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>
31 #include <linux/rbtree_latch.h>
32 #include <linux/kallsyms.h>
33 #include <linux/rcupdate.h>
35 #include <asm/unaligned.h>
38 #define BPF_R0 regs[BPF_REG_0]
39 #define BPF_R1 regs[BPF_REG_1]
40 #define BPF_R2 regs[BPF_REG_2]
41 #define BPF_R3 regs[BPF_REG_3]
42 #define BPF_R4 regs[BPF_REG_4]
43 #define BPF_R5 regs[BPF_REG_5]
44 #define BPF_R6 regs[BPF_REG_6]
45 #define BPF_R7 regs[BPF_REG_7]
46 #define BPF_R8 regs[BPF_REG_8]
47 #define BPF_R9 regs[BPF_REG_9]
48 #define BPF_R10 regs[BPF_REG_10]
51 #define DST regs[insn->dst_reg]
52 #define SRC regs[insn->src_reg]
53 #define FP regs[BPF_REG_FP]
54 #define ARG1 regs[BPF_REG_ARG1]
55 #define CTX regs[BPF_REG_CTX]
58 /* No hurry in this branch
60 * Exported for the bpf jit load helper.
62 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
67 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
68 else if (k >= SKF_LL_OFF)
69 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
71 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
77 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
79 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
81 struct bpf_prog_aux *aux;
84 size = round_up(size, PAGE_SIZE);
85 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
89 kmemcheck_annotate_bitfield(fp, meta);
91 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
97 fp->pages = size / PAGE_SIZE;
101 INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode);
105 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
107 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
108 gfp_t gfp_extra_flags)
110 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
116 BUG_ON(fp_old == NULL);
118 size = round_up(size, PAGE_SIZE);
119 pages = size / PAGE_SIZE;
120 if (pages <= fp_old->pages)
123 delta = pages - fp_old->pages;
124 ret = __bpf_prog_charge(fp_old->aux->user, delta);
128 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
130 __bpf_prog_uncharge(fp_old->aux->user, delta);
132 kmemcheck_annotate_bitfield(fp, meta);
134 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
138 /* We keep fp->aux from fp_old around in the new
139 * reallocated structure.
142 __bpf_prog_free(fp_old);
148 void __bpf_prog_free(struct bpf_prog *fp)
154 int bpf_prog_calc_tag(struct bpf_prog *fp)
156 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
157 u32 raw_size = bpf_prog_tag_scratch_size(fp);
158 u32 digest[SHA_DIGEST_WORDS];
159 u32 ws[SHA_WORKSPACE_WORDS];
160 u32 i, bsize, psize, blocks;
161 struct bpf_insn *dst;
167 raw = vmalloc(raw_size);
172 memset(ws, 0, sizeof(ws));
174 /* We need to take out the map fd for the digest calculation
175 * since they are unstable from user space side.
178 for (i = 0, was_ld_map = false; i < fp->len; i++) {
179 dst[i] = fp->insnsi[i];
181 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
182 dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
185 } else if (was_ld_map &&
187 dst[i].dst_reg == 0 &&
188 dst[i].src_reg == 0 &&
197 psize = bpf_prog_insn_size(fp);
198 memset(&raw[psize], 0, raw_size - psize);
201 bsize = round_up(psize, SHA_MESSAGE_BYTES);
202 blocks = bsize / SHA_MESSAGE_BYTES;
204 if (bsize - psize >= sizeof(__be64)) {
205 bits = (__be64 *)(todo + bsize - sizeof(__be64));
207 bits = (__be64 *)(todo + bsize + bits_offset);
210 *bits = cpu_to_be64((psize - 1) << 3);
213 sha_transform(digest, todo, ws);
214 todo += SHA_MESSAGE_BYTES;
217 result = (__force __be32 *)digest;
218 for (i = 0; i < SHA_DIGEST_WORDS; i++)
219 result[i] = cpu_to_be32(digest[i]);
220 memcpy(fp->tag, result, sizeof(fp->tag));
226 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
228 return BPF_CLASS(insn->code) == BPF_JMP &&
229 /* Call and Exit are both special jumps with no
230 * target inside the BPF instruction image.
232 BPF_OP(insn->code) != BPF_CALL &&
233 BPF_OP(insn->code) != BPF_EXIT;
236 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
238 struct bpf_insn *insn = prog->insnsi;
239 u32 i, insn_cnt = prog->len;
241 for (i = 0; i < insn_cnt; i++, insn++) {
242 if (!bpf_is_jmp_and_has_target(insn))
245 /* Adjust offset of jmps if we cross boundaries. */
246 if (i < pos && i + insn->off + 1 > pos)
248 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
253 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
254 const struct bpf_insn *patch, u32 len)
256 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
257 struct bpf_prog *prog_adj;
259 /* Since our patchlet doesn't expand the image, we're done. */
260 if (insn_delta == 0) {
261 memcpy(prog->insnsi + off, patch, sizeof(*patch));
265 insn_adj_cnt = prog->len + insn_delta;
267 /* Several new instructions need to be inserted. Make room
268 * for them. Likely, there's no need for a new allocation as
269 * last page could have large enough tailroom.
271 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
276 prog_adj->len = insn_adj_cnt;
278 /* Patching happens in 3 steps:
280 * 1) Move over tail of insnsi from next instruction onwards,
281 * so we can patch the single target insn with one or more
282 * new ones (patching is always from 1 to n insns, n > 0).
283 * 2) Inject new instructions at the target location.
284 * 3) Adjust branch offsets if necessary.
286 insn_rest = insn_adj_cnt - off - len;
288 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
289 sizeof(*patch) * insn_rest);
290 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
292 bpf_adj_branches(prog_adj, off, insn_delta);
297 #ifdef CONFIG_BPF_JIT
298 static __always_inline void
299 bpf_get_prog_addr_region(const struct bpf_prog *prog,
300 unsigned long *symbol_start,
301 unsigned long *symbol_end)
303 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
304 unsigned long addr = (unsigned long)hdr;
306 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
308 *symbol_start = addr;
309 *symbol_end = addr + hdr->pages * PAGE_SIZE;
312 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym)
314 BUILD_BUG_ON(sizeof("bpf_prog_") +
315 sizeof(prog->tag) * 2 + 1 > KSYM_NAME_LEN);
317 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
318 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
322 static __always_inline unsigned long
323 bpf_get_prog_addr_start(struct latch_tree_node *n)
325 unsigned long symbol_start, symbol_end;
326 const struct bpf_prog_aux *aux;
328 aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
329 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
334 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
335 struct latch_tree_node *b)
337 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b);
340 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
342 unsigned long val = (unsigned long)key;
343 unsigned long symbol_start, symbol_end;
344 const struct bpf_prog_aux *aux;
346 aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
347 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
349 if (val < symbol_start)
351 if (val >= symbol_end)
357 static const struct latch_tree_ops bpf_tree_ops = {
358 .less = bpf_tree_less,
359 .comp = bpf_tree_comp,
362 static DEFINE_SPINLOCK(bpf_lock);
363 static LIST_HEAD(bpf_kallsyms);
364 static struct latch_tree_root bpf_tree __cacheline_aligned;
366 int bpf_jit_kallsyms __read_mostly;
368 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux)
370 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode));
371 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms);
372 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
375 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux)
377 if (list_empty(&aux->ksym_lnode))
380 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
381 list_del_rcu(&aux->ksym_lnode);
384 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
386 return fp->jited && !bpf_prog_was_classic(fp);
389 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
391 return list_empty(&fp->aux->ksym_lnode) ||
392 fp->aux->ksym_lnode.prev == LIST_POISON2;
395 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
397 if (!bpf_prog_kallsyms_candidate(fp) ||
398 !capable(CAP_SYS_ADMIN))
401 spin_lock_bh(&bpf_lock);
402 bpf_prog_ksym_node_add(fp->aux);
403 spin_unlock_bh(&bpf_lock);
406 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
408 if (!bpf_prog_kallsyms_candidate(fp))
411 spin_lock_bh(&bpf_lock);
412 bpf_prog_ksym_node_del(fp->aux);
413 spin_unlock_bh(&bpf_lock);
416 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr)
418 struct latch_tree_node *n;
420 if (!bpf_jit_kallsyms_enabled())
423 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
425 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog :
429 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
430 unsigned long *off, char *sym)
432 unsigned long symbol_start, symbol_end;
433 struct bpf_prog *prog;
437 prog = bpf_prog_kallsyms_find(addr);
439 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end);
440 bpf_get_prog_name(prog, sym);
444 *size = symbol_end - symbol_start;
446 *off = addr - symbol_start;
453 bool is_bpf_text_address(unsigned long addr)
458 ret = bpf_prog_kallsyms_find(addr) != NULL;
464 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
467 unsigned long symbol_start, symbol_end;
468 struct bpf_prog_aux *aux;
472 if (!bpf_jit_kallsyms_enabled())
476 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) {
480 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
481 bpf_get_prog_name(aux->prog, sym);
483 *value = symbol_start;
484 *type = BPF_SYM_ELF_TYPE;
494 struct bpf_binary_header *
495 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
496 unsigned int alignment,
497 bpf_jit_fill_hole_t bpf_fill_ill_insns)
499 struct bpf_binary_header *hdr;
500 unsigned int size, hole, start;
502 /* Most of BPF filters are really small, but if some of them
503 * fill a page, allow at least 128 extra bytes to insert a
504 * random section of illegal instructions.
506 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
507 hdr = module_alloc(size);
511 /* Fill space with illegal/arch-dep instructions. */
512 bpf_fill_ill_insns(hdr, size);
514 hdr->pages = size / PAGE_SIZE;
515 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
516 PAGE_SIZE - sizeof(*hdr));
517 start = (get_random_int() % hole) & ~(alignment - 1);
519 /* Leave a random number of instructions before BPF code. */
520 *image_ptr = &hdr->image[start];
525 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
530 /* This symbol is only overridden by archs that have different
531 * requirements than the usual eBPF JITs, f.e. when they only
532 * implement cBPF JIT, do not set images read-only, etc.
534 void __weak bpf_jit_free(struct bpf_prog *fp)
537 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
539 bpf_jit_binary_unlock_ro(hdr);
540 bpf_jit_binary_free(hdr);
542 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
545 bpf_prog_unlock_free(fp);
548 int bpf_jit_harden __read_mostly;
550 static int bpf_jit_blind_insn(const struct bpf_insn *from,
551 const struct bpf_insn *aux,
552 struct bpf_insn *to_buff)
554 struct bpf_insn *to = to_buff;
555 u32 imm_rnd = get_random_int();
558 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
559 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
561 if (from->imm == 0 &&
562 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
563 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
564 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
568 switch (from->code) {
569 case BPF_ALU | BPF_ADD | BPF_K:
570 case BPF_ALU | BPF_SUB | BPF_K:
571 case BPF_ALU | BPF_AND | BPF_K:
572 case BPF_ALU | BPF_OR | BPF_K:
573 case BPF_ALU | BPF_XOR | BPF_K:
574 case BPF_ALU | BPF_MUL | BPF_K:
575 case BPF_ALU | BPF_MOV | BPF_K:
576 case BPF_ALU | BPF_DIV | BPF_K:
577 case BPF_ALU | BPF_MOD | BPF_K:
578 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
579 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
580 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
583 case BPF_ALU64 | BPF_ADD | BPF_K:
584 case BPF_ALU64 | BPF_SUB | BPF_K:
585 case BPF_ALU64 | BPF_AND | BPF_K:
586 case BPF_ALU64 | BPF_OR | BPF_K:
587 case BPF_ALU64 | BPF_XOR | BPF_K:
588 case BPF_ALU64 | BPF_MUL | BPF_K:
589 case BPF_ALU64 | BPF_MOV | BPF_K:
590 case BPF_ALU64 | BPF_DIV | BPF_K:
591 case BPF_ALU64 | BPF_MOD | BPF_K:
592 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
593 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
594 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
597 case BPF_JMP | BPF_JEQ | BPF_K:
598 case BPF_JMP | BPF_JNE | BPF_K:
599 case BPF_JMP | BPF_JGT | BPF_K:
600 case BPF_JMP | BPF_JGE | BPF_K:
601 case BPF_JMP | BPF_JSGT | BPF_K:
602 case BPF_JMP | BPF_JSGE | BPF_K:
603 case BPF_JMP | BPF_JSET | BPF_K:
604 /* Accommodate for extra offset in case of a backjump. */
608 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
609 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
610 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
613 case BPF_LD | BPF_ABS | BPF_W:
614 case BPF_LD | BPF_ABS | BPF_H:
615 case BPF_LD | BPF_ABS | BPF_B:
616 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
617 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
618 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
621 case BPF_LD | BPF_IND | BPF_W:
622 case BPF_LD | BPF_IND | BPF_H:
623 case BPF_LD | BPF_IND | BPF_B:
624 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
625 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
626 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
627 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
630 case BPF_LD | BPF_IMM | BPF_DW:
631 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
632 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
633 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
634 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
636 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
637 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
638 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
639 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
642 case BPF_ST | BPF_MEM | BPF_DW:
643 case BPF_ST | BPF_MEM | BPF_W:
644 case BPF_ST | BPF_MEM | BPF_H:
645 case BPF_ST | BPF_MEM | BPF_B:
646 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
647 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
648 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
655 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
656 gfp_t gfp_extra_flags)
658 gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
662 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
664 kmemcheck_annotate_bitfield(fp, meta);
666 /* aux->prog still points to the fp_other one, so
667 * when promoting the clone to the real program,
668 * this still needs to be adapted.
670 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
676 static void bpf_prog_clone_free(struct bpf_prog *fp)
678 /* aux was stolen by the other clone, so we cannot free
679 * it from this path! It will be freed eventually by the
680 * other program on release.
682 * At this point, we don't need a deferred release since
683 * clone is guaranteed to not be locked.
689 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
691 /* We have to repoint aux->prog to self, as we don't
692 * know whether fp here is the clone or the original.
695 bpf_prog_clone_free(fp_other);
698 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
700 struct bpf_insn insn_buff[16], aux[2];
701 struct bpf_prog *clone, *tmp;
702 int insn_delta, insn_cnt;
703 struct bpf_insn *insn;
706 if (!bpf_jit_blinding_enabled())
709 clone = bpf_prog_clone_create(prog, GFP_USER);
711 return ERR_PTR(-ENOMEM);
713 insn_cnt = clone->len;
714 insn = clone->insnsi;
716 for (i = 0; i < insn_cnt; i++, insn++) {
717 /* We temporarily need to hold the original ld64 insn
718 * so that we can still access the first part in the
719 * second blinding run.
721 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
723 memcpy(aux, insn, sizeof(aux));
725 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
729 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
731 /* Patching may have repointed aux->prog during
732 * realloc from the original one, so we need to
733 * fix it up here on error.
735 bpf_jit_prog_release_other(prog, clone);
736 return ERR_PTR(-ENOMEM);
740 insn_delta = rewritten - 1;
742 /* Walk new program and skip insns we just inserted. */
743 insn = clone->insnsi + i + insn_delta;
744 insn_cnt += insn_delta;
750 #endif /* CONFIG_BPF_JIT */
752 /* Base function for offset calculation. Needs to go into .text section,
753 * therefore keeping it non-static as well; will also be used by JITs
754 * anyway later on, so do not let the compiler omit it.
756 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
760 EXPORT_SYMBOL_GPL(__bpf_call_base);
763 * __bpf_prog_run - run eBPF program on a given context
764 * @ctx: is the data we are operating on
765 * @insn: is the array of eBPF instructions
767 * Decode and execute eBPF instructions.
769 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
771 u64 stack[MAX_BPF_STACK / sizeof(u64)];
772 u64 regs[MAX_BPF_REG], tmp;
773 static const void *jumptable[256] = {
774 [0 ... 255] = &&default_label,
775 /* Now overwrite non-defaults ... */
776 /* 32 bit ALU operations */
777 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
778 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
779 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
780 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
781 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
782 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
783 [BPF_ALU | BPF_OR | BPF_X] = &&ALU_OR_X,
784 [BPF_ALU | BPF_OR | BPF_K] = &&ALU_OR_K,
785 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
786 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
787 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
788 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
789 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
790 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
791 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
792 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
793 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
794 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
795 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
796 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
797 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
798 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
799 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
800 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
801 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
802 /* 64 bit ALU operations */
803 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
804 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
805 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
806 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
807 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
808 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
809 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
810 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
811 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
812 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
813 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
814 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
815 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
816 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
817 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
818 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
819 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
820 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
821 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
822 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
823 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
824 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
825 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
826 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
827 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
828 /* Call instruction */
829 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
830 [BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
832 [BPF_JMP | BPF_JA] = &&JMP_JA,
833 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
834 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
835 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
836 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
837 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
838 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
839 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
840 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
841 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
842 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
843 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
844 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
845 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
846 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
848 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
849 /* Store instructions */
850 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
851 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
852 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
853 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
854 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
855 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
856 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
857 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
858 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
859 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
860 /* Load instructions */
861 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
862 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
863 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
864 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
865 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
866 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
867 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
868 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
869 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
870 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
871 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
873 u32 tail_call_cnt = 0;
877 #define CONT ({ insn++; goto select_insn; })
878 #define CONT_JMP ({ insn++; goto select_insn; })
880 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
881 ARG1 = (u64) (unsigned long) ctx;
884 goto *jumptable[insn->code];
887 #define ALU(OPCODE, OP) \
888 ALU64_##OPCODE##_X: \
892 DST = (u32) DST OP (u32) SRC; \
894 ALU64_##OPCODE##_K: \
898 DST = (u32) DST OP (u32) IMM; \
929 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
933 (*(s64 *) &DST) >>= SRC;
936 (*(s64 *) &DST) >>= IMM;
939 if (unlikely(SRC == 0))
941 div64_u64_rem(DST, SRC, &tmp);
945 if (unlikely(SRC == 0))
948 DST = do_div(tmp, (u32) SRC);
951 div64_u64_rem(DST, IMM, &tmp);
956 DST = do_div(tmp, (u32) IMM);
959 if (unlikely(SRC == 0))
961 DST = div64_u64(DST, SRC);
964 if (unlikely(SRC == 0))
967 do_div(tmp, (u32) SRC);
971 DST = div64_u64(DST, IMM);
975 do_div(tmp, (u32) IMM);
981 DST = (__force u16) cpu_to_be16(DST);
984 DST = (__force u32) cpu_to_be32(DST);
987 DST = (__force u64) cpu_to_be64(DST);
994 DST = (__force u16) cpu_to_le16(DST);
997 DST = (__force u32) cpu_to_le32(DST);
1000 DST = (__force u64) cpu_to_le64(DST);
1007 /* Function call scratches BPF_R1-BPF_R5 registers,
1008 * preserves BPF_R6-BPF_R9, and stores return value
1011 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1016 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1017 struct bpf_array *array = container_of(map, struct bpf_array, map);
1018 struct bpf_prog *prog;
1021 if (unlikely(index >= array->map.max_entries))
1023 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1028 prog = READ_ONCE(array->ptrs[index]);
1032 /* ARG1 at this point is guaranteed to point to CTX from
1033 * the verifier side due to the fact that the tail call is
1034 * handeled like a helper, that is, bpf_tail_call_proto,
1035 * where arg1_type is ARG_PTR_TO_CTX.
1037 insn = prog->insnsi;
1095 if (((s64) DST) > ((s64) SRC)) {
1101 if (((s64) DST) > ((s64) IMM)) {
1107 if (((s64) DST) >= ((s64) SRC)) {
1113 if (((s64) DST) >= ((s64) IMM)) {
1133 /* STX and ST and LDX*/
1134 #define LDST(SIZEOP, SIZE) \
1136 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1139 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1142 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1150 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
1151 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
1154 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
1155 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
1158 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
1161 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are only
1162 * appearing in the programs where ctx == skb
1163 * (see may_access_skb() in the verifier). All programs
1164 * keep 'ctx' in regs[BPF_REG_CTX] == BPF_R6,
1165 * bpf_convert_filter() saves it in BPF_R6, internal BPF
1166 * verifier will check that BPF_R6 == ctx.
1168 * BPF_ABS and BPF_IND are wrappers of function calls,
1169 * so they scratch BPF_R1-BPF_R5 registers, preserve
1170 * BPF_R6-BPF_R9, and store return value into BPF_R0.
1173 * ctx == skb == BPF_R6 == CTX
1176 * SRC == any register
1177 * IMM == 32-bit immediate
1180 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
1183 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
1184 if (likely(ptr != NULL)) {
1185 BPF_R0 = get_unaligned_be32(ptr);
1190 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
1193 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
1194 if (likely(ptr != NULL)) {
1195 BPF_R0 = get_unaligned_be16(ptr);
1200 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
1203 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
1204 if (likely(ptr != NULL)) {
1205 BPF_R0 = *(u8 *)ptr;
1210 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
1213 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
1216 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
1221 /* If we ever reach this, we have a bug somewhere. */
1222 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
1225 STACK_FRAME_NON_STANDARD(__bpf_prog_run); /* jump table */
1227 bool bpf_prog_array_compatible(struct bpf_array *array,
1228 const struct bpf_prog *fp)
1230 if (!array->owner_prog_type) {
1231 /* There's no owner yet where we could check for
1234 array->owner_prog_type = fp->type;
1235 array->owner_jited = fp->jited;
1240 return array->owner_prog_type == fp->type &&
1241 array->owner_jited == fp->jited;
1244 static int bpf_check_tail_call(const struct bpf_prog *fp)
1246 struct bpf_prog_aux *aux = fp->aux;
1249 for (i = 0; i < aux->used_map_cnt; i++) {
1250 struct bpf_map *map = aux->used_maps[i];
1251 struct bpf_array *array;
1253 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1256 array = container_of(map, struct bpf_array, map);
1257 if (!bpf_prog_array_compatible(array, fp))
1265 * bpf_prog_select_runtime - select exec runtime for BPF program
1266 * @fp: bpf_prog populated with internal BPF program
1267 * @err: pointer to error variable
1269 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1270 * The BPF program will be executed via BPF_PROG_RUN() macro.
1272 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1274 fp->bpf_func = (void *) __bpf_prog_run;
1276 /* eBPF JITs can rewrite the program in case constant
1277 * blinding is active. However, in case of error during
1278 * blinding, bpf_int_jit_compile() must always return a
1279 * valid program, which in this case would simply not
1280 * be JITed, but falls back to the interpreter.
1282 fp = bpf_int_jit_compile(fp);
1283 bpf_prog_lock_ro(fp);
1285 /* The tail call compatibility check can only be done at
1286 * this late stage as we need to determine, if we deal
1287 * with JITed or non JITed program concatenations and not
1288 * all eBPF JITs might immediately support all features.
1290 *err = bpf_check_tail_call(fp);
1294 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1296 static void bpf_prog_free_deferred(struct work_struct *work)
1298 struct bpf_prog_aux *aux;
1300 aux = container_of(work, struct bpf_prog_aux, work);
1301 bpf_jit_free(aux->prog);
1304 /* Free internal BPF program */
1305 void bpf_prog_free(struct bpf_prog *fp)
1307 struct bpf_prog_aux *aux = fp->aux;
1309 INIT_WORK(&aux->work, bpf_prog_free_deferred);
1310 schedule_work(&aux->work);
1312 EXPORT_SYMBOL_GPL(bpf_prog_free);
1314 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1315 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1317 void bpf_user_rnd_init_once(void)
1319 prandom_init_once(&bpf_user_rnd_state);
1322 BPF_CALL_0(bpf_user_rnd_u32)
1324 /* Should someone ever have the rather unwise idea to use some
1325 * of the registers passed into this function, then note that
1326 * this function is called from native eBPF and classic-to-eBPF
1327 * transformations. Register assignments from both sides are
1328 * different, f.e. classic always sets fn(ctx, A, X) here.
1330 struct rnd_state *state;
1333 state = &get_cpu_var(bpf_user_rnd_state);
1334 res = prandom_u32_state(state);
1335 put_cpu_var(bpf_user_rnd_state);
1340 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1341 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1342 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1343 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1345 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1346 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1347 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1348 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1350 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1351 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1352 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1354 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1360 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1361 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1366 /* Always built-in helper functions. */
1367 const struct bpf_func_proto bpf_tail_call_proto = {
1370 .ret_type = RET_VOID,
1371 .arg1_type = ARG_PTR_TO_CTX,
1372 .arg2_type = ARG_CONST_MAP_PTR,
1373 .arg3_type = ARG_ANYTHING,
1376 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
1377 * It is encouraged to implement bpf_int_jit_compile() instead, so that
1378 * eBPF and implicitly also cBPF can get JITed!
1380 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1385 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
1386 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
1388 void __weak bpf_jit_compile(struct bpf_prog *prog)
1392 bool __weak bpf_helper_changes_pkt_data(void *func)
1397 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1398 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1400 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
1406 /* All definitions of tracepoints related to BPF. */
1407 #define CREATE_TRACE_POINTS
1408 #include <linux/bpf_trace.h>
1410 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
1412 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_get_type);
1413 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_put_rcu);