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
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/objtool.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 <linux/bpf_verifier.h>
36 #include <linux/nodemask.h>
37 #include <linux/bpf_mem_alloc.h>
39 #include <asm/barrier.h>
40 #include <asm/unaligned.h>
43 #define BPF_R0 regs[BPF_REG_0]
44 #define BPF_R1 regs[BPF_REG_1]
45 #define BPF_R2 regs[BPF_REG_2]
46 #define BPF_R3 regs[BPF_REG_3]
47 #define BPF_R4 regs[BPF_REG_4]
48 #define BPF_R5 regs[BPF_REG_5]
49 #define BPF_R6 regs[BPF_REG_6]
50 #define BPF_R7 regs[BPF_REG_7]
51 #define BPF_R8 regs[BPF_REG_8]
52 #define BPF_R9 regs[BPF_REG_9]
53 #define BPF_R10 regs[BPF_REG_10]
56 #define DST regs[insn->dst_reg]
57 #define SRC regs[insn->src_reg]
58 #define FP regs[BPF_REG_FP]
59 #define AX regs[BPF_REG_AX]
60 #define ARG1 regs[BPF_REG_ARG1]
61 #define CTX regs[BPF_REG_CTX]
64 struct bpf_mem_alloc bpf_global_ma;
65 bool bpf_global_ma_set;
67 /* No hurry in this branch
69 * Exported for the bpf jit load helper.
71 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
75 if (k >= SKF_NET_OFF) {
76 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
77 } else if (k >= SKF_LL_OFF) {
78 if (unlikely(!skb_mac_header_was_set(skb)))
80 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
82 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
88 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
90 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
91 struct bpf_prog_aux *aux;
94 size = round_up(size, PAGE_SIZE);
95 fp = __vmalloc(size, gfp_flags);
99 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags);
104 fp->active = alloc_percpu_gfp(int, GFP_KERNEL_ACCOUNT | gfp_extra_flags);
111 fp->pages = size / PAGE_SIZE;
114 fp->jit_requested = ebpf_jit_enabled();
115 fp->blinding_requested = bpf_jit_blinding_enabled(fp);
116 #ifdef CONFIG_CGROUP_BPF
117 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
120 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
121 mutex_init(&fp->aux->used_maps_mutex);
122 mutex_init(&fp->aux->dst_mutex);
127 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
129 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
130 struct bpf_prog *prog;
133 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
137 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
139 free_percpu(prog->active);
145 for_each_possible_cpu(cpu) {
146 struct bpf_prog_stats *pstats;
148 pstats = per_cpu_ptr(prog->stats, cpu);
149 u64_stats_init(&pstats->syncp);
153 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
155 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
157 if (!prog->aux->nr_linfo || !prog->jit_requested)
160 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
161 sizeof(*prog->aux->jited_linfo),
162 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
163 if (!prog->aux->jited_linfo)
169 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
171 if (prog->aux->jited_linfo &&
172 (!prog->jited || !prog->aux->jited_linfo[0])) {
173 kvfree(prog->aux->jited_linfo);
174 prog->aux->jited_linfo = NULL;
177 kfree(prog->aux->kfunc_tab);
178 prog->aux->kfunc_tab = NULL;
181 /* The jit engine is responsible to provide an array
182 * for insn_off to the jited_off mapping (insn_to_jit_off).
184 * The idx to this array is the insn_off. Hence, the insn_off
185 * here is relative to the prog itself instead of the main prog.
186 * This array has one entry for each xlated bpf insn.
188 * jited_off is the byte off to the end of the jited insn.
192 * The first bpf insn off of the prog. The insn off
193 * here is relative to the main prog.
194 * e.g. if prog is a subprog, insn_start > 0
196 * The prog's idx to prog->aux->linfo and jited_linfo
198 * jited_linfo[linfo_idx] = prog->bpf_func
202 * jited_linfo[i] = prog->bpf_func +
203 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
205 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
206 const u32 *insn_to_jit_off)
208 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
209 const struct bpf_line_info *linfo;
212 if (!prog->aux->jited_linfo)
213 /* Userspace did not provide linfo */
216 linfo_idx = prog->aux->linfo_idx;
217 linfo = &prog->aux->linfo[linfo_idx];
218 insn_start = linfo[0].insn_off;
219 insn_end = insn_start + prog->len;
221 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
222 jited_linfo[0] = prog->bpf_func;
224 nr_linfo = prog->aux->nr_linfo - linfo_idx;
226 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
227 /* The verifier ensures that linfo[i].insn_off is
228 * strictly increasing
230 jited_linfo[i] = prog->bpf_func +
231 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
234 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
235 gfp_t gfp_extra_flags)
237 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
241 size = round_up(size, PAGE_SIZE);
242 pages = size / PAGE_SIZE;
243 if (pages <= fp_old->pages)
246 fp = __vmalloc(size, gfp_flags);
248 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
252 /* We keep fp->aux from fp_old around in the new
253 * reallocated structure.
256 fp_old->stats = NULL;
257 fp_old->active = NULL;
258 __bpf_prog_free(fp_old);
264 void __bpf_prog_free(struct bpf_prog *fp)
267 mutex_destroy(&fp->aux->used_maps_mutex);
268 mutex_destroy(&fp->aux->dst_mutex);
269 kfree(fp->aux->poke_tab);
272 free_percpu(fp->stats);
273 free_percpu(fp->active);
277 int bpf_prog_calc_tag(struct bpf_prog *fp)
279 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
280 u32 raw_size = bpf_prog_tag_scratch_size(fp);
281 u32 digest[SHA1_DIGEST_WORDS];
282 u32 ws[SHA1_WORKSPACE_WORDS];
283 u32 i, bsize, psize, blocks;
284 struct bpf_insn *dst;
290 raw = vmalloc(raw_size);
295 memset(ws, 0, sizeof(ws));
297 /* We need to take out the map fd for the digest calculation
298 * since they are unstable from user space side.
301 for (i = 0, was_ld_map = false; i < fp->len; i++) {
302 dst[i] = fp->insnsi[i];
304 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
305 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
306 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
309 } else if (was_ld_map &&
311 dst[i].dst_reg == 0 &&
312 dst[i].src_reg == 0 &&
321 psize = bpf_prog_insn_size(fp);
322 memset(&raw[psize], 0, raw_size - psize);
325 bsize = round_up(psize, SHA1_BLOCK_SIZE);
326 blocks = bsize / SHA1_BLOCK_SIZE;
328 if (bsize - psize >= sizeof(__be64)) {
329 bits = (__be64 *)(todo + bsize - sizeof(__be64));
331 bits = (__be64 *)(todo + bsize + bits_offset);
334 *bits = cpu_to_be64((psize - 1) << 3);
337 sha1_transform(digest, todo, ws);
338 todo += SHA1_BLOCK_SIZE;
341 result = (__force __be32 *)digest;
342 for (i = 0; i < SHA1_DIGEST_WORDS; i++)
343 result[i] = cpu_to_be32(digest[i]);
344 memcpy(fp->tag, result, sizeof(fp->tag));
350 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
351 s32 end_new, s32 curr, const bool probe_pass)
353 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
354 s32 delta = end_new - end_old;
357 if (curr < pos && curr + imm + 1 >= end_old)
359 else if (curr >= end_new && curr + imm + 1 < end_new)
361 if (imm < imm_min || imm > imm_max)
368 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
369 s32 end_new, s32 curr, const bool probe_pass)
371 const s32 off_min = S16_MIN, off_max = S16_MAX;
372 s32 delta = end_new - end_old;
375 if (curr < pos && curr + off + 1 >= end_old)
377 else if (curr >= end_new && curr + off + 1 < end_new)
379 if (off < off_min || off > off_max)
386 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
387 s32 end_new, const bool probe_pass)
389 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
390 struct bpf_insn *insn = prog->insnsi;
393 for (i = 0; i < insn_cnt; i++, insn++) {
396 /* In the probing pass we still operate on the original,
397 * unpatched image in order to check overflows before we
398 * do any other adjustments. Therefore skip the patchlet.
400 if (probe_pass && i == pos) {
402 insn = prog->insnsi + end_old;
404 if (bpf_pseudo_func(insn)) {
405 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
406 end_new, i, probe_pass);
412 if ((BPF_CLASS(code) != BPF_JMP &&
413 BPF_CLASS(code) != BPF_JMP32) ||
414 BPF_OP(code) == BPF_EXIT)
416 /* Adjust offset of jmps if we cross patch boundaries. */
417 if (BPF_OP(code) == BPF_CALL) {
418 if (insn->src_reg != BPF_PSEUDO_CALL)
420 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
421 end_new, i, probe_pass);
423 ret = bpf_adj_delta_to_off(insn, pos, end_old,
424 end_new, i, probe_pass);
433 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
435 struct bpf_line_info *linfo;
438 nr_linfo = prog->aux->nr_linfo;
439 if (!nr_linfo || !delta)
442 linfo = prog->aux->linfo;
444 for (i = 0; i < nr_linfo; i++)
445 if (off < linfo[i].insn_off)
448 /* Push all off < linfo[i].insn_off by delta */
449 for (; i < nr_linfo; i++)
450 linfo[i].insn_off += delta;
453 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
454 const struct bpf_insn *patch, u32 len)
456 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
457 const u32 cnt_max = S16_MAX;
458 struct bpf_prog *prog_adj;
461 /* Since our patchlet doesn't expand the image, we're done. */
462 if (insn_delta == 0) {
463 memcpy(prog->insnsi + off, patch, sizeof(*patch));
467 insn_adj_cnt = prog->len + insn_delta;
469 /* Reject anything that would potentially let the insn->off
470 * target overflow when we have excessive program expansions.
471 * We need to probe here before we do any reallocation where
472 * we afterwards may not fail anymore.
474 if (insn_adj_cnt > cnt_max &&
475 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
478 /* Several new instructions need to be inserted. Make room
479 * for them. Likely, there's no need for a new allocation as
480 * last page could have large enough tailroom.
482 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
485 return ERR_PTR(-ENOMEM);
487 prog_adj->len = insn_adj_cnt;
489 /* Patching happens in 3 steps:
491 * 1) Move over tail of insnsi from next instruction onwards,
492 * so we can patch the single target insn with one or more
493 * new ones (patching is always from 1 to n insns, n > 0).
494 * 2) Inject new instructions at the target location.
495 * 3) Adjust branch offsets if necessary.
497 insn_rest = insn_adj_cnt - off - len;
499 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
500 sizeof(*patch) * insn_rest);
501 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
503 /* We are guaranteed to not fail at this point, otherwise
504 * the ship has sailed to reverse to the original state. An
505 * overflow cannot happen at this point.
507 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
509 bpf_adj_linfo(prog_adj, off, insn_delta);
514 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
516 /* Branch offsets can't overflow when program is shrinking, no need
517 * to call bpf_adj_branches(..., true) here
519 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
520 sizeof(struct bpf_insn) * (prog->len - off - cnt));
523 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
526 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
530 for (i = 0; i < fp->aux->func_cnt; i++)
531 bpf_prog_kallsyms_del(fp->aux->func[i]);
534 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
536 bpf_prog_kallsyms_del_subprogs(fp);
537 bpf_prog_kallsyms_del(fp);
540 #ifdef CONFIG_BPF_JIT
541 /* All BPF JIT sysctl knobs here. */
542 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
543 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
544 int bpf_jit_harden __read_mostly;
545 long bpf_jit_limit __read_mostly;
546 long bpf_jit_limit_max __read_mostly;
549 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
551 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
553 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
554 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len;
558 bpf_prog_ksym_set_name(struct bpf_prog *prog)
560 char *sym = prog->aux->ksym.name;
561 const char *end = sym + KSYM_NAME_LEN;
562 const struct btf_type *type;
563 const char *func_name;
565 BUILD_BUG_ON(sizeof("bpf_prog_") +
566 sizeof(prog->tag) * 2 +
567 /* name has been null terminated.
568 * We should need +1 for the '_' preceding
569 * the name. However, the null character
570 * is double counted between the name and the
571 * sizeof("bpf_prog_") above, so we omit
574 sizeof(prog->aux->name) > KSYM_NAME_LEN);
576 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
577 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
579 /* prog->aux->name will be ignored if full btf name is available */
580 if (prog->aux->func_info_cnt) {
581 type = btf_type_by_id(prog->aux->btf,
582 prog->aux->func_info[prog->aux->func_idx].type_id);
583 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
584 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
588 if (prog->aux->name[0])
589 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
594 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
596 return container_of(n, struct bpf_ksym, tnode)->start;
599 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
600 struct latch_tree_node *b)
602 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
605 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
607 unsigned long val = (unsigned long)key;
608 const struct bpf_ksym *ksym;
610 ksym = container_of(n, struct bpf_ksym, tnode);
612 if (val < ksym->start)
614 if (val >= ksym->end)
620 static const struct latch_tree_ops bpf_tree_ops = {
621 .less = bpf_tree_less,
622 .comp = bpf_tree_comp,
625 static DEFINE_SPINLOCK(bpf_lock);
626 static LIST_HEAD(bpf_kallsyms);
627 static struct latch_tree_root bpf_tree __cacheline_aligned;
629 void bpf_ksym_add(struct bpf_ksym *ksym)
631 spin_lock_bh(&bpf_lock);
632 WARN_ON_ONCE(!list_empty(&ksym->lnode));
633 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
634 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
635 spin_unlock_bh(&bpf_lock);
638 static void __bpf_ksym_del(struct bpf_ksym *ksym)
640 if (list_empty(&ksym->lnode))
643 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
644 list_del_rcu(&ksym->lnode);
647 void bpf_ksym_del(struct bpf_ksym *ksym)
649 spin_lock_bh(&bpf_lock);
650 __bpf_ksym_del(ksym);
651 spin_unlock_bh(&bpf_lock);
654 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
656 return fp->jited && !bpf_prog_was_classic(fp);
659 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
661 if (!bpf_prog_kallsyms_candidate(fp) ||
665 bpf_prog_ksym_set_addr(fp);
666 bpf_prog_ksym_set_name(fp);
667 fp->aux->ksym.prog = true;
669 bpf_ksym_add(&fp->aux->ksym);
672 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
674 if (!bpf_prog_kallsyms_candidate(fp))
677 bpf_ksym_del(&fp->aux->ksym);
680 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
682 struct latch_tree_node *n;
684 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
685 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
688 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
689 unsigned long *off, char *sym)
691 struct bpf_ksym *ksym;
695 ksym = bpf_ksym_find(addr);
697 unsigned long symbol_start = ksym->start;
698 unsigned long symbol_end = ksym->end;
700 strncpy(sym, ksym->name, KSYM_NAME_LEN);
704 *size = symbol_end - symbol_start;
706 *off = addr - symbol_start;
713 bool is_bpf_text_address(unsigned long addr)
718 ret = bpf_ksym_find(addr) != NULL;
724 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
726 struct bpf_ksym *ksym = bpf_ksym_find(addr);
728 return ksym && ksym->prog ?
729 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
733 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
735 const struct exception_table_entry *e = NULL;
736 struct bpf_prog *prog;
739 prog = bpf_prog_ksym_find(addr);
742 if (!prog->aux->num_exentries)
745 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
751 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
754 struct bpf_ksym *ksym;
758 if (!bpf_jit_kallsyms_enabled())
762 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
766 strncpy(sym, ksym->name, KSYM_NAME_LEN);
768 *value = ksym->start;
769 *type = BPF_SYM_ELF_TYPE;
779 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
780 struct bpf_jit_poke_descriptor *poke)
782 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
783 static const u32 poke_tab_max = 1024;
784 u32 slot = prog->aux->size_poke_tab;
787 if (size > poke_tab_max)
789 if (poke->tailcall_target || poke->tailcall_target_stable ||
790 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
793 switch (poke->reason) {
794 case BPF_POKE_REASON_TAIL_CALL:
795 if (!poke->tail_call.map)
802 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
806 memcpy(&tab[slot], poke, sizeof(*poke));
807 prog->aux->size_poke_tab = size;
808 prog->aux->poke_tab = tab;
814 * BPF program pack allocator.
816 * Most BPF programs are pretty small. Allocating a hole page for each
817 * program is sometime a waste. Many small bpf program also adds pressure
818 * to instruction TLB. To solve this issue, we introduce a BPF program pack
819 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
820 * to host BPF programs.
822 #define BPF_PROG_CHUNK_SHIFT 6
823 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT)
824 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1))
826 struct bpf_prog_pack {
827 struct list_head list;
829 unsigned long bitmap[];
832 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size)
834 memset(area, 0, size);
837 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
839 static DEFINE_MUTEX(pack_mutex);
840 static LIST_HEAD(pack_list);
842 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
843 * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
846 #define BPF_PROG_PACK_SIZE (PMD_SIZE * num_possible_nodes())
848 #define BPF_PROG_PACK_SIZE PAGE_SIZE
851 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
853 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
855 struct bpf_prog_pack *pack;
857 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
861 pack->ptr = module_alloc(BPF_PROG_PACK_SIZE);
866 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
867 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
868 list_add_tail(&pack->list, &pack_list);
870 set_vm_flush_reset_perms(pack->ptr);
871 set_memory_rox((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
875 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
877 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
878 struct bpf_prog_pack *pack;
882 mutex_lock(&pack_mutex);
883 if (size > BPF_PROG_PACK_SIZE) {
884 size = round_up(size, PAGE_SIZE);
885 ptr = module_alloc(size);
887 bpf_fill_ill_insns(ptr, size);
888 set_vm_flush_reset_perms(ptr);
889 set_memory_rox((unsigned long)ptr, size / PAGE_SIZE);
893 list_for_each_entry(pack, &pack_list, list) {
894 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
896 if (pos < BPF_PROG_CHUNK_COUNT)
897 goto found_free_area;
900 pack = alloc_new_pack(bpf_fill_ill_insns);
907 bitmap_set(pack->bitmap, pos, nbits);
908 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
911 mutex_unlock(&pack_mutex);
915 void bpf_prog_pack_free(struct bpf_binary_header *hdr)
917 struct bpf_prog_pack *pack = NULL, *tmp;
921 mutex_lock(&pack_mutex);
922 if (hdr->size > BPF_PROG_PACK_SIZE) {
927 list_for_each_entry(tmp, &pack_list, list) {
928 if ((void *)hdr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > (void *)hdr) {
934 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
937 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size);
938 pos = ((unsigned long)hdr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
940 WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size),
941 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
943 bitmap_clear(pack->bitmap, pos, nbits);
944 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
945 BPF_PROG_CHUNK_COUNT, 0) == 0) {
946 list_del(&pack->list);
947 module_memfree(pack->ptr);
951 mutex_unlock(&pack_mutex);
954 static atomic_long_t bpf_jit_current;
956 /* Can be overridden by an arch's JIT compiler if it has a custom,
957 * dedicated BPF backend memory area, or if neither of the two
960 u64 __weak bpf_jit_alloc_exec_limit(void)
962 #if defined(MODULES_VADDR)
963 return MODULES_END - MODULES_VADDR;
965 return VMALLOC_END - VMALLOC_START;
969 static int __init bpf_jit_charge_init(void)
971 /* Only used as heuristic here to derive limit. */
972 bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
973 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2,
974 PAGE_SIZE), LONG_MAX);
977 pure_initcall(bpf_jit_charge_init);
979 int bpf_jit_charge_modmem(u32 size)
981 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
982 if (!bpf_capable()) {
983 atomic_long_sub(size, &bpf_jit_current);
991 void bpf_jit_uncharge_modmem(u32 size)
993 atomic_long_sub(size, &bpf_jit_current);
996 void *__weak bpf_jit_alloc_exec(unsigned long size)
998 return module_alloc(size);
1001 void __weak bpf_jit_free_exec(void *addr)
1003 module_memfree(addr);
1006 struct bpf_binary_header *
1007 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1008 unsigned int alignment,
1009 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1011 struct bpf_binary_header *hdr;
1012 u32 size, hole, start;
1014 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1015 alignment > BPF_IMAGE_ALIGNMENT);
1017 /* Most of BPF filters are really small, but if some of them
1018 * fill a page, allow at least 128 extra bytes to insert a
1019 * random section of illegal instructions.
1021 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1023 if (bpf_jit_charge_modmem(size))
1025 hdr = bpf_jit_alloc_exec(size);
1027 bpf_jit_uncharge_modmem(size);
1031 /* Fill space with illegal/arch-dep instructions. */
1032 bpf_fill_ill_insns(hdr, size);
1035 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1036 PAGE_SIZE - sizeof(*hdr));
1037 start = get_random_u32_below(hole) & ~(alignment - 1);
1039 /* Leave a random number of instructions before BPF code. */
1040 *image_ptr = &hdr->image[start];
1045 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1047 u32 size = hdr->size;
1049 bpf_jit_free_exec(hdr);
1050 bpf_jit_uncharge_modmem(size);
1053 /* Allocate jit binary from bpf_prog_pack allocator.
1054 * Since the allocated memory is RO+X, the JIT engine cannot write directly
1055 * to the memory. To solve this problem, a RW buffer is also allocated at
1056 * as the same time. The JIT engine should calculate offsets based on the
1057 * RO memory address, but write JITed program to the RW buffer. Once the
1058 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1059 * the JITed program to the RO memory.
1061 struct bpf_binary_header *
1062 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1063 unsigned int alignment,
1064 struct bpf_binary_header **rw_header,
1066 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1068 struct bpf_binary_header *ro_header;
1069 u32 size, hole, start;
1071 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1072 alignment > BPF_IMAGE_ALIGNMENT);
1074 /* add 16 bytes for a random section of illegal instructions */
1075 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1077 if (bpf_jit_charge_modmem(size))
1079 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1081 bpf_jit_uncharge_modmem(size);
1085 *rw_header = kvmalloc(size, GFP_KERNEL);
1087 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size));
1088 bpf_prog_pack_free(ro_header);
1089 bpf_jit_uncharge_modmem(size);
1093 /* Fill space with illegal/arch-dep instructions. */
1094 bpf_fill_ill_insns(*rw_header, size);
1095 (*rw_header)->size = size;
1097 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1098 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1099 start = get_random_u32_below(hole) & ~(alignment - 1);
1101 *image_ptr = &ro_header->image[start];
1102 *rw_image = &(*rw_header)->image[start];
1107 /* Copy JITed text from rw_header to its final location, the ro_header. */
1108 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog,
1109 struct bpf_binary_header *ro_header,
1110 struct bpf_binary_header *rw_header)
1114 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1119 bpf_prog_pack_free(ro_header);
1120 return PTR_ERR(ptr);
1125 /* bpf_jit_binary_pack_free is called in two different scenarios:
1126 * 1) when the program is freed after;
1127 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1128 * For case 2), we need to free both the RO memory and the RW buffer.
1130 * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1131 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1132 * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1133 * bpf_arch_text_copy (when jit fails).
1135 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1136 struct bpf_binary_header *rw_header)
1138 u32 size = ro_header->size;
1140 bpf_prog_pack_free(ro_header);
1142 bpf_jit_uncharge_modmem(size);
1145 struct bpf_binary_header *
1146 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1148 unsigned long real_start = (unsigned long)fp->bpf_func;
1151 addr = real_start & BPF_PROG_CHUNK_MASK;
1152 return (void *)addr;
1155 static inline struct bpf_binary_header *
1156 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1158 unsigned long real_start = (unsigned long)fp->bpf_func;
1161 addr = real_start & PAGE_MASK;
1162 return (void *)addr;
1165 /* This symbol is only overridden by archs that have different
1166 * requirements than the usual eBPF JITs, f.e. when they only
1167 * implement cBPF JIT, do not set images read-only, etc.
1169 void __weak bpf_jit_free(struct bpf_prog *fp)
1172 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1174 bpf_jit_binary_free(hdr);
1175 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1178 bpf_prog_unlock_free(fp);
1181 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1182 const struct bpf_insn *insn, bool extra_pass,
1183 u64 *func_addr, bool *func_addr_fixed)
1185 s16 off = insn->off;
1186 s32 imm = insn->imm;
1189 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1190 if (!*func_addr_fixed) {
1191 /* Place-holder address till the last pass has collected
1192 * all addresses for JITed subprograms in which case we
1193 * can pick them up from prog->aux.
1197 else if (prog->aux->func &&
1198 off >= 0 && off < prog->aux->func_cnt)
1199 addr = (u8 *)prog->aux->func[off]->bpf_func;
1203 /* Address of a BPF helper call. Since part of the core
1204 * kernel, it's always at a fixed location. __bpf_call_base
1205 * and the helper with imm relative to it are both in core
1208 addr = (u8 *)__bpf_call_base + imm;
1211 *func_addr = (unsigned long)addr;
1215 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1216 const struct bpf_insn *aux,
1217 struct bpf_insn *to_buff,
1220 struct bpf_insn *to = to_buff;
1221 u32 imm_rnd = get_random_u32();
1224 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
1225 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1227 /* Constraints on AX register:
1229 * AX register is inaccessible from user space. It is mapped in
1230 * all JITs, and used here for constant blinding rewrites. It is
1231 * typically "stateless" meaning its contents are only valid within
1232 * the executed instruction, but not across several instructions.
1233 * There are a few exceptions however which are further detailed
1236 * Constant blinding is only used by JITs, not in the interpreter.
1237 * The interpreter uses AX in some occasions as a local temporary
1238 * register e.g. in DIV or MOD instructions.
1240 * In restricted circumstances, the verifier can also use the AX
1241 * register for rewrites as long as they do not interfere with
1244 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1247 if (from->imm == 0 &&
1248 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
1249 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1250 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1254 switch (from->code) {
1255 case BPF_ALU | BPF_ADD | BPF_K:
1256 case BPF_ALU | BPF_SUB | BPF_K:
1257 case BPF_ALU | BPF_AND | BPF_K:
1258 case BPF_ALU | BPF_OR | BPF_K:
1259 case BPF_ALU | BPF_XOR | BPF_K:
1260 case BPF_ALU | BPF_MUL | BPF_K:
1261 case BPF_ALU | BPF_MOV | BPF_K:
1262 case BPF_ALU | BPF_DIV | BPF_K:
1263 case BPF_ALU | BPF_MOD | BPF_K:
1264 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1265 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1266 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1269 case BPF_ALU64 | BPF_ADD | BPF_K:
1270 case BPF_ALU64 | BPF_SUB | BPF_K:
1271 case BPF_ALU64 | BPF_AND | BPF_K:
1272 case BPF_ALU64 | BPF_OR | BPF_K:
1273 case BPF_ALU64 | BPF_XOR | BPF_K:
1274 case BPF_ALU64 | BPF_MUL | BPF_K:
1275 case BPF_ALU64 | BPF_MOV | BPF_K:
1276 case BPF_ALU64 | BPF_DIV | BPF_K:
1277 case BPF_ALU64 | BPF_MOD | BPF_K:
1278 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1279 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1280 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1283 case BPF_JMP | BPF_JEQ | BPF_K:
1284 case BPF_JMP | BPF_JNE | BPF_K:
1285 case BPF_JMP | BPF_JGT | BPF_K:
1286 case BPF_JMP | BPF_JLT | BPF_K:
1287 case BPF_JMP | BPF_JGE | BPF_K:
1288 case BPF_JMP | BPF_JLE | BPF_K:
1289 case BPF_JMP | BPF_JSGT | BPF_K:
1290 case BPF_JMP | BPF_JSLT | BPF_K:
1291 case BPF_JMP | BPF_JSGE | BPF_K:
1292 case BPF_JMP | BPF_JSLE | BPF_K:
1293 case BPF_JMP | BPF_JSET | BPF_K:
1294 /* Accommodate for extra offset in case of a backjump. */
1298 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1299 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1300 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1303 case BPF_JMP32 | BPF_JEQ | BPF_K:
1304 case BPF_JMP32 | BPF_JNE | BPF_K:
1305 case BPF_JMP32 | BPF_JGT | BPF_K:
1306 case BPF_JMP32 | BPF_JLT | BPF_K:
1307 case BPF_JMP32 | BPF_JGE | BPF_K:
1308 case BPF_JMP32 | BPF_JLE | BPF_K:
1309 case BPF_JMP32 | BPF_JSGT | BPF_K:
1310 case BPF_JMP32 | BPF_JSLT | BPF_K:
1311 case BPF_JMP32 | BPF_JSGE | BPF_K:
1312 case BPF_JMP32 | BPF_JSLE | BPF_K:
1313 case BPF_JMP32 | BPF_JSET | BPF_K:
1314 /* Accommodate for extra offset in case of a backjump. */
1318 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1319 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1320 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1324 case BPF_LD | BPF_IMM | BPF_DW:
1325 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1326 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1327 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1328 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1330 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1331 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1332 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1334 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1335 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1338 case BPF_ST | BPF_MEM | BPF_DW:
1339 case BPF_ST | BPF_MEM | BPF_W:
1340 case BPF_ST | BPF_MEM | BPF_H:
1341 case BPF_ST | BPF_MEM | BPF_B:
1342 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1343 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1344 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1348 return to - to_buff;
1351 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1352 gfp_t gfp_extra_flags)
1354 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1355 struct bpf_prog *fp;
1357 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1359 /* aux->prog still points to the fp_other one, so
1360 * when promoting the clone to the real program,
1361 * this still needs to be adapted.
1363 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1369 static void bpf_prog_clone_free(struct bpf_prog *fp)
1371 /* aux was stolen by the other clone, so we cannot free
1372 * it from this path! It will be freed eventually by the
1373 * other program on release.
1375 * At this point, we don't need a deferred release since
1376 * clone is guaranteed to not be locked.
1381 __bpf_prog_free(fp);
1384 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1386 /* We have to repoint aux->prog to self, as we don't
1387 * know whether fp here is the clone or the original.
1390 bpf_prog_clone_free(fp_other);
1393 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1395 struct bpf_insn insn_buff[16], aux[2];
1396 struct bpf_prog *clone, *tmp;
1397 int insn_delta, insn_cnt;
1398 struct bpf_insn *insn;
1401 if (!prog->blinding_requested || prog->blinded)
1404 clone = bpf_prog_clone_create(prog, GFP_USER);
1406 return ERR_PTR(-ENOMEM);
1408 insn_cnt = clone->len;
1409 insn = clone->insnsi;
1411 for (i = 0; i < insn_cnt; i++, insn++) {
1412 if (bpf_pseudo_func(insn)) {
1413 /* ld_imm64 with an address of bpf subprog is not
1414 * a user controlled constant. Don't randomize it,
1415 * since it will conflict with jit_subprogs() logic.
1422 /* We temporarily need to hold the original ld64 insn
1423 * so that we can still access the first part in the
1424 * second blinding run.
1426 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1428 memcpy(aux, insn, sizeof(aux));
1430 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1431 clone->aux->verifier_zext);
1435 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1437 /* Patching may have repointed aux->prog during
1438 * realloc from the original one, so we need to
1439 * fix it up here on error.
1441 bpf_jit_prog_release_other(prog, clone);
1446 insn_delta = rewritten - 1;
1448 /* Walk new program and skip insns we just inserted. */
1449 insn = clone->insnsi + i + insn_delta;
1450 insn_cnt += insn_delta;
1457 #endif /* CONFIG_BPF_JIT */
1459 /* Base function for offset calculation. Needs to go into .text section,
1460 * therefore keeping it non-static as well; will also be used by JITs
1461 * anyway later on, so do not let the compiler omit it. This also needs
1462 * to go into kallsyms for correlation from e.g. bpftool, so naming
1465 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1469 EXPORT_SYMBOL_GPL(__bpf_call_base);
1471 /* All UAPI available opcodes. */
1472 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1473 /* 32 bit ALU operations. */ \
1474 /* Register based. */ \
1475 INSN_3(ALU, ADD, X), \
1476 INSN_3(ALU, SUB, X), \
1477 INSN_3(ALU, AND, X), \
1478 INSN_3(ALU, OR, X), \
1479 INSN_3(ALU, LSH, X), \
1480 INSN_3(ALU, RSH, X), \
1481 INSN_3(ALU, XOR, X), \
1482 INSN_3(ALU, MUL, X), \
1483 INSN_3(ALU, MOV, X), \
1484 INSN_3(ALU, ARSH, X), \
1485 INSN_3(ALU, DIV, X), \
1486 INSN_3(ALU, MOD, X), \
1488 INSN_3(ALU, END, TO_BE), \
1489 INSN_3(ALU, END, TO_LE), \
1490 /* Immediate based. */ \
1491 INSN_3(ALU, ADD, K), \
1492 INSN_3(ALU, SUB, K), \
1493 INSN_3(ALU, AND, K), \
1494 INSN_3(ALU, OR, K), \
1495 INSN_3(ALU, LSH, K), \
1496 INSN_3(ALU, RSH, K), \
1497 INSN_3(ALU, XOR, K), \
1498 INSN_3(ALU, MUL, K), \
1499 INSN_3(ALU, MOV, K), \
1500 INSN_3(ALU, ARSH, K), \
1501 INSN_3(ALU, DIV, K), \
1502 INSN_3(ALU, MOD, K), \
1503 /* 64 bit ALU operations. */ \
1504 /* Register based. */ \
1505 INSN_3(ALU64, ADD, X), \
1506 INSN_3(ALU64, SUB, X), \
1507 INSN_3(ALU64, AND, X), \
1508 INSN_3(ALU64, OR, X), \
1509 INSN_3(ALU64, LSH, X), \
1510 INSN_3(ALU64, RSH, X), \
1511 INSN_3(ALU64, XOR, X), \
1512 INSN_3(ALU64, MUL, X), \
1513 INSN_3(ALU64, MOV, X), \
1514 INSN_3(ALU64, ARSH, X), \
1515 INSN_3(ALU64, DIV, X), \
1516 INSN_3(ALU64, MOD, X), \
1517 INSN_2(ALU64, NEG), \
1518 /* Immediate based. */ \
1519 INSN_3(ALU64, ADD, K), \
1520 INSN_3(ALU64, SUB, K), \
1521 INSN_3(ALU64, AND, K), \
1522 INSN_3(ALU64, OR, K), \
1523 INSN_3(ALU64, LSH, K), \
1524 INSN_3(ALU64, RSH, K), \
1525 INSN_3(ALU64, XOR, K), \
1526 INSN_3(ALU64, MUL, K), \
1527 INSN_3(ALU64, MOV, K), \
1528 INSN_3(ALU64, ARSH, K), \
1529 INSN_3(ALU64, DIV, K), \
1530 INSN_3(ALU64, MOD, K), \
1531 /* Call instruction. */ \
1532 INSN_2(JMP, CALL), \
1533 /* Exit instruction. */ \
1534 INSN_2(JMP, EXIT), \
1535 /* 32-bit Jump instructions. */ \
1536 /* Register based. */ \
1537 INSN_3(JMP32, JEQ, X), \
1538 INSN_3(JMP32, JNE, X), \
1539 INSN_3(JMP32, JGT, X), \
1540 INSN_3(JMP32, JLT, X), \
1541 INSN_3(JMP32, JGE, X), \
1542 INSN_3(JMP32, JLE, X), \
1543 INSN_3(JMP32, JSGT, X), \
1544 INSN_3(JMP32, JSLT, X), \
1545 INSN_3(JMP32, JSGE, X), \
1546 INSN_3(JMP32, JSLE, X), \
1547 INSN_3(JMP32, JSET, X), \
1548 /* Immediate based. */ \
1549 INSN_3(JMP32, JEQ, K), \
1550 INSN_3(JMP32, JNE, K), \
1551 INSN_3(JMP32, JGT, K), \
1552 INSN_3(JMP32, JLT, K), \
1553 INSN_3(JMP32, JGE, K), \
1554 INSN_3(JMP32, JLE, K), \
1555 INSN_3(JMP32, JSGT, K), \
1556 INSN_3(JMP32, JSLT, K), \
1557 INSN_3(JMP32, JSGE, K), \
1558 INSN_3(JMP32, JSLE, K), \
1559 INSN_3(JMP32, JSET, K), \
1560 /* Jump instructions. */ \
1561 /* Register based. */ \
1562 INSN_3(JMP, JEQ, X), \
1563 INSN_3(JMP, JNE, X), \
1564 INSN_3(JMP, JGT, X), \
1565 INSN_3(JMP, JLT, X), \
1566 INSN_3(JMP, JGE, X), \
1567 INSN_3(JMP, JLE, X), \
1568 INSN_3(JMP, JSGT, X), \
1569 INSN_3(JMP, JSLT, X), \
1570 INSN_3(JMP, JSGE, X), \
1571 INSN_3(JMP, JSLE, X), \
1572 INSN_3(JMP, JSET, X), \
1573 /* Immediate based. */ \
1574 INSN_3(JMP, JEQ, K), \
1575 INSN_3(JMP, JNE, K), \
1576 INSN_3(JMP, JGT, K), \
1577 INSN_3(JMP, JLT, K), \
1578 INSN_3(JMP, JGE, K), \
1579 INSN_3(JMP, JLE, K), \
1580 INSN_3(JMP, JSGT, K), \
1581 INSN_3(JMP, JSLT, K), \
1582 INSN_3(JMP, JSGE, K), \
1583 INSN_3(JMP, JSLE, K), \
1584 INSN_3(JMP, JSET, K), \
1586 /* Store instructions. */ \
1587 /* Register based. */ \
1588 INSN_3(STX, MEM, B), \
1589 INSN_3(STX, MEM, H), \
1590 INSN_3(STX, MEM, W), \
1591 INSN_3(STX, MEM, DW), \
1592 INSN_3(STX, ATOMIC, W), \
1593 INSN_3(STX, ATOMIC, DW), \
1594 /* Immediate based. */ \
1595 INSN_3(ST, MEM, B), \
1596 INSN_3(ST, MEM, H), \
1597 INSN_3(ST, MEM, W), \
1598 INSN_3(ST, MEM, DW), \
1599 /* Load instructions. */ \
1600 /* Register based. */ \
1601 INSN_3(LDX, MEM, B), \
1602 INSN_3(LDX, MEM, H), \
1603 INSN_3(LDX, MEM, W), \
1604 INSN_3(LDX, MEM, DW), \
1605 /* Immediate based. */ \
1608 bool bpf_opcode_in_insntable(u8 code)
1610 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1611 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1612 static const bool public_insntable[256] = {
1613 [0 ... 255] = false,
1614 /* Now overwrite non-defaults ... */
1615 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1616 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1617 [BPF_LD | BPF_ABS | BPF_B] = true,
1618 [BPF_LD | BPF_ABS | BPF_H] = true,
1619 [BPF_LD | BPF_ABS | BPF_W] = true,
1620 [BPF_LD | BPF_IND | BPF_B] = true,
1621 [BPF_LD | BPF_IND | BPF_H] = true,
1622 [BPF_LD | BPF_IND | BPF_W] = true,
1624 #undef BPF_INSN_3_TBL
1625 #undef BPF_INSN_2_TBL
1626 return public_insntable[code];
1629 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1630 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1632 memset(dst, 0, size);
1637 * ___bpf_prog_run - run eBPF program on a given context
1638 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1639 * @insn: is the array of eBPF instructions
1641 * Decode and execute eBPF instructions.
1643 * Return: whatever value is in %BPF_R0 at program exit
1645 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1647 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1648 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1649 static const void * const jumptable[256] __annotate_jump_table = {
1650 [0 ... 255] = &&default_label,
1651 /* Now overwrite non-defaults ... */
1652 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1653 /* Non-UAPI available opcodes. */
1654 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1655 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1656 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1657 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1658 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1659 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1660 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1662 #undef BPF_INSN_3_LBL
1663 #undef BPF_INSN_2_LBL
1664 u32 tail_call_cnt = 0;
1666 #define CONT ({ insn++; goto select_insn; })
1667 #define CONT_JMP ({ insn++; goto select_insn; })
1670 goto *jumptable[insn->code];
1672 /* Explicitly mask the register-based shift amounts with 63 or 31
1673 * to avoid undefined behavior. Normally this won't affect the
1674 * generated code, for example, in case of native 64 bit archs such
1675 * as x86-64 or arm64, the compiler is optimizing the AND away for
1676 * the interpreter. In case of JITs, each of the JIT backends compiles
1677 * the BPF shift operations to machine instructions which produce
1678 * implementation-defined results in such a case; the resulting
1679 * contents of the register may be arbitrary, but program behaviour
1680 * as a whole remains defined. In other words, in case of JIT backends,
1681 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1684 #define SHT(OPCODE, OP) \
1685 ALU64_##OPCODE##_X: \
1686 DST = DST OP (SRC & 63); \
1689 DST = (u32) DST OP ((u32) SRC & 31); \
1691 ALU64_##OPCODE##_K: \
1695 DST = (u32) DST OP (u32) IMM; \
1698 #define ALU(OPCODE, OP) \
1699 ALU64_##OPCODE##_X: \
1703 DST = (u32) DST OP (u32) SRC; \
1705 ALU64_##OPCODE##_K: \
1709 DST = (u32) DST OP (u32) IMM; \
1740 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1744 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1747 DST = (u64) (u32) (((s32) DST) >> IMM);
1750 (*(s64 *) &DST) >>= (SRC & 63);
1753 (*(s64 *) &DST) >>= IMM;
1756 div64_u64_rem(DST, SRC, &AX);
1761 DST = do_div(AX, (u32) SRC);
1764 div64_u64_rem(DST, IMM, &AX);
1769 DST = do_div(AX, (u32) IMM);
1772 DST = div64_u64(DST, SRC);
1776 do_div(AX, (u32) SRC);
1780 DST = div64_u64(DST, IMM);
1784 do_div(AX, (u32) IMM);
1790 DST = (__force u16) cpu_to_be16(DST);
1793 DST = (__force u32) cpu_to_be32(DST);
1796 DST = (__force u64) cpu_to_be64(DST);
1803 DST = (__force u16) cpu_to_le16(DST);
1806 DST = (__force u32) cpu_to_le32(DST);
1809 DST = (__force u64) cpu_to_le64(DST);
1816 /* Function call scratches BPF_R1-BPF_R5 registers,
1817 * preserves BPF_R6-BPF_R9, and stores return value
1820 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1825 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1828 insn + insn->off + 1);
1832 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1833 struct bpf_array *array = container_of(map, struct bpf_array, map);
1834 struct bpf_prog *prog;
1837 if (unlikely(index >= array->map.max_entries))
1840 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
1845 prog = READ_ONCE(array->ptrs[index]);
1849 /* ARG1 at this point is guaranteed to point to CTX from
1850 * the verifier side due to the fact that the tail call is
1851 * handled like a helper, that is, bpf_tail_call_proto,
1852 * where arg1_type is ARG_PTR_TO_CTX.
1854 insn = prog->insnsi;
1865 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1867 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1868 insn += insn->off; \
1872 JMP32_##OPCODE##_X: \
1873 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1874 insn += insn->off; \
1879 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1880 insn += insn->off; \
1884 JMP32_##OPCODE##_K: \
1885 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1886 insn += insn->off; \
1890 COND_JMP(u, JEQ, ==)
1891 COND_JMP(u, JNE, !=)
1894 COND_JMP(u, JGE, >=)
1895 COND_JMP(u, JLE, <=)
1896 COND_JMP(u, JSET, &)
1897 COND_JMP(s, JSGT, >)
1898 COND_JMP(s, JSLT, <)
1899 COND_JMP(s, JSGE, >=)
1900 COND_JMP(s, JSLE, <=)
1902 /* ST, STX and LDX*/
1904 /* Speculation barrier for mitigating Speculative Store Bypass.
1905 * In case of arm64, we rely on the firmware mitigation as
1906 * controlled via the ssbd kernel parameter. Whenever the
1907 * mitigation is enabled, it works for all of the kernel code
1908 * with no need to provide any additional instructions here.
1909 * In case of x86, we use 'lfence' insn for mitigation. We
1910 * reuse preexisting logic from Spectre v1 mitigation that
1911 * happens to produce the required code on x86 for v4 as well.
1917 #define LDST(SIZEOP, SIZE) \
1919 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1922 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1925 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1927 LDX_PROBE_MEM_##SIZEOP: \
1928 bpf_probe_read_kernel(&DST, sizeof(SIZE), \
1929 (const void *)(long) (SRC + insn->off)); \
1930 DST = *((SIZE *)&DST); \
1939 #define ATOMIC_ALU_OP(BOP, KOP) \
1941 if (BPF_SIZE(insn->code) == BPF_W) \
1942 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
1943 (DST + insn->off)); \
1945 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
1946 (DST + insn->off)); \
1948 case BOP | BPF_FETCH: \
1949 if (BPF_SIZE(insn->code) == BPF_W) \
1950 SRC = (u32) atomic_fetch_##KOP( \
1952 (atomic_t *)(unsigned long) (DST + insn->off)); \
1954 SRC = (u64) atomic64_fetch_##KOP( \
1956 (atomic64_t *)(unsigned long) (DST + insn->off)); \
1962 ATOMIC_ALU_OP(BPF_ADD, add)
1963 ATOMIC_ALU_OP(BPF_AND, and)
1964 ATOMIC_ALU_OP(BPF_OR, or)
1965 ATOMIC_ALU_OP(BPF_XOR, xor)
1966 #undef ATOMIC_ALU_OP
1969 if (BPF_SIZE(insn->code) == BPF_W)
1970 SRC = (u32) atomic_xchg(
1971 (atomic_t *)(unsigned long) (DST + insn->off),
1974 SRC = (u64) atomic64_xchg(
1975 (atomic64_t *)(unsigned long) (DST + insn->off),
1979 if (BPF_SIZE(insn->code) == BPF_W)
1980 BPF_R0 = (u32) atomic_cmpxchg(
1981 (atomic_t *)(unsigned long) (DST + insn->off),
1982 (u32) BPF_R0, (u32) SRC);
1984 BPF_R0 = (u64) atomic64_cmpxchg(
1985 (atomic64_t *)(unsigned long) (DST + insn->off),
1986 (u64) BPF_R0, (u64) SRC);
1995 /* If we ever reach this, we have a bug somewhere. Die hard here
1996 * instead of just returning 0; we could be somewhere in a subprog,
1997 * so execution could continue otherwise which we do /not/ want.
1999 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
2001 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
2002 insn->code, insn->imm);
2007 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2008 #define DEFINE_BPF_PROG_RUN(stack_size) \
2009 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2011 u64 stack[stack_size / sizeof(u64)]; \
2012 u64 regs[MAX_BPF_EXT_REG] = {}; \
2014 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2015 ARG1 = (u64) (unsigned long) ctx; \
2016 return ___bpf_prog_run(regs, insn); \
2019 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2020 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2021 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2022 const struct bpf_insn *insn) \
2024 u64 stack[stack_size / sizeof(u64)]; \
2025 u64 regs[MAX_BPF_EXT_REG]; \
2027 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2033 return ___bpf_prog_run(regs, insn); \
2036 #define EVAL1(FN, X) FN(X)
2037 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2038 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2039 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2040 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2041 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2043 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2044 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2045 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2047 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2048 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2049 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2051 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2053 static unsigned int (*interpreters[])(const void *ctx,
2054 const struct bpf_insn *insn) = {
2055 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2056 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2057 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2059 #undef PROG_NAME_LIST
2060 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2061 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2062 const struct bpf_insn *insn) = {
2063 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2064 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2065 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2067 #undef PROG_NAME_LIST
2069 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2071 stack_depth = max_t(u32, stack_depth, 1);
2072 insn->off = (s16) insn->imm;
2073 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2074 __bpf_call_base_args;
2075 insn->code = BPF_JMP | BPF_CALL_ARGS;
2079 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2080 const struct bpf_insn *insn)
2082 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2083 * is not working properly, so warn about it!
2090 bool bpf_prog_map_compatible(struct bpf_map *map,
2091 const struct bpf_prog *fp)
2093 enum bpf_prog_type prog_type = resolve_prog_type(fp);
2096 if (fp->kprobe_override)
2099 spin_lock(&map->owner.lock);
2100 if (!map->owner.type) {
2101 /* There's no owner yet where we could check for
2104 map->owner.type = prog_type;
2105 map->owner.jited = fp->jited;
2106 map->owner.xdp_has_frags = fp->aux->xdp_has_frags;
2109 ret = map->owner.type == prog_type &&
2110 map->owner.jited == fp->jited &&
2111 map->owner.xdp_has_frags == fp->aux->xdp_has_frags;
2113 spin_unlock(&map->owner.lock);
2118 static int bpf_check_tail_call(const struct bpf_prog *fp)
2120 struct bpf_prog_aux *aux = fp->aux;
2123 mutex_lock(&aux->used_maps_mutex);
2124 for (i = 0; i < aux->used_map_cnt; i++) {
2125 struct bpf_map *map = aux->used_maps[i];
2127 if (!map_type_contains_progs(map))
2130 if (!bpf_prog_map_compatible(map, fp)) {
2137 mutex_unlock(&aux->used_maps_mutex);
2141 static void bpf_prog_select_func(struct bpf_prog *fp)
2143 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2144 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2146 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2148 fp->bpf_func = __bpf_prog_ret0_warn;
2153 * bpf_prog_select_runtime - select exec runtime for BPF program
2154 * @fp: bpf_prog populated with BPF program
2155 * @err: pointer to error variable
2157 * Try to JIT eBPF program, if JIT is not available, use interpreter.
2158 * The BPF program will be executed via bpf_prog_run() function.
2160 * Return: the &fp argument along with &err set to 0 for success or
2161 * a negative errno code on failure
2163 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2165 /* In case of BPF to BPF calls, verifier did all the prep
2166 * work with regards to JITing, etc.
2168 bool jit_needed = false;
2173 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2174 bpf_prog_has_kfunc_call(fp))
2177 bpf_prog_select_func(fp);
2179 /* eBPF JITs can rewrite the program in case constant
2180 * blinding is active. However, in case of error during
2181 * blinding, bpf_int_jit_compile() must always return a
2182 * valid program, which in this case would simply not
2183 * be JITed, but falls back to the interpreter.
2185 if (!bpf_prog_is_dev_bound(fp->aux)) {
2186 *err = bpf_prog_alloc_jited_linfo(fp);
2190 fp = bpf_int_jit_compile(fp);
2191 bpf_prog_jit_attempt_done(fp);
2192 if (!fp->jited && jit_needed) {
2197 *err = bpf_prog_offload_compile(fp);
2203 bpf_prog_lock_ro(fp);
2205 /* The tail call compatibility check can only be done at
2206 * this late stage as we need to determine, if we deal
2207 * with JITed or non JITed program concatenations and not
2208 * all eBPF JITs might immediately support all features.
2210 *err = bpf_check_tail_call(fp);
2214 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2216 static unsigned int __bpf_prog_ret1(const void *ctx,
2217 const struct bpf_insn *insn)
2222 static struct bpf_prog_dummy {
2223 struct bpf_prog prog;
2224 } dummy_bpf_prog = {
2226 .bpf_func = __bpf_prog_ret1,
2230 struct bpf_empty_prog_array bpf_empty_prog_array = {
2233 EXPORT_SYMBOL(bpf_empty_prog_array);
2235 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2238 return kzalloc(sizeof(struct bpf_prog_array) +
2239 sizeof(struct bpf_prog_array_item) *
2243 return &bpf_empty_prog_array.hdr;
2246 void bpf_prog_array_free(struct bpf_prog_array *progs)
2248 if (!progs || progs == &bpf_empty_prog_array.hdr)
2250 kfree_rcu(progs, rcu);
2253 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2255 struct bpf_prog_array *progs;
2257 /* If RCU Tasks Trace grace period implies RCU grace period, there is
2258 * no need to call kfree_rcu(), just call kfree() directly.
2260 progs = container_of(rcu, struct bpf_prog_array, rcu);
2261 if (rcu_trace_implies_rcu_gp())
2264 kfree_rcu(progs, rcu);
2267 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2269 if (!progs || progs == &bpf_empty_prog_array.hdr)
2271 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2274 int bpf_prog_array_length(struct bpf_prog_array *array)
2276 struct bpf_prog_array_item *item;
2279 for (item = array->items; item->prog; item++)
2280 if (item->prog != &dummy_bpf_prog.prog)
2285 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2287 struct bpf_prog_array_item *item;
2289 for (item = array->items; item->prog; item++)
2290 if (item->prog != &dummy_bpf_prog.prog)
2295 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2299 struct bpf_prog_array_item *item;
2302 for (item = array->items; item->prog; item++) {
2303 if (item->prog == &dummy_bpf_prog.prog)
2305 prog_ids[i] = item->prog->aux->id;
2306 if (++i == request_cnt) {
2312 return !!(item->prog);
2315 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2316 __u32 __user *prog_ids, u32 cnt)
2318 unsigned long err = 0;
2322 /* users of this function are doing:
2323 * cnt = bpf_prog_array_length();
2325 * bpf_prog_array_copy_to_user(..., cnt);
2326 * so below kcalloc doesn't need extra cnt > 0 check.
2328 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2331 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2332 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2341 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2342 struct bpf_prog *old_prog)
2344 struct bpf_prog_array_item *item;
2346 for (item = array->items; item->prog; item++)
2347 if (item->prog == old_prog) {
2348 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2354 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2355 * index into the program array with
2356 * a dummy no-op program.
2357 * @array: a bpf_prog_array
2358 * @index: the index of the program to replace
2360 * Skips over dummy programs, by not counting them, when calculating
2361 * the position of the program to replace.
2365 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2366 * * -ENOENT - Index out of range
2368 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2370 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2374 * bpf_prog_array_update_at() - Updates the program at the given index
2375 * into the program array.
2376 * @array: a bpf_prog_array
2377 * @index: the index of the program to update
2378 * @prog: the program to insert into the array
2380 * Skips over dummy programs, by not counting them, when calculating
2381 * the position of the program to update.
2385 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2386 * * -ENOENT - Index out of range
2388 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2389 struct bpf_prog *prog)
2391 struct bpf_prog_array_item *item;
2393 if (unlikely(index < 0))
2396 for (item = array->items; item->prog; item++) {
2397 if (item->prog == &dummy_bpf_prog.prog)
2400 WRITE_ONCE(item->prog, prog);
2408 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2409 struct bpf_prog *exclude_prog,
2410 struct bpf_prog *include_prog,
2412 struct bpf_prog_array **new_array)
2414 int new_prog_cnt, carry_prog_cnt = 0;
2415 struct bpf_prog_array_item *existing, *new;
2416 struct bpf_prog_array *array;
2417 bool found_exclude = false;
2419 /* Figure out how many existing progs we need to carry over to
2423 existing = old_array->items;
2424 for (; existing->prog; existing++) {
2425 if (existing->prog == exclude_prog) {
2426 found_exclude = true;
2429 if (existing->prog != &dummy_bpf_prog.prog)
2431 if (existing->prog == include_prog)
2436 if (exclude_prog && !found_exclude)
2439 /* How many progs (not NULL) will be in the new array? */
2440 new_prog_cnt = carry_prog_cnt;
2444 /* Do we have any prog (not NULL) in the new array? */
2445 if (!new_prog_cnt) {
2450 /* +1 as the end of prog_array is marked with NULL */
2451 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2456 /* Fill in the new prog array */
2457 if (carry_prog_cnt) {
2458 existing = old_array->items;
2459 for (; existing->prog; existing++) {
2460 if (existing->prog == exclude_prog ||
2461 existing->prog == &dummy_bpf_prog.prog)
2464 new->prog = existing->prog;
2465 new->bpf_cookie = existing->bpf_cookie;
2470 new->prog = include_prog;
2471 new->bpf_cookie = bpf_cookie;
2479 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2480 u32 *prog_ids, u32 request_cnt,
2486 cnt = bpf_prog_array_length(array);
2490 /* return early if user requested only program count or nothing to copy */
2491 if (!request_cnt || !cnt)
2494 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2495 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2499 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2500 struct bpf_map **used_maps, u32 len)
2502 struct bpf_map *map;
2505 for (i = 0; i < len; i++) {
2507 if (map->ops->map_poke_untrack)
2508 map->ops->map_poke_untrack(map, aux);
2513 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2515 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2516 kfree(aux->used_maps);
2519 void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2520 struct btf_mod_pair *used_btfs, u32 len)
2522 #ifdef CONFIG_BPF_SYSCALL
2523 struct btf_mod_pair *btf_mod;
2526 for (i = 0; i < len; i++) {
2527 btf_mod = &used_btfs[i];
2528 if (btf_mod->module)
2529 module_put(btf_mod->module);
2530 btf_put(btf_mod->btf);
2535 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2537 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2538 kfree(aux->used_btfs);
2541 static void bpf_prog_free_deferred(struct work_struct *work)
2543 struct bpf_prog_aux *aux;
2546 aux = container_of(work, struct bpf_prog_aux, work);
2547 #ifdef CONFIG_BPF_SYSCALL
2548 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2550 #ifdef CONFIG_CGROUP_BPF
2551 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2552 bpf_cgroup_atype_put(aux->cgroup_atype);
2554 bpf_free_used_maps(aux);
2555 bpf_free_used_btfs(aux);
2556 if (bpf_prog_is_dev_bound(aux))
2557 bpf_prog_offload_destroy(aux->prog);
2558 #ifdef CONFIG_PERF_EVENTS
2559 if (aux->prog->has_callchain_buf)
2560 put_callchain_buffers();
2562 if (aux->dst_trampoline)
2563 bpf_trampoline_put(aux->dst_trampoline);
2564 for (i = 0; i < aux->func_cnt; i++) {
2565 /* We can just unlink the subprog poke descriptor table as
2566 * it was originally linked to the main program and is also
2567 * released along with it.
2569 aux->func[i]->aux->poke_tab = NULL;
2570 bpf_jit_free(aux->func[i]);
2572 if (aux->func_cnt) {
2574 bpf_prog_unlock_free(aux->prog);
2576 bpf_jit_free(aux->prog);
2580 void bpf_prog_free(struct bpf_prog *fp)
2582 struct bpf_prog_aux *aux = fp->aux;
2585 bpf_prog_put(aux->dst_prog);
2586 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2587 schedule_work(&aux->work);
2589 EXPORT_SYMBOL_GPL(bpf_prog_free);
2591 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2592 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2594 void bpf_user_rnd_init_once(void)
2596 prandom_init_once(&bpf_user_rnd_state);
2599 BPF_CALL_0(bpf_user_rnd_u32)
2601 /* Should someone ever have the rather unwise idea to use some
2602 * of the registers passed into this function, then note that
2603 * this function is called from native eBPF and classic-to-eBPF
2604 * transformations. Register assignments from both sides are
2605 * different, f.e. classic always sets fn(ctx, A, X) here.
2607 struct rnd_state *state;
2610 state = &get_cpu_var(bpf_user_rnd_state);
2611 res = prandom_u32_state(state);
2612 put_cpu_var(bpf_user_rnd_state);
2617 BPF_CALL_0(bpf_get_raw_cpu_id)
2619 return raw_smp_processor_id();
2622 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2623 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2624 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2625 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2626 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2627 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2628 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2629 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2630 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2631 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2632 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2634 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2635 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2636 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2637 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2638 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2639 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2640 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak;
2642 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2643 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2644 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2645 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2646 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2647 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2648 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2649 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2650 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2651 const struct bpf_func_proto bpf_set_retval_proto __weak;
2652 const struct bpf_func_proto bpf_get_retval_proto __weak;
2654 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2659 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2665 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2666 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2670 EXPORT_SYMBOL_GPL(bpf_event_output);
2672 /* Always built-in helper functions. */
2673 const struct bpf_func_proto bpf_tail_call_proto = {
2676 .ret_type = RET_VOID,
2677 .arg1_type = ARG_PTR_TO_CTX,
2678 .arg2_type = ARG_CONST_MAP_PTR,
2679 .arg3_type = ARG_ANYTHING,
2682 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2683 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2684 * eBPF and implicitly also cBPF can get JITed!
2686 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2691 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2692 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2694 void __weak bpf_jit_compile(struct bpf_prog *prog)
2698 bool __weak bpf_helper_changes_pkt_data(void *func)
2703 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2704 * analysis code and wants explicit zero extension inserted by verifier.
2705 * Otherwise, return FALSE.
2707 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2708 * you don't override this. JITs that don't want these extra insns can detect
2709 * them using insn_is_zext.
2711 bool __weak bpf_jit_needs_zext(void)
2716 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2717 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2722 bool __weak bpf_jit_supports_kfunc_call(void)
2727 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2728 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2730 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2736 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2737 void *addr1, void *addr2)
2742 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
2744 return ERR_PTR(-ENOTSUPP);
2747 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
2752 #ifdef CONFIG_BPF_SYSCALL
2753 static int __init bpf_global_ma_init(void)
2757 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false);
2758 bpf_global_ma_set = !ret;
2761 late_initcall(bpf_global_ma_init);
2764 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2765 EXPORT_SYMBOL(bpf_stats_enabled_key);
2767 /* All definitions of tracepoints related to BPF. */
2768 #define CREATE_TRACE_POINTS
2769 #include <linux/bpf_trace.h>
2771 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2772 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);