_insn;
22567 			}
22568 
22569 			BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
22570 				     (void *(*)(struct bpf_map *map, void *key))NULL));
22571 			BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
22572 				     (long (*)(struct bpf_map *map, void *key))NULL));
22573 			BUILD_BUG_ON(!__same_type(ops->map_update_elem,
22574 				     (long (*)(struct bpf_map *map, void *key, void *value,
22575 					      u64 flags))NULL));
22576 			BUILD_BUG_ON(!__same_type(ops->map_push_elem,
22577 				     (long (*)(struct bpf_map *map, void *value,
22578 					      u64 flags))NULL));
22579 			BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
22580 				     (long (*)(struct bpf_map *map, void *value))NULL));
22581 			BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
22582 				     (long (*)(struct bpf_map *map, void *value))NULL));
22583 			BUILD_BUG_ON(!__same_type(ops->map_redirect,
22584 				     (long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
22585 			BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
22586 				     (long (*)(struct bpf_map *map,
22587 					      bpf_callback_t callback_fn,
22588 					      void *callback_ctx,
22589 					      u64 flags))NULL));
22590 			BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
22591 				     (void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
22592 
22593 patch_map_ops_generic:
22594 			switch (insn->imm) {
22595 			case BPF_FUNC_map_lookup_elem:
22596 				insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
22597 				goto next_insn;
22598 			case BPF_FUNC_map_update_elem:
22599 				insn->imm = BPF_CALL_IMM(ops->map_update_elem);
22600 				goto next_insn;
22601 			case BPF_FUNC_map_delete_elem:
22602 				insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
22603 				goto next_insn;
22604 			case BPF_FUNC_map_push_elem:
22605 				insn->imm = BPF_CALL_IMM(ops->map_push_elem);
22606 				goto next_insn;
22607 			case BPF_FUNC_map_pop_elem:
22608 				insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
22609 				goto next_insn;
22610 			case BPF_FUNC_map_peek_elem:
22611 				insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
22612 				goto next_insn;
22613 			case BPF_FUNC_redirect_map:
22614 				insn->imm = BPF_CALL_IMM(ops->map_redirect);
22615 				goto next_insn;
22616 			case BPF_FUNC_for_each_map_elem:
22617 				insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
22618 				goto next_insn;
22619 			case BPF_FUNC_map_lookup_percpu_elem:
22620 				insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
22621 				goto next_insn;
22622 			}
22623 
22624 			goto patch_call_imm;
22625 		}
22626 
22627 		/* Implement bpf_jiffies64 inline. */
22628 		if (prog->jit_requested && BITS_PER_LONG == 64 &&
22629 		    insn->imm == BPF_FUNC_jiffies64) {
22630 			struct bpf_insn ld_jiffies_addr[2] = {
22631 				BPF_LD_IMM64(BPF_REG_0,
22632 					     (unsigned long)&jiffies),
22633 			};
22634 
22635 			insn_buf[0] = ld_jiffies_addr[0];
22636 			insn_buf[1] = ld_jiffies_addr[1];
22637 			insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
22638 						  BPF_REG_0, 0);
22639 			cnt = 3;
22640 
22641 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
22642 						       cnt);
22643 			if (!new_prog)
22644 				return -ENOMEM;
22645 
22646 			delta    += cnt - 1;
22647 			env->prog = prog = new_prog;
22648 			insn      = new_prog->insnsi + i + delta;
22649 			goto next_insn;
22650 		}
22651 
22652 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
22653 		/* Implement bpf_get_smp_processor_id() inline. */
22654 		if (insn->imm == BPF_FUNC_get_smp_processor_id &&
22655 		    verifier_inlines_helper_call(env, insn->imm)) {
22656 			/* BPF_FUNC_get_smp_processor_id inlining is an
22657 			 * optimization, so if cpu_number is ever
22658 			 * changed in some incompatible and hard to support
22659 			 * way, it's fine to back out this inlining logic
22660 			 */
22661 #ifdef CONFIG_SMP
22662 			insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, (u32)(unsigned long)&cpu_number);
22663 			insn_buf[1] = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0);
22664 			insn_buf[2] = BPF_LDX_MEM(BPF_W, BPF_REG_0, BPF_REG_0, 0);
22665 			cnt = 3;
22666 #else
22667 			insn_buf[0] = BPF_ALU32_REG(BPF_XOR, BPF_REG_0, BPF_REG_0);
22668 			cnt = 1;
22669 #endif
22670 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22671 			if (!new_prog)
22672 				return -ENOMEM;
22673 
22674 			delta    += cnt - 1;
22675 			env->prog = prog = new_prog;
22676 			insn      = new_prog->insnsi + i + delta;
22677 			goto next_insn;
22678 		}
22679 #endif
22680 		/* Implement bpf_get_func_arg inline. */
22681 		if (prog_type == BPF_PROG_TYPE_TRACING &&
22682 		    insn->imm == BPF_FUNC_get_func_arg) {
22683 			/* Load nr_args from ctx - 8 */
22684 			insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
22685 			insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
22686 			insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
22687 			insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
22688 			insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
22689 			insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
22690 			insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
22691 			insn_buf[7] = BPF_JMP_A(1);
22692 			insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
22693 			cnt = 9;
22694 
22695 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22696 			if (!new_prog)
22697 				return -ENOMEM;
22698 
22699 			delta    += cnt - 1;
22700 			env->prog = prog = new_prog;
22701 			insn      = new_prog->insnsi + i + delta;
22702 			goto next_insn;
22703 		}
22704 
22705 		/* Implement bpf_get_func_ret inline. */
22706 		if (prog_type == BPF_PROG_TYPE_TRACING &&
22707 		    insn->imm == BPF_FUNC_get_func_ret) {
22708 			if (eatype == BPF_TRACE_FEXIT ||
22709 			    eatype == BPF_MODIFY_RETURN) {
22710 				/* Load nr_args from ctx - 8 */
22711 				insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
22712 				insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
22713 				insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
22714 				insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
22715 				insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
22716 				insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
22717 				cnt = 6;
22718 			} else {
22719 				insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
22720 				cnt = 1;
22721 			}
22722 
22723 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22724 			if (!new_prog)
22725 				return -ENOMEM;
22726 
22727 			delta    += cnt - 1;
22728 			env->prog = prog = new_prog;
22729 			insn      = new_prog->insnsi + i + delta;
22730 			goto next_insn;
22731 		}
22732 
22733 		/* Implement get_func_arg_cnt inline. */
22734 		if (prog_type == BPF_PROG_TYPE_TRACING &&
22735 		    insn->imm == BPF_FUNC_get_func_arg_cnt) {
22736 			/* Load nr_args from ctx - 8 */
22737 			insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
22738 
22739 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
22740 			if (!new_prog)
22741 				return -ENOMEM;
22742 
22743 			env->prog = prog = new_prog;
22744 			insn      = new_prog->insnsi + i + delta;
22745 			goto next_insn;
22746 		}
22747 
22748 		/* Implement bpf_get_func_ip inline. */
22749 		if (prog_type == BPF_PROG_TYPE_TRACING &&
22750 		    insn->imm == BPF_FUNC_get_func_ip) {
22751 			/* Load IP address from ctx - 16 */
22752 			insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
22753 
22754 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
22755 			if (!new_prog)
22756 				return -ENOMEM;
22757 
22758 			env->prog = prog = new_prog;
22759 			insn      = new_prog->insnsi + i + delta;
22760 			goto next_insn;
22761 		}
22762 
22763 		/* Implement bpf_get_branch_snapshot inline. */
22764 		if (IS_ENABLED(CONFIG_PERF_EVENTS) &&
22765 		    prog->jit_requested && BITS_PER_LONG == 64 &&
22766 		    insn->imm == BPF_FUNC_get_branch_snapshot) {
22767 			/* We are dealing with the following func protos:
22768 			 * u64 bpf_get_branch_snapshot(void *buf, u32 size, u64 flags);
22769 			 * int perf_snapshot_branch_stack(struct perf_branch_entry *entries, u32 cnt);
22770 			 */
22771 			const u32 br_entry_size = sizeof(struct perf_branch_entry);
22772 
22773 			/* struct perf_branch_entry is part of UAPI and is
22774 			 * used as an array element, so extremely unlikely to
22775 			 * ever grow or shrink
22776 			 */
22777 			BUILD_BUG_ON(br_entry_size != 24);
22778 
22779 			/* if (unlikely(flags)) return -EINVAL */
22780 			insn_buf[0] = BPF_JMP_IMM(BPF_JNE, BPF_REG_3, 0, 7);
22781 
22782 			/* Transform size (bytes) into number of entries (cnt = size / 24).
22783 			 * But to avoid expensive division instruction, we implement
22784 			 * divide-by-3 through multiplication, followed by further
22785 			 * division by 8 through 3-bit right shift.
22786 			 * Refer to book "Hacker's Delight, 2nd ed." by Henry S. Warren, Jr.,
22787 			 * p. 227, chapter "Unsigned Division by 3" for details and proofs.
22788 			 *
22789 			 * N / 3 <=> M * N / 2^33, where M = (2^33 + 1) / 3 = 0xaaaaaaab.
22790 			 */
22791 			insn_buf[1] = BPF_MOV32_IMM(BPF_REG_0, 0xaaaaaaab);
22792 			insn_buf[2] = BPF_ALU64_REG(BPF_MUL, BPF_REG_2, BPF_REG_0);
22793 			insn_buf[3] = BPF_ALU64_IMM(BPF_RSH, BPF_REG_2, 36);
22794 
22795 			/* call perf_snapshot_branch_stack implementation */
22796 			insn_buf[4] = BPF_EMIT_CALL(static_call_query(perf_snapshot_branch_stack));
22797 			/* if (entry_cnt == 0) return -ENOENT */
22798 			insn_buf[5] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 4);
22799 			/* return entry_cnt * sizeof(struct perf_branch_entry) */
22800 			insn_buf[6] = BPF_ALU32_IMM(BPF_MUL, BPF_REG_0, br_entry_size);
22801 			insn_buf[7] = BPF_JMP_A(3);
22802 			/* return -EINVAL; */
22803 			insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
22804 			insn_buf[9] = BPF_JMP_A(1);
22805 			/* return -ENOENT; */
22806 			insn_buf[10] = BPF_MOV64_IMM(BPF_REG_0, -ENOENT);
22807 			cnt = 11;
22808 
22809 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22810 			if (!new_prog)
22811 				return -ENOMEM;
22812 
22813 			delta    += cnt - 1;
22814 			env->prog = prog = new_prog;
22815 			insn      = new_prog->insnsi + i + delta;
22816 			goto next_insn;
22817 		}
22818 
22819 		/* Implement bpf_kptr_xchg inline */
22820 		if (prog->jit_requested && BITS_PER_LONG == 64 &&
22821 		    insn->imm == BPF_FUNC_kptr_xchg &&
22822 		    bpf_jit_supports_ptr_xchg()) {
22823 			insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_2);
22824 			insn_buf[1] = BPF_ATOMIC_OP(BPF_DW, BPF_XCHG, BPF_REG_1, BPF_REG_0, 0);
22825 			cnt = 2;
22826 
22827 			new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
22828 			if (!new_prog)
22829 				return -ENOMEM;
22830 
22831 			delta    += cnt - 1;
22832 			env->prog = prog = new_prog;
22833 			insn      = new_prog->insnsi + i + delta;
22834 			goto next_insn;
22835 		}
22836 patch_call_imm:
22837 		fn = env->ops->get_func_proto(insn->imm, env->prog);
22838 		/* all functions that have prototype and verifier allowed
22839 		 * programs to call them, must be real in-kernel functions
22840 		 */
22841 		if (!fn->func) {
22842 			verifier_bug(env,
22843 				     "not inlined functions %s#%d is missing func",
22844 				     func_id_name(insn->imm), insn->imm);
22845 			return -EFAULT;
22846 		}
22847 		insn->imm = fn->func - __bpf_call_base;
22848 next_insn:
22849 		if (subprogs[cur_subprog + 1].start == i + delta + 1) {
22850 			subprogs[cur_subprog].stack_depth += stack_depth_extra;
22851 			subprogs[cur_subprog].stack_extra = stack_depth_extra;
22852 
22853 			stack_depth = subprogs[cur_subprog].stack_depth;
22854 			if (stack_depth > MAX_BPF_STACK && !prog->jit_requested) {
22855 				verbose(env, "stack size %d(extra %d) is too large\n",
22856 					stack_depth, stack_depth_extra);
22857 				return -EINVAL;
22858 			}
22859 			cur_subprog++;
22860 			stack_depth = subprogs[cur_subprog].stack_depth;
22861 			stack_depth_extra = 0;
22862 		}
22863 		i++;
22864 		insn++;
22865 	}
22866 
22867 	env->prog->aux->stack_depth = subprogs[0].stack_depth;
22868 	for (i = 0; i < env->subprog_cnt; i++) {
22869 		int delta = bpf_jit_supports_timed_may_goto() ? 2 : 1;
22870 		int subprog_start = subprogs[i].start;
22871 		int stack_slots = subprogs[i].stack_extra / 8;
22872 		int slots = delta, cnt = 0;
22873 
22874 		if (!stack_slots)
22875 			continue;
22876 		/* We need two slots in case timed may_goto is supported. */
22877 		if (stack_slots > slots) {
22878 			verifier_bug(env, "stack_slots supports may_goto only");
22879 			return -EFAULT;
22880 		}
22881 
22882 		stack_depth = subprogs[i].stack_depth;
22883 		if (bpf_jit_supports_timed_may_goto()) {
22884 			insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth,
22885 						     BPF_MAX_TIMED_LOOPS);
22886 			insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth + 8, 0);
22887 		} else {
22888 			/* Add ST insn to subprog prologue to init extra stack */
22889 			insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth,
22890 						     BPF_MAX_LOOPS);
22891 		}
22892 		/* Copy first actual insn to preserve it */
22893 		insn_buf[cnt++] = env->prog->insnsi[subprog_start];
22894 
22895 		new_prog = bpf_patch_insn_data(env, subprog_start, insn_buf, cnt);
22896 		if (!new_prog)
22897 			return -ENOMEM;
22898 		env->prog = prog = new_prog;
22899 		/*
22900 		 * If may_goto is a first insn of a prog there could be a jmp
22901 		 * insn that points to it, hence adjust all such jmps to point
22902 		 * to insn after BPF_ST that inits may_goto count.
22903 		 * Adjustment will succeed because bpf_patch_insn_data() didn't fail.
22904 		 */
22905 		WARN_ON(adjust_jmp_off(env->prog, subprog_start, delta));
22906 	}
22907 
22908 	/* Since poke tab is now finalized, publish aux to tracker. */
22909 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
22910 		map_ptr = prog->aux->poke_tab[i].tail_call.map;
22911 		if (!map_ptr->ops->map_poke_track ||
22912 		    !map_ptr->ops->map_poke_untrack ||
22913 		    !map_ptr->ops->map_poke_run) {
22914 			verifier_bug(env, "poke tab is misconfigured");
22915 			return -EFAULT;
22916 		}
22917 
22918 		ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
22919 		if (ret < 0) {
22920 			verbose(env, "tracking tail call prog failed\n");
22921 			return ret;
22922 		}
22923 	}
22924 
22925 	sort_kfunc_descs_by_imm_off(env->prog);
22926 
22927 	return 0;
22928 }
22929 
inline_bpf_loop(struct bpf_verifier_env * env,int position,s32 stack_base,u32 callback_subprogno,u32 * total_cnt)22930 static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
22931 					int position,
22932 					s32 stack_base,
22933 					u32 callback_subprogno,
22934 					u32 *total_cnt)
22935 {
22936 	s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
22937 	s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
22938 	s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
22939 	int reg_loop_max = BPF_REG_6;
22940 	int reg_loop_cnt = BPF_REG_7;
22941 	int reg_loop_ctx = BPF_REG_8;
22942 
22943 	struct bpf_insn *insn_buf = env->insn_buf;
22944 	struct bpf_prog *new_prog;
22945 	u32 callback_start;
22946 	u32 call_insn_offset;
22947 	s32 callback_offset;
22948 	u32 cnt = 0;
22949 
22950 	/* This represents an inlined version of bpf_iter.c:bpf_loop,
22951 	 * be careful to modify this code in sync.
22952 	 */
22953 
22954 	/* Return error and jump to the end of the patch if
22955 	 * expected number of iterations is too big.
22956 	 */
22957 	insn_buf[cnt++] = BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2);
22958 	insn_buf[cnt++] = BPF_MOV32_IMM(BPF_REG_0, -E2BIG);
22959 	insn_buf[cnt++] = BPF_JMP_IMM(BPF_JA, 0, 0, 16);
22960 	/* spill R6, R7, R8 to use these as loop vars */
22961 	insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset);
22962 	insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset);
22963 	insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset);
22964 	/* initialize loop vars */
22965 	insn_buf[cnt++] = BPF_MOV64_REG(reg_loop_max, BPF_REG_1);
22966 	insn_buf[cnt++] = BPF_MOV32_IMM(reg_loop_cnt, 0);
22967 	insn_buf[cnt++] = BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3);
22968 	/* loop header,
22969 	 * if reg_loop_cnt >= reg_loop_max skip the loop body
22970 	 */
22971 	insn_buf[cnt++] = BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5);
22972 	/* callback call,
22973 	 * correct callback offset would be set after patching
22974 	 */
22975 	insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt);
22976 	insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx);
22977 	insn_buf[cnt++] = BPF_CALL_REL(0);
22978 	/* increment loop counter */
22979 	insn_buf[cnt++] = BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1);
22980 	/* jump to loop header if callback returned 0 */
22981 	insn_buf[cnt++] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6);
22982 	/* return value of bpf_loop,
22983 	 * set R0 to the number of iterations
22984 	 */
22985 	insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt);
22986 	/* restore original values of R6, R7, R8 */
22987 	insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset);
22988 	insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset);
22989 	insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset);
22990 
22991 	*total_cnt = cnt;
22992 	new_prog = bpf_patch_insn_data(env, position, insn_buf, cnt);
22993 	if (!new_prog)
22994 		return new_prog;
22995 
22996 	/* callback start is known only after patching */
22997 	callback_start = env->subprog_info[callback_subprogno].start;
22998 	/* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
22999 	call_insn_offset = position + 12;
23000 	callback_offset = callback_start - call_insn_offset - 1;
23001 	new_prog->insnsi[call_insn_offset].imm = callback_offset;
23002 
23003 	return new_prog;
23004 }
23005 
is_bpf_loop_call(struct bpf_insn * insn)23006 static bool is_bpf_loop_call(struct bpf_insn *insn)
23007 {
23008 	return insn->code == (BPF_JMP | BPF_CALL) &&
23009 		insn->src_reg == 0 &&
23010 		insn->imm == BPF_FUNC_loop;
23011 }
23012 
23013 /* For all sub-programs in the program (including main) check
23014  * insn_aux_data to see if there are bpf_loop calls that require
23015  * inlining. If such calls are found the calls are replaced with a
23016  * sequence of instructions produced by `inline_bpf_loop` function and
23017  * subprog stack_depth is increased by the size of 3 registers.
23018  * This stack space is used to spill values of the R6, R7, R8.  These
23019  * registers are used to store the loop bound, counter and context
23020  * variables.
23021  */
optimize_bpf_loop(struct bpf_verifier_env * env)23022 static int optimize_bpf_loop(struct bpf_verifier_env *env)
23023 {
23024 	struct bpf_subprog_info *subprogs = env->subprog_info;
23025 	int i, cur_subprog = 0, cnt, delta = 0;
23026 	struct bpf_insn *insn = env->prog->insnsi;
23027 	int insn_cnt = env->prog->len;
23028 	u16 stack_depth = subprogs[cur_subprog].stack_depth;
23029 	u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
23030 	u16 stack_depth_extra = 0;
23031 
23032 	for (i = 0; i < insn_cnt; i++, insn++) {
23033 		struct bpf_loop_inline_state *inline_state =
23034 			&env->insn_aux_data[i + delta].loop_inline_state;
23035 
23036 		if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
23037 			struct bpf_prog *new_prog;
23038 
23039 			stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
23040 			new_prog = inline_bpf_loop(env,
23041 						   i + delta,
23042 						   -(stack_depth + stack_depth_extra),
23043 						   inline_state->callback_subprogno,
23044 						   &cnt);
23045 			if (!new_prog)
23046 				return -ENOMEM;
23047 
23048 			delta     += cnt - 1;
23049 			env->prog  = new_prog;
23050 			insn       = new_prog->insnsi + i + delta;
23051 		}
23052 
23053 		if (subprogs[cur_subprog + 1].start == i + delta + 1) {
23054 			subprogs[cur_subprog].stack_depth += stack_depth_extra;
23055 			cur_subprog++;
23056 			stack_depth = subprogs[cur_subprog].stack_depth;
23057 			stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
23058 			stack_depth_extra = 0;
23059 		}
23060 	}
23061 
23062 	env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
23063 
23064 	return 0;
23065 }
23066 
23067 /* Remove unnecessary spill/fill pairs, members of fastcall pattern,
23068  * adjust subprograms stack depth when possible.
23069  */
remove_fastcall_spills_fills(struct bpf_verifier_env * env)23070 static int remove_fastcall_spills_fills(struct bpf_verifier_env *env)
23071 {
23072 	struct bpf_subprog_info *subprog = env->subprog_info;
23073 	struct bpf_insn_aux_data *aux = env->insn_aux_data;
23074 	struct bpf_insn *insn = env->prog->insnsi;
23075 	int insn_cnt = env->prog->len;
23076 	u32 spills_num;
23077 	bool modified = false;
23078 	int i, j;
23079 
23080 	for (i = 0; i < insn_cnt; i++, insn++) {
23081 		if (aux[i].fastcall_spills_num > 0) {
23082 			spills_num = aux[i].fastcall_spills_num;
23083 			/* NOPs would be removed by opt_remove_nops() */
23084 			for (j = 1; j <= spills_num; ++j) {
23085 				*(insn - j) = NOP;
23086 				*(insn + j) = NOP;
23087 			}
23088 			modified = true;
23089 		}
23090 		if ((subprog + 1)->start == i + 1) {
23091 			if (modified && !subprog->keep_fastcall_stack)
23092 				subprog->stack_depth = -subprog->fastcall_stack_off;
23093 			subprog++;
23094 			modified = false;
23095 		}
23096 	}
23097 
23098 	return 0;
23099 }
23100 
free_states(struct bpf_verifier_env * env)23101 static void free_states(struct bpf_verifier_env *env)
23102 {
23103 	struct bpf_verifier_state_list *sl;
23104 	struct list_head *head, *pos, *tmp;
23105 	struct bpf_scc_info *info;
23106 	int i, j;
23107 
23108 	free_verifier_state(env->cur_state, true);
23109 	env->cur_state = NULL;
23110 	while (!pop_stack(env, NULL, NULL, false));
23111 
23112 	list_for_each_safe(pos, tmp, &env->free_list) {
23113 		sl = container_of(pos, struct bpf_verifier_state_list, node);
23114 		free_verifier_state(&sl->state, false);
23115 		kfree(sl);
23116 	}
23117 	INIT_LIST_HEAD(&env->free_list);
23118 
23119 	for (i = 0; i < env->scc_cnt; ++i) {
23120 		info = env->scc_info[i];
23121 		if (!info)
23122 			continue;
23123 		for (j = 0; j < info->num_visits; j++)
23124 			free_backedges(&info->visits[j]);
23125 		kvfree(info);
23126 		env->scc_info[i] = NULL;
23127 	}
23128 
23129 	if (!env->explored_states)
23130 		return;
23131 
23132 	for (i = 0; i < state_htab_size(env); i++) {
23133 		head = &env->explored_states[i];
23134 
23135 		list_for_each_safe(pos, tmp, head) {
23136 			sl = container_of(pos, struct bpf_verifier_state_list, node);
23137 			free_verifier_state(&sl->state, false);
23138 			kfree(sl);
23139 		}
23140 		INIT_LIST_HEAD(&env->explored_states[i]);
23141 	}
23142 }
23143 
do_check_common(struct bpf_verifier_env * env,int subprog)23144 static int do_check_common(struct bpf_verifier_env *env, int subprog)
23145 {
23146 	bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
23147 	struct bpf_subprog_info *sub = subprog_info(env, subprog);
23148 	struct bpf_prog_aux *aux = env->prog->aux;
23149 	struct bpf_verifier_state *state;
23150 	struct bpf_reg_state *regs;
23151 	int ret, i;
23152 
23153 	env->prev_linfo = NULL;
23154 	env->pass_cnt++;
23155 
23156 	state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL_ACCOUNT);
23157 	if (!state)
23158 		return -ENOMEM;
23159 	state->curframe = 0;
23160 	state->speculative = false;
23161 	state->branches = 1;
23162 	state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL_ACCOUNT);
23163 	if (!state->frame[0]) {
23164 		kfree(state);
23165 		return -ENOMEM;
23166 	}
23167 	env->cur_state = state;
23168 	init_func_state(env, state->frame[0],
23169 			BPF_MAIN_FUNC /* callsite */,
23170 			0 /* frameno */,
23171 			subprog);
23172 	state->first_insn_idx = env->subprog_info[subprog].start;
23173 	state->last_insn_idx = -1;
23174 
23175 	regs = state->frame[state->curframe]->regs;
23176 	if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
23177 		const char *sub_name = subprog_name(env, subprog);
23178 		struct bpf_subprog_arg_info *arg;
23179 		struct bpf_reg_state *reg;
23180 
23181 		verbose(env, "Validating %s() func#%d...\n", sub_name, subprog);
23182 		ret = btf_prepare_func_args(env, subprog);
23183 		if (ret)
23184 			goto out;
23185 
23186 		if (subprog_is_exc_cb(env, subprog)) {
23187 			state->frame[0]->in_exception_callback_fn = true;
23188 			/* We have already ensured that the callback returns an integer, just
23189 			 * like all global subprogs. We need to determine it only has a single
23190 			 * scalar argument.
23191 			 */
23192 			if (sub->arg_cnt != 1 || sub->args[0].arg_type != ARG_ANYTHING) {
23193 				verbose(env, "exception cb only supports single integer argument\n");
23194 				ret = -EINVAL;
23195 				goto out;
23196 			}
23197 		}
23198 		for (i = BPF_REG_1; i <= sub->arg_cnt; i++) {
23199 			arg = &sub->args[i - BPF_REG_1];
23200 			reg = &regs[i];
23201 
23202 			if (arg->arg_type == ARG_PTR_TO_CTX) {
23203 				reg->type = PTR_TO_CTX;
23204 				mark_reg_known_zero(env, regs, i);
23205 			} else if (arg->arg_type == ARG_ANYTHING) {
23206 				reg->type = SCALAR_VALUE;
23207 				mark_reg_unknown(env, regs, i);
23208 			} else if (arg->arg_type == (ARG_PTR_TO_DYNPTR | MEM_RDONLY)) {
23209 				/* assume unspecial LOCAL dynptr type */
23210 				__mark_dynptr_reg(reg, BPF_DYNPTR_TYPE_LOCAL, true, ++env->id_gen);
23211 			} else if (base_type(arg->arg_type) == ARG_PTR_TO_MEM) {
23212 				reg->type = PTR_TO_MEM;
23213 				reg->type |= arg->arg_type &
23214 					     (PTR_MAYBE_NULL | PTR_UNTRUSTED | MEM_RDONLY);
23215 				mark_reg_known_zero(env, regs, i);
23216 				reg->mem_size = arg->mem_size;
23217 				if (arg->arg_type & PTR_MAYBE_NULL)
23218 					reg->id = ++env->id_gen;
23219 			} else if (base_type(arg->arg_type) == ARG_PTR_TO_BTF_ID) {
23220 				reg->type = PTR_TO_BTF_ID;
23221 				if (arg->arg_type & PTR_MAYBE_NULL)
23222 					reg->type |= PTR_MAYBE_NULL;
23223 				if (arg->arg_type & PTR_UNTRUSTED)
23224 					reg->type |= PTR_UNTRUSTED;
23225 				if (arg->arg_type & PTR_TRUSTED)
23226 					reg->type |= PTR_TRUSTED;
23227 				mark_reg_known_zero(env, regs, i);
23228 				reg->btf = bpf_get_btf_vmlinux(); /* can't fail at this point */
23229 				reg->btf_id = arg->btf_id;
23230 				reg->id = ++env->id_gen;
23231 			} else if (base_type(arg->arg_type) == ARG_PTR_TO_ARENA) {
23232 				/* caller can pass either PTR_TO_ARENA or SCALAR */
23233 				mark_reg_unknown(env, regs, i);
23234 			} else {
23235 				verifier_bug(env, "unhandled arg#%d type %d",
23236 					     i - BPF_REG_1, arg->arg_type);
23237 				ret = -EFAULT;
23238 				goto out;
23239 			}
23240 		}
23241 	} else {
23242 		/* if main BPF program has associated BTF info, validate that
23243 		 * it's matching expected signature, and otherwise mark BTF
23244 		 * info for main program as unreliable
23245 		 */
23246 		if (env->prog->aux->func_info_aux) {
23247 			ret = btf_prepare_func_args(env, 0);
23248 			if (ret || sub->arg_cnt != 1 || sub->args[0].arg_type != ARG_PTR_TO_CTX)
23249 				env->prog->aux->func_info_aux[0].unreliable = true;
23250 		}
23251 
23252 		/* 1st arg to a function */
23253 		regs[BPF_REG_1].type = PTR_TO_CTX;
23254 		mark_reg_known_zero(env, regs, BPF_REG_1);
23255 	}
23256 
23257 	/* Acquire references for struct_ops program arguments tagged with "__ref" */
23258 	if (!subprog && env->prog->type == BPF_PROG_TYPE_STRUCT_OPS) {
23259 		for (i = 0; i < aux->ctx_arg_info_size; i++)
23260 			aux->ctx_arg_info[i].ref_obj_id = aux->ctx_arg_info[i].refcounted ?
23261 							  acquire_reference(env, 0) : 0;
23262 	}
23263 
23264 	ret = do_check(env);
23265 out:
23266 	if (!ret && pop_log)
23267 		bpf_vlog_reset(&env->log, 0);
23268 	free_states(env);
23269 	return ret;
23270 }
23271 
23272 /* Lazily verify all global functions based on their BTF, if they are called
23273  * from main BPF program or any of subprograms transitively.
23274  * BPF global subprogs called from dead code are not validated.
23275  * All callable global functions must pass verification.
23276  * Otherwise the whole program is rejected.
23277  * Consider:
23278  * int bar(int);
23279  * int foo(int f)
23280  * {
23281  *    return bar(f);
23282  * }
23283  * int bar(int b)
23284  * {
23285  *    ...
23286  * }
23287  * foo() will be verified first for R1=any_scalar_value. During verification it
23288  * will be assumed that bar() already verified successfully and call to bar()
23289  * from foo() will be checked for type match only. Later bar() will be verified
23290  * independently to check that it's safe for R1=any_scalar_value.
23291  */
do_check_subprogs(struct bpf_verifier_env * env)23292 static int do_check_subprogs(struct bpf_verifier_env *env)
23293 {
23294 	struct bpf_prog_aux *aux = env->prog->aux;
23295 	struct bpf_func_info_aux *sub_aux;
23296 	int i, ret, new_cnt;
23297 
23298 	if (!aux->func_info)
23299 		return 0;
23300 
23301 	/* exception callback is presumed to be always called */
23302 	if (env->exception_callback_subprog)
23303 		subprog_aux(env, env->exception_callback_subprog)->called = true;
23304 
23305 again:
23306 	new_cnt = 0;
23307 	for (i = 1; i < env->subprog_cnt; i++) {
23308 		if (!subprog_is_global(env, i))
23309 			continue;
23310 
23311 		sub_aux = subprog_aux(env, i);
23312 		if (!sub_aux->called || sub_aux->verified)
23313 			continue;
23314 
23315 		env->insn_idx = env->subprog_info[i].start;
23316 		WARN_ON_ONCE(env->insn_idx == 0);
23317 		ret = do_check_common(env, i);
23318 		if (ret) {
23319 			return ret;
23320 		} else if (env->log.level & BPF_LOG_LEVEL) {
23321 			verbose(env, "Func#%d ('%s') is safe for any args that match its prototype\n",
23322 				i, subprog_name(env, i));
23323 		}
23324 
23325 		/* We verified new global subprog, it might have called some
23326 		 * more global subprogs that we haven't verified yet, so we
23327 		 * need to do another pass over subprogs to verify those.
23328 		 */
23329 		sub_aux->verified = true;
23330 		new_cnt++;
23331 	}
23332 
23333 	/* We can't loop forever as we verify at least one global subprog on
23334 	 * each pass.
23335 	 */
23336 	if (new_cnt)
23337 		goto again;
23338 
23339 	return 0;
23340 }
23341 
do_check_main(struct bpf_verifier_env * env)23342 static int do_check_main(struct bpf_verifier_env *env)
23343 {
23344 	int ret;
23345 
23346 	env->insn_idx = 0;
23347 	ret = do_check_common(env, 0);
23348 	if (!ret)
23349 		env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
23350 	return ret;
23351 }
23352 
23353 
print_verification_stats(struct bpf_verifier_env * env)23354 static void print_verification_stats(struct bpf_verifier_env *env)
23355 {
23356 	int i;
23357 
23358 	if (env->log.level & BPF_LOG_STATS) {
23359 		verbose(env, "verification time %lld usec\n",
23360 			div_u64(env->verification_time, 1000));
23361 		verbose(env, "stack depth ");
23362 		for (i = 0; i < env->subprog_cnt; i++) {
23363 			u32 depth = env->subprog_info[i].stack_depth;
23364 
23365 			verbose(env, "%d", depth);
23366 			if (i + 1 < env->subprog_cnt)
23367 				verbose(env, "+");
23368 		}
23369 		verbose(env, "\n");
23370 	}
23371 	verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
23372 		"total_states %d peak_states %d mark_read %d\n",
23373 		env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
23374 		env->max_states_per_insn, env->total_states,
23375 		env->peak_states, env->longest_mark_read_walk);
23376 }
23377 
bpf_prog_ctx_arg_info_init(struct bpf_prog * prog,const struct bpf_ctx_arg_aux * info,u32 cnt)23378 int bpf_prog_ctx_arg_info_init(struct bpf_prog *prog,
23379 			       const struct bpf_ctx_arg_aux *info, u32 cnt)
23380 {
23381 	prog->aux->ctx_arg_info = kmemdup_array(info, cnt, sizeof(*info), GFP_KERNEL_ACCOUNT);
23382 	prog->aux->ctx_arg_info_size = cnt;
23383 
23384 	return prog->aux->ctx_arg_info ? 0 : -ENOMEM;
23385 }
23386 
check_struct_ops_btf_id(struct bpf_verifier_env * env)23387 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
23388 {
23389 	const struct btf_type *t, *func_proto;
23390 	const struct bpf_struct_ops_desc *st_ops_desc;
23391 	const struct bpf_struct_ops *st_ops;
23392 	const struct btf_member *member;
23393 	struct bpf_prog *prog = env->prog;
23394 	bool has_refcounted_arg = false;
23395 	u32 btf_id, member_idx, member_off;
23396 	struct btf *btf;
23397 	const char *mname;
23398 	int i, err;
23399 
23400 	if (!prog->gpl_compatible) {
23401 		verbose(env, "struct ops programs must have a GPL compatible license\n");
23402 		return -EINVAL;
23403 	}
23404 
23405 	if (!prog->aux->attach_btf_id)
23406 		return -ENOTSUPP;
23407 
23408 	btf = prog->aux->attach_btf;
23409 	if (btf_is_module(btf)) {
23410 		/* Make sure st_ops is valid through the lifetime of env */
23411 		env->attach_btf_mod = btf_try_get_module(btf);
23412 		if (!env->attach_btf_mod) {
23413 			verbose(env, "struct_ops module %s is not found\n",
23414 				btf_get_name(btf));
23415 			return -ENOTSUPP;
23416 		}
23417 	}
23418 
23419 	btf_id = prog->aux->attach_btf_id;
23420 	st_ops_desc = bpf_struct_ops_find(btf, btf_id);
23421 	if (!st_ops_desc) {
23422 		verbose(env, "attach_btf_id %u is not a supported struct\n",
23423 			btf_id);
23424 		return -ENOTSUPP;
23425 	}
23426 	st_ops = st_ops_desc->st_ops;
23427 
23428 	t = st_ops_desc->type;
23429 	member_idx = prog->expected_attach_type;
23430 	if (member_idx >= btf_type_vlen(t)) {
23431 		verbose(env, "attach to invalid member idx %u of struct %s\n",
23432 			member_idx, st_ops->name);
23433 		return -EINVAL;
23434 	}
23435 
23436 	member = &btf_type_member(t)[member_idx];
23437 	mname = btf_name_by_offset(btf, member->name_off);
23438 	func_proto = btf_type_resolve_func_ptr(btf, member->type,
23439 					       NULL);
23440 	if (!func_proto) {
23441 		verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
23442 			mname, member_idx, st_ops->name);
23443 		return -EINVAL;
23444 	}
23445 
23446 	member_off = __btf_member_bit_offset(t, member) / 8;
23447 	err = bpf_struct_ops_supported(st_ops, member_off);
23448 	if (err) {
23449 		verbose(env, "attach to unsupported member %s of struct %s\n",
23450 			mname, st_ops->name);
23451 		return err;
23452 	}
23453 
23454 	if (st_ops->check_member) {
23455 		err = st_ops->check_member(t, member, prog);
23456 
23457 		if (err) {
23458 			verbose(env, "attach to unsupported member %s of struct %s\n",
23459 				mname, st_ops->name);
23460 			return err;
23461 		}
23462 	}
23463 
23464 	if (prog->aux->priv_stack_requested && !bpf_jit_supports_private_stack()) {
23465 		verbose(env, "Private stack not supported by jit\n");
23466 		return -EACCES;
23467 	}
23468 
23469 	for (i = 0; i < st_ops_desc->arg_info[member_idx].cnt; i++) {
23470 		if (st_ops_desc->arg_info[member_idx].info->refcounted) {
23471 			has_refcounted_arg = true;
23472 			break;
23473 		}
23474 	}
23475 
23476 	/* Tail call is not allowed for programs with refcounted arguments since we
23477 	 * cannot guarantee that valid refcounted kptrs will be passed to the callee.
23478 	 */
23479 	for (i = 0; i < env->subprog_cnt; i++) {
23480 		if (has_refcounted_arg && env->subprog_info[i].has_tail_call) {
23481 			verbose(env, "program with __ref argument cannot tail call\n");
23482 			return -EINVAL;
23483 		}
23484 	}
23485 
23486 	prog->aux->st_ops = st_ops;
23487 	prog->aux->attach_st_ops_member_off = member_off;
23488 
23489 	prog->aux->attach_func_proto = func_proto;
23490 	prog->aux->attach_func_name = mname;
23491 	env->ops = st_ops->verifier_ops;
23492 
23493 	return bpf_prog_ctx_arg_info_init(prog, st_ops_desc->arg_info[member_idx].info,
23494 					  st_ops_desc->arg_info[member_idx].cnt);
23495 }
23496 #define SECURITY_PREFIX "security_"
23497 
check_attach_modify_return(unsigned long addr,const char * func_name)23498 static int check_attach_modify_return(unsigned long addr, const char *func_name)
23499 {
23500 	if (within_error_injection_list(addr) ||
23501 	    !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
23502 		return 0;
23503 
23504 	return -EINVAL;
23505 }
23506 
23507 /* list of non-sleepable functions that are otherwise on
23508  * ALLOW_ERROR_INJECTION list
23509  */
23510 BTF_SET_START(btf_non_sleepable_error_inject)
23511 /* Three functions below can be called from sleepable and non-sleepable context.
23512  * Assume non-sleepable from bpf safety point of view.
23513  */
BTF_ID(func,__filemap_add_folio)23514 BTF_ID(func, __filemap_add_folio)
23515 #ifdef CONFIG_FAIL_PAGE_ALLOC
23516 BTF_ID(func, should_fail_alloc_page)
23517 #endif
23518 #ifdef CONFIG_FAILSLAB
23519 BTF_ID(func, should_failslab)
23520 #endif
23521 BTF_SET_END(btf_non_sleepable_error_inject)
23522 
23523 static int check_non_sleepable_error_inject(u32 btf_id)
23524 {
23525 	return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
23526 }
23527 
bpf_check_attach_target(struct bpf_verifier_log * log,const struct bpf_prog * prog,const struct bpf_prog * tgt_prog,u32 btf_id,struct bpf_attach_target_info * tgt_info)23528 int bpf_check_attach_target(struct bpf_verifier_log *log,
23529 			    const struct bpf_prog *prog,
23530 			    const struct bpf_prog *tgt_prog,
23531 			    u32 btf_id,
23532 			    struct bpf_attach_target_info *tgt_info)
23533 {
23534 	bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
23535 	bool prog_tracing = prog->type == BPF_PROG_TYPE_TRACING;
23536 	char trace_symbol[KSYM_SYMBOL_LEN];
23537 	const char prefix[] = "btf_trace_";
23538 	struct bpf_raw_event_map *btp;
23539 	int ret = 0, subprog = -1, i;
23540 	const struct btf_type *t;
23541 	bool conservative = true;
23542 	const char *tname, *fname;
23543 	struct btf *btf;
23544 	long addr = 0;
23545 	struct module *mod = NULL;
23546 
23547 	if (!btf_id) {
23548 		bpf_log(log, "Tracing programs must provide btf_id\n");
23549 		return -EINVAL;
23550 	}
23551 	btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
23552 	if (!btf) {
23553 		bpf_log(log,
23554 			"FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
23555 		return -EINVAL;
23556 	}
23557 	t = btf_type_by_id(btf, btf_id);
23558 	if (!t) {
23559 		bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
23560 		return -EINVAL;
23561 	}
23562 	tname = btf_name_by_offset(btf, t->name_off);
23563 	if (!tname) {
23564 		bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
23565 		return -EINVAL;
23566 	}
23567 	if (tgt_prog) {
23568 		struct bpf_prog_aux *aux = tgt_prog->aux;
23569 		bool tgt_changes_pkt_data;
23570 		bool tgt_might_sleep;
23571 
23572 		if (bpf_prog_is_dev_bound(prog->aux) &&
23573 		    !bpf_prog_dev_bound_match(prog, tgt_prog)) {
23574 			bpf_log(log, "Target program bound device mismatch");
23575 			return -EINVAL;
23576 		}
23577 
23578 		for (i = 0; i < aux->func_info_cnt; i++)
23579 			if (aux->func_info[i].type_id == btf_id) {
23580 				subprog = i;
23581 				break;
23582 			}
23583 		if (subprog == -1) {
23584 			bpf_log(log, "Subprog %s doesn't exist\n", tname);
23585 			return -EINVAL;
23586 		}
23587 		if (aux->func && aux->func[subprog]->aux->exception_cb) {
23588 			bpf_log(log,
23589 				"%s programs cannot attach to exception callback\n",
23590 				prog_extension ? "Extension" : "FENTRY/FEXIT");
23591 			return -EINVAL;
23592 		}
23593 		conservative = aux->func_info_aux[subprog].unreliable;
23594 		if (prog_extension) {
23595 			if (conservative) {
23596 				bpf_log(log,
23597 					"Cannot replace static functions\n");
23598 				return -EINVAL;
23599 			}
23600 			if (!prog->jit_requested) {
23601 				bpf_log(log,
23602 					"Extension programs should be JITed\n");
23603 				return -EINVAL;
23604 			}
23605 			tgt_changes_pkt_data = aux->func
23606 					       ? aux->func[subprog]->aux->changes_pkt_data
23607 					       : aux->changes_pkt_data;
23608 			if (prog->aux->changes_pkt_data && !tgt_changes_pkt_data) {
23609 				bpf_log(log,
23610 					"Extension program changes packet data, while original does not\n");
23611 				return -EINVAL;
23612 			}
23613 
23614 			tgt_might_sleep = aux->func
23615 					  ? aux->func[subprog]->aux->might_sleep
23616 					  : aux->might_sleep;
23617 			if (prog->aux->might_sleep && !tgt_might_sleep) {
23618 				bpf_log(log,
23619 					"Extension program may sleep, while original does not\n");
23620 				return -EINVAL;
23621 			}
23622 		}
23623 		if (!tgt_prog->jited) {
23624 			bpf_log(log, "Can attach to only JITed progs\n");
23625 			return -EINVAL;
23626 		}
23627 		if (prog_tracing) {
23628 			if (aux->attach_tracing_prog) {
23629 				/*
23630 				 * Target program is an fentry/fexit which is already attached
23631 				 * to another tracing program. More levels of nesting
23632 				 * attachment are not allowed.
23633 				 */
23634 				bpf_log(log, "Cannot nest tracing program attach more than once\n");
23635 				return -EINVAL;
23636 			}
23637 		} else if (tgt_prog->type == prog->type) {
23638 			/*
23639 			 * To avoid potential call chain cycles, prevent attaching of a
23640 			 * program extension to another extension. It's ok to attach
23641 			 * fentry/fexit to extension program.
23642 			 */
23643 			bpf_log(log, "Cannot recursively attach\n");
23644 			return -EINVAL;
23645 		}
23646 		if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
23647 		    prog_extension &&
23648 		    (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
23649 		     tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
23650 			/* Program extensions can extend all program types
23651 			 * except fentry/fexit. The reason is the following.
23652 			 * The fentry/fexit programs are used for performance
23653 			 * analysis, stats and can be attached to any program
23654 			 * type. When extension program is replacing XDP function
23655 			 * it is necessary to allow performance analysis of all
23656 			 * functions. Both original XDP program and its program
23657 			 * extension. Hence attaching fentry/fexit to
23658 			 * BPF_PROG_TYPE_EXT is allowed. If extending of
23659 			 * fentry/fexit was allowed it would be possible to create
23660 			 * long call chain fentry->extension->fentry->extension
23661 			 * beyond reasonable stack size. Hence extending fentry
23662 			 * is not allowed.
23663 			 */
23664 			bpf_log(log, "Cannot extend fentry/fexit\n");
23665 			return -EINVAL;
23666 		}
23667 	} else {
23668 		if (prog_extension) {
23669 			bpf_log(log, "Cannot replace kernel functions\n");
23670 			return -EINVAL;
23671 		}
23672 	}
23673 
23674 	switch (prog->expected_attach_type) {
23675 	case BPF_TRACE_RAW_TP:
23676 		if (tgt_prog) {
23677 			bpf_log(log,
23678 				"Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
23679 			return -EINVAL;
23680 		}
23681 		if (!btf_type_is_typedef(t)) {
23682 			bpf_log(log, "attach_btf_id %u is not a typedef\n",
23683 				btf_id);
23684 			return -EINVAL;
23685 		}
23686 		if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
23687 			bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
23688 				btf_id, tname);
23689 			return -EINVAL;
23690 		}
23691 		tname += sizeof(prefix) - 1;
23692 
23693 		/* The func_proto of "btf_trace_##tname" is generated from typedef without argument
23694 		 * names. Thus using bpf_raw_event_map to get argument names.
23695 		 */
23696 		btp = bpf_get_raw_tracepoint(tname);
23697 		if (!btp)
23698 			return -EINVAL;
23699 		fname = kallsyms_lookup((unsigned long)btp->bpf_func, NULL, NULL, NULL,
23700 					trace_symbol);
23701 		bpf_put_raw_tracepoint(btp);
23702 
23703 		if (fname)
23704 			ret = btf_find_by_name_kind(btf, fname, BTF_KIND_FUNC);
23705 
23706 		if (!fname || ret < 0) {
23707 			bpf_log(log, "Cannot find btf of tracepoint template, fall back to %s%s.\n",
23708 				prefix, tname);
23709 			t = btf_type_by_id(btf, t->type);
23710 			if (!btf_type_is_ptr(t))
23711 				/* should never happen in valid vmlinux build */
23712 				return -EINVAL;
23713 		} else {
23714 			t = btf_type_by_id(btf, ret);
23715 			if (!btf_type_is_func(t))
23716 				/* should never happen in valid vmlinux build */
23717 				return -EINVAL;
23718 		}
23719 
23720 		t = btf_type_by_id(btf, t->type);
23721 		if (!btf_type_is_func_proto(t))
23722 			/* should never happen in valid vmlinux build */
23723 			return -EINVAL;
23724 
23725 		break;
23726 	case BPF_TRACE_ITER:
23727 		if (!btf_type_is_func(t)) {
23728 			bpf_log(log, "attach_btf_id %u is not a function\n",
23729 				btf_id);
23730 			return -EINVAL;
23731 		}
23732 		t = btf_type_by_id(btf, t->type);
23733 		if (!btf_type_is_func_proto(t))
23734 			return -EINVAL;
23735 		ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
23736 		if (ret)
23737 			return ret;
23738 		break;
23739 	default:
23740 		if (!prog_extension)
23741 			return -EINVAL;
23742 		fallthrough;
23743 	case BPF_MODIFY_RETURN:
23744 	case BPF_LSM_MAC:
23745 	case BPF_LSM_CGROUP:
23746 	case BPF_TRACE_FENTRY:
23747 	case BPF_TRACE_FEXIT:
23748 		if (!btf_type_is_func(t)) {
23749 			bpf_log(log, "attach_btf_id %u is not a function\n",
23750 				btf_id);
23751 			return -EINVAL;
23752 		}
23753 		if (prog_extension &&
23754 		    btf_check_type_match(log, prog, btf, t))
23755 			return -EINVAL;
23756 		t = btf_type_by_id(btf, t->type);
23757 		if (!btf_type_is_func_proto(t))
23758 			return -EINVAL;
23759 
23760 		if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
23761 		    (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
23762 		     prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
23763 			return -EINVAL;
23764 
23765 		if (tgt_prog && conservative)
23766 			t = NULL;
23767 
23768 		ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
23769 		if (ret < 0)
23770 			return ret;
23771 
23772 		if (tgt_prog) {
23773 			if (subprog == 0)
23774 				addr = (long) tgt_prog->bpf_func;
23775 			else
23776 				addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
23777 		} else {
23778 			if (btf_is_module(btf)) {
23779 				mod = btf_try_get_module(btf);
23780 				if (mod)
23781 					addr = find_kallsyms_symbol_value(mod, tname);
23782 				else
23783 					addr = 0;
23784 			} else {
23785 				addr = kallsyms_lookup_name(tname);
23786 			}
23787 			if (!addr) {
23788 				module_put(mod);
23789 				bpf_log(log,
23790 					"The address of function %s cannot be found\n",
23791 					tname);
23792 				return -ENOENT;
23793 			}
23794 		}
23795 
23796 		if (prog->sleepable) {
23797 			ret = -EINVAL;
23798 			switch (prog->type) {
23799 			case BPF_PROG_TYPE_TRACING:
23800 
23801 				/* fentry/fexit/fmod_ret progs can be sleepable if they are
23802 				 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
23803 				 */
23804 				if (!check_non_sleepable_error_inject(btf_id) &&
23805 				    within_error_injection_list(addr))
23806 					ret = 0;
23807 				/* fentry/fexit/fmod_ret progs can also be sleepable if they are
23808 				 * in the fmodret id set with the KF_SLEEPABLE flag.
23809 				 */
23810 				else {
23811 					u32 *flags = btf_kfunc_is_modify_return(btf, btf_id,
23812 										prog);
23813 
23814 					if (flags && (*flags & KF_SLEEPABLE))
23815 						ret = 0;
23816 				}
23817 				break;
23818 			case BPF_PROG_TYPE_LSM:
23819 				/* LSM progs check that they are attached to bpf_lsm_*() funcs.
23820 				 * Only some of them are sleepable.
23821 				 */
23822 				if (bpf_lsm_is_sleepable_hook(btf_id))
23823 					ret = 0;
23824 				break;
23825 			default:
23826 				break;
23827 			}
23828 			if (ret) {
23829 				module_put(mod);
23830 				bpf_log(log, "%s is not sleepable\n", tname);
23831 				return ret;
23832 			}
23833 		} else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
23834 			if (tgt_prog) {
23835 				module_put(mod);
23836 				bpf_log(log, "can't modify return codes of BPF programs\n");
23837 				return -EINVAL;
23838 			}
23839 			ret = -EINVAL;
23840 			if (btf_kfunc_is_modify_return(btf, btf_id, prog) ||
23841 			    !check_attach_modify_return(addr, tname))
23842 				ret = 0;
23843 			if (ret) {
23844 				module_put(mod);
23845 				bpf_log(log, "%s() is not modifiable\n", tname);
23846 				return ret;
23847 			}
23848 		}
23849 
23850 		break;
23851 	}
23852 	tgt_info->tgt_addr = addr;
23853 	tgt_info->tgt_name = tname;
23854 	tgt_info->tgt_type = t;
23855 	tgt_info->tgt_mod = mod;
23856 	return 0;
23857 }
23858 
BTF_SET_START(btf_id_deny)23859 BTF_SET_START(btf_id_deny)
23860 BTF_ID_UNUSED
23861 #ifdef CONFIG_SMP
23862 BTF_ID(func, migrate_disable)
23863 BTF_ID(func, migrate_enable)
23864 #endif
23865 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
23866 BTF_ID(func, rcu_read_unlock_strict)
23867 #endif
23868 #if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
23869 BTF_ID(func, preempt_count_add)
23870 BTF_ID(func, preempt_count_sub)
23871 #endif
23872 #ifdef CONFIG_PREEMPT_RCU
23873 BTF_ID(func, __rcu_read_lock)
23874 BTF_ID(func, __rcu_read_unlock)
23875 #endif
23876 BTF_SET_END(btf_id_deny)
23877 
23878 /* fexit and fmod_ret can't be used to attach to __noreturn functions.
23879  * Currently, we must manually list all __noreturn functions here. Once a more
23880  * robust solution is implemented, this workaround can be removed.
23881  */
23882 BTF_SET_START(noreturn_deny)
23883 #ifdef CONFIG_IA32_EMULATION
23884 BTF_ID(func, __ia32_sys_exit)
23885 BTF_ID(func, __ia32_sys_exit_group)
23886 #endif
23887 #ifdef CONFIG_KUNIT
23888 BTF_ID(func, __kunit_abort)
23889 BTF_ID(func, kunit_try_catch_throw)
23890 #endif
23891 #ifdef CONFIG_MODULES
23892 BTF_ID(func, __module_put_and_kthread_exit)
23893 #endif
23894 #ifdef CONFIG_X86_64
23895 BTF_ID(func, __x64_sys_exit)
23896 BTF_ID(func, __x64_sys_exit_group)
23897 #endif
23898 BTF_ID(func, do_exit)
23899 BTF_ID(func, do_group_exit)
23900 BTF_ID(func, kthread_complete_and_exit)
23901 BTF_ID(func, kthread_exit)
23902 BTF_ID(func, make_task_dead)
23903 BTF_SET_END(noreturn_deny)
23904 
23905 static bool can_be_sleepable(struct bpf_prog *prog)
23906 {
23907 	if (prog->type == BPF_PROG_TYPE_TRACING) {
23908 		switch (prog->expected_attach_type) {
23909 		case BPF_TRACE_FENTRY:
23910 		case BPF_TRACE_FEXIT:
23911 		case BPF_MODIFY_RETURN:
23912 		case BPF_TRACE_ITER:
23913 			return true;
23914 		default:
23915 			return false;
23916 		}
23917 	}
23918 	return prog->type == BPF_PROG_TYPE_LSM ||
23919 	       prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
23920 	       prog->type == BPF_PROG_TYPE_STRUCT_OPS;
23921 }
23922 
check_attach_btf_id(struct bpf_verifier_env * env)23923 static int check_attach_btf_id(struct bpf_verifier_env *env)
23924 {
23925 	struct bpf_prog *prog = env->prog;
23926 	struct bpf_prog *tgt_prog = prog->aux->dst_prog;
23927 	struct bpf_attach_target_info tgt_info = {};
23928 	u32 btf_id = prog->aux->attach_btf_id;
23929 	struct bpf_trampoline *tr;
23930 	int ret;
23931 	u64 key;
23932 
23933 	if (prog->type == BPF_PROG_TYPE_SYSCALL) {
23934 		if (prog->sleepable)
23935 			/* attach_btf_id checked to be zero already */
23936 			return 0;
23937 		verbose(env, "Syscall programs can only be sleepable\n");
23938 		return -EINVAL;
23939 	}
23940 
23941 	if (prog->sleepable && !can_be_sleepable(prog)) {
23942 		verbose(env, "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
23943 		return -EINVAL;
23944 	}
23945 
23946 	if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
23947 		return check_struct_ops_btf_id(env);
23948 
23949 	if (prog->type != BPF_PROG_TYPE_TRACING &&
23950 	    prog->type != BPF_PROG_TYPE_LSM &&
23951 	    prog->type != BPF_PROG_TYPE_EXT)
23952 		return 0;
23953 
23954 	ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
23955 	if (ret)
23956 		return ret;
23957 
23958 	if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
23959 		/* to make freplace equivalent to their targets, they need to
23960 		 * inherit env->ops and expected_attach_type for the rest of the
23961 		 * verification
23962 		 */
23963 		env->ops = bpf_verifier_ops[tgt_prog->type];
23964 		prog->expected_attach_type = tgt_prog->expected_attach_type;
23965 	}
23966 
23967 	/* store info about the attachment target that will be used later */
23968 	prog->aux->attach_func_proto = tgt_info.tgt_type;
23969 	prog->aux->attach_func_name = tgt_info.tgt_name;
23970 	prog->aux->mod = tgt_info.tgt_mod;
23971 
23972 	if (tgt_prog) {
23973 		prog->aux->saved_dst_prog_type = tgt_prog->type;
23974 		prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
23975 	}
23976 
23977 	if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
23978 		prog->aux->attach_btf_trace = true;
23979 		return 0;
23980 	} else if (prog->expected_attach_type == BPF_TRACE_ITER) {
23981 		return bpf_iter_prog_supported(prog);
23982 	}
23983 
23984 	if (prog->type == BPF_PROG_TYPE_LSM) {
23985 		ret = bpf_lsm_verify_prog(&env->log, prog);
23986 		if (ret < 0)
23987 			return ret;
23988 	} else if (prog->type == BPF_PROG_TYPE_TRACING &&
23989 		   btf_id_set_contains(&btf_id_deny, btf_id)) {
23990 		verbose(env, "Attaching tracing programs to function '%s' is rejected.\n",
23991 			tgt_info.tgt_name);
23992 		return -EINVAL;
23993 	} else if ((prog->expected_attach_type == BPF_TRACE_FEXIT ||
23994 		   prog->expected_attach_type == BPF_MODIFY_RETURN) &&
23995 		   btf_id_set_contains(&noreturn_deny, btf_id)) {
23996 		verbose(env, "Attaching fexit/fmod_ret to __noreturn function '%s' is rejected.\n",
23997 			tgt_info.tgt_name);
23998 		return -EINVAL;
23999 	}
24000 
24001 	key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
24002 	tr = bpf_trampoline_get(key, &tgt_info);
24003 	if (!tr)
24004 		return -ENOMEM;
24005 
24006 	if (tgt_prog && tgt_prog->aux->tail_call_reachable)
24007 		tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
24008 
24009 	prog->aux->dst_trampoline = tr;
24010 	return 0;
24011 }
24012 
bpf_get_btf_vmlinux(void)24013 struct btf *bpf_get_btf_vmlinux(void)
24014 {
24015 	if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
24016 		mutex_lock(&bpf_verifier_lock);
24017 		if (!btf_vmlinux)
24018 			btf_vmlinux = btf_parse_vmlinux();
24019 		mutex_unlock(&bpf_verifier_lock);
24020 	}
24021 	return btf_vmlinux;
24022 }
24023 
24024 /*
24025  * The add_fd_from_fd_array() is executed only if fd_array_cnt is non-zero. In
24026  * this case expect that every file descriptor in the array is either a map or
24027  * a BTF. Everything else is considered to be trash.
24028  */
add_fd_from_fd_array(struct bpf_verifier_env * env,int fd)24029 static int add_fd_from_fd_array(struct bpf_verifier_env *env, int fd)
24030 {
24031 	struct bpf_map *map;
24032 	struct btf *btf;
24033 	CLASS(fd, f)(fd);
24034 	int err;
24035 
24036 	map = __bpf_map_get(f);
24037 	if (!IS_ERR(map)) {
24038 		err = __add_used_map(env, map);
24039 		if (err < 0)
24040 			return err;
24041 		return 0;
24042 	}
24043 
24044 	btf = __btf_get_by_fd(f);
24045 	if (!IS_ERR(btf)) {
24046 		err = __add_used_btf(env, btf);
24047 		if (err < 0)
24048 			return err;
24049 		return 0;
24050 	}
24051 
24052 	verbose(env, "fd %d is not pointing to valid bpf_map or btf\n", fd);
24053 	return PTR_ERR(map);
24054 }
24055 
process_fd_array(struct bpf_verifier_env * env,union bpf_attr * attr,bpfptr_t uattr)24056 static int process_fd_array(struct bpf_verifier_env *env, union bpf_attr *attr, bpfptr_t uattr)
24057 {
24058 	size_t size = sizeof(int);
24059 	int ret;
24060 	int fd;
24061 	u32 i;
24062 
24063 	env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
24064 
24065 	/*
24066 	 * The only difference between old (no fd_array_cnt is given) and new
24067 	 * APIs is that in the latter case the fd_array is expected to be
24068 	 * continuous and is scanned for map fds right away
24069 	 */
24070 	if (!attr->fd_array_cnt)
24071 		return 0;
24072 
24073 	/* Check for integer overflow */
24074 	if (attr->fd_array_cnt >= (U32_MAX / size)) {
24075 		verbose(env, "fd_array_cnt is too big (%u)\n", attr->fd_array_cnt);
24076 		return -EINVAL;
24077 	}
24078 
24079 	for (i = 0; i < attr->fd_array_cnt; i++) {
24080 		if (copy_from_bpfptr_offset(&fd, env->fd_array, i * size, size))
24081 			return -EFAULT;
24082 
24083 		ret = add_fd_from_fd_array(env, fd);
24084 		if (ret)
24085 			return ret;
24086 	}
24087 
24088 	return 0;
24089 }
24090 
can_fallthrough(struct bpf_insn * insn)24091 static bool can_fallthrough(struct bpf_insn *insn)
24092 {
24093 	u8 class = BPF_CLASS(insn->code);
24094 	u8 opcode = BPF_OP(insn->code);
24095 
24096 	if (class != BPF_JMP && class != BPF_JMP32)
24097 		return true;
24098 
24099 	if (opcode == BPF_EXIT || opcode == BPF_JA)
24100 		return false;
24101 
24102 	return true;
24103 }
24104 
can_jump(struct bpf_insn * insn)24105 static bool can_jump(struct bpf_insn *insn)
24106 {
24107 	u8 class = BPF_CLASS(insn->code);
24108 	u8 opcode = BPF_OP(insn->code);
24109 
24110 	if (class != BPF_JMP && class != BPF_JMP32)
24111 		return false;
24112 
24113 	switch (opcode) {
24114 	case BPF_JA:
24115 	case BPF_JEQ:
24116 	case BPF_JNE:
24117 	case BPF_JLT:
24118 	case BPF_JLE:
24119 	case BPF_JGT:
24120 	case BPF_JGE:
24121 	case BPF_JSGT:
24122 	case BPF_JSGE:
24123 	case BPF_JSLT:
24124 	case BPF_JSLE:
24125 	case BPF_JCOND:
24126 	case BPF_JSET:
24127 		return true;
24128 	}
24129 
24130 	return false;
24131 }
24132 
insn_successors(struct bpf_prog * prog,u32 idx,u32 succ[2])24133 static int insn_successors(struct bpf_prog *prog, u32 idx, u32 succ[2])
24134 {
24135 	struct bpf_insn *insn = &prog->insnsi[idx];
24136 	int i = 0, insn_sz;
24137 	u32 dst;
24138 
24139 	insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
24140 	if (can_fallthrough(insn) && idx + 1 < prog->len)
24141 		succ[i++] = idx + insn_sz;
24142 
24143 	if (can_jump(insn)) {
24144 		dst = idx + jmp_offset(insn) + 1;
24145 		if (i == 0 || succ[0] != dst)
24146 			succ[i++] = dst;
24147 	}
24148 
24149 	return i;
24150 }
24151 
24152 /* Each field is a register bitmask */
24153 struct insn_live_regs {
24154 	u16 use;	/* registers read by instruction */
24155 	u16 def;	/* registers written by instruction */
24156 	u16 in;		/* registers that may be alive before instruction */
24157 	u16 out;	/* registers that may be alive after instruction */
24158 };
24159 
24160 /* Bitmask with 1s for all caller saved registers */
24161 #define ALL_CALLER_SAVED_REGS ((1u << CALLER_SAVED_REGS) - 1)
24162 
24163 /* Compute info->{use,def} fields for the instruction */
compute_insn_live_regs(struct bpf_verifier_env * env,struct bpf_insn * insn,struct insn_live_regs * info)24164 static void compute_insn_live_regs(struct bpf_verifier_env *env,
24165 				   struct bpf_insn *insn,
24166 				   struct insn_live_regs *info)
24167 {
24168 	struct call_summary cs;
24169 	u8 class = BPF_CLASS(insn->code);
24170 	u8 code = BPF_OP(insn->code);
24171 	u8 mode = BPF_MODE(insn->code);
24172 	u16 src = BIT(insn->src_reg);
24173 	u16 dst = BIT(insn->dst_reg);
24174 	u16 r0  = BIT(0);
24175 	u16 def = 0;
24176 	u16 use = 0xffff;
24177 
24178 	switch (class) {
24179 	case BPF_LD:
24180 		switch (mode) {
24181 		case BPF_IMM:
24182 			if (BPF_SIZE(insn->code) == BPF_DW) {
24183 				def = dst;
24184 				use = 0;
24185 			}
24186 			break;
24187 		case BPF_LD | BPF_ABS:
24188 		case BPF_LD | BPF_IND:
24189 			/* stick with defaults */
24190 			break;
24191 		}
24192 		break;
24193 	case BPF_LDX:
24194 		switch (mode) {
24195 		case BPF_MEM:
24196 		case BPF_MEMSX:
24197 			def = dst;
24198 			use = src;
24199 			break;
24200 		}
24201 		break;
24202 	case BPF_ST:
24203 		switch (mode) {
24204 		case BPF_MEM:
24205 			def = 0;
24206 			use = dst;
24207 			break;
24208 		}
24209 		break;
24210 	case BPF_STX:
24211 		switch (mode) {
24212 		case BPF_MEM:
24213 			def = 0;
24214 			use = dst | src;
24215 			break;
24216 		case BPF_ATOMIC:
24217 			switch (insn->imm) {
24218 			case BPF_CMPXCHG:
24219 				use = r0 | dst | src;
24220 				def = r0;
24221 				break;
24222 			case BPF_LOAD_ACQ:
24223 				def = dst;
24224 				use = src;
24225 				break;
24226 			case BPF_STORE_REL:
24227 				def = 0;
24228 				use = dst | src;
24229 				break;
24230 			default:
24231 				use = dst | src;
24232 				if (insn->imm & BPF_FETCH)
24233 					def = src;
24234 				else
24235 					def = 0;
24236 			}
24237 			break;
24238 		}
24239 		break;
24240 	case BPF_ALU:
24241 	case BPF_ALU64:
24242 		switch (code) {
24243 		case BPF_END:
24244 			use = dst;
24245 			def = dst;
24246 			break;
24247 		case BPF_MOV:
24248 			def = dst;
24249 			if (BPF_SRC(insn->code) == BPF_K)
24250 				use = 0;
24251 			else
24252 				use = src;
24253 			break;
24254 		default:
24255 			def = dst;
24256 			if (BPF_SRC(insn->code) == BPF_K)
24257 				use = dst;
24258 			else
24259 				use = dst | src;
24260 		}
24261 		break;
24262 	case BPF_JMP:
24263 	case BPF_JMP32:
24264 		switch (code) {
24265 		case BPF_JA:
24266 		case BPF_JCOND:
24267 			def = 0;
24268 			use = 0;
24269 			break;
24270 		case BPF_EXIT:
24271 			def = 0;
24272 			use = r0;
24273 			break;
24274 		case BPF_CALL:
24275 			def = ALL_CALLER_SAVED_REGS;
24276 			use = def & ~BIT(BPF_REG_0);
24277 			if (get_call_summary(env, insn, &cs))
24278 				use = GENMASK(cs.num_params, 1);
24279 			break;
24280 		default:
24281 			def = 0;
24282 			if (BPF_SRC(insn->code) == BPF_K)
24283 				use = dst;
24284 			else
24285 				use = dst | src;
24286 		}
24287 		break;
24288 	}
24289 
24290 	info->def = def;
24291 	info->use = use;
24292 }
24293 
24294 /* Compute may-live registers after each instruction in the program.
24295  * The register is live after the instruction I if it is read by some
24296  * instruction S following I during program execution and is not
24297  * overwritten between I and S.
24298  *
24299  * Store result in env->insn_aux_data[i].live_regs.
24300  */
compute_live_registers(struct bpf_verifier_env * env)24301 static int compute_live_registers(struct bpf_verifier_env *env)
24302 {
24303 	struct bpf_insn_aux_data *insn_aux = env->insn_aux_data;
24304 	struct bpf_insn *insns = env->prog->insnsi;
24305 	struct insn_live_regs *state;
24306 	int insn_cnt = env->prog->len;
24307 	int err = 0, i, j;
24308 	bool changed;
24309 
24310 	/* Use the following algorithm:
24311 	 * - define the following:
24312 	 *   - I.use : a set of all registers read by instruction I;
24313 	 *   - I.def : a set of all registers written by instruction I;
24314 	 *   - I.in  : a set of all registers that may be alive before I execution;
24315 	 *   - I.out : a set of all registers that may be alive after I execution;
24316 	 *   - insn_successors(I): a set of instructions S that might immediately
24317 	 *                         follow I for some program execution;
24318 	 * - associate separate empty sets 'I.in' and 'I.out' with each instruction;
24319 	 * - visit each instruction in a postorder and update
24320 	 *   state[i].in, state[i].out as follows:
24321 	 *
24322 	 *       state[i].out = U [state[s].in for S in insn_successors(i)]
24323 	 *       state[i].in  = (state[i].out / state[i].def) U state[i].use
24324 	 *
24325 	 *   (where U stands for set union, / stands for set difference)
24326 	 * - repeat the computation while {in,out} fields changes for
24327 	 *   any instruction.
24328 	 */
24329 	state = kvcalloc(insn_cnt, sizeof(*state), GFP_KERNEL_ACCOUNT);
24330 	if (!state) {
24331 		err = -ENOMEM;
24332 		goto out;
24333 	}
24334 
24335 	for (i = 0; i < insn_cnt; ++i)
24336 		compute_insn_live_regs(env, &insns[i], &state[i]);
24337 
24338 	changed = true;
24339 	while (changed) {
24340 		changed = false;
24341 		for (i = 0; i < env->cfg.cur_postorder; ++i) {
24342 			int insn_idx = env->cfg.insn_postorder[i];
24343 			struct insn_live_regs *live = &state[insn_idx];
24344 			int succ_num;
24345 			u32 succ[2];
24346 			u16 new_out = 0;
24347 			u16 new_in = 0;
24348 
24349 			succ_num = insn_successors(env->prog, insn_idx, succ);
24350 			for (int s = 0; s < succ_num; ++s)
24351 				new_out |= state[succ[s]].in;
24352 			new_in = (new_out & ~live->def) | live->use;
24353 			if (new_out != live->out || new_in != live->in) {
24354 				live->in = new_in;
24355 				live->out = new_out;
24356 				changed = true;
24357 			}
24358 		}
24359 	}
24360 
24361 	for (i = 0; i < insn_cnt; ++i)
24362 		insn_aux[i].live_regs_before = state[i].in;
24363 
24364 	if (env->log.level & BPF_LOG_LEVEL2) {
24365 		verbose(env, "Live regs before insn:\n");
24366 		for (i = 0; i < insn_cnt; ++i) {
24367 			if (env->insn_aux_data[i].scc)
24368 				verbose(env, "%3d ", env->insn_aux_data[i].scc);
24369 			else
24370 				verbose(env, "    ");
24371 			verbose(env, "%3d: ", i);
24372 			for (j = BPF_REG_0; j < BPF_REG_10; ++j)
24373 				if (insn_aux[i].live_regs_before & BIT(j))
24374 					verbose(env, "%d", j);
24375 				else
24376 					verbose(env, ".");
24377 			verbose(env, " ");
24378 			verbose_insn(env, &insns[i]);
24379 			if (bpf_is_ldimm64(&insns[i]))
24380 				i++;
24381 		}
24382 	}
24383 
24384 out:
24385 	kvfree(state);
24386 	kvfree(env->cfg.insn_postorder);
24387 	env->cfg.insn_postorder = NULL;
24388 	env->cfg.cur_postorder = 0;
24389 	return err;
24390 }
24391 
24392 /*
24393  * Compute strongly connected components (SCCs) on the CFG.
24394  * Assign an SCC number to each instruction, recorded in env->insn_aux[*].scc.
24395  * If instruction is a sole member of its SCC and there are no self edges,
24396  * assign it SCC number of zero.
24397  * Uses a non-recursive adaptation of Tarjan's algorithm for SCC computation.
24398  */
compute_scc(struct bpf_verifier_env * env)24399 static int compute_scc(struct bpf_verifier_env *env)
24400 {
24401 	const u32 NOT_ON_STACK = U32_MAX;
24402 
24403 	struct bpf_insn_aux_data *aux = env->insn_aux_data;
24404 	const u32 insn_cnt = env->prog->len;
24405 	int stack_sz, dfs_sz, err = 0;
24406 	u32 *stack, *pre, *low, *dfs;
24407 	u32 succ_cnt, i, j, t, w;
24408 	u32 next_preorder_num;
24409 	u32 next_scc_id;
24410 	bool assign_scc;
24411 	u32 succ[2];
24412 
24413 	next_preorder_num = 1;
24414 	next_scc_id = 1;
24415 	/*
24416 	 * - 'stack' accumulates vertices in DFS order, see invariant comment below;
24417 	 * - 'pre[t] == p' => preorder number of vertex 't' is 'p';
24418 	 * - 'low[t] == n' => smallest preorder number of the vertex reachable from 't' is 'n';
24419 	 * - 'dfs' DFS traversal stack, used to emulate explicit recursion.
24420 	 */
24421 	stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
24422 	pre = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
24423 	low = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
24424 	dfs = kvcalloc(insn_cnt, sizeof(*dfs), GFP_KERNEL_ACCOUNT);
24425 	if (!stack || !pre || !low || !dfs) {
24426 		err = -ENOMEM;
24427 		goto exit;
24428 	}
24429 	/*
24430 	 * References:
24431 	 * [1] R. Tarjan "Depth-First Search and Linear Graph Algorithms"
24432 	 * [2] D. J. Pearce "A Space-Efficient Algorithm for Finding Strongly Connected Components"
24433 	 *
24434 	 * The algorithm maintains the following invariant:
24435 	 * - suppose there is a path 'u' ~> 'v', such that 'pre[v] < pre[u]';
24436 	 * - then, vertex 'u' remains on stack while vertex 'v' is on stack.
24437 	 *
24438 	 * Consequently:
24439 	 * - If 'low[v] < pre[v]', there is a path from 'v' to some vertex 'u',
24440 	 *   such that 'pre[u] == low[v]'; vertex 'u' is currently on the stack,
24441 	 *   and thus there is an SCC (loop) containing both 'u' and 'v'.
24442 	 * - If 'low[v] == pre[v]', loops containing 'v' have been explored,
24443 	 *   and 'v' can be considered the root of some SCC.
24444 	 *
24445 	 * Here is a pseudo-code for an explicitly recursive version of the algorithm:
24446 	 *
24447 	 *    NOT_ON_STACK = insn_cnt + 1
24448 	 *    pre = [0] * insn_cnt
24449 	 *    low = [0] * insn_cnt
24450 	 *    scc = [0] * insn_cnt
24451 	 *    stack = []
24452 	 *
24453 	 *    next_preorder_num = 1
24454 	 *    next_scc_id = 1
24455 	 *
24456 	 *    def recur(w):
24457 	 *        nonlocal next_preorder_num
24458 	 *        nonlocal next_scc_id
24459 	 *
24460 	 *        pre[w] = next_preorder_num
24461 	 *        low[w] = next_preorder_num
24462 	 *        next_preorder_num += 1
24463 	 *        stack.append(w)
24464 	 *        for s in successors(w):
24465 	 *            # Note: for classic algorithm the block below should look as:
24466 	 *            #
24467 	 *            # if pre[s] == 0:
24468 	 *            #     recur(s)
24469 	 *            #	    low[w] = min(low[w], low[s])
24470 	 *            # elif low[s] != NOT_ON_STACK:
24471 	 *            #     low[w] = min(low[w], pre[s])
24472 	 *            #
24473 	 *            # But replacing both 'min' instructions with 'low[w] = min(low[w], low[s])'
24474 	 *            # does not break the invariant and makes itartive version of the algorithm
24475 	 *            # simpler. See 'Algorithm #3' from [2].
24476 	 *
24477 	 *            # 's' not yet visited
24478 	 *            if pre[s] == 0:
24479 	 *                recur(s)
24480 	 *            # if 's' is on stack, pick lowest reachable preorder number from it;
24481 	 *            # if 's' is not on stack 'low[s] == NOT_ON_STACK > low[w]',
24482 	 *            # so 'min' would be a noop.
24483 	 *            low[w] = min(low[w], low[s])
24484 	 *
24485 	 *        if low[w] == pre[w]:
24486 	 *            # 'w' is the root of an SCC, pop all vertices
24487 	 *            # below 'w' on stack and assign same SCC to them.
24488 	 *            while True:
24489 	 *                t = stack.pop()
24490 	 *                low[t] = NOT_ON_STACK
24491 	 *                scc[t] = next_scc_id
24492 	 *                if t == w:
24493 	 *                    break
24494 	 *            next_scc_id += 1
24495 	 *
24496 	 *    for i in range(0, insn_cnt):
24497 	 *        if pre[i] == 0:
24498 	 *            recur(i)
24499 	 *
24500 	 * Below implementation replaces explicit recursion with array 'dfs'.
24501 	 */
24502 	for (i = 0; i < insn_cnt; i++) {
24503 		if (pre[i])
24504 			continue;
24505 		stack_sz = 0;
24506 		dfs_sz = 1;
24507 		dfs[0] = i;
24508 dfs_continue:
24509 		while (dfs_sz) {
24510 			w = dfs[dfs_sz - 1];
24511 			if (pre[w] == 0) {
24512 				low[w] = next_preorder_num;
24513 				pre[w] = next_preorder_num;
24514 				next_preorder_num++;
24515 				stack[stack_sz++] = w;
24516 			}
24517 			/* Visit 'w' successors */
24518 			succ_cnt = insn_successors(env->prog, w, succ);
24519 			for (j = 0; j < succ_cnt; ++j) {
24520 				if (pre[succ[j]]) {
24521 					low[w] = min(low[w], low[succ[j]]);
24522 				} else {
24523 					dfs[dfs_sz++] = succ[j];
24524 					goto dfs_continue;
24525 				}
24526 			}
24527 			/*
24528 			 * Preserve the invariant: if some vertex above in the stack
24529 			 * is reachable from 'w', keep 'w' on the stack.
24530 			 */
24531 			if (low[w] < pre[w]) {
24532 				dfs_sz--;
24533 				goto dfs_continue;
24534 			}
24535 			/*
24536 			 * Assign SCC number only if component has two or more elements,
24537 			 * or if component has a self reference.
24538 			 */
24539 			assign_scc = stack[stack_sz - 1] != w;
24540 			for (j = 0; j < succ_cnt; ++j) {
24541 				if (succ[j] == w) {
24542 					assign_scc = true;
24543 					break;
24544 				}
24545 			}
24546 			/* Pop component elements from stack */
24547 			do {
24548 				t = stack[--stack_sz];
24549 				low[t] = NOT_ON_STACK;
24550 				if (assign_scc)
24551 					aux[t].scc = next_scc_id;
24552 			} while (t != w);
24553 			if (assign_scc)
24554 				next_scc_id++;
24555 			dfs_sz--;
24556 		}
24557 	}
24558 	env->scc_info = kvcalloc(next_scc_id, sizeof(*env->scc_info), GFP_KERNEL_ACCOUNT);
24559 	if (!env->scc_info) {
24560 		err = -ENOMEM;
24561 		goto exit;
24562 	}
24563 	env->scc_cnt = next_scc_id;
24564 exit:
24565 	kvfree(stack);
24566 	kvfree(pre);
24567 	kvfree(low);
24568 	kvfree(dfs);
24569 	return err;
24570 }
24571 
bpf_check(struct bpf_prog ** prog,union bpf_attr * attr,bpfptr_t uattr,__u32 uattr_size)24572 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
24573 {
24574 	u64 start_time = ktime_get_ns();
24575 	struct bpf_verifier_env *env;
24576 	int i, len, ret = -EINVAL, err;
24577 	u32 log_true_size;
24578 	bool is_priv;
24579 
24580 	BTF_TYPE_EMIT(enum bpf_features);
24581 
24582 	/* no program is valid */
24583 	if (ARRAY_SIZE(bpf_verifier_ops) == 0)
24584 		return -EINVAL;
24585 
24586 	/* 'struct bpf_verifier_env' can be global, but since it's not small,
24587 	 * allocate/free it every time bpf_check() is called
24588 	 */
24589 	env = kvzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL_ACCOUNT);
24590 	if (!env)
24591 		return -ENOMEM;
24592 
24593 	env->bt.env = env;
24594 
24595 	len = (*prog)->len;
24596 	env->insn_aux_data =
24597 		vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
24598 	ret = -ENOMEM;
24599 	if (!env->insn_aux_data)
24600 		goto err_free_env;
24601 	for (i = 0; i < len; i++)
24602 		env->insn_aux_data[i].orig_idx = i;
24603 	env->prog = *prog;
24604 	env->ops = bpf_verifier_ops[env->prog->type];
24605 
24606 	env->allow_ptr_leaks = bpf_allow_ptr_leaks(env->prog->aux->token);
24607 	env->allow_uninit_stack = bpf_allow_uninit_stack(env->prog->aux->token);
24608 	env->bypass_spec_v1 = bpf_bypass_spec_v1(env->prog->aux->token);
24609 	env->bypass_spec_v4 = bpf_bypass_spec_v4(env->prog->aux->token);
24610 	env->bpf_capable = is_priv = bpf_token_capable(env->prog->aux->token, CAP_BPF);
24611 
24612 	bpf_get_btf_vmlinux();
24613 
24614 	/* grab the mutex to protect few globals used by verifier */
24615 	if (!is_priv)
24616 		mutex_lock(&bpf_verifier_lock);
24617 
24618 	/* user could have requested verbose verifier output
24619 	 * and supplied buffer to store the verification trace
24620 	 */
24621 	ret = bpf_vlog_init(&env->log, attr->log_level,
24622 			    (char __user *) (unsigned long) attr->log_buf,
24623 			    attr->log_size);
24624 	if (ret)
24625 		goto err_unlock;
24626 
24627 	ret = process_fd_array(env, attr, uattr);
24628 	if (ret)
24629 		goto skip_full_check;
24630 
24631 	mark_verifier_state_clean(env);
24632 
24633 	if (IS_ERR(btf_vmlinux)) {
24634 		/* Either gcc or pahole or kernel are broken. */
24635 		verbose(env, "in-kernel BTF is malformed\n");
24636 		ret = PTR_ERR(btf_vmlinux);
24637 		goto skip_full_check;
24638 	}
24639 
24640 	env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
24641 	if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
24642 		env->strict_alignment = true;
24643 	if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
24644 		env->strict_alignment = false;
24645 
24646 	if (is_priv)
24647 		env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
24648 	env->test_reg_invariants = attr->prog_flags & BPF_F_TEST_REG_INVARIANTS;
24649 
24650 	env->explored_states = kvcalloc(state_htab_size(env),
24651 				       sizeof(struct list_head),
24652 				       GFP_KERNEL_ACCOUNT);
24653 	ret = -ENOMEM;
24654 	if (!env->explored_states)
24655 		goto skip_full_check;
24656 
24657 	for (i = 0; i < state_htab_size(env); i++)
24658 		INIT_LIST_HEAD(&env->explored_states[i]);
24659 	INIT_LIST_HEAD(&env->free_list);
24660 
24661 	ret = check_btf_info_early(env, attr, uattr);
24662 	if (ret < 0)
24663 		goto skip_full_check;
24664 
24665 	ret = add_subprog_and_kfunc(env);
24666 	if (ret < 0)
24667 		goto skip_full_check;
24668 
24669 	ret = check_subprogs(env);
24670 	if (ret < 0)
24671 		goto skip_full_check;
24672 
24673 	ret = check_btf_info(env, attr, uattr);
24674 	if (ret < 0)
24675 		goto skip_full_check;
24676 
24677 	ret = resolve_pseudo_ldimm64(env);
24678 	if (ret < 0)
24679 		goto skip_full_check;
24680 
24681 	if (bpf_prog_is_offloaded(env->prog->aux)) {
24682 		ret = bpf_prog_offload_verifier_prep(env->prog);
24683 		if (ret)
24684 			goto skip_full_check;
24685 	}
24686 
24687 	ret = check_cfg(env);
24688 	if (ret < 0)
24689 		goto skip_full_check;
24690 
24691 	ret = check_attach_btf_id(env);
24692 	if (ret)
24693 		goto skip_full_check;
24694 
24695 	ret = compute_scc(env);
24696 	if (ret < 0)
24697 		goto skip_full_check;
24698 
24699 	ret = compute_live_registers(env);
24700 	if (ret < 0)
24701 		goto skip_full_check;
24702 
24703 	ret = mark_fastcall_patterns(env);
24704 	if (ret < 0)
24705 		goto skip_full_check;
24706 
24707 	ret = do_check_main(env);
24708 	ret = ret ?: do_check_subprogs(env);
24709 
24710 	if (ret == 0 && bpf_prog_is_offloaded(env->prog->aux))
24711 		ret = bpf_prog_offload_finalize(env);
24712 
24713 skip_full_check:
24714 	kvfree(env->explored_states);
24715 
24716 	/* might decrease stack depth, keep it before passes that
24717 	 * allocate additional slots.
24718 	 */
24719 	if (ret == 0)
24720 		ret = remove_fastcall_spills_fills(env);
24721 
24722 	if (ret == 0)
24723 		ret = check_max_stack_depth(env);
24724 
24725 	/* instruction rewrites happen after this point */
24726 	if (ret == 0)
24727 		ret = optimize_bpf_loop(env);
24728 
24729 	if (is_priv) {
24730 		if (ret == 0)
24731 			opt_hard_wire_dead_code_branches(env);
24732 		if (ret == 0)
24733 			ret = opt_remove_dead_code(env);
24734 		if (ret == 0)
24735 			ret = opt_remove_nops(env);
24736 	} else {
24737 		if (ret == 0)
24738 			sanitize_dead_code(env);
24739 	}
24740 
24741 	if (ret == 0)
24742 		/* program is valid, convert *(u32*)(ctx + off) accesses */
24743 		ret = convert_ctx_accesses(env);
24744 
24745 	if (ret == 0)
24746 		ret = do_misc_fixups(env);
24747 
24748 	/* do 32-bit optimization after insn patching has done so those patched
24749 	 * insns could be handled correctly.
24750 	 */
24751 	if (ret == 0 && !bpf_prog_is_offloaded(env->prog->aux)) {
24752 		ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
24753 		env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
24754 								     : false;
24755 	}
24756 
24757 	if (ret == 0)
24758 		ret = fixup_call_args(env);
24759 
24760 	env->verification_time = ktime_get_ns() - start_time;
24761 	print_verification_stats(env);
24762 	env->prog->aux->verified_insns = env->insn_processed;
24763 
24764 	/* preserve original error even if log finalization is successful */
24765 	err = bpf_vlog_finalize(&env->log, &log_true_size);
24766 	if (err)
24767 		ret = err;
24768 
24769 	if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
24770 	    copy_to_bpfptr_offset(uattr, offsetof(union bpf_attr, log_true_size),
24771 				  &log_true_size, sizeof(log_true_size))) {
24772 		ret = -EFAULT;
24773 		goto err_release_maps;
24774 	}
24775 
24776 	if (ret)
24777 		goto err_release_maps;
24778 
24779 	if (env->used_map_cnt) {
24780 		/* if program passed verifier, update used_maps in bpf_prog_info */
24781 		env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
24782 							  sizeof(env->used_maps[0]),
24783 							  GFP_KERNEL_ACCOUNT);
24784 
24785 		if (!env->prog->aux->used_maps) {
24786 			ret = -ENOMEM;
24787 			goto err_release_maps;
24788 		}
24789 
24790 		memcpy(env->prog->aux->used_maps, env->used_maps,
24791 		       sizeof(env->used_maps[0]) * env->used_map_cnt);
24792 		env->prog->aux->used_map_cnt = env->used_map_cnt;
24793 	}
24794 	if (env->used_btf_cnt) {
24795 		/* if program passed verifier, update used_btfs in bpf_prog_aux */
24796 		env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
24797 							  sizeof(env->used_btfs[0]),
24798 							  GFP_KERNEL_ACCOUNT);
24799 		if (!env->prog->aux->used_btfs) {
24800 			ret = -ENOMEM;
24801 			goto err_release_maps;
24802 		}
24803 
24804 		memcpy(env->prog->aux->used_btfs, env->used_btfs,
24805 		       sizeof(env->used_btfs[0]) * env->used_btf_cnt);
24806 		env->prog->aux->used_btf_cnt = env->used_btf_cnt;
24807 	}
24808 	if (env->used_map_cnt || env->used_btf_cnt) {
24809 		/* program is valid. Convert pseudo bpf_ld_imm64 into generic
24810 		 * bpf_ld_imm64 instructions
24811 		 */
24812 		convert_pseudo_ld_imm64(env);
24813 	}
24814 
24815 	adjust_btf_func(env);
24816 
24817 err_release_maps:
24818 	if (!env->prog->aux->used_maps)
24819 		/* if we didn't copy map pointers into bpf_prog_info, release
24820 		 * them now. Otherwise free_used_maps() will release them.
24821 		 */
24822 		release_maps(env);
24823 	if (!env->prog->aux->used_btfs)
24824 		release_btfs(env);
24825 
24826 	/* extension progs temporarily inherit the attach_type of their targets
24827 	   for verification purposes, so set it back to zero before returning
24828 	 */
24829 	if (env->prog->type == BPF_PROG_TYPE_EXT)
24830 		env->prog->expected_attach_type = 0;
24831 
24832 	*prog = env->prog;
24833 
24834 	module_put(env->attach_btf_mod);
24835 err_unlock:
24836 	if (!is_priv)
24837 		mutex_unlock(&bpf_verifier_lock);
24838 	vfree(env->insn_aux_data);
24839 err_free_env:
24840 	kvfree(env->cfg.insn_postorder);
24841 	kvfree(env->scc_info);
24842 	kvfree(env);
24843 	return ret;
24844 }
24845