1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Linux Socket Filter - Kernel level socket filtering
4 *
5 * Based on the design of the Berkeley Packet Filter. The new
6 * internal format has been designed by PLUMgrid:
7 *
8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
9 *
10 * Authors:
11 *
12 * Jay Schulist <jschlst@samba.org>
13 * Alexei Starovoitov <ast@plumgrid.com>
14 * Daniel Borkmann <dborkman@redhat.com>
15 *
16 * Andi Kleen - Fix a few bad bugs and races.
17 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
18 */
19
20 #include <uapi/linux/btf.h>
21 #include <crypto/sha1.h>
22 #include <linux/filter.h>
23 #include <linux/skbuff.h>
24 #include <linux/vmalloc.h>
25 #include <linux/prandom.h>
26 #include <linux/bpf.h>
27 #include <linux/btf.h>
28 #include <linux/objtool.h>
29 #include <linux/overflow.h>
30 #include <linux/rbtree_latch.h>
31 #include <linux/kallsyms.h>
32 #include <linux/rcupdate.h>
33 #include <linux/perf_event.h>
34 #include <linux/extable.h>
35 #include <linux/log2.h>
36 #include <linux/bpf_verifier.h>
37 #include <linux/nodemask.h>
38 #include <linux/nospec.h>
39 #include <linux/bpf_mem_alloc.h>
40 #include <linux/memcontrol.h>
41 #include <linux/execmem.h>
42 #include <crypto/sha2.h>
43
44 #include <asm/barrier.h>
45 #include <linux/unaligned.h>
46
47 /* Registers */
48 #define BPF_R0 regs[BPF_REG_0]
49 #define BPF_R1 regs[BPF_REG_1]
50 #define BPF_R2 regs[BPF_REG_2]
51 #define BPF_R3 regs[BPF_REG_3]
52 #define BPF_R4 regs[BPF_REG_4]
53 #define BPF_R5 regs[BPF_REG_5]
54 #define BPF_R6 regs[BPF_REG_6]
55 #define BPF_R7 regs[BPF_REG_7]
56 #define BPF_R8 regs[BPF_REG_8]
57 #define BPF_R9 regs[BPF_REG_9]
58 #define BPF_R10 regs[BPF_REG_10]
59
60 /* Named registers */
61 #define DST regs[insn->dst_reg]
62 #define SRC regs[insn->src_reg]
63 #define FP regs[BPF_REG_FP]
64 #define AX regs[BPF_REG_AX]
65 #define ARG1 regs[BPF_REG_ARG1]
66 #define CTX regs[BPF_REG_CTX]
67 #define OFF insn->off
68 #define IMM insn->imm
69
70 struct bpf_mem_alloc bpf_global_ma;
71 bool bpf_global_ma_set;
72
73 /* No hurry in this branch
74 *
75 * Exported for the bpf jit load helper.
76 */
bpf_internal_load_pointer_neg_helper(const struct sk_buff * skb,int k,unsigned int size)77 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
78 {
79 u8 *ptr = NULL;
80
81 if (k >= SKF_NET_OFF) {
82 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
83 } else if (k >= SKF_LL_OFF) {
84 if (unlikely(!skb_mac_header_was_set(skb)))
85 return NULL;
86 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
87 }
88 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
89 return ptr;
90
91 return NULL;
92 }
93
94 /* tell bpf programs that include vmlinux.h kernel's PAGE_SIZE */
95 enum page_size_enum {
96 __PAGE_SIZE = PAGE_SIZE
97 };
98
bpf_prog_alloc_no_stats(unsigned int size,gfp_t gfp_extra_flags)99 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
100 {
101 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
102 struct bpf_prog_aux *aux;
103 struct bpf_prog *fp;
104
105 size = round_up(size, __PAGE_SIZE);
106 fp = __vmalloc(size, gfp_flags);
107 if (fp == NULL)
108 return NULL;
109
110 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
111 if (aux == NULL) {
112 vfree(fp);
113 return NULL;
114 }
115 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
116 if (!fp->active) {
117 vfree(fp);
118 kfree(aux);
119 return NULL;
120 }
121
122 fp->pages = size / PAGE_SIZE;
123 fp->aux = aux;
124 fp->aux->main_prog_aux = aux;
125 fp->aux->prog = fp;
126 fp->jit_requested = ebpf_jit_enabled();
127 fp->blinding_requested = bpf_jit_blinding_enabled(fp);
128 #ifdef CONFIG_CGROUP_BPF
129 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
130 #endif
131
132 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
133 #ifdef CONFIG_FINEIBT
134 INIT_LIST_HEAD_RCU(&fp->aux->ksym_prefix.lnode);
135 #endif
136 mutex_init(&fp->aux->used_maps_mutex);
137 mutex_init(&fp->aux->ext_mutex);
138 mutex_init(&fp->aux->dst_mutex);
139
140 #ifdef CONFIG_BPF_SYSCALL
141 bpf_prog_stream_init(fp);
142 #endif
143
144 return fp;
145 }
146
bpf_prog_alloc(unsigned int size,gfp_t gfp_extra_flags)147 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
148 {
149 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
150 struct bpf_prog *prog;
151 int cpu;
152
153 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
154 if (!prog)
155 return NULL;
156
157 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
158 if (!prog->stats) {
159 free_percpu(prog->active);
160 kfree(prog->aux);
161 vfree(prog);
162 return NULL;
163 }
164
165 for_each_possible_cpu(cpu) {
166 struct bpf_prog_stats *pstats;
167
168 pstats = per_cpu_ptr(prog->stats, cpu);
169 u64_stats_init(&pstats->syncp);
170 }
171 return prog;
172 }
173 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
174
bpf_prog_alloc_jited_linfo(struct bpf_prog * prog)175 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
176 {
177 if (!prog->aux->nr_linfo || !prog->jit_requested)
178 return 0;
179
180 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
181 sizeof(*prog->aux->jited_linfo),
182 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN));
183 if (!prog->aux->jited_linfo)
184 return -ENOMEM;
185
186 return 0;
187 }
188
bpf_prog_jit_attempt_done(struct bpf_prog * prog)189 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
190 {
191 if (prog->aux->jited_linfo &&
192 (!prog->jited || !prog->aux->jited_linfo[0])) {
193 kvfree(prog->aux->jited_linfo);
194 prog->aux->jited_linfo = NULL;
195 }
196
197 kfree(prog->aux->kfunc_tab);
198 prog->aux->kfunc_tab = NULL;
199 }
200
201 /* The jit engine is responsible to provide an array
202 * for insn_off to the jited_off mapping (insn_to_jit_off).
203 *
204 * The idx to this array is the insn_off. Hence, the insn_off
205 * here is relative to the prog itself instead of the main prog.
206 * This array has one entry for each xlated bpf insn.
207 *
208 * jited_off is the byte off to the end of the jited insn.
209 *
210 * Hence, with
211 * insn_start:
212 * The first bpf insn off of the prog. The insn off
213 * here is relative to the main prog.
214 * e.g. if prog is a subprog, insn_start > 0
215 * linfo_idx:
216 * The prog's idx to prog->aux->linfo and jited_linfo
217 *
218 * jited_linfo[linfo_idx] = prog->bpf_func
219 *
220 * For i > linfo_idx,
221 *
222 * jited_linfo[i] = prog->bpf_func +
223 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
224 */
bpf_prog_fill_jited_linfo(struct bpf_prog * prog,const u32 * insn_to_jit_off)225 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
226 const u32 *insn_to_jit_off)
227 {
228 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
229 const struct bpf_line_info *linfo;
230 void **jited_linfo;
231
232 if (!prog->aux->jited_linfo || prog->aux->func_idx > prog->aux->func_cnt)
233 /* Userspace did not provide linfo */
234 return;
235
236 linfo_idx = prog->aux->linfo_idx;
237 linfo = &prog->aux->linfo[linfo_idx];
238 insn_start = linfo[0].insn_off;
239 insn_end = insn_start + prog->len;
240
241 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
242 jited_linfo[0] = prog->bpf_func;
243
244 nr_linfo = prog->aux->nr_linfo - linfo_idx;
245
246 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
247 /* The verifier ensures that linfo[i].insn_off is
248 * strictly increasing
249 */
250 jited_linfo[i] = prog->bpf_func +
251 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
252 }
253
bpf_prog_realloc(struct bpf_prog * fp_old,unsigned int size,gfp_t gfp_extra_flags)254 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
255 gfp_t gfp_extra_flags)
256 {
257 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
258 struct bpf_prog *fp;
259 u32 pages;
260
261 size = round_up(size, PAGE_SIZE);
262 pages = size / PAGE_SIZE;
263 if (pages <= fp_old->pages)
264 return fp_old;
265
266 fp = __vmalloc(size, gfp_flags);
267 if (fp) {
268 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
269 fp->pages = pages;
270 fp->aux->prog = fp;
271
272 /* We keep fp->aux from fp_old around in the new
273 * reallocated structure.
274 */
275 fp_old->aux = NULL;
276 fp_old->stats = NULL;
277 fp_old->active = NULL;
278 __bpf_prog_free(fp_old);
279 }
280
281 return fp;
282 }
283
__bpf_prog_free(struct bpf_prog * fp)284 void __bpf_prog_free(struct bpf_prog *fp)
285 {
286 if (fp->aux) {
287 mutex_destroy(&fp->aux->used_maps_mutex);
288 mutex_destroy(&fp->aux->dst_mutex);
289 kfree(fp->aux->poke_tab);
290 kfree(fp->aux);
291 }
292 free_percpu(fp->stats);
293 free_percpu(fp->active);
294 vfree(fp);
295 }
296
bpf_prog_calc_tag(struct bpf_prog * fp)297 int bpf_prog_calc_tag(struct bpf_prog *fp)
298 {
299 size_t size = bpf_prog_insn_size(fp);
300 struct bpf_insn *dst;
301 bool was_ld_map;
302 u32 i;
303
304 dst = vmalloc(size);
305 if (!dst)
306 return -ENOMEM;
307
308 /* We need to take out the map fd for the digest calculation
309 * since they are unstable from user space side.
310 */
311 for (i = 0, was_ld_map = false; i < fp->len; i++) {
312 dst[i] = fp->insnsi[i];
313 if (!was_ld_map &&
314 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
315 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
316 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
317 was_ld_map = true;
318 dst[i].imm = 0;
319 } else if (was_ld_map &&
320 dst[i].code == 0 &&
321 dst[i].dst_reg == 0 &&
322 dst[i].src_reg == 0 &&
323 dst[i].off == 0) {
324 was_ld_map = false;
325 dst[i].imm = 0;
326 } else {
327 was_ld_map = false;
328 }
329 }
330 sha256((u8 *)dst, size, fp->digest);
331 vfree(dst);
332 return 0;
333 }
334
bpf_adj_delta_to_imm(struct bpf_insn * insn,u32 pos,s32 end_old,s32 end_new,s32 curr,const bool probe_pass)335 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
336 s32 end_new, s32 curr, const bool probe_pass)
337 {
338 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
339 s32 delta = end_new - end_old;
340 s64 imm = insn->imm;
341
342 if (curr < pos && curr + imm + 1 >= end_old)
343 imm += delta;
344 else if (curr >= end_new && curr + imm + 1 < end_new)
345 imm -= delta;
346 if (imm < imm_min || imm > imm_max)
347 return -ERANGE;
348 if (!probe_pass)
349 insn->imm = imm;
350 return 0;
351 }
352
bpf_adj_delta_to_off(struct bpf_insn * insn,u32 pos,s32 end_old,s32 end_new,s32 curr,const bool probe_pass)353 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
354 s32 end_new, s32 curr, const bool probe_pass)
355 {
356 s64 off_min, off_max, off;
357 s32 delta = end_new - end_old;
358
359 if (insn->code == (BPF_JMP32 | BPF_JA)) {
360 off = insn->imm;
361 off_min = S32_MIN;
362 off_max = S32_MAX;
363 } else {
364 off = insn->off;
365 off_min = S16_MIN;
366 off_max = S16_MAX;
367 }
368
369 if (curr < pos && curr + off + 1 >= end_old)
370 off += delta;
371 else if (curr >= end_new && curr + off + 1 < end_new)
372 off -= delta;
373 if (off < off_min || off > off_max)
374 return -ERANGE;
375 if (!probe_pass) {
376 if (insn->code == (BPF_JMP32 | BPF_JA))
377 insn->imm = off;
378 else
379 insn->off = off;
380 }
381 return 0;
382 }
383
bpf_adj_branches(struct bpf_prog * prog,u32 pos,s32 end_old,s32 end_new,const bool probe_pass)384 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
385 s32 end_new, const bool probe_pass)
386 {
387 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
388 struct bpf_insn *insn = prog->insnsi;
389 int ret = 0;
390
391 for (i = 0; i < insn_cnt; i++, insn++) {
392 u8 code;
393
394 /* In the probing pass we still operate on the original,
395 * unpatched image in order to check overflows before we
396 * do any other adjustments. Therefore skip the patchlet.
397 */
398 if (probe_pass && i == pos) {
399 i = end_new;
400 insn = prog->insnsi + end_old;
401 }
402 if (bpf_pseudo_func(insn)) {
403 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
404 end_new, i, probe_pass);
405 if (ret)
406 return ret;
407 continue;
408 }
409 code = insn->code;
410 if ((BPF_CLASS(code) != BPF_JMP &&
411 BPF_CLASS(code) != BPF_JMP32) ||
412 BPF_OP(code) == BPF_EXIT)
413 continue;
414 /* Adjust offset of jmps if we cross patch boundaries. */
415 if (BPF_OP(code) == BPF_CALL) {
416 if (insn->src_reg != BPF_PSEUDO_CALL)
417 continue;
418 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
419 end_new, i, probe_pass);
420 } else {
421 ret = bpf_adj_delta_to_off(insn, pos, end_old,
422 end_new, i, probe_pass);
423 }
424 if (ret)
425 break;
426 }
427
428 return ret;
429 }
430
bpf_adj_linfo(struct bpf_prog * prog,u32 off,u32 delta)431 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
432 {
433 struct bpf_line_info *linfo;
434 u32 i, nr_linfo;
435
436 nr_linfo = prog->aux->nr_linfo;
437 if (!nr_linfo || !delta)
438 return;
439
440 linfo = prog->aux->linfo;
441
442 for (i = 0; i < nr_linfo; i++)
443 if (off < linfo[i].insn_off)
444 break;
445
446 /* Push all off < linfo[i].insn_off by delta */
447 for (; i < nr_linfo; i++)
448 linfo[i].insn_off += delta;
449 }
450
bpf_patch_insn_single(struct bpf_prog * prog,u32 off,const struct bpf_insn * patch,u32 len)451 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
452 const struct bpf_insn *patch, u32 len)
453 {
454 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
455 const u32 cnt_max = S16_MAX;
456 struct bpf_prog *prog_adj;
457 int err;
458
459 /* Since our patchlet doesn't expand the image, we're done. */
460 if (insn_delta == 0) {
461 memcpy(prog->insnsi + off, patch, sizeof(*patch));
462 return prog;
463 }
464
465 insn_adj_cnt = prog->len + insn_delta;
466
467 /* Reject anything that would potentially let the insn->off
468 * target overflow when we have excessive program expansions.
469 * We need to probe here before we do any reallocation where
470 * we afterwards may not fail anymore.
471 */
472 if (insn_adj_cnt > cnt_max &&
473 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
474 return ERR_PTR(err);
475
476 /* Several new instructions need to be inserted. Make room
477 * for them. Likely, there's no need for a new allocation as
478 * last page could have large enough tailroom.
479 */
480 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
481 GFP_USER);
482 if (!prog_adj)
483 return ERR_PTR(-ENOMEM);
484
485 prog_adj->len = insn_adj_cnt;
486
487 /* Patching happens in 3 steps:
488 *
489 * 1) Move over tail of insnsi from next instruction onwards,
490 * so we can patch the single target insn with one or more
491 * new ones (patching is always from 1 to n insns, n > 0).
492 * 2) Inject new instructions at the target location.
493 * 3) Adjust branch offsets if necessary.
494 */
495 insn_rest = insn_adj_cnt - off - len;
496
497 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
498 sizeof(*patch) * insn_rest);
499 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
500
501 /* We are guaranteed to not fail at this point, otherwise
502 * the ship has sailed to reverse to the original state. An
503 * overflow cannot happen at this point.
504 */
505 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
506
507 bpf_adj_linfo(prog_adj, off, insn_delta);
508
509 return prog_adj;
510 }
511
bpf_remove_insns(struct bpf_prog * prog,u32 off,u32 cnt)512 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
513 {
514 int err;
515
516 /* Branch offsets can't overflow when program is shrinking, no need
517 * to call bpf_adj_branches(..., true) here
518 */
519 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
520 sizeof(struct bpf_insn) * (prog->len - off - cnt));
521 prog->len -= cnt;
522
523 err = bpf_adj_branches(prog, off, off + cnt, off, false);
524 WARN_ON_ONCE(err);
525 return err;
526 }
527
bpf_prog_kallsyms_del_subprogs(struct bpf_prog * fp)528 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
529 {
530 int i;
531
532 for (i = 0; i < fp->aux->real_func_cnt; i++)
533 bpf_prog_kallsyms_del(fp->aux->func[i]);
534 }
535
bpf_prog_kallsyms_del_all(struct bpf_prog * fp)536 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
537 {
538 bpf_prog_kallsyms_del_subprogs(fp);
539 bpf_prog_kallsyms_del(fp);
540 }
541
542 #ifdef CONFIG_BPF_JIT
543 /* All BPF JIT sysctl knobs here. */
544 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
545 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
546 int bpf_jit_harden __read_mostly;
547 long bpf_jit_limit __read_mostly;
548 long bpf_jit_limit_max __read_mostly;
549
550 static void
bpf_prog_ksym_set_addr(struct bpf_prog * prog)551 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
552 {
553 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
554
555 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
556 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len;
557 }
558
559 static void
bpf_prog_ksym_set_name(struct bpf_prog * prog)560 bpf_prog_ksym_set_name(struct bpf_prog *prog)
561 {
562 char *sym = prog->aux->ksym.name;
563 const char *end = sym + KSYM_NAME_LEN;
564 const struct btf_type *type;
565 const char *func_name;
566
567 BUILD_BUG_ON(sizeof("bpf_prog_") +
568 sizeof(prog->tag) * 2 +
569 /* name has been null terminated.
570 * We should need +1 for the '_' preceding
571 * the name. However, the null character
572 * is double counted between the name and the
573 * sizeof("bpf_prog_") above, so we omit
574 * the +1 here.
575 */
576 sizeof(prog->aux->name) > KSYM_NAME_LEN);
577
578 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
579 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
580
581 /* prog->aux->name will be ignored if full btf name is available */
582 if (prog->aux->func_info_cnt && prog->aux->func_idx < prog->aux->func_info_cnt) {
583 type = btf_type_by_id(prog->aux->btf,
584 prog->aux->func_info[prog->aux->func_idx].type_id);
585 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
586 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
587 return;
588 }
589
590 if (prog->aux->name[0])
591 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
592 else
593 *sym = 0;
594 }
595
bpf_get_ksym_start(struct latch_tree_node * n)596 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
597 {
598 return container_of(n, struct bpf_ksym, tnode)->start;
599 }
600
bpf_tree_less(struct latch_tree_node * a,struct latch_tree_node * b)601 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
602 struct latch_tree_node *b)
603 {
604 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
605 }
606
bpf_tree_comp(void * key,struct latch_tree_node * n)607 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
608 {
609 unsigned long val = (unsigned long)key;
610 const struct bpf_ksym *ksym;
611
612 ksym = container_of(n, struct bpf_ksym, tnode);
613
614 if (val < ksym->start)
615 return -1;
616 /* Ensure that we detect return addresses as part of the program, when
617 * the final instruction is a call for a program part of the stack
618 * trace. Therefore, do val > ksym->end instead of val >= ksym->end.
619 */
620 if (val > ksym->end)
621 return 1;
622
623 return 0;
624 }
625
626 static const struct latch_tree_ops bpf_tree_ops = {
627 .less = bpf_tree_less,
628 .comp = bpf_tree_comp,
629 };
630
631 static DEFINE_SPINLOCK(bpf_lock);
632 static LIST_HEAD(bpf_kallsyms);
633 static struct latch_tree_root bpf_tree __cacheline_aligned;
634
bpf_ksym_add(struct bpf_ksym * ksym)635 void bpf_ksym_add(struct bpf_ksym *ksym)
636 {
637 spin_lock_bh(&bpf_lock);
638 WARN_ON_ONCE(!list_empty(&ksym->lnode));
639 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
640 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
641 spin_unlock_bh(&bpf_lock);
642 }
643
__bpf_ksym_del(struct bpf_ksym * ksym)644 static void __bpf_ksym_del(struct bpf_ksym *ksym)
645 {
646 if (list_empty(&ksym->lnode))
647 return;
648
649 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
650 list_del_rcu(&ksym->lnode);
651 }
652
bpf_ksym_del(struct bpf_ksym * ksym)653 void bpf_ksym_del(struct bpf_ksym *ksym)
654 {
655 spin_lock_bh(&bpf_lock);
656 __bpf_ksym_del(ksym);
657 spin_unlock_bh(&bpf_lock);
658 }
659
bpf_prog_kallsyms_candidate(const struct bpf_prog * fp)660 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
661 {
662 return fp->jited && !bpf_prog_was_classic(fp);
663 }
664
bpf_prog_kallsyms_add(struct bpf_prog * fp)665 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
666 {
667 if (!bpf_prog_kallsyms_candidate(fp) ||
668 !bpf_token_capable(fp->aux->token, CAP_BPF))
669 return;
670
671 bpf_prog_ksym_set_addr(fp);
672 bpf_prog_ksym_set_name(fp);
673 fp->aux->ksym.prog = true;
674
675 bpf_ksym_add(&fp->aux->ksym);
676
677 #ifdef CONFIG_FINEIBT
678 /*
679 * When FineIBT, code in the __cfi_foo() symbols can get executed
680 * and hence unwinder needs help.
681 */
682 if (cfi_mode != CFI_FINEIBT)
683 return;
684
685 snprintf(fp->aux->ksym_prefix.name, KSYM_NAME_LEN,
686 "__cfi_%s", fp->aux->ksym.name);
687
688 fp->aux->ksym_prefix.start = (unsigned long) fp->bpf_func - 16;
689 fp->aux->ksym_prefix.end = (unsigned long) fp->bpf_func;
690
691 bpf_ksym_add(&fp->aux->ksym_prefix);
692 #endif
693 }
694
bpf_prog_kallsyms_del(struct bpf_prog * fp)695 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
696 {
697 if (!bpf_prog_kallsyms_candidate(fp))
698 return;
699
700 bpf_ksym_del(&fp->aux->ksym);
701 #ifdef CONFIG_FINEIBT
702 if (cfi_mode != CFI_FINEIBT)
703 return;
704 bpf_ksym_del(&fp->aux->ksym_prefix);
705 #endif
706 }
707
bpf_ksym_find(unsigned long addr)708 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
709 {
710 struct latch_tree_node *n;
711
712 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
713 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
714 }
715
__bpf_address_lookup(unsigned long addr,unsigned long * size,unsigned long * off,char * sym)716 int __bpf_address_lookup(unsigned long addr, unsigned long *size,
717 unsigned long *off, char *sym)
718 {
719 struct bpf_ksym *ksym;
720 int ret = 0;
721
722 rcu_read_lock();
723 ksym = bpf_ksym_find(addr);
724 if (ksym) {
725 unsigned long symbol_start = ksym->start;
726 unsigned long symbol_end = ksym->end;
727
728 ret = strscpy(sym, ksym->name, KSYM_NAME_LEN);
729
730 if (size)
731 *size = symbol_end - symbol_start;
732 if (off)
733 *off = addr - symbol_start;
734 }
735 rcu_read_unlock();
736
737 return ret;
738 }
739
is_bpf_text_address(unsigned long addr)740 bool is_bpf_text_address(unsigned long addr)
741 {
742 bool ret;
743
744 rcu_read_lock();
745 ret = bpf_ksym_find(addr) != NULL;
746 rcu_read_unlock();
747
748 return ret;
749 }
750
bpf_prog_ksym_find(unsigned long addr)751 struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
752 {
753 struct bpf_ksym *ksym;
754
755 WARN_ON_ONCE(!rcu_read_lock_held());
756 ksym = bpf_ksym_find(addr);
757
758 return ksym && ksym->prog ?
759 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
760 NULL;
761 }
762
search_bpf_extables(unsigned long addr)763 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
764 {
765 const struct exception_table_entry *e = NULL;
766 struct bpf_prog *prog;
767
768 rcu_read_lock();
769 prog = bpf_prog_ksym_find(addr);
770 if (!prog)
771 goto out;
772 if (!prog->aux->num_exentries)
773 goto out;
774
775 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
776 out:
777 rcu_read_unlock();
778 return e;
779 }
780
bpf_get_kallsym(unsigned int symnum,unsigned long * value,char * type,char * sym)781 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
782 char *sym)
783 {
784 struct bpf_ksym *ksym;
785 unsigned int it = 0;
786 int ret = -ERANGE;
787
788 if (!bpf_jit_kallsyms_enabled())
789 return ret;
790
791 rcu_read_lock();
792 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
793 if (it++ != symnum)
794 continue;
795
796 strscpy(sym, ksym->name, KSYM_NAME_LEN);
797
798 *value = ksym->start;
799 *type = BPF_SYM_ELF_TYPE;
800
801 ret = 0;
802 break;
803 }
804 rcu_read_unlock();
805
806 return ret;
807 }
808
bpf_jit_add_poke_descriptor(struct bpf_prog * prog,struct bpf_jit_poke_descriptor * poke)809 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
810 struct bpf_jit_poke_descriptor *poke)
811 {
812 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
813 static const u32 poke_tab_max = 1024;
814 u32 slot = prog->aux->size_poke_tab;
815 u32 size = slot + 1;
816
817 if (size > poke_tab_max)
818 return -ENOSPC;
819 if (poke->tailcall_target || poke->tailcall_target_stable ||
820 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
821 return -EINVAL;
822
823 switch (poke->reason) {
824 case BPF_POKE_REASON_TAIL_CALL:
825 if (!poke->tail_call.map)
826 return -EINVAL;
827 break;
828 default:
829 return -EINVAL;
830 }
831
832 tab = krealloc_array(tab, size, sizeof(*poke), GFP_KERNEL);
833 if (!tab)
834 return -ENOMEM;
835
836 memcpy(&tab[slot], poke, sizeof(*poke));
837 prog->aux->size_poke_tab = size;
838 prog->aux->poke_tab = tab;
839
840 return slot;
841 }
842
843 /*
844 * BPF program pack allocator.
845 *
846 * Most BPF programs are pretty small. Allocating a hole page for each
847 * program is sometime a waste. Many small bpf program also adds pressure
848 * to instruction TLB. To solve this issue, we introduce a BPF program pack
849 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
850 * to host BPF programs.
851 */
852 #define BPF_PROG_CHUNK_SHIFT 6
853 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT)
854 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1))
855
856 struct bpf_prog_pack {
857 struct list_head list;
858 void *ptr;
859 unsigned long bitmap[];
860 };
861
bpf_jit_fill_hole_with_zero(void * area,unsigned int size)862 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size)
863 {
864 memset(area, 0, size);
865 }
866
867 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
868
869 static DEFINE_MUTEX(pack_mutex);
870 static LIST_HEAD(pack_list);
871
872 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
873 * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
874 */
875 #ifdef PMD_SIZE
876 /* PMD_SIZE is really big for some archs. It doesn't make sense to
877 * reserve too much memory in one allocation. Hardcode BPF_PROG_PACK_SIZE to
878 * 2MiB * num_possible_nodes(). On most architectures PMD_SIZE will be
879 * greater than or equal to 2MB.
880 */
881 #define BPF_PROG_PACK_SIZE (SZ_2M * num_possible_nodes())
882 #else
883 #define BPF_PROG_PACK_SIZE PAGE_SIZE
884 #endif
885
886 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
887
alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)888 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
889 {
890 struct bpf_prog_pack *pack;
891 int err;
892
893 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
894 GFP_KERNEL);
895 if (!pack)
896 return NULL;
897 pack->ptr = bpf_jit_alloc_exec(BPF_PROG_PACK_SIZE);
898 if (!pack->ptr)
899 goto out;
900 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
901 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
902
903 set_vm_flush_reset_perms(pack->ptr);
904 err = set_memory_rox((unsigned long)pack->ptr,
905 BPF_PROG_PACK_SIZE / PAGE_SIZE);
906 if (err)
907 goto out;
908 list_add_tail(&pack->list, &pack_list);
909 return pack;
910
911 out:
912 bpf_jit_free_exec(pack->ptr);
913 kfree(pack);
914 return NULL;
915 }
916
bpf_prog_pack_alloc(u32 size,bpf_jit_fill_hole_t bpf_fill_ill_insns)917 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
918 {
919 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
920 struct bpf_prog_pack *pack;
921 unsigned long pos;
922 void *ptr = NULL;
923
924 mutex_lock(&pack_mutex);
925 if (size > BPF_PROG_PACK_SIZE) {
926 size = round_up(size, PAGE_SIZE);
927 ptr = bpf_jit_alloc_exec(size);
928 if (ptr) {
929 int err;
930
931 bpf_fill_ill_insns(ptr, size);
932 set_vm_flush_reset_perms(ptr);
933 err = set_memory_rox((unsigned long)ptr,
934 size / PAGE_SIZE);
935 if (err) {
936 bpf_jit_free_exec(ptr);
937 ptr = NULL;
938 }
939 }
940 goto out;
941 }
942 list_for_each_entry(pack, &pack_list, list) {
943 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
944 nbits, 0);
945 if (pos < BPF_PROG_CHUNK_COUNT)
946 goto found_free_area;
947 }
948
949 pack = alloc_new_pack(bpf_fill_ill_insns);
950 if (!pack)
951 goto out;
952
953 pos = 0;
954
955 found_free_area:
956 bitmap_set(pack->bitmap, pos, nbits);
957 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
958
959 out:
960 mutex_unlock(&pack_mutex);
961 return ptr;
962 }
963
bpf_prog_pack_free(void * ptr,u32 size)964 void bpf_prog_pack_free(void *ptr, u32 size)
965 {
966 struct bpf_prog_pack *pack = NULL, *tmp;
967 unsigned int nbits;
968 unsigned long pos;
969
970 mutex_lock(&pack_mutex);
971 if (size > BPF_PROG_PACK_SIZE) {
972 bpf_jit_free_exec(ptr);
973 goto out;
974 }
975
976 list_for_each_entry(tmp, &pack_list, list) {
977 if (ptr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > ptr) {
978 pack = tmp;
979 break;
980 }
981 }
982
983 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
984 goto out;
985
986 nbits = BPF_PROG_SIZE_TO_NBITS(size);
987 pos = ((unsigned long)ptr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
988
989 WARN_ONCE(bpf_arch_text_invalidate(ptr, size),
990 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
991
992 bitmap_clear(pack->bitmap, pos, nbits);
993 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
994 BPF_PROG_CHUNK_COUNT, 0) == 0) {
995 list_del(&pack->list);
996 bpf_jit_free_exec(pack->ptr);
997 kfree(pack);
998 }
999 out:
1000 mutex_unlock(&pack_mutex);
1001 }
1002
1003 static atomic_long_t bpf_jit_current;
1004
1005 /* Can be overridden by an arch's JIT compiler if it has a custom,
1006 * dedicated BPF backend memory area, or if neither of the two
1007 * below apply.
1008 */
bpf_jit_alloc_exec_limit(void)1009 u64 __weak bpf_jit_alloc_exec_limit(void)
1010 {
1011 #if defined(MODULES_VADDR)
1012 return MODULES_END - MODULES_VADDR;
1013 #else
1014 return VMALLOC_END - VMALLOC_START;
1015 #endif
1016 }
1017
bpf_jit_charge_init(void)1018 static int __init bpf_jit_charge_init(void)
1019 {
1020 /* Only used as heuristic here to derive limit. */
1021 bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
1022 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1,
1023 PAGE_SIZE), LONG_MAX);
1024 return 0;
1025 }
1026 pure_initcall(bpf_jit_charge_init);
1027
bpf_jit_charge_modmem(u32 size)1028 int bpf_jit_charge_modmem(u32 size)
1029 {
1030 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
1031 if (!bpf_capable()) {
1032 atomic_long_sub(size, &bpf_jit_current);
1033 return -EPERM;
1034 }
1035 }
1036
1037 return 0;
1038 }
1039
bpf_jit_uncharge_modmem(u32 size)1040 void bpf_jit_uncharge_modmem(u32 size)
1041 {
1042 atomic_long_sub(size, &bpf_jit_current);
1043 }
1044
bpf_jit_alloc_exec(unsigned long size)1045 void *__weak bpf_jit_alloc_exec(unsigned long size)
1046 {
1047 return execmem_alloc(EXECMEM_BPF, size);
1048 }
1049
bpf_jit_free_exec(void * addr)1050 void __weak bpf_jit_free_exec(void *addr)
1051 {
1052 execmem_free(addr);
1053 }
1054
1055 struct bpf_binary_header *
bpf_jit_binary_alloc(unsigned int proglen,u8 ** image_ptr,unsigned int alignment,bpf_jit_fill_hole_t bpf_fill_ill_insns)1056 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1057 unsigned int alignment,
1058 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1059 {
1060 struct bpf_binary_header *hdr;
1061 u32 size, hole, start;
1062
1063 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1064 alignment > BPF_IMAGE_ALIGNMENT);
1065
1066 /* Most of BPF filters are really small, but if some of them
1067 * fill a page, allow at least 128 extra bytes to insert a
1068 * random section of illegal instructions.
1069 */
1070 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1071
1072 if (bpf_jit_charge_modmem(size))
1073 return NULL;
1074 hdr = bpf_jit_alloc_exec(size);
1075 if (!hdr) {
1076 bpf_jit_uncharge_modmem(size);
1077 return NULL;
1078 }
1079
1080 /* Fill space with illegal/arch-dep instructions. */
1081 bpf_fill_ill_insns(hdr, size);
1082
1083 hdr->size = size;
1084 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1085 PAGE_SIZE - sizeof(*hdr));
1086 start = get_random_u32_below(hole) & ~(alignment - 1);
1087
1088 /* Leave a random number of instructions before BPF code. */
1089 *image_ptr = &hdr->image[start];
1090
1091 return hdr;
1092 }
1093
bpf_jit_binary_free(struct bpf_binary_header * hdr)1094 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1095 {
1096 u32 size = hdr->size;
1097
1098 bpf_jit_free_exec(hdr);
1099 bpf_jit_uncharge_modmem(size);
1100 }
1101
1102 /* Allocate jit binary from bpf_prog_pack allocator.
1103 * Since the allocated memory is RO+X, the JIT engine cannot write directly
1104 * to the memory. To solve this problem, a RW buffer is also allocated at
1105 * as the same time. The JIT engine should calculate offsets based on the
1106 * RO memory address, but write JITed program to the RW buffer. Once the
1107 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1108 * the JITed program to the RO memory.
1109 */
1110 struct bpf_binary_header *
bpf_jit_binary_pack_alloc(unsigned int proglen,u8 ** image_ptr,unsigned int alignment,struct bpf_binary_header ** rw_header,u8 ** rw_image,bpf_jit_fill_hole_t bpf_fill_ill_insns)1111 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1112 unsigned int alignment,
1113 struct bpf_binary_header **rw_header,
1114 u8 **rw_image,
1115 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1116 {
1117 struct bpf_binary_header *ro_header;
1118 u32 size, hole, start;
1119
1120 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1121 alignment > BPF_IMAGE_ALIGNMENT);
1122
1123 /* add 16 bytes for a random section of illegal instructions */
1124 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1125
1126 if (bpf_jit_charge_modmem(size))
1127 return NULL;
1128 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1129 if (!ro_header) {
1130 bpf_jit_uncharge_modmem(size);
1131 return NULL;
1132 }
1133
1134 *rw_header = kvmalloc(size, GFP_KERNEL);
1135 if (!*rw_header) {
1136 bpf_prog_pack_free(ro_header, size);
1137 bpf_jit_uncharge_modmem(size);
1138 return NULL;
1139 }
1140
1141 /* Fill space with illegal/arch-dep instructions. */
1142 bpf_fill_ill_insns(*rw_header, size);
1143 (*rw_header)->size = size;
1144
1145 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1146 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1147 start = get_random_u32_below(hole) & ~(alignment - 1);
1148
1149 *image_ptr = &ro_header->image[start];
1150 *rw_image = &(*rw_header)->image[start];
1151
1152 return ro_header;
1153 }
1154
1155 /* Copy JITed text from rw_header to its final location, the ro_header. */
bpf_jit_binary_pack_finalize(struct bpf_binary_header * ro_header,struct bpf_binary_header * rw_header)1156 int bpf_jit_binary_pack_finalize(struct bpf_binary_header *ro_header,
1157 struct bpf_binary_header *rw_header)
1158 {
1159 void *ptr;
1160
1161 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1162
1163 kvfree(rw_header);
1164
1165 if (IS_ERR(ptr)) {
1166 bpf_prog_pack_free(ro_header, ro_header->size);
1167 return PTR_ERR(ptr);
1168 }
1169 return 0;
1170 }
1171
1172 /* bpf_jit_binary_pack_free is called in two different scenarios:
1173 * 1) when the program is freed after;
1174 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1175 * For case 2), we need to free both the RO memory and the RW buffer.
1176 *
1177 * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1178 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1179 * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1180 * bpf_arch_text_copy (when jit fails).
1181 */
bpf_jit_binary_pack_free(struct bpf_binary_header * ro_header,struct bpf_binary_header * rw_header)1182 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1183 struct bpf_binary_header *rw_header)
1184 {
1185 u32 size = ro_header->size;
1186
1187 bpf_prog_pack_free(ro_header, size);
1188 kvfree(rw_header);
1189 bpf_jit_uncharge_modmem(size);
1190 }
1191
1192 struct bpf_binary_header *
bpf_jit_binary_pack_hdr(const struct bpf_prog * fp)1193 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1194 {
1195 unsigned long real_start = (unsigned long)fp->bpf_func;
1196 unsigned long addr;
1197
1198 addr = real_start & BPF_PROG_CHUNK_MASK;
1199 return (void *)addr;
1200 }
1201
1202 static inline struct bpf_binary_header *
bpf_jit_binary_hdr(const struct bpf_prog * fp)1203 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1204 {
1205 unsigned long real_start = (unsigned long)fp->bpf_func;
1206 unsigned long addr;
1207
1208 addr = real_start & PAGE_MASK;
1209 return (void *)addr;
1210 }
1211
1212 /* This symbol is only overridden by archs that have different
1213 * requirements than the usual eBPF JITs, f.e. when they only
1214 * implement cBPF JIT, do not set images read-only, etc.
1215 */
bpf_jit_free(struct bpf_prog * fp)1216 void __weak bpf_jit_free(struct bpf_prog *fp)
1217 {
1218 if (fp->jited) {
1219 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1220
1221 bpf_jit_binary_free(hdr);
1222 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1223 }
1224
1225 bpf_prog_unlock_free(fp);
1226 }
1227
bpf_jit_get_func_addr(const struct bpf_prog * prog,const struct bpf_insn * insn,bool extra_pass,u64 * func_addr,bool * func_addr_fixed)1228 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1229 const struct bpf_insn *insn, bool extra_pass,
1230 u64 *func_addr, bool *func_addr_fixed)
1231 {
1232 s16 off = insn->off;
1233 s32 imm = insn->imm;
1234 u8 *addr;
1235 int err;
1236
1237 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1238 if (!*func_addr_fixed) {
1239 /* Place-holder address till the last pass has collected
1240 * all addresses for JITed subprograms in which case we
1241 * can pick them up from prog->aux.
1242 */
1243 if (!extra_pass)
1244 addr = NULL;
1245 else if (prog->aux->func &&
1246 off >= 0 && off < prog->aux->real_func_cnt)
1247 addr = (u8 *)prog->aux->func[off]->bpf_func;
1248 else
1249 return -EINVAL;
1250 } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
1251 bpf_jit_supports_far_kfunc_call()) {
1252 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr);
1253 if (err)
1254 return err;
1255 } else {
1256 /* Address of a BPF helper call. Since part of the core
1257 * kernel, it's always at a fixed location. __bpf_call_base
1258 * and the helper with imm relative to it are both in core
1259 * kernel.
1260 */
1261 addr = (u8 *)__bpf_call_base + imm;
1262 }
1263
1264 *func_addr = (unsigned long)addr;
1265 return 0;
1266 }
1267
bpf_jit_get_prog_name(struct bpf_prog * prog)1268 const char *bpf_jit_get_prog_name(struct bpf_prog *prog)
1269 {
1270 if (prog->aux->ksym.prog)
1271 return prog->aux->ksym.name;
1272 return prog->aux->name;
1273 }
1274
bpf_jit_blind_insn(const struct bpf_insn * from,const struct bpf_insn * aux,struct bpf_insn * to_buff,bool emit_zext)1275 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1276 const struct bpf_insn *aux,
1277 struct bpf_insn *to_buff,
1278 bool emit_zext)
1279 {
1280 struct bpf_insn *to = to_buff;
1281 u32 imm_rnd = get_random_u32();
1282 s16 off;
1283
1284 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
1285 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1286
1287 /* Constraints on AX register:
1288 *
1289 * AX register is inaccessible from user space. It is mapped in
1290 * all JITs, and used here for constant blinding rewrites. It is
1291 * typically "stateless" meaning its contents are only valid within
1292 * the executed instruction, but not across several instructions.
1293 * There are a few exceptions however which are further detailed
1294 * below.
1295 *
1296 * Constant blinding is only used by JITs, not in the interpreter.
1297 * The interpreter uses AX in some occasions as a local temporary
1298 * register e.g. in DIV or MOD instructions.
1299 *
1300 * In restricted circumstances, the verifier can also use the AX
1301 * register for rewrites as long as they do not interfere with
1302 * the above cases!
1303 */
1304 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1305 goto out;
1306
1307 if (from->imm == 0 &&
1308 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
1309 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1310 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1311 goto out;
1312 }
1313
1314 switch (from->code) {
1315 case BPF_ALU | BPF_ADD | BPF_K:
1316 case BPF_ALU | BPF_SUB | BPF_K:
1317 case BPF_ALU | BPF_AND | BPF_K:
1318 case BPF_ALU | BPF_OR | BPF_K:
1319 case BPF_ALU | BPF_XOR | BPF_K:
1320 case BPF_ALU | BPF_MUL | BPF_K:
1321 case BPF_ALU | BPF_MOV | BPF_K:
1322 case BPF_ALU | BPF_DIV | BPF_K:
1323 case BPF_ALU | BPF_MOD | BPF_K:
1324 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1325 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1326 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off);
1327 break;
1328
1329 case BPF_ALU64 | BPF_ADD | BPF_K:
1330 case BPF_ALU64 | BPF_SUB | BPF_K:
1331 case BPF_ALU64 | BPF_AND | BPF_K:
1332 case BPF_ALU64 | BPF_OR | BPF_K:
1333 case BPF_ALU64 | BPF_XOR | BPF_K:
1334 case BPF_ALU64 | BPF_MUL | BPF_K:
1335 case BPF_ALU64 | BPF_MOV | BPF_K:
1336 case BPF_ALU64 | BPF_DIV | BPF_K:
1337 case BPF_ALU64 | BPF_MOD | BPF_K:
1338 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1339 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1340 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off);
1341 break;
1342
1343 case BPF_JMP | BPF_JEQ | BPF_K:
1344 case BPF_JMP | BPF_JNE | BPF_K:
1345 case BPF_JMP | BPF_JGT | BPF_K:
1346 case BPF_JMP | BPF_JLT | BPF_K:
1347 case BPF_JMP | BPF_JGE | BPF_K:
1348 case BPF_JMP | BPF_JLE | BPF_K:
1349 case BPF_JMP | BPF_JSGT | BPF_K:
1350 case BPF_JMP | BPF_JSLT | BPF_K:
1351 case BPF_JMP | BPF_JSGE | BPF_K:
1352 case BPF_JMP | BPF_JSLE | BPF_K:
1353 case BPF_JMP | BPF_JSET | BPF_K:
1354 /* Accommodate for extra offset in case of a backjump. */
1355 off = from->off;
1356 if (off < 0)
1357 off -= 2;
1358 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1359 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1360 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1361 break;
1362
1363 case BPF_JMP32 | BPF_JEQ | BPF_K:
1364 case BPF_JMP32 | BPF_JNE | BPF_K:
1365 case BPF_JMP32 | BPF_JGT | BPF_K:
1366 case BPF_JMP32 | BPF_JLT | BPF_K:
1367 case BPF_JMP32 | BPF_JGE | BPF_K:
1368 case BPF_JMP32 | BPF_JLE | BPF_K:
1369 case BPF_JMP32 | BPF_JSGT | BPF_K:
1370 case BPF_JMP32 | BPF_JSLT | BPF_K:
1371 case BPF_JMP32 | BPF_JSGE | BPF_K:
1372 case BPF_JMP32 | BPF_JSLE | BPF_K:
1373 case BPF_JMP32 | BPF_JSET | BPF_K:
1374 /* Accommodate for extra offset in case of a backjump. */
1375 off = from->off;
1376 if (off < 0)
1377 off -= 2;
1378 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1379 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1380 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1381 off);
1382 break;
1383
1384 case BPF_LD | BPF_IMM | BPF_DW:
1385 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1386 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1387 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1388 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1389 break;
1390 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1391 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1392 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1393 if (emit_zext)
1394 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1395 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1396 break;
1397
1398 case BPF_ST | BPF_MEM | BPF_DW:
1399 case BPF_ST | BPF_MEM | BPF_W:
1400 case BPF_ST | BPF_MEM | BPF_H:
1401 case BPF_ST | BPF_MEM | BPF_B:
1402 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1403 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1404 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1405 break;
1406 }
1407 out:
1408 return to - to_buff;
1409 }
1410
bpf_prog_clone_create(struct bpf_prog * fp_other,gfp_t gfp_extra_flags)1411 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1412 gfp_t gfp_extra_flags)
1413 {
1414 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1415 struct bpf_prog *fp;
1416
1417 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1418 if (fp != NULL) {
1419 /* aux->prog still points to the fp_other one, so
1420 * when promoting the clone to the real program,
1421 * this still needs to be adapted.
1422 */
1423 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1424 }
1425
1426 return fp;
1427 }
1428
bpf_prog_clone_free(struct bpf_prog * fp)1429 static void bpf_prog_clone_free(struct bpf_prog *fp)
1430 {
1431 /* aux was stolen by the other clone, so we cannot free
1432 * it from this path! It will be freed eventually by the
1433 * other program on release.
1434 *
1435 * At this point, we don't need a deferred release since
1436 * clone is guaranteed to not be locked.
1437 */
1438 fp->aux = NULL;
1439 fp->stats = NULL;
1440 fp->active = NULL;
1441 __bpf_prog_free(fp);
1442 }
1443
bpf_jit_prog_release_other(struct bpf_prog * fp,struct bpf_prog * fp_other)1444 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1445 {
1446 /* We have to repoint aux->prog to self, as we don't
1447 * know whether fp here is the clone or the original.
1448 */
1449 fp->aux->prog = fp;
1450 bpf_prog_clone_free(fp_other);
1451 }
1452
bpf_jit_blind_constants(struct bpf_prog * prog)1453 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1454 {
1455 struct bpf_insn insn_buff[16], aux[2];
1456 struct bpf_prog *clone, *tmp;
1457 int insn_delta, insn_cnt;
1458 struct bpf_insn *insn;
1459 int i, rewritten;
1460
1461 if (!prog->blinding_requested || prog->blinded)
1462 return prog;
1463
1464 clone = bpf_prog_clone_create(prog, GFP_USER);
1465 if (!clone)
1466 return ERR_PTR(-ENOMEM);
1467
1468 insn_cnt = clone->len;
1469 insn = clone->insnsi;
1470
1471 for (i = 0; i < insn_cnt; i++, insn++) {
1472 if (bpf_pseudo_func(insn)) {
1473 /* ld_imm64 with an address of bpf subprog is not
1474 * a user controlled constant. Don't randomize it,
1475 * since it will conflict with jit_subprogs() logic.
1476 */
1477 insn++;
1478 i++;
1479 continue;
1480 }
1481
1482 /* We temporarily need to hold the original ld64 insn
1483 * so that we can still access the first part in the
1484 * second blinding run.
1485 */
1486 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1487 insn[1].code == 0)
1488 memcpy(aux, insn, sizeof(aux));
1489
1490 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1491 clone->aux->verifier_zext);
1492 if (!rewritten)
1493 continue;
1494
1495 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1496 if (IS_ERR(tmp)) {
1497 /* Patching may have repointed aux->prog during
1498 * realloc from the original one, so we need to
1499 * fix it up here on error.
1500 */
1501 bpf_jit_prog_release_other(prog, clone);
1502 return tmp;
1503 }
1504
1505 clone = tmp;
1506 insn_delta = rewritten - 1;
1507
1508 /* Walk new program and skip insns we just inserted. */
1509 insn = clone->insnsi + i + insn_delta;
1510 insn_cnt += insn_delta;
1511 i += insn_delta;
1512 }
1513
1514 clone->blinded = 1;
1515 return clone;
1516 }
1517 #endif /* CONFIG_BPF_JIT */
1518
1519 /* Base function for offset calculation. Needs to go into .text section,
1520 * therefore keeping it non-static as well; will also be used by JITs
1521 * anyway later on, so do not let the compiler omit it. This also needs
1522 * to go into kallsyms for correlation from e.g. bpftool, so naming
1523 * must not change.
1524 */
__bpf_call_base(u64 r1,u64 r2,u64 r3,u64 r4,u64 r5)1525 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1526 {
1527 return 0;
1528 }
1529 EXPORT_SYMBOL_GPL(__bpf_call_base);
1530
1531 /* All UAPI available opcodes. */
1532 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1533 /* 32 bit ALU operations. */ \
1534 /* Register based. */ \
1535 INSN_3(ALU, ADD, X), \
1536 INSN_3(ALU, SUB, X), \
1537 INSN_3(ALU, AND, X), \
1538 INSN_3(ALU, OR, X), \
1539 INSN_3(ALU, LSH, X), \
1540 INSN_3(ALU, RSH, X), \
1541 INSN_3(ALU, XOR, X), \
1542 INSN_3(ALU, MUL, X), \
1543 INSN_3(ALU, MOV, X), \
1544 INSN_3(ALU, ARSH, X), \
1545 INSN_3(ALU, DIV, X), \
1546 INSN_3(ALU, MOD, X), \
1547 INSN_2(ALU, NEG), \
1548 INSN_3(ALU, END, TO_BE), \
1549 INSN_3(ALU, END, TO_LE), \
1550 /* Immediate based. */ \
1551 INSN_3(ALU, ADD, K), \
1552 INSN_3(ALU, SUB, K), \
1553 INSN_3(ALU, AND, K), \
1554 INSN_3(ALU, OR, K), \
1555 INSN_3(ALU, LSH, K), \
1556 INSN_3(ALU, RSH, K), \
1557 INSN_3(ALU, XOR, K), \
1558 INSN_3(ALU, MUL, K), \
1559 INSN_3(ALU, MOV, K), \
1560 INSN_3(ALU, ARSH, K), \
1561 INSN_3(ALU, DIV, K), \
1562 INSN_3(ALU, MOD, K), \
1563 /* 64 bit ALU operations. */ \
1564 /* Register based. */ \
1565 INSN_3(ALU64, ADD, X), \
1566 INSN_3(ALU64, SUB, X), \
1567 INSN_3(ALU64, AND, X), \
1568 INSN_3(ALU64, OR, X), \
1569 INSN_3(ALU64, LSH, X), \
1570 INSN_3(ALU64, RSH, X), \
1571 INSN_3(ALU64, XOR, X), \
1572 INSN_3(ALU64, MUL, X), \
1573 INSN_3(ALU64, MOV, X), \
1574 INSN_3(ALU64, ARSH, X), \
1575 INSN_3(ALU64, DIV, X), \
1576 INSN_3(ALU64, MOD, X), \
1577 INSN_2(ALU64, NEG), \
1578 INSN_3(ALU64, END, TO_LE), \
1579 /* Immediate based. */ \
1580 INSN_3(ALU64, ADD, K), \
1581 INSN_3(ALU64, SUB, K), \
1582 INSN_3(ALU64, AND, K), \
1583 INSN_3(ALU64, OR, K), \
1584 INSN_3(ALU64, LSH, K), \
1585 INSN_3(ALU64, RSH, K), \
1586 INSN_3(ALU64, XOR, K), \
1587 INSN_3(ALU64, MUL, K), \
1588 INSN_3(ALU64, MOV, K), \
1589 INSN_3(ALU64, ARSH, K), \
1590 INSN_3(ALU64, DIV, K), \
1591 INSN_3(ALU64, MOD, K), \
1592 /* Call instruction. */ \
1593 INSN_2(JMP, CALL), \
1594 /* Exit instruction. */ \
1595 INSN_2(JMP, EXIT), \
1596 /* 32-bit Jump instructions. */ \
1597 /* Register based. */ \
1598 INSN_3(JMP32, JEQ, X), \
1599 INSN_3(JMP32, JNE, X), \
1600 INSN_3(JMP32, JGT, X), \
1601 INSN_3(JMP32, JLT, X), \
1602 INSN_3(JMP32, JGE, X), \
1603 INSN_3(JMP32, JLE, X), \
1604 INSN_3(JMP32, JSGT, X), \
1605 INSN_3(JMP32, JSLT, X), \
1606 INSN_3(JMP32, JSGE, X), \
1607 INSN_3(JMP32, JSLE, X), \
1608 INSN_3(JMP32, JSET, X), \
1609 /* Immediate based. */ \
1610 INSN_3(JMP32, JEQ, K), \
1611 INSN_3(JMP32, JNE, K), \
1612 INSN_3(JMP32, JGT, K), \
1613 INSN_3(JMP32, JLT, K), \
1614 INSN_3(JMP32, JGE, K), \
1615 INSN_3(JMP32, JLE, K), \
1616 INSN_3(JMP32, JSGT, K), \
1617 INSN_3(JMP32, JSLT, K), \
1618 INSN_3(JMP32, JSGE, K), \
1619 INSN_3(JMP32, JSLE, K), \
1620 INSN_3(JMP32, JSET, K), \
1621 /* Jump instructions. */ \
1622 /* Register based. */ \
1623 INSN_3(JMP, JEQ, X), \
1624 INSN_3(JMP, JNE, X), \
1625 INSN_3(JMP, JGT, X), \
1626 INSN_3(JMP, JLT, X), \
1627 INSN_3(JMP, JGE, X), \
1628 INSN_3(JMP, JLE, X), \
1629 INSN_3(JMP, JSGT, X), \
1630 INSN_3(JMP, JSLT, X), \
1631 INSN_3(JMP, JSGE, X), \
1632 INSN_3(JMP, JSLE, X), \
1633 INSN_3(JMP, JSET, X), \
1634 /* Immediate based. */ \
1635 INSN_3(JMP, JEQ, K), \
1636 INSN_3(JMP, JNE, K), \
1637 INSN_3(JMP, JGT, K), \
1638 INSN_3(JMP, JLT, K), \
1639 INSN_3(JMP, JGE, K), \
1640 INSN_3(JMP, JLE, K), \
1641 INSN_3(JMP, JSGT, K), \
1642 INSN_3(JMP, JSLT, K), \
1643 INSN_3(JMP, JSGE, K), \
1644 INSN_3(JMP, JSLE, K), \
1645 INSN_3(JMP, JSET, K), \
1646 INSN_2(JMP, JA), \
1647 INSN_2(JMP32, JA), \
1648 /* Atomic operations. */ \
1649 INSN_3(STX, ATOMIC, B), \
1650 INSN_3(STX, ATOMIC, H), \
1651 INSN_3(STX, ATOMIC, W), \
1652 INSN_3(STX, ATOMIC, DW), \
1653 /* Store instructions. */ \
1654 /* Register based. */ \
1655 INSN_3(STX, MEM, B), \
1656 INSN_3(STX, MEM, H), \
1657 INSN_3(STX, MEM, W), \
1658 INSN_3(STX, MEM, DW), \
1659 /* Immediate based. */ \
1660 INSN_3(ST, MEM, B), \
1661 INSN_3(ST, MEM, H), \
1662 INSN_3(ST, MEM, W), \
1663 INSN_3(ST, MEM, DW), \
1664 /* Load instructions. */ \
1665 /* Register based. */ \
1666 INSN_3(LDX, MEM, B), \
1667 INSN_3(LDX, MEM, H), \
1668 INSN_3(LDX, MEM, W), \
1669 INSN_3(LDX, MEM, DW), \
1670 INSN_3(LDX, MEMSX, B), \
1671 INSN_3(LDX, MEMSX, H), \
1672 INSN_3(LDX, MEMSX, W), \
1673 /* Immediate based. */ \
1674 INSN_3(LD, IMM, DW)
1675
bpf_opcode_in_insntable(u8 code)1676 bool bpf_opcode_in_insntable(u8 code)
1677 {
1678 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1679 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1680 static const bool public_insntable[256] = {
1681 [0 ... 255] = false,
1682 /* Now overwrite non-defaults ... */
1683 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1684 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1685 [BPF_LD | BPF_ABS | BPF_B] = true,
1686 [BPF_LD | BPF_ABS | BPF_H] = true,
1687 [BPF_LD | BPF_ABS | BPF_W] = true,
1688 [BPF_LD | BPF_IND | BPF_B] = true,
1689 [BPF_LD | BPF_IND | BPF_H] = true,
1690 [BPF_LD | BPF_IND | BPF_W] = true,
1691 [BPF_JMP | BPF_JCOND] = true,
1692 };
1693 #undef BPF_INSN_3_TBL
1694 #undef BPF_INSN_2_TBL
1695 return public_insntable[code];
1696 }
1697
1698 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1699 /**
1700 * ___bpf_prog_run - run eBPF program on a given context
1701 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1702 * @insn: is the array of eBPF instructions
1703 *
1704 * Decode and execute eBPF instructions.
1705 *
1706 * Return: whatever value is in %BPF_R0 at program exit
1707 */
___bpf_prog_run(u64 * regs,const struct bpf_insn * insn)1708 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1709 {
1710 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1711 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1712 static const void * const jumptable[256] __annotate_jump_table = {
1713 [0 ... 255] = &&default_label,
1714 /* Now overwrite non-defaults ... */
1715 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1716 /* Non-UAPI available opcodes. */
1717 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1718 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1719 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1720 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1721 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1722 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1723 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1724 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B,
1725 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H,
1726 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W,
1727 };
1728 #undef BPF_INSN_3_LBL
1729 #undef BPF_INSN_2_LBL
1730 u32 tail_call_cnt = 0;
1731
1732 #define CONT ({ insn++; goto select_insn; })
1733 #define CONT_JMP ({ insn++; goto select_insn; })
1734
1735 select_insn:
1736 goto *jumptable[insn->code];
1737
1738 /* Explicitly mask the register-based shift amounts with 63 or 31
1739 * to avoid undefined behavior. Normally this won't affect the
1740 * generated code, for example, in case of native 64 bit archs such
1741 * as x86-64 or arm64, the compiler is optimizing the AND away for
1742 * the interpreter. In case of JITs, each of the JIT backends compiles
1743 * the BPF shift operations to machine instructions which produce
1744 * implementation-defined results in such a case; the resulting
1745 * contents of the register may be arbitrary, but program behaviour
1746 * as a whole remains defined. In other words, in case of JIT backends,
1747 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1748 */
1749 /* ALU (shifts) */
1750 #define SHT(OPCODE, OP) \
1751 ALU64_##OPCODE##_X: \
1752 DST = DST OP (SRC & 63); \
1753 CONT; \
1754 ALU_##OPCODE##_X: \
1755 DST = (u32) DST OP ((u32) SRC & 31); \
1756 CONT; \
1757 ALU64_##OPCODE##_K: \
1758 DST = DST OP IMM; \
1759 CONT; \
1760 ALU_##OPCODE##_K: \
1761 DST = (u32) DST OP (u32) IMM; \
1762 CONT;
1763 /* ALU (rest) */
1764 #define ALU(OPCODE, OP) \
1765 ALU64_##OPCODE##_X: \
1766 DST = DST OP SRC; \
1767 CONT; \
1768 ALU_##OPCODE##_X: \
1769 DST = (u32) DST OP (u32) SRC; \
1770 CONT; \
1771 ALU64_##OPCODE##_K: \
1772 DST = DST OP IMM; \
1773 CONT; \
1774 ALU_##OPCODE##_K: \
1775 DST = (u32) DST OP (u32) IMM; \
1776 CONT;
1777 ALU(ADD, +)
1778 ALU(SUB, -)
1779 ALU(AND, &)
1780 ALU(OR, |)
1781 ALU(XOR, ^)
1782 ALU(MUL, *)
1783 SHT(LSH, <<)
1784 SHT(RSH, >>)
1785 #undef SHT
1786 #undef ALU
1787 ALU_NEG:
1788 DST = (u32) -DST;
1789 CONT;
1790 ALU64_NEG:
1791 DST = -DST;
1792 CONT;
1793 ALU_MOV_X:
1794 switch (OFF) {
1795 case 0:
1796 DST = (u32) SRC;
1797 break;
1798 case 8:
1799 DST = (u32)(s8) SRC;
1800 break;
1801 case 16:
1802 DST = (u32)(s16) SRC;
1803 break;
1804 }
1805 CONT;
1806 ALU_MOV_K:
1807 DST = (u32) IMM;
1808 CONT;
1809 ALU64_MOV_X:
1810 switch (OFF) {
1811 case 0:
1812 DST = SRC;
1813 break;
1814 case 8:
1815 DST = (s8) SRC;
1816 break;
1817 case 16:
1818 DST = (s16) SRC;
1819 break;
1820 case 32:
1821 DST = (s32) SRC;
1822 break;
1823 }
1824 CONT;
1825 ALU64_MOV_K:
1826 DST = IMM;
1827 CONT;
1828 LD_IMM_DW:
1829 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1830 insn++;
1831 CONT;
1832 ALU_ARSH_X:
1833 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1834 CONT;
1835 ALU_ARSH_K:
1836 DST = (u64) (u32) (((s32) DST) >> IMM);
1837 CONT;
1838 ALU64_ARSH_X:
1839 (*(s64 *) &DST) >>= (SRC & 63);
1840 CONT;
1841 ALU64_ARSH_K:
1842 (*(s64 *) &DST) >>= IMM;
1843 CONT;
1844 ALU64_MOD_X:
1845 switch (OFF) {
1846 case 0:
1847 div64_u64_rem(DST, SRC, &AX);
1848 DST = AX;
1849 break;
1850 case 1:
1851 AX = div64_s64(DST, SRC);
1852 DST = DST - AX * SRC;
1853 break;
1854 }
1855 CONT;
1856 ALU_MOD_X:
1857 switch (OFF) {
1858 case 0:
1859 AX = (u32) DST;
1860 DST = do_div(AX, (u32) SRC);
1861 break;
1862 case 1:
1863 AX = abs((s32)DST);
1864 AX = do_div(AX, abs((s32)SRC));
1865 if ((s32)DST < 0)
1866 DST = (u32)-AX;
1867 else
1868 DST = (u32)AX;
1869 break;
1870 }
1871 CONT;
1872 ALU64_MOD_K:
1873 switch (OFF) {
1874 case 0:
1875 div64_u64_rem(DST, IMM, &AX);
1876 DST = AX;
1877 break;
1878 case 1:
1879 AX = div64_s64(DST, IMM);
1880 DST = DST - AX * IMM;
1881 break;
1882 }
1883 CONT;
1884 ALU_MOD_K:
1885 switch (OFF) {
1886 case 0:
1887 AX = (u32) DST;
1888 DST = do_div(AX, (u32) IMM);
1889 break;
1890 case 1:
1891 AX = abs((s32)DST);
1892 AX = do_div(AX, abs((s32)IMM));
1893 if ((s32)DST < 0)
1894 DST = (u32)-AX;
1895 else
1896 DST = (u32)AX;
1897 break;
1898 }
1899 CONT;
1900 ALU64_DIV_X:
1901 switch (OFF) {
1902 case 0:
1903 DST = div64_u64(DST, SRC);
1904 break;
1905 case 1:
1906 DST = div64_s64(DST, SRC);
1907 break;
1908 }
1909 CONT;
1910 ALU_DIV_X:
1911 switch (OFF) {
1912 case 0:
1913 AX = (u32) DST;
1914 do_div(AX, (u32) SRC);
1915 DST = (u32) AX;
1916 break;
1917 case 1:
1918 AX = abs((s32)DST);
1919 do_div(AX, abs((s32)SRC));
1920 if (((s32)DST < 0) == ((s32)SRC < 0))
1921 DST = (u32)AX;
1922 else
1923 DST = (u32)-AX;
1924 break;
1925 }
1926 CONT;
1927 ALU64_DIV_K:
1928 switch (OFF) {
1929 case 0:
1930 DST = div64_u64(DST, IMM);
1931 break;
1932 case 1:
1933 DST = div64_s64(DST, IMM);
1934 break;
1935 }
1936 CONT;
1937 ALU_DIV_K:
1938 switch (OFF) {
1939 case 0:
1940 AX = (u32) DST;
1941 do_div(AX, (u32) IMM);
1942 DST = (u32) AX;
1943 break;
1944 case 1:
1945 AX = abs((s32)DST);
1946 do_div(AX, abs((s32)IMM));
1947 if (((s32)DST < 0) == ((s32)IMM < 0))
1948 DST = (u32)AX;
1949 else
1950 DST = (u32)-AX;
1951 break;
1952 }
1953 CONT;
1954 ALU_END_TO_BE:
1955 switch (IMM) {
1956 case 16:
1957 DST = (__force u16) cpu_to_be16(DST);
1958 break;
1959 case 32:
1960 DST = (__force u32) cpu_to_be32(DST);
1961 break;
1962 case 64:
1963 DST = (__force u64) cpu_to_be64(DST);
1964 break;
1965 }
1966 CONT;
1967 ALU_END_TO_LE:
1968 switch (IMM) {
1969 case 16:
1970 DST = (__force u16) cpu_to_le16(DST);
1971 break;
1972 case 32:
1973 DST = (__force u32) cpu_to_le32(DST);
1974 break;
1975 case 64:
1976 DST = (__force u64) cpu_to_le64(DST);
1977 break;
1978 }
1979 CONT;
1980 ALU64_END_TO_LE:
1981 switch (IMM) {
1982 case 16:
1983 DST = (__force u16) __swab16(DST);
1984 break;
1985 case 32:
1986 DST = (__force u32) __swab32(DST);
1987 break;
1988 case 64:
1989 DST = (__force u64) __swab64(DST);
1990 break;
1991 }
1992 CONT;
1993
1994 /* CALL */
1995 JMP_CALL:
1996 /* Function call scratches BPF_R1-BPF_R5 registers,
1997 * preserves BPF_R6-BPF_R9, and stores return value
1998 * into BPF_R0.
1999 */
2000 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
2001 BPF_R4, BPF_R5);
2002 CONT;
2003
2004 JMP_CALL_ARGS:
2005 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
2006 BPF_R3, BPF_R4,
2007 BPF_R5,
2008 insn + insn->off + 1);
2009 CONT;
2010
2011 JMP_TAIL_CALL: {
2012 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
2013 struct bpf_array *array = container_of(map, struct bpf_array, map);
2014 struct bpf_prog *prog;
2015 u32 index = BPF_R3;
2016
2017 if (unlikely(index >= array->map.max_entries))
2018 goto out;
2019
2020 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
2021 goto out;
2022
2023 tail_call_cnt++;
2024
2025 prog = READ_ONCE(array->ptrs[index]);
2026 if (!prog)
2027 goto out;
2028
2029 /* ARG1 at this point is guaranteed to point to CTX from
2030 * the verifier side due to the fact that the tail call is
2031 * handled like a helper, that is, bpf_tail_call_proto,
2032 * where arg1_type is ARG_PTR_TO_CTX.
2033 */
2034 insn = prog->insnsi;
2035 goto select_insn;
2036 out:
2037 CONT;
2038 }
2039 JMP_JA:
2040 insn += insn->off;
2041 CONT;
2042 JMP32_JA:
2043 insn += insn->imm;
2044 CONT;
2045 JMP_EXIT:
2046 return BPF_R0;
2047 /* JMP */
2048 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
2049 JMP_##OPCODE##_X: \
2050 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
2051 insn += insn->off; \
2052 CONT_JMP; \
2053 } \
2054 CONT; \
2055 JMP32_##OPCODE##_X: \
2056 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
2057 insn += insn->off; \
2058 CONT_JMP; \
2059 } \
2060 CONT; \
2061 JMP_##OPCODE##_K: \
2062 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
2063 insn += insn->off; \
2064 CONT_JMP; \
2065 } \
2066 CONT; \
2067 JMP32_##OPCODE##_K: \
2068 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
2069 insn += insn->off; \
2070 CONT_JMP; \
2071 } \
2072 CONT;
2073 COND_JMP(u, JEQ, ==)
2074 COND_JMP(u, JNE, !=)
2075 COND_JMP(u, JGT, >)
2076 COND_JMP(u, JLT, <)
2077 COND_JMP(u, JGE, >=)
2078 COND_JMP(u, JLE, <=)
2079 COND_JMP(u, JSET, &)
2080 COND_JMP(s, JSGT, >)
2081 COND_JMP(s, JSLT, <)
2082 COND_JMP(s, JSGE, >=)
2083 COND_JMP(s, JSLE, <=)
2084 #undef COND_JMP
2085 /* ST, STX and LDX*/
2086 ST_NOSPEC:
2087 /* Speculation barrier for mitigating Speculative Store Bypass,
2088 * Bounds-Check Bypass and Type Confusion. In case of arm64, we
2089 * rely on the firmware mitigation as controlled via the ssbd
2090 * kernel parameter. Whenever the mitigation is enabled, it
2091 * works for all of the kernel code with no need to provide any
2092 * additional instructions here. In case of x86, we use 'lfence'
2093 * insn for mitigation. We reuse preexisting logic from Spectre
2094 * v1 mitigation that happens to produce the required code on
2095 * x86 for v4 as well.
2096 */
2097 barrier_nospec();
2098 CONT;
2099 #define LDST(SIZEOP, SIZE) \
2100 STX_MEM_##SIZEOP: \
2101 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
2102 CONT; \
2103 ST_MEM_##SIZEOP: \
2104 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
2105 CONT; \
2106 LDX_MEM_##SIZEOP: \
2107 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
2108 CONT; \
2109 LDX_PROBE_MEM_##SIZEOP: \
2110 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \
2111 (const void *)(long) (SRC + insn->off)); \
2112 DST = *((SIZE *)&DST); \
2113 CONT;
2114
2115 LDST(B, u8)
2116 LDST(H, u16)
2117 LDST(W, u32)
2118 LDST(DW, u64)
2119 #undef LDST
2120
2121 #define LDSX(SIZEOP, SIZE) \
2122 LDX_MEMSX_##SIZEOP: \
2123 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
2124 CONT; \
2125 LDX_PROBE_MEMSX_##SIZEOP: \
2126 bpf_probe_read_kernel_common(&DST, sizeof(SIZE), \
2127 (const void *)(long) (SRC + insn->off)); \
2128 DST = *((SIZE *)&DST); \
2129 CONT;
2130
2131 LDSX(B, s8)
2132 LDSX(H, s16)
2133 LDSX(W, s32)
2134 #undef LDSX
2135
2136 #define ATOMIC_ALU_OP(BOP, KOP) \
2137 case BOP: \
2138 if (BPF_SIZE(insn->code) == BPF_W) \
2139 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
2140 (DST + insn->off)); \
2141 else if (BPF_SIZE(insn->code) == BPF_DW) \
2142 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
2143 (DST + insn->off)); \
2144 else \
2145 goto default_label; \
2146 break; \
2147 case BOP | BPF_FETCH: \
2148 if (BPF_SIZE(insn->code) == BPF_W) \
2149 SRC = (u32) atomic_fetch_##KOP( \
2150 (u32) SRC, \
2151 (atomic_t *)(unsigned long) (DST + insn->off)); \
2152 else if (BPF_SIZE(insn->code) == BPF_DW) \
2153 SRC = (u64) atomic64_fetch_##KOP( \
2154 (u64) SRC, \
2155 (atomic64_t *)(unsigned long) (DST + insn->off)); \
2156 else \
2157 goto default_label; \
2158 break;
2159
2160 STX_ATOMIC_DW:
2161 STX_ATOMIC_W:
2162 STX_ATOMIC_H:
2163 STX_ATOMIC_B:
2164 switch (IMM) {
2165 /* Atomic read-modify-write instructions support only W and DW
2166 * size modifiers.
2167 */
2168 ATOMIC_ALU_OP(BPF_ADD, add)
2169 ATOMIC_ALU_OP(BPF_AND, and)
2170 ATOMIC_ALU_OP(BPF_OR, or)
2171 ATOMIC_ALU_OP(BPF_XOR, xor)
2172 #undef ATOMIC_ALU_OP
2173
2174 case BPF_XCHG:
2175 if (BPF_SIZE(insn->code) == BPF_W)
2176 SRC = (u32) atomic_xchg(
2177 (atomic_t *)(unsigned long) (DST + insn->off),
2178 (u32) SRC);
2179 else if (BPF_SIZE(insn->code) == BPF_DW)
2180 SRC = (u64) atomic64_xchg(
2181 (atomic64_t *)(unsigned long) (DST + insn->off),
2182 (u64) SRC);
2183 else
2184 goto default_label;
2185 break;
2186 case BPF_CMPXCHG:
2187 if (BPF_SIZE(insn->code) == BPF_W)
2188 BPF_R0 = (u32) atomic_cmpxchg(
2189 (atomic_t *)(unsigned long) (DST + insn->off),
2190 (u32) BPF_R0, (u32) SRC);
2191 else if (BPF_SIZE(insn->code) == BPF_DW)
2192 BPF_R0 = (u64) atomic64_cmpxchg(
2193 (atomic64_t *)(unsigned long) (DST + insn->off),
2194 (u64) BPF_R0, (u64) SRC);
2195 else
2196 goto default_label;
2197 break;
2198 /* Atomic load and store instructions support all size
2199 * modifiers.
2200 */
2201 case BPF_LOAD_ACQ:
2202 switch (BPF_SIZE(insn->code)) {
2203 #define LOAD_ACQUIRE(SIZEOP, SIZE) \
2204 case BPF_##SIZEOP: \
2205 DST = (SIZE)smp_load_acquire( \
2206 (SIZE *)(unsigned long)(SRC + insn->off)); \
2207 break;
2208 LOAD_ACQUIRE(B, u8)
2209 LOAD_ACQUIRE(H, u16)
2210 LOAD_ACQUIRE(W, u32)
2211 #ifdef CONFIG_64BIT
2212 LOAD_ACQUIRE(DW, u64)
2213 #endif
2214 #undef LOAD_ACQUIRE
2215 default:
2216 goto default_label;
2217 }
2218 break;
2219 case BPF_STORE_REL:
2220 switch (BPF_SIZE(insn->code)) {
2221 #define STORE_RELEASE(SIZEOP, SIZE) \
2222 case BPF_##SIZEOP: \
2223 smp_store_release( \
2224 (SIZE *)(unsigned long)(DST + insn->off), (SIZE)SRC); \
2225 break;
2226 STORE_RELEASE(B, u8)
2227 STORE_RELEASE(H, u16)
2228 STORE_RELEASE(W, u32)
2229 #ifdef CONFIG_64BIT
2230 STORE_RELEASE(DW, u64)
2231 #endif
2232 #undef STORE_RELEASE
2233 default:
2234 goto default_label;
2235 }
2236 break;
2237
2238 default:
2239 goto default_label;
2240 }
2241 CONT;
2242
2243 default_label:
2244 /* If we ever reach this, we have a bug somewhere. Die hard here
2245 * instead of just returning 0; we could be somewhere in a subprog,
2246 * so execution could continue otherwise which we do /not/ want.
2247 *
2248 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
2249 */
2250 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
2251 insn->code, insn->imm);
2252 BUG_ON(1);
2253 return 0;
2254 }
2255
2256 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2257 #define DEFINE_BPF_PROG_RUN(stack_size) \
2258 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2259 { \
2260 u64 stack[stack_size / sizeof(u64)]; \
2261 u64 regs[MAX_BPF_EXT_REG] = {}; \
2262 \
2263 kmsan_unpoison_memory(stack, sizeof(stack)); \
2264 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2265 ARG1 = (u64) (unsigned long) ctx; \
2266 return ___bpf_prog_run(regs, insn); \
2267 }
2268
2269 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2270 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2271 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2272 const struct bpf_insn *insn) \
2273 { \
2274 u64 stack[stack_size / sizeof(u64)]; \
2275 u64 regs[MAX_BPF_EXT_REG]; \
2276 \
2277 kmsan_unpoison_memory(stack, sizeof(stack)); \
2278 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2279 BPF_R1 = r1; \
2280 BPF_R2 = r2; \
2281 BPF_R3 = r3; \
2282 BPF_R4 = r4; \
2283 BPF_R5 = r5; \
2284 return ___bpf_prog_run(regs, insn); \
2285 }
2286
2287 #define EVAL1(FN, X) FN(X)
2288 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2289 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2290 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2291 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2292 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2293
2294 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2295 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2296 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2297
2298 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2299 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2300 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2301
2302 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2303
2304 static unsigned int (*interpreters[])(const void *ctx,
2305 const struct bpf_insn *insn) = {
2306 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2307 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2308 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2309 };
2310 #undef PROG_NAME_LIST
2311 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2312 static __maybe_unused
2313 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2314 const struct bpf_insn *insn) = {
2315 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2316 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2317 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2318 };
2319 #undef PROG_NAME_LIST
2320
2321 #ifdef CONFIG_BPF_SYSCALL
bpf_patch_call_args(struct bpf_insn * insn,u32 stack_depth)2322 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2323 {
2324 stack_depth = max_t(u32, stack_depth, 1);
2325 insn->off = (s16) insn->imm;
2326 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2327 __bpf_call_base_args;
2328 insn->code = BPF_JMP | BPF_CALL_ARGS;
2329 }
2330 #endif
2331 #endif
2332
__bpf_prog_ret0_warn(const void * ctx,const struct bpf_insn * insn)2333 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2334 const struct bpf_insn *insn)
2335 {
2336 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2337 * is not working properly, so warn about it!
2338 */
2339 WARN_ON_ONCE(1);
2340 return 0;
2341 }
2342
__bpf_prog_map_compatible(struct bpf_map * map,const struct bpf_prog * fp)2343 static bool __bpf_prog_map_compatible(struct bpf_map *map,
2344 const struct bpf_prog *fp)
2345 {
2346 enum bpf_prog_type prog_type = resolve_prog_type(fp);
2347 struct bpf_prog_aux *aux = fp->aux;
2348 enum bpf_cgroup_storage_type i;
2349 bool ret = false;
2350 u64 cookie;
2351
2352 if (fp->kprobe_override)
2353 return ret;
2354
2355 spin_lock(&map->owner_lock);
2356 /* There's no owner yet where we could check for compatibility. */
2357 if (!map->owner) {
2358 map->owner = bpf_map_owner_alloc(map);
2359 if (!map->owner)
2360 goto err;
2361 map->owner->type = prog_type;
2362 map->owner->jited = fp->jited;
2363 map->owner->xdp_has_frags = aux->xdp_has_frags;
2364 map->owner->expected_attach_type = fp->expected_attach_type;
2365 map->owner->attach_func_proto = aux->attach_func_proto;
2366 for_each_cgroup_storage_type(i) {
2367 map->owner->storage_cookie[i] =
2368 aux->cgroup_storage[i] ?
2369 aux->cgroup_storage[i]->cookie : 0;
2370 }
2371 ret = true;
2372 } else {
2373 ret = map->owner->type == prog_type &&
2374 map->owner->jited == fp->jited &&
2375 map->owner->xdp_has_frags == aux->xdp_has_frags;
2376 if (ret &&
2377 map->map_type == BPF_MAP_TYPE_PROG_ARRAY &&
2378 map->owner->expected_attach_type != fp->expected_attach_type)
2379 ret = false;
2380 for_each_cgroup_storage_type(i) {
2381 if (!ret)
2382 break;
2383 cookie = aux->cgroup_storage[i] ?
2384 aux->cgroup_storage[i]->cookie : 0;
2385 ret = map->owner->storage_cookie[i] == cookie ||
2386 !cookie;
2387 }
2388 if (ret &&
2389 map->owner->attach_func_proto != aux->attach_func_proto) {
2390 switch (prog_type) {
2391 case BPF_PROG_TYPE_TRACING:
2392 case BPF_PROG_TYPE_LSM:
2393 case BPF_PROG_TYPE_EXT:
2394 case BPF_PROG_TYPE_STRUCT_OPS:
2395 ret = false;
2396 break;
2397 default:
2398 break;
2399 }
2400 }
2401 }
2402 err:
2403 spin_unlock(&map->owner_lock);
2404 return ret;
2405 }
2406
bpf_prog_map_compatible(struct bpf_map * map,const struct bpf_prog * fp)2407 bool bpf_prog_map_compatible(struct bpf_map *map, const struct bpf_prog *fp)
2408 {
2409 /* XDP programs inserted into maps are not guaranteed to run on
2410 * a particular netdev (and can run outside driver context entirely
2411 * in the case of devmap and cpumap). Until device checks
2412 * are implemented, prohibit adding dev-bound programs to program maps.
2413 */
2414 if (bpf_prog_is_dev_bound(fp->aux))
2415 return false;
2416
2417 return __bpf_prog_map_compatible(map, fp);
2418 }
2419
bpf_check_tail_call(const struct bpf_prog * fp)2420 static int bpf_check_tail_call(const struct bpf_prog *fp)
2421 {
2422 struct bpf_prog_aux *aux = fp->aux;
2423 int i, ret = 0;
2424
2425 mutex_lock(&aux->used_maps_mutex);
2426 for (i = 0; i < aux->used_map_cnt; i++) {
2427 struct bpf_map *map = aux->used_maps[i];
2428
2429 if (!map_type_contains_progs(map))
2430 continue;
2431
2432 if (!__bpf_prog_map_compatible(map, fp)) {
2433 ret = -EINVAL;
2434 goto out;
2435 }
2436 }
2437
2438 out:
2439 mutex_unlock(&aux->used_maps_mutex);
2440 return ret;
2441 }
2442
bpf_prog_select_interpreter(struct bpf_prog * fp)2443 static bool bpf_prog_select_interpreter(struct bpf_prog *fp)
2444 {
2445 bool select_interpreter = false;
2446 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2447 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2448 u32 idx = (round_up(stack_depth, 32) / 32) - 1;
2449
2450 /* may_goto may cause stack size > 512, leading to idx out-of-bounds.
2451 * But for non-JITed programs, we don't need bpf_func, so no bounds
2452 * check needed.
2453 */
2454 if (idx < ARRAY_SIZE(interpreters)) {
2455 fp->bpf_func = interpreters[idx];
2456 select_interpreter = true;
2457 } else {
2458 fp->bpf_func = __bpf_prog_ret0_warn;
2459 }
2460 #else
2461 fp->bpf_func = __bpf_prog_ret0_warn;
2462 #endif
2463 return select_interpreter;
2464 }
2465
2466 /**
2467 * bpf_prog_select_runtime - select exec runtime for BPF program
2468 * @fp: bpf_prog populated with BPF program
2469 * @err: pointer to error variable
2470 *
2471 * Try to JIT eBPF program, if JIT is not available, use interpreter.
2472 * The BPF program will be executed via bpf_prog_run() function.
2473 *
2474 * Return: the &fp argument along with &err set to 0 for success or
2475 * a negative errno code on failure
2476 */
bpf_prog_select_runtime(struct bpf_prog * fp,int * err)2477 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2478 {
2479 /* In case of BPF to BPF calls, verifier did all the prep
2480 * work with regards to JITing, etc.
2481 */
2482 bool jit_needed = false;
2483
2484 if (fp->bpf_func)
2485 goto finalize;
2486
2487 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2488 bpf_prog_has_kfunc_call(fp))
2489 jit_needed = true;
2490
2491 if (!bpf_prog_select_interpreter(fp))
2492 jit_needed = true;
2493
2494 /* eBPF JITs can rewrite the program in case constant
2495 * blinding is active. However, in case of error during
2496 * blinding, bpf_int_jit_compile() must always return a
2497 * valid program, which in this case would simply not
2498 * be JITed, but falls back to the interpreter.
2499 */
2500 if (!bpf_prog_is_offloaded(fp->aux)) {
2501 *err = bpf_prog_alloc_jited_linfo(fp);
2502 if (*err)
2503 return fp;
2504
2505 fp = bpf_int_jit_compile(fp);
2506 bpf_prog_jit_attempt_done(fp);
2507 if (!fp->jited && jit_needed) {
2508 *err = -ENOTSUPP;
2509 return fp;
2510 }
2511 } else {
2512 *err = bpf_prog_offload_compile(fp);
2513 if (*err)
2514 return fp;
2515 }
2516
2517 finalize:
2518 *err = bpf_prog_lock_ro(fp);
2519 if (*err)
2520 return fp;
2521
2522 /* The tail call compatibility check can only be done at
2523 * this late stage as we need to determine, if we deal
2524 * with JITed or non JITed program concatenations and not
2525 * all eBPF JITs might immediately support all features.
2526 */
2527 *err = bpf_check_tail_call(fp);
2528
2529 return fp;
2530 }
2531 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2532
__bpf_prog_ret1(const void * ctx,const struct bpf_insn * insn)2533 static unsigned int __bpf_prog_ret1(const void *ctx,
2534 const struct bpf_insn *insn)
2535 {
2536 return 1;
2537 }
2538
2539 static struct bpf_prog_dummy {
2540 struct bpf_prog prog;
2541 } dummy_bpf_prog = {
2542 .prog = {
2543 .bpf_func = __bpf_prog_ret1,
2544 },
2545 };
2546
2547 struct bpf_empty_prog_array bpf_empty_prog_array = {
2548 .null_prog = NULL,
2549 };
2550 EXPORT_SYMBOL(bpf_empty_prog_array);
2551
bpf_prog_array_alloc(u32 prog_cnt,gfp_t flags)2552 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2553 {
2554 struct bpf_prog_array *p;
2555
2556 if (prog_cnt)
2557 p = kzalloc(struct_size(p, items, prog_cnt + 1), flags);
2558 else
2559 p = &bpf_empty_prog_array.hdr;
2560
2561 return p;
2562 }
2563
bpf_prog_array_free(struct bpf_prog_array * progs)2564 void bpf_prog_array_free(struct bpf_prog_array *progs)
2565 {
2566 if (!progs || progs == &bpf_empty_prog_array.hdr)
2567 return;
2568 kfree_rcu(progs, rcu);
2569 }
2570
__bpf_prog_array_free_sleepable_cb(struct rcu_head * rcu)2571 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2572 {
2573 struct bpf_prog_array *progs;
2574
2575 /* If RCU Tasks Trace grace period implies RCU grace period, there is
2576 * no need to call kfree_rcu(), just call kfree() directly.
2577 */
2578 progs = container_of(rcu, struct bpf_prog_array, rcu);
2579 if (rcu_trace_implies_rcu_gp())
2580 kfree(progs);
2581 else
2582 kfree_rcu(progs, rcu);
2583 }
2584
bpf_prog_array_free_sleepable(struct bpf_prog_array * progs)2585 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2586 {
2587 if (!progs || progs == &bpf_empty_prog_array.hdr)
2588 return;
2589 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2590 }
2591
bpf_prog_array_length(struct bpf_prog_array * array)2592 int bpf_prog_array_length(struct bpf_prog_array *array)
2593 {
2594 struct bpf_prog_array_item *item;
2595 u32 cnt = 0;
2596
2597 for (item = array->items; item->prog; item++)
2598 if (item->prog != &dummy_bpf_prog.prog)
2599 cnt++;
2600 return cnt;
2601 }
2602
bpf_prog_array_is_empty(struct bpf_prog_array * array)2603 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2604 {
2605 struct bpf_prog_array_item *item;
2606
2607 for (item = array->items; item->prog; item++)
2608 if (item->prog != &dummy_bpf_prog.prog)
2609 return false;
2610 return true;
2611 }
2612
bpf_prog_array_copy_core(struct bpf_prog_array * array,u32 * prog_ids,u32 request_cnt)2613 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2614 u32 *prog_ids,
2615 u32 request_cnt)
2616 {
2617 struct bpf_prog_array_item *item;
2618 int i = 0;
2619
2620 for (item = array->items; item->prog; item++) {
2621 if (item->prog == &dummy_bpf_prog.prog)
2622 continue;
2623 prog_ids[i] = item->prog->aux->id;
2624 if (++i == request_cnt) {
2625 item++;
2626 break;
2627 }
2628 }
2629
2630 return !!(item->prog);
2631 }
2632
bpf_prog_array_copy_to_user(struct bpf_prog_array * array,__u32 __user * prog_ids,u32 cnt)2633 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2634 __u32 __user *prog_ids, u32 cnt)
2635 {
2636 unsigned long err = 0;
2637 bool nospc;
2638 u32 *ids;
2639
2640 /* users of this function are doing:
2641 * cnt = bpf_prog_array_length();
2642 * if (cnt > 0)
2643 * bpf_prog_array_copy_to_user(..., cnt);
2644 * so below kcalloc doesn't need extra cnt > 0 check.
2645 */
2646 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2647 if (!ids)
2648 return -ENOMEM;
2649 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2650 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2651 kfree(ids);
2652 if (err)
2653 return -EFAULT;
2654 if (nospc)
2655 return -ENOSPC;
2656 return 0;
2657 }
2658
bpf_prog_array_delete_safe(struct bpf_prog_array * array,struct bpf_prog * old_prog)2659 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2660 struct bpf_prog *old_prog)
2661 {
2662 struct bpf_prog_array_item *item;
2663
2664 for (item = array->items; item->prog; item++)
2665 if (item->prog == old_prog) {
2666 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2667 break;
2668 }
2669 }
2670
2671 /**
2672 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2673 * index into the program array with
2674 * a dummy no-op program.
2675 * @array: a bpf_prog_array
2676 * @index: the index of the program to replace
2677 *
2678 * Skips over dummy programs, by not counting them, when calculating
2679 * the position of the program to replace.
2680 *
2681 * Return:
2682 * * 0 - Success
2683 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2684 * * -ENOENT - Index out of range
2685 */
bpf_prog_array_delete_safe_at(struct bpf_prog_array * array,int index)2686 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2687 {
2688 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2689 }
2690
2691 /**
2692 * bpf_prog_array_update_at() - Updates the program at the given index
2693 * into the program array.
2694 * @array: a bpf_prog_array
2695 * @index: the index of the program to update
2696 * @prog: the program to insert into the array
2697 *
2698 * Skips over dummy programs, by not counting them, when calculating
2699 * the position of the program to update.
2700 *
2701 * Return:
2702 * * 0 - Success
2703 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2704 * * -ENOENT - Index out of range
2705 */
bpf_prog_array_update_at(struct bpf_prog_array * array,int index,struct bpf_prog * prog)2706 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2707 struct bpf_prog *prog)
2708 {
2709 struct bpf_prog_array_item *item;
2710
2711 if (unlikely(index < 0))
2712 return -EINVAL;
2713
2714 for (item = array->items; item->prog; item++) {
2715 if (item->prog == &dummy_bpf_prog.prog)
2716 continue;
2717 if (!index) {
2718 WRITE_ONCE(item->prog, prog);
2719 return 0;
2720 }
2721 index--;
2722 }
2723 return -ENOENT;
2724 }
2725
bpf_prog_array_copy(struct bpf_prog_array * old_array,struct bpf_prog * exclude_prog,struct bpf_prog * include_prog,u64 bpf_cookie,struct bpf_prog_array ** new_array)2726 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2727 struct bpf_prog *exclude_prog,
2728 struct bpf_prog *include_prog,
2729 u64 bpf_cookie,
2730 struct bpf_prog_array **new_array)
2731 {
2732 int new_prog_cnt, carry_prog_cnt = 0;
2733 struct bpf_prog_array_item *existing, *new;
2734 struct bpf_prog_array *array;
2735 bool found_exclude = false;
2736
2737 /* Figure out how many existing progs we need to carry over to
2738 * the new array.
2739 */
2740 if (old_array) {
2741 existing = old_array->items;
2742 for (; existing->prog; existing++) {
2743 if (existing->prog == exclude_prog) {
2744 found_exclude = true;
2745 continue;
2746 }
2747 if (existing->prog != &dummy_bpf_prog.prog)
2748 carry_prog_cnt++;
2749 if (existing->prog == include_prog)
2750 return -EEXIST;
2751 }
2752 }
2753
2754 if (exclude_prog && !found_exclude)
2755 return -ENOENT;
2756
2757 /* How many progs (not NULL) will be in the new array? */
2758 new_prog_cnt = carry_prog_cnt;
2759 if (include_prog)
2760 new_prog_cnt += 1;
2761
2762 /* Do we have any prog (not NULL) in the new array? */
2763 if (!new_prog_cnt) {
2764 *new_array = NULL;
2765 return 0;
2766 }
2767
2768 /* +1 as the end of prog_array is marked with NULL */
2769 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2770 if (!array)
2771 return -ENOMEM;
2772 new = array->items;
2773
2774 /* Fill in the new prog array */
2775 if (carry_prog_cnt) {
2776 existing = old_array->items;
2777 for (; existing->prog; existing++) {
2778 if (existing->prog == exclude_prog ||
2779 existing->prog == &dummy_bpf_prog.prog)
2780 continue;
2781
2782 new->prog = existing->prog;
2783 new->bpf_cookie = existing->bpf_cookie;
2784 new++;
2785 }
2786 }
2787 if (include_prog) {
2788 new->prog = include_prog;
2789 new->bpf_cookie = bpf_cookie;
2790 new++;
2791 }
2792 new->prog = NULL;
2793 *new_array = array;
2794 return 0;
2795 }
2796
bpf_prog_array_copy_info(struct bpf_prog_array * array,u32 * prog_ids,u32 request_cnt,u32 * prog_cnt)2797 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2798 u32 *prog_ids, u32 request_cnt,
2799 u32 *prog_cnt)
2800 {
2801 u32 cnt = 0;
2802
2803 if (array)
2804 cnt = bpf_prog_array_length(array);
2805
2806 *prog_cnt = cnt;
2807
2808 /* return early if user requested only program count or nothing to copy */
2809 if (!request_cnt || !cnt)
2810 return 0;
2811
2812 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2813 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2814 : 0;
2815 }
2816
__bpf_free_used_maps(struct bpf_prog_aux * aux,struct bpf_map ** used_maps,u32 len)2817 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2818 struct bpf_map **used_maps, u32 len)
2819 {
2820 struct bpf_map *map;
2821 bool sleepable;
2822 u32 i;
2823
2824 sleepable = aux->prog->sleepable;
2825 for (i = 0; i < len; i++) {
2826 map = used_maps[i];
2827 if (map->ops->map_poke_untrack)
2828 map->ops->map_poke_untrack(map, aux);
2829 if (sleepable)
2830 atomic64_dec(&map->sleepable_refcnt);
2831 bpf_map_put(map);
2832 }
2833 }
2834
bpf_free_used_maps(struct bpf_prog_aux * aux)2835 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2836 {
2837 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2838 kfree(aux->used_maps);
2839 }
2840
__bpf_free_used_btfs(struct btf_mod_pair * used_btfs,u32 len)2841 void __bpf_free_used_btfs(struct btf_mod_pair *used_btfs, u32 len)
2842 {
2843 #ifdef CONFIG_BPF_SYSCALL
2844 struct btf_mod_pair *btf_mod;
2845 u32 i;
2846
2847 for (i = 0; i < len; i++) {
2848 btf_mod = &used_btfs[i];
2849 if (btf_mod->module)
2850 module_put(btf_mod->module);
2851 btf_put(btf_mod->btf);
2852 }
2853 #endif
2854 }
2855
bpf_free_used_btfs(struct bpf_prog_aux * aux)2856 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2857 {
2858 __bpf_free_used_btfs(aux->used_btfs, aux->used_btf_cnt);
2859 kfree(aux->used_btfs);
2860 }
2861
bpf_prog_free_deferred(struct work_struct * work)2862 static void bpf_prog_free_deferred(struct work_struct *work)
2863 {
2864 struct bpf_prog_aux *aux;
2865 int i;
2866
2867 aux = container_of(work, struct bpf_prog_aux, work);
2868 #ifdef CONFIG_BPF_SYSCALL
2869 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2870 bpf_prog_stream_free(aux->prog);
2871 #endif
2872 #ifdef CONFIG_CGROUP_BPF
2873 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2874 bpf_cgroup_atype_put(aux->cgroup_atype);
2875 #endif
2876 bpf_free_used_maps(aux);
2877 bpf_free_used_btfs(aux);
2878 if (bpf_prog_is_dev_bound(aux))
2879 bpf_prog_dev_bound_destroy(aux->prog);
2880 #ifdef CONFIG_PERF_EVENTS
2881 if (aux->prog->has_callchain_buf)
2882 put_callchain_buffers();
2883 #endif
2884 if (aux->dst_trampoline)
2885 bpf_trampoline_put(aux->dst_trampoline);
2886 for (i = 0; i < aux->real_func_cnt; i++) {
2887 /* We can just unlink the subprog poke descriptor table as
2888 * it was originally linked to the main program and is also
2889 * released along with it.
2890 */
2891 aux->func[i]->aux->poke_tab = NULL;
2892 bpf_jit_free(aux->func[i]);
2893 }
2894 if (aux->real_func_cnt) {
2895 kfree(aux->func);
2896 bpf_prog_unlock_free(aux->prog);
2897 } else {
2898 bpf_jit_free(aux->prog);
2899 }
2900 }
2901
bpf_prog_free(struct bpf_prog * fp)2902 void bpf_prog_free(struct bpf_prog *fp)
2903 {
2904 struct bpf_prog_aux *aux = fp->aux;
2905
2906 if (aux->dst_prog)
2907 bpf_prog_put(aux->dst_prog);
2908 bpf_token_put(aux->token);
2909 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2910 schedule_work(&aux->work);
2911 }
2912 EXPORT_SYMBOL_GPL(bpf_prog_free);
2913
2914 /* RNG for unprivileged user space with separated state from prandom_u32(). */
2915 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2916
bpf_user_rnd_init_once(void)2917 void bpf_user_rnd_init_once(void)
2918 {
2919 prandom_init_once(&bpf_user_rnd_state);
2920 }
2921
BPF_CALL_0(bpf_user_rnd_u32)2922 BPF_CALL_0(bpf_user_rnd_u32)
2923 {
2924 /* Should someone ever have the rather unwise idea to use some
2925 * of the registers passed into this function, then note that
2926 * this function is called from native eBPF and classic-to-eBPF
2927 * transformations. Register assignments from both sides are
2928 * different, f.e. classic always sets fn(ctx, A, X) here.
2929 */
2930 struct rnd_state *state;
2931 u32 res;
2932
2933 state = &get_cpu_var(bpf_user_rnd_state);
2934 res = prandom_u32_state(state);
2935 put_cpu_var(bpf_user_rnd_state);
2936
2937 return res;
2938 }
2939
BPF_CALL_0(bpf_get_raw_cpu_id)2940 BPF_CALL_0(bpf_get_raw_cpu_id)
2941 {
2942 return raw_smp_processor_id();
2943 }
2944
2945 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2946 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2947 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2948 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2949 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2950 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2951 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2952 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2953 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2954 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2955 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2956
2957 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2958 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2959 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2960 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2961 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2962 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2963 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak;
2964
2965 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2966 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2967 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2968 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2969 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2970 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2971 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2972 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2973 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2974 const struct bpf_func_proto bpf_set_retval_proto __weak;
2975 const struct bpf_func_proto bpf_get_retval_proto __weak;
2976
bpf_get_trace_printk_proto(void)2977 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2978 {
2979 return NULL;
2980 }
2981
bpf_get_trace_vprintk_proto(void)2982 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2983 {
2984 return NULL;
2985 }
2986
bpf_get_perf_event_read_value_proto(void)2987 const struct bpf_func_proto * __weak bpf_get_perf_event_read_value_proto(void)
2988 {
2989 return NULL;
2990 }
2991
2992 u64 __weak
bpf_event_output(struct bpf_map * map,u64 flags,void * meta,u64 meta_size,void * ctx,u64 ctx_size,bpf_ctx_copy_t ctx_copy)2993 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2994 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2995 {
2996 return -ENOTSUPP;
2997 }
2998 EXPORT_SYMBOL_GPL(bpf_event_output);
2999
3000 /* Always built-in helper functions. */
3001 const struct bpf_func_proto bpf_tail_call_proto = {
3002 /* func is unused for tail_call, we set it to pass the
3003 * get_helper_proto check
3004 */
3005 .func = BPF_PTR_POISON,
3006 .gpl_only = false,
3007 .ret_type = RET_VOID,
3008 .arg1_type = ARG_PTR_TO_CTX,
3009 .arg2_type = ARG_CONST_MAP_PTR,
3010 .arg3_type = ARG_ANYTHING,
3011 };
3012
3013 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
3014 * It is encouraged to implement bpf_int_jit_compile() instead, so that
3015 * eBPF and implicitly also cBPF can get JITed!
3016 */
bpf_int_jit_compile(struct bpf_prog * prog)3017 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
3018 {
3019 return prog;
3020 }
3021
3022 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
3023 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
3024 */
bpf_jit_compile(struct bpf_prog * prog)3025 void __weak bpf_jit_compile(struct bpf_prog *prog)
3026 {
3027 }
3028
bpf_helper_changes_pkt_data(enum bpf_func_id func_id)3029 bool __weak bpf_helper_changes_pkt_data(enum bpf_func_id func_id)
3030 {
3031 return false;
3032 }
3033
3034 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
3035 * analysis code and wants explicit zero extension inserted by verifier.
3036 * Otherwise, return FALSE.
3037 *
3038 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
3039 * you don't override this. JITs that don't want these extra insns can detect
3040 * them using insn_is_zext.
3041 */
bpf_jit_needs_zext(void)3042 bool __weak bpf_jit_needs_zext(void)
3043 {
3044 return false;
3045 }
3046
3047 /* By default, enable the verifier's mitigations against Spectre v1 and v4 for
3048 * all archs. The value returned must not change at runtime as there is
3049 * currently no support for reloading programs that were loaded without
3050 * mitigations.
3051 */
bpf_jit_bypass_spec_v1(void)3052 bool __weak bpf_jit_bypass_spec_v1(void)
3053 {
3054 return false;
3055 }
3056
bpf_jit_bypass_spec_v4(void)3057 bool __weak bpf_jit_bypass_spec_v4(void)
3058 {
3059 return false;
3060 }
3061
3062 /* Return true if the JIT inlines the call to the helper corresponding to
3063 * the imm.
3064 *
3065 * The verifier will not patch the insn->imm for the call to the helper if
3066 * this returns true.
3067 */
bpf_jit_inlines_helper_call(s32 imm)3068 bool __weak bpf_jit_inlines_helper_call(s32 imm)
3069 {
3070 return false;
3071 }
3072
3073 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
bpf_jit_supports_subprog_tailcalls(void)3074 bool __weak bpf_jit_supports_subprog_tailcalls(void)
3075 {
3076 return false;
3077 }
3078
bpf_jit_supports_percpu_insn(void)3079 bool __weak bpf_jit_supports_percpu_insn(void)
3080 {
3081 return false;
3082 }
3083
bpf_jit_supports_kfunc_call(void)3084 bool __weak bpf_jit_supports_kfunc_call(void)
3085 {
3086 return false;
3087 }
3088
bpf_jit_supports_far_kfunc_call(void)3089 bool __weak bpf_jit_supports_far_kfunc_call(void)
3090 {
3091 return false;
3092 }
3093
bpf_jit_supports_arena(void)3094 bool __weak bpf_jit_supports_arena(void)
3095 {
3096 return false;
3097 }
3098
bpf_jit_supports_insn(struct bpf_insn * insn,bool in_arena)3099 bool __weak bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
3100 {
3101 return false;
3102 }
3103
bpf_arch_uaddress_limit(void)3104 u64 __weak bpf_arch_uaddress_limit(void)
3105 {
3106 #if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE)
3107 return TASK_SIZE;
3108 #else
3109 return 0;
3110 #endif
3111 }
3112
3113 /* Return TRUE if the JIT backend satisfies the following two conditions:
3114 * 1) JIT backend supports atomic_xchg() on pointer-sized words.
3115 * 2) Under the specific arch, the implementation of xchg() is the same
3116 * as atomic_xchg() on pointer-sized words.
3117 */
bpf_jit_supports_ptr_xchg(void)3118 bool __weak bpf_jit_supports_ptr_xchg(void)
3119 {
3120 return false;
3121 }
3122
3123 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
3124 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
3125 */
skb_copy_bits(const struct sk_buff * skb,int offset,void * to,int len)3126 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
3127 int len)
3128 {
3129 return -EFAULT;
3130 }
3131
bpf_arch_text_poke(void * ip,enum bpf_text_poke_type t,void * addr1,void * addr2)3132 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
3133 void *addr1, void *addr2)
3134 {
3135 return -ENOTSUPP;
3136 }
3137
bpf_arch_text_copy(void * dst,void * src,size_t len)3138 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
3139 {
3140 return ERR_PTR(-ENOTSUPP);
3141 }
3142
bpf_arch_text_invalidate(void * dst,size_t len)3143 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
3144 {
3145 return -ENOTSUPP;
3146 }
3147
bpf_jit_supports_exceptions(void)3148 bool __weak bpf_jit_supports_exceptions(void)
3149 {
3150 return false;
3151 }
3152
bpf_jit_supports_private_stack(void)3153 bool __weak bpf_jit_supports_private_stack(void)
3154 {
3155 return false;
3156 }
3157
arch_bpf_stack_walk(bool (* consume_fn)(void * cookie,u64 ip,u64 sp,u64 bp),void * cookie)3158 void __weak arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
3159 {
3160 }
3161
bpf_jit_supports_timed_may_goto(void)3162 bool __weak bpf_jit_supports_timed_may_goto(void)
3163 {
3164 return false;
3165 }
3166
arch_bpf_timed_may_goto(void)3167 u64 __weak arch_bpf_timed_may_goto(void)
3168 {
3169 return 0;
3170 }
3171
bpf_prog_report_may_goto_violation(void)3172 static noinline void bpf_prog_report_may_goto_violation(void)
3173 {
3174 #ifdef CONFIG_BPF_SYSCALL
3175 struct bpf_stream_stage ss;
3176 struct bpf_prog *prog;
3177
3178 prog = bpf_prog_find_from_stack();
3179 if (!prog)
3180 return;
3181 bpf_stream_stage(ss, prog, BPF_STDERR, ({
3182 bpf_stream_printk(ss, "ERROR: Timeout detected for may_goto instruction\n");
3183 bpf_stream_dump_stack(ss);
3184 }));
3185 #endif
3186 }
3187
bpf_check_timed_may_goto(struct bpf_timed_may_goto * p)3188 u64 bpf_check_timed_may_goto(struct bpf_timed_may_goto *p)
3189 {
3190 u64 time = ktime_get_mono_fast_ns();
3191
3192 /* Populate the timestamp for this stack frame, and refresh count. */
3193 if (!p->timestamp) {
3194 p->timestamp = time;
3195 return BPF_MAX_TIMED_LOOPS;
3196 }
3197 /* Check if we've exhausted our time slice, and zero count. */
3198 if (unlikely(time - p->timestamp >= (NSEC_PER_SEC / 4))) {
3199 bpf_prog_report_may_goto_violation();
3200 return 0;
3201 }
3202 /* Refresh the count for the stack frame. */
3203 return BPF_MAX_TIMED_LOOPS;
3204 }
3205
3206 /* for configs without MMU or 32-bit */
3207 __weak const struct bpf_map_ops arena_map_ops;
bpf_arena_get_user_vm_start(struct bpf_arena * arena)3208 __weak u64 bpf_arena_get_user_vm_start(struct bpf_arena *arena)
3209 {
3210 return 0;
3211 }
bpf_arena_get_kern_vm_start(struct bpf_arena * arena)3212 __weak u64 bpf_arena_get_kern_vm_start(struct bpf_arena *arena)
3213 {
3214 return 0;
3215 }
3216
3217 #ifdef CONFIG_BPF_SYSCALL
bpf_global_ma_init(void)3218 static int __init bpf_global_ma_init(void)
3219 {
3220 int ret;
3221
3222 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false);
3223 bpf_global_ma_set = !ret;
3224 return ret;
3225 }
3226 late_initcall(bpf_global_ma_init);
3227 #endif
3228
3229 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
3230 EXPORT_SYMBOL(bpf_stats_enabled_key);
3231
3232 /* All definitions of tracepoints related to BPF. */
3233 #define CREATE_TRACE_POINTS
3234 #include <linux/bpf_trace.h>
3235
3236 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
3237 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);
3238
3239 #ifdef CONFIG_BPF_SYSCALL
3240
bpf_prog_get_file_line(struct bpf_prog * prog,unsigned long ip,const char ** filep,const char ** linep,int * nump)3241 int bpf_prog_get_file_line(struct bpf_prog *prog, unsigned long ip, const char **filep,
3242 const char **linep, int *nump)
3243 {
3244 int idx = -1, insn_start, insn_end, len;
3245 struct bpf_line_info *linfo;
3246 void **jited_linfo;
3247 struct btf *btf;
3248 int nr_linfo;
3249
3250 btf = prog->aux->btf;
3251 linfo = prog->aux->linfo;
3252 jited_linfo = prog->aux->jited_linfo;
3253
3254 if (!btf || !linfo || !jited_linfo)
3255 return -EINVAL;
3256 len = prog->aux->func ? prog->aux->func[prog->aux->func_idx]->len : prog->len;
3257
3258 linfo = &prog->aux->linfo[prog->aux->linfo_idx];
3259 jited_linfo = &prog->aux->jited_linfo[prog->aux->linfo_idx];
3260
3261 insn_start = linfo[0].insn_off;
3262 insn_end = insn_start + len;
3263 nr_linfo = prog->aux->nr_linfo - prog->aux->linfo_idx;
3264
3265 for (int i = 0; i < nr_linfo &&
3266 linfo[i].insn_off >= insn_start && linfo[i].insn_off < insn_end; i++) {
3267 if (jited_linfo[i] >= (void *)ip)
3268 break;
3269 idx = i;
3270 }
3271
3272 if (idx == -1)
3273 return -ENOENT;
3274
3275 /* Get base component of the file path. */
3276 *filep = btf_name_by_offset(btf, linfo[idx].file_name_off);
3277 *filep = kbasename(*filep);
3278 /* Obtain the source line, and strip whitespace in prefix. */
3279 *linep = btf_name_by_offset(btf, linfo[idx].line_off);
3280 while (isspace(**linep))
3281 *linep += 1;
3282 *nump = BPF_LINE_INFO_LINE_NUM(linfo[idx].line_col);
3283 return 0;
3284 }
3285
3286 struct walk_stack_ctx {
3287 struct bpf_prog *prog;
3288 };
3289
find_from_stack_cb(void * cookie,u64 ip,u64 sp,u64 bp)3290 static bool find_from_stack_cb(void *cookie, u64 ip, u64 sp, u64 bp)
3291 {
3292 struct walk_stack_ctx *ctxp = cookie;
3293 struct bpf_prog *prog;
3294
3295 /*
3296 * The RCU read lock is held to safely traverse the latch tree, but we
3297 * don't need its protection when accessing the prog, since it has an
3298 * active stack frame on the current stack trace, and won't disappear.
3299 */
3300 rcu_read_lock();
3301 prog = bpf_prog_ksym_find(ip);
3302 rcu_read_unlock();
3303 if (!prog)
3304 return true;
3305 /* Make sure we return the main prog if we found a subprog */
3306 ctxp->prog = prog->aux->main_prog_aux->prog;
3307 return false;
3308 }
3309
bpf_prog_find_from_stack(void)3310 struct bpf_prog *bpf_prog_find_from_stack(void)
3311 {
3312 struct walk_stack_ctx ctx = {};
3313
3314 arch_bpf_stack_walk(find_from_stack_cb, &ctx);
3315 return ctx.prog;
3316 }
3317
3318 #endif
3319