Loading...
1/*
2 * Linux Socket Filter - Kernel level socket filtering
3 *
4 * Based on the design of the Berkeley Packet Filter. The new
5 * internal format has been designed by PLUMgrid:
6 *
7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
8 *
9 * Authors:
10 *
11 * Jay Schulist <jschlst@samba.org>
12 * Alexei Starovoitov <ast@plumgrid.com>
13 * Daniel Borkmann <dborkman@redhat.com>
14 *
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
19 *
20 * Andi Kleen - Fix a few bad bugs and races.
21 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
22 */
23
24#include <linux/filter.h>
25#include <linux/skbuff.h>
26#include <linux/vmalloc.h>
27#include <linux/random.h>
28#include <linux/moduleloader.h>
29#include <linux/bpf.h>
30#include <linux/frame.h>
31
32#include <asm/unaligned.h>
33
34/* Registers */
35#define BPF_R0 regs[BPF_REG_0]
36#define BPF_R1 regs[BPF_REG_1]
37#define BPF_R2 regs[BPF_REG_2]
38#define BPF_R3 regs[BPF_REG_3]
39#define BPF_R4 regs[BPF_REG_4]
40#define BPF_R5 regs[BPF_REG_5]
41#define BPF_R6 regs[BPF_REG_6]
42#define BPF_R7 regs[BPF_REG_7]
43#define BPF_R8 regs[BPF_REG_8]
44#define BPF_R9 regs[BPF_REG_9]
45#define BPF_R10 regs[BPF_REG_10]
46
47/* Named registers */
48#define DST regs[insn->dst_reg]
49#define SRC regs[insn->src_reg]
50#define FP regs[BPF_REG_FP]
51#define ARG1 regs[BPF_REG_ARG1]
52#define CTX regs[BPF_REG_CTX]
53#define IMM insn->imm
54
55/* No hurry in this branch
56 *
57 * Exported for the bpf jit load helper.
58 */
59void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
60{
61 u8 *ptr = NULL;
62
63 if (k >= SKF_NET_OFF)
64 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
65 else if (k >= SKF_LL_OFF)
66 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
67
68 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
69 return ptr;
70
71 return NULL;
72}
73
74struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
75{
76 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
77 gfp_extra_flags;
78 struct bpf_prog_aux *aux;
79 struct bpf_prog *fp;
80
81 size = round_up(size, PAGE_SIZE);
82 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
83 if (fp == NULL)
84 return NULL;
85
86 kmemcheck_annotate_bitfield(fp, meta);
87
88 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
89 if (aux == NULL) {
90 vfree(fp);
91 return NULL;
92 }
93
94 fp->pages = size / PAGE_SIZE;
95 fp->aux = aux;
96 fp->aux->prog = fp;
97
98 return fp;
99}
100EXPORT_SYMBOL_GPL(bpf_prog_alloc);
101
102struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
103 gfp_t gfp_extra_flags)
104{
105 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
106 gfp_extra_flags;
107 struct bpf_prog *fp;
108 u32 pages, delta;
109 int ret;
110
111 BUG_ON(fp_old == NULL);
112
113 size = round_up(size, PAGE_SIZE);
114 pages = size / PAGE_SIZE;
115 if (pages <= fp_old->pages)
116 return fp_old;
117
118 delta = pages - fp_old->pages;
119 ret = __bpf_prog_charge(fp_old->aux->user, delta);
120 if (ret)
121 return NULL;
122
123 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
124 if (fp == NULL) {
125 __bpf_prog_uncharge(fp_old->aux->user, delta);
126 } else {
127 kmemcheck_annotate_bitfield(fp, meta);
128
129 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
130 fp->pages = pages;
131 fp->aux->prog = fp;
132
133 /* We keep fp->aux from fp_old around in the new
134 * reallocated structure.
135 */
136 fp_old->aux = NULL;
137 __bpf_prog_free(fp_old);
138 }
139
140 return fp;
141}
142
143void __bpf_prog_free(struct bpf_prog *fp)
144{
145 kfree(fp->aux);
146 vfree(fp);
147}
148
149int bpf_prog_calc_tag(struct bpf_prog *fp)
150{
151 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
152 u32 raw_size = bpf_prog_tag_scratch_size(fp);
153 u32 digest[SHA_DIGEST_WORDS];
154 u32 ws[SHA_WORKSPACE_WORDS];
155 u32 i, bsize, psize, blocks;
156 struct bpf_insn *dst;
157 bool was_ld_map;
158 u8 *raw, *todo;
159 __be32 *result;
160 __be64 *bits;
161
162 raw = vmalloc(raw_size);
163 if (!raw)
164 return -ENOMEM;
165
166 sha_init(digest);
167 memset(ws, 0, sizeof(ws));
168
169 /* We need to take out the map fd for the digest calculation
170 * since they are unstable from user space side.
171 */
172 dst = (void *)raw;
173 for (i = 0, was_ld_map = false; i < fp->len; i++) {
174 dst[i] = fp->insnsi[i];
175 if (!was_ld_map &&
176 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
177 dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
178 was_ld_map = true;
179 dst[i].imm = 0;
180 } else if (was_ld_map &&
181 dst[i].code == 0 &&
182 dst[i].dst_reg == 0 &&
183 dst[i].src_reg == 0 &&
184 dst[i].off == 0) {
185 was_ld_map = false;
186 dst[i].imm = 0;
187 } else {
188 was_ld_map = false;
189 }
190 }
191
192 psize = bpf_prog_insn_size(fp);
193 memset(&raw[psize], 0, raw_size - psize);
194 raw[psize++] = 0x80;
195
196 bsize = round_up(psize, SHA_MESSAGE_BYTES);
197 blocks = bsize / SHA_MESSAGE_BYTES;
198 todo = raw;
199 if (bsize - psize >= sizeof(__be64)) {
200 bits = (__be64 *)(todo + bsize - sizeof(__be64));
201 } else {
202 bits = (__be64 *)(todo + bsize + bits_offset);
203 blocks++;
204 }
205 *bits = cpu_to_be64((psize - 1) << 3);
206
207 while (blocks--) {
208 sha_transform(digest, todo, ws);
209 todo += SHA_MESSAGE_BYTES;
210 }
211
212 result = (__force __be32 *)digest;
213 for (i = 0; i < SHA_DIGEST_WORDS; i++)
214 result[i] = cpu_to_be32(digest[i]);
215 memcpy(fp->tag, result, sizeof(fp->tag));
216
217 vfree(raw);
218 return 0;
219}
220
221static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
222{
223 return BPF_CLASS(insn->code) == BPF_JMP &&
224 /* Call and Exit are both special jumps with no
225 * target inside the BPF instruction image.
226 */
227 BPF_OP(insn->code) != BPF_CALL &&
228 BPF_OP(insn->code) != BPF_EXIT;
229}
230
231static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
232{
233 struct bpf_insn *insn = prog->insnsi;
234 u32 i, insn_cnt = prog->len;
235
236 for (i = 0; i < insn_cnt; i++, insn++) {
237 if (!bpf_is_jmp_and_has_target(insn))
238 continue;
239
240 /* Adjust offset of jmps if we cross boundaries. */
241 if (i < pos && i + insn->off + 1 > pos)
242 insn->off += delta;
243 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
244 insn->off -= delta;
245 }
246}
247
248struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
249 const struct bpf_insn *patch, u32 len)
250{
251 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
252 struct bpf_prog *prog_adj;
253
254 /* Since our patchlet doesn't expand the image, we're done. */
255 if (insn_delta == 0) {
256 memcpy(prog->insnsi + off, patch, sizeof(*patch));
257 return prog;
258 }
259
260 insn_adj_cnt = prog->len + insn_delta;
261
262 /* Several new instructions need to be inserted. Make room
263 * for them. Likely, there's no need for a new allocation as
264 * last page could have large enough tailroom.
265 */
266 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
267 GFP_USER);
268 if (!prog_adj)
269 return NULL;
270
271 prog_adj->len = insn_adj_cnt;
272
273 /* Patching happens in 3 steps:
274 *
275 * 1) Move over tail of insnsi from next instruction onwards,
276 * so we can patch the single target insn with one or more
277 * new ones (patching is always from 1 to n insns, n > 0).
278 * 2) Inject new instructions at the target location.
279 * 3) Adjust branch offsets if necessary.
280 */
281 insn_rest = insn_adj_cnt - off - len;
282
283 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
284 sizeof(*patch) * insn_rest);
285 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
286
287 bpf_adj_branches(prog_adj, off, insn_delta);
288
289 return prog_adj;
290}
291
292#ifdef CONFIG_BPF_JIT
293struct bpf_binary_header *
294bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
295 unsigned int alignment,
296 bpf_jit_fill_hole_t bpf_fill_ill_insns)
297{
298 struct bpf_binary_header *hdr;
299 unsigned int size, hole, start;
300
301 /* Most of BPF filters are really small, but if some of them
302 * fill a page, allow at least 128 extra bytes to insert a
303 * random section of illegal instructions.
304 */
305 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
306 hdr = module_alloc(size);
307 if (hdr == NULL)
308 return NULL;
309
310 /* Fill space with illegal/arch-dep instructions. */
311 bpf_fill_ill_insns(hdr, size);
312
313 hdr->pages = size / PAGE_SIZE;
314 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
315 PAGE_SIZE - sizeof(*hdr));
316 start = (get_random_int() % hole) & ~(alignment - 1);
317
318 /* Leave a random number of instructions before BPF code. */
319 *image_ptr = &hdr->image[start];
320
321 return hdr;
322}
323
324void bpf_jit_binary_free(struct bpf_binary_header *hdr)
325{
326 module_memfree(hdr);
327}
328
329int bpf_jit_harden __read_mostly;
330
331static int bpf_jit_blind_insn(const struct bpf_insn *from,
332 const struct bpf_insn *aux,
333 struct bpf_insn *to_buff)
334{
335 struct bpf_insn *to = to_buff;
336 u32 imm_rnd = get_random_int();
337 s16 off;
338
339 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
340 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
341
342 if (from->imm == 0 &&
343 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
344 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
345 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
346 goto out;
347 }
348
349 switch (from->code) {
350 case BPF_ALU | BPF_ADD | BPF_K:
351 case BPF_ALU | BPF_SUB | BPF_K:
352 case BPF_ALU | BPF_AND | BPF_K:
353 case BPF_ALU | BPF_OR | BPF_K:
354 case BPF_ALU | BPF_XOR | BPF_K:
355 case BPF_ALU | BPF_MUL | BPF_K:
356 case BPF_ALU | BPF_MOV | BPF_K:
357 case BPF_ALU | BPF_DIV | BPF_K:
358 case BPF_ALU | BPF_MOD | BPF_K:
359 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
360 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
361 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
362 break;
363
364 case BPF_ALU64 | BPF_ADD | BPF_K:
365 case BPF_ALU64 | BPF_SUB | BPF_K:
366 case BPF_ALU64 | BPF_AND | BPF_K:
367 case BPF_ALU64 | BPF_OR | BPF_K:
368 case BPF_ALU64 | BPF_XOR | BPF_K:
369 case BPF_ALU64 | BPF_MUL | BPF_K:
370 case BPF_ALU64 | BPF_MOV | BPF_K:
371 case BPF_ALU64 | BPF_DIV | BPF_K:
372 case BPF_ALU64 | BPF_MOD | BPF_K:
373 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
374 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
375 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
376 break;
377
378 case BPF_JMP | BPF_JEQ | BPF_K:
379 case BPF_JMP | BPF_JNE | BPF_K:
380 case BPF_JMP | BPF_JGT | BPF_K:
381 case BPF_JMP | BPF_JGE | BPF_K:
382 case BPF_JMP | BPF_JSGT | BPF_K:
383 case BPF_JMP | BPF_JSGE | BPF_K:
384 case BPF_JMP | BPF_JSET | BPF_K:
385 /* Accommodate for extra offset in case of a backjump. */
386 off = from->off;
387 if (off < 0)
388 off -= 2;
389 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
390 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
391 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
392 break;
393
394 case BPF_LD | BPF_ABS | BPF_W:
395 case BPF_LD | BPF_ABS | BPF_H:
396 case BPF_LD | BPF_ABS | BPF_B:
397 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
398 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
399 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
400 break;
401
402 case BPF_LD | BPF_IND | BPF_W:
403 case BPF_LD | BPF_IND | BPF_H:
404 case BPF_LD | BPF_IND | BPF_B:
405 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
406 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
407 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
408 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
409 break;
410
411 case BPF_LD | BPF_IMM | BPF_DW:
412 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
413 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
414 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
415 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
416 break;
417 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
418 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
419 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
420 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
421 break;
422
423 case BPF_ST | BPF_MEM | BPF_DW:
424 case BPF_ST | BPF_MEM | BPF_W:
425 case BPF_ST | BPF_MEM | BPF_H:
426 case BPF_ST | BPF_MEM | BPF_B:
427 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
428 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
429 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
430 break;
431 }
432out:
433 return to - to_buff;
434}
435
436static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
437 gfp_t gfp_extra_flags)
438{
439 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
440 gfp_extra_flags;
441 struct bpf_prog *fp;
442
443 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
444 if (fp != NULL) {
445 kmemcheck_annotate_bitfield(fp, meta);
446
447 /* aux->prog still points to the fp_other one, so
448 * when promoting the clone to the real program,
449 * this still needs to be adapted.
450 */
451 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
452 }
453
454 return fp;
455}
456
457static void bpf_prog_clone_free(struct bpf_prog *fp)
458{
459 /* aux was stolen by the other clone, so we cannot free
460 * it from this path! It will be freed eventually by the
461 * other program on release.
462 *
463 * At this point, we don't need a deferred release since
464 * clone is guaranteed to not be locked.
465 */
466 fp->aux = NULL;
467 __bpf_prog_free(fp);
468}
469
470void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
471{
472 /* We have to repoint aux->prog to self, as we don't
473 * know whether fp here is the clone or the original.
474 */
475 fp->aux->prog = fp;
476 bpf_prog_clone_free(fp_other);
477}
478
479struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
480{
481 struct bpf_insn insn_buff[16], aux[2];
482 struct bpf_prog *clone, *tmp;
483 int insn_delta, insn_cnt;
484 struct bpf_insn *insn;
485 int i, rewritten;
486
487 if (!bpf_jit_blinding_enabled())
488 return prog;
489
490 clone = bpf_prog_clone_create(prog, GFP_USER);
491 if (!clone)
492 return ERR_PTR(-ENOMEM);
493
494 insn_cnt = clone->len;
495 insn = clone->insnsi;
496
497 for (i = 0; i < insn_cnt; i++, insn++) {
498 /* We temporarily need to hold the original ld64 insn
499 * so that we can still access the first part in the
500 * second blinding run.
501 */
502 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
503 insn[1].code == 0)
504 memcpy(aux, insn, sizeof(aux));
505
506 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
507 if (!rewritten)
508 continue;
509
510 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
511 if (!tmp) {
512 /* Patching may have repointed aux->prog during
513 * realloc from the original one, so we need to
514 * fix it up here on error.
515 */
516 bpf_jit_prog_release_other(prog, clone);
517 return ERR_PTR(-ENOMEM);
518 }
519
520 clone = tmp;
521 insn_delta = rewritten - 1;
522
523 /* Walk new program and skip insns we just inserted. */
524 insn = clone->insnsi + i + insn_delta;
525 insn_cnt += insn_delta;
526 i += insn_delta;
527 }
528
529 return clone;
530}
531#endif /* CONFIG_BPF_JIT */
532
533/* Base function for offset calculation. Needs to go into .text section,
534 * therefore keeping it non-static as well; will also be used by JITs
535 * anyway later on, so do not let the compiler omit it.
536 */
537noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
538{
539 return 0;
540}
541EXPORT_SYMBOL_GPL(__bpf_call_base);
542
543/**
544 * __bpf_prog_run - run eBPF program on a given context
545 * @ctx: is the data we are operating on
546 * @insn: is the array of eBPF instructions
547 *
548 * Decode and execute eBPF instructions.
549 */
550static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
551{
552 u64 stack[MAX_BPF_STACK / sizeof(u64)];
553 u64 regs[MAX_BPF_REG], tmp;
554 static const void *jumptable[256] = {
555 [0 ... 255] = &&default_label,
556 /* Now overwrite non-defaults ... */
557 /* 32 bit ALU operations */
558 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
559 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
560 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
561 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
562 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
563 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
564 [BPF_ALU | BPF_OR | BPF_X] = &&ALU_OR_X,
565 [BPF_ALU | BPF_OR | BPF_K] = &&ALU_OR_K,
566 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
567 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
568 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
569 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
570 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
571 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
572 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
573 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
574 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
575 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
576 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
577 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
578 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
579 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
580 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
581 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
582 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
583 /* 64 bit ALU operations */
584 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
585 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
586 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
587 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
588 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
589 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
590 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
591 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
592 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
593 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
594 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
595 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
596 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
597 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
598 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
599 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
600 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
601 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
602 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
603 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
604 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
605 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
606 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
607 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
608 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
609 /* Call instruction */
610 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
611 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
612 /* Jumps */
613 [BPF_JMP | BPF_JA] = &&JMP_JA,
614 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
615 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
616 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
617 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
618 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
619 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
620 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
621 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
622 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
623 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
624 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
625 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
626 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
627 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
628 /* Program return */
629 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
630 /* Store instructions */
631 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
632 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
633 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
634 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
635 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
636 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
637 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
638 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
639 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
640 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
641 /* Load instructions */
642 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
643 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
644 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
645 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
646 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
647 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
648 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
649 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
650 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
651 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
652 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
653 };
654 u32 tail_call_cnt = 0;
655 void *ptr;
656 int off;
657
658#define CONT ({ insn++; goto select_insn; })
659#define CONT_JMP ({ insn++; goto select_insn; })
660
661 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
662 ARG1 = (u64) (unsigned long) ctx;
663
664select_insn:
665 goto *jumptable[insn->code];
666
667 /* ALU */
668#define ALU(OPCODE, OP) \
669 ALU64_##OPCODE##_X: \
670 DST = DST OP SRC; \
671 CONT; \
672 ALU_##OPCODE##_X: \
673 DST = (u32) DST OP (u32) SRC; \
674 CONT; \
675 ALU64_##OPCODE##_K: \
676 DST = DST OP IMM; \
677 CONT; \
678 ALU_##OPCODE##_K: \
679 DST = (u32) DST OP (u32) IMM; \
680 CONT;
681
682 ALU(ADD, +)
683 ALU(SUB, -)
684 ALU(AND, &)
685 ALU(OR, |)
686 ALU(LSH, <<)
687 ALU(RSH, >>)
688 ALU(XOR, ^)
689 ALU(MUL, *)
690#undef ALU
691 ALU_NEG:
692 DST = (u32) -DST;
693 CONT;
694 ALU64_NEG:
695 DST = -DST;
696 CONT;
697 ALU_MOV_X:
698 DST = (u32) SRC;
699 CONT;
700 ALU_MOV_K:
701 DST = (u32) IMM;
702 CONT;
703 ALU64_MOV_X:
704 DST = SRC;
705 CONT;
706 ALU64_MOV_K:
707 DST = IMM;
708 CONT;
709 LD_IMM_DW:
710 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
711 insn++;
712 CONT;
713 ALU64_ARSH_X:
714 (*(s64 *) &DST) >>= SRC;
715 CONT;
716 ALU64_ARSH_K:
717 (*(s64 *) &DST) >>= IMM;
718 CONT;
719 ALU64_MOD_X:
720 if (unlikely(SRC == 0))
721 return 0;
722 div64_u64_rem(DST, SRC, &tmp);
723 DST = tmp;
724 CONT;
725 ALU_MOD_X:
726 if (unlikely(SRC == 0))
727 return 0;
728 tmp = (u32) DST;
729 DST = do_div(tmp, (u32) SRC);
730 CONT;
731 ALU64_MOD_K:
732 div64_u64_rem(DST, IMM, &tmp);
733 DST = tmp;
734 CONT;
735 ALU_MOD_K:
736 tmp = (u32) DST;
737 DST = do_div(tmp, (u32) IMM);
738 CONT;
739 ALU64_DIV_X:
740 if (unlikely(SRC == 0))
741 return 0;
742 DST = div64_u64(DST, SRC);
743 CONT;
744 ALU_DIV_X:
745 if (unlikely(SRC == 0))
746 return 0;
747 tmp = (u32) DST;
748 do_div(tmp, (u32) SRC);
749 DST = (u32) tmp;
750 CONT;
751 ALU64_DIV_K:
752 DST = div64_u64(DST, IMM);
753 CONT;
754 ALU_DIV_K:
755 tmp = (u32) DST;
756 do_div(tmp, (u32) IMM);
757 DST = (u32) tmp;
758 CONT;
759 ALU_END_TO_BE:
760 switch (IMM) {
761 case 16:
762 DST = (__force u16) cpu_to_be16(DST);
763 break;
764 case 32:
765 DST = (__force u32) cpu_to_be32(DST);
766 break;
767 case 64:
768 DST = (__force u64) cpu_to_be64(DST);
769 break;
770 }
771 CONT;
772 ALU_END_TO_LE:
773 switch (IMM) {
774 case 16:
775 DST = (__force u16) cpu_to_le16(DST);
776 break;
777 case 32:
778 DST = (__force u32) cpu_to_le32(DST);
779 break;
780 case 64:
781 DST = (__force u64) cpu_to_le64(DST);
782 break;
783 }
784 CONT;
785
786 /* CALL */
787 JMP_CALL:
788 /* Function call scratches BPF_R1-BPF_R5 registers,
789 * preserves BPF_R6-BPF_R9, and stores return value
790 * into BPF_R0.
791 */
792 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
793 BPF_R4, BPF_R5);
794 CONT;
795
796 JMP_TAIL_CALL: {
797 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
798 struct bpf_array *array = container_of(map, struct bpf_array, map);
799 struct bpf_prog *prog;
800 u64 index = BPF_R3;
801
802 if (unlikely(index >= array->map.max_entries))
803 goto out;
804 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
805 goto out;
806
807 tail_call_cnt++;
808
809 prog = READ_ONCE(array->ptrs[index]);
810 if (!prog)
811 goto out;
812
813 /* ARG1 at this point is guaranteed to point to CTX from
814 * the verifier side due to the fact that the tail call is
815 * handeled like a helper, that is, bpf_tail_call_proto,
816 * where arg1_type is ARG_PTR_TO_CTX.
817 */
818 insn = prog->insnsi;
819 goto select_insn;
820out:
821 CONT;
822 }
823 /* JMP */
824 JMP_JA:
825 insn += insn->off;
826 CONT;
827 JMP_JEQ_X:
828 if (DST == SRC) {
829 insn += insn->off;
830 CONT_JMP;
831 }
832 CONT;
833 JMP_JEQ_K:
834 if (DST == IMM) {
835 insn += insn->off;
836 CONT_JMP;
837 }
838 CONT;
839 JMP_JNE_X:
840 if (DST != SRC) {
841 insn += insn->off;
842 CONT_JMP;
843 }
844 CONT;
845 JMP_JNE_K:
846 if (DST != IMM) {
847 insn += insn->off;
848 CONT_JMP;
849 }
850 CONT;
851 JMP_JGT_X:
852 if (DST > SRC) {
853 insn += insn->off;
854 CONT_JMP;
855 }
856 CONT;
857 JMP_JGT_K:
858 if (DST > IMM) {
859 insn += insn->off;
860 CONT_JMP;
861 }
862 CONT;
863 JMP_JGE_X:
864 if (DST >= SRC) {
865 insn += insn->off;
866 CONT_JMP;
867 }
868 CONT;
869 JMP_JGE_K:
870 if (DST >= IMM) {
871 insn += insn->off;
872 CONT_JMP;
873 }
874 CONT;
875 JMP_JSGT_X:
876 if (((s64) DST) > ((s64) SRC)) {
877 insn += insn->off;
878 CONT_JMP;
879 }
880 CONT;
881 JMP_JSGT_K:
882 if (((s64) DST) > ((s64) IMM)) {
883 insn += insn->off;
884 CONT_JMP;
885 }
886 CONT;
887 JMP_JSGE_X:
888 if (((s64) DST) >= ((s64) SRC)) {
889 insn += insn->off;
890 CONT_JMP;
891 }
892 CONT;
893 JMP_JSGE_K:
894 if (((s64) DST) >= ((s64) IMM)) {
895 insn += insn->off;
896 CONT_JMP;
897 }
898 CONT;
899 JMP_JSET_X:
900 if (DST & SRC) {
901 insn += insn->off;
902 CONT_JMP;
903 }
904 CONT;
905 JMP_JSET_K:
906 if (DST & IMM) {
907 insn += insn->off;
908 CONT_JMP;
909 }
910 CONT;
911 JMP_EXIT:
912 return BPF_R0;
913
914 /* STX and ST and LDX*/
915#define LDST(SIZEOP, SIZE) \
916 STX_MEM_##SIZEOP: \
917 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
918 CONT; \
919 ST_MEM_##SIZEOP: \
920 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
921 CONT; \
922 LDX_MEM_##SIZEOP: \
923 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
924 CONT;
925
926 LDST(B, u8)
927 LDST(H, u16)
928 LDST(W, u32)
929 LDST(DW, u64)
930#undef LDST
931 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
932 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
933 (DST + insn->off));
934 CONT;
935 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
936 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
937 (DST + insn->off));
938 CONT;
939 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
940 off = IMM;
941load_word:
942 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
943 * only appearing in the programs where ctx ==
944 * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
945 * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
946 * internal BPF verifier will check that BPF_R6 ==
947 * ctx.
948 *
949 * BPF_ABS and BPF_IND are wrappers of function calls,
950 * so they scratch BPF_R1-BPF_R5 registers, preserve
951 * BPF_R6-BPF_R9, and store return value into BPF_R0.
952 *
953 * Implicit input:
954 * ctx == skb == BPF_R6 == CTX
955 *
956 * Explicit input:
957 * SRC == any register
958 * IMM == 32-bit immediate
959 *
960 * Output:
961 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
962 */
963
964 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
965 if (likely(ptr != NULL)) {
966 BPF_R0 = get_unaligned_be32(ptr);
967 CONT;
968 }
969
970 return 0;
971 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
972 off = IMM;
973load_half:
974 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
975 if (likely(ptr != NULL)) {
976 BPF_R0 = get_unaligned_be16(ptr);
977 CONT;
978 }
979
980 return 0;
981 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
982 off = IMM;
983load_byte:
984 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
985 if (likely(ptr != NULL)) {
986 BPF_R0 = *(u8 *)ptr;
987 CONT;
988 }
989
990 return 0;
991 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
992 off = IMM + SRC;
993 goto load_word;
994 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
995 off = IMM + SRC;
996 goto load_half;
997 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
998 off = IMM + SRC;
999 goto load_byte;
1000
1001 default_label:
1002 /* If we ever reach this, we have a bug somewhere. */
1003 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
1004 return 0;
1005}
1006STACK_FRAME_NON_STANDARD(__bpf_prog_run); /* jump table */
1007
1008bool bpf_prog_array_compatible(struct bpf_array *array,
1009 const struct bpf_prog *fp)
1010{
1011 if (!array->owner_prog_type) {
1012 /* There's no owner yet where we could check for
1013 * compatibility.
1014 */
1015 array->owner_prog_type = fp->type;
1016 array->owner_jited = fp->jited;
1017
1018 return true;
1019 }
1020
1021 return array->owner_prog_type == fp->type &&
1022 array->owner_jited == fp->jited;
1023}
1024
1025static int bpf_check_tail_call(const struct bpf_prog *fp)
1026{
1027 struct bpf_prog_aux *aux = fp->aux;
1028 int i;
1029
1030 for (i = 0; i < aux->used_map_cnt; i++) {
1031 struct bpf_map *map = aux->used_maps[i];
1032 struct bpf_array *array;
1033
1034 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1035 continue;
1036
1037 array = container_of(map, struct bpf_array, map);
1038 if (!bpf_prog_array_compatible(array, fp))
1039 return -EINVAL;
1040 }
1041
1042 return 0;
1043}
1044
1045/**
1046 * bpf_prog_select_runtime - select exec runtime for BPF program
1047 * @fp: bpf_prog populated with internal BPF program
1048 * @err: pointer to error variable
1049 *
1050 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1051 * The BPF program will be executed via BPF_PROG_RUN() macro.
1052 */
1053struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1054{
1055 fp->bpf_func = (void *) __bpf_prog_run;
1056
1057 /* eBPF JITs can rewrite the program in case constant
1058 * blinding is active. However, in case of error during
1059 * blinding, bpf_int_jit_compile() must always return a
1060 * valid program, which in this case would simply not
1061 * be JITed, but falls back to the interpreter.
1062 */
1063 fp = bpf_int_jit_compile(fp);
1064 bpf_prog_lock_ro(fp);
1065
1066 /* The tail call compatibility check can only be done at
1067 * this late stage as we need to determine, if we deal
1068 * with JITed or non JITed program concatenations and not
1069 * all eBPF JITs might immediately support all features.
1070 */
1071 *err = bpf_check_tail_call(fp);
1072
1073 return fp;
1074}
1075EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1076
1077static void bpf_prog_free_deferred(struct work_struct *work)
1078{
1079 struct bpf_prog_aux *aux;
1080
1081 aux = container_of(work, struct bpf_prog_aux, work);
1082 bpf_jit_free(aux->prog);
1083}
1084
1085/* Free internal BPF program */
1086void bpf_prog_free(struct bpf_prog *fp)
1087{
1088 struct bpf_prog_aux *aux = fp->aux;
1089
1090 INIT_WORK(&aux->work, bpf_prog_free_deferred);
1091 schedule_work(&aux->work);
1092}
1093EXPORT_SYMBOL_GPL(bpf_prog_free);
1094
1095/* RNG for unpriviledged user space with separated state from prandom_u32(). */
1096static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1097
1098void bpf_user_rnd_init_once(void)
1099{
1100 prandom_init_once(&bpf_user_rnd_state);
1101}
1102
1103BPF_CALL_0(bpf_user_rnd_u32)
1104{
1105 /* Should someone ever have the rather unwise idea to use some
1106 * of the registers passed into this function, then note that
1107 * this function is called from native eBPF and classic-to-eBPF
1108 * transformations. Register assignments from both sides are
1109 * different, f.e. classic always sets fn(ctx, A, X) here.
1110 */
1111 struct rnd_state *state;
1112 u32 res;
1113
1114 state = &get_cpu_var(bpf_user_rnd_state);
1115 res = prandom_u32_state(state);
1116 put_cpu_var(bpf_user_rnd_state);
1117
1118 return res;
1119}
1120
1121/* Weak definitions of helper functions in case we don't have bpf syscall. */
1122const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1123const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1124const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1125
1126const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1127const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1128const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1129const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1130
1131const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1132const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1133const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1134
1135const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1136{
1137 return NULL;
1138}
1139
1140u64 __weak
1141bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1142 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1143{
1144 return -ENOTSUPP;
1145}
1146
1147/* Always built-in helper functions. */
1148const struct bpf_func_proto bpf_tail_call_proto = {
1149 .func = NULL,
1150 .gpl_only = false,
1151 .ret_type = RET_VOID,
1152 .arg1_type = ARG_PTR_TO_CTX,
1153 .arg2_type = ARG_CONST_MAP_PTR,
1154 .arg3_type = ARG_ANYTHING,
1155};
1156
1157/* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
1158struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1159{
1160 return prog;
1161}
1162
1163bool __weak bpf_helper_changes_pkt_data(void *func)
1164{
1165 return false;
1166}
1167
1168/* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1169 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1170 */
1171int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
1172 int len)
1173{
1174 return -EFAULT;
1175}
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 <linux/filter.h>
22#include <linux/skbuff.h>
23#include <linux/vmalloc.h>
24#include <linux/random.h>
25#include <linux/moduleloader.h>
26#include <linux/bpf.h>
27#include <linux/btf.h>
28#include <linux/objtool.h>
29#include <linux/rbtree_latch.h>
30#include <linux/kallsyms.h>
31#include <linux/rcupdate.h>
32#include <linux/perf_event.h>
33#include <linux/extable.h>
34#include <linux/log2.h>
35
36#include <asm/barrier.h>
37#include <asm/unaligned.h>
38
39/* Registers */
40#define BPF_R0 regs[BPF_REG_0]
41#define BPF_R1 regs[BPF_REG_1]
42#define BPF_R2 regs[BPF_REG_2]
43#define BPF_R3 regs[BPF_REG_3]
44#define BPF_R4 regs[BPF_REG_4]
45#define BPF_R5 regs[BPF_REG_5]
46#define BPF_R6 regs[BPF_REG_6]
47#define BPF_R7 regs[BPF_REG_7]
48#define BPF_R8 regs[BPF_REG_8]
49#define BPF_R9 regs[BPF_REG_9]
50#define BPF_R10 regs[BPF_REG_10]
51
52/* Named registers */
53#define DST regs[insn->dst_reg]
54#define SRC regs[insn->src_reg]
55#define FP regs[BPF_REG_FP]
56#define AX regs[BPF_REG_AX]
57#define ARG1 regs[BPF_REG_ARG1]
58#define CTX regs[BPF_REG_CTX]
59#define IMM insn->imm
60
61/* No hurry in this branch
62 *
63 * Exported for the bpf jit load helper.
64 */
65void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
66{
67 u8 *ptr = NULL;
68
69 if (k >= SKF_NET_OFF)
70 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
71 else if (k >= SKF_LL_OFF)
72 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
73
74 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
75 return ptr;
76
77 return NULL;
78}
79
80struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
81{
82 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
83 struct bpf_prog_aux *aux;
84 struct bpf_prog *fp;
85
86 size = round_up(size, PAGE_SIZE);
87 fp = __vmalloc(size, gfp_flags);
88 if (fp == NULL)
89 return NULL;
90
91 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags);
92 if (aux == NULL) {
93 vfree(fp);
94 return NULL;
95 }
96 fp->active = alloc_percpu_gfp(int, GFP_KERNEL_ACCOUNT | gfp_extra_flags);
97 if (!fp->active) {
98 vfree(fp);
99 kfree(aux);
100 return NULL;
101 }
102
103 fp->pages = size / PAGE_SIZE;
104 fp->aux = aux;
105 fp->aux->prog = fp;
106 fp->jit_requested = ebpf_jit_enabled();
107
108 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
109 mutex_init(&fp->aux->used_maps_mutex);
110 mutex_init(&fp->aux->dst_mutex);
111
112 return fp;
113}
114
115struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
116{
117 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
118 struct bpf_prog *prog;
119 int cpu;
120
121 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
122 if (!prog)
123 return NULL;
124
125 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
126 if (!prog->stats) {
127 free_percpu(prog->active);
128 kfree(prog->aux);
129 vfree(prog);
130 return NULL;
131 }
132
133 for_each_possible_cpu(cpu) {
134 struct bpf_prog_stats *pstats;
135
136 pstats = per_cpu_ptr(prog->stats, cpu);
137 u64_stats_init(&pstats->syncp);
138 }
139 return prog;
140}
141EXPORT_SYMBOL_GPL(bpf_prog_alloc);
142
143int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
144{
145 if (!prog->aux->nr_linfo || !prog->jit_requested)
146 return 0;
147
148 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
149 sizeof(*prog->aux->jited_linfo),
150 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
151 if (!prog->aux->jited_linfo)
152 return -ENOMEM;
153
154 return 0;
155}
156
157void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
158{
159 if (prog->aux->jited_linfo &&
160 (!prog->jited || !prog->aux->jited_linfo[0])) {
161 kvfree(prog->aux->jited_linfo);
162 prog->aux->jited_linfo = NULL;
163 }
164
165 kfree(prog->aux->kfunc_tab);
166 prog->aux->kfunc_tab = NULL;
167}
168
169/* The jit engine is responsible to provide an array
170 * for insn_off to the jited_off mapping (insn_to_jit_off).
171 *
172 * The idx to this array is the insn_off. Hence, the insn_off
173 * here is relative to the prog itself instead of the main prog.
174 * This array has one entry for each xlated bpf insn.
175 *
176 * jited_off is the byte off to the last byte of the jited insn.
177 *
178 * Hence, with
179 * insn_start:
180 * The first bpf insn off of the prog. The insn off
181 * here is relative to the main prog.
182 * e.g. if prog is a subprog, insn_start > 0
183 * linfo_idx:
184 * The prog's idx to prog->aux->linfo and jited_linfo
185 *
186 * jited_linfo[linfo_idx] = prog->bpf_func
187 *
188 * For i > linfo_idx,
189 *
190 * jited_linfo[i] = prog->bpf_func +
191 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
192 */
193void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
194 const u32 *insn_to_jit_off)
195{
196 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
197 const struct bpf_line_info *linfo;
198 void **jited_linfo;
199
200 if (!prog->aux->jited_linfo)
201 /* Userspace did not provide linfo */
202 return;
203
204 linfo_idx = prog->aux->linfo_idx;
205 linfo = &prog->aux->linfo[linfo_idx];
206 insn_start = linfo[0].insn_off;
207 insn_end = insn_start + prog->len;
208
209 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
210 jited_linfo[0] = prog->bpf_func;
211
212 nr_linfo = prog->aux->nr_linfo - linfo_idx;
213
214 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
215 /* The verifier ensures that linfo[i].insn_off is
216 * strictly increasing
217 */
218 jited_linfo[i] = prog->bpf_func +
219 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
220}
221
222struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
223 gfp_t gfp_extra_flags)
224{
225 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
226 struct bpf_prog *fp;
227 u32 pages;
228
229 size = round_up(size, PAGE_SIZE);
230 pages = size / PAGE_SIZE;
231 if (pages <= fp_old->pages)
232 return fp_old;
233
234 fp = __vmalloc(size, gfp_flags);
235 if (fp) {
236 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
237 fp->pages = pages;
238 fp->aux->prog = fp;
239
240 /* We keep fp->aux from fp_old around in the new
241 * reallocated structure.
242 */
243 fp_old->aux = NULL;
244 fp_old->stats = NULL;
245 fp_old->active = NULL;
246 __bpf_prog_free(fp_old);
247 }
248
249 return fp;
250}
251
252void __bpf_prog_free(struct bpf_prog *fp)
253{
254 if (fp->aux) {
255 mutex_destroy(&fp->aux->used_maps_mutex);
256 mutex_destroy(&fp->aux->dst_mutex);
257 kfree(fp->aux->poke_tab);
258 kfree(fp->aux);
259 }
260 free_percpu(fp->stats);
261 free_percpu(fp->active);
262 vfree(fp);
263}
264
265int bpf_prog_calc_tag(struct bpf_prog *fp)
266{
267 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
268 u32 raw_size = bpf_prog_tag_scratch_size(fp);
269 u32 digest[SHA1_DIGEST_WORDS];
270 u32 ws[SHA1_WORKSPACE_WORDS];
271 u32 i, bsize, psize, blocks;
272 struct bpf_insn *dst;
273 bool was_ld_map;
274 u8 *raw, *todo;
275 __be32 *result;
276 __be64 *bits;
277
278 raw = vmalloc(raw_size);
279 if (!raw)
280 return -ENOMEM;
281
282 sha1_init(digest);
283 memset(ws, 0, sizeof(ws));
284
285 /* We need to take out the map fd for the digest calculation
286 * since they are unstable from user space side.
287 */
288 dst = (void *)raw;
289 for (i = 0, was_ld_map = false; i < fp->len; i++) {
290 dst[i] = fp->insnsi[i];
291 if (!was_ld_map &&
292 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
293 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
294 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
295 was_ld_map = true;
296 dst[i].imm = 0;
297 } else if (was_ld_map &&
298 dst[i].code == 0 &&
299 dst[i].dst_reg == 0 &&
300 dst[i].src_reg == 0 &&
301 dst[i].off == 0) {
302 was_ld_map = false;
303 dst[i].imm = 0;
304 } else {
305 was_ld_map = false;
306 }
307 }
308
309 psize = bpf_prog_insn_size(fp);
310 memset(&raw[psize], 0, raw_size - psize);
311 raw[psize++] = 0x80;
312
313 bsize = round_up(psize, SHA1_BLOCK_SIZE);
314 blocks = bsize / SHA1_BLOCK_SIZE;
315 todo = raw;
316 if (bsize - psize >= sizeof(__be64)) {
317 bits = (__be64 *)(todo + bsize - sizeof(__be64));
318 } else {
319 bits = (__be64 *)(todo + bsize + bits_offset);
320 blocks++;
321 }
322 *bits = cpu_to_be64((psize - 1) << 3);
323
324 while (blocks--) {
325 sha1_transform(digest, todo, ws);
326 todo += SHA1_BLOCK_SIZE;
327 }
328
329 result = (__force __be32 *)digest;
330 for (i = 0; i < SHA1_DIGEST_WORDS; i++)
331 result[i] = cpu_to_be32(digest[i]);
332 memcpy(fp->tag, result, sizeof(fp->tag));
333
334 vfree(raw);
335 return 0;
336}
337
338static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
339 s32 end_new, s32 curr, const bool probe_pass)
340{
341 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
342 s32 delta = end_new - end_old;
343 s64 imm = insn->imm;
344
345 if (curr < pos && curr + imm + 1 >= end_old)
346 imm += delta;
347 else if (curr >= end_new && curr + imm + 1 < end_new)
348 imm -= delta;
349 if (imm < imm_min || imm > imm_max)
350 return -ERANGE;
351 if (!probe_pass)
352 insn->imm = imm;
353 return 0;
354}
355
356static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
357 s32 end_new, s32 curr, const bool probe_pass)
358{
359 const s32 off_min = S16_MIN, off_max = S16_MAX;
360 s32 delta = end_new - end_old;
361 s32 off = insn->off;
362
363 if (curr < pos && curr + off + 1 >= end_old)
364 off += delta;
365 else if (curr >= end_new && curr + off + 1 < end_new)
366 off -= delta;
367 if (off < off_min || off > off_max)
368 return -ERANGE;
369 if (!probe_pass)
370 insn->off = off;
371 return 0;
372}
373
374static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
375 s32 end_new, const bool probe_pass)
376{
377 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
378 struct bpf_insn *insn = prog->insnsi;
379 int ret = 0;
380
381 for (i = 0; i < insn_cnt; i++, insn++) {
382 u8 code;
383
384 /* In the probing pass we still operate on the original,
385 * unpatched image in order to check overflows before we
386 * do any other adjustments. Therefore skip the patchlet.
387 */
388 if (probe_pass && i == pos) {
389 i = end_new;
390 insn = prog->insnsi + end_old;
391 }
392 code = insn->code;
393 if ((BPF_CLASS(code) != BPF_JMP &&
394 BPF_CLASS(code) != BPF_JMP32) ||
395 BPF_OP(code) == BPF_EXIT)
396 continue;
397 /* Adjust offset of jmps if we cross patch boundaries. */
398 if (BPF_OP(code) == BPF_CALL) {
399 if (insn->src_reg != BPF_PSEUDO_CALL)
400 continue;
401 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
402 end_new, i, probe_pass);
403 } else {
404 ret = bpf_adj_delta_to_off(insn, pos, end_old,
405 end_new, i, probe_pass);
406 }
407 if (ret)
408 break;
409 }
410
411 return ret;
412}
413
414static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
415{
416 struct bpf_line_info *linfo;
417 u32 i, nr_linfo;
418
419 nr_linfo = prog->aux->nr_linfo;
420 if (!nr_linfo || !delta)
421 return;
422
423 linfo = prog->aux->linfo;
424
425 for (i = 0; i < nr_linfo; i++)
426 if (off < linfo[i].insn_off)
427 break;
428
429 /* Push all off < linfo[i].insn_off by delta */
430 for (; i < nr_linfo; i++)
431 linfo[i].insn_off += delta;
432}
433
434struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
435 const struct bpf_insn *patch, u32 len)
436{
437 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
438 const u32 cnt_max = S16_MAX;
439 struct bpf_prog *prog_adj;
440 int err;
441
442 /* Since our patchlet doesn't expand the image, we're done. */
443 if (insn_delta == 0) {
444 memcpy(prog->insnsi + off, patch, sizeof(*patch));
445 return prog;
446 }
447
448 insn_adj_cnt = prog->len + insn_delta;
449
450 /* Reject anything that would potentially let the insn->off
451 * target overflow when we have excessive program expansions.
452 * We need to probe here before we do any reallocation where
453 * we afterwards may not fail anymore.
454 */
455 if (insn_adj_cnt > cnt_max &&
456 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
457 return ERR_PTR(err);
458
459 /* Several new instructions need to be inserted. Make room
460 * for them. Likely, there's no need for a new allocation as
461 * last page could have large enough tailroom.
462 */
463 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
464 GFP_USER);
465 if (!prog_adj)
466 return ERR_PTR(-ENOMEM);
467
468 prog_adj->len = insn_adj_cnt;
469
470 /* Patching happens in 3 steps:
471 *
472 * 1) Move over tail of insnsi from next instruction onwards,
473 * so we can patch the single target insn with one or more
474 * new ones (patching is always from 1 to n insns, n > 0).
475 * 2) Inject new instructions at the target location.
476 * 3) Adjust branch offsets if necessary.
477 */
478 insn_rest = insn_adj_cnt - off - len;
479
480 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
481 sizeof(*patch) * insn_rest);
482 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
483
484 /* We are guaranteed to not fail at this point, otherwise
485 * the ship has sailed to reverse to the original state. An
486 * overflow cannot happen at this point.
487 */
488 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
489
490 bpf_adj_linfo(prog_adj, off, insn_delta);
491
492 return prog_adj;
493}
494
495int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
496{
497 /* Branch offsets can't overflow when program is shrinking, no need
498 * to call bpf_adj_branches(..., true) here
499 */
500 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
501 sizeof(struct bpf_insn) * (prog->len - off - cnt));
502 prog->len -= cnt;
503
504 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
505}
506
507static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
508{
509 int i;
510
511 for (i = 0; i < fp->aux->func_cnt; i++)
512 bpf_prog_kallsyms_del(fp->aux->func[i]);
513}
514
515void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
516{
517 bpf_prog_kallsyms_del_subprogs(fp);
518 bpf_prog_kallsyms_del(fp);
519}
520
521#ifdef CONFIG_BPF_JIT
522/* All BPF JIT sysctl knobs here. */
523int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
524int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
525int bpf_jit_harden __read_mostly;
526long bpf_jit_limit __read_mostly;
527
528static void
529bpf_prog_ksym_set_addr(struct bpf_prog *prog)
530{
531 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
532 unsigned long addr = (unsigned long)hdr;
533
534 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
535
536 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
537 prog->aux->ksym.end = addr + hdr->pages * PAGE_SIZE;
538}
539
540static void
541bpf_prog_ksym_set_name(struct bpf_prog *prog)
542{
543 char *sym = prog->aux->ksym.name;
544 const char *end = sym + KSYM_NAME_LEN;
545 const struct btf_type *type;
546 const char *func_name;
547
548 BUILD_BUG_ON(sizeof("bpf_prog_") +
549 sizeof(prog->tag) * 2 +
550 /* name has been null terminated.
551 * We should need +1 for the '_' preceding
552 * the name. However, the null character
553 * is double counted between the name and the
554 * sizeof("bpf_prog_") above, so we omit
555 * the +1 here.
556 */
557 sizeof(prog->aux->name) > KSYM_NAME_LEN);
558
559 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
560 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
561
562 /* prog->aux->name will be ignored if full btf name is available */
563 if (prog->aux->func_info_cnt) {
564 type = btf_type_by_id(prog->aux->btf,
565 prog->aux->func_info[prog->aux->func_idx].type_id);
566 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
567 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
568 return;
569 }
570
571 if (prog->aux->name[0])
572 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
573 else
574 *sym = 0;
575}
576
577static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
578{
579 return container_of(n, struct bpf_ksym, tnode)->start;
580}
581
582static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
583 struct latch_tree_node *b)
584{
585 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
586}
587
588static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
589{
590 unsigned long val = (unsigned long)key;
591 const struct bpf_ksym *ksym;
592
593 ksym = container_of(n, struct bpf_ksym, tnode);
594
595 if (val < ksym->start)
596 return -1;
597 if (val >= ksym->end)
598 return 1;
599
600 return 0;
601}
602
603static const struct latch_tree_ops bpf_tree_ops = {
604 .less = bpf_tree_less,
605 .comp = bpf_tree_comp,
606};
607
608static DEFINE_SPINLOCK(bpf_lock);
609static LIST_HEAD(bpf_kallsyms);
610static struct latch_tree_root bpf_tree __cacheline_aligned;
611
612void bpf_ksym_add(struct bpf_ksym *ksym)
613{
614 spin_lock_bh(&bpf_lock);
615 WARN_ON_ONCE(!list_empty(&ksym->lnode));
616 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
617 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
618 spin_unlock_bh(&bpf_lock);
619}
620
621static void __bpf_ksym_del(struct bpf_ksym *ksym)
622{
623 if (list_empty(&ksym->lnode))
624 return;
625
626 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
627 list_del_rcu(&ksym->lnode);
628}
629
630void bpf_ksym_del(struct bpf_ksym *ksym)
631{
632 spin_lock_bh(&bpf_lock);
633 __bpf_ksym_del(ksym);
634 spin_unlock_bh(&bpf_lock);
635}
636
637static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
638{
639 return fp->jited && !bpf_prog_was_classic(fp);
640}
641
642static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
643{
644 return list_empty(&fp->aux->ksym.lnode) ||
645 fp->aux->ksym.lnode.prev == LIST_POISON2;
646}
647
648void bpf_prog_kallsyms_add(struct bpf_prog *fp)
649{
650 if (!bpf_prog_kallsyms_candidate(fp) ||
651 !bpf_capable())
652 return;
653
654 bpf_prog_ksym_set_addr(fp);
655 bpf_prog_ksym_set_name(fp);
656 fp->aux->ksym.prog = true;
657
658 bpf_ksym_add(&fp->aux->ksym);
659}
660
661void bpf_prog_kallsyms_del(struct bpf_prog *fp)
662{
663 if (!bpf_prog_kallsyms_candidate(fp))
664 return;
665
666 bpf_ksym_del(&fp->aux->ksym);
667}
668
669static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
670{
671 struct latch_tree_node *n;
672
673 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
674 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
675}
676
677const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
678 unsigned long *off, char *sym)
679{
680 struct bpf_ksym *ksym;
681 char *ret = NULL;
682
683 rcu_read_lock();
684 ksym = bpf_ksym_find(addr);
685 if (ksym) {
686 unsigned long symbol_start = ksym->start;
687 unsigned long symbol_end = ksym->end;
688
689 strncpy(sym, ksym->name, KSYM_NAME_LEN);
690
691 ret = sym;
692 if (size)
693 *size = symbol_end - symbol_start;
694 if (off)
695 *off = addr - symbol_start;
696 }
697 rcu_read_unlock();
698
699 return ret;
700}
701
702bool is_bpf_text_address(unsigned long addr)
703{
704 bool ret;
705
706 rcu_read_lock();
707 ret = bpf_ksym_find(addr) != NULL;
708 rcu_read_unlock();
709
710 return ret;
711}
712
713static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
714{
715 struct bpf_ksym *ksym = bpf_ksym_find(addr);
716
717 return ksym && ksym->prog ?
718 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
719 NULL;
720}
721
722const struct exception_table_entry *search_bpf_extables(unsigned long addr)
723{
724 const struct exception_table_entry *e = NULL;
725 struct bpf_prog *prog;
726
727 rcu_read_lock();
728 prog = bpf_prog_ksym_find(addr);
729 if (!prog)
730 goto out;
731 if (!prog->aux->num_exentries)
732 goto out;
733
734 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
735out:
736 rcu_read_unlock();
737 return e;
738}
739
740int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
741 char *sym)
742{
743 struct bpf_ksym *ksym;
744 unsigned int it = 0;
745 int ret = -ERANGE;
746
747 if (!bpf_jit_kallsyms_enabled())
748 return ret;
749
750 rcu_read_lock();
751 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
752 if (it++ != symnum)
753 continue;
754
755 strncpy(sym, ksym->name, KSYM_NAME_LEN);
756
757 *value = ksym->start;
758 *type = BPF_SYM_ELF_TYPE;
759
760 ret = 0;
761 break;
762 }
763 rcu_read_unlock();
764
765 return ret;
766}
767
768int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
769 struct bpf_jit_poke_descriptor *poke)
770{
771 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
772 static const u32 poke_tab_max = 1024;
773 u32 slot = prog->aux->size_poke_tab;
774 u32 size = slot + 1;
775
776 if (size > poke_tab_max)
777 return -ENOSPC;
778 if (poke->tailcall_target || poke->tailcall_target_stable ||
779 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
780 return -EINVAL;
781
782 switch (poke->reason) {
783 case BPF_POKE_REASON_TAIL_CALL:
784 if (!poke->tail_call.map)
785 return -EINVAL;
786 break;
787 default:
788 return -EINVAL;
789 }
790
791 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
792 if (!tab)
793 return -ENOMEM;
794
795 memcpy(&tab[slot], poke, sizeof(*poke));
796 prog->aux->size_poke_tab = size;
797 prog->aux->poke_tab = tab;
798
799 return slot;
800}
801
802static atomic_long_t bpf_jit_current;
803
804/* Can be overridden by an arch's JIT compiler if it has a custom,
805 * dedicated BPF backend memory area, or if neither of the two
806 * below apply.
807 */
808u64 __weak bpf_jit_alloc_exec_limit(void)
809{
810#if defined(MODULES_VADDR)
811 return MODULES_END - MODULES_VADDR;
812#else
813 return VMALLOC_END - VMALLOC_START;
814#endif
815}
816
817static int __init bpf_jit_charge_init(void)
818{
819 /* Only used as heuristic here to derive limit. */
820 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2,
821 PAGE_SIZE), LONG_MAX);
822 return 0;
823}
824pure_initcall(bpf_jit_charge_init);
825
826int bpf_jit_charge_modmem(u32 pages)
827{
828 if (atomic_long_add_return(pages, &bpf_jit_current) >
829 (bpf_jit_limit >> PAGE_SHIFT)) {
830 if (!bpf_capable()) {
831 atomic_long_sub(pages, &bpf_jit_current);
832 return -EPERM;
833 }
834 }
835
836 return 0;
837}
838
839void bpf_jit_uncharge_modmem(u32 pages)
840{
841 atomic_long_sub(pages, &bpf_jit_current);
842}
843
844void *__weak bpf_jit_alloc_exec(unsigned long size)
845{
846 return module_alloc(size);
847}
848
849void __weak bpf_jit_free_exec(void *addr)
850{
851 module_memfree(addr);
852}
853
854struct bpf_binary_header *
855bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
856 unsigned int alignment,
857 bpf_jit_fill_hole_t bpf_fill_ill_insns)
858{
859 struct bpf_binary_header *hdr;
860 u32 size, hole, start, pages;
861
862 WARN_ON_ONCE(!is_power_of_2(alignment) ||
863 alignment > BPF_IMAGE_ALIGNMENT);
864
865 /* Most of BPF filters are really small, but if some of them
866 * fill a page, allow at least 128 extra bytes to insert a
867 * random section of illegal instructions.
868 */
869 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
870 pages = size / PAGE_SIZE;
871
872 if (bpf_jit_charge_modmem(pages))
873 return NULL;
874 hdr = bpf_jit_alloc_exec(size);
875 if (!hdr) {
876 bpf_jit_uncharge_modmem(pages);
877 return NULL;
878 }
879
880 /* Fill space with illegal/arch-dep instructions. */
881 bpf_fill_ill_insns(hdr, size);
882
883 hdr->pages = pages;
884 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
885 PAGE_SIZE - sizeof(*hdr));
886 start = (get_random_int() % hole) & ~(alignment - 1);
887
888 /* Leave a random number of instructions before BPF code. */
889 *image_ptr = &hdr->image[start];
890
891 return hdr;
892}
893
894void bpf_jit_binary_free(struct bpf_binary_header *hdr)
895{
896 u32 pages = hdr->pages;
897
898 bpf_jit_free_exec(hdr);
899 bpf_jit_uncharge_modmem(pages);
900}
901
902/* This symbol is only overridden by archs that have different
903 * requirements than the usual eBPF JITs, f.e. when they only
904 * implement cBPF JIT, do not set images read-only, etc.
905 */
906void __weak bpf_jit_free(struct bpf_prog *fp)
907{
908 if (fp->jited) {
909 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
910
911 bpf_jit_binary_free(hdr);
912
913 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
914 }
915
916 bpf_prog_unlock_free(fp);
917}
918
919int bpf_jit_get_func_addr(const struct bpf_prog *prog,
920 const struct bpf_insn *insn, bool extra_pass,
921 u64 *func_addr, bool *func_addr_fixed)
922{
923 s16 off = insn->off;
924 s32 imm = insn->imm;
925 u8 *addr;
926
927 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
928 if (!*func_addr_fixed) {
929 /* Place-holder address till the last pass has collected
930 * all addresses for JITed subprograms in which case we
931 * can pick them up from prog->aux.
932 */
933 if (!extra_pass)
934 addr = NULL;
935 else if (prog->aux->func &&
936 off >= 0 && off < prog->aux->func_cnt)
937 addr = (u8 *)prog->aux->func[off]->bpf_func;
938 else
939 return -EINVAL;
940 } else {
941 /* Address of a BPF helper call. Since part of the core
942 * kernel, it's always at a fixed location. __bpf_call_base
943 * and the helper with imm relative to it are both in core
944 * kernel.
945 */
946 addr = (u8 *)__bpf_call_base + imm;
947 }
948
949 *func_addr = (unsigned long)addr;
950 return 0;
951}
952
953static int bpf_jit_blind_insn(const struct bpf_insn *from,
954 const struct bpf_insn *aux,
955 struct bpf_insn *to_buff,
956 bool emit_zext)
957{
958 struct bpf_insn *to = to_buff;
959 u32 imm_rnd = get_random_int();
960 s16 off;
961
962 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
963 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
964
965 /* Constraints on AX register:
966 *
967 * AX register is inaccessible from user space. It is mapped in
968 * all JITs, and used here for constant blinding rewrites. It is
969 * typically "stateless" meaning its contents are only valid within
970 * the executed instruction, but not across several instructions.
971 * There are a few exceptions however which are further detailed
972 * below.
973 *
974 * Constant blinding is only used by JITs, not in the interpreter.
975 * The interpreter uses AX in some occasions as a local temporary
976 * register e.g. in DIV or MOD instructions.
977 *
978 * In restricted circumstances, the verifier can also use the AX
979 * register for rewrites as long as they do not interfere with
980 * the above cases!
981 */
982 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
983 goto out;
984
985 if (from->imm == 0 &&
986 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
987 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
988 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
989 goto out;
990 }
991
992 switch (from->code) {
993 case BPF_ALU | BPF_ADD | BPF_K:
994 case BPF_ALU | BPF_SUB | BPF_K:
995 case BPF_ALU | BPF_AND | BPF_K:
996 case BPF_ALU | BPF_OR | BPF_K:
997 case BPF_ALU | BPF_XOR | BPF_K:
998 case BPF_ALU | BPF_MUL | BPF_K:
999 case BPF_ALU | BPF_MOV | BPF_K:
1000 case BPF_ALU | BPF_DIV | BPF_K:
1001 case BPF_ALU | BPF_MOD | BPF_K:
1002 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1003 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1004 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1005 break;
1006
1007 case BPF_ALU64 | BPF_ADD | BPF_K:
1008 case BPF_ALU64 | BPF_SUB | BPF_K:
1009 case BPF_ALU64 | BPF_AND | BPF_K:
1010 case BPF_ALU64 | BPF_OR | BPF_K:
1011 case BPF_ALU64 | BPF_XOR | BPF_K:
1012 case BPF_ALU64 | BPF_MUL | BPF_K:
1013 case BPF_ALU64 | BPF_MOV | BPF_K:
1014 case BPF_ALU64 | BPF_DIV | BPF_K:
1015 case BPF_ALU64 | BPF_MOD | BPF_K:
1016 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1017 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1018 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1019 break;
1020
1021 case BPF_JMP | BPF_JEQ | BPF_K:
1022 case BPF_JMP | BPF_JNE | BPF_K:
1023 case BPF_JMP | BPF_JGT | BPF_K:
1024 case BPF_JMP | BPF_JLT | BPF_K:
1025 case BPF_JMP | BPF_JGE | BPF_K:
1026 case BPF_JMP | BPF_JLE | BPF_K:
1027 case BPF_JMP | BPF_JSGT | BPF_K:
1028 case BPF_JMP | BPF_JSLT | BPF_K:
1029 case BPF_JMP | BPF_JSGE | BPF_K:
1030 case BPF_JMP | BPF_JSLE | BPF_K:
1031 case BPF_JMP | BPF_JSET | BPF_K:
1032 /* Accommodate for extra offset in case of a backjump. */
1033 off = from->off;
1034 if (off < 0)
1035 off -= 2;
1036 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1037 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1038 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1039 break;
1040
1041 case BPF_JMP32 | BPF_JEQ | BPF_K:
1042 case BPF_JMP32 | BPF_JNE | BPF_K:
1043 case BPF_JMP32 | BPF_JGT | BPF_K:
1044 case BPF_JMP32 | BPF_JLT | BPF_K:
1045 case BPF_JMP32 | BPF_JGE | BPF_K:
1046 case BPF_JMP32 | BPF_JLE | BPF_K:
1047 case BPF_JMP32 | BPF_JSGT | BPF_K:
1048 case BPF_JMP32 | BPF_JSLT | BPF_K:
1049 case BPF_JMP32 | BPF_JSGE | BPF_K:
1050 case BPF_JMP32 | BPF_JSLE | BPF_K:
1051 case BPF_JMP32 | BPF_JSET | BPF_K:
1052 /* Accommodate for extra offset in case of a backjump. */
1053 off = from->off;
1054 if (off < 0)
1055 off -= 2;
1056 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1057 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1058 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1059 off);
1060 break;
1061
1062 case BPF_LD | BPF_IMM | BPF_DW:
1063 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1064 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1065 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1066 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1067 break;
1068 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1069 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1070 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1071 if (emit_zext)
1072 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1073 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1074 break;
1075
1076 case BPF_ST | BPF_MEM | BPF_DW:
1077 case BPF_ST | BPF_MEM | BPF_W:
1078 case BPF_ST | BPF_MEM | BPF_H:
1079 case BPF_ST | BPF_MEM | BPF_B:
1080 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1081 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1082 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1083 break;
1084 }
1085out:
1086 return to - to_buff;
1087}
1088
1089static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1090 gfp_t gfp_extra_flags)
1091{
1092 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1093 struct bpf_prog *fp;
1094
1095 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1096 if (fp != NULL) {
1097 /* aux->prog still points to the fp_other one, so
1098 * when promoting the clone to the real program,
1099 * this still needs to be adapted.
1100 */
1101 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1102 }
1103
1104 return fp;
1105}
1106
1107static void bpf_prog_clone_free(struct bpf_prog *fp)
1108{
1109 /* aux was stolen by the other clone, so we cannot free
1110 * it from this path! It will be freed eventually by the
1111 * other program on release.
1112 *
1113 * At this point, we don't need a deferred release since
1114 * clone is guaranteed to not be locked.
1115 */
1116 fp->aux = NULL;
1117 fp->stats = NULL;
1118 fp->active = NULL;
1119 __bpf_prog_free(fp);
1120}
1121
1122void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1123{
1124 /* We have to repoint aux->prog to self, as we don't
1125 * know whether fp here is the clone or the original.
1126 */
1127 fp->aux->prog = fp;
1128 bpf_prog_clone_free(fp_other);
1129}
1130
1131struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1132{
1133 struct bpf_insn insn_buff[16], aux[2];
1134 struct bpf_prog *clone, *tmp;
1135 int insn_delta, insn_cnt;
1136 struct bpf_insn *insn;
1137 int i, rewritten;
1138
1139 if (!bpf_jit_blinding_enabled(prog) || prog->blinded)
1140 return prog;
1141
1142 clone = bpf_prog_clone_create(prog, GFP_USER);
1143 if (!clone)
1144 return ERR_PTR(-ENOMEM);
1145
1146 insn_cnt = clone->len;
1147 insn = clone->insnsi;
1148
1149 for (i = 0; i < insn_cnt; i++, insn++) {
1150 /* We temporarily need to hold the original ld64 insn
1151 * so that we can still access the first part in the
1152 * second blinding run.
1153 */
1154 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1155 insn[1].code == 0)
1156 memcpy(aux, insn, sizeof(aux));
1157
1158 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1159 clone->aux->verifier_zext);
1160 if (!rewritten)
1161 continue;
1162
1163 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1164 if (IS_ERR(tmp)) {
1165 /* Patching may have repointed aux->prog during
1166 * realloc from the original one, so we need to
1167 * fix it up here on error.
1168 */
1169 bpf_jit_prog_release_other(prog, clone);
1170 return tmp;
1171 }
1172
1173 clone = tmp;
1174 insn_delta = rewritten - 1;
1175
1176 /* Walk new program and skip insns we just inserted. */
1177 insn = clone->insnsi + i + insn_delta;
1178 insn_cnt += insn_delta;
1179 i += insn_delta;
1180 }
1181
1182 clone->blinded = 1;
1183 return clone;
1184}
1185#endif /* CONFIG_BPF_JIT */
1186
1187/* Base function for offset calculation. Needs to go into .text section,
1188 * therefore keeping it non-static as well; will also be used by JITs
1189 * anyway later on, so do not let the compiler omit it. This also needs
1190 * to go into kallsyms for correlation from e.g. bpftool, so naming
1191 * must not change.
1192 */
1193noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1194{
1195 return 0;
1196}
1197EXPORT_SYMBOL_GPL(__bpf_call_base);
1198
1199/* All UAPI available opcodes. */
1200#define BPF_INSN_MAP(INSN_2, INSN_3) \
1201 /* 32 bit ALU operations. */ \
1202 /* Register based. */ \
1203 INSN_3(ALU, ADD, X), \
1204 INSN_3(ALU, SUB, X), \
1205 INSN_3(ALU, AND, X), \
1206 INSN_3(ALU, OR, X), \
1207 INSN_3(ALU, LSH, X), \
1208 INSN_3(ALU, RSH, X), \
1209 INSN_3(ALU, XOR, X), \
1210 INSN_3(ALU, MUL, X), \
1211 INSN_3(ALU, MOV, X), \
1212 INSN_3(ALU, ARSH, X), \
1213 INSN_3(ALU, DIV, X), \
1214 INSN_3(ALU, MOD, X), \
1215 INSN_2(ALU, NEG), \
1216 INSN_3(ALU, END, TO_BE), \
1217 INSN_3(ALU, END, TO_LE), \
1218 /* Immediate based. */ \
1219 INSN_3(ALU, ADD, K), \
1220 INSN_3(ALU, SUB, K), \
1221 INSN_3(ALU, AND, K), \
1222 INSN_3(ALU, OR, K), \
1223 INSN_3(ALU, LSH, K), \
1224 INSN_3(ALU, RSH, K), \
1225 INSN_3(ALU, XOR, K), \
1226 INSN_3(ALU, MUL, K), \
1227 INSN_3(ALU, MOV, K), \
1228 INSN_3(ALU, ARSH, K), \
1229 INSN_3(ALU, DIV, K), \
1230 INSN_3(ALU, MOD, K), \
1231 /* 64 bit ALU operations. */ \
1232 /* Register based. */ \
1233 INSN_3(ALU64, ADD, X), \
1234 INSN_3(ALU64, SUB, X), \
1235 INSN_3(ALU64, AND, X), \
1236 INSN_3(ALU64, OR, X), \
1237 INSN_3(ALU64, LSH, X), \
1238 INSN_3(ALU64, RSH, X), \
1239 INSN_3(ALU64, XOR, X), \
1240 INSN_3(ALU64, MUL, X), \
1241 INSN_3(ALU64, MOV, X), \
1242 INSN_3(ALU64, ARSH, X), \
1243 INSN_3(ALU64, DIV, X), \
1244 INSN_3(ALU64, MOD, X), \
1245 INSN_2(ALU64, NEG), \
1246 /* Immediate based. */ \
1247 INSN_3(ALU64, ADD, K), \
1248 INSN_3(ALU64, SUB, K), \
1249 INSN_3(ALU64, AND, K), \
1250 INSN_3(ALU64, OR, K), \
1251 INSN_3(ALU64, LSH, K), \
1252 INSN_3(ALU64, RSH, K), \
1253 INSN_3(ALU64, XOR, K), \
1254 INSN_3(ALU64, MUL, K), \
1255 INSN_3(ALU64, MOV, K), \
1256 INSN_3(ALU64, ARSH, K), \
1257 INSN_3(ALU64, DIV, K), \
1258 INSN_3(ALU64, MOD, K), \
1259 /* Call instruction. */ \
1260 INSN_2(JMP, CALL), \
1261 /* Exit instruction. */ \
1262 INSN_2(JMP, EXIT), \
1263 /* 32-bit Jump instructions. */ \
1264 /* Register based. */ \
1265 INSN_3(JMP32, JEQ, X), \
1266 INSN_3(JMP32, JNE, X), \
1267 INSN_3(JMP32, JGT, X), \
1268 INSN_3(JMP32, JLT, X), \
1269 INSN_3(JMP32, JGE, X), \
1270 INSN_3(JMP32, JLE, X), \
1271 INSN_3(JMP32, JSGT, X), \
1272 INSN_3(JMP32, JSLT, X), \
1273 INSN_3(JMP32, JSGE, X), \
1274 INSN_3(JMP32, JSLE, X), \
1275 INSN_3(JMP32, JSET, X), \
1276 /* Immediate based. */ \
1277 INSN_3(JMP32, JEQ, K), \
1278 INSN_3(JMP32, JNE, K), \
1279 INSN_3(JMP32, JGT, K), \
1280 INSN_3(JMP32, JLT, K), \
1281 INSN_3(JMP32, JGE, K), \
1282 INSN_3(JMP32, JLE, K), \
1283 INSN_3(JMP32, JSGT, K), \
1284 INSN_3(JMP32, JSLT, K), \
1285 INSN_3(JMP32, JSGE, K), \
1286 INSN_3(JMP32, JSLE, K), \
1287 INSN_3(JMP32, JSET, K), \
1288 /* Jump instructions. */ \
1289 /* Register based. */ \
1290 INSN_3(JMP, JEQ, X), \
1291 INSN_3(JMP, JNE, X), \
1292 INSN_3(JMP, JGT, X), \
1293 INSN_3(JMP, JLT, X), \
1294 INSN_3(JMP, JGE, X), \
1295 INSN_3(JMP, JLE, X), \
1296 INSN_3(JMP, JSGT, X), \
1297 INSN_3(JMP, JSLT, X), \
1298 INSN_3(JMP, JSGE, X), \
1299 INSN_3(JMP, JSLE, X), \
1300 INSN_3(JMP, JSET, X), \
1301 /* Immediate based. */ \
1302 INSN_3(JMP, JEQ, K), \
1303 INSN_3(JMP, JNE, K), \
1304 INSN_3(JMP, JGT, K), \
1305 INSN_3(JMP, JLT, K), \
1306 INSN_3(JMP, JGE, K), \
1307 INSN_3(JMP, JLE, K), \
1308 INSN_3(JMP, JSGT, K), \
1309 INSN_3(JMP, JSLT, K), \
1310 INSN_3(JMP, JSGE, K), \
1311 INSN_3(JMP, JSLE, K), \
1312 INSN_3(JMP, JSET, K), \
1313 INSN_2(JMP, JA), \
1314 /* Store instructions. */ \
1315 /* Register based. */ \
1316 INSN_3(STX, MEM, B), \
1317 INSN_3(STX, MEM, H), \
1318 INSN_3(STX, MEM, W), \
1319 INSN_3(STX, MEM, DW), \
1320 INSN_3(STX, ATOMIC, W), \
1321 INSN_3(STX, ATOMIC, DW), \
1322 /* Immediate based. */ \
1323 INSN_3(ST, MEM, B), \
1324 INSN_3(ST, MEM, H), \
1325 INSN_3(ST, MEM, W), \
1326 INSN_3(ST, MEM, DW), \
1327 /* Load instructions. */ \
1328 /* Register based. */ \
1329 INSN_3(LDX, MEM, B), \
1330 INSN_3(LDX, MEM, H), \
1331 INSN_3(LDX, MEM, W), \
1332 INSN_3(LDX, MEM, DW), \
1333 /* Immediate based. */ \
1334 INSN_3(LD, IMM, DW)
1335
1336bool bpf_opcode_in_insntable(u8 code)
1337{
1338#define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1339#define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1340 static const bool public_insntable[256] = {
1341 [0 ... 255] = false,
1342 /* Now overwrite non-defaults ... */
1343 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1344 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1345 [BPF_LD | BPF_ABS | BPF_B] = true,
1346 [BPF_LD | BPF_ABS | BPF_H] = true,
1347 [BPF_LD | BPF_ABS | BPF_W] = true,
1348 [BPF_LD | BPF_IND | BPF_B] = true,
1349 [BPF_LD | BPF_IND | BPF_H] = true,
1350 [BPF_LD | BPF_IND | BPF_W] = true,
1351 };
1352#undef BPF_INSN_3_TBL
1353#undef BPF_INSN_2_TBL
1354 return public_insntable[code];
1355}
1356
1357#ifndef CONFIG_BPF_JIT_ALWAYS_ON
1358u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1359{
1360 memset(dst, 0, size);
1361 return -EFAULT;
1362}
1363
1364/**
1365 * ___bpf_prog_run - run eBPF program on a given context
1366 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1367 * @insn: is the array of eBPF instructions
1368 *
1369 * Decode and execute eBPF instructions.
1370 *
1371 * Return: whatever value is in %BPF_R0 at program exit
1372 */
1373static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1374{
1375#define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1376#define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1377 static const void * const jumptable[256] __annotate_jump_table = {
1378 [0 ... 255] = &&default_label,
1379 /* Now overwrite non-defaults ... */
1380 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1381 /* Non-UAPI available opcodes. */
1382 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1383 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1384 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1385 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1386 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1387 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1388 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1389 };
1390#undef BPF_INSN_3_LBL
1391#undef BPF_INSN_2_LBL
1392 u32 tail_call_cnt = 0;
1393
1394#define CONT ({ insn++; goto select_insn; })
1395#define CONT_JMP ({ insn++; goto select_insn; })
1396
1397select_insn:
1398 goto *jumptable[insn->code];
1399
1400 /* Explicitly mask the register-based shift amounts with 63 or 31
1401 * to avoid undefined behavior. Normally this won't affect the
1402 * generated code, for example, in case of native 64 bit archs such
1403 * as x86-64 or arm64, the compiler is optimizing the AND away for
1404 * the interpreter. In case of JITs, each of the JIT backends compiles
1405 * the BPF shift operations to machine instructions which produce
1406 * implementation-defined results in such a case; the resulting
1407 * contents of the register may be arbitrary, but program behaviour
1408 * as a whole remains defined. In other words, in case of JIT backends,
1409 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1410 */
1411 /* ALU (shifts) */
1412#define SHT(OPCODE, OP) \
1413 ALU64_##OPCODE##_X: \
1414 DST = DST OP (SRC & 63); \
1415 CONT; \
1416 ALU_##OPCODE##_X: \
1417 DST = (u32) DST OP ((u32) SRC & 31); \
1418 CONT; \
1419 ALU64_##OPCODE##_K: \
1420 DST = DST OP IMM; \
1421 CONT; \
1422 ALU_##OPCODE##_K: \
1423 DST = (u32) DST OP (u32) IMM; \
1424 CONT;
1425 /* ALU (rest) */
1426#define ALU(OPCODE, OP) \
1427 ALU64_##OPCODE##_X: \
1428 DST = DST OP SRC; \
1429 CONT; \
1430 ALU_##OPCODE##_X: \
1431 DST = (u32) DST OP (u32) SRC; \
1432 CONT; \
1433 ALU64_##OPCODE##_K: \
1434 DST = DST OP IMM; \
1435 CONT; \
1436 ALU_##OPCODE##_K: \
1437 DST = (u32) DST OP (u32) IMM; \
1438 CONT;
1439 ALU(ADD, +)
1440 ALU(SUB, -)
1441 ALU(AND, &)
1442 ALU(OR, |)
1443 ALU(XOR, ^)
1444 ALU(MUL, *)
1445 SHT(LSH, <<)
1446 SHT(RSH, >>)
1447#undef SHT
1448#undef ALU
1449 ALU_NEG:
1450 DST = (u32) -DST;
1451 CONT;
1452 ALU64_NEG:
1453 DST = -DST;
1454 CONT;
1455 ALU_MOV_X:
1456 DST = (u32) SRC;
1457 CONT;
1458 ALU_MOV_K:
1459 DST = (u32) IMM;
1460 CONT;
1461 ALU64_MOV_X:
1462 DST = SRC;
1463 CONT;
1464 ALU64_MOV_K:
1465 DST = IMM;
1466 CONT;
1467 LD_IMM_DW:
1468 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1469 insn++;
1470 CONT;
1471 ALU_ARSH_X:
1472 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1473 CONT;
1474 ALU_ARSH_K:
1475 DST = (u64) (u32) (((s32) DST) >> IMM);
1476 CONT;
1477 ALU64_ARSH_X:
1478 (*(s64 *) &DST) >>= (SRC & 63);
1479 CONT;
1480 ALU64_ARSH_K:
1481 (*(s64 *) &DST) >>= IMM;
1482 CONT;
1483 ALU64_MOD_X:
1484 div64_u64_rem(DST, SRC, &AX);
1485 DST = AX;
1486 CONT;
1487 ALU_MOD_X:
1488 AX = (u32) DST;
1489 DST = do_div(AX, (u32) SRC);
1490 CONT;
1491 ALU64_MOD_K:
1492 div64_u64_rem(DST, IMM, &AX);
1493 DST = AX;
1494 CONT;
1495 ALU_MOD_K:
1496 AX = (u32) DST;
1497 DST = do_div(AX, (u32) IMM);
1498 CONT;
1499 ALU64_DIV_X:
1500 DST = div64_u64(DST, SRC);
1501 CONT;
1502 ALU_DIV_X:
1503 AX = (u32) DST;
1504 do_div(AX, (u32) SRC);
1505 DST = (u32) AX;
1506 CONT;
1507 ALU64_DIV_K:
1508 DST = div64_u64(DST, IMM);
1509 CONT;
1510 ALU_DIV_K:
1511 AX = (u32) DST;
1512 do_div(AX, (u32) IMM);
1513 DST = (u32) AX;
1514 CONT;
1515 ALU_END_TO_BE:
1516 switch (IMM) {
1517 case 16:
1518 DST = (__force u16) cpu_to_be16(DST);
1519 break;
1520 case 32:
1521 DST = (__force u32) cpu_to_be32(DST);
1522 break;
1523 case 64:
1524 DST = (__force u64) cpu_to_be64(DST);
1525 break;
1526 }
1527 CONT;
1528 ALU_END_TO_LE:
1529 switch (IMM) {
1530 case 16:
1531 DST = (__force u16) cpu_to_le16(DST);
1532 break;
1533 case 32:
1534 DST = (__force u32) cpu_to_le32(DST);
1535 break;
1536 case 64:
1537 DST = (__force u64) cpu_to_le64(DST);
1538 break;
1539 }
1540 CONT;
1541
1542 /* CALL */
1543 JMP_CALL:
1544 /* Function call scratches BPF_R1-BPF_R5 registers,
1545 * preserves BPF_R6-BPF_R9, and stores return value
1546 * into BPF_R0.
1547 */
1548 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1549 BPF_R4, BPF_R5);
1550 CONT;
1551
1552 JMP_CALL_ARGS:
1553 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1554 BPF_R3, BPF_R4,
1555 BPF_R5,
1556 insn + insn->off + 1);
1557 CONT;
1558
1559 JMP_TAIL_CALL: {
1560 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1561 struct bpf_array *array = container_of(map, struct bpf_array, map);
1562 struct bpf_prog *prog;
1563 u32 index = BPF_R3;
1564
1565 if (unlikely(index >= array->map.max_entries))
1566 goto out;
1567 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1568 goto out;
1569
1570 tail_call_cnt++;
1571
1572 prog = READ_ONCE(array->ptrs[index]);
1573 if (!prog)
1574 goto out;
1575
1576 /* ARG1 at this point is guaranteed to point to CTX from
1577 * the verifier side due to the fact that the tail call is
1578 * handled like a helper, that is, bpf_tail_call_proto,
1579 * where arg1_type is ARG_PTR_TO_CTX.
1580 */
1581 insn = prog->insnsi;
1582 goto select_insn;
1583out:
1584 CONT;
1585 }
1586 JMP_JA:
1587 insn += insn->off;
1588 CONT;
1589 JMP_EXIT:
1590 return BPF_R0;
1591 /* JMP */
1592#define COND_JMP(SIGN, OPCODE, CMP_OP) \
1593 JMP_##OPCODE##_X: \
1594 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1595 insn += insn->off; \
1596 CONT_JMP; \
1597 } \
1598 CONT; \
1599 JMP32_##OPCODE##_X: \
1600 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1601 insn += insn->off; \
1602 CONT_JMP; \
1603 } \
1604 CONT; \
1605 JMP_##OPCODE##_K: \
1606 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1607 insn += insn->off; \
1608 CONT_JMP; \
1609 } \
1610 CONT; \
1611 JMP32_##OPCODE##_K: \
1612 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1613 insn += insn->off; \
1614 CONT_JMP; \
1615 } \
1616 CONT;
1617 COND_JMP(u, JEQ, ==)
1618 COND_JMP(u, JNE, !=)
1619 COND_JMP(u, JGT, >)
1620 COND_JMP(u, JLT, <)
1621 COND_JMP(u, JGE, >=)
1622 COND_JMP(u, JLE, <=)
1623 COND_JMP(u, JSET, &)
1624 COND_JMP(s, JSGT, >)
1625 COND_JMP(s, JSLT, <)
1626 COND_JMP(s, JSGE, >=)
1627 COND_JMP(s, JSLE, <=)
1628#undef COND_JMP
1629 /* ST, STX and LDX*/
1630 ST_NOSPEC:
1631 /* Speculation barrier for mitigating Speculative Store Bypass.
1632 * In case of arm64, we rely on the firmware mitigation as
1633 * controlled via the ssbd kernel parameter. Whenever the
1634 * mitigation is enabled, it works for all of the kernel code
1635 * with no need to provide any additional instructions here.
1636 * In case of x86, we use 'lfence' insn for mitigation. We
1637 * reuse preexisting logic from Spectre v1 mitigation that
1638 * happens to produce the required code on x86 for v4 as well.
1639 */
1640#ifdef CONFIG_X86
1641 barrier_nospec();
1642#endif
1643 CONT;
1644#define LDST(SIZEOP, SIZE) \
1645 STX_MEM_##SIZEOP: \
1646 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1647 CONT; \
1648 ST_MEM_##SIZEOP: \
1649 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1650 CONT; \
1651 LDX_MEM_##SIZEOP: \
1652 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1653 CONT;
1654
1655 LDST(B, u8)
1656 LDST(H, u16)
1657 LDST(W, u32)
1658 LDST(DW, u64)
1659#undef LDST
1660#define LDX_PROBE(SIZEOP, SIZE) \
1661 LDX_PROBE_MEM_##SIZEOP: \
1662 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \
1663 CONT;
1664 LDX_PROBE(B, 1)
1665 LDX_PROBE(H, 2)
1666 LDX_PROBE(W, 4)
1667 LDX_PROBE(DW, 8)
1668#undef LDX_PROBE
1669
1670#define ATOMIC_ALU_OP(BOP, KOP) \
1671 case BOP: \
1672 if (BPF_SIZE(insn->code) == BPF_W) \
1673 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
1674 (DST + insn->off)); \
1675 else \
1676 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
1677 (DST + insn->off)); \
1678 break; \
1679 case BOP | BPF_FETCH: \
1680 if (BPF_SIZE(insn->code) == BPF_W) \
1681 SRC = (u32) atomic_fetch_##KOP( \
1682 (u32) SRC, \
1683 (atomic_t *)(unsigned long) (DST + insn->off)); \
1684 else \
1685 SRC = (u64) atomic64_fetch_##KOP( \
1686 (u64) SRC, \
1687 (atomic64_t *)(unsigned long) (DST + insn->off)); \
1688 break;
1689
1690 STX_ATOMIC_DW:
1691 STX_ATOMIC_W:
1692 switch (IMM) {
1693 ATOMIC_ALU_OP(BPF_ADD, add)
1694 ATOMIC_ALU_OP(BPF_AND, and)
1695 ATOMIC_ALU_OP(BPF_OR, or)
1696 ATOMIC_ALU_OP(BPF_XOR, xor)
1697#undef ATOMIC_ALU_OP
1698
1699 case BPF_XCHG:
1700 if (BPF_SIZE(insn->code) == BPF_W)
1701 SRC = (u32) atomic_xchg(
1702 (atomic_t *)(unsigned long) (DST + insn->off),
1703 (u32) SRC);
1704 else
1705 SRC = (u64) atomic64_xchg(
1706 (atomic64_t *)(unsigned long) (DST + insn->off),
1707 (u64) SRC);
1708 break;
1709 case BPF_CMPXCHG:
1710 if (BPF_SIZE(insn->code) == BPF_W)
1711 BPF_R0 = (u32) atomic_cmpxchg(
1712 (atomic_t *)(unsigned long) (DST + insn->off),
1713 (u32) BPF_R0, (u32) SRC);
1714 else
1715 BPF_R0 = (u64) atomic64_cmpxchg(
1716 (atomic64_t *)(unsigned long) (DST + insn->off),
1717 (u64) BPF_R0, (u64) SRC);
1718 break;
1719
1720 default:
1721 goto default_label;
1722 }
1723 CONT;
1724
1725 default_label:
1726 /* If we ever reach this, we have a bug somewhere. Die hard here
1727 * instead of just returning 0; we could be somewhere in a subprog,
1728 * so execution could continue otherwise which we do /not/ want.
1729 *
1730 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
1731 */
1732 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
1733 insn->code, insn->imm);
1734 BUG_ON(1);
1735 return 0;
1736}
1737
1738#define PROG_NAME(stack_size) __bpf_prog_run##stack_size
1739#define DEFINE_BPF_PROG_RUN(stack_size) \
1740static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
1741{ \
1742 u64 stack[stack_size / sizeof(u64)]; \
1743 u64 regs[MAX_BPF_EXT_REG]; \
1744\
1745 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1746 ARG1 = (u64) (unsigned long) ctx; \
1747 return ___bpf_prog_run(regs, insn); \
1748}
1749
1750#define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
1751#define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
1752static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
1753 const struct bpf_insn *insn) \
1754{ \
1755 u64 stack[stack_size / sizeof(u64)]; \
1756 u64 regs[MAX_BPF_EXT_REG]; \
1757\
1758 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1759 BPF_R1 = r1; \
1760 BPF_R2 = r2; \
1761 BPF_R3 = r3; \
1762 BPF_R4 = r4; \
1763 BPF_R5 = r5; \
1764 return ___bpf_prog_run(regs, insn); \
1765}
1766
1767#define EVAL1(FN, X) FN(X)
1768#define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
1769#define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
1770#define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
1771#define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
1772#define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
1773
1774EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
1775EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
1776EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
1777
1778EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
1779EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
1780EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
1781
1782#define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
1783
1784static unsigned int (*interpreters[])(const void *ctx,
1785 const struct bpf_insn *insn) = {
1786EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1787EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1788EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1789};
1790#undef PROG_NAME_LIST
1791#define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
1792static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
1793 const struct bpf_insn *insn) = {
1794EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1795EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1796EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1797};
1798#undef PROG_NAME_LIST
1799
1800void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
1801{
1802 stack_depth = max_t(u32, stack_depth, 1);
1803 insn->off = (s16) insn->imm;
1804 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
1805 __bpf_call_base_args;
1806 insn->code = BPF_JMP | BPF_CALL_ARGS;
1807}
1808
1809#else
1810static unsigned int __bpf_prog_ret0_warn(const void *ctx,
1811 const struct bpf_insn *insn)
1812{
1813 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
1814 * is not working properly, so warn about it!
1815 */
1816 WARN_ON_ONCE(1);
1817 return 0;
1818}
1819#endif
1820
1821bool bpf_prog_array_compatible(struct bpf_array *array,
1822 const struct bpf_prog *fp)
1823{
1824 if (fp->kprobe_override)
1825 return false;
1826
1827 if (!array->aux->type) {
1828 /* There's no owner yet where we could check for
1829 * compatibility.
1830 */
1831 array->aux->type = fp->type;
1832 array->aux->jited = fp->jited;
1833 return true;
1834 }
1835
1836 return array->aux->type == fp->type &&
1837 array->aux->jited == fp->jited;
1838}
1839
1840static int bpf_check_tail_call(const struct bpf_prog *fp)
1841{
1842 struct bpf_prog_aux *aux = fp->aux;
1843 int i, ret = 0;
1844
1845 mutex_lock(&aux->used_maps_mutex);
1846 for (i = 0; i < aux->used_map_cnt; i++) {
1847 struct bpf_map *map = aux->used_maps[i];
1848 struct bpf_array *array;
1849
1850 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1851 continue;
1852
1853 array = container_of(map, struct bpf_array, map);
1854 if (!bpf_prog_array_compatible(array, fp)) {
1855 ret = -EINVAL;
1856 goto out;
1857 }
1858 }
1859
1860out:
1861 mutex_unlock(&aux->used_maps_mutex);
1862 return ret;
1863}
1864
1865static void bpf_prog_select_func(struct bpf_prog *fp)
1866{
1867#ifndef CONFIG_BPF_JIT_ALWAYS_ON
1868 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
1869
1870 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
1871#else
1872 fp->bpf_func = __bpf_prog_ret0_warn;
1873#endif
1874}
1875
1876/**
1877 * bpf_prog_select_runtime - select exec runtime for BPF program
1878 * @fp: bpf_prog populated with internal BPF program
1879 * @err: pointer to error variable
1880 *
1881 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1882 * The BPF program will be executed via BPF_PROG_RUN() macro.
1883 *
1884 * Return: the &fp argument along with &err set to 0 for success or
1885 * a negative errno code on failure
1886 */
1887struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1888{
1889 /* In case of BPF to BPF calls, verifier did all the prep
1890 * work with regards to JITing, etc.
1891 */
1892 bool jit_needed = false;
1893
1894 if (fp->bpf_func)
1895 goto finalize;
1896
1897 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
1898 bpf_prog_has_kfunc_call(fp))
1899 jit_needed = true;
1900
1901 bpf_prog_select_func(fp);
1902
1903 /* eBPF JITs can rewrite the program in case constant
1904 * blinding is active. However, in case of error during
1905 * blinding, bpf_int_jit_compile() must always return a
1906 * valid program, which in this case would simply not
1907 * be JITed, but falls back to the interpreter.
1908 */
1909 if (!bpf_prog_is_dev_bound(fp->aux)) {
1910 *err = bpf_prog_alloc_jited_linfo(fp);
1911 if (*err)
1912 return fp;
1913
1914 fp = bpf_int_jit_compile(fp);
1915 bpf_prog_jit_attempt_done(fp);
1916 if (!fp->jited && jit_needed) {
1917 *err = -ENOTSUPP;
1918 return fp;
1919 }
1920 } else {
1921 *err = bpf_prog_offload_compile(fp);
1922 if (*err)
1923 return fp;
1924 }
1925
1926finalize:
1927 bpf_prog_lock_ro(fp);
1928
1929 /* The tail call compatibility check can only be done at
1930 * this late stage as we need to determine, if we deal
1931 * with JITed or non JITed program concatenations and not
1932 * all eBPF JITs might immediately support all features.
1933 */
1934 *err = bpf_check_tail_call(fp);
1935
1936 return fp;
1937}
1938EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1939
1940static unsigned int __bpf_prog_ret1(const void *ctx,
1941 const struct bpf_insn *insn)
1942{
1943 return 1;
1944}
1945
1946static struct bpf_prog_dummy {
1947 struct bpf_prog prog;
1948} dummy_bpf_prog = {
1949 .prog = {
1950 .bpf_func = __bpf_prog_ret1,
1951 },
1952};
1953
1954/* to avoid allocating empty bpf_prog_array for cgroups that
1955 * don't have bpf program attached use one global 'empty_prog_array'
1956 * It will not be modified the caller of bpf_prog_array_alloc()
1957 * (since caller requested prog_cnt == 0)
1958 * that pointer should be 'freed' by bpf_prog_array_free()
1959 */
1960static struct {
1961 struct bpf_prog_array hdr;
1962 struct bpf_prog *null_prog;
1963} empty_prog_array = {
1964 .null_prog = NULL,
1965};
1966
1967struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
1968{
1969 if (prog_cnt)
1970 return kzalloc(sizeof(struct bpf_prog_array) +
1971 sizeof(struct bpf_prog_array_item) *
1972 (prog_cnt + 1),
1973 flags);
1974
1975 return &empty_prog_array.hdr;
1976}
1977
1978void bpf_prog_array_free(struct bpf_prog_array *progs)
1979{
1980 if (!progs || progs == &empty_prog_array.hdr)
1981 return;
1982 kfree_rcu(progs, rcu);
1983}
1984
1985int bpf_prog_array_length(struct bpf_prog_array *array)
1986{
1987 struct bpf_prog_array_item *item;
1988 u32 cnt = 0;
1989
1990 for (item = array->items; item->prog; item++)
1991 if (item->prog != &dummy_bpf_prog.prog)
1992 cnt++;
1993 return cnt;
1994}
1995
1996bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
1997{
1998 struct bpf_prog_array_item *item;
1999
2000 for (item = array->items; item->prog; item++)
2001 if (item->prog != &dummy_bpf_prog.prog)
2002 return false;
2003 return true;
2004}
2005
2006static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2007 u32 *prog_ids,
2008 u32 request_cnt)
2009{
2010 struct bpf_prog_array_item *item;
2011 int i = 0;
2012
2013 for (item = array->items; item->prog; item++) {
2014 if (item->prog == &dummy_bpf_prog.prog)
2015 continue;
2016 prog_ids[i] = item->prog->aux->id;
2017 if (++i == request_cnt) {
2018 item++;
2019 break;
2020 }
2021 }
2022
2023 return !!(item->prog);
2024}
2025
2026int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2027 __u32 __user *prog_ids, u32 cnt)
2028{
2029 unsigned long err = 0;
2030 bool nospc;
2031 u32 *ids;
2032
2033 /* users of this function are doing:
2034 * cnt = bpf_prog_array_length();
2035 * if (cnt > 0)
2036 * bpf_prog_array_copy_to_user(..., cnt);
2037 * so below kcalloc doesn't need extra cnt > 0 check.
2038 */
2039 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2040 if (!ids)
2041 return -ENOMEM;
2042 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2043 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2044 kfree(ids);
2045 if (err)
2046 return -EFAULT;
2047 if (nospc)
2048 return -ENOSPC;
2049 return 0;
2050}
2051
2052void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2053 struct bpf_prog *old_prog)
2054{
2055 struct bpf_prog_array_item *item;
2056
2057 for (item = array->items; item->prog; item++)
2058 if (item->prog == old_prog) {
2059 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2060 break;
2061 }
2062}
2063
2064/**
2065 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2066 * index into the program array with
2067 * a dummy no-op program.
2068 * @array: a bpf_prog_array
2069 * @index: the index of the program to replace
2070 *
2071 * Skips over dummy programs, by not counting them, when calculating
2072 * the position of the program to replace.
2073 *
2074 * Return:
2075 * * 0 - Success
2076 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2077 * * -ENOENT - Index out of range
2078 */
2079int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2080{
2081 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2082}
2083
2084/**
2085 * bpf_prog_array_update_at() - Updates the program at the given index
2086 * into the program array.
2087 * @array: a bpf_prog_array
2088 * @index: the index of the program to update
2089 * @prog: the program to insert into the array
2090 *
2091 * Skips over dummy programs, by not counting them, when calculating
2092 * the position of the program to update.
2093 *
2094 * Return:
2095 * * 0 - Success
2096 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2097 * * -ENOENT - Index out of range
2098 */
2099int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2100 struct bpf_prog *prog)
2101{
2102 struct bpf_prog_array_item *item;
2103
2104 if (unlikely(index < 0))
2105 return -EINVAL;
2106
2107 for (item = array->items; item->prog; item++) {
2108 if (item->prog == &dummy_bpf_prog.prog)
2109 continue;
2110 if (!index) {
2111 WRITE_ONCE(item->prog, prog);
2112 return 0;
2113 }
2114 index--;
2115 }
2116 return -ENOENT;
2117}
2118
2119int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2120 struct bpf_prog *exclude_prog,
2121 struct bpf_prog *include_prog,
2122 struct bpf_prog_array **new_array)
2123{
2124 int new_prog_cnt, carry_prog_cnt = 0;
2125 struct bpf_prog_array_item *existing;
2126 struct bpf_prog_array *array;
2127 bool found_exclude = false;
2128 int new_prog_idx = 0;
2129
2130 /* Figure out how many existing progs we need to carry over to
2131 * the new array.
2132 */
2133 if (old_array) {
2134 existing = old_array->items;
2135 for (; existing->prog; existing++) {
2136 if (existing->prog == exclude_prog) {
2137 found_exclude = true;
2138 continue;
2139 }
2140 if (existing->prog != &dummy_bpf_prog.prog)
2141 carry_prog_cnt++;
2142 if (existing->prog == include_prog)
2143 return -EEXIST;
2144 }
2145 }
2146
2147 if (exclude_prog && !found_exclude)
2148 return -ENOENT;
2149
2150 /* How many progs (not NULL) will be in the new array? */
2151 new_prog_cnt = carry_prog_cnt;
2152 if (include_prog)
2153 new_prog_cnt += 1;
2154
2155 /* Do we have any prog (not NULL) in the new array? */
2156 if (!new_prog_cnt) {
2157 *new_array = NULL;
2158 return 0;
2159 }
2160
2161 /* +1 as the end of prog_array is marked with NULL */
2162 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2163 if (!array)
2164 return -ENOMEM;
2165
2166 /* Fill in the new prog array */
2167 if (carry_prog_cnt) {
2168 existing = old_array->items;
2169 for (; existing->prog; existing++)
2170 if (existing->prog != exclude_prog &&
2171 existing->prog != &dummy_bpf_prog.prog) {
2172 array->items[new_prog_idx++].prog =
2173 existing->prog;
2174 }
2175 }
2176 if (include_prog)
2177 array->items[new_prog_idx++].prog = include_prog;
2178 array->items[new_prog_idx].prog = NULL;
2179 *new_array = array;
2180 return 0;
2181}
2182
2183int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2184 u32 *prog_ids, u32 request_cnt,
2185 u32 *prog_cnt)
2186{
2187 u32 cnt = 0;
2188
2189 if (array)
2190 cnt = bpf_prog_array_length(array);
2191
2192 *prog_cnt = cnt;
2193
2194 /* return early if user requested only program count or nothing to copy */
2195 if (!request_cnt || !cnt)
2196 return 0;
2197
2198 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2199 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2200 : 0;
2201}
2202
2203void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2204 struct bpf_map **used_maps, u32 len)
2205{
2206 struct bpf_map *map;
2207 u32 i;
2208
2209 for (i = 0; i < len; i++) {
2210 map = used_maps[i];
2211 if (map->ops->map_poke_untrack)
2212 map->ops->map_poke_untrack(map, aux);
2213 bpf_map_put(map);
2214 }
2215}
2216
2217static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2218{
2219 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2220 kfree(aux->used_maps);
2221}
2222
2223void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2224 struct btf_mod_pair *used_btfs, u32 len)
2225{
2226#ifdef CONFIG_BPF_SYSCALL
2227 struct btf_mod_pair *btf_mod;
2228 u32 i;
2229
2230 for (i = 0; i < len; i++) {
2231 btf_mod = &used_btfs[i];
2232 if (btf_mod->module)
2233 module_put(btf_mod->module);
2234 btf_put(btf_mod->btf);
2235 }
2236#endif
2237}
2238
2239static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2240{
2241 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2242 kfree(aux->used_btfs);
2243}
2244
2245static void bpf_prog_free_deferred(struct work_struct *work)
2246{
2247 struct bpf_prog_aux *aux;
2248 int i;
2249
2250 aux = container_of(work, struct bpf_prog_aux, work);
2251 bpf_free_used_maps(aux);
2252 bpf_free_used_btfs(aux);
2253 if (bpf_prog_is_dev_bound(aux))
2254 bpf_prog_offload_destroy(aux->prog);
2255#ifdef CONFIG_PERF_EVENTS
2256 if (aux->prog->has_callchain_buf)
2257 put_callchain_buffers();
2258#endif
2259 if (aux->dst_trampoline)
2260 bpf_trampoline_put(aux->dst_trampoline);
2261 for (i = 0; i < aux->func_cnt; i++) {
2262 /* We can just unlink the subprog poke descriptor table as
2263 * it was originally linked to the main program and is also
2264 * released along with it.
2265 */
2266 aux->func[i]->aux->poke_tab = NULL;
2267 bpf_jit_free(aux->func[i]);
2268 }
2269 if (aux->func_cnt) {
2270 kfree(aux->func);
2271 bpf_prog_unlock_free(aux->prog);
2272 } else {
2273 bpf_jit_free(aux->prog);
2274 }
2275}
2276
2277/* Free internal BPF program */
2278void bpf_prog_free(struct bpf_prog *fp)
2279{
2280 struct bpf_prog_aux *aux = fp->aux;
2281
2282 if (aux->dst_prog)
2283 bpf_prog_put(aux->dst_prog);
2284 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2285 schedule_work(&aux->work);
2286}
2287EXPORT_SYMBOL_GPL(bpf_prog_free);
2288
2289/* RNG for unpriviledged user space with separated state from prandom_u32(). */
2290static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2291
2292void bpf_user_rnd_init_once(void)
2293{
2294 prandom_init_once(&bpf_user_rnd_state);
2295}
2296
2297BPF_CALL_0(bpf_user_rnd_u32)
2298{
2299 /* Should someone ever have the rather unwise idea to use some
2300 * of the registers passed into this function, then note that
2301 * this function is called from native eBPF and classic-to-eBPF
2302 * transformations. Register assignments from both sides are
2303 * different, f.e. classic always sets fn(ctx, A, X) here.
2304 */
2305 struct rnd_state *state;
2306 u32 res;
2307
2308 state = &get_cpu_var(bpf_user_rnd_state);
2309 res = prandom_u32_state(state);
2310 put_cpu_var(bpf_user_rnd_state);
2311
2312 return res;
2313}
2314
2315BPF_CALL_0(bpf_get_raw_cpu_id)
2316{
2317 return raw_smp_processor_id();
2318}
2319
2320/* Weak definitions of helper functions in case we don't have bpf syscall. */
2321const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2322const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2323const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2324const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2325const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2326const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2327const struct bpf_func_proto bpf_spin_lock_proto __weak;
2328const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2329const struct bpf_func_proto bpf_jiffies64_proto __weak;
2330
2331const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2332const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2333const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2334const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2335const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2336const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2337
2338const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2339const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2340const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2341const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2342const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2343const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2344const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2345const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2346const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2347
2348const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2349{
2350 return NULL;
2351}
2352
2353u64 __weak
2354bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2355 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2356{
2357 return -ENOTSUPP;
2358}
2359EXPORT_SYMBOL_GPL(bpf_event_output);
2360
2361/* Always built-in helper functions. */
2362const struct bpf_func_proto bpf_tail_call_proto = {
2363 .func = NULL,
2364 .gpl_only = false,
2365 .ret_type = RET_VOID,
2366 .arg1_type = ARG_PTR_TO_CTX,
2367 .arg2_type = ARG_CONST_MAP_PTR,
2368 .arg3_type = ARG_ANYTHING,
2369};
2370
2371/* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2372 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2373 * eBPF and implicitly also cBPF can get JITed!
2374 */
2375struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2376{
2377 return prog;
2378}
2379
2380/* Stub for JITs that support eBPF. All cBPF code gets transformed into
2381 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2382 */
2383void __weak bpf_jit_compile(struct bpf_prog *prog)
2384{
2385}
2386
2387bool __weak bpf_helper_changes_pkt_data(void *func)
2388{
2389 return false;
2390}
2391
2392/* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2393 * analysis code and wants explicit zero extension inserted by verifier.
2394 * Otherwise, return FALSE.
2395 *
2396 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2397 * you don't override this. JITs that don't want these extra insns can detect
2398 * them using insn_is_zext.
2399 */
2400bool __weak bpf_jit_needs_zext(void)
2401{
2402 return false;
2403}
2404
2405bool __weak bpf_jit_supports_kfunc_call(void)
2406{
2407 return false;
2408}
2409
2410/* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2411 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2412 */
2413int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2414 int len)
2415{
2416 return -EFAULT;
2417}
2418
2419int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2420 void *addr1, void *addr2)
2421{
2422 return -ENOTSUPP;
2423}
2424
2425DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2426EXPORT_SYMBOL(bpf_stats_enabled_key);
2427
2428/* All definitions of tracepoints related to BPF. */
2429#define CREATE_TRACE_POINTS
2430#include <linux/bpf_trace.h>
2431
2432EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2433EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);