xref: /linux/kernel/bpf/verifier.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2  *
3  * This program is free software; you can redistribute it and/or
4  * modify it under the terms of version 2 of the GNU General Public
5  * License as published by the Free Software Foundation.
6  *
7  * This program is distributed in the hope that it will be useful, but
8  * WITHOUT ANY WARRANTY; without even the implied warranty of
9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
10  * General Public License for more details.
11  */
12 #include <linux/kernel.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/bpf.h>
16 #include <linux/filter.h>
17 #include <net/netlink.h>
18 #include <linux/file.h>
19 #include <linux/vmalloc.h>
20 
21 /* bpf_check() is a static code analyzer that walks eBPF program
22  * instruction by instruction and updates register/stack state.
23  * All paths of conditional branches are analyzed until 'bpf_exit' insn.
24  *
25  * The first pass is depth-first-search to check that the program is a DAG.
26  * It rejects the following programs:
27  * - larger than BPF_MAXINSNS insns
28  * - if loop is present (detected via back-edge)
29  * - unreachable insns exist (shouldn't be a forest. program = one function)
30  * - out of bounds or malformed jumps
31  * The second pass is all possible path descent from the 1st insn.
32  * Since it's analyzing all pathes through the program, the length of the
33  * analysis is limited to 32k insn, which may be hit even if total number of
34  * insn is less then 4K, but there are too many branches that change stack/regs.
35  * Number of 'branches to be analyzed' is limited to 1k
36  *
37  * On entry to each instruction, each register has a type, and the instruction
38  * changes the types of the registers depending on instruction semantics.
39  * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
40  * copied to R1.
41  *
42  * All registers are 64-bit.
43  * R0 - return register
44  * R1-R5 argument passing registers
45  * R6-R9 callee saved registers
46  * R10 - frame pointer read-only
47  *
48  * At the start of BPF program the register R1 contains a pointer to bpf_context
49  * and has type PTR_TO_CTX.
50  *
51  * Verifier tracks arithmetic operations on pointers in case:
52  *    BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
53  *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
54  * 1st insn copies R10 (which has FRAME_PTR) type into R1
55  * and 2nd arithmetic instruction is pattern matched to recognize
56  * that it wants to construct a pointer to some element within stack.
57  * So after 2nd insn, the register R1 has type PTR_TO_STACK
58  * (and -20 constant is saved for further stack bounds checking).
59  * Meaning that this reg is a pointer to stack plus known immediate constant.
60  *
61  * Most of the time the registers have UNKNOWN_VALUE type, which
62  * means the register has some value, but it's not a valid pointer.
63  * (like pointer plus pointer becomes UNKNOWN_VALUE type)
64  *
65  * When verifier sees load or store instructions the type of base register
66  * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, FRAME_PTR. These are three pointer
67  * types recognized by check_mem_access() function.
68  *
69  * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
70  * and the range of [ptr, ptr + map's value_size) is accessible.
71  *
72  * registers used to pass values to function calls are checked against
73  * function argument constraints.
74  *
75  * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
76  * It means that the register type passed to this function must be
77  * PTR_TO_STACK and it will be used inside the function as
78  * 'pointer to map element key'
79  *
80  * For example the argument constraints for bpf_map_lookup_elem():
81  *   .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
82  *   .arg1_type = ARG_CONST_MAP_PTR,
83  *   .arg2_type = ARG_PTR_TO_MAP_KEY,
84  *
85  * ret_type says that this function returns 'pointer to map elem value or null'
86  * function expects 1st argument to be a const pointer to 'struct bpf_map' and
87  * 2nd argument should be a pointer to stack, which will be used inside
88  * the helper function as a pointer to map element key.
89  *
90  * On the kernel side the helper function looks like:
91  * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
92  * {
93  *    struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
94  *    void *key = (void *) (unsigned long) r2;
95  *    void *value;
96  *
97  *    here kernel can access 'key' and 'map' pointers safely, knowing that
98  *    [key, key + map->key_size) bytes are valid and were initialized on
99  *    the stack of eBPF program.
100  * }
101  *
102  * Corresponding eBPF program may look like:
103  *    BPF_MOV64_REG(BPF_REG_2, BPF_REG_10),  // after this insn R2 type is FRAME_PTR
104  *    BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
105  *    BPF_LD_MAP_FD(BPF_REG_1, map_fd),      // after this insn R1 type is CONST_PTR_TO_MAP
106  *    BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
107  * here verifier looks at prototype of map_lookup_elem() and sees:
108  * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
109  * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
110  *
111  * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
112  * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
113  * and were initialized prior to this call.
114  * If it's ok, then verifier allows this BPF_CALL insn and looks at
115  * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
116  * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
117  * returns ether pointer to map value or NULL.
118  *
119  * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
120  * insn, the register holding that pointer in the true branch changes state to
121  * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
122  * branch. See check_cond_jmp_op().
123  *
124  * After the call R0 is set to return type of the function and registers R1-R5
125  * are set to NOT_INIT to indicate that they are no longer readable.
126  */
127 
128 /* types of values stored in eBPF registers */
129 enum bpf_reg_type {
130 	NOT_INIT = 0,		 /* nothing was written into register */
131 	UNKNOWN_VALUE,		 /* reg doesn't contain a valid pointer */
132 	PTR_TO_CTX,		 /* reg points to bpf_context */
133 	CONST_PTR_TO_MAP,	 /* reg points to struct bpf_map */
134 	PTR_TO_MAP_VALUE,	 /* reg points to map element value */
135 	PTR_TO_MAP_VALUE_OR_NULL,/* points to map elem value or NULL */
136 	FRAME_PTR,		 /* reg == frame_pointer */
137 	PTR_TO_STACK,		 /* reg == frame_pointer + imm */
138 	CONST_IMM,		 /* constant integer value */
139 };
140 
141 struct reg_state {
142 	enum bpf_reg_type type;
143 	union {
144 		/* valid when type == CONST_IMM | PTR_TO_STACK */
145 		int imm;
146 
147 		/* valid when type == CONST_PTR_TO_MAP | PTR_TO_MAP_VALUE |
148 		 *   PTR_TO_MAP_VALUE_OR_NULL
149 		 */
150 		struct bpf_map *map_ptr;
151 	};
152 };
153 
154 enum bpf_stack_slot_type {
155 	STACK_INVALID,    /* nothing was stored in this stack slot */
156 	STACK_SPILL,      /* register spilled into stack */
157 	STACK_MISC	  /* BPF program wrote some data into this slot */
158 };
159 
160 #define BPF_REG_SIZE 8	/* size of eBPF register in bytes */
161 
162 /* state of the program:
163  * type of all registers and stack info
164  */
165 struct verifier_state {
166 	struct reg_state regs[MAX_BPF_REG];
167 	u8 stack_slot_type[MAX_BPF_STACK];
168 	struct reg_state spilled_regs[MAX_BPF_STACK / BPF_REG_SIZE];
169 };
170 
171 /* linked list of verifier states used to prune search */
172 struct verifier_state_list {
173 	struct verifier_state state;
174 	struct verifier_state_list *next;
175 };
176 
177 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
178 struct verifier_stack_elem {
179 	/* verifer state is 'st'
180 	 * before processing instruction 'insn_idx'
181 	 * and after processing instruction 'prev_insn_idx'
182 	 */
183 	struct verifier_state st;
184 	int insn_idx;
185 	int prev_insn_idx;
186 	struct verifier_stack_elem *next;
187 };
188 
189 #define MAX_USED_MAPS 64 /* max number of maps accessed by one eBPF program */
190 
191 /* single container for all structs
192  * one verifier_env per bpf_check() call
193  */
194 struct verifier_env {
195 	struct bpf_prog *prog;		/* eBPF program being verified */
196 	struct verifier_stack_elem *head; /* stack of verifier states to be processed */
197 	int stack_size;			/* number of states to be processed */
198 	struct verifier_state cur_state; /* current verifier state */
199 	struct verifier_state_list **explored_states; /* search pruning optimization */
200 	struct bpf_map *used_maps[MAX_USED_MAPS]; /* array of map's used by eBPF program */
201 	u32 used_map_cnt;		/* number of used maps */
202 };
203 
204 /* verbose verifier prints what it's seeing
205  * bpf_check() is called under lock, so no race to access these global vars
206  */
207 static u32 log_level, log_size, log_len;
208 static char *log_buf;
209 
210 static DEFINE_MUTEX(bpf_verifier_lock);
211 
212 /* log_level controls verbosity level of eBPF verifier.
213  * verbose() is used to dump the verification trace to the log, so the user
214  * can figure out what's wrong with the program
215  */
216 static void verbose(const char *fmt, ...)
217 {
218 	va_list args;
219 
220 	if (log_level == 0 || log_len >= log_size - 1)
221 		return;
222 
223 	va_start(args, fmt);
224 	log_len += vscnprintf(log_buf + log_len, log_size - log_len, fmt, args);
225 	va_end(args);
226 }
227 
228 /* string representation of 'enum bpf_reg_type' */
229 static const char * const reg_type_str[] = {
230 	[NOT_INIT]		= "?",
231 	[UNKNOWN_VALUE]		= "inv",
232 	[PTR_TO_CTX]		= "ctx",
233 	[CONST_PTR_TO_MAP]	= "map_ptr",
234 	[PTR_TO_MAP_VALUE]	= "map_value",
235 	[PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
236 	[FRAME_PTR]		= "fp",
237 	[PTR_TO_STACK]		= "fp",
238 	[CONST_IMM]		= "imm",
239 };
240 
241 static const struct {
242 	int map_type;
243 	int func_id;
244 } func_limit[] = {
245 	{BPF_MAP_TYPE_PROG_ARRAY, BPF_FUNC_tail_call},
246 	{BPF_MAP_TYPE_PERF_EVENT_ARRAY, BPF_FUNC_perf_event_read},
247 };
248 
249 static void print_verifier_state(struct verifier_env *env)
250 {
251 	enum bpf_reg_type t;
252 	int i;
253 
254 	for (i = 0; i < MAX_BPF_REG; i++) {
255 		t = env->cur_state.regs[i].type;
256 		if (t == NOT_INIT)
257 			continue;
258 		verbose(" R%d=%s", i, reg_type_str[t]);
259 		if (t == CONST_IMM || t == PTR_TO_STACK)
260 			verbose("%d", env->cur_state.regs[i].imm);
261 		else if (t == CONST_PTR_TO_MAP || t == PTR_TO_MAP_VALUE ||
262 			 t == PTR_TO_MAP_VALUE_OR_NULL)
263 			verbose("(ks=%d,vs=%d)",
264 				env->cur_state.regs[i].map_ptr->key_size,
265 				env->cur_state.regs[i].map_ptr->value_size);
266 	}
267 	for (i = 0; i < MAX_BPF_STACK; i += BPF_REG_SIZE) {
268 		if (env->cur_state.stack_slot_type[i] == STACK_SPILL)
269 			verbose(" fp%d=%s", -MAX_BPF_STACK + i,
270 				reg_type_str[env->cur_state.spilled_regs[i / BPF_REG_SIZE].type]);
271 	}
272 	verbose("\n");
273 }
274 
275 static const char *const bpf_class_string[] = {
276 	[BPF_LD]    = "ld",
277 	[BPF_LDX]   = "ldx",
278 	[BPF_ST]    = "st",
279 	[BPF_STX]   = "stx",
280 	[BPF_ALU]   = "alu",
281 	[BPF_JMP]   = "jmp",
282 	[BPF_RET]   = "BUG",
283 	[BPF_ALU64] = "alu64",
284 };
285 
286 static const char *const bpf_alu_string[16] = {
287 	[BPF_ADD >> 4]  = "+=",
288 	[BPF_SUB >> 4]  = "-=",
289 	[BPF_MUL >> 4]  = "*=",
290 	[BPF_DIV >> 4]  = "/=",
291 	[BPF_OR  >> 4]  = "|=",
292 	[BPF_AND >> 4]  = "&=",
293 	[BPF_LSH >> 4]  = "<<=",
294 	[BPF_RSH >> 4]  = ">>=",
295 	[BPF_NEG >> 4]  = "neg",
296 	[BPF_MOD >> 4]  = "%=",
297 	[BPF_XOR >> 4]  = "^=",
298 	[BPF_MOV >> 4]  = "=",
299 	[BPF_ARSH >> 4] = "s>>=",
300 	[BPF_END >> 4]  = "endian",
301 };
302 
303 static const char *const bpf_ldst_string[] = {
304 	[BPF_W >> 3]  = "u32",
305 	[BPF_H >> 3]  = "u16",
306 	[BPF_B >> 3]  = "u8",
307 	[BPF_DW >> 3] = "u64",
308 };
309 
310 static const char *const bpf_jmp_string[16] = {
311 	[BPF_JA >> 4]   = "jmp",
312 	[BPF_JEQ >> 4]  = "==",
313 	[BPF_JGT >> 4]  = ">",
314 	[BPF_JGE >> 4]  = ">=",
315 	[BPF_JSET >> 4] = "&",
316 	[BPF_JNE >> 4]  = "!=",
317 	[BPF_JSGT >> 4] = "s>",
318 	[BPF_JSGE >> 4] = "s>=",
319 	[BPF_CALL >> 4] = "call",
320 	[BPF_EXIT >> 4] = "exit",
321 };
322 
323 static void print_bpf_insn(struct bpf_insn *insn)
324 {
325 	u8 class = BPF_CLASS(insn->code);
326 
327 	if (class == BPF_ALU || class == BPF_ALU64) {
328 		if (BPF_SRC(insn->code) == BPF_X)
329 			verbose("(%02x) %sr%d %s %sr%d\n",
330 				insn->code, class == BPF_ALU ? "(u32) " : "",
331 				insn->dst_reg,
332 				bpf_alu_string[BPF_OP(insn->code) >> 4],
333 				class == BPF_ALU ? "(u32) " : "",
334 				insn->src_reg);
335 		else
336 			verbose("(%02x) %sr%d %s %s%d\n",
337 				insn->code, class == BPF_ALU ? "(u32) " : "",
338 				insn->dst_reg,
339 				bpf_alu_string[BPF_OP(insn->code) >> 4],
340 				class == BPF_ALU ? "(u32) " : "",
341 				insn->imm);
342 	} else if (class == BPF_STX) {
343 		if (BPF_MODE(insn->code) == BPF_MEM)
344 			verbose("(%02x) *(%s *)(r%d %+d) = r%d\n",
345 				insn->code,
346 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
347 				insn->dst_reg,
348 				insn->off, insn->src_reg);
349 		else if (BPF_MODE(insn->code) == BPF_XADD)
350 			verbose("(%02x) lock *(%s *)(r%d %+d) += r%d\n",
351 				insn->code,
352 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
353 				insn->dst_reg, insn->off,
354 				insn->src_reg);
355 		else
356 			verbose("BUG_%02x\n", insn->code);
357 	} else if (class == BPF_ST) {
358 		if (BPF_MODE(insn->code) != BPF_MEM) {
359 			verbose("BUG_st_%02x\n", insn->code);
360 			return;
361 		}
362 		verbose("(%02x) *(%s *)(r%d %+d) = %d\n",
363 			insn->code,
364 			bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
365 			insn->dst_reg,
366 			insn->off, insn->imm);
367 	} else if (class == BPF_LDX) {
368 		if (BPF_MODE(insn->code) != BPF_MEM) {
369 			verbose("BUG_ldx_%02x\n", insn->code);
370 			return;
371 		}
372 		verbose("(%02x) r%d = *(%s *)(r%d %+d)\n",
373 			insn->code, insn->dst_reg,
374 			bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
375 			insn->src_reg, insn->off);
376 	} else if (class == BPF_LD) {
377 		if (BPF_MODE(insn->code) == BPF_ABS) {
378 			verbose("(%02x) r0 = *(%s *)skb[%d]\n",
379 				insn->code,
380 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
381 				insn->imm);
382 		} else if (BPF_MODE(insn->code) == BPF_IND) {
383 			verbose("(%02x) r0 = *(%s *)skb[r%d + %d]\n",
384 				insn->code,
385 				bpf_ldst_string[BPF_SIZE(insn->code) >> 3],
386 				insn->src_reg, insn->imm);
387 		} else if (BPF_MODE(insn->code) == BPF_IMM) {
388 			verbose("(%02x) r%d = 0x%x\n",
389 				insn->code, insn->dst_reg, insn->imm);
390 		} else {
391 			verbose("BUG_ld_%02x\n", insn->code);
392 			return;
393 		}
394 	} else if (class == BPF_JMP) {
395 		u8 opcode = BPF_OP(insn->code);
396 
397 		if (opcode == BPF_CALL) {
398 			verbose("(%02x) call %d\n", insn->code, insn->imm);
399 		} else if (insn->code == (BPF_JMP | BPF_JA)) {
400 			verbose("(%02x) goto pc%+d\n",
401 				insn->code, insn->off);
402 		} else if (insn->code == (BPF_JMP | BPF_EXIT)) {
403 			verbose("(%02x) exit\n", insn->code);
404 		} else if (BPF_SRC(insn->code) == BPF_X) {
405 			verbose("(%02x) if r%d %s r%d goto pc%+d\n",
406 				insn->code, insn->dst_reg,
407 				bpf_jmp_string[BPF_OP(insn->code) >> 4],
408 				insn->src_reg, insn->off);
409 		} else {
410 			verbose("(%02x) if r%d %s 0x%x goto pc%+d\n",
411 				insn->code, insn->dst_reg,
412 				bpf_jmp_string[BPF_OP(insn->code) >> 4],
413 				insn->imm, insn->off);
414 		}
415 	} else {
416 		verbose("(%02x) %s\n", insn->code, bpf_class_string[class]);
417 	}
418 }
419 
420 static int pop_stack(struct verifier_env *env, int *prev_insn_idx)
421 {
422 	struct verifier_stack_elem *elem;
423 	int insn_idx;
424 
425 	if (env->head == NULL)
426 		return -1;
427 
428 	memcpy(&env->cur_state, &env->head->st, sizeof(env->cur_state));
429 	insn_idx = env->head->insn_idx;
430 	if (prev_insn_idx)
431 		*prev_insn_idx = env->head->prev_insn_idx;
432 	elem = env->head->next;
433 	kfree(env->head);
434 	env->head = elem;
435 	env->stack_size--;
436 	return insn_idx;
437 }
438 
439 static struct verifier_state *push_stack(struct verifier_env *env, int insn_idx,
440 					 int prev_insn_idx)
441 {
442 	struct verifier_stack_elem *elem;
443 
444 	elem = kmalloc(sizeof(struct verifier_stack_elem), GFP_KERNEL);
445 	if (!elem)
446 		goto err;
447 
448 	memcpy(&elem->st, &env->cur_state, sizeof(env->cur_state));
449 	elem->insn_idx = insn_idx;
450 	elem->prev_insn_idx = prev_insn_idx;
451 	elem->next = env->head;
452 	env->head = elem;
453 	env->stack_size++;
454 	if (env->stack_size > 1024) {
455 		verbose("BPF program is too complex\n");
456 		goto err;
457 	}
458 	return &elem->st;
459 err:
460 	/* pop all elements and return */
461 	while (pop_stack(env, NULL) >= 0);
462 	return NULL;
463 }
464 
465 #define CALLER_SAVED_REGS 6
466 static const int caller_saved[CALLER_SAVED_REGS] = {
467 	BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
468 };
469 
470 static void init_reg_state(struct reg_state *regs)
471 {
472 	int i;
473 
474 	for (i = 0; i < MAX_BPF_REG; i++) {
475 		regs[i].type = NOT_INIT;
476 		regs[i].imm = 0;
477 		regs[i].map_ptr = NULL;
478 	}
479 
480 	/* frame pointer */
481 	regs[BPF_REG_FP].type = FRAME_PTR;
482 
483 	/* 1st arg to a function */
484 	regs[BPF_REG_1].type = PTR_TO_CTX;
485 }
486 
487 static void mark_reg_unknown_value(struct reg_state *regs, u32 regno)
488 {
489 	BUG_ON(regno >= MAX_BPF_REG);
490 	regs[regno].type = UNKNOWN_VALUE;
491 	regs[regno].imm = 0;
492 	regs[regno].map_ptr = NULL;
493 }
494 
495 enum reg_arg_type {
496 	SRC_OP,		/* register is used as source operand */
497 	DST_OP,		/* register is used as destination operand */
498 	DST_OP_NO_MARK	/* same as above, check only, don't mark */
499 };
500 
501 static int check_reg_arg(struct reg_state *regs, u32 regno,
502 			 enum reg_arg_type t)
503 {
504 	if (regno >= MAX_BPF_REG) {
505 		verbose("R%d is invalid\n", regno);
506 		return -EINVAL;
507 	}
508 
509 	if (t == SRC_OP) {
510 		/* check whether register used as source operand can be read */
511 		if (regs[regno].type == NOT_INIT) {
512 			verbose("R%d !read_ok\n", regno);
513 			return -EACCES;
514 		}
515 	} else {
516 		/* check whether register used as dest operand can be written to */
517 		if (regno == BPF_REG_FP) {
518 			verbose("frame pointer is read only\n");
519 			return -EACCES;
520 		}
521 		if (t == DST_OP)
522 			mark_reg_unknown_value(regs, regno);
523 	}
524 	return 0;
525 }
526 
527 static int bpf_size_to_bytes(int bpf_size)
528 {
529 	if (bpf_size == BPF_W)
530 		return 4;
531 	else if (bpf_size == BPF_H)
532 		return 2;
533 	else if (bpf_size == BPF_B)
534 		return 1;
535 	else if (bpf_size == BPF_DW)
536 		return 8;
537 	else
538 		return -EINVAL;
539 }
540 
541 /* check_stack_read/write functions track spill/fill of registers,
542  * stack boundary and alignment are checked in check_mem_access()
543  */
544 static int check_stack_write(struct verifier_state *state, int off, int size,
545 			     int value_regno)
546 {
547 	int i;
548 	/* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
549 	 * so it's aligned access and [off, off + size) are within stack limits
550 	 */
551 
552 	if (value_regno >= 0 &&
553 	    (state->regs[value_regno].type == PTR_TO_MAP_VALUE ||
554 	     state->regs[value_regno].type == PTR_TO_STACK ||
555 	     state->regs[value_regno].type == PTR_TO_CTX)) {
556 
557 		/* register containing pointer is being spilled into stack */
558 		if (size != BPF_REG_SIZE) {
559 			verbose("invalid size of register spill\n");
560 			return -EACCES;
561 		}
562 
563 		/* save register state */
564 		state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] =
565 			state->regs[value_regno];
566 
567 		for (i = 0; i < BPF_REG_SIZE; i++)
568 			state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_SPILL;
569 	} else {
570 		/* regular write of data into stack */
571 		state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE] =
572 			(struct reg_state) {};
573 
574 		for (i = 0; i < size; i++)
575 			state->stack_slot_type[MAX_BPF_STACK + off + i] = STACK_MISC;
576 	}
577 	return 0;
578 }
579 
580 static int check_stack_read(struct verifier_state *state, int off, int size,
581 			    int value_regno)
582 {
583 	u8 *slot_type;
584 	int i;
585 
586 	slot_type = &state->stack_slot_type[MAX_BPF_STACK + off];
587 
588 	if (slot_type[0] == STACK_SPILL) {
589 		if (size != BPF_REG_SIZE) {
590 			verbose("invalid size of register spill\n");
591 			return -EACCES;
592 		}
593 		for (i = 1; i < BPF_REG_SIZE; i++) {
594 			if (slot_type[i] != STACK_SPILL) {
595 				verbose("corrupted spill memory\n");
596 				return -EACCES;
597 			}
598 		}
599 
600 		if (value_regno >= 0)
601 			/* restore register state from stack */
602 			state->regs[value_regno] =
603 				state->spilled_regs[(MAX_BPF_STACK + off) / BPF_REG_SIZE];
604 		return 0;
605 	} else {
606 		for (i = 0; i < size; i++) {
607 			if (slot_type[i] != STACK_MISC) {
608 				verbose("invalid read from stack off %d+%d size %d\n",
609 					off, i, size);
610 				return -EACCES;
611 			}
612 		}
613 		if (value_regno >= 0)
614 			/* have read misc data from the stack */
615 			mark_reg_unknown_value(state->regs, value_regno);
616 		return 0;
617 	}
618 }
619 
620 /* check read/write into map element returned by bpf_map_lookup_elem() */
621 static int check_map_access(struct verifier_env *env, u32 regno, int off,
622 			    int size)
623 {
624 	struct bpf_map *map = env->cur_state.regs[regno].map_ptr;
625 
626 	if (off < 0 || off + size > map->value_size) {
627 		verbose("invalid access to map value, value_size=%d off=%d size=%d\n",
628 			map->value_size, off, size);
629 		return -EACCES;
630 	}
631 	return 0;
632 }
633 
634 /* check access to 'struct bpf_context' fields */
635 static int check_ctx_access(struct verifier_env *env, int off, int size,
636 			    enum bpf_access_type t)
637 {
638 	if (env->prog->aux->ops->is_valid_access &&
639 	    env->prog->aux->ops->is_valid_access(off, size, t))
640 		return 0;
641 
642 	verbose("invalid bpf_context access off=%d size=%d\n", off, size);
643 	return -EACCES;
644 }
645 
646 /* check whether memory at (regno + off) is accessible for t = (read | write)
647  * if t==write, value_regno is a register which value is stored into memory
648  * if t==read, value_regno is a register which will receive the value from memory
649  * if t==write && value_regno==-1, some unknown value is stored into memory
650  * if t==read && value_regno==-1, don't care what we read from memory
651  */
652 static int check_mem_access(struct verifier_env *env, u32 regno, int off,
653 			    int bpf_size, enum bpf_access_type t,
654 			    int value_regno)
655 {
656 	struct verifier_state *state = &env->cur_state;
657 	int size, err = 0;
658 
659 	if (state->regs[regno].type == PTR_TO_STACK)
660 		off += state->regs[regno].imm;
661 
662 	size = bpf_size_to_bytes(bpf_size);
663 	if (size < 0)
664 		return size;
665 
666 	if (off % size != 0) {
667 		verbose("misaligned access off %d size %d\n", off, size);
668 		return -EACCES;
669 	}
670 
671 	if (state->regs[regno].type == PTR_TO_MAP_VALUE) {
672 		err = check_map_access(env, regno, off, size);
673 		if (!err && t == BPF_READ && value_regno >= 0)
674 			mark_reg_unknown_value(state->regs, value_regno);
675 
676 	} else if (state->regs[regno].type == PTR_TO_CTX) {
677 		err = check_ctx_access(env, off, size, t);
678 		if (!err && t == BPF_READ && value_regno >= 0)
679 			mark_reg_unknown_value(state->regs, value_regno);
680 
681 	} else if (state->regs[regno].type == FRAME_PTR ||
682 		   state->regs[regno].type == PTR_TO_STACK) {
683 		if (off >= 0 || off < -MAX_BPF_STACK) {
684 			verbose("invalid stack off=%d size=%d\n", off, size);
685 			return -EACCES;
686 		}
687 		if (t == BPF_WRITE)
688 			err = check_stack_write(state, off, size, value_regno);
689 		else
690 			err = check_stack_read(state, off, size, value_regno);
691 	} else {
692 		verbose("R%d invalid mem access '%s'\n",
693 			regno, reg_type_str[state->regs[regno].type]);
694 		return -EACCES;
695 	}
696 	return err;
697 }
698 
699 static int check_xadd(struct verifier_env *env, struct bpf_insn *insn)
700 {
701 	struct reg_state *regs = env->cur_state.regs;
702 	int err;
703 
704 	if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
705 	    insn->imm != 0) {
706 		verbose("BPF_XADD uses reserved fields\n");
707 		return -EINVAL;
708 	}
709 
710 	/* check src1 operand */
711 	err = check_reg_arg(regs, insn->src_reg, SRC_OP);
712 	if (err)
713 		return err;
714 
715 	/* check src2 operand */
716 	err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
717 	if (err)
718 		return err;
719 
720 	/* check whether atomic_add can read the memory */
721 	err = check_mem_access(env, insn->dst_reg, insn->off,
722 			       BPF_SIZE(insn->code), BPF_READ, -1);
723 	if (err)
724 		return err;
725 
726 	/* check whether atomic_add can write into the same memory */
727 	return check_mem_access(env, insn->dst_reg, insn->off,
728 				BPF_SIZE(insn->code), BPF_WRITE, -1);
729 }
730 
731 /* when register 'regno' is passed into function that will read 'access_size'
732  * bytes from that pointer, make sure that it's within stack boundary
733  * and all elements of stack are initialized
734  */
735 static int check_stack_boundary(struct verifier_env *env,
736 				int regno, int access_size)
737 {
738 	struct verifier_state *state = &env->cur_state;
739 	struct reg_state *regs = state->regs;
740 	int off, i;
741 
742 	if (regs[regno].type != PTR_TO_STACK)
743 		return -EACCES;
744 
745 	off = regs[regno].imm;
746 	if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
747 	    access_size <= 0) {
748 		verbose("invalid stack type R%d off=%d access_size=%d\n",
749 			regno, off, access_size);
750 		return -EACCES;
751 	}
752 
753 	for (i = 0; i < access_size; i++) {
754 		if (state->stack_slot_type[MAX_BPF_STACK + off + i] != STACK_MISC) {
755 			verbose("invalid indirect read from stack off %d+%d size %d\n",
756 				off, i, access_size);
757 			return -EACCES;
758 		}
759 	}
760 	return 0;
761 }
762 
763 static int check_func_arg(struct verifier_env *env, u32 regno,
764 			  enum bpf_arg_type arg_type, struct bpf_map **mapp)
765 {
766 	struct reg_state *reg = env->cur_state.regs + regno;
767 	enum bpf_reg_type expected_type;
768 	int err = 0;
769 
770 	if (arg_type == ARG_DONTCARE)
771 		return 0;
772 
773 	if (reg->type == NOT_INIT) {
774 		verbose("R%d !read_ok\n", regno);
775 		return -EACCES;
776 	}
777 
778 	if (arg_type == ARG_ANYTHING)
779 		return 0;
780 
781 	if (arg_type == ARG_PTR_TO_STACK || arg_type == ARG_PTR_TO_MAP_KEY ||
782 	    arg_type == ARG_PTR_TO_MAP_VALUE) {
783 		expected_type = PTR_TO_STACK;
784 	} else if (arg_type == ARG_CONST_STACK_SIZE) {
785 		expected_type = CONST_IMM;
786 	} else if (arg_type == ARG_CONST_MAP_PTR) {
787 		expected_type = CONST_PTR_TO_MAP;
788 	} else if (arg_type == ARG_PTR_TO_CTX) {
789 		expected_type = PTR_TO_CTX;
790 	} else {
791 		verbose("unsupported arg_type %d\n", arg_type);
792 		return -EFAULT;
793 	}
794 
795 	if (reg->type != expected_type) {
796 		verbose("R%d type=%s expected=%s\n", regno,
797 			reg_type_str[reg->type], reg_type_str[expected_type]);
798 		return -EACCES;
799 	}
800 
801 	if (arg_type == ARG_CONST_MAP_PTR) {
802 		/* bpf_map_xxx(map_ptr) call: remember that map_ptr */
803 		*mapp = reg->map_ptr;
804 
805 	} else if (arg_type == ARG_PTR_TO_MAP_KEY) {
806 		/* bpf_map_xxx(..., map_ptr, ..., key) call:
807 		 * check that [key, key + map->key_size) are within
808 		 * stack limits and initialized
809 		 */
810 		if (!*mapp) {
811 			/* in function declaration map_ptr must come before
812 			 * map_key, so that it's verified and known before
813 			 * we have to check map_key here. Otherwise it means
814 			 * that kernel subsystem misconfigured verifier
815 			 */
816 			verbose("invalid map_ptr to access map->key\n");
817 			return -EACCES;
818 		}
819 		err = check_stack_boundary(env, regno, (*mapp)->key_size);
820 
821 	} else if (arg_type == ARG_PTR_TO_MAP_VALUE) {
822 		/* bpf_map_xxx(..., map_ptr, ..., value) call:
823 		 * check [value, value + map->value_size) validity
824 		 */
825 		if (!*mapp) {
826 			/* kernel subsystem misconfigured verifier */
827 			verbose("invalid map_ptr to access map->value\n");
828 			return -EACCES;
829 		}
830 		err = check_stack_boundary(env, regno, (*mapp)->value_size);
831 
832 	} else if (arg_type == ARG_CONST_STACK_SIZE) {
833 		/* bpf_xxx(..., buf, len) call will access 'len' bytes
834 		 * from stack pointer 'buf'. Check it
835 		 * note: regno == len, regno - 1 == buf
836 		 */
837 		if (regno == 0) {
838 			/* kernel subsystem misconfigured verifier */
839 			verbose("ARG_CONST_STACK_SIZE cannot be first argument\n");
840 			return -EACCES;
841 		}
842 		err = check_stack_boundary(env, regno - 1, reg->imm);
843 	}
844 
845 	return err;
846 }
847 
848 static int check_map_func_compatibility(struct bpf_map *map, int func_id)
849 {
850 	bool bool_map, bool_func;
851 	int i;
852 
853 	if (!map)
854 		return 0;
855 
856 	for (i = 0; i < ARRAY_SIZE(func_limit); i++) {
857 		bool_map = (map->map_type == func_limit[i].map_type);
858 		bool_func = (func_id == func_limit[i].func_id);
859 		/* only when map & func pair match it can continue.
860 		 * don't allow any other map type to be passed into
861 		 * the special func;
862 		 */
863 		if (bool_map != bool_func)
864 			return -EINVAL;
865 	}
866 
867 	return 0;
868 }
869 
870 static int check_call(struct verifier_env *env, int func_id)
871 {
872 	struct verifier_state *state = &env->cur_state;
873 	const struct bpf_func_proto *fn = NULL;
874 	struct reg_state *regs = state->regs;
875 	struct bpf_map *map = NULL;
876 	struct reg_state *reg;
877 	int i, err;
878 
879 	/* find function prototype */
880 	if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
881 		verbose("invalid func %d\n", func_id);
882 		return -EINVAL;
883 	}
884 
885 	if (env->prog->aux->ops->get_func_proto)
886 		fn = env->prog->aux->ops->get_func_proto(func_id);
887 
888 	if (!fn) {
889 		verbose("unknown func %d\n", func_id);
890 		return -EINVAL;
891 	}
892 
893 	/* eBPF programs must be GPL compatible to use GPL-ed functions */
894 	if (!env->prog->gpl_compatible && fn->gpl_only) {
895 		verbose("cannot call GPL only function from proprietary program\n");
896 		return -EINVAL;
897 	}
898 
899 	/* check args */
900 	err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &map);
901 	if (err)
902 		return err;
903 	err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &map);
904 	if (err)
905 		return err;
906 	err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &map);
907 	if (err)
908 		return err;
909 	err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &map);
910 	if (err)
911 		return err;
912 	err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &map);
913 	if (err)
914 		return err;
915 
916 	/* reset caller saved regs */
917 	for (i = 0; i < CALLER_SAVED_REGS; i++) {
918 		reg = regs + caller_saved[i];
919 		reg->type = NOT_INIT;
920 		reg->imm = 0;
921 	}
922 
923 	/* update return register */
924 	if (fn->ret_type == RET_INTEGER) {
925 		regs[BPF_REG_0].type = UNKNOWN_VALUE;
926 	} else if (fn->ret_type == RET_VOID) {
927 		regs[BPF_REG_0].type = NOT_INIT;
928 	} else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL) {
929 		regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
930 		/* remember map_ptr, so that check_map_access()
931 		 * can check 'value_size' boundary of memory access
932 		 * to map element returned from bpf_map_lookup_elem()
933 		 */
934 		if (map == NULL) {
935 			verbose("kernel subsystem misconfigured verifier\n");
936 			return -EINVAL;
937 		}
938 		regs[BPF_REG_0].map_ptr = map;
939 	} else {
940 		verbose("unknown return type %d of func %d\n",
941 			fn->ret_type, func_id);
942 		return -EINVAL;
943 	}
944 
945 	err = check_map_func_compatibility(map, func_id);
946 	if (err)
947 		return err;
948 
949 	return 0;
950 }
951 
952 /* check validity of 32-bit and 64-bit arithmetic operations */
953 static int check_alu_op(struct reg_state *regs, struct bpf_insn *insn)
954 {
955 	u8 opcode = BPF_OP(insn->code);
956 	int err;
957 
958 	if (opcode == BPF_END || opcode == BPF_NEG) {
959 		if (opcode == BPF_NEG) {
960 			if (BPF_SRC(insn->code) != 0 ||
961 			    insn->src_reg != BPF_REG_0 ||
962 			    insn->off != 0 || insn->imm != 0) {
963 				verbose("BPF_NEG uses reserved fields\n");
964 				return -EINVAL;
965 			}
966 		} else {
967 			if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
968 			    (insn->imm != 16 && insn->imm != 32 && insn->imm != 64)) {
969 				verbose("BPF_END uses reserved fields\n");
970 				return -EINVAL;
971 			}
972 		}
973 
974 		/* check src operand */
975 		err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
976 		if (err)
977 			return err;
978 
979 		/* check dest operand */
980 		err = check_reg_arg(regs, insn->dst_reg, DST_OP);
981 		if (err)
982 			return err;
983 
984 	} else if (opcode == BPF_MOV) {
985 
986 		if (BPF_SRC(insn->code) == BPF_X) {
987 			if (insn->imm != 0 || insn->off != 0) {
988 				verbose("BPF_MOV uses reserved fields\n");
989 				return -EINVAL;
990 			}
991 
992 			/* check src operand */
993 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
994 			if (err)
995 				return err;
996 		} else {
997 			if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
998 				verbose("BPF_MOV uses reserved fields\n");
999 				return -EINVAL;
1000 			}
1001 		}
1002 
1003 		/* check dest operand */
1004 		err = check_reg_arg(regs, insn->dst_reg, DST_OP);
1005 		if (err)
1006 			return err;
1007 
1008 		if (BPF_SRC(insn->code) == BPF_X) {
1009 			if (BPF_CLASS(insn->code) == BPF_ALU64) {
1010 				/* case: R1 = R2
1011 				 * copy register state to dest reg
1012 				 */
1013 				regs[insn->dst_reg] = regs[insn->src_reg];
1014 			} else {
1015 				regs[insn->dst_reg].type = UNKNOWN_VALUE;
1016 				regs[insn->dst_reg].map_ptr = NULL;
1017 			}
1018 		} else {
1019 			/* case: R = imm
1020 			 * remember the value we stored into this reg
1021 			 */
1022 			regs[insn->dst_reg].type = CONST_IMM;
1023 			regs[insn->dst_reg].imm = insn->imm;
1024 		}
1025 
1026 	} else if (opcode > BPF_END) {
1027 		verbose("invalid BPF_ALU opcode %x\n", opcode);
1028 		return -EINVAL;
1029 
1030 	} else {	/* all other ALU ops: and, sub, xor, add, ... */
1031 
1032 		bool stack_relative = false;
1033 
1034 		if (BPF_SRC(insn->code) == BPF_X) {
1035 			if (insn->imm != 0 || insn->off != 0) {
1036 				verbose("BPF_ALU uses reserved fields\n");
1037 				return -EINVAL;
1038 			}
1039 			/* check src1 operand */
1040 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1041 			if (err)
1042 				return err;
1043 		} else {
1044 			if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
1045 				verbose("BPF_ALU uses reserved fields\n");
1046 				return -EINVAL;
1047 			}
1048 		}
1049 
1050 		/* check src2 operand */
1051 		err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1052 		if (err)
1053 			return err;
1054 
1055 		if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
1056 		    BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
1057 			verbose("div by zero\n");
1058 			return -EINVAL;
1059 		}
1060 
1061 		/* pattern match 'bpf_add Rx, imm' instruction */
1062 		if (opcode == BPF_ADD && BPF_CLASS(insn->code) == BPF_ALU64 &&
1063 		    regs[insn->dst_reg].type == FRAME_PTR &&
1064 		    BPF_SRC(insn->code) == BPF_K)
1065 			stack_relative = true;
1066 
1067 		/* check dest operand */
1068 		err = check_reg_arg(regs, insn->dst_reg, DST_OP);
1069 		if (err)
1070 			return err;
1071 
1072 		if (stack_relative) {
1073 			regs[insn->dst_reg].type = PTR_TO_STACK;
1074 			regs[insn->dst_reg].imm = insn->imm;
1075 		}
1076 	}
1077 
1078 	return 0;
1079 }
1080 
1081 static int check_cond_jmp_op(struct verifier_env *env,
1082 			     struct bpf_insn *insn, int *insn_idx)
1083 {
1084 	struct reg_state *regs = env->cur_state.regs;
1085 	struct verifier_state *other_branch;
1086 	u8 opcode = BPF_OP(insn->code);
1087 	int err;
1088 
1089 	if (opcode > BPF_EXIT) {
1090 		verbose("invalid BPF_JMP opcode %x\n", opcode);
1091 		return -EINVAL;
1092 	}
1093 
1094 	if (BPF_SRC(insn->code) == BPF_X) {
1095 		if (insn->imm != 0) {
1096 			verbose("BPF_JMP uses reserved fields\n");
1097 			return -EINVAL;
1098 		}
1099 
1100 		/* check src1 operand */
1101 		err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1102 		if (err)
1103 			return err;
1104 	} else {
1105 		if (insn->src_reg != BPF_REG_0) {
1106 			verbose("BPF_JMP uses reserved fields\n");
1107 			return -EINVAL;
1108 		}
1109 	}
1110 
1111 	/* check src2 operand */
1112 	err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1113 	if (err)
1114 		return err;
1115 
1116 	/* detect if R == 0 where R was initialized to zero earlier */
1117 	if (BPF_SRC(insn->code) == BPF_K &&
1118 	    (opcode == BPF_JEQ || opcode == BPF_JNE) &&
1119 	    regs[insn->dst_reg].type == CONST_IMM &&
1120 	    regs[insn->dst_reg].imm == insn->imm) {
1121 		if (opcode == BPF_JEQ) {
1122 			/* if (imm == imm) goto pc+off;
1123 			 * only follow the goto, ignore fall-through
1124 			 */
1125 			*insn_idx += insn->off;
1126 			return 0;
1127 		} else {
1128 			/* if (imm != imm) goto pc+off;
1129 			 * only follow fall-through branch, since
1130 			 * that's where the program will go
1131 			 */
1132 			return 0;
1133 		}
1134 	}
1135 
1136 	other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
1137 	if (!other_branch)
1138 		return -EFAULT;
1139 
1140 	/* detect if R == 0 where R is returned value from bpf_map_lookup_elem() */
1141 	if (BPF_SRC(insn->code) == BPF_K &&
1142 	    insn->imm == 0 && (opcode == BPF_JEQ ||
1143 			       opcode == BPF_JNE) &&
1144 	    regs[insn->dst_reg].type == PTR_TO_MAP_VALUE_OR_NULL) {
1145 		if (opcode == BPF_JEQ) {
1146 			/* next fallthrough insn can access memory via
1147 			 * this register
1148 			 */
1149 			regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
1150 			/* branch targer cannot access it, since reg == 0 */
1151 			other_branch->regs[insn->dst_reg].type = CONST_IMM;
1152 			other_branch->regs[insn->dst_reg].imm = 0;
1153 		} else {
1154 			other_branch->regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
1155 			regs[insn->dst_reg].type = CONST_IMM;
1156 			regs[insn->dst_reg].imm = 0;
1157 		}
1158 	} else if (BPF_SRC(insn->code) == BPF_K &&
1159 		   (opcode == BPF_JEQ || opcode == BPF_JNE)) {
1160 
1161 		if (opcode == BPF_JEQ) {
1162 			/* detect if (R == imm) goto
1163 			 * and in the target state recognize that R = imm
1164 			 */
1165 			other_branch->regs[insn->dst_reg].type = CONST_IMM;
1166 			other_branch->regs[insn->dst_reg].imm = insn->imm;
1167 		} else {
1168 			/* detect if (R != imm) goto
1169 			 * and in the fall-through state recognize that R = imm
1170 			 */
1171 			regs[insn->dst_reg].type = CONST_IMM;
1172 			regs[insn->dst_reg].imm = insn->imm;
1173 		}
1174 	}
1175 	if (log_level)
1176 		print_verifier_state(env);
1177 	return 0;
1178 }
1179 
1180 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
1181 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
1182 {
1183 	u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
1184 
1185 	return (struct bpf_map *) (unsigned long) imm64;
1186 }
1187 
1188 /* verify BPF_LD_IMM64 instruction */
1189 static int check_ld_imm(struct verifier_env *env, struct bpf_insn *insn)
1190 {
1191 	struct reg_state *regs = env->cur_state.regs;
1192 	int err;
1193 
1194 	if (BPF_SIZE(insn->code) != BPF_DW) {
1195 		verbose("invalid BPF_LD_IMM insn\n");
1196 		return -EINVAL;
1197 	}
1198 	if (insn->off != 0) {
1199 		verbose("BPF_LD_IMM64 uses reserved fields\n");
1200 		return -EINVAL;
1201 	}
1202 
1203 	err = check_reg_arg(regs, insn->dst_reg, DST_OP);
1204 	if (err)
1205 		return err;
1206 
1207 	if (insn->src_reg == 0)
1208 		/* generic move 64-bit immediate into a register */
1209 		return 0;
1210 
1211 	/* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
1212 	BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
1213 
1214 	regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
1215 	regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
1216 	return 0;
1217 }
1218 
1219 static bool may_access_skb(enum bpf_prog_type type)
1220 {
1221 	switch (type) {
1222 	case BPF_PROG_TYPE_SOCKET_FILTER:
1223 	case BPF_PROG_TYPE_SCHED_CLS:
1224 	case BPF_PROG_TYPE_SCHED_ACT:
1225 		return true;
1226 	default:
1227 		return false;
1228 	}
1229 }
1230 
1231 /* verify safety of LD_ABS|LD_IND instructions:
1232  * - they can only appear in the programs where ctx == skb
1233  * - since they are wrappers of function calls, they scratch R1-R5 registers,
1234  *   preserve R6-R9, and store return value into R0
1235  *
1236  * Implicit input:
1237  *   ctx == skb == R6 == CTX
1238  *
1239  * Explicit input:
1240  *   SRC == any register
1241  *   IMM == 32-bit immediate
1242  *
1243  * Output:
1244  *   R0 - 8/16/32-bit skb data converted to cpu endianness
1245  */
1246 static int check_ld_abs(struct verifier_env *env, struct bpf_insn *insn)
1247 {
1248 	struct reg_state *regs = env->cur_state.regs;
1249 	u8 mode = BPF_MODE(insn->code);
1250 	struct reg_state *reg;
1251 	int i, err;
1252 
1253 	if (!may_access_skb(env->prog->type)) {
1254 		verbose("BPF_LD_ABS|IND instructions not allowed for this program type\n");
1255 		return -EINVAL;
1256 	}
1257 
1258 	if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
1259 	    (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
1260 		verbose("BPF_LD_ABS uses reserved fields\n");
1261 		return -EINVAL;
1262 	}
1263 
1264 	/* check whether implicit source operand (register R6) is readable */
1265 	err = check_reg_arg(regs, BPF_REG_6, SRC_OP);
1266 	if (err)
1267 		return err;
1268 
1269 	if (regs[BPF_REG_6].type != PTR_TO_CTX) {
1270 		verbose("at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
1271 		return -EINVAL;
1272 	}
1273 
1274 	if (mode == BPF_IND) {
1275 		/* check explicit source operand */
1276 		err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1277 		if (err)
1278 			return err;
1279 	}
1280 
1281 	/* reset caller saved regs to unreadable */
1282 	for (i = 0; i < CALLER_SAVED_REGS; i++) {
1283 		reg = regs + caller_saved[i];
1284 		reg->type = NOT_INIT;
1285 		reg->imm = 0;
1286 	}
1287 
1288 	/* mark destination R0 register as readable, since it contains
1289 	 * the value fetched from the packet
1290 	 */
1291 	regs[BPF_REG_0].type = UNKNOWN_VALUE;
1292 	return 0;
1293 }
1294 
1295 /* non-recursive DFS pseudo code
1296  * 1  procedure DFS-iterative(G,v):
1297  * 2      label v as discovered
1298  * 3      let S be a stack
1299  * 4      S.push(v)
1300  * 5      while S is not empty
1301  * 6            t <- S.pop()
1302  * 7            if t is what we're looking for:
1303  * 8                return t
1304  * 9            for all edges e in G.adjacentEdges(t) do
1305  * 10               if edge e is already labelled
1306  * 11                   continue with the next edge
1307  * 12               w <- G.adjacentVertex(t,e)
1308  * 13               if vertex w is not discovered and not explored
1309  * 14                   label e as tree-edge
1310  * 15                   label w as discovered
1311  * 16                   S.push(w)
1312  * 17                   continue at 5
1313  * 18               else if vertex w is discovered
1314  * 19                   label e as back-edge
1315  * 20               else
1316  * 21                   // vertex w is explored
1317  * 22                   label e as forward- or cross-edge
1318  * 23           label t as explored
1319  * 24           S.pop()
1320  *
1321  * convention:
1322  * 0x10 - discovered
1323  * 0x11 - discovered and fall-through edge labelled
1324  * 0x12 - discovered and fall-through and branch edges labelled
1325  * 0x20 - explored
1326  */
1327 
1328 enum {
1329 	DISCOVERED = 0x10,
1330 	EXPLORED = 0x20,
1331 	FALLTHROUGH = 1,
1332 	BRANCH = 2,
1333 };
1334 
1335 #define STATE_LIST_MARK ((struct verifier_state_list *) -1L)
1336 
1337 static int *insn_stack;	/* stack of insns to process */
1338 static int cur_stack;	/* current stack index */
1339 static int *insn_state;
1340 
1341 /* t, w, e - match pseudo-code above:
1342  * t - index of current instruction
1343  * w - next instruction
1344  * e - edge
1345  */
1346 static int push_insn(int t, int w, int e, struct verifier_env *env)
1347 {
1348 	if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
1349 		return 0;
1350 
1351 	if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
1352 		return 0;
1353 
1354 	if (w < 0 || w >= env->prog->len) {
1355 		verbose("jump out of range from insn %d to %d\n", t, w);
1356 		return -EINVAL;
1357 	}
1358 
1359 	if (e == BRANCH)
1360 		/* mark branch target for state pruning */
1361 		env->explored_states[w] = STATE_LIST_MARK;
1362 
1363 	if (insn_state[w] == 0) {
1364 		/* tree-edge */
1365 		insn_state[t] = DISCOVERED | e;
1366 		insn_state[w] = DISCOVERED;
1367 		if (cur_stack >= env->prog->len)
1368 			return -E2BIG;
1369 		insn_stack[cur_stack++] = w;
1370 		return 1;
1371 	} else if ((insn_state[w] & 0xF0) == DISCOVERED) {
1372 		verbose("back-edge from insn %d to %d\n", t, w);
1373 		return -EINVAL;
1374 	} else if (insn_state[w] == EXPLORED) {
1375 		/* forward- or cross-edge */
1376 		insn_state[t] = DISCOVERED | e;
1377 	} else {
1378 		verbose("insn state internal bug\n");
1379 		return -EFAULT;
1380 	}
1381 	return 0;
1382 }
1383 
1384 /* non-recursive depth-first-search to detect loops in BPF program
1385  * loop == back-edge in directed graph
1386  */
1387 static int check_cfg(struct verifier_env *env)
1388 {
1389 	struct bpf_insn *insns = env->prog->insnsi;
1390 	int insn_cnt = env->prog->len;
1391 	int ret = 0;
1392 	int i, t;
1393 
1394 	insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
1395 	if (!insn_state)
1396 		return -ENOMEM;
1397 
1398 	insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
1399 	if (!insn_stack) {
1400 		kfree(insn_state);
1401 		return -ENOMEM;
1402 	}
1403 
1404 	insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
1405 	insn_stack[0] = 0; /* 0 is the first instruction */
1406 	cur_stack = 1;
1407 
1408 peek_stack:
1409 	if (cur_stack == 0)
1410 		goto check_state;
1411 	t = insn_stack[cur_stack - 1];
1412 
1413 	if (BPF_CLASS(insns[t].code) == BPF_JMP) {
1414 		u8 opcode = BPF_OP(insns[t].code);
1415 
1416 		if (opcode == BPF_EXIT) {
1417 			goto mark_explored;
1418 		} else if (opcode == BPF_CALL) {
1419 			ret = push_insn(t, t + 1, FALLTHROUGH, env);
1420 			if (ret == 1)
1421 				goto peek_stack;
1422 			else if (ret < 0)
1423 				goto err_free;
1424 		} else if (opcode == BPF_JA) {
1425 			if (BPF_SRC(insns[t].code) != BPF_K) {
1426 				ret = -EINVAL;
1427 				goto err_free;
1428 			}
1429 			/* unconditional jump with single edge */
1430 			ret = push_insn(t, t + insns[t].off + 1,
1431 					FALLTHROUGH, env);
1432 			if (ret == 1)
1433 				goto peek_stack;
1434 			else if (ret < 0)
1435 				goto err_free;
1436 			/* tell verifier to check for equivalent states
1437 			 * after every call and jump
1438 			 */
1439 			if (t + 1 < insn_cnt)
1440 				env->explored_states[t + 1] = STATE_LIST_MARK;
1441 		} else {
1442 			/* conditional jump with two edges */
1443 			ret = push_insn(t, t + 1, FALLTHROUGH, env);
1444 			if (ret == 1)
1445 				goto peek_stack;
1446 			else if (ret < 0)
1447 				goto err_free;
1448 
1449 			ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
1450 			if (ret == 1)
1451 				goto peek_stack;
1452 			else if (ret < 0)
1453 				goto err_free;
1454 		}
1455 	} else {
1456 		/* all other non-branch instructions with single
1457 		 * fall-through edge
1458 		 */
1459 		ret = push_insn(t, t + 1, FALLTHROUGH, env);
1460 		if (ret == 1)
1461 			goto peek_stack;
1462 		else if (ret < 0)
1463 			goto err_free;
1464 	}
1465 
1466 mark_explored:
1467 	insn_state[t] = EXPLORED;
1468 	if (cur_stack-- <= 0) {
1469 		verbose("pop stack internal bug\n");
1470 		ret = -EFAULT;
1471 		goto err_free;
1472 	}
1473 	goto peek_stack;
1474 
1475 check_state:
1476 	for (i = 0; i < insn_cnt; i++) {
1477 		if (insn_state[i] != EXPLORED) {
1478 			verbose("unreachable insn %d\n", i);
1479 			ret = -EINVAL;
1480 			goto err_free;
1481 		}
1482 	}
1483 	ret = 0; /* cfg looks good */
1484 
1485 err_free:
1486 	kfree(insn_state);
1487 	kfree(insn_stack);
1488 	return ret;
1489 }
1490 
1491 /* compare two verifier states
1492  *
1493  * all states stored in state_list are known to be valid, since
1494  * verifier reached 'bpf_exit' instruction through them
1495  *
1496  * this function is called when verifier exploring different branches of
1497  * execution popped from the state stack. If it sees an old state that has
1498  * more strict register state and more strict stack state then this execution
1499  * branch doesn't need to be explored further, since verifier already
1500  * concluded that more strict state leads to valid finish.
1501  *
1502  * Therefore two states are equivalent if register state is more conservative
1503  * and explored stack state is more conservative than the current one.
1504  * Example:
1505  *       explored                   current
1506  * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
1507  * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
1508  *
1509  * In other words if current stack state (one being explored) has more
1510  * valid slots than old one that already passed validation, it means
1511  * the verifier can stop exploring and conclude that current state is valid too
1512  *
1513  * Similarly with registers. If explored state has register type as invalid
1514  * whereas register type in current state is meaningful, it means that
1515  * the current state will reach 'bpf_exit' instruction safely
1516  */
1517 static bool states_equal(struct verifier_state *old, struct verifier_state *cur)
1518 {
1519 	int i;
1520 
1521 	for (i = 0; i < MAX_BPF_REG; i++) {
1522 		if (memcmp(&old->regs[i], &cur->regs[i],
1523 			   sizeof(old->regs[0])) != 0) {
1524 			if (old->regs[i].type == NOT_INIT ||
1525 			    (old->regs[i].type == UNKNOWN_VALUE &&
1526 			     cur->regs[i].type != NOT_INIT))
1527 				continue;
1528 			return false;
1529 		}
1530 	}
1531 
1532 	for (i = 0; i < MAX_BPF_STACK; i++) {
1533 		if (old->stack_slot_type[i] == STACK_INVALID)
1534 			continue;
1535 		if (old->stack_slot_type[i] != cur->stack_slot_type[i])
1536 			/* Ex: old explored (safe) state has STACK_SPILL in
1537 			 * this stack slot, but current has has STACK_MISC ->
1538 			 * this verifier states are not equivalent,
1539 			 * return false to continue verification of this path
1540 			 */
1541 			return false;
1542 		if (i % BPF_REG_SIZE)
1543 			continue;
1544 		if (memcmp(&old->spilled_regs[i / BPF_REG_SIZE],
1545 			   &cur->spilled_regs[i / BPF_REG_SIZE],
1546 			   sizeof(old->spilled_regs[0])))
1547 			/* when explored and current stack slot types are
1548 			 * the same, check that stored pointers types
1549 			 * are the same as well.
1550 			 * Ex: explored safe path could have stored
1551 			 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -8}
1552 			 * but current path has stored:
1553 			 * (struct reg_state) {.type = PTR_TO_STACK, .imm = -16}
1554 			 * such verifier states are not equivalent.
1555 			 * return false to continue verification of this path
1556 			 */
1557 			return false;
1558 		else
1559 			continue;
1560 	}
1561 	return true;
1562 }
1563 
1564 static int is_state_visited(struct verifier_env *env, int insn_idx)
1565 {
1566 	struct verifier_state_list *new_sl;
1567 	struct verifier_state_list *sl;
1568 
1569 	sl = env->explored_states[insn_idx];
1570 	if (!sl)
1571 		/* this 'insn_idx' instruction wasn't marked, so we will not
1572 		 * be doing state search here
1573 		 */
1574 		return 0;
1575 
1576 	while (sl != STATE_LIST_MARK) {
1577 		if (states_equal(&sl->state, &env->cur_state))
1578 			/* reached equivalent register/stack state,
1579 			 * prune the search
1580 			 */
1581 			return 1;
1582 		sl = sl->next;
1583 	}
1584 
1585 	/* there were no equivalent states, remember current one.
1586 	 * technically the current state is not proven to be safe yet,
1587 	 * but it will either reach bpf_exit (which means it's safe) or
1588 	 * it will be rejected. Since there are no loops, we won't be
1589 	 * seeing this 'insn_idx' instruction again on the way to bpf_exit
1590 	 */
1591 	new_sl = kmalloc(sizeof(struct verifier_state_list), GFP_USER);
1592 	if (!new_sl)
1593 		return -ENOMEM;
1594 
1595 	/* add new state to the head of linked list */
1596 	memcpy(&new_sl->state, &env->cur_state, sizeof(env->cur_state));
1597 	new_sl->next = env->explored_states[insn_idx];
1598 	env->explored_states[insn_idx] = new_sl;
1599 	return 0;
1600 }
1601 
1602 static int do_check(struct verifier_env *env)
1603 {
1604 	struct verifier_state *state = &env->cur_state;
1605 	struct bpf_insn *insns = env->prog->insnsi;
1606 	struct reg_state *regs = state->regs;
1607 	int insn_cnt = env->prog->len;
1608 	int insn_idx, prev_insn_idx = 0;
1609 	int insn_processed = 0;
1610 	bool do_print_state = false;
1611 
1612 	init_reg_state(regs);
1613 	insn_idx = 0;
1614 	for (;;) {
1615 		struct bpf_insn *insn;
1616 		u8 class;
1617 		int err;
1618 
1619 		if (insn_idx >= insn_cnt) {
1620 			verbose("invalid insn idx %d insn_cnt %d\n",
1621 				insn_idx, insn_cnt);
1622 			return -EFAULT;
1623 		}
1624 
1625 		insn = &insns[insn_idx];
1626 		class = BPF_CLASS(insn->code);
1627 
1628 		if (++insn_processed > 32768) {
1629 			verbose("BPF program is too large. Proccessed %d insn\n",
1630 				insn_processed);
1631 			return -E2BIG;
1632 		}
1633 
1634 		err = is_state_visited(env, insn_idx);
1635 		if (err < 0)
1636 			return err;
1637 		if (err == 1) {
1638 			/* found equivalent state, can prune the search */
1639 			if (log_level) {
1640 				if (do_print_state)
1641 					verbose("\nfrom %d to %d: safe\n",
1642 						prev_insn_idx, insn_idx);
1643 				else
1644 					verbose("%d: safe\n", insn_idx);
1645 			}
1646 			goto process_bpf_exit;
1647 		}
1648 
1649 		if (log_level && do_print_state) {
1650 			verbose("\nfrom %d to %d:", prev_insn_idx, insn_idx);
1651 			print_verifier_state(env);
1652 			do_print_state = false;
1653 		}
1654 
1655 		if (log_level) {
1656 			verbose("%d: ", insn_idx);
1657 			print_bpf_insn(insn);
1658 		}
1659 
1660 		if (class == BPF_ALU || class == BPF_ALU64) {
1661 			err = check_alu_op(regs, insn);
1662 			if (err)
1663 				return err;
1664 
1665 		} else if (class == BPF_LDX) {
1666 			enum bpf_reg_type src_reg_type;
1667 
1668 			/* check for reserved fields is already done */
1669 
1670 			/* check src operand */
1671 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1672 			if (err)
1673 				return err;
1674 
1675 			err = check_reg_arg(regs, insn->dst_reg, DST_OP_NO_MARK);
1676 			if (err)
1677 				return err;
1678 
1679 			src_reg_type = regs[insn->src_reg].type;
1680 
1681 			/* check that memory (src_reg + off) is readable,
1682 			 * the state of dst_reg will be updated by this func
1683 			 */
1684 			err = check_mem_access(env, insn->src_reg, insn->off,
1685 					       BPF_SIZE(insn->code), BPF_READ,
1686 					       insn->dst_reg);
1687 			if (err)
1688 				return err;
1689 
1690 			if (BPF_SIZE(insn->code) != BPF_W) {
1691 				insn_idx++;
1692 				continue;
1693 			}
1694 
1695 			if (insn->imm == 0) {
1696 				/* saw a valid insn
1697 				 * dst_reg = *(u32 *)(src_reg + off)
1698 				 * use reserved 'imm' field to mark this insn
1699 				 */
1700 				insn->imm = src_reg_type;
1701 
1702 			} else if (src_reg_type != insn->imm &&
1703 				   (src_reg_type == PTR_TO_CTX ||
1704 				    insn->imm == PTR_TO_CTX)) {
1705 				/* ABuser program is trying to use the same insn
1706 				 * dst_reg = *(u32*) (src_reg + off)
1707 				 * with different pointer types:
1708 				 * src_reg == ctx in one branch and
1709 				 * src_reg == stack|map in some other branch.
1710 				 * Reject it.
1711 				 */
1712 				verbose("same insn cannot be used with different pointers\n");
1713 				return -EINVAL;
1714 			}
1715 
1716 		} else if (class == BPF_STX) {
1717 			enum bpf_reg_type dst_reg_type;
1718 
1719 			if (BPF_MODE(insn->code) == BPF_XADD) {
1720 				err = check_xadd(env, insn);
1721 				if (err)
1722 					return err;
1723 				insn_idx++;
1724 				continue;
1725 			}
1726 
1727 			/* check src1 operand */
1728 			err = check_reg_arg(regs, insn->src_reg, SRC_OP);
1729 			if (err)
1730 				return err;
1731 			/* check src2 operand */
1732 			err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1733 			if (err)
1734 				return err;
1735 
1736 			dst_reg_type = regs[insn->dst_reg].type;
1737 
1738 			/* check that memory (dst_reg + off) is writeable */
1739 			err = check_mem_access(env, insn->dst_reg, insn->off,
1740 					       BPF_SIZE(insn->code), BPF_WRITE,
1741 					       insn->src_reg);
1742 			if (err)
1743 				return err;
1744 
1745 			if (insn->imm == 0) {
1746 				insn->imm = dst_reg_type;
1747 			} else if (dst_reg_type != insn->imm &&
1748 				   (dst_reg_type == PTR_TO_CTX ||
1749 				    insn->imm == PTR_TO_CTX)) {
1750 				verbose("same insn cannot be used with different pointers\n");
1751 				return -EINVAL;
1752 			}
1753 
1754 		} else if (class == BPF_ST) {
1755 			if (BPF_MODE(insn->code) != BPF_MEM ||
1756 			    insn->src_reg != BPF_REG_0) {
1757 				verbose("BPF_ST uses reserved fields\n");
1758 				return -EINVAL;
1759 			}
1760 			/* check src operand */
1761 			err = check_reg_arg(regs, insn->dst_reg, SRC_OP);
1762 			if (err)
1763 				return err;
1764 
1765 			/* check that memory (dst_reg + off) is writeable */
1766 			err = check_mem_access(env, insn->dst_reg, insn->off,
1767 					       BPF_SIZE(insn->code), BPF_WRITE,
1768 					       -1);
1769 			if (err)
1770 				return err;
1771 
1772 		} else if (class == BPF_JMP) {
1773 			u8 opcode = BPF_OP(insn->code);
1774 
1775 			if (opcode == BPF_CALL) {
1776 				if (BPF_SRC(insn->code) != BPF_K ||
1777 				    insn->off != 0 ||
1778 				    insn->src_reg != BPF_REG_0 ||
1779 				    insn->dst_reg != BPF_REG_0) {
1780 					verbose("BPF_CALL uses reserved fields\n");
1781 					return -EINVAL;
1782 				}
1783 
1784 				err = check_call(env, insn->imm);
1785 				if (err)
1786 					return err;
1787 
1788 			} else if (opcode == BPF_JA) {
1789 				if (BPF_SRC(insn->code) != BPF_K ||
1790 				    insn->imm != 0 ||
1791 				    insn->src_reg != BPF_REG_0 ||
1792 				    insn->dst_reg != BPF_REG_0) {
1793 					verbose("BPF_JA uses reserved fields\n");
1794 					return -EINVAL;
1795 				}
1796 
1797 				insn_idx += insn->off + 1;
1798 				continue;
1799 
1800 			} else if (opcode == BPF_EXIT) {
1801 				if (BPF_SRC(insn->code) != BPF_K ||
1802 				    insn->imm != 0 ||
1803 				    insn->src_reg != BPF_REG_0 ||
1804 				    insn->dst_reg != BPF_REG_0) {
1805 					verbose("BPF_EXIT uses reserved fields\n");
1806 					return -EINVAL;
1807 				}
1808 
1809 				/* eBPF calling convetion is such that R0 is used
1810 				 * to return the value from eBPF program.
1811 				 * Make sure that it's readable at this time
1812 				 * of bpf_exit, which means that program wrote
1813 				 * something into it earlier
1814 				 */
1815 				err = check_reg_arg(regs, BPF_REG_0, SRC_OP);
1816 				if (err)
1817 					return err;
1818 
1819 process_bpf_exit:
1820 				insn_idx = pop_stack(env, &prev_insn_idx);
1821 				if (insn_idx < 0) {
1822 					break;
1823 				} else {
1824 					do_print_state = true;
1825 					continue;
1826 				}
1827 			} else {
1828 				err = check_cond_jmp_op(env, insn, &insn_idx);
1829 				if (err)
1830 					return err;
1831 			}
1832 		} else if (class == BPF_LD) {
1833 			u8 mode = BPF_MODE(insn->code);
1834 
1835 			if (mode == BPF_ABS || mode == BPF_IND) {
1836 				err = check_ld_abs(env, insn);
1837 				if (err)
1838 					return err;
1839 
1840 			} else if (mode == BPF_IMM) {
1841 				err = check_ld_imm(env, insn);
1842 				if (err)
1843 					return err;
1844 
1845 				insn_idx++;
1846 			} else {
1847 				verbose("invalid BPF_LD mode\n");
1848 				return -EINVAL;
1849 			}
1850 		} else {
1851 			verbose("unknown insn class %d\n", class);
1852 			return -EINVAL;
1853 		}
1854 
1855 		insn_idx++;
1856 	}
1857 
1858 	return 0;
1859 }
1860 
1861 /* look for pseudo eBPF instructions that access map FDs and
1862  * replace them with actual map pointers
1863  */
1864 static int replace_map_fd_with_map_ptr(struct verifier_env *env)
1865 {
1866 	struct bpf_insn *insn = env->prog->insnsi;
1867 	int insn_cnt = env->prog->len;
1868 	int i, j;
1869 
1870 	for (i = 0; i < insn_cnt; i++, insn++) {
1871 		if (BPF_CLASS(insn->code) == BPF_LDX &&
1872 		    (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
1873 			verbose("BPF_LDX uses reserved fields\n");
1874 			return -EINVAL;
1875 		}
1876 
1877 		if (BPF_CLASS(insn->code) == BPF_STX &&
1878 		    ((BPF_MODE(insn->code) != BPF_MEM &&
1879 		      BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
1880 			verbose("BPF_STX uses reserved fields\n");
1881 			return -EINVAL;
1882 		}
1883 
1884 		if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
1885 			struct bpf_map *map;
1886 			struct fd f;
1887 
1888 			if (i == insn_cnt - 1 || insn[1].code != 0 ||
1889 			    insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
1890 			    insn[1].off != 0) {
1891 				verbose("invalid bpf_ld_imm64 insn\n");
1892 				return -EINVAL;
1893 			}
1894 
1895 			if (insn->src_reg == 0)
1896 				/* valid generic load 64-bit imm */
1897 				goto next_insn;
1898 
1899 			if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
1900 				verbose("unrecognized bpf_ld_imm64 insn\n");
1901 				return -EINVAL;
1902 			}
1903 
1904 			f = fdget(insn->imm);
1905 
1906 			map = bpf_map_get(f);
1907 			if (IS_ERR(map)) {
1908 				verbose("fd %d is not pointing to valid bpf_map\n",
1909 					insn->imm);
1910 				fdput(f);
1911 				return PTR_ERR(map);
1912 			}
1913 
1914 			/* store map pointer inside BPF_LD_IMM64 instruction */
1915 			insn[0].imm = (u32) (unsigned long) map;
1916 			insn[1].imm = ((u64) (unsigned long) map) >> 32;
1917 
1918 			/* check whether we recorded this map already */
1919 			for (j = 0; j < env->used_map_cnt; j++)
1920 				if (env->used_maps[j] == map) {
1921 					fdput(f);
1922 					goto next_insn;
1923 				}
1924 
1925 			if (env->used_map_cnt >= MAX_USED_MAPS) {
1926 				fdput(f);
1927 				return -E2BIG;
1928 			}
1929 
1930 			/* remember this map */
1931 			env->used_maps[env->used_map_cnt++] = map;
1932 
1933 			/* hold the map. If the program is rejected by verifier,
1934 			 * the map will be released by release_maps() or it
1935 			 * will be used by the valid program until it's unloaded
1936 			 * and all maps are released in free_bpf_prog_info()
1937 			 */
1938 			atomic_inc(&map->refcnt);
1939 
1940 			fdput(f);
1941 next_insn:
1942 			insn++;
1943 			i++;
1944 		}
1945 	}
1946 
1947 	/* now all pseudo BPF_LD_IMM64 instructions load valid
1948 	 * 'struct bpf_map *' into a register instead of user map_fd.
1949 	 * These pointers will be used later by verifier to validate map access.
1950 	 */
1951 	return 0;
1952 }
1953 
1954 /* drop refcnt of maps used by the rejected program */
1955 static void release_maps(struct verifier_env *env)
1956 {
1957 	int i;
1958 
1959 	for (i = 0; i < env->used_map_cnt; i++)
1960 		bpf_map_put(env->used_maps[i]);
1961 }
1962 
1963 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
1964 static void convert_pseudo_ld_imm64(struct verifier_env *env)
1965 {
1966 	struct bpf_insn *insn = env->prog->insnsi;
1967 	int insn_cnt = env->prog->len;
1968 	int i;
1969 
1970 	for (i = 0; i < insn_cnt; i++, insn++)
1971 		if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
1972 			insn->src_reg = 0;
1973 }
1974 
1975 static void adjust_branches(struct bpf_prog *prog, int pos, int delta)
1976 {
1977 	struct bpf_insn *insn = prog->insnsi;
1978 	int insn_cnt = prog->len;
1979 	int i;
1980 
1981 	for (i = 0; i < insn_cnt; i++, insn++) {
1982 		if (BPF_CLASS(insn->code) != BPF_JMP ||
1983 		    BPF_OP(insn->code) == BPF_CALL ||
1984 		    BPF_OP(insn->code) == BPF_EXIT)
1985 			continue;
1986 
1987 		/* adjust offset of jmps if necessary */
1988 		if (i < pos && i + insn->off + 1 > pos)
1989 			insn->off += delta;
1990 		else if (i > pos && i + insn->off + 1 < pos)
1991 			insn->off -= delta;
1992 	}
1993 }
1994 
1995 /* convert load instructions that access fields of 'struct __sk_buff'
1996  * into sequence of instructions that access fields of 'struct sk_buff'
1997  */
1998 static int convert_ctx_accesses(struct verifier_env *env)
1999 {
2000 	struct bpf_insn *insn = env->prog->insnsi;
2001 	int insn_cnt = env->prog->len;
2002 	struct bpf_insn insn_buf[16];
2003 	struct bpf_prog *new_prog;
2004 	u32 cnt;
2005 	int i;
2006 	enum bpf_access_type type;
2007 
2008 	if (!env->prog->aux->ops->convert_ctx_access)
2009 		return 0;
2010 
2011 	for (i = 0; i < insn_cnt; i++, insn++) {
2012 		if (insn->code == (BPF_LDX | BPF_MEM | BPF_W))
2013 			type = BPF_READ;
2014 		else if (insn->code == (BPF_STX | BPF_MEM | BPF_W))
2015 			type = BPF_WRITE;
2016 		else
2017 			continue;
2018 
2019 		if (insn->imm != PTR_TO_CTX) {
2020 			/* clear internal mark */
2021 			insn->imm = 0;
2022 			continue;
2023 		}
2024 
2025 		cnt = env->prog->aux->ops->
2026 			convert_ctx_access(type, insn->dst_reg, insn->src_reg,
2027 					   insn->off, insn_buf);
2028 		if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
2029 			verbose("bpf verifier is misconfigured\n");
2030 			return -EINVAL;
2031 		}
2032 
2033 		if (cnt == 1) {
2034 			memcpy(insn, insn_buf, sizeof(*insn));
2035 			continue;
2036 		}
2037 
2038 		/* several new insns need to be inserted. Make room for them */
2039 		insn_cnt += cnt - 1;
2040 		new_prog = bpf_prog_realloc(env->prog,
2041 					    bpf_prog_size(insn_cnt),
2042 					    GFP_USER);
2043 		if (!new_prog)
2044 			return -ENOMEM;
2045 
2046 		new_prog->len = insn_cnt;
2047 
2048 		memmove(new_prog->insnsi + i + cnt, new_prog->insns + i + 1,
2049 			sizeof(*insn) * (insn_cnt - i - cnt));
2050 
2051 		/* copy substitute insns in place of load instruction */
2052 		memcpy(new_prog->insnsi + i, insn_buf, sizeof(*insn) * cnt);
2053 
2054 		/* adjust branches in the whole program */
2055 		adjust_branches(new_prog, i, cnt - 1);
2056 
2057 		/* keep walking new program and skip insns we just inserted */
2058 		env->prog = new_prog;
2059 		insn = new_prog->insnsi + i + cnt - 1;
2060 		i += cnt - 1;
2061 	}
2062 
2063 	return 0;
2064 }
2065 
2066 static void free_states(struct verifier_env *env)
2067 {
2068 	struct verifier_state_list *sl, *sln;
2069 	int i;
2070 
2071 	if (!env->explored_states)
2072 		return;
2073 
2074 	for (i = 0; i < env->prog->len; i++) {
2075 		sl = env->explored_states[i];
2076 
2077 		if (sl)
2078 			while (sl != STATE_LIST_MARK) {
2079 				sln = sl->next;
2080 				kfree(sl);
2081 				sl = sln;
2082 			}
2083 	}
2084 
2085 	kfree(env->explored_states);
2086 }
2087 
2088 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
2089 {
2090 	char __user *log_ubuf = NULL;
2091 	struct verifier_env *env;
2092 	int ret = -EINVAL;
2093 
2094 	if ((*prog)->len <= 0 || (*prog)->len > BPF_MAXINSNS)
2095 		return -E2BIG;
2096 
2097 	/* 'struct verifier_env' can be global, but since it's not small,
2098 	 * allocate/free it every time bpf_check() is called
2099 	 */
2100 	env = kzalloc(sizeof(struct verifier_env), GFP_KERNEL);
2101 	if (!env)
2102 		return -ENOMEM;
2103 
2104 	env->prog = *prog;
2105 
2106 	/* grab the mutex to protect few globals used by verifier */
2107 	mutex_lock(&bpf_verifier_lock);
2108 
2109 	if (attr->log_level || attr->log_buf || attr->log_size) {
2110 		/* user requested verbose verifier output
2111 		 * and supplied buffer to store the verification trace
2112 		 */
2113 		log_level = attr->log_level;
2114 		log_ubuf = (char __user *) (unsigned long) attr->log_buf;
2115 		log_size = attr->log_size;
2116 		log_len = 0;
2117 
2118 		ret = -EINVAL;
2119 		/* log_* values have to be sane */
2120 		if (log_size < 128 || log_size > UINT_MAX >> 8 ||
2121 		    log_level == 0 || log_ubuf == NULL)
2122 			goto free_env;
2123 
2124 		ret = -ENOMEM;
2125 		log_buf = vmalloc(log_size);
2126 		if (!log_buf)
2127 			goto free_env;
2128 	} else {
2129 		log_level = 0;
2130 	}
2131 
2132 	ret = replace_map_fd_with_map_ptr(env);
2133 	if (ret < 0)
2134 		goto skip_full_check;
2135 
2136 	env->explored_states = kcalloc(env->prog->len,
2137 				       sizeof(struct verifier_state_list *),
2138 				       GFP_USER);
2139 	ret = -ENOMEM;
2140 	if (!env->explored_states)
2141 		goto skip_full_check;
2142 
2143 	ret = check_cfg(env);
2144 	if (ret < 0)
2145 		goto skip_full_check;
2146 
2147 	ret = do_check(env);
2148 
2149 skip_full_check:
2150 	while (pop_stack(env, NULL) >= 0);
2151 	free_states(env);
2152 
2153 	if (ret == 0)
2154 		/* program is valid, convert *(u32*)(ctx + off) accesses */
2155 		ret = convert_ctx_accesses(env);
2156 
2157 	if (log_level && log_len >= log_size - 1) {
2158 		BUG_ON(log_len >= log_size);
2159 		/* verifier log exceeded user supplied buffer */
2160 		ret = -ENOSPC;
2161 		/* fall through to return what was recorded */
2162 	}
2163 
2164 	/* copy verifier log back to user space including trailing zero */
2165 	if (log_level && copy_to_user(log_ubuf, log_buf, log_len + 1) != 0) {
2166 		ret = -EFAULT;
2167 		goto free_log_buf;
2168 	}
2169 
2170 	if (ret == 0 && env->used_map_cnt) {
2171 		/* if program passed verifier, update used_maps in bpf_prog_info */
2172 		env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
2173 							  sizeof(env->used_maps[0]),
2174 							  GFP_KERNEL);
2175 
2176 		if (!env->prog->aux->used_maps) {
2177 			ret = -ENOMEM;
2178 			goto free_log_buf;
2179 		}
2180 
2181 		memcpy(env->prog->aux->used_maps, env->used_maps,
2182 		       sizeof(env->used_maps[0]) * env->used_map_cnt);
2183 		env->prog->aux->used_map_cnt = env->used_map_cnt;
2184 
2185 		/* program is valid. Convert pseudo bpf_ld_imm64 into generic
2186 		 * bpf_ld_imm64 instructions
2187 		 */
2188 		convert_pseudo_ld_imm64(env);
2189 	}
2190 
2191 free_log_buf:
2192 	if (log_level)
2193 		vfree(log_buf);
2194 free_env:
2195 	if (!env->prog->aux->used_maps)
2196 		/* if we didn't copy map pointers into bpf_prog_info, release
2197 		 * them now. Otherwise free_bpf_prog_info() will release them.
2198 		 */
2199 		release_maps(env);
2200 	*prog = env->prog;
2201 	kfree(env);
2202 	mutex_unlock(&bpf_verifier_lock);
2203 	return ret;
2204 }
2205