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 int bpf_prog_calc_digest(struct bpf_prog *fp)
141 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
142 u32 raw_size = bpf_prog_digest_scratch_size(fp);
143 u32 ws[SHA_WORKSPACE_WORDS];
144 u32 i, bsize, psize, blocks;
145 struct bpf_insn *dst;
151 raw = vmalloc(raw_size);
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.
162 for (i = 0, was_ld_map = false; i < fp->len; i++) {
163 dst[i] = fp->insnsi[i];
165 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
166 dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
169 } else if (was_ld_map &&
171 dst[i].dst_reg == 0 &&
172 dst[i].src_reg == 0 &&
181 psize = bpf_prog_insn_size(fp);
182 memset(&raw[psize], 0, raw_size - psize);
185 bsize = round_up(psize, SHA_MESSAGE_BYTES);
186 blocks = bsize / SHA_MESSAGE_BYTES;
188 if (bsize - psize >= sizeof(__be64)) {
189 bits = (__be64 *)(todo + bsize - sizeof(__be64));
191 bits = (__be64 *)(todo + bsize + bits_offset);
194 *bits = cpu_to_be64((psize - 1) << 3);
197 sha_transform(fp->digest, todo, ws);
198 todo += SHA_MESSAGE_BYTES;
201 result = (__force __be32 *)fp->digest;
202 for (i = 0; i < SHA_DIGEST_WORDS; i++)
203 result[i] = cpu_to_be32(fp->digest[i]);
209 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
211 return BPF_CLASS(insn->code) == BPF_JMP &&
212 /* Call and Exit are both special jumps with no
213 * target inside the BPF instruction image.
215 BPF_OP(insn->code) != BPF_CALL &&
216 BPF_OP(insn->code) != BPF_EXIT;
219 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
221 struct bpf_insn *insn = prog->insnsi;
222 u32 i, insn_cnt = prog->len;
224 for (i = 0; i < insn_cnt; i++, insn++) {
225 if (!bpf_is_jmp_and_has_target(insn))
228 /* Adjust offset of jmps if we cross boundaries. */
229 if (i < pos && i + insn->off + 1 > pos)
231 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
236 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
237 const struct bpf_insn *patch, u32 len)
239 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
240 struct bpf_prog *prog_adj;
242 /* Since our patchlet doesn't expand the image, we're done. */
243 if (insn_delta == 0) {
244 memcpy(prog->insnsi + off, patch, sizeof(*patch));
248 insn_adj_cnt = prog->len + insn_delta;
250 /* Several new instructions need to be inserted. Make room
251 * for them. Likely, there's no need for a new allocation as
252 * last page could have large enough tailroom.
254 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
259 prog_adj->len = insn_adj_cnt;
261 /* Patching happens in 3 steps:
263 * 1) Move over tail of insnsi from next instruction onwards,
264 * so we can patch the single target insn with one or more
265 * new ones (patching is always from 1 to n insns, n > 0).
266 * 2) Inject new instructions at the target location.
267 * 3) Adjust branch offsets if necessary.
269 insn_rest = insn_adj_cnt - off - len;
271 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
272 sizeof(*patch) * insn_rest);
273 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
275 bpf_adj_branches(prog_adj, off, insn_delta);
280 #ifdef CONFIG_BPF_JIT
281 struct bpf_binary_header *
282 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
283 unsigned int alignment,
284 bpf_jit_fill_hole_t bpf_fill_ill_insns)
286 struct bpf_binary_header *hdr;
287 unsigned int size, hole, start;
289 /* Most of BPF filters are really small, but if some of them
290 * fill a page, allow at least 128 extra bytes to insert a
291 * random section of illegal instructions.
293 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
294 hdr = module_alloc(size);
298 /* Fill space with illegal/arch-dep instructions. */
299 bpf_fill_ill_insns(hdr, size);
301 hdr->pages = size / PAGE_SIZE;
302 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
303 PAGE_SIZE - sizeof(*hdr));
304 start = (get_random_int() % hole) & ~(alignment - 1);
306 /* Leave a random number of instructions before BPF code. */
307 *image_ptr = &hdr->image[start];
312 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
317 int bpf_jit_harden __read_mostly;
319 static int bpf_jit_blind_insn(const struct bpf_insn *from,
320 const struct bpf_insn *aux,
321 struct bpf_insn *to_buff)
323 struct bpf_insn *to = to_buff;
324 u32 imm_rnd = get_random_int();
327 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
328 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
330 if (from->imm == 0 &&
331 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
332 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
333 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
337 switch (from->code) {
338 case BPF_ALU | BPF_ADD | BPF_K:
339 case BPF_ALU | BPF_SUB | BPF_K:
340 case BPF_ALU | BPF_AND | BPF_K:
341 case BPF_ALU | BPF_OR | BPF_K:
342 case BPF_ALU | BPF_XOR | BPF_K:
343 case BPF_ALU | BPF_MUL | BPF_K:
344 case BPF_ALU | BPF_MOV | BPF_K:
345 case BPF_ALU | BPF_DIV | BPF_K:
346 case BPF_ALU | BPF_MOD | BPF_K:
347 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
348 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
349 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
352 case BPF_ALU64 | BPF_ADD | BPF_K:
353 case BPF_ALU64 | BPF_SUB | BPF_K:
354 case BPF_ALU64 | BPF_AND | BPF_K:
355 case BPF_ALU64 | BPF_OR | BPF_K:
356 case BPF_ALU64 | BPF_XOR | BPF_K:
357 case BPF_ALU64 | BPF_MUL | BPF_K:
358 case BPF_ALU64 | BPF_MOV | BPF_K:
359 case BPF_ALU64 | BPF_DIV | BPF_K:
360 case BPF_ALU64 | BPF_MOD | BPF_K:
361 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
362 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
363 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
366 case BPF_JMP | BPF_JEQ | BPF_K:
367 case BPF_JMP | BPF_JNE | BPF_K:
368 case BPF_JMP | BPF_JGT | BPF_K:
369 case BPF_JMP | BPF_JGE | BPF_K:
370 case BPF_JMP | BPF_JSGT | BPF_K:
371 case BPF_JMP | BPF_JSGE | BPF_K:
372 case BPF_JMP | BPF_JSET | BPF_K:
373 /* Accommodate for extra offset in case of a backjump. */
377 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
378 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
379 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
382 case BPF_LD | BPF_ABS | BPF_W:
383 case BPF_LD | BPF_ABS | BPF_H:
384 case BPF_LD | BPF_ABS | BPF_B:
385 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
386 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
387 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
390 case BPF_LD | BPF_IND | BPF_W:
391 case BPF_LD | BPF_IND | BPF_H:
392 case BPF_LD | BPF_IND | BPF_B:
393 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
394 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
395 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
396 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
399 case BPF_LD | BPF_IMM | BPF_DW:
400 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
401 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
402 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
403 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
405 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
406 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
407 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
408 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
411 case BPF_ST | BPF_MEM | BPF_DW:
412 case BPF_ST | BPF_MEM | BPF_W:
413 case BPF_ST | BPF_MEM | BPF_H:
414 case BPF_ST | BPF_MEM | BPF_B:
415 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
416 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
417 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
424 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
425 gfp_t gfp_extra_flags)
427 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
431 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
433 kmemcheck_annotate_bitfield(fp, meta);
435 /* aux->prog still points to the fp_other one, so
436 * when promoting the clone to the real program,
437 * this still needs to be adapted.
439 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
445 static void bpf_prog_clone_free(struct bpf_prog *fp)
447 /* aux was stolen by the other clone, so we cannot free
448 * it from this path! It will be freed eventually by the
449 * other program on release.
451 * At this point, we don't need a deferred release since
452 * clone is guaranteed to not be locked.
458 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
460 /* We have to repoint aux->prog to self, as we don't
461 * know whether fp here is the clone or the original.
464 bpf_prog_clone_free(fp_other);
467 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
469 struct bpf_insn insn_buff[16], aux[2];
470 struct bpf_prog *clone, *tmp;
471 int insn_delta, insn_cnt;
472 struct bpf_insn *insn;
475 if (!bpf_jit_blinding_enabled())
478 clone = bpf_prog_clone_create(prog, GFP_USER);
480 return ERR_PTR(-ENOMEM);
482 insn_cnt = clone->len;
483 insn = clone->insnsi;
485 for (i = 0; i < insn_cnt; i++, insn++) {
486 /* We temporarily need to hold the original ld64 insn
487 * so that we can still access the first part in the
488 * second blinding run.
490 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
492 memcpy(aux, insn, sizeof(aux));
494 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
498 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
500 /* Patching may have repointed aux->prog during
501 * realloc from the original one, so we need to
502 * fix it up here on error.
504 bpf_jit_prog_release_other(prog, clone);
505 return ERR_PTR(-ENOMEM);
509 insn_delta = rewritten - 1;
511 /* Walk new program and skip insns we just inserted. */
512 insn = clone->insnsi + i + insn_delta;
513 insn_cnt += insn_delta;
519 #endif /* CONFIG_BPF_JIT */
521 /* Base function for offset calculation. Needs to go into .text section,
522 * therefore keeping it non-static as well; will also be used by JITs
523 * anyway later on, so do not let the compiler omit it.
525 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
529 EXPORT_SYMBOL_GPL(__bpf_call_base);
532 * __bpf_prog_run - run eBPF program on a given context
533 * @ctx: is the data we are operating on
534 * @insn: is the array of eBPF instructions
536 * Decode and execute eBPF instructions.
538 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
540 u64 stack[MAX_BPF_STACK / sizeof(u64)];
541 u64 regs[MAX_BPF_REG], tmp;
542 static const void *jumptable[256] = {
543 [0 ... 255] = &&default_label,
544 /* Now overwrite non-defaults ... */
545 /* 32 bit ALU operations */
546 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
547 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
548 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
549 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
550 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
551 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
552 [BPF_ALU | BPF_OR | BPF_X] = &&ALU_OR_X,
553 [BPF_ALU | BPF_OR | BPF_K] = &&ALU_OR_K,
554 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
555 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
556 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
557 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
558 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
559 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
560 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
561 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
562 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
563 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
564 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
565 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
566 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
567 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
568 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
569 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
570 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
571 /* 64 bit ALU operations */
572 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
573 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
574 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
575 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
576 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
577 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
578 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
579 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
580 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
581 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
582 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
583 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
584 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
585 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
586 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
587 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
588 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
589 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
590 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
591 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
592 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
593 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
594 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
595 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
596 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
597 /* Call instruction */
598 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
599 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
601 [BPF_JMP | BPF_JA] = &&JMP_JA,
602 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
603 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
604 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
605 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
606 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
607 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
608 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
609 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
610 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
611 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
612 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
613 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
614 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
615 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
617 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
618 /* Store instructions */
619 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
620 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
621 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
622 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
623 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
624 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
625 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
626 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
627 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
628 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
629 /* Load instructions */
630 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
631 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
632 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
633 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
634 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
635 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
636 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
637 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
638 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
639 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
640 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
642 u32 tail_call_cnt = 0;
646 #define CONT ({ insn++; goto select_insn; })
647 #define CONT_JMP ({ insn++; goto select_insn; })
649 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
650 ARG1 = (u64) (unsigned long) ctx;
653 goto *jumptable[insn->code];
656 #define ALU(OPCODE, OP) \
657 ALU64_##OPCODE##_X: \
661 DST = (u32) DST OP (u32) SRC; \
663 ALU64_##OPCODE##_K: \
667 DST = (u32) DST OP (u32) IMM; \
698 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
702 (*(s64 *) &DST) >>= SRC;
705 (*(s64 *) &DST) >>= IMM;
708 if (unlikely(SRC == 0))
710 div64_u64_rem(DST, SRC, &tmp);
714 if (unlikely(SRC == 0))
717 DST = do_div(tmp, (u32) SRC);
720 div64_u64_rem(DST, IMM, &tmp);
725 DST = do_div(tmp, (u32) IMM);
728 if (unlikely(SRC == 0))
730 DST = div64_u64(DST, SRC);
733 if (unlikely(SRC == 0))
736 do_div(tmp, (u32) SRC);
740 DST = div64_u64(DST, IMM);
744 do_div(tmp, (u32) IMM);
750 DST = (__force u16) cpu_to_be16(DST);
753 DST = (__force u32) cpu_to_be32(DST);
756 DST = (__force u64) cpu_to_be64(DST);
763 DST = (__force u16) cpu_to_le16(DST);
766 DST = (__force u32) cpu_to_le32(DST);
769 DST = (__force u64) cpu_to_le64(DST);
776 /* Function call scratches BPF_R1-BPF_R5 registers,
777 * preserves BPF_R6-BPF_R9, and stores return value
780 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
785 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
786 struct bpf_array *array = container_of(map, struct bpf_array, map);
787 struct bpf_prog *prog;
790 if (unlikely(index >= array->map.max_entries))
792 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
797 prog = READ_ONCE(array->ptrs[index]);
801 /* ARG1 at this point is guaranteed to point to CTX from
802 * the verifier side due to the fact that the tail call is
803 * handeled like a helper, that is, bpf_tail_call_proto,
804 * where arg1_type is ARG_PTR_TO_CTX.
864 if (((s64) DST) > ((s64) SRC)) {
870 if (((s64) DST) > ((s64) IMM)) {
876 if (((s64) DST) >= ((s64) SRC)) {
882 if (((s64) DST) >= ((s64) IMM)) {
902 /* STX and ST and LDX*/
903 #define LDST(SIZEOP, SIZE) \
905 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
908 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
911 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
919 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
920 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
923 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
924 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
927 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
930 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
931 * only appearing in the programs where ctx ==
932 * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
933 * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
934 * internal BPF verifier will check that BPF_R6 ==
937 * BPF_ABS and BPF_IND are wrappers of function calls,
938 * so they scratch BPF_R1-BPF_R5 registers, preserve
939 * BPF_R6-BPF_R9, and store return value into BPF_R0.
942 * ctx == skb == BPF_R6 == CTX
945 * SRC == any register
946 * IMM == 32-bit immediate
949 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
952 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
953 if (likely(ptr != NULL)) {
954 BPF_R0 = get_unaligned_be32(ptr);
959 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
962 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
963 if (likely(ptr != NULL)) {
964 BPF_R0 = get_unaligned_be16(ptr);
969 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
972 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
973 if (likely(ptr != NULL)) {
979 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
982 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
985 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
990 /* If we ever reach this, we have a bug somewhere. */
991 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
994 STACK_FRAME_NON_STANDARD(__bpf_prog_run); /* jump table */
996 bool bpf_prog_array_compatible(struct bpf_array *array,
997 const struct bpf_prog *fp)
999 if (!array->owner_prog_type) {
1000 /* There's no owner yet where we could check for
1003 array->owner_prog_type = fp->type;
1004 array->owner_jited = fp->jited;
1009 return array->owner_prog_type == fp->type &&
1010 array->owner_jited == fp->jited;
1013 static int bpf_check_tail_call(const struct bpf_prog *fp)
1015 struct bpf_prog_aux *aux = fp->aux;
1018 for (i = 0; i < aux->used_map_cnt; i++) {
1019 struct bpf_map *map = aux->used_maps[i];
1020 struct bpf_array *array;
1022 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1025 array = container_of(map, struct bpf_array, map);
1026 if (!bpf_prog_array_compatible(array, fp))
1034 * bpf_prog_select_runtime - select exec runtime for BPF program
1035 * @fp: bpf_prog populated with internal BPF program
1036 * @err: pointer to error variable
1038 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1039 * The BPF program will be executed via BPF_PROG_RUN() macro.
1041 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1043 fp->bpf_func = (void *) __bpf_prog_run;
1045 /* eBPF JITs can rewrite the program in case constant
1046 * blinding is active. However, in case of error during
1047 * blinding, bpf_int_jit_compile() must always return a
1048 * valid program, which in this case would simply not
1049 * be JITed, but falls back to the interpreter.
1051 fp = bpf_int_jit_compile(fp);
1052 bpf_prog_lock_ro(fp);
1054 /* The tail call compatibility check can only be done at
1055 * this late stage as we need to determine, if we deal
1056 * with JITed or non JITed program concatenations and not
1057 * all eBPF JITs might immediately support all features.
1059 *err = bpf_check_tail_call(fp);
1063 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1065 static void bpf_prog_free_deferred(struct work_struct *work)
1067 struct bpf_prog_aux *aux;
1069 aux = container_of(work, struct bpf_prog_aux, work);
1070 bpf_jit_free(aux->prog);
1073 /* Free internal BPF program */
1074 void bpf_prog_free(struct bpf_prog *fp)
1076 struct bpf_prog_aux *aux = fp->aux;
1078 INIT_WORK(&aux->work, bpf_prog_free_deferred);
1079 schedule_work(&aux->work);
1081 EXPORT_SYMBOL_GPL(bpf_prog_free);
1083 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1084 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1086 void bpf_user_rnd_init_once(void)
1088 prandom_init_once(&bpf_user_rnd_state);
1091 BPF_CALL_0(bpf_user_rnd_u32)
1093 /* Should someone ever have the rather unwise idea to use some
1094 * of the registers passed into this function, then note that
1095 * this function is called from native eBPF and classic-to-eBPF
1096 * transformations. Register assignments from both sides are
1097 * different, f.e. classic always sets fn(ctx, A, X) here.
1099 struct rnd_state *state;
1102 state = &get_cpu_var(bpf_user_rnd_state);
1103 res = prandom_u32_state(state);
1104 put_cpu_var(bpf_user_rnd_state);
1109 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1110 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1111 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1112 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1114 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1115 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1116 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1117 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1119 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1120 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1121 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1123 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1129 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1130 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1135 /* Always built-in helper functions. */
1136 const struct bpf_func_proto bpf_tail_call_proto = {
1139 .ret_type = RET_VOID,
1140 .arg1_type = ARG_PTR_TO_CTX,
1141 .arg2_type = ARG_CONST_MAP_PTR,
1142 .arg3_type = ARG_ANYTHING,
1145 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
1146 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1151 bool __weak bpf_helper_changes_pkt_data(void *func)
1156 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1157 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1159 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,