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_ZERO | gfp_extra_flags;
80 struct bpf_prog_aux *aux;
83 size = round_up(size, PAGE_SIZE);
84 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
88 kmemcheck_annotate_bitfield(fp, meta);
90 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
96 fp->pages = size / PAGE_SIZE;
100 INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode);
104 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
106 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
107 gfp_t gfp_extra_flags)
109 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
114 BUG_ON(fp_old == NULL);
116 size = round_up(size, PAGE_SIZE);
117 pages = size / PAGE_SIZE;
118 if (pages <= fp_old->pages)
121 delta = pages - fp_old->pages;
122 ret = __bpf_prog_charge(fp_old->aux->user, delta);
126 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
128 __bpf_prog_uncharge(fp_old->aux->user, delta);
130 kmemcheck_annotate_bitfield(fp, meta);
132 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
136 /* We keep fp->aux from fp_old around in the new
137 * reallocated structure.
140 __bpf_prog_free(fp_old);
146 void __bpf_prog_free(struct bpf_prog *fp)
152 int bpf_prog_calc_tag(struct bpf_prog *fp)
154 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
155 u32 raw_size = bpf_prog_tag_scratch_size(fp);
156 u32 digest[SHA_DIGEST_WORDS];
157 u32 ws[SHA_WORKSPACE_WORDS];
158 u32 i, bsize, psize, blocks;
159 struct bpf_insn *dst;
165 raw = vmalloc(raw_size);
170 memset(ws, 0, sizeof(ws));
172 /* We need to take out the map fd for the digest calculation
173 * since they are unstable from user space side.
176 for (i = 0, was_ld_map = false; i < fp->len; i++) {
177 dst[i] = fp->insnsi[i];
179 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
180 dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
183 } else if (was_ld_map &&
185 dst[i].dst_reg == 0 &&
186 dst[i].src_reg == 0 &&
195 psize = bpf_prog_insn_size(fp);
196 memset(&raw[psize], 0, raw_size - psize);
199 bsize = round_up(psize, SHA_MESSAGE_BYTES);
200 blocks = bsize / SHA_MESSAGE_BYTES;
202 if (bsize - psize >= sizeof(__be64)) {
203 bits = (__be64 *)(todo + bsize - sizeof(__be64));
205 bits = (__be64 *)(todo + bsize + bits_offset);
208 *bits = cpu_to_be64((psize - 1) << 3);
211 sha_transform(digest, todo, ws);
212 todo += SHA_MESSAGE_BYTES;
215 result = (__force __be32 *)digest;
216 for (i = 0; i < SHA_DIGEST_WORDS; i++)
217 result[i] = cpu_to_be32(digest[i]);
218 memcpy(fp->tag, result, sizeof(fp->tag));
224 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
226 return BPF_CLASS(insn->code) == BPF_JMP &&
227 /* Call and Exit are both special jumps with no
228 * target inside the BPF instruction image.
230 BPF_OP(insn->code) != BPF_CALL &&
231 BPF_OP(insn->code) != BPF_EXIT;
234 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
236 struct bpf_insn *insn = prog->insnsi;
237 u32 i, insn_cnt = prog->len;
239 for (i = 0; i < insn_cnt; i++, insn++) {
240 if (!bpf_is_jmp_and_has_target(insn))
243 /* Adjust offset of jmps if we cross boundaries. */
244 if (i < pos && i + insn->off + 1 > pos)
246 else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
251 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
252 const struct bpf_insn *patch, u32 len)
254 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
255 struct bpf_prog *prog_adj;
257 /* Since our patchlet doesn't expand the image, we're done. */
258 if (insn_delta == 0) {
259 memcpy(prog->insnsi + off, patch, sizeof(*patch));
263 insn_adj_cnt = prog->len + insn_delta;
265 /* Several new instructions need to be inserted. Make room
266 * for them. Likely, there's no need for a new allocation as
267 * last page could have large enough tailroom.
269 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
274 prog_adj->len = insn_adj_cnt;
276 /* Patching happens in 3 steps:
278 * 1) Move over tail of insnsi from next instruction onwards,
279 * so we can patch the single target insn with one or more
280 * new ones (patching is always from 1 to n insns, n > 0).
281 * 2) Inject new instructions at the target location.
282 * 3) Adjust branch offsets if necessary.
284 insn_rest = insn_adj_cnt - off - len;
286 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
287 sizeof(*patch) * insn_rest);
288 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
290 bpf_adj_branches(prog_adj, off, insn_delta);
295 #ifdef CONFIG_BPF_JIT
296 static __always_inline void
297 bpf_get_prog_addr_region(const struct bpf_prog *prog,
298 unsigned long *symbol_start,
299 unsigned long *symbol_end)
301 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
302 unsigned long addr = (unsigned long)hdr;
304 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
306 *symbol_start = addr;
307 *symbol_end = addr + hdr->pages * PAGE_SIZE;
310 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym)
312 BUILD_BUG_ON(sizeof("bpf_prog_") +
313 sizeof(prog->tag) * 2 + 1 > KSYM_NAME_LEN);
315 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
316 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
320 static __always_inline unsigned long
321 bpf_get_prog_addr_start(struct latch_tree_node *n)
323 unsigned long symbol_start, symbol_end;
324 const struct bpf_prog_aux *aux;
326 aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
327 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
332 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
333 struct latch_tree_node *b)
335 return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b);
338 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
340 unsigned long val = (unsigned long)key;
341 unsigned long symbol_start, symbol_end;
342 const struct bpf_prog_aux *aux;
344 aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
345 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
347 if (val < symbol_start)
349 if (val >= symbol_end)
355 static const struct latch_tree_ops bpf_tree_ops = {
356 .less = bpf_tree_less,
357 .comp = bpf_tree_comp,
360 static DEFINE_SPINLOCK(bpf_lock);
361 static LIST_HEAD(bpf_kallsyms);
362 static struct latch_tree_root bpf_tree __cacheline_aligned;
364 int bpf_jit_kallsyms __read_mostly;
366 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux)
368 WARN_ON_ONCE(!list_empty(&aux->ksym_lnode));
369 list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms);
370 latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
373 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux)
375 if (list_empty(&aux->ksym_lnode))
378 latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
379 list_del_rcu(&aux->ksym_lnode);
382 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
384 return fp->jited && !bpf_prog_was_classic(fp);
387 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
389 return list_empty(&fp->aux->ksym_lnode) ||
390 fp->aux->ksym_lnode.prev == LIST_POISON2;
393 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
395 if (!bpf_prog_kallsyms_candidate(fp) ||
396 !capable(CAP_SYS_ADMIN))
399 spin_lock_bh(&bpf_lock);
400 bpf_prog_ksym_node_add(fp->aux);
401 spin_unlock_bh(&bpf_lock);
404 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
406 if (!bpf_prog_kallsyms_candidate(fp))
409 spin_lock_bh(&bpf_lock);
410 bpf_prog_ksym_node_del(fp->aux);
411 spin_unlock_bh(&bpf_lock);
414 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr)
416 struct latch_tree_node *n;
418 if (!bpf_jit_kallsyms_enabled())
421 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
423 container_of(n, struct bpf_prog_aux, ksym_tnode)->prog :
427 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
428 unsigned long *off, char *sym)
430 unsigned long symbol_start, symbol_end;
431 struct bpf_prog *prog;
435 prog = bpf_prog_kallsyms_find(addr);
437 bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end);
438 bpf_get_prog_name(prog, sym);
442 *size = symbol_end - symbol_start;
444 *off = addr - symbol_start;
451 bool is_bpf_text_address(unsigned long addr)
456 ret = bpf_prog_kallsyms_find(addr) != NULL;
462 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
465 unsigned long symbol_start, symbol_end;
466 struct bpf_prog_aux *aux;
470 if (!bpf_jit_kallsyms_enabled())
474 list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) {
478 bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
479 bpf_get_prog_name(aux->prog, sym);
481 *value = symbol_start;
482 *type = BPF_SYM_ELF_TYPE;
492 struct bpf_binary_header *
493 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
494 unsigned int alignment,
495 bpf_jit_fill_hole_t bpf_fill_ill_insns)
497 struct bpf_binary_header *hdr;
498 unsigned int size, hole, start;
500 /* Most of BPF filters are really small, but if some of them
501 * fill a page, allow at least 128 extra bytes to insert a
502 * random section of illegal instructions.
504 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
505 hdr = module_alloc(size);
509 /* Fill space with illegal/arch-dep instructions. */
510 bpf_fill_ill_insns(hdr, size);
512 hdr->pages = size / PAGE_SIZE;
513 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
514 PAGE_SIZE - sizeof(*hdr));
515 start = (get_random_int() % hole) & ~(alignment - 1);
517 /* Leave a random number of instructions before BPF code. */
518 *image_ptr = &hdr->image[start];
523 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
528 /* This symbol is only overridden by archs that have different
529 * requirements than the usual eBPF JITs, f.e. when they only
530 * implement cBPF JIT, do not set images read-only, etc.
532 void __weak bpf_jit_free(struct bpf_prog *fp)
535 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
537 bpf_jit_binary_unlock_ro(hdr);
538 bpf_jit_binary_free(hdr);
540 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
543 bpf_prog_unlock_free(fp);
546 int bpf_jit_harden __read_mostly;
548 static int bpf_jit_blind_insn(const struct bpf_insn *from,
549 const struct bpf_insn *aux,
550 struct bpf_insn *to_buff)
552 struct bpf_insn *to = to_buff;
553 u32 imm_rnd = get_random_int();
556 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
557 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
559 if (from->imm == 0 &&
560 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
561 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
562 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
566 switch (from->code) {
567 case BPF_ALU | BPF_ADD | BPF_K:
568 case BPF_ALU | BPF_SUB | BPF_K:
569 case BPF_ALU | BPF_AND | BPF_K:
570 case BPF_ALU | BPF_OR | BPF_K:
571 case BPF_ALU | BPF_XOR | BPF_K:
572 case BPF_ALU | BPF_MUL | BPF_K:
573 case BPF_ALU | BPF_MOV | BPF_K:
574 case BPF_ALU | BPF_DIV | BPF_K:
575 case BPF_ALU | BPF_MOD | BPF_K:
576 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
577 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
578 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
581 case BPF_ALU64 | BPF_ADD | BPF_K:
582 case BPF_ALU64 | BPF_SUB | BPF_K:
583 case BPF_ALU64 | BPF_AND | BPF_K:
584 case BPF_ALU64 | BPF_OR | BPF_K:
585 case BPF_ALU64 | BPF_XOR | BPF_K:
586 case BPF_ALU64 | BPF_MUL | BPF_K:
587 case BPF_ALU64 | BPF_MOV | BPF_K:
588 case BPF_ALU64 | BPF_DIV | BPF_K:
589 case BPF_ALU64 | BPF_MOD | BPF_K:
590 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
591 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
592 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
595 case BPF_JMP | BPF_JEQ | BPF_K:
596 case BPF_JMP | BPF_JNE | BPF_K:
597 case BPF_JMP | BPF_JGT | BPF_K:
598 case BPF_JMP | BPF_JGE | BPF_K:
599 case BPF_JMP | BPF_JSGT | BPF_K:
600 case BPF_JMP | BPF_JSGE | BPF_K:
601 case BPF_JMP | BPF_JSET | BPF_K:
602 /* Accommodate for extra offset in case of a backjump. */
606 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
607 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
608 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
611 case BPF_LD | BPF_ABS | BPF_W:
612 case BPF_LD | BPF_ABS | BPF_H:
613 case BPF_LD | BPF_ABS | BPF_B:
614 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
615 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
616 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
619 case BPF_LD | BPF_IND | BPF_W:
620 case BPF_LD | BPF_IND | BPF_H:
621 case BPF_LD | BPF_IND | BPF_B:
622 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
623 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
624 *to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
625 *to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
628 case BPF_LD | BPF_IMM | BPF_DW:
629 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
630 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
631 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
632 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
634 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
635 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
636 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
637 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
640 case BPF_ST | BPF_MEM | BPF_DW:
641 case BPF_ST | BPF_MEM | BPF_W:
642 case BPF_ST | BPF_MEM | BPF_H:
643 case BPF_ST | BPF_MEM | BPF_B:
644 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
645 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
646 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
653 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
654 gfp_t gfp_extra_flags)
656 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
659 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
661 kmemcheck_annotate_bitfield(fp, meta);
663 /* aux->prog still points to the fp_other one, so
664 * when promoting the clone to the real program,
665 * this still needs to be adapted.
667 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
673 static void bpf_prog_clone_free(struct bpf_prog *fp)
675 /* aux was stolen by the other clone, so we cannot free
676 * it from this path! It will be freed eventually by the
677 * other program on release.
679 * At this point, we don't need a deferred release since
680 * clone is guaranteed to not be locked.
686 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
688 /* We have to repoint aux->prog to self, as we don't
689 * know whether fp here is the clone or the original.
692 bpf_prog_clone_free(fp_other);
695 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
697 struct bpf_insn insn_buff[16], aux[2];
698 struct bpf_prog *clone, *tmp;
699 int insn_delta, insn_cnt;
700 struct bpf_insn *insn;
703 if (!bpf_jit_blinding_enabled())
706 clone = bpf_prog_clone_create(prog, GFP_USER);
708 return ERR_PTR(-ENOMEM);
710 insn_cnt = clone->len;
711 insn = clone->insnsi;
713 for (i = 0; i < insn_cnt; i++, insn++) {
714 /* We temporarily need to hold the original ld64 insn
715 * so that we can still access the first part in the
716 * second blinding run.
718 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
720 memcpy(aux, insn, sizeof(aux));
722 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
726 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
728 /* Patching may have repointed aux->prog during
729 * realloc from the original one, so we need to
730 * fix it up here on error.
732 bpf_jit_prog_release_other(prog, clone);
733 return ERR_PTR(-ENOMEM);
737 insn_delta = rewritten - 1;
739 /* Walk new program and skip insns we just inserted. */
740 insn = clone->insnsi + i + insn_delta;
741 insn_cnt += insn_delta;
747 #endif /* CONFIG_BPF_JIT */
749 /* Base function for offset calculation. Needs to go into .text section,
750 * therefore keeping it non-static as well; will also be used by JITs
751 * anyway later on, so do not let the compiler omit it.
753 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
757 EXPORT_SYMBOL_GPL(__bpf_call_base);
760 * __bpf_prog_run - run eBPF program on a given context
761 * @ctx: is the data we are operating on
762 * @insn: is the array of eBPF instructions
764 * Decode and execute eBPF instructions.
766 static unsigned int ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn,
770 static const void *jumptable[256] = {
771 [0 ... 255] = &&default_label,
772 /* Now overwrite non-defaults ... */
773 /* 32 bit ALU operations */
774 [BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
775 [BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
776 [BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
777 [BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
778 [BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
779 [BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
780 [BPF_ALU | BPF_OR | BPF_X] = &&ALU_OR_X,
781 [BPF_ALU | BPF_OR | BPF_K] = &&ALU_OR_K,
782 [BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
783 [BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
784 [BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
785 [BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
786 [BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
787 [BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
788 [BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
789 [BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
790 [BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
791 [BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
792 [BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
793 [BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
794 [BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
795 [BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
796 [BPF_ALU | BPF_NEG] = &&ALU_NEG,
797 [BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
798 [BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
799 /* 64 bit ALU operations */
800 [BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
801 [BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
802 [BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
803 [BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
804 [BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
805 [BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
806 [BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
807 [BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
808 [BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
809 [BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
810 [BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
811 [BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
812 [BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
813 [BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
814 [BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
815 [BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
816 [BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
817 [BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
818 [BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
819 [BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
820 [BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
821 [BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
822 [BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
823 [BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
824 [BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
825 /* Call instruction */
826 [BPF_JMP | BPF_CALL] = &&JMP_CALL,
827 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
829 [BPF_JMP | BPF_JA] = &&JMP_JA,
830 [BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
831 [BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
832 [BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
833 [BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
834 [BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
835 [BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
836 [BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
837 [BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
838 [BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
839 [BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
840 [BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
841 [BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
842 [BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
843 [BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
845 [BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
846 /* Store instructions */
847 [BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
848 [BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
849 [BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
850 [BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
851 [BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
852 [BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
853 [BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
854 [BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
855 [BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
856 [BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
857 /* Load instructions */
858 [BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
859 [BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
860 [BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
861 [BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
862 [BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
863 [BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
864 [BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
865 [BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
866 [BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
867 [BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
868 [BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
870 u32 tail_call_cnt = 0;
874 #define CONT ({ insn++; goto select_insn; })
875 #define CONT_JMP ({ insn++; goto select_insn; })
878 goto *jumptable[insn->code];
881 #define ALU(OPCODE, OP) \
882 ALU64_##OPCODE##_X: \
886 DST = (u32) DST OP (u32) SRC; \
888 ALU64_##OPCODE##_K: \
892 DST = (u32) DST OP (u32) IMM; \
923 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
927 (*(s64 *) &DST) >>= SRC;
930 (*(s64 *) &DST) >>= IMM;
933 if (unlikely(SRC == 0))
935 div64_u64_rem(DST, SRC, &tmp);
939 if (unlikely(SRC == 0))
942 DST = do_div(tmp, (u32) SRC);
945 div64_u64_rem(DST, IMM, &tmp);
950 DST = do_div(tmp, (u32) IMM);
953 if (unlikely(SRC == 0))
955 DST = div64_u64(DST, SRC);
958 if (unlikely(SRC == 0))
961 do_div(tmp, (u32) SRC);
965 DST = div64_u64(DST, IMM);
969 do_div(tmp, (u32) IMM);
975 DST = (__force u16) cpu_to_be16(DST);
978 DST = (__force u32) cpu_to_be32(DST);
981 DST = (__force u64) cpu_to_be64(DST);
988 DST = (__force u16) cpu_to_le16(DST);
991 DST = (__force u32) cpu_to_le32(DST);
994 DST = (__force u64) cpu_to_le64(DST);
1001 /* Function call scratches BPF_R1-BPF_R5 registers,
1002 * preserves BPF_R6-BPF_R9, and stores return value
1005 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1010 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1011 struct bpf_array *array = container_of(map, struct bpf_array, map);
1012 struct bpf_prog *prog;
1015 if (unlikely(index >= array->map.max_entries))
1017 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1022 prog = READ_ONCE(array->ptrs[index]);
1026 /* ARG1 at this point is guaranteed to point to CTX from
1027 * the verifier side due to the fact that the tail call is
1028 * handeled like a helper, that is, bpf_tail_call_proto,
1029 * where arg1_type is ARG_PTR_TO_CTX.
1031 insn = prog->insnsi;
1089 if (((s64) DST) > ((s64) SRC)) {
1095 if (((s64) DST) > ((s64) IMM)) {
1101 if (((s64) DST) >= ((s64) SRC)) {
1107 if (((s64) DST) >= ((s64) IMM)) {
1127 /* STX and ST and LDX*/
1128 #define LDST(SIZEOP, SIZE) \
1130 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1133 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1136 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1144 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
1145 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
1148 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
1149 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
1152 LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
1155 /* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are only
1156 * appearing in the programs where ctx == skb
1157 * (see may_access_skb() in the verifier). All programs
1158 * keep 'ctx' in regs[BPF_REG_CTX] == BPF_R6,
1159 * bpf_convert_filter() saves it in BPF_R6, internal BPF
1160 * verifier will check that BPF_R6 == ctx.
1162 * BPF_ABS and BPF_IND are wrappers of function calls,
1163 * so they scratch BPF_R1-BPF_R5 registers, preserve
1164 * BPF_R6-BPF_R9, and store return value into BPF_R0.
1167 * ctx == skb == BPF_R6 == CTX
1170 * SRC == any register
1171 * IMM == 32-bit immediate
1174 * BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
1177 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
1178 if (likely(ptr != NULL)) {
1179 BPF_R0 = get_unaligned_be32(ptr);
1184 LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
1187 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
1188 if (likely(ptr != NULL)) {
1189 BPF_R0 = get_unaligned_be16(ptr);
1194 LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
1197 ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
1198 if (likely(ptr != NULL)) {
1199 BPF_R0 = *(u8 *)ptr;
1204 LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
1207 LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
1210 LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
1215 /* If we ever reach this, we have a bug somewhere. */
1216 WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
1219 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */
1221 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
1222 #define DEFINE_BPF_PROG_RUN(stack_size) \
1223 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
1225 u64 stack[stack_size / sizeof(u64)]; \
1226 u64 regs[MAX_BPF_REG]; \
1228 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1229 ARG1 = (u64) (unsigned long) ctx; \
1230 return ___bpf_prog_run(regs, insn, stack); \
1233 #define EVAL1(FN, X) FN(X)
1234 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
1235 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
1236 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
1237 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
1238 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
1240 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
1241 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
1242 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
1244 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
1246 static unsigned int (*interpreters[])(const void *ctx,
1247 const struct bpf_insn *insn) = {
1248 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1249 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1250 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1253 bool bpf_prog_array_compatible(struct bpf_array *array,
1254 const struct bpf_prog *fp)
1256 if (!array->owner_prog_type) {
1257 /* There's no owner yet where we could check for
1260 array->owner_prog_type = fp->type;
1261 array->owner_jited = fp->jited;
1266 return array->owner_prog_type == fp->type &&
1267 array->owner_jited == fp->jited;
1270 static int bpf_check_tail_call(const struct bpf_prog *fp)
1272 struct bpf_prog_aux *aux = fp->aux;
1275 for (i = 0; i < aux->used_map_cnt; i++) {
1276 struct bpf_map *map = aux->used_maps[i];
1277 struct bpf_array *array;
1279 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1282 array = container_of(map, struct bpf_array, map);
1283 if (!bpf_prog_array_compatible(array, fp))
1291 * bpf_prog_select_runtime - select exec runtime for BPF program
1292 * @fp: bpf_prog populated with internal BPF program
1293 * @err: pointer to error variable
1295 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1296 * The BPF program will be executed via BPF_PROG_RUN() macro.
1298 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1300 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
1302 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
1304 /* eBPF JITs can rewrite the program in case constant
1305 * blinding is active. However, in case of error during
1306 * blinding, bpf_int_jit_compile() must always return a
1307 * valid program, which in this case would simply not
1308 * be JITed, but falls back to the interpreter.
1310 fp = bpf_int_jit_compile(fp);
1311 bpf_prog_lock_ro(fp);
1313 /* The tail call compatibility check can only be done at
1314 * this late stage as we need to determine, if we deal
1315 * with JITed or non JITed program concatenations and not
1316 * all eBPF JITs might immediately support all features.
1318 *err = bpf_check_tail_call(fp);
1322 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1324 static void bpf_prog_free_deferred(struct work_struct *work)
1326 struct bpf_prog_aux *aux;
1328 aux = container_of(work, struct bpf_prog_aux, work);
1329 bpf_jit_free(aux->prog);
1332 /* Free internal BPF program */
1333 void bpf_prog_free(struct bpf_prog *fp)
1335 struct bpf_prog_aux *aux = fp->aux;
1337 INIT_WORK(&aux->work, bpf_prog_free_deferred);
1338 schedule_work(&aux->work);
1340 EXPORT_SYMBOL_GPL(bpf_prog_free);
1342 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1343 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1345 void bpf_user_rnd_init_once(void)
1347 prandom_init_once(&bpf_user_rnd_state);
1350 BPF_CALL_0(bpf_user_rnd_u32)
1352 /* Should someone ever have the rather unwise idea to use some
1353 * of the registers passed into this function, then note that
1354 * this function is called from native eBPF and classic-to-eBPF
1355 * transformations. Register assignments from both sides are
1356 * different, f.e. classic always sets fn(ctx, A, X) here.
1358 struct rnd_state *state;
1361 state = &get_cpu_var(bpf_user_rnd_state);
1362 res = prandom_u32_state(state);
1363 put_cpu_var(bpf_user_rnd_state);
1368 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1369 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1370 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1371 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1373 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1374 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1375 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1376 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1378 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1379 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1380 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1382 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1388 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1389 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1394 /* Always built-in helper functions. */
1395 const struct bpf_func_proto bpf_tail_call_proto = {
1398 .ret_type = RET_VOID,
1399 .arg1_type = ARG_PTR_TO_CTX,
1400 .arg2_type = ARG_CONST_MAP_PTR,
1401 .arg3_type = ARG_ANYTHING,
1404 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
1405 * It is encouraged to implement bpf_int_jit_compile() instead, so that
1406 * eBPF and implicitly also cBPF can get JITed!
1408 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1413 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
1414 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
1416 void __weak bpf_jit_compile(struct bpf_prog *prog)
1420 bool __weak bpf_helper_changes_pkt_data(void *func)
1425 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1426 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1428 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
1434 /* All definitions of tracepoints related to BPF. */
1435 #define CREATE_TRACE_POINTS
1436 #include <linux/bpf_trace.h>
1438 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
1440 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_get_type);
1441 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_put_rcu);