2 * Linux Socket Filter - Kernel level socket filtering
4 * Based on the design of the Berkeley Packet Filter. The new
5 * internal format has been designed by PLUMgrid:
7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
11 * Jay Schulist <jschlst@samba.org>
12 * Alexei Starovoitov <ast@plumgrid.com>
13 * Daniel Borkmann <dborkman@redhat.com>
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.
20 * Andi Kleen - Fix a few bad bugs and races.
21 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
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>
32 #include <asm/unaligned.h>
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]
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]
55 /* No hurry in this branch
57 * Exported for the bpf jit load helper.
59 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
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;
68 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
74 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
76 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
78 struct bpf_prog_aux *aux;
81 size = round_up(size, PAGE_SIZE);
82 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
86 kmemcheck_annotate_bitfield(fp, meta);
88 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
94 fp->pages = size / PAGE_SIZE;
100 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
102 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
103 gfp_t gfp_extra_flags)
105 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
109 BUG_ON(fp_old == NULL);
111 size = round_up(size, PAGE_SIZE);
112 if (size <= fp_old->pages * PAGE_SIZE)
115 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
117 kmemcheck_annotate_bitfield(fp, meta);
119 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
120 fp->pages = size / PAGE_SIZE;
123 /* We keep fp->aux from fp_old around in the new
124 * reallocated structure.
127 __bpf_prog_free(fp_old);
133 void __bpf_prog_free(struct bpf_prog *fp)
139 #define SHA_BPF_RAW_SIZE \
140 round_up(MAX_BPF_SIZE + sizeof(__be64) + 1, SHA_MESSAGE_BYTES)
142 /* Called under verifier mutex. */
143 void bpf_prog_calc_digest(struct bpf_prog *fp)
145 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
146 static u32 ws[SHA_WORKSPACE_WORDS];
147 static u8 raw[SHA_BPF_RAW_SIZE];
148 struct bpf_insn *dst = (void *)raw;
149 u32 i, bsize, psize, blocks;
155 sha_init(fp->digest);
156 memset(ws, 0, sizeof(ws));
158 /* We need to take out the map fd for the digest calculation
159 * since they are unstable from user space side.
161 for (i = 0, was_ld_map = false; i < fp->len; i++) {
162 dst[i] = fp->insnsi[i];
164 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
165 dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
168 } else if (was_ld_map &&
170 dst[i].dst_reg == 0 &&
171 dst[i].src_reg == 0 &&
180 psize = fp->len * sizeof(struct bpf_insn);
181 memset(&raw[psize], 0, sizeof(raw) - psize);
184 bsize = round_up(psize, SHA_MESSAGE_BYTES);
185 blocks = bsize / SHA_MESSAGE_BYTES;
186 if (bsize - psize >= sizeof(__be64)) {
187 bits = (__be64 *)(todo + bsize - sizeof(__be64));
189 bits = (__be64 *)(todo + bsize + bits_offset);
192 *bits = cpu_to_be64((psize - 1) << 3);
195 sha_transform(fp->digest, todo, ws);
196 todo += SHA_MESSAGE_BYTES;
199 result = (__force __be32 *)fp->digest;
200 for (i = 0; i < SHA_DIGEST_WORDS; i++)
201 result[i] = cpu_to_be32(fp->digest[i]);
204 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
206 return BPF_CLASS(insn->code) == BPF_JMP &&
207 /* Call and Exit are both special jumps with no
208 * target inside the BPF instruction image.
210 BPF_OP(insn->code) != BPF_CALL &&
211 BPF_OP(insn->code) != BPF_EXIT;
214 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
216 struct bpf_insn *insn = prog->insnsi;
217 u32 i, insn_cnt = prog->len;
219 for (i = 0; i < insn_cnt; i++, insn++) {
220 if (!bpf_is_jmp_and_has_target(insn))
223 /* Adjust offset of jmps if we cross boundaries. */
224 if (i < pos && i + insn->off + 1 > pos)
226 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
231 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
232 const struct bpf_insn *patch, u32 len)
234 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
235 struct bpf_prog *prog_adj;
237 /* Since our patchlet doesn't expand the image, we're done. */
238 if (insn_delta == 0) {
239 memcpy(prog->insnsi + off, patch, sizeof(*patch));
243 insn_adj_cnt = prog->len + insn_delta;
245 /* Several new instructions need to be inserted. Make room
246 * for them. Likely, there's no need for a new allocation as
247 * last page could have large enough tailroom.
249 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
254 prog_adj->len = insn_adj_cnt;
256 /* Patching happens in 3 steps:
258 * 1) Move over tail of insnsi from next instruction onwards,
259 * so we can patch the single target insn with one or more
260 * new ones (patching is always from 1 to n insns, n > 0).
261 * 2) Inject new instructions at the target location.
262 * 3) Adjust branch offsets if necessary.
264 insn_rest = insn_adj_cnt - off - len;
266 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
267 sizeof(*patch) * insn_rest);
268 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
270 bpf_adj_branches(prog_adj, off, insn_delta);
275 #ifdef CONFIG_BPF_JIT
276 struct bpf_binary_header *
277 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
278 unsigned int alignment,
279 bpf_jit_fill_hole_t bpf_fill_ill_insns)
281 struct bpf_binary_header *hdr;
282 unsigned int size, hole, start;
284 /* Most of BPF filters are really small, but if some of them
285 * fill a page, allow at least 128 extra bytes to insert a
286 * random section of illegal instructions.
288 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
289 hdr = module_alloc(size);
293 /* Fill space with illegal/arch-dep instructions. */
294 bpf_fill_ill_insns(hdr, size);
296 hdr->pages = size / PAGE_SIZE;
297 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
298 PAGE_SIZE - sizeof(*hdr));
299 start = (get_random_int() % hole) & ~(alignment - 1);
301 /* Leave a random number of instructions before BPF code. */
302 *image_ptr = &hdr->image[start];
307 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
312 int bpf_jit_harden __read_mostly;
314 static int bpf_jit_blind_insn(const struct bpf_insn *from,
315 const struct bpf_insn *aux,
316 struct bpf_insn *to_buff)
318 struct bpf_insn *to = to_buff;
319 u32 imm_rnd = get_random_int();
322 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
323 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
325 if (from->imm == 0 &&
326 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
327 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
328 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
332 switch (from->code) {
333 case BPF_ALU | BPF_ADD | BPF_K:
334 case BPF_ALU | BPF_SUB | BPF_K:
335 case BPF_ALU | BPF_AND | BPF_K:
336 case BPF_ALU | BPF_OR | BPF_K:
337 case BPF_ALU | BPF_XOR | BPF_K:
338 case BPF_ALU | BPF_MUL | BPF_K:
339 case BPF_ALU | BPF_MOV | BPF_K:
340 case BPF_ALU | BPF_DIV | BPF_K:
341 case BPF_ALU | BPF_MOD | BPF_K:
342 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
343 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
344 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
347 case BPF_ALU64 | BPF_ADD | BPF_K:
348 case BPF_ALU64 | BPF_SUB | BPF_K:
349 case BPF_ALU64 | BPF_AND | BPF_K:
350 case BPF_ALU64 | BPF_OR | BPF_K:
351 case BPF_ALU64 | BPF_XOR | BPF_K:
352 case BPF_ALU64 | BPF_MUL | BPF_K:
353 case BPF_ALU64 | BPF_MOV | BPF_K:
354 case BPF_ALU64 | BPF_DIV | BPF_K:
355 case BPF_ALU64 | BPF_MOD | BPF_K:
356 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
357 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
358 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
361 case BPF_JMP | BPF_JEQ | BPF_K:
362 case BPF_JMP | BPF_JNE | BPF_K:
363 case BPF_JMP | BPF_JGT | BPF_K:
364 case BPF_JMP | BPF_JGE | BPF_K:
365 case BPF_JMP | BPF_JSGT | BPF_K:
366 case BPF_JMP | BPF_JSGE | BPF_K:
367 case BPF_JMP | BPF_JSET | BPF_K:
368 /* Accommodate for extra offset in case of a backjump. */
372 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
373 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
374 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
377 case BPF_LD | BPF_ABS | BPF_W:
378 case BPF_LD | BPF_ABS | BPF_H:
379 case BPF_LD | BPF_ABS | BPF_B:
380 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
381 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
382 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
385 case BPF_LD | BPF_IND | BPF_W:
386 case BPF_LD | BPF_IND | BPF_H:
387 case BPF_LD | BPF_IND | BPF_B:
388 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
389 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
390 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
391 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
394 case BPF_LD | BPF_IMM | BPF_DW:
395 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
396 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
397 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
398 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
400 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
401 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
402 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
403 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
406 case BPF_ST | BPF_MEM | BPF_DW:
407 case BPF_ST | BPF_MEM | BPF_W:
408 case BPF_ST | BPF_MEM | BPF_H:
409 case BPF_ST | BPF_MEM | BPF_B:
410 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
411 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
412 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
419 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
420 gfp_t gfp_extra_flags)
422 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
426 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
428 kmemcheck_annotate_bitfield(fp, meta);
430 /* aux->prog still points to the fp_other one, so
431 * when promoting the clone to the real program,
432 * this still needs to be adapted.
434 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
440 static void bpf_prog_clone_free(struct bpf_prog *fp)
442 /* aux was stolen by the other clone, so we cannot free
443 * it from this path! It will be freed eventually by the
444 * other program on release.
446 * At this point, we don't need a deferred release since
447 * clone is guaranteed to not be locked.
453 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
455 /* We have to repoint aux->prog to self, as we don't
456 * know whether fp here is the clone or the original.
459 bpf_prog_clone_free(fp_other);
462 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
464 struct bpf_insn insn_buff[16], aux[2];
465 struct bpf_prog *clone, *tmp;
466 int insn_delta, insn_cnt;
467 struct bpf_insn *insn;
470 if (!bpf_jit_blinding_enabled())
473 clone = bpf_prog_clone_create(prog, GFP_USER);
475 return ERR_PTR(-ENOMEM);
477 insn_cnt = clone->len;
478 insn = clone->insnsi;
480 for (i = 0; i < insn_cnt; i++, insn++) {
481 /* We temporarily need to hold the original ld64 insn
482 * so that we can still access the first part in the
483 * second blinding run.
485 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
487 memcpy(aux, insn, sizeof(aux));
489 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
493 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
495 /* Patching may have repointed aux->prog during
496 * realloc from the original one, so we need to
497 * fix it up here on error.
499 bpf_jit_prog_release_other(prog, clone);
500 return ERR_PTR(-ENOMEM);
504 insn_delta = rewritten - 1;
506 /* Walk new program and skip insns we just inserted. */
507 insn = clone->insnsi + i + insn_delta;
508 insn_cnt += insn_delta;
514 #endif /* CONFIG_BPF_JIT */
516 /* Base function for offset calculation. Needs to go into .text section,
517 * therefore keeping it non-static as well; will also be used by JITs
518 * anyway later on, so do not let the compiler omit it.
520 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
524 EXPORT_SYMBOL_GPL(__bpf_call_base);
527 * __bpf_prog_run - run eBPF program on a given context
528 * @ctx: is the data we are operating on
529 * @insn: is the array of eBPF instructions
531 * Decode and execute eBPF instructions.
533 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
535 u64 stack[MAX_BPF_STACK / sizeof(u64)];
536 u64 regs[MAX_BPF_REG], tmp;
537 static const void *jumptable[256] = {
538 [0 ... 255] = &&default_label,
539 /* Now overwrite non-defaults ... */
540 /* 32 bit ALU operations */
541 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
542 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
543 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
544 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
545 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
546 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
547 [BPF_ALU | BPF_OR | BPF_X] = &&ALU_OR_X,
548 [BPF_ALU | BPF_OR | BPF_K] = &&ALU_OR_K,
549 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
550 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
551 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
552 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
553 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
554 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
555 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
556 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
557 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
558 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
559 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
560 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
561 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
562 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
563 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
564 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
565 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
566 /* 64 bit ALU operations */
567 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
568 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
569 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
570 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
571 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
572 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
573 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
574 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
575 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
576 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
577 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
578 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
579 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
580 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
581 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
582 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
583 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
584 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
585 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
586 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
587 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
588 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
589 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
590 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
591 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
592 /* Call instruction */
593 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
594 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
596 [BPF_JMP | BPF_JA] = &&JMP_JA,
597 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
598 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
599 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
600 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
601 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
602 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
603 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
604 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
605 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
606 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
607 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
608 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
609 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
610 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
612 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
613 /* Store instructions */
614 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
615 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
616 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
617 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
618 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
619 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
620 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
621 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
622 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
623 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
624 /* Load instructions */
625 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
626 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
627 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
628 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
629 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
630 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
631 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
632 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
633 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
634 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
635 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
637 u32 tail_call_cnt = 0;
641 #define CONT ({ insn++; goto select_insn; })
642 #define CONT_JMP ({ insn++; goto select_insn; })
644 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
645 ARG1 = (u64) (unsigned long) ctx;
648 goto *jumptable[insn->code];
651 #define ALU(OPCODE, OP) \
652 ALU64_##OPCODE##_X: \
656 DST = (u32) DST OP (u32) SRC; \
658 ALU64_##OPCODE##_K: \
662 DST = (u32) DST OP (u32) IMM; \
693 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
697 (*(s64 *) &DST) >>= SRC;
700 (*(s64 *) &DST) >>= IMM;
703 if (unlikely(SRC == 0))
705 div64_u64_rem(DST, SRC, &tmp);
709 if (unlikely(SRC == 0))
712 DST = do_div(tmp, (u32) SRC);
715 div64_u64_rem(DST, IMM, &tmp);
720 DST = do_div(tmp, (u32) IMM);
723 if (unlikely(SRC == 0))
725 DST = div64_u64(DST, SRC);
728 if (unlikely(SRC == 0))
731 do_div(tmp, (u32) SRC);
735 DST = div64_u64(DST, IMM);
739 do_div(tmp, (u32) IMM);
745 DST = (__force u16) cpu_to_be16(DST);
748 DST = (__force u32) cpu_to_be32(DST);
751 DST = (__force u64) cpu_to_be64(DST);
758 DST = (__force u16) cpu_to_le16(DST);
761 DST = (__force u32) cpu_to_le32(DST);
764 DST = (__force u64) cpu_to_le64(DST);
771 /* Function call scratches BPF_R1-BPF_R5 registers,
772 * preserves BPF_R6-BPF_R9, and stores return value
775 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
780 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
781 struct bpf_array *array = container_of(map, struct bpf_array, map);
782 struct bpf_prog *prog;
785 if (unlikely(index >= array->map.max_entries))
787 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
792 prog = READ_ONCE(array->ptrs[index]);
796 /* ARG1 at this point is guaranteed to point to CTX from
797 * the verifier side due to the fact that the tail call is
798 * handeled like a helper, that is, bpf_tail_call_proto,
799 * where arg1_type is ARG_PTR_TO_CTX.
859 if (((s64) DST) > ((s64) SRC)) {
865 if (((s64) DST) > ((s64) IMM)) {
871 if (((s64) DST) >= ((s64) SRC)) {
877 if (((s64) DST) >= ((s64) IMM)) {
897 /* STX and ST and LDX*/
898 #define LDST(SIZEOP, SIZE) \
900 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
903 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
906 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
914 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
915 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
918 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
919 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
922 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
925 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
926 * only appearing in the programs where ctx ==
927 * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
928 * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
929 * internal BPF verifier will check that BPF_R6 ==
932 * BPF_ABS and BPF_IND are wrappers of function calls,
933 * so they scratch BPF_R1-BPF_R5 registers, preserve
934 * BPF_R6-BPF_R9, and store return value into BPF_R0.
937 * ctx == skb == BPF_R6 == CTX
940 * SRC == any register
941 * IMM == 32-bit immediate
944 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
947 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
948 if (likely(ptr != NULL)) {
949 BPF_R0 = get_unaligned_be32(ptr);
954 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
957 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
958 if (likely(ptr != NULL)) {
959 BPF_R0 = get_unaligned_be16(ptr);
964 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
967 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
968 if (likely(ptr != NULL)) {
974 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
977 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
980 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
985 /* If we ever reach this, we have a bug somewhere. */
986 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
989 STACK_FRAME_NON_STANDARD(__bpf_prog_run); /* jump table */
991 bool bpf_prog_array_compatible(struct bpf_array *array,
992 const struct bpf_prog *fp)
994 if (!array->owner_prog_type) {
995 /* There's no owner yet where we could check for
998 array->owner_prog_type = fp->type;
999 array->owner_jited = fp->jited;
1004 return array->owner_prog_type == fp->type &&
1005 array->owner_jited == fp->jited;
1008 static int bpf_check_tail_call(const struct bpf_prog *fp)
1010 struct bpf_prog_aux *aux = fp->aux;
1013 for (i = 0; i < aux->used_map_cnt; i++) {
1014 struct bpf_map *map = aux->used_maps[i];
1015 struct bpf_array *array;
1017 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1020 array = container_of(map, struct bpf_array, map);
1021 if (!bpf_prog_array_compatible(array, fp))
1029 * bpf_prog_select_runtime - select exec runtime for BPF program
1030 * @fp: bpf_prog populated with internal BPF program
1031 * @err: pointer to error variable
1033 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1034 * The BPF program will be executed via BPF_PROG_RUN() macro.
1036 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1038 fp->bpf_func = (void *) __bpf_prog_run;
1040 /* eBPF JITs can rewrite the program in case constant
1041 * blinding is active. However, in case of error during
1042 * blinding, bpf_int_jit_compile() must always return a
1043 * valid program, which in this case would simply not
1044 * be JITed, but falls back to the interpreter.
1046 fp = bpf_int_jit_compile(fp);
1047 bpf_prog_lock_ro(fp);
1049 /* The tail call compatibility check can only be done at
1050 * this late stage as we need to determine, if we deal
1051 * with JITed or non JITed program concatenations and not
1052 * all eBPF JITs might immediately support all features.
1054 *err = bpf_check_tail_call(fp);
1058 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1060 static void bpf_prog_free_deferred(struct work_struct *work)
1062 struct bpf_prog_aux *aux;
1064 aux = container_of(work, struct bpf_prog_aux, work);
1065 bpf_jit_free(aux->prog);
1068 /* Free internal BPF program */
1069 void bpf_prog_free(struct bpf_prog *fp)
1071 struct bpf_prog_aux *aux = fp->aux;
1073 INIT_WORK(&aux->work, bpf_prog_free_deferred);
1074 schedule_work(&aux->work);
1076 EXPORT_SYMBOL_GPL(bpf_prog_free);
1078 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1079 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1081 void bpf_user_rnd_init_once(void)
1083 prandom_init_once(&bpf_user_rnd_state);
1086 BPF_CALL_0(bpf_user_rnd_u32)
1088 /* Should someone ever have the rather unwise idea to use some
1089 * of the registers passed into this function, then note that
1090 * this function is called from native eBPF and classic-to-eBPF
1091 * transformations. Register assignments from both sides are
1092 * different, f.e. classic always sets fn(ctx, A, X) here.
1094 struct rnd_state *state;
1097 state = &get_cpu_var(bpf_user_rnd_state);
1098 res = prandom_u32_state(state);
1099 put_cpu_var(bpf_user_rnd_state);
1104 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1105 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1106 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1107 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1109 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1110 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1111 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1112 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1114 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1115 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1116 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1118 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1124 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1125 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1130 /* Always built-in helper functions. */
1131 const struct bpf_func_proto bpf_tail_call_proto = {
1134 .ret_type = RET_VOID,
1135 .arg1_type = ARG_PTR_TO_CTX,
1136 .arg2_type = ARG_CONST_MAP_PTR,
1137 .arg3_type = ARG_ANYTHING,
1140 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
1141 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1146 bool __weak bpf_helper_changes_skb_data(void *func)
1151 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1152 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1154 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,