1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Linux Socket Filter - Kernel level socket filtering
5 * Based on the design of the Berkeley Packet Filter. The new
6 * internal format has been designed by PLUMgrid:
8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
12 * Jay Schulist <jschlst@samba.org>
13 * Alexei Starovoitov <ast@plumgrid.com>
14 * Daniel Borkmann <dborkman@redhat.com>
16 * Andi Kleen - Fix a few bad bugs and races.
17 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
20 #include <uapi/linux/btf.h>
21 #include <linux/filter.h>
22 #include <linux/skbuff.h>
23 #include <linux/vmalloc.h>
24 #include <linux/random.h>
25 #include <linux/moduleloader.h>
26 #include <linux/bpf.h>
27 #include <linux/btf.h>
28 #include <linux/frame.h>
29 #include <linux/rbtree_latch.h>
30 #include <linux/kallsyms.h>
31 #include <linux/rcupdate.h>
32 #include <linux/perf_event.h>
33 #include <linux/extable.h>
34 #include <linux/log2.h>
35 #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 AX regs[BPF_REG_AX]
55 #define ARG1 regs[BPF_REG_ARG1]
56 #define CTX regs[BPF_REG_CTX]
59 /* No hurry in this branch
61 * Exported for the bpf jit load helper.
63 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
68 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
69 else if (k >= SKF_LL_OFF)
70 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
72 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
78 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
80 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
81 struct bpf_prog_aux *aux;
84 size = round_up(size, PAGE_SIZE);
85 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
89 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
95 fp->pages = size / PAGE_SIZE;
98 fp->jit_requested = ebpf_jit_enabled();
100 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
105 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
107 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
108 struct bpf_prog *prog;
111 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
115 prog->aux->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
116 if (!prog->aux->stats) {
122 for_each_possible_cpu(cpu) {
123 struct bpf_prog_stats *pstats;
125 pstats = per_cpu_ptr(prog->aux->stats, cpu);
126 u64_stats_init(&pstats->syncp);
130 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
132 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
134 if (!prog->aux->nr_linfo || !prog->jit_requested)
137 prog->aux->jited_linfo = kcalloc(prog->aux->nr_linfo,
138 sizeof(*prog->aux->jited_linfo),
139 GFP_KERNEL | __GFP_NOWARN);
140 if (!prog->aux->jited_linfo)
146 void bpf_prog_free_jited_linfo(struct bpf_prog *prog)
148 kfree(prog->aux->jited_linfo);
149 prog->aux->jited_linfo = NULL;
152 void bpf_prog_free_unused_jited_linfo(struct bpf_prog *prog)
154 if (prog->aux->jited_linfo && !prog->aux->jited_linfo[0])
155 bpf_prog_free_jited_linfo(prog);
158 /* The jit engine is responsible to provide an array
159 * for insn_off to the jited_off mapping (insn_to_jit_off).
161 * The idx to this array is the insn_off. Hence, the insn_off
162 * here is relative to the prog itself instead of the main prog.
163 * This array has one entry for each xlated bpf insn.
165 * jited_off is the byte off to the last byte of the jited insn.
169 * The first bpf insn off of the prog. The insn off
170 * here is relative to the main prog.
171 * e.g. if prog is a subprog, insn_start > 0
173 * The prog's idx to prog->aux->linfo and jited_linfo
175 * jited_linfo[linfo_idx] = prog->bpf_func
179 * jited_linfo[i] = prog->bpf_func +
180 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
182 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
183 const u32 *insn_to_jit_off)
185 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
186 const struct bpf_line_info *linfo;
189 if (!prog->aux->jited_linfo)
190 /* Userspace did not provide linfo */
193 linfo_idx = prog->aux->linfo_idx;
194 linfo = &prog->aux->linfo[linfo_idx];
195 insn_start = linfo[0].insn_off;
196 insn_end = insn_start + prog->len;
198 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
199 jited_linfo[0] = prog->bpf_func;
201 nr_linfo = prog->aux->nr_linfo - linfo_idx;
203 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
204 /* The verifier ensures that linfo[i].insn_off is
205 * strictly increasing
207 jited_linfo[i] = prog->bpf_func +
208 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
211 void bpf_prog_free_linfo(struct bpf_prog *prog)
213 bpf_prog_free_jited_linfo(prog);
214 kvfree(prog->aux->linfo);
217 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
218 gfp_t gfp_extra_flags)
220 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
225 size = round_up(size, PAGE_SIZE);
226 pages = size / PAGE_SIZE;
227 if (pages <= fp_old->pages)
230 delta = pages - fp_old->pages;
231 ret = __bpf_prog_charge(fp_old->aux->user, delta);
235 fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
237 __bpf_prog_uncharge(fp_old->aux->user, delta);
239 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
243 /* We keep fp->aux from fp_old around in the new
244 * reallocated structure.
247 __bpf_prog_free(fp_old);
253 void __bpf_prog_free(struct bpf_prog *fp)
256 free_percpu(fp->aux->stats);
257 kfree(fp->aux->poke_tab);
263 int bpf_prog_calc_tag(struct bpf_prog *fp)
265 const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
266 u32 raw_size = bpf_prog_tag_scratch_size(fp);
267 u32 digest[SHA_DIGEST_WORDS];
268 u32 ws[SHA_WORKSPACE_WORDS];
269 u32 i, bsize, psize, blocks;
270 struct bpf_insn *dst;
276 raw = vmalloc(raw_size);
281 memset(ws, 0, sizeof(ws));
283 /* We need to take out the map fd for the digest calculation
284 * since they are unstable from user space side.
287 for (i = 0, was_ld_map = false; i < fp->len; i++) {
288 dst[i] = fp->insnsi[i];
290 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
291 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
292 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
295 } else if (was_ld_map &&
297 dst[i].dst_reg == 0 &&
298 dst[i].src_reg == 0 &&
307 psize = bpf_prog_insn_size(fp);
308 memset(&raw[psize], 0, raw_size - psize);
311 bsize = round_up(psize, SHA_MESSAGE_BYTES);
312 blocks = bsize / SHA_MESSAGE_BYTES;
314 if (bsize - psize >= sizeof(__be64)) {
315 bits = (__be64 *)(todo + bsize - sizeof(__be64));
317 bits = (__be64 *)(todo + bsize + bits_offset);
320 *bits = cpu_to_be64((psize - 1) << 3);
323 sha_transform(digest, todo, ws);
324 todo += SHA_MESSAGE_BYTES;
327 result = (__force __be32 *)digest;
328 for (i = 0; i < SHA_DIGEST_WORDS; i++)
329 result[i] = cpu_to_be32(digest[i]);
330 memcpy(fp->tag, result, sizeof(fp->tag));
336 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
337 s32 end_new, s32 curr, const bool probe_pass)
339 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
340 s32 delta = end_new - end_old;
343 if (curr < pos && curr + imm + 1 >= end_old)
345 else if (curr >= end_new && curr + imm + 1 < end_new)
347 if (imm < imm_min || imm > imm_max)
354 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
355 s32 end_new, s32 curr, const bool probe_pass)
357 const s32 off_min = S16_MIN, off_max = S16_MAX;
358 s32 delta = end_new - end_old;
361 if (curr < pos && curr + off + 1 >= end_old)
363 else if (curr >= end_new && curr + off + 1 < end_new)
365 if (off < off_min || off > off_max)
372 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
373 s32 end_new, const bool probe_pass)
375 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
376 struct bpf_insn *insn = prog->insnsi;
379 for (i = 0; i < insn_cnt; i++, insn++) {
382 /* In the probing pass we still operate on the original,
383 * unpatched image in order to check overflows before we
384 * do any other adjustments. Therefore skip the patchlet.
386 if (probe_pass && i == pos) {
388 insn = prog->insnsi + end_old;
391 if ((BPF_CLASS(code) != BPF_JMP &&
392 BPF_CLASS(code) != BPF_JMP32) ||
393 BPF_OP(code) == BPF_EXIT)
395 /* Adjust offset of jmps if we cross patch boundaries. */
396 if (BPF_OP(code) == BPF_CALL) {
397 if (insn->src_reg != BPF_PSEUDO_CALL)
399 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
400 end_new, i, probe_pass);
402 ret = bpf_adj_delta_to_off(insn, pos, end_old,
403 end_new, i, probe_pass);
412 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
414 struct bpf_line_info *linfo;
417 nr_linfo = prog->aux->nr_linfo;
418 if (!nr_linfo || !delta)
421 linfo = prog->aux->linfo;
423 for (i = 0; i < nr_linfo; i++)
424 if (off < linfo[i].insn_off)
427 /* Push all off < linfo[i].insn_off by delta */
428 for (; i < nr_linfo; i++)
429 linfo[i].insn_off += delta;
432 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
433 const struct bpf_insn *patch, u32 len)
435 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
436 const u32 cnt_max = S16_MAX;
437 struct bpf_prog *prog_adj;
440 /* Since our patchlet doesn't expand the image, we're done. */
441 if (insn_delta == 0) {
442 memcpy(prog->insnsi + off, patch, sizeof(*patch));
446 insn_adj_cnt = prog->len + insn_delta;
448 /* Reject anything that would potentially let the insn->off
449 * target overflow when we have excessive program expansions.
450 * We need to probe here before we do any reallocation where
451 * we afterwards may not fail anymore.
453 if (insn_adj_cnt > cnt_max &&
454 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
457 /* Several new instructions need to be inserted. Make room
458 * for them. Likely, there's no need for a new allocation as
459 * last page could have large enough tailroom.
461 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
464 return ERR_PTR(-ENOMEM);
466 prog_adj->len = insn_adj_cnt;
468 /* Patching happens in 3 steps:
470 * 1) Move over tail of insnsi from next instruction onwards,
471 * so we can patch the single target insn with one or more
472 * new ones (patching is always from 1 to n insns, n > 0).
473 * 2) Inject new instructions at the target location.
474 * 3) Adjust branch offsets if necessary.
476 insn_rest = insn_adj_cnt - off - len;
478 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
479 sizeof(*patch) * insn_rest);
480 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
482 /* We are guaranteed to not fail at this point, otherwise
483 * the ship has sailed to reverse to the original state. An
484 * overflow cannot happen at this point.
486 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
488 bpf_adj_linfo(prog_adj, off, insn_delta);
493 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
495 /* Branch offsets can't overflow when program is shrinking, no need
496 * to call bpf_adj_branches(..., true) here
498 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
499 sizeof(struct bpf_insn) * (prog->len - off - cnt));
502 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
505 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
509 for (i = 0; i < fp->aux->func_cnt; i++)
510 bpf_prog_kallsyms_del(fp->aux->func[i]);
513 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
515 bpf_prog_kallsyms_del_subprogs(fp);
516 bpf_prog_kallsyms_del(fp);
519 #ifdef CONFIG_BPF_JIT
520 /* All BPF JIT sysctl knobs here. */
521 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
522 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
523 int bpf_jit_harden __read_mostly;
524 long bpf_jit_limit __read_mostly;
527 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
529 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
530 unsigned long addr = (unsigned long)hdr;
532 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
534 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
535 prog->aux->ksym.end = addr + hdr->pages * PAGE_SIZE;
539 bpf_prog_ksym_set_name(struct bpf_prog *prog)
541 char *sym = prog->aux->ksym.name;
542 const char *end = sym + KSYM_NAME_LEN;
543 const struct btf_type *type;
544 const char *func_name;
546 BUILD_BUG_ON(sizeof("bpf_prog_") +
547 sizeof(prog->tag) * 2 +
548 /* name has been null terminated.
549 * We should need +1 for the '_' preceding
550 * the name. However, the null character
551 * is double counted between the name and the
552 * sizeof("bpf_prog_") above, so we omit
555 sizeof(prog->aux->name) > KSYM_NAME_LEN);
557 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
558 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
560 /* prog->aux->name will be ignored if full btf name is available */
561 if (prog->aux->func_info_cnt) {
562 type = btf_type_by_id(prog->aux->btf,
563 prog->aux->func_info[prog->aux->func_idx].type_id);
564 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
565 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
569 if (prog->aux->name[0])
570 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
575 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
577 return container_of(n, struct bpf_ksym, tnode)->start;
580 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
581 struct latch_tree_node *b)
583 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
586 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
588 unsigned long val = (unsigned long)key;
589 const struct bpf_ksym *ksym;
591 ksym = container_of(n, struct bpf_ksym, tnode);
593 if (val < ksym->start)
595 if (val >= ksym->end)
601 static const struct latch_tree_ops bpf_tree_ops = {
602 .less = bpf_tree_less,
603 .comp = bpf_tree_comp,
606 static DEFINE_SPINLOCK(bpf_lock);
607 static LIST_HEAD(bpf_kallsyms);
608 static struct latch_tree_root bpf_tree __cacheline_aligned;
610 void bpf_ksym_add(struct bpf_ksym *ksym)
612 spin_lock_bh(&bpf_lock);
613 WARN_ON_ONCE(!list_empty(&ksym->lnode));
614 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
615 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
616 spin_unlock_bh(&bpf_lock);
619 static void __bpf_ksym_del(struct bpf_ksym *ksym)
621 if (list_empty(&ksym->lnode))
624 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
625 list_del_rcu(&ksym->lnode);
628 void bpf_ksym_del(struct bpf_ksym *ksym)
630 spin_lock_bh(&bpf_lock);
631 __bpf_ksym_del(ksym);
632 spin_unlock_bh(&bpf_lock);
635 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
637 return fp->jited && !bpf_prog_was_classic(fp);
640 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
642 return list_empty(&fp->aux->ksym.lnode) ||
643 fp->aux->ksym.lnode.prev == LIST_POISON2;
646 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
648 if (!bpf_prog_kallsyms_candidate(fp) ||
652 bpf_prog_ksym_set_addr(fp);
653 bpf_prog_ksym_set_name(fp);
654 fp->aux->ksym.prog = true;
656 bpf_ksym_add(&fp->aux->ksym);
659 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
661 if (!bpf_prog_kallsyms_candidate(fp))
664 bpf_ksym_del(&fp->aux->ksym);
667 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
669 struct latch_tree_node *n;
671 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
672 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
675 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
676 unsigned long *off, char *sym)
678 struct bpf_ksym *ksym;
682 ksym = bpf_ksym_find(addr);
684 unsigned long symbol_start = ksym->start;
685 unsigned long symbol_end = ksym->end;
687 strncpy(sym, ksym->name, KSYM_NAME_LEN);
691 *size = symbol_end - symbol_start;
693 *off = addr - symbol_start;
700 bool is_bpf_text_address(unsigned long addr)
705 ret = bpf_ksym_find(addr) != NULL;
711 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
713 struct bpf_ksym *ksym = bpf_ksym_find(addr);
715 return ksym && ksym->prog ?
716 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
720 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
722 const struct exception_table_entry *e = NULL;
723 struct bpf_prog *prog;
726 prog = bpf_prog_ksym_find(addr);
729 if (!prog->aux->num_exentries)
732 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
738 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
741 struct bpf_ksym *ksym;
745 if (!bpf_jit_kallsyms_enabled())
749 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
753 strncpy(sym, ksym->name, KSYM_NAME_LEN);
755 *value = ksym->start;
756 *type = BPF_SYM_ELF_TYPE;
766 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
767 struct bpf_jit_poke_descriptor *poke)
769 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
770 static const u32 poke_tab_max = 1024;
771 u32 slot = prog->aux->size_poke_tab;
774 if (size > poke_tab_max)
776 if (poke->ip || poke->ip_stable || poke->adj_off)
779 switch (poke->reason) {
780 case BPF_POKE_REASON_TAIL_CALL:
781 if (!poke->tail_call.map)
788 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
792 memcpy(&tab[slot], poke, sizeof(*poke));
793 prog->aux->size_poke_tab = size;
794 prog->aux->poke_tab = tab;
799 static atomic_long_t bpf_jit_current;
801 /* Can be overridden by an arch's JIT compiler if it has a custom,
802 * dedicated BPF backend memory area, or if neither of the two
805 u64 __weak bpf_jit_alloc_exec_limit(void)
807 #if defined(MODULES_VADDR)
808 return MODULES_END - MODULES_VADDR;
810 return VMALLOC_END - VMALLOC_START;
814 static int __init bpf_jit_charge_init(void)
816 /* Only used as heuristic here to derive limit. */
817 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2,
818 PAGE_SIZE), LONG_MAX);
821 pure_initcall(bpf_jit_charge_init);
823 static int bpf_jit_charge_modmem(u32 pages)
825 if (atomic_long_add_return(pages, &bpf_jit_current) >
826 (bpf_jit_limit >> PAGE_SHIFT)) {
827 if (!capable(CAP_SYS_ADMIN)) {
828 atomic_long_sub(pages, &bpf_jit_current);
836 static void bpf_jit_uncharge_modmem(u32 pages)
838 atomic_long_sub(pages, &bpf_jit_current);
841 void *__weak bpf_jit_alloc_exec(unsigned long size)
843 return module_alloc(size);
846 void __weak bpf_jit_free_exec(void *addr)
848 module_memfree(addr);
851 struct bpf_binary_header *
852 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
853 unsigned int alignment,
854 bpf_jit_fill_hole_t bpf_fill_ill_insns)
856 struct bpf_binary_header *hdr;
857 u32 size, hole, start, pages;
859 WARN_ON_ONCE(!is_power_of_2(alignment) ||
860 alignment > BPF_IMAGE_ALIGNMENT);
862 /* Most of BPF filters are really small, but if some of them
863 * fill a page, allow at least 128 extra bytes to insert a
864 * random section of illegal instructions.
866 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
867 pages = size / PAGE_SIZE;
869 if (bpf_jit_charge_modmem(pages))
871 hdr = bpf_jit_alloc_exec(size);
873 bpf_jit_uncharge_modmem(pages);
877 /* Fill space with illegal/arch-dep instructions. */
878 bpf_fill_ill_insns(hdr, size);
881 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
882 PAGE_SIZE - sizeof(*hdr));
883 start = (get_random_int() % hole) & ~(alignment - 1);
885 /* Leave a random number of instructions before BPF code. */
886 *image_ptr = &hdr->image[start];
891 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
893 u32 pages = hdr->pages;
895 bpf_jit_free_exec(hdr);
896 bpf_jit_uncharge_modmem(pages);
899 /* This symbol is only overridden by archs that have different
900 * requirements than the usual eBPF JITs, f.e. when they only
901 * implement cBPF JIT, do not set images read-only, etc.
903 void __weak bpf_jit_free(struct bpf_prog *fp)
906 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
908 bpf_jit_binary_free(hdr);
910 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
913 bpf_prog_unlock_free(fp);
916 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
917 const struct bpf_insn *insn, bool extra_pass,
918 u64 *func_addr, bool *func_addr_fixed)
924 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
925 if (!*func_addr_fixed) {
926 /* Place-holder address till the last pass has collected
927 * all addresses for JITed subprograms in which case we
928 * can pick them up from prog->aux.
932 else if (prog->aux->func &&
933 off >= 0 && off < prog->aux->func_cnt)
934 addr = (u8 *)prog->aux->func[off]->bpf_func;
938 /* Address of a BPF helper call. Since part of the core
939 * kernel, it's always at a fixed location. __bpf_call_base
940 * and the helper with imm relative to it are both in core
943 addr = (u8 *)__bpf_call_base + imm;
946 *func_addr = (unsigned long)addr;
950 static int bpf_jit_blind_insn(const struct bpf_insn *from,
951 const struct bpf_insn *aux,
952 struct bpf_insn *to_buff,
955 struct bpf_insn *to = to_buff;
956 u32 imm_rnd = get_random_int();
959 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
960 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
962 /* Constraints on AX register:
964 * AX register is inaccessible from user space. It is mapped in
965 * all JITs, and used here for constant blinding rewrites. It is
966 * typically "stateless" meaning its contents are only valid within
967 * the executed instruction, but not across several instructions.
968 * There are a few exceptions however which are further detailed
971 * Constant blinding is only used by JITs, not in the interpreter.
972 * The interpreter uses AX in some occasions as a local temporary
973 * register e.g. in DIV or MOD instructions.
975 * In restricted circumstances, the verifier can also use the AX
976 * register for rewrites as long as they do not interfere with
979 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
982 if (from->imm == 0 &&
983 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
984 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
985 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
989 switch (from->code) {
990 case BPF_ALU | BPF_ADD | BPF_K:
991 case BPF_ALU | BPF_SUB | BPF_K:
992 case BPF_ALU | BPF_AND | BPF_K:
993 case BPF_ALU | BPF_OR | BPF_K:
994 case BPF_ALU | BPF_XOR | BPF_K:
995 case BPF_ALU | BPF_MUL | BPF_K:
996 case BPF_ALU | BPF_MOV | BPF_K:
997 case BPF_ALU | BPF_DIV | BPF_K:
998 case BPF_ALU | BPF_MOD | BPF_K:
999 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1000 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1001 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1004 case BPF_ALU64 | BPF_ADD | BPF_K:
1005 case BPF_ALU64 | BPF_SUB | BPF_K:
1006 case BPF_ALU64 | BPF_AND | BPF_K:
1007 case BPF_ALU64 | BPF_OR | BPF_K:
1008 case BPF_ALU64 | BPF_XOR | BPF_K:
1009 case BPF_ALU64 | BPF_MUL | BPF_K:
1010 case BPF_ALU64 | BPF_MOV | BPF_K:
1011 case BPF_ALU64 | BPF_DIV | BPF_K:
1012 case BPF_ALU64 | BPF_MOD | BPF_K:
1013 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1014 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1015 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1018 case BPF_JMP | BPF_JEQ | BPF_K:
1019 case BPF_JMP | BPF_JNE | BPF_K:
1020 case BPF_JMP | BPF_JGT | BPF_K:
1021 case BPF_JMP | BPF_JLT | BPF_K:
1022 case BPF_JMP | BPF_JGE | BPF_K:
1023 case BPF_JMP | BPF_JLE | BPF_K:
1024 case BPF_JMP | BPF_JSGT | BPF_K:
1025 case BPF_JMP | BPF_JSLT | BPF_K:
1026 case BPF_JMP | BPF_JSGE | BPF_K:
1027 case BPF_JMP | BPF_JSLE | BPF_K:
1028 case BPF_JMP | BPF_JSET | BPF_K:
1029 /* Accommodate for extra offset in case of a backjump. */
1033 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1034 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1035 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1038 case BPF_JMP32 | BPF_JEQ | BPF_K:
1039 case BPF_JMP32 | BPF_JNE | BPF_K:
1040 case BPF_JMP32 | BPF_JGT | BPF_K:
1041 case BPF_JMP32 | BPF_JLT | BPF_K:
1042 case BPF_JMP32 | BPF_JGE | BPF_K:
1043 case BPF_JMP32 | BPF_JLE | BPF_K:
1044 case BPF_JMP32 | BPF_JSGT | BPF_K:
1045 case BPF_JMP32 | BPF_JSLT | BPF_K:
1046 case BPF_JMP32 | BPF_JSGE | BPF_K:
1047 case BPF_JMP32 | BPF_JSLE | BPF_K:
1048 case BPF_JMP32 | BPF_JSET | BPF_K:
1049 /* Accommodate for extra offset in case of a backjump. */
1053 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1054 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1055 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1059 case BPF_LD | BPF_IMM | BPF_DW:
1060 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1061 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1062 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1063 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1065 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1066 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1067 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1069 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1070 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1073 case BPF_ST | BPF_MEM | BPF_DW:
1074 case BPF_ST | BPF_MEM | BPF_W:
1075 case BPF_ST | BPF_MEM | BPF_H:
1076 case BPF_ST | BPF_MEM | BPF_B:
1077 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1078 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1079 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1083 return to - to_buff;
1086 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1087 gfp_t gfp_extra_flags)
1089 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1090 struct bpf_prog *fp;
1092 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
1094 /* aux->prog still points to the fp_other one, so
1095 * when promoting the clone to the real program,
1096 * this still needs to be adapted.
1098 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1104 static void bpf_prog_clone_free(struct bpf_prog *fp)
1106 /* aux was stolen by the other clone, so we cannot free
1107 * it from this path! It will be freed eventually by the
1108 * other program on release.
1110 * At this point, we don't need a deferred release since
1111 * clone is guaranteed to not be locked.
1114 __bpf_prog_free(fp);
1117 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1119 /* We have to repoint aux->prog to self, as we don't
1120 * know whether fp here is the clone or the original.
1123 bpf_prog_clone_free(fp_other);
1126 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1128 struct bpf_insn insn_buff[16], aux[2];
1129 struct bpf_prog *clone, *tmp;
1130 int insn_delta, insn_cnt;
1131 struct bpf_insn *insn;
1134 if (!bpf_jit_blinding_enabled(prog) || prog->blinded)
1137 clone = bpf_prog_clone_create(prog, GFP_USER);
1139 return ERR_PTR(-ENOMEM);
1141 insn_cnt = clone->len;
1142 insn = clone->insnsi;
1144 for (i = 0; i < insn_cnt; i++, insn++) {
1145 /* We temporarily need to hold the original ld64 insn
1146 * so that we can still access the first part in the
1147 * second blinding run.
1149 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1151 memcpy(aux, insn, sizeof(aux));
1153 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1154 clone->aux->verifier_zext);
1158 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1160 /* Patching may have repointed aux->prog during
1161 * realloc from the original one, so we need to
1162 * fix it up here on error.
1164 bpf_jit_prog_release_other(prog, clone);
1169 insn_delta = rewritten - 1;
1171 /* Walk new program and skip insns we just inserted. */
1172 insn = clone->insnsi + i + insn_delta;
1173 insn_cnt += insn_delta;
1180 #endif /* CONFIG_BPF_JIT */
1182 /* Base function for offset calculation. Needs to go into .text section,
1183 * therefore keeping it non-static as well; will also be used by JITs
1184 * anyway later on, so do not let the compiler omit it. This also needs
1185 * to go into kallsyms for correlation from e.g. bpftool, so naming
1188 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1192 EXPORT_SYMBOL_GPL(__bpf_call_base);
1194 /* All UAPI available opcodes. */
1195 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1196 /* 32 bit ALU operations. */ \
1197 /* Register based. */ \
1198 INSN_3(ALU, ADD, X), \
1199 INSN_3(ALU, SUB, X), \
1200 INSN_3(ALU, AND, X), \
1201 INSN_3(ALU, OR, X), \
1202 INSN_3(ALU, LSH, X), \
1203 INSN_3(ALU, RSH, X), \
1204 INSN_3(ALU, XOR, X), \
1205 INSN_3(ALU, MUL, X), \
1206 INSN_3(ALU, MOV, X), \
1207 INSN_3(ALU, ARSH, X), \
1208 INSN_3(ALU, DIV, X), \
1209 INSN_3(ALU, MOD, X), \
1211 INSN_3(ALU, END, TO_BE), \
1212 INSN_3(ALU, END, TO_LE), \
1213 /* Immediate based. */ \
1214 INSN_3(ALU, ADD, K), \
1215 INSN_3(ALU, SUB, K), \
1216 INSN_3(ALU, AND, K), \
1217 INSN_3(ALU, OR, K), \
1218 INSN_3(ALU, LSH, K), \
1219 INSN_3(ALU, RSH, K), \
1220 INSN_3(ALU, XOR, K), \
1221 INSN_3(ALU, MUL, K), \
1222 INSN_3(ALU, MOV, K), \
1223 INSN_3(ALU, ARSH, K), \
1224 INSN_3(ALU, DIV, K), \
1225 INSN_3(ALU, MOD, K), \
1226 /* 64 bit ALU operations. */ \
1227 /* Register based. */ \
1228 INSN_3(ALU64, ADD, X), \
1229 INSN_3(ALU64, SUB, X), \
1230 INSN_3(ALU64, AND, X), \
1231 INSN_3(ALU64, OR, X), \
1232 INSN_3(ALU64, LSH, X), \
1233 INSN_3(ALU64, RSH, X), \
1234 INSN_3(ALU64, XOR, X), \
1235 INSN_3(ALU64, MUL, X), \
1236 INSN_3(ALU64, MOV, X), \
1237 INSN_3(ALU64, ARSH, X), \
1238 INSN_3(ALU64, DIV, X), \
1239 INSN_3(ALU64, MOD, X), \
1240 INSN_2(ALU64, NEG), \
1241 /* Immediate based. */ \
1242 INSN_3(ALU64, ADD, K), \
1243 INSN_3(ALU64, SUB, K), \
1244 INSN_3(ALU64, AND, K), \
1245 INSN_3(ALU64, OR, K), \
1246 INSN_3(ALU64, LSH, K), \
1247 INSN_3(ALU64, RSH, K), \
1248 INSN_3(ALU64, XOR, K), \
1249 INSN_3(ALU64, MUL, K), \
1250 INSN_3(ALU64, MOV, K), \
1251 INSN_3(ALU64, ARSH, K), \
1252 INSN_3(ALU64, DIV, K), \
1253 INSN_3(ALU64, MOD, K), \
1254 /* Call instruction. */ \
1255 INSN_2(JMP, CALL), \
1256 /* Exit instruction. */ \
1257 INSN_2(JMP, EXIT), \
1258 /* 32-bit Jump instructions. */ \
1259 /* Register based. */ \
1260 INSN_3(JMP32, JEQ, X), \
1261 INSN_3(JMP32, JNE, X), \
1262 INSN_3(JMP32, JGT, X), \
1263 INSN_3(JMP32, JLT, X), \
1264 INSN_3(JMP32, JGE, X), \
1265 INSN_3(JMP32, JLE, X), \
1266 INSN_3(JMP32, JSGT, X), \
1267 INSN_3(JMP32, JSLT, X), \
1268 INSN_3(JMP32, JSGE, X), \
1269 INSN_3(JMP32, JSLE, X), \
1270 INSN_3(JMP32, JSET, X), \
1271 /* Immediate based. */ \
1272 INSN_3(JMP32, JEQ, K), \
1273 INSN_3(JMP32, JNE, K), \
1274 INSN_3(JMP32, JGT, K), \
1275 INSN_3(JMP32, JLT, K), \
1276 INSN_3(JMP32, JGE, K), \
1277 INSN_3(JMP32, JLE, K), \
1278 INSN_3(JMP32, JSGT, K), \
1279 INSN_3(JMP32, JSLT, K), \
1280 INSN_3(JMP32, JSGE, K), \
1281 INSN_3(JMP32, JSLE, K), \
1282 INSN_3(JMP32, JSET, K), \
1283 /* Jump instructions. */ \
1284 /* Register based. */ \
1285 INSN_3(JMP, JEQ, X), \
1286 INSN_3(JMP, JNE, X), \
1287 INSN_3(JMP, JGT, X), \
1288 INSN_3(JMP, JLT, X), \
1289 INSN_3(JMP, JGE, X), \
1290 INSN_3(JMP, JLE, X), \
1291 INSN_3(JMP, JSGT, X), \
1292 INSN_3(JMP, JSLT, X), \
1293 INSN_3(JMP, JSGE, X), \
1294 INSN_3(JMP, JSLE, X), \
1295 INSN_3(JMP, JSET, X), \
1296 /* Immediate based. */ \
1297 INSN_3(JMP, JEQ, K), \
1298 INSN_3(JMP, JNE, K), \
1299 INSN_3(JMP, JGT, K), \
1300 INSN_3(JMP, JLT, K), \
1301 INSN_3(JMP, JGE, K), \
1302 INSN_3(JMP, JLE, K), \
1303 INSN_3(JMP, JSGT, K), \
1304 INSN_3(JMP, JSLT, K), \
1305 INSN_3(JMP, JSGE, K), \
1306 INSN_3(JMP, JSLE, K), \
1307 INSN_3(JMP, JSET, K), \
1309 /* Store instructions. */ \
1310 /* Register based. */ \
1311 INSN_3(STX, MEM, B), \
1312 INSN_3(STX, MEM, H), \
1313 INSN_3(STX, MEM, W), \
1314 INSN_3(STX, MEM, DW), \
1315 INSN_3(STX, XADD, W), \
1316 INSN_3(STX, XADD, DW), \
1317 /* Immediate based. */ \
1318 INSN_3(ST, MEM, B), \
1319 INSN_3(ST, MEM, H), \
1320 INSN_3(ST, MEM, W), \
1321 INSN_3(ST, MEM, DW), \
1322 /* Load instructions. */ \
1323 /* Register based. */ \
1324 INSN_3(LDX, MEM, B), \
1325 INSN_3(LDX, MEM, H), \
1326 INSN_3(LDX, MEM, W), \
1327 INSN_3(LDX, MEM, DW), \
1328 /* Immediate based. */ \
1331 bool bpf_opcode_in_insntable(u8 code)
1333 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1334 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1335 static const bool public_insntable[256] = {
1336 [0 ... 255] = false,
1337 /* Now overwrite non-defaults ... */
1338 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1339 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1340 [BPF_LD | BPF_ABS | BPF_B] = true,
1341 [BPF_LD | BPF_ABS | BPF_H] = true,
1342 [BPF_LD | BPF_ABS | BPF_W] = true,
1343 [BPF_LD | BPF_IND | BPF_B] = true,
1344 [BPF_LD | BPF_IND | BPF_H] = true,
1345 [BPF_LD | BPF_IND | BPF_W] = true,
1347 #undef BPF_INSN_3_TBL
1348 #undef BPF_INSN_2_TBL
1349 return public_insntable[code];
1352 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1353 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1355 memset(dst, 0, size);
1360 * __bpf_prog_run - run eBPF program on a given context
1361 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1362 * @insn: is the array of eBPF instructions
1363 * @stack: is the eBPF storage stack
1365 * Decode and execute eBPF instructions.
1367 static u64 __no_fgcse ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack)
1369 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1370 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1371 static const void * const jumptable[256] __annotate_jump_table = {
1372 [0 ... 255] = &&default_label,
1373 /* Now overwrite non-defaults ... */
1374 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1375 /* Non-UAPI available opcodes. */
1376 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1377 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1378 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1379 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1380 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1381 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1383 #undef BPF_INSN_3_LBL
1384 #undef BPF_INSN_2_LBL
1385 u32 tail_call_cnt = 0;
1387 #define CONT ({ insn++; goto select_insn; })
1388 #define CONT_JMP ({ insn++; goto select_insn; })
1391 goto *jumptable[insn->code];
1394 #define ALU(OPCODE, OP) \
1395 ALU64_##OPCODE##_X: \
1399 DST = (u32) DST OP (u32) SRC; \
1401 ALU64_##OPCODE##_K: \
1405 DST = (u32) DST OP (u32) IMM; \
1436 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1440 DST = (u64) (u32) (((s32) DST) >> SRC);
1443 DST = (u64) (u32) (((s32) DST) >> IMM);
1446 (*(s64 *) &DST) >>= SRC;
1449 (*(s64 *) &DST) >>= IMM;
1452 div64_u64_rem(DST, SRC, &AX);
1457 DST = do_div(AX, (u32) SRC);
1460 div64_u64_rem(DST, IMM, &AX);
1465 DST = do_div(AX, (u32) IMM);
1468 DST = div64_u64(DST, SRC);
1472 do_div(AX, (u32) SRC);
1476 DST = div64_u64(DST, IMM);
1480 do_div(AX, (u32) IMM);
1486 DST = (__force u16) cpu_to_be16(DST);
1489 DST = (__force u32) cpu_to_be32(DST);
1492 DST = (__force u64) cpu_to_be64(DST);
1499 DST = (__force u16) cpu_to_le16(DST);
1502 DST = (__force u32) cpu_to_le32(DST);
1505 DST = (__force u64) cpu_to_le64(DST);
1512 /* Function call scratches BPF_R1-BPF_R5 registers,
1513 * preserves BPF_R6-BPF_R9, and stores return value
1516 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1521 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1524 insn + insn->off + 1);
1528 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1529 struct bpf_array *array = container_of(map, struct bpf_array, map);
1530 struct bpf_prog *prog;
1533 if (unlikely(index >= array->map.max_entries))
1535 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1540 prog = READ_ONCE(array->ptrs[index]);
1544 /* ARG1 at this point is guaranteed to point to CTX from
1545 * the verifier side due to the fact that the tail call is
1546 * handeled like a helper, that is, bpf_tail_call_proto,
1547 * where arg1_type is ARG_PTR_TO_CTX.
1549 insn = prog->insnsi;
1560 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1562 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1563 insn += insn->off; \
1567 JMP32_##OPCODE##_X: \
1568 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1569 insn += insn->off; \
1574 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1575 insn += insn->off; \
1579 JMP32_##OPCODE##_K: \
1580 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1581 insn += insn->off; \
1585 COND_JMP(u, JEQ, ==)
1586 COND_JMP(u, JNE, !=)
1589 COND_JMP(u, JGE, >=)
1590 COND_JMP(u, JLE, <=)
1591 COND_JMP(u, JSET, &)
1592 COND_JMP(s, JSGT, >)
1593 COND_JMP(s, JSLT, <)
1594 COND_JMP(s, JSGE, >=)
1595 COND_JMP(s, JSLE, <=)
1597 /* STX and ST and LDX*/
1598 #define LDST(SIZEOP, SIZE) \
1600 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1603 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1606 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1614 #define LDX_PROBE(SIZEOP, SIZE) \
1615 LDX_PROBE_MEM_##SIZEOP: \
1616 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \
1624 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
1625 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
1628 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
1629 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
1634 /* If we ever reach this, we have a bug somewhere. Die hard here
1635 * instead of just returning 0; we could be somewhere in a subprog,
1636 * so execution could continue otherwise which we do /not/ want.
1638 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
1640 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code);
1645 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
1646 #define DEFINE_BPF_PROG_RUN(stack_size) \
1647 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
1649 u64 stack[stack_size / sizeof(u64)]; \
1650 u64 regs[MAX_BPF_EXT_REG]; \
1652 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1653 ARG1 = (u64) (unsigned long) ctx; \
1654 return ___bpf_prog_run(regs, insn, stack); \
1657 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
1658 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
1659 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
1660 const struct bpf_insn *insn) \
1662 u64 stack[stack_size / sizeof(u64)]; \
1663 u64 regs[MAX_BPF_EXT_REG]; \
1665 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1671 return ___bpf_prog_run(regs, insn, stack); \
1674 #define EVAL1(FN, X) FN(X)
1675 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
1676 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
1677 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
1678 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
1679 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
1681 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
1682 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
1683 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
1685 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
1686 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
1687 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
1689 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
1691 static unsigned int (*interpreters[])(const void *ctx,
1692 const struct bpf_insn *insn) = {
1693 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1694 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1695 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1697 #undef PROG_NAME_LIST
1698 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
1699 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
1700 const struct bpf_insn *insn) = {
1701 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1702 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1703 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1705 #undef PROG_NAME_LIST
1707 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
1709 stack_depth = max_t(u32, stack_depth, 1);
1710 insn->off = (s16) insn->imm;
1711 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
1712 __bpf_call_base_args;
1713 insn->code = BPF_JMP | BPF_CALL_ARGS;
1717 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
1718 const struct bpf_insn *insn)
1720 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
1721 * is not working properly, so warn about it!
1728 bool bpf_prog_array_compatible(struct bpf_array *array,
1729 const struct bpf_prog *fp)
1731 if (fp->kprobe_override)
1734 if (!array->aux->type) {
1735 /* There's no owner yet where we could check for
1738 array->aux->type = fp->type;
1739 array->aux->jited = fp->jited;
1743 return array->aux->type == fp->type &&
1744 array->aux->jited == fp->jited;
1747 static int bpf_check_tail_call(const struct bpf_prog *fp)
1749 struct bpf_prog_aux *aux = fp->aux;
1752 for (i = 0; i < aux->used_map_cnt; i++) {
1753 struct bpf_map *map = aux->used_maps[i];
1754 struct bpf_array *array;
1756 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1759 array = container_of(map, struct bpf_array, map);
1760 if (!bpf_prog_array_compatible(array, fp))
1767 static void bpf_prog_select_func(struct bpf_prog *fp)
1769 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1770 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
1772 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
1774 fp->bpf_func = __bpf_prog_ret0_warn;
1779 * bpf_prog_select_runtime - select exec runtime for BPF program
1780 * @fp: bpf_prog populated with internal BPF program
1781 * @err: pointer to error variable
1783 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1784 * The BPF program will be executed via BPF_PROG_RUN() macro.
1786 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1788 /* In case of BPF to BPF calls, verifier did all the prep
1789 * work with regards to JITing, etc.
1794 bpf_prog_select_func(fp);
1796 /* eBPF JITs can rewrite the program in case constant
1797 * blinding is active. However, in case of error during
1798 * blinding, bpf_int_jit_compile() must always return a
1799 * valid program, which in this case would simply not
1800 * be JITed, but falls back to the interpreter.
1802 if (!bpf_prog_is_dev_bound(fp->aux)) {
1803 *err = bpf_prog_alloc_jited_linfo(fp);
1807 fp = bpf_int_jit_compile(fp);
1809 bpf_prog_free_jited_linfo(fp);
1810 #ifdef CONFIG_BPF_JIT_ALWAYS_ON
1815 bpf_prog_free_unused_jited_linfo(fp);
1818 *err = bpf_prog_offload_compile(fp);
1824 bpf_prog_lock_ro(fp);
1826 /* The tail call compatibility check can only be done at
1827 * this late stage as we need to determine, if we deal
1828 * with JITed or non JITed program concatenations and not
1829 * all eBPF JITs might immediately support all features.
1831 *err = bpf_check_tail_call(fp);
1835 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1837 static unsigned int __bpf_prog_ret1(const void *ctx,
1838 const struct bpf_insn *insn)
1843 static struct bpf_prog_dummy {
1844 struct bpf_prog prog;
1845 } dummy_bpf_prog = {
1847 .bpf_func = __bpf_prog_ret1,
1851 /* to avoid allocating empty bpf_prog_array for cgroups that
1852 * don't have bpf program attached use one global 'empty_prog_array'
1853 * It will not be modified the caller of bpf_prog_array_alloc()
1854 * (since caller requested prog_cnt == 0)
1855 * that pointer should be 'freed' by bpf_prog_array_free()
1858 struct bpf_prog_array hdr;
1859 struct bpf_prog *null_prog;
1860 } empty_prog_array = {
1864 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
1867 return kzalloc(sizeof(struct bpf_prog_array) +
1868 sizeof(struct bpf_prog_array_item) *
1872 return &empty_prog_array.hdr;
1875 void bpf_prog_array_free(struct bpf_prog_array *progs)
1877 if (!progs || progs == &empty_prog_array.hdr)
1879 kfree_rcu(progs, rcu);
1882 int bpf_prog_array_length(struct bpf_prog_array *array)
1884 struct bpf_prog_array_item *item;
1887 for (item = array->items; item->prog; item++)
1888 if (item->prog != &dummy_bpf_prog.prog)
1893 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
1895 struct bpf_prog_array_item *item;
1897 for (item = array->items; item->prog; item++)
1898 if (item->prog != &dummy_bpf_prog.prog)
1903 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
1907 struct bpf_prog_array_item *item;
1910 for (item = array->items; item->prog; item++) {
1911 if (item->prog == &dummy_bpf_prog.prog)
1913 prog_ids[i] = item->prog->aux->id;
1914 if (++i == request_cnt) {
1920 return !!(item->prog);
1923 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
1924 __u32 __user *prog_ids, u32 cnt)
1926 unsigned long err = 0;
1930 /* users of this function are doing:
1931 * cnt = bpf_prog_array_length();
1933 * bpf_prog_array_copy_to_user(..., cnt);
1934 * so below kcalloc doesn't need extra cnt > 0 check.
1936 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
1939 nospc = bpf_prog_array_copy_core(array, ids, cnt);
1940 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
1949 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
1950 struct bpf_prog *old_prog)
1952 struct bpf_prog_array_item *item;
1954 for (item = array->items; item->prog; item++)
1955 if (item->prog == old_prog) {
1956 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
1961 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
1962 struct bpf_prog *exclude_prog,
1963 struct bpf_prog *include_prog,
1964 struct bpf_prog_array **new_array)
1966 int new_prog_cnt, carry_prog_cnt = 0;
1967 struct bpf_prog_array_item *existing;
1968 struct bpf_prog_array *array;
1969 bool found_exclude = false;
1970 int new_prog_idx = 0;
1972 /* Figure out how many existing progs we need to carry over to
1976 existing = old_array->items;
1977 for (; existing->prog; existing++) {
1978 if (existing->prog == exclude_prog) {
1979 found_exclude = true;
1982 if (existing->prog != &dummy_bpf_prog.prog)
1984 if (existing->prog == include_prog)
1989 if (exclude_prog && !found_exclude)
1992 /* How many progs (not NULL) will be in the new array? */
1993 new_prog_cnt = carry_prog_cnt;
1997 /* Do we have any prog (not NULL) in the new array? */
1998 if (!new_prog_cnt) {
2003 /* +1 as the end of prog_array is marked with NULL */
2004 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2008 /* Fill in the new prog array */
2009 if (carry_prog_cnt) {
2010 existing = old_array->items;
2011 for (; existing->prog; existing++)
2012 if (existing->prog != exclude_prog &&
2013 existing->prog != &dummy_bpf_prog.prog) {
2014 array->items[new_prog_idx++].prog =
2019 array->items[new_prog_idx++].prog = include_prog;
2020 array->items[new_prog_idx].prog = NULL;
2025 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2026 u32 *prog_ids, u32 request_cnt,
2032 cnt = bpf_prog_array_length(array);
2036 /* return early if user requested only program count or nothing to copy */
2037 if (!request_cnt || !cnt)
2040 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2041 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2045 static void bpf_free_cgroup_storage(struct bpf_prog_aux *aux)
2047 enum bpf_cgroup_storage_type stype;
2049 for_each_cgroup_storage_type(stype) {
2050 if (!aux->cgroup_storage[stype])
2052 bpf_cgroup_storage_release(aux, aux->cgroup_storage[stype]);
2056 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2057 struct bpf_map **used_maps, u32 len)
2059 struct bpf_map *map;
2062 bpf_free_cgroup_storage(aux);
2063 for (i = 0; i < len; i++) {
2065 if (map->ops->map_poke_untrack)
2066 map->ops->map_poke_untrack(map, aux);
2071 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2073 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2074 kfree(aux->used_maps);
2077 static void bpf_prog_free_deferred(struct work_struct *work)
2079 struct bpf_prog_aux *aux;
2082 aux = container_of(work, struct bpf_prog_aux, work);
2083 bpf_free_used_maps(aux);
2084 if (bpf_prog_is_dev_bound(aux))
2085 bpf_prog_offload_destroy(aux->prog);
2086 #ifdef CONFIG_PERF_EVENTS
2087 if (aux->prog->has_callchain_buf)
2088 put_callchain_buffers();
2090 bpf_trampoline_put(aux->trampoline);
2091 for (i = 0; i < aux->func_cnt; i++)
2092 bpf_jit_free(aux->func[i]);
2093 if (aux->func_cnt) {
2095 bpf_prog_unlock_free(aux->prog);
2097 bpf_jit_free(aux->prog);
2101 /* Free internal BPF program */
2102 void bpf_prog_free(struct bpf_prog *fp)
2104 struct bpf_prog_aux *aux = fp->aux;
2106 if (aux->linked_prog)
2107 bpf_prog_put(aux->linked_prog);
2108 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2109 schedule_work(&aux->work);
2111 EXPORT_SYMBOL_GPL(bpf_prog_free);
2113 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2114 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2116 void bpf_user_rnd_init_once(void)
2118 prandom_init_once(&bpf_user_rnd_state);
2121 BPF_CALL_0(bpf_user_rnd_u32)
2123 /* Should someone ever have the rather unwise idea to use some
2124 * of the registers passed into this function, then note that
2125 * this function is called from native eBPF and classic-to-eBPF
2126 * transformations. Register assignments from both sides are
2127 * different, f.e. classic always sets fn(ctx, A, X) here.
2129 struct rnd_state *state;
2132 state = &get_cpu_var(bpf_user_rnd_state);
2133 res = prandom_u32_state(state);
2134 put_cpu_var(bpf_user_rnd_state);
2139 BPF_CALL_0(bpf_get_raw_cpu_id)
2141 return raw_smp_processor_id();
2144 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2145 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2146 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2147 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2148 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2149 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2150 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2151 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2152 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2153 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2155 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2156 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2157 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2158 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2159 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2161 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2162 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2163 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2164 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2165 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2166 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2167 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2169 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2175 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2176 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2180 EXPORT_SYMBOL_GPL(bpf_event_output);
2182 /* Always built-in helper functions. */
2183 const struct bpf_func_proto bpf_tail_call_proto = {
2186 .ret_type = RET_VOID,
2187 .arg1_type = ARG_PTR_TO_CTX,
2188 .arg2_type = ARG_CONST_MAP_PTR,
2189 .arg3_type = ARG_ANYTHING,
2192 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2193 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2194 * eBPF and implicitly also cBPF can get JITed!
2196 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2201 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2202 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2204 void __weak bpf_jit_compile(struct bpf_prog *prog)
2208 bool __weak bpf_helper_changes_pkt_data(void *func)
2213 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2214 * analysis code and wants explicit zero extension inserted by verifier.
2215 * Otherwise, return FALSE.
2217 bool __weak bpf_jit_needs_zext(void)
2222 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2223 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2225 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2231 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2232 void *addr1, void *addr2)
2237 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2238 EXPORT_SYMBOL(bpf_stats_enabled_key);
2240 /* All definitions of tracepoints related to BPF. */
2241 #define CREATE_TRACE_POINTS
2242 #include <linux/bpf_trace.h>
2244 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2245 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);