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/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>
38 #include <asm/barrier.h>
39 #include <asm/unaligned.h>
42 #define BPF_R0 regs[BPF_REG_0]
43 #define BPF_R1 regs[BPF_REG_1]
44 #define BPF_R2 regs[BPF_REG_2]
45 #define BPF_R3 regs[BPF_REG_3]
46 #define BPF_R4 regs[BPF_REG_4]
47 #define BPF_R5 regs[BPF_REG_5]
48 #define BPF_R6 regs[BPF_REG_6]
49 #define BPF_R7 regs[BPF_REG_7]
50 #define BPF_R8 regs[BPF_REG_8]
51 #define BPF_R9 regs[BPF_REG_9]
52 #define BPF_R10 regs[BPF_REG_10]
55 #define DST regs[insn->dst_reg]
56 #define SRC regs[insn->src_reg]
57 #define FP regs[BPF_REG_FP]
58 #define AX regs[BPF_REG_AX]
59 #define ARG1 regs[BPF_REG_ARG1]
60 #define CTX regs[BPF_REG_CTX]
63 /* No hurry in this branch
65 * Exported for the bpf jit load helper.
67 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
71 if (k >= SKF_NET_OFF) {
72 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
73 } else if (k >= SKF_LL_OFF) {
74 if (unlikely(!skb_mac_header_was_set(skb)))
76 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
78 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
84 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
86 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
87 struct bpf_prog_aux *aux;
90 size = round_up(size, PAGE_SIZE);
91 fp = __vmalloc(size, gfp_flags);
95 aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT | gfp_extra_flags);
100 fp->active = alloc_percpu_gfp(int, GFP_KERNEL_ACCOUNT | gfp_extra_flags);
107 fp->pages = size / PAGE_SIZE;
110 fp->jit_requested = ebpf_jit_enabled();
111 fp->blinding_requested = bpf_jit_blinding_enabled(fp);
112 #ifdef CONFIG_CGROUP_BPF
113 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
116 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
117 mutex_init(&fp->aux->used_maps_mutex);
118 mutex_init(&fp->aux->dst_mutex);
123 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
125 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
126 struct bpf_prog *prog;
129 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
133 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
135 free_percpu(prog->active);
141 for_each_possible_cpu(cpu) {
142 struct bpf_prog_stats *pstats;
144 pstats = per_cpu_ptr(prog->stats, cpu);
145 u64_stats_init(&pstats->syncp);
149 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
151 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
153 if (!prog->aux->nr_linfo || !prog->jit_requested)
156 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
157 sizeof(*prog->aux->jited_linfo),
158 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
159 if (!prog->aux->jited_linfo)
165 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
167 if (prog->aux->jited_linfo &&
168 (!prog->jited || !prog->aux->jited_linfo[0])) {
169 kvfree(prog->aux->jited_linfo);
170 prog->aux->jited_linfo = NULL;
173 kfree(prog->aux->kfunc_tab);
174 prog->aux->kfunc_tab = NULL;
177 /* The jit engine is responsible to provide an array
178 * for insn_off to the jited_off mapping (insn_to_jit_off).
180 * The idx to this array is the insn_off. Hence, the insn_off
181 * here is relative to the prog itself instead of the main prog.
182 * This array has one entry for each xlated bpf insn.
184 * jited_off is the byte off to the end of the jited insn.
188 * The first bpf insn off of the prog. The insn off
189 * here is relative to the main prog.
190 * e.g. if prog is a subprog, insn_start > 0
192 * The prog's idx to prog->aux->linfo and jited_linfo
194 * jited_linfo[linfo_idx] = prog->bpf_func
198 * jited_linfo[i] = prog->bpf_func +
199 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
201 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
202 const u32 *insn_to_jit_off)
204 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
205 const struct bpf_line_info *linfo;
208 if (!prog->aux->jited_linfo)
209 /* Userspace did not provide linfo */
212 linfo_idx = prog->aux->linfo_idx;
213 linfo = &prog->aux->linfo[linfo_idx];
214 insn_start = linfo[0].insn_off;
215 insn_end = insn_start + prog->len;
217 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
218 jited_linfo[0] = prog->bpf_func;
220 nr_linfo = prog->aux->nr_linfo - linfo_idx;
222 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
223 /* The verifier ensures that linfo[i].insn_off is
224 * strictly increasing
226 jited_linfo[i] = prog->bpf_func +
227 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
230 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
231 gfp_t gfp_extra_flags)
233 gfp_t gfp_flags = GFP_KERNEL_ACCOUNT | __GFP_ZERO | gfp_extra_flags;
237 size = round_up(size, PAGE_SIZE);
238 pages = size / PAGE_SIZE;
239 if (pages <= fp_old->pages)
242 fp = __vmalloc(size, gfp_flags);
244 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
248 /* We keep fp->aux from fp_old around in the new
249 * reallocated structure.
252 fp_old->stats = NULL;
253 fp_old->active = NULL;
254 __bpf_prog_free(fp_old);
260 void __bpf_prog_free(struct bpf_prog *fp)
263 mutex_destroy(&fp->aux->used_maps_mutex);
264 mutex_destroy(&fp->aux->dst_mutex);
265 kfree(fp->aux->poke_tab);
268 free_percpu(fp->stats);
269 free_percpu(fp->active);
273 int bpf_prog_calc_tag(struct bpf_prog *fp)
275 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
276 u32 raw_size = bpf_prog_tag_scratch_size(fp);
277 u32 digest[SHA1_DIGEST_WORDS];
278 u32 ws[SHA1_WORKSPACE_WORDS];
279 u32 i, bsize, psize, blocks;
280 struct bpf_insn *dst;
286 raw = vmalloc(raw_size);
291 memset(ws, 0, sizeof(ws));
293 /* We need to take out the map fd for the digest calculation
294 * since they are unstable from user space side.
297 for (i = 0, was_ld_map = false; i < fp->len; i++) {
298 dst[i] = fp->insnsi[i];
300 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
301 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
302 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
305 } else if (was_ld_map &&
307 dst[i].dst_reg == 0 &&
308 dst[i].src_reg == 0 &&
317 psize = bpf_prog_insn_size(fp);
318 memset(&raw[psize], 0, raw_size - psize);
321 bsize = round_up(psize, SHA1_BLOCK_SIZE);
322 blocks = bsize / SHA1_BLOCK_SIZE;
324 if (bsize - psize >= sizeof(__be64)) {
325 bits = (__be64 *)(todo + bsize - sizeof(__be64));
327 bits = (__be64 *)(todo + bsize + bits_offset);
330 *bits = cpu_to_be64((psize - 1) << 3);
333 sha1_transform(digest, todo, ws);
334 todo += SHA1_BLOCK_SIZE;
337 result = (__force __be32 *)digest;
338 for (i = 0; i < SHA1_DIGEST_WORDS; i++)
339 result[i] = cpu_to_be32(digest[i]);
340 memcpy(fp->tag, result, sizeof(fp->tag));
346 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
347 s32 end_new, s32 curr, const bool probe_pass)
349 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
350 s32 delta = end_new - end_old;
353 if (curr < pos && curr + imm + 1 >= end_old)
355 else if (curr >= end_new && curr + imm + 1 < end_new)
357 if (imm < imm_min || imm > imm_max)
364 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
365 s32 end_new, s32 curr, const bool probe_pass)
367 const s32 off_min = S16_MIN, off_max = S16_MAX;
368 s32 delta = end_new - end_old;
371 if (curr < pos && curr + off + 1 >= end_old)
373 else if (curr >= end_new && curr + off + 1 < end_new)
375 if (off < off_min || off > off_max)
382 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
383 s32 end_new, const bool probe_pass)
385 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
386 struct bpf_insn *insn = prog->insnsi;
389 for (i = 0; i < insn_cnt; i++, insn++) {
392 /* In the probing pass we still operate on the original,
393 * unpatched image in order to check overflows before we
394 * do any other adjustments. Therefore skip the patchlet.
396 if (probe_pass && i == pos) {
398 insn = prog->insnsi + end_old;
400 if (bpf_pseudo_func(insn)) {
401 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
402 end_new, i, probe_pass);
408 if ((BPF_CLASS(code) != BPF_JMP &&
409 BPF_CLASS(code) != BPF_JMP32) ||
410 BPF_OP(code) == BPF_EXIT)
412 /* Adjust offset of jmps if we cross patch boundaries. */
413 if (BPF_OP(code) == BPF_CALL) {
414 if (insn->src_reg != BPF_PSEUDO_CALL)
416 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
417 end_new, i, probe_pass);
419 ret = bpf_adj_delta_to_off(insn, pos, end_old,
420 end_new, i, probe_pass);
429 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
431 struct bpf_line_info *linfo;
434 nr_linfo = prog->aux->nr_linfo;
435 if (!nr_linfo || !delta)
438 linfo = prog->aux->linfo;
440 for (i = 0; i < nr_linfo; i++)
441 if (off < linfo[i].insn_off)
444 /* Push all off < linfo[i].insn_off by delta */
445 for (; i < nr_linfo; i++)
446 linfo[i].insn_off += delta;
449 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
450 const struct bpf_insn *patch, u32 len)
452 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
453 const u32 cnt_max = S16_MAX;
454 struct bpf_prog *prog_adj;
457 /* Since our patchlet doesn't expand the image, we're done. */
458 if (insn_delta == 0) {
459 memcpy(prog->insnsi + off, patch, sizeof(*patch));
463 insn_adj_cnt = prog->len + insn_delta;
465 /* Reject anything that would potentially let the insn->off
466 * target overflow when we have excessive program expansions.
467 * We need to probe here before we do any reallocation where
468 * we afterwards may not fail anymore.
470 if (insn_adj_cnt > cnt_max &&
471 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
474 /* Several new instructions need to be inserted. Make room
475 * for them. Likely, there's no need for a new allocation as
476 * last page could have large enough tailroom.
478 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
481 return ERR_PTR(-ENOMEM);
483 prog_adj->len = insn_adj_cnt;
485 /* Patching happens in 3 steps:
487 * 1) Move over tail of insnsi from next instruction onwards,
488 * so we can patch the single target insn with one or more
489 * new ones (patching is always from 1 to n insns, n > 0).
490 * 2) Inject new instructions at the target location.
491 * 3) Adjust branch offsets if necessary.
493 insn_rest = insn_adj_cnt - off - len;
495 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
496 sizeof(*patch) * insn_rest);
497 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
499 /* We are guaranteed to not fail at this point, otherwise
500 * the ship has sailed to reverse to the original state. An
501 * overflow cannot happen at this point.
503 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
505 bpf_adj_linfo(prog_adj, off, insn_delta);
510 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
512 /* Branch offsets can't overflow when program is shrinking, no need
513 * to call bpf_adj_branches(..., true) here
515 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
516 sizeof(struct bpf_insn) * (prog->len - off - cnt));
519 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
522 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
526 for (i = 0; i < fp->aux->func_cnt; i++)
527 bpf_prog_kallsyms_del(fp->aux->func[i]);
530 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
532 bpf_prog_kallsyms_del_subprogs(fp);
533 bpf_prog_kallsyms_del(fp);
536 #ifdef CONFIG_BPF_JIT
537 /* All BPF JIT sysctl knobs here. */
538 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
539 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
540 int bpf_jit_harden __read_mostly;
541 long bpf_jit_limit __read_mostly;
542 long bpf_jit_limit_max __read_mostly;
545 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
547 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
549 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
550 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len;
554 bpf_prog_ksym_set_name(struct bpf_prog *prog)
556 char *sym = prog->aux->ksym.name;
557 const char *end = sym + KSYM_NAME_LEN;
558 const struct btf_type *type;
559 const char *func_name;
561 BUILD_BUG_ON(sizeof("bpf_prog_") +
562 sizeof(prog->tag) * 2 +
563 /* name has been null terminated.
564 * We should need +1 for the '_' preceding
565 * the name. However, the null character
566 * is double counted between the name and the
567 * sizeof("bpf_prog_") above, so we omit
570 sizeof(prog->aux->name) > KSYM_NAME_LEN);
572 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
573 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
575 /* prog->aux->name will be ignored if full btf name is available */
576 if (prog->aux->func_info_cnt) {
577 type = btf_type_by_id(prog->aux->btf,
578 prog->aux->func_info[prog->aux->func_idx].type_id);
579 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
580 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
584 if (prog->aux->name[0])
585 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
590 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
592 return container_of(n, struct bpf_ksym, tnode)->start;
595 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
596 struct latch_tree_node *b)
598 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
601 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
603 unsigned long val = (unsigned long)key;
604 const struct bpf_ksym *ksym;
606 ksym = container_of(n, struct bpf_ksym, tnode);
608 if (val < ksym->start)
610 if (val >= ksym->end)
616 static const struct latch_tree_ops bpf_tree_ops = {
617 .less = bpf_tree_less,
618 .comp = bpf_tree_comp,
621 static DEFINE_SPINLOCK(bpf_lock);
622 static LIST_HEAD(bpf_kallsyms);
623 static struct latch_tree_root bpf_tree __cacheline_aligned;
625 void bpf_ksym_add(struct bpf_ksym *ksym)
627 spin_lock_bh(&bpf_lock);
628 WARN_ON_ONCE(!list_empty(&ksym->lnode));
629 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
630 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
631 spin_unlock_bh(&bpf_lock);
634 static void __bpf_ksym_del(struct bpf_ksym *ksym)
636 if (list_empty(&ksym->lnode))
639 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
640 list_del_rcu(&ksym->lnode);
643 void bpf_ksym_del(struct bpf_ksym *ksym)
645 spin_lock_bh(&bpf_lock);
646 __bpf_ksym_del(ksym);
647 spin_unlock_bh(&bpf_lock);
650 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
652 return fp->jited && !bpf_prog_was_classic(fp);
655 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
657 return list_empty(&fp->aux->ksym.lnode) ||
658 fp->aux->ksym.lnode.prev == LIST_POISON2;
661 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
663 if (!bpf_prog_kallsyms_candidate(fp) ||
667 bpf_prog_ksym_set_addr(fp);
668 bpf_prog_ksym_set_name(fp);
669 fp->aux->ksym.prog = true;
671 bpf_ksym_add(&fp->aux->ksym);
674 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
676 if (!bpf_prog_kallsyms_candidate(fp))
679 bpf_ksym_del(&fp->aux->ksym);
682 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
684 struct latch_tree_node *n;
686 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
687 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
690 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
691 unsigned long *off, char *sym)
693 struct bpf_ksym *ksym;
697 ksym = bpf_ksym_find(addr);
699 unsigned long symbol_start = ksym->start;
700 unsigned long symbol_end = ksym->end;
702 strncpy(sym, ksym->name, KSYM_NAME_LEN);
706 *size = symbol_end - symbol_start;
708 *off = addr - symbol_start;
715 bool is_bpf_text_address(unsigned long addr)
720 ret = bpf_ksym_find(addr) != NULL;
726 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
728 struct bpf_ksym *ksym = bpf_ksym_find(addr);
730 return ksym && ksym->prog ?
731 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
735 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
737 const struct exception_table_entry *e = NULL;
738 struct bpf_prog *prog;
741 prog = bpf_prog_ksym_find(addr);
744 if (!prog->aux->num_exentries)
747 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
753 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
756 struct bpf_ksym *ksym;
760 if (!bpf_jit_kallsyms_enabled())
764 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
768 strncpy(sym, ksym->name, KSYM_NAME_LEN);
770 *value = ksym->start;
771 *type = BPF_SYM_ELF_TYPE;
781 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
782 struct bpf_jit_poke_descriptor *poke)
784 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
785 static const u32 poke_tab_max = 1024;
786 u32 slot = prog->aux->size_poke_tab;
789 if (size > poke_tab_max)
791 if (poke->tailcall_target || poke->tailcall_target_stable ||
792 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
795 switch (poke->reason) {
796 case BPF_POKE_REASON_TAIL_CALL:
797 if (!poke->tail_call.map)
804 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
808 memcpy(&tab[slot], poke, sizeof(*poke));
809 prog->aux->size_poke_tab = size;
810 prog->aux->poke_tab = tab;
816 * BPF program pack allocator.
818 * Most BPF programs are pretty small. Allocating a hole page for each
819 * program is sometime a waste. Many small bpf program also adds pressure
820 * to instruction TLB. To solve this issue, we introduce a BPF program pack
821 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
822 * to host BPF programs.
824 #define BPF_PROG_CHUNK_SHIFT 6
825 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT)
826 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1))
828 struct bpf_prog_pack {
829 struct list_head list;
831 unsigned long bitmap[];
834 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
836 static size_t bpf_prog_pack_size = -1;
837 static size_t bpf_prog_pack_mask = -1;
839 static int bpf_prog_chunk_count(void)
841 WARN_ON_ONCE(bpf_prog_pack_size == -1);
842 return bpf_prog_pack_size / BPF_PROG_CHUNK_SIZE;
845 static DEFINE_MUTEX(pack_mutex);
846 static LIST_HEAD(pack_list);
848 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
849 * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
852 #define BPF_HPAGE_SIZE PMD_SIZE
853 #define BPF_HPAGE_MASK PMD_MASK
855 #define BPF_HPAGE_SIZE PAGE_SIZE
856 #define BPF_HPAGE_MASK PAGE_MASK
859 static size_t select_bpf_prog_pack_size(void)
864 size = BPF_HPAGE_SIZE * num_online_nodes();
865 ptr = module_alloc(size);
867 /* Test whether we can get huge pages. If not just use PAGE_SIZE
870 if (!ptr || !is_vm_area_hugepages(ptr)) {
872 bpf_prog_pack_mask = PAGE_MASK;
874 bpf_prog_pack_mask = BPF_HPAGE_MASK;
881 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
883 struct bpf_prog_pack *pack;
885 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(bpf_prog_chunk_count())),
889 pack->ptr = module_alloc(bpf_prog_pack_size);
894 bpf_fill_ill_insns(pack->ptr, bpf_prog_pack_size);
895 bitmap_zero(pack->bitmap, bpf_prog_pack_size / BPF_PROG_CHUNK_SIZE);
896 list_add_tail(&pack->list, &pack_list);
898 set_vm_flush_reset_perms(pack->ptr);
899 set_memory_ro((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE);
900 set_memory_x((unsigned long)pack->ptr, bpf_prog_pack_size / PAGE_SIZE);
904 static void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
906 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
907 struct bpf_prog_pack *pack;
911 mutex_lock(&pack_mutex);
912 if (bpf_prog_pack_size == -1)
913 bpf_prog_pack_size = select_bpf_prog_pack_size();
915 if (size > bpf_prog_pack_size) {
916 size = round_up(size, PAGE_SIZE);
917 ptr = module_alloc(size);
919 bpf_fill_ill_insns(ptr, size);
920 set_vm_flush_reset_perms(ptr);
921 set_memory_ro((unsigned long)ptr, size / PAGE_SIZE);
922 set_memory_x((unsigned long)ptr, size / PAGE_SIZE);
926 list_for_each_entry(pack, &pack_list, list) {
927 pos = bitmap_find_next_zero_area(pack->bitmap, bpf_prog_chunk_count(), 0,
929 if (pos < bpf_prog_chunk_count())
930 goto found_free_area;
933 pack = alloc_new_pack(bpf_fill_ill_insns);
940 bitmap_set(pack->bitmap, pos, nbits);
941 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
944 mutex_unlock(&pack_mutex);
948 static void bpf_prog_pack_free(struct bpf_binary_header *hdr)
950 struct bpf_prog_pack *pack = NULL, *tmp;
955 mutex_lock(&pack_mutex);
956 if (hdr->size > bpf_prog_pack_size) {
961 pack_ptr = (void *)((unsigned long)hdr & bpf_prog_pack_mask);
963 list_for_each_entry(tmp, &pack_list, list) {
964 if (tmp->ptr == pack_ptr) {
970 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
973 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size);
974 pos = ((unsigned long)hdr - (unsigned long)pack_ptr) >> BPF_PROG_CHUNK_SHIFT;
976 WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size),
977 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
979 bitmap_clear(pack->bitmap, pos, nbits);
980 if (bitmap_find_next_zero_area(pack->bitmap, bpf_prog_chunk_count(), 0,
981 bpf_prog_chunk_count(), 0) == 0) {
982 list_del(&pack->list);
983 module_memfree(pack->ptr);
987 mutex_unlock(&pack_mutex);
990 static atomic_long_t bpf_jit_current;
992 /* Can be overridden by an arch's JIT compiler if it has a custom,
993 * dedicated BPF backend memory area, or if neither of the two
996 u64 __weak bpf_jit_alloc_exec_limit(void)
998 #if defined(MODULES_VADDR)
999 return MODULES_END - MODULES_VADDR;
1001 return VMALLOC_END - VMALLOC_START;
1005 static int __init bpf_jit_charge_init(void)
1007 /* Only used as heuristic here to derive limit. */
1008 bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
1009 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2,
1010 PAGE_SIZE), LONG_MAX);
1013 pure_initcall(bpf_jit_charge_init);
1015 int bpf_jit_charge_modmem(u32 size)
1017 if (atomic_long_add_return(size, &bpf_jit_current) > bpf_jit_limit) {
1018 if (!bpf_capable()) {
1019 atomic_long_sub(size, &bpf_jit_current);
1027 void bpf_jit_uncharge_modmem(u32 size)
1029 atomic_long_sub(size, &bpf_jit_current);
1032 void *__weak bpf_jit_alloc_exec(unsigned long size)
1034 return module_alloc(size);
1037 void __weak bpf_jit_free_exec(void *addr)
1039 module_memfree(addr);
1042 struct bpf_binary_header *
1043 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1044 unsigned int alignment,
1045 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1047 struct bpf_binary_header *hdr;
1048 u32 size, hole, start;
1050 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1051 alignment > BPF_IMAGE_ALIGNMENT);
1053 /* Most of BPF filters are really small, but if some of them
1054 * fill a page, allow at least 128 extra bytes to insert a
1055 * random section of illegal instructions.
1057 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1059 if (bpf_jit_charge_modmem(size))
1061 hdr = bpf_jit_alloc_exec(size);
1063 bpf_jit_uncharge_modmem(size);
1067 /* Fill space with illegal/arch-dep instructions. */
1068 bpf_fill_ill_insns(hdr, size);
1071 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1072 PAGE_SIZE - sizeof(*hdr));
1073 start = (get_random_int() % hole) & ~(alignment - 1);
1075 /* Leave a random number of instructions before BPF code. */
1076 *image_ptr = &hdr->image[start];
1081 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1083 u32 size = hdr->size;
1085 bpf_jit_free_exec(hdr);
1086 bpf_jit_uncharge_modmem(size);
1089 /* Allocate jit binary from bpf_prog_pack allocator.
1090 * Since the allocated memory is RO+X, the JIT engine cannot write directly
1091 * to the memory. To solve this problem, a RW buffer is also allocated at
1092 * as the same time. The JIT engine should calculate offsets based on the
1093 * RO memory address, but write JITed program to the RW buffer. Once the
1094 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1095 * the JITed program to the RO memory.
1097 struct bpf_binary_header *
1098 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1099 unsigned int alignment,
1100 struct bpf_binary_header **rw_header,
1102 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1104 struct bpf_binary_header *ro_header;
1105 u32 size, hole, start;
1107 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1108 alignment > BPF_IMAGE_ALIGNMENT);
1110 /* add 16 bytes for a random section of illegal instructions */
1111 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1113 if (bpf_jit_charge_modmem(size))
1115 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1117 bpf_jit_uncharge_modmem(size);
1121 *rw_header = kvmalloc(size, GFP_KERNEL);
1123 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size));
1124 bpf_prog_pack_free(ro_header);
1125 bpf_jit_uncharge_modmem(size);
1129 /* Fill space with illegal/arch-dep instructions. */
1130 bpf_fill_ill_insns(*rw_header, size);
1131 (*rw_header)->size = size;
1133 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1134 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1135 start = (get_random_int() % hole) & ~(alignment - 1);
1137 *image_ptr = &ro_header->image[start];
1138 *rw_image = &(*rw_header)->image[start];
1143 /* Copy JITed text from rw_header to its final location, the ro_header. */
1144 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog,
1145 struct bpf_binary_header *ro_header,
1146 struct bpf_binary_header *rw_header)
1150 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1155 bpf_prog_pack_free(ro_header);
1156 return PTR_ERR(ptr);
1158 prog->aux->use_bpf_prog_pack = true;
1162 /* bpf_jit_binary_pack_free is called in two different scenarios:
1163 * 1) when the program is freed after;
1164 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1165 * For case 2), we need to free both the RO memory and the RW buffer.
1167 * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1168 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1169 * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1170 * bpf_arch_text_copy (when jit fails).
1172 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1173 struct bpf_binary_header *rw_header)
1175 u32 size = ro_header->size;
1177 bpf_prog_pack_free(ro_header);
1179 bpf_jit_uncharge_modmem(size);
1182 static inline struct bpf_binary_header *
1183 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1185 unsigned long real_start = (unsigned long)fp->bpf_func;
1188 if (fp->aux->use_bpf_prog_pack)
1189 addr = real_start & BPF_PROG_CHUNK_MASK;
1191 addr = real_start & PAGE_MASK;
1193 return (void *)addr;
1196 /* This symbol is only overridden by archs that have different
1197 * requirements than the usual eBPF JITs, f.e. when they only
1198 * implement cBPF JIT, do not set images read-only, etc.
1200 void __weak bpf_jit_free(struct bpf_prog *fp)
1203 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1205 if (fp->aux->use_bpf_prog_pack)
1206 bpf_jit_binary_pack_free(hdr, NULL /* rw_buffer */);
1208 bpf_jit_binary_free(hdr);
1210 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1213 bpf_prog_unlock_free(fp);
1216 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1217 const struct bpf_insn *insn, bool extra_pass,
1218 u64 *func_addr, bool *func_addr_fixed)
1220 s16 off = insn->off;
1221 s32 imm = insn->imm;
1224 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1225 if (!*func_addr_fixed) {
1226 /* Place-holder address till the last pass has collected
1227 * all addresses for JITed subprograms in which case we
1228 * can pick them up from prog->aux.
1232 else if (prog->aux->func &&
1233 off >= 0 && off < prog->aux->func_cnt)
1234 addr = (u8 *)prog->aux->func[off]->bpf_func;
1238 /* Address of a BPF helper call. Since part of the core
1239 * kernel, it's always at a fixed location. __bpf_call_base
1240 * and the helper with imm relative to it are both in core
1243 addr = (u8 *)__bpf_call_base + imm;
1246 *func_addr = (unsigned long)addr;
1250 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1251 const struct bpf_insn *aux,
1252 struct bpf_insn *to_buff,
1255 struct bpf_insn *to = to_buff;
1256 u32 imm_rnd = get_random_int();
1259 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
1260 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1262 /* Constraints on AX register:
1264 * AX register is inaccessible from user space. It is mapped in
1265 * all JITs, and used here for constant blinding rewrites. It is
1266 * typically "stateless" meaning its contents are only valid within
1267 * the executed instruction, but not across several instructions.
1268 * There are a few exceptions however which are further detailed
1271 * Constant blinding is only used by JITs, not in the interpreter.
1272 * The interpreter uses AX in some occasions as a local temporary
1273 * register e.g. in DIV or MOD instructions.
1275 * In restricted circumstances, the verifier can also use the AX
1276 * register for rewrites as long as they do not interfere with
1279 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1282 if (from->imm == 0 &&
1283 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
1284 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1285 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1289 switch (from->code) {
1290 case BPF_ALU | BPF_ADD | BPF_K:
1291 case BPF_ALU | BPF_SUB | BPF_K:
1292 case BPF_ALU | BPF_AND | BPF_K:
1293 case BPF_ALU | BPF_OR | BPF_K:
1294 case BPF_ALU | BPF_XOR | BPF_K:
1295 case BPF_ALU | BPF_MUL | BPF_K:
1296 case BPF_ALU | BPF_MOV | BPF_K:
1297 case BPF_ALU | BPF_DIV | BPF_K:
1298 case BPF_ALU | BPF_MOD | BPF_K:
1299 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1300 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1301 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1304 case BPF_ALU64 | BPF_ADD | BPF_K:
1305 case BPF_ALU64 | BPF_SUB | BPF_K:
1306 case BPF_ALU64 | BPF_AND | BPF_K:
1307 case BPF_ALU64 | BPF_OR | BPF_K:
1308 case BPF_ALU64 | BPF_XOR | BPF_K:
1309 case BPF_ALU64 | BPF_MUL | BPF_K:
1310 case BPF_ALU64 | BPF_MOV | BPF_K:
1311 case BPF_ALU64 | BPF_DIV | BPF_K:
1312 case BPF_ALU64 | BPF_MOD | BPF_K:
1313 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1314 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1315 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1318 case BPF_JMP | BPF_JEQ | BPF_K:
1319 case BPF_JMP | BPF_JNE | BPF_K:
1320 case BPF_JMP | BPF_JGT | BPF_K:
1321 case BPF_JMP | BPF_JLT | BPF_K:
1322 case BPF_JMP | BPF_JGE | BPF_K:
1323 case BPF_JMP | BPF_JLE | BPF_K:
1324 case BPF_JMP | BPF_JSGT | BPF_K:
1325 case BPF_JMP | BPF_JSLT | BPF_K:
1326 case BPF_JMP | BPF_JSGE | BPF_K:
1327 case BPF_JMP | BPF_JSLE | BPF_K:
1328 case BPF_JMP | BPF_JSET | BPF_K:
1329 /* Accommodate for extra offset in case of a backjump. */
1333 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1334 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1335 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1338 case BPF_JMP32 | BPF_JEQ | BPF_K:
1339 case BPF_JMP32 | BPF_JNE | BPF_K:
1340 case BPF_JMP32 | BPF_JGT | BPF_K:
1341 case BPF_JMP32 | BPF_JLT | BPF_K:
1342 case BPF_JMP32 | BPF_JGE | BPF_K:
1343 case BPF_JMP32 | BPF_JLE | BPF_K:
1344 case BPF_JMP32 | BPF_JSGT | BPF_K:
1345 case BPF_JMP32 | BPF_JSLT | BPF_K:
1346 case BPF_JMP32 | BPF_JSGE | BPF_K:
1347 case BPF_JMP32 | BPF_JSLE | BPF_K:
1348 case BPF_JMP32 | BPF_JSET | BPF_K:
1349 /* Accommodate for extra offset in case of a backjump. */
1353 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1354 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1355 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1359 case BPF_LD | BPF_IMM | BPF_DW:
1360 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1361 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1362 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1363 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1365 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1366 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1367 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1369 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1370 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1373 case BPF_ST | BPF_MEM | BPF_DW:
1374 case BPF_ST | BPF_MEM | BPF_W:
1375 case BPF_ST | BPF_MEM | BPF_H:
1376 case BPF_ST | BPF_MEM | BPF_B:
1377 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1378 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1379 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1383 return to - to_buff;
1386 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1387 gfp_t gfp_extra_flags)
1389 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1390 struct bpf_prog *fp;
1392 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1394 /* aux->prog still points to the fp_other one, so
1395 * when promoting the clone to the real program,
1396 * this still needs to be adapted.
1398 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1404 static void bpf_prog_clone_free(struct bpf_prog *fp)
1406 /* aux was stolen by the other clone, so we cannot free
1407 * it from this path! It will be freed eventually by the
1408 * other program on release.
1410 * At this point, we don't need a deferred release since
1411 * clone is guaranteed to not be locked.
1416 __bpf_prog_free(fp);
1419 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1421 /* We have to repoint aux->prog to self, as we don't
1422 * know whether fp here is the clone or the original.
1425 bpf_prog_clone_free(fp_other);
1428 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1430 struct bpf_insn insn_buff[16], aux[2];
1431 struct bpf_prog *clone, *tmp;
1432 int insn_delta, insn_cnt;
1433 struct bpf_insn *insn;
1436 if (!prog->blinding_requested || prog->blinded)
1439 clone = bpf_prog_clone_create(prog, GFP_USER);
1441 return ERR_PTR(-ENOMEM);
1443 insn_cnt = clone->len;
1444 insn = clone->insnsi;
1446 for (i = 0; i < insn_cnt; i++, insn++) {
1447 if (bpf_pseudo_func(insn)) {
1448 /* ld_imm64 with an address of bpf subprog is not
1449 * a user controlled constant. Don't randomize it,
1450 * since it will conflict with jit_subprogs() logic.
1457 /* We temporarily need to hold the original ld64 insn
1458 * so that we can still access the first part in the
1459 * second blinding run.
1461 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1463 memcpy(aux, insn, sizeof(aux));
1465 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1466 clone->aux->verifier_zext);
1470 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1472 /* Patching may have repointed aux->prog during
1473 * realloc from the original one, so we need to
1474 * fix it up here on error.
1476 bpf_jit_prog_release_other(prog, clone);
1481 insn_delta = rewritten - 1;
1483 /* Walk new program and skip insns we just inserted. */
1484 insn = clone->insnsi + i + insn_delta;
1485 insn_cnt += insn_delta;
1492 #endif /* CONFIG_BPF_JIT */
1494 /* Base function for offset calculation. Needs to go into .text section,
1495 * therefore keeping it non-static as well; will also be used by JITs
1496 * anyway later on, so do not let the compiler omit it. This also needs
1497 * to go into kallsyms for correlation from e.g. bpftool, so naming
1500 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1504 EXPORT_SYMBOL_GPL(__bpf_call_base);
1506 /* All UAPI available opcodes. */
1507 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1508 /* 32 bit ALU operations. */ \
1509 /* Register based. */ \
1510 INSN_3(ALU, ADD, X), \
1511 INSN_3(ALU, SUB, X), \
1512 INSN_3(ALU, AND, X), \
1513 INSN_3(ALU, OR, X), \
1514 INSN_3(ALU, LSH, X), \
1515 INSN_3(ALU, RSH, X), \
1516 INSN_3(ALU, XOR, X), \
1517 INSN_3(ALU, MUL, X), \
1518 INSN_3(ALU, MOV, X), \
1519 INSN_3(ALU, ARSH, X), \
1520 INSN_3(ALU, DIV, X), \
1521 INSN_3(ALU, MOD, X), \
1523 INSN_3(ALU, END, TO_BE), \
1524 INSN_3(ALU, END, TO_LE), \
1525 /* Immediate based. */ \
1526 INSN_3(ALU, ADD, K), \
1527 INSN_3(ALU, SUB, K), \
1528 INSN_3(ALU, AND, K), \
1529 INSN_3(ALU, OR, K), \
1530 INSN_3(ALU, LSH, K), \
1531 INSN_3(ALU, RSH, K), \
1532 INSN_3(ALU, XOR, K), \
1533 INSN_3(ALU, MUL, K), \
1534 INSN_3(ALU, MOV, K), \
1535 INSN_3(ALU, ARSH, K), \
1536 INSN_3(ALU, DIV, K), \
1537 INSN_3(ALU, MOD, K), \
1538 /* 64 bit ALU operations. */ \
1539 /* Register based. */ \
1540 INSN_3(ALU64, ADD, X), \
1541 INSN_3(ALU64, SUB, X), \
1542 INSN_3(ALU64, AND, X), \
1543 INSN_3(ALU64, OR, X), \
1544 INSN_3(ALU64, LSH, X), \
1545 INSN_3(ALU64, RSH, X), \
1546 INSN_3(ALU64, XOR, X), \
1547 INSN_3(ALU64, MUL, X), \
1548 INSN_3(ALU64, MOV, X), \
1549 INSN_3(ALU64, ARSH, X), \
1550 INSN_3(ALU64, DIV, X), \
1551 INSN_3(ALU64, MOD, X), \
1552 INSN_2(ALU64, NEG), \
1553 /* Immediate based. */ \
1554 INSN_3(ALU64, ADD, K), \
1555 INSN_3(ALU64, SUB, K), \
1556 INSN_3(ALU64, AND, K), \
1557 INSN_3(ALU64, OR, K), \
1558 INSN_3(ALU64, LSH, K), \
1559 INSN_3(ALU64, RSH, K), \
1560 INSN_3(ALU64, XOR, K), \
1561 INSN_3(ALU64, MUL, K), \
1562 INSN_3(ALU64, MOV, K), \
1563 INSN_3(ALU64, ARSH, K), \
1564 INSN_3(ALU64, DIV, K), \
1565 INSN_3(ALU64, MOD, K), \
1566 /* Call instruction. */ \
1567 INSN_2(JMP, CALL), \
1568 /* Exit instruction. */ \
1569 INSN_2(JMP, EXIT), \
1570 /* 32-bit Jump instructions. */ \
1571 /* Register based. */ \
1572 INSN_3(JMP32, JEQ, X), \
1573 INSN_3(JMP32, JNE, X), \
1574 INSN_3(JMP32, JGT, X), \
1575 INSN_3(JMP32, JLT, X), \
1576 INSN_3(JMP32, JGE, X), \
1577 INSN_3(JMP32, JLE, X), \
1578 INSN_3(JMP32, JSGT, X), \
1579 INSN_3(JMP32, JSLT, X), \
1580 INSN_3(JMP32, JSGE, X), \
1581 INSN_3(JMP32, JSLE, X), \
1582 INSN_3(JMP32, JSET, X), \
1583 /* Immediate based. */ \
1584 INSN_3(JMP32, JEQ, K), \
1585 INSN_3(JMP32, JNE, K), \
1586 INSN_3(JMP32, JGT, K), \
1587 INSN_3(JMP32, JLT, K), \
1588 INSN_3(JMP32, JGE, K), \
1589 INSN_3(JMP32, JLE, K), \
1590 INSN_3(JMP32, JSGT, K), \
1591 INSN_3(JMP32, JSLT, K), \
1592 INSN_3(JMP32, JSGE, K), \
1593 INSN_3(JMP32, JSLE, K), \
1594 INSN_3(JMP32, JSET, K), \
1595 /* Jump instructions. */ \
1596 /* Register based. */ \
1597 INSN_3(JMP, JEQ, X), \
1598 INSN_3(JMP, JNE, X), \
1599 INSN_3(JMP, JGT, X), \
1600 INSN_3(JMP, JLT, X), \
1601 INSN_3(JMP, JGE, X), \
1602 INSN_3(JMP, JLE, X), \
1603 INSN_3(JMP, JSGT, X), \
1604 INSN_3(JMP, JSLT, X), \
1605 INSN_3(JMP, JSGE, X), \
1606 INSN_3(JMP, JSLE, X), \
1607 INSN_3(JMP, JSET, X), \
1608 /* Immediate based. */ \
1609 INSN_3(JMP, JEQ, K), \
1610 INSN_3(JMP, JNE, K), \
1611 INSN_3(JMP, JGT, K), \
1612 INSN_3(JMP, JLT, K), \
1613 INSN_3(JMP, JGE, K), \
1614 INSN_3(JMP, JLE, K), \
1615 INSN_3(JMP, JSGT, K), \
1616 INSN_3(JMP, JSLT, K), \
1617 INSN_3(JMP, JSGE, K), \
1618 INSN_3(JMP, JSLE, K), \
1619 INSN_3(JMP, JSET, K), \
1621 /* Store instructions. */ \
1622 /* Register based. */ \
1623 INSN_3(STX, MEM, B), \
1624 INSN_3(STX, MEM, H), \
1625 INSN_3(STX, MEM, W), \
1626 INSN_3(STX, MEM, DW), \
1627 INSN_3(STX, ATOMIC, W), \
1628 INSN_3(STX, ATOMIC, DW), \
1629 /* Immediate based. */ \
1630 INSN_3(ST, MEM, B), \
1631 INSN_3(ST, MEM, H), \
1632 INSN_3(ST, MEM, W), \
1633 INSN_3(ST, MEM, DW), \
1634 /* Load instructions. */ \
1635 /* Register based. */ \
1636 INSN_3(LDX, MEM, B), \
1637 INSN_3(LDX, MEM, H), \
1638 INSN_3(LDX, MEM, W), \
1639 INSN_3(LDX, MEM, DW), \
1640 /* Immediate based. */ \
1643 bool bpf_opcode_in_insntable(u8 code)
1645 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1646 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1647 static const bool public_insntable[256] = {
1648 [0 ... 255] = false,
1649 /* Now overwrite non-defaults ... */
1650 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1651 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1652 [BPF_LD | BPF_ABS | BPF_B] = true,
1653 [BPF_LD | BPF_ABS | BPF_H] = true,
1654 [BPF_LD | BPF_ABS | BPF_W] = true,
1655 [BPF_LD | BPF_IND | BPF_B] = true,
1656 [BPF_LD | BPF_IND | BPF_H] = true,
1657 [BPF_LD | BPF_IND | BPF_W] = true,
1659 #undef BPF_INSN_3_TBL
1660 #undef BPF_INSN_2_TBL
1661 return public_insntable[code];
1664 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1665 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1667 memset(dst, 0, size);
1672 * ___bpf_prog_run - run eBPF program on a given context
1673 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1674 * @insn: is the array of eBPF instructions
1676 * Decode and execute eBPF instructions.
1678 * Return: whatever value is in %BPF_R0 at program exit
1680 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1682 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1683 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1684 static const void * const jumptable[256] __annotate_jump_table = {
1685 [0 ... 255] = &&default_label,
1686 /* Now overwrite non-defaults ... */
1687 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1688 /* Non-UAPI available opcodes. */
1689 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1690 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1691 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1692 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1693 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1694 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1695 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1697 #undef BPF_INSN_3_LBL
1698 #undef BPF_INSN_2_LBL
1699 u32 tail_call_cnt = 0;
1701 #define CONT ({ insn++; goto select_insn; })
1702 #define CONT_JMP ({ insn++; goto select_insn; })
1705 goto *jumptable[insn->code];
1707 /* Explicitly mask the register-based shift amounts with 63 or 31
1708 * to avoid undefined behavior. Normally this won't affect the
1709 * generated code, for example, in case of native 64 bit archs such
1710 * as x86-64 or arm64, the compiler is optimizing the AND away for
1711 * the interpreter. In case of JITs, each of the JIT backends compiles
1712 * the BPF shift operations to machine instructions which produce
1713 * implementation-defined results in such a case; the resulting
1714 * contents of the register may be arbitrary, but program behaviour
1715 * as a whole remains defined. In other words, in case of JIT backends,
1716 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1719 #define SHT(OPCODE, OP) \
1720 ALU64_##OPCODE##_X: \
1721 DST = DST OP (SRC & 63); \
1724 DST = (u32) DST OP ((u32) SRC & 31); \
1726 ALU64_##OPCODE##_K: \
1730 DST = (u32) DST OP (u32) IMM; \
1733 #define ALU(OPCODE, OP) \
1734 ALU64_##OPCODE##_X: \
1738 DST = (u32) DST OP (u32) SRC; \
1740 ALU64_##OPCODE##_K: \
1744 DST = (u32) DST OP (u32) IMM; \
1775 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1779 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1782 DST = (u64) (u32) (((s32) DST) >> IMM);
1785 (*(s64 *) &DST) >>= (SRC & 63);
1788 (*(s64 *) &DST) >>= IMM;
1791 div64_u64_rem(DST, SRC, &AX);
1796 DST = do_div(AX, (u32) SRC);
1799 div64_u64_rem(DST, IMM, &AX);
1804 DST = do_div(AX, (u32) IMM);
1807 DST = div64_u64(DST, SRC);
1811 do_div(AX, (u32) SRC);
1815 DST = div64_u64(DST, IMM);
1819 do_div(AX, (u32) IMM);
1825 DST = (__force u16) cpu_to_be16(DST);
1828 DST = (__force u32) cpu_to_be32(DST);
1831 DST = (__force u64) cpu_to_be64(DST);
1838 DST = (__force u16) cpu_to_le16(DST);
1841 DST = (__force u32) cpu_to_le32(DST);
1844 DST = (__force u64) cpu_to_le64(DST);
1851 /* Function call scratches BPF_R1-BPF_R5 registers,
1852 * preserves BPF_R6-BPF_R9, and stores return value
1855 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1860 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1863 insn + insn->off + 1);
1867 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1868 struct bpf_array *array = container_of(map, struct bpf_array, map);
1869 struct bpf_prog *prog;
1872 if (unlikely(index >= array->map.max_entries))
1875 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
1880 prog = READ_ONCE(array->ptrs[index]);
1884 /* ARG1 at this point is guaranteed to point to CTX from
1885 * the verifier side due to the fact that the tail call is
1886 * handled like a helper, that is, bpf_tail_call_proto,
1887 * where arg1_type is ARG_PTR_TO_CTX.
1889 insn = prog->insnsi;
1900 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1902 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1903 insn += insn->off; \
1907 JMP32_##OPCODE##_X: \
1908 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1909 insn += insn->off; \
1914 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1915 insn += insn->off; \
1919 JMP32_##OPCODE##_K: \
1920 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1921 insn += insn->off; \
1925 COND_JMP(u, JEQ, ==)
1926 COND_JMP(u, JNE, !=)
1929 COND_JMP(u, JGE, >=)
1930 COND_JMP(u, JLE, <=)
1931 COND_JMP(u, JSET, &)
1932 COND_JMP(s, JSGT, >)
1933 COND_JMP(s, JSLT, <)
1934 COND_JMP(s, JSGE, >=)
1935 COND_JMP(s, JSLE, <=)
1937 /* ST, STX and LDX*/
1939 /* Speculation barrier for mitigating Speculative Store Bypass.
1940 * In case of arm64, we rely on the firmware mitigation as
1941 * controlled via the ssbd kernel parameter. Whenever the
1942 * mitigation is enabled, it works for all of the kernel code
1943 * with no need to provide any additional instructions here.
1944 * In case of x86, we use 'lfence' insn for mitigation. We
1945 * reuse preexisting logic from Spectre v1 mitigation that
1946 * happens to produce the required code on x86 for v4 as well.
1952 #define LDST(SIZEOP, SIZE) \
1954 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1957 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1960 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1962 LDX_PROBE_MEM_##SIZEOP: \
1963 bpf_probe_read_kernel(&DST, sizeof(SIZE), \
1964 (const void *)(long) (SRC + insn->off)); \
1965 DST = *((SIZE *)&DST); \
1974 #define ATOMIC_ALU_OP(BOP, KOP) \
1976 if (BPF_SIZE(insn->code) == BPF_W) \
1977 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
1978 (DST + insn->off)); \
1980 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
1981 (DST + insn->off)); \
1983 case BOP | BPF_FETCH: \
1984 if (BPF_SIZE(insn->code) == BPF_W) \
1985 SRC = (u32) atomic_fetch_##KOP( \
1987 (atomic_t *)(unsigned long) (DST + insn->off)); \
1989 SRC = (u64) atomic64_fetch_##KOP( \
1991 (atomic64_t *)(unsigned long) (DST + insn->off)); \
1997 ATOMIC_ALU_OP(BPF_ADD, add)
1998 ATOMIC_ALU_OP(BPF_AND, and)
1999 ATOMIC_ALU_OP(BPF_OR, or)
2000 ATOMIC_ALU_OP(BPF_XOR, xor)
2001 #undef ATOMIC_ALU_OP
2004 if (BPF_SIZE(insn->code) == BPF_W)
2005 SRC = (u32) atomic_xchg(
2006 (atomic_t *)(unsigned long) (DST + insn->off),
2009 SRC = (u64) atomic64_xchg(
2010 (atomic64_t *)(unsigned long) (DST + insn->off),
2014 if (BPF_SIZE(insn->code) == BPF_W)
2015 BPF_R0 = (u32) atomic_cmpxchg(
2016 (atomic_t *)(unsigned long) (DST + insn->off),
2017 (u32) BPF_R0, (u32) SRC);
2019 BPF_R0 = (u64) atomic64_cmpxchg(
2020 (atomic64_t *)(unsigned long) (DST + insn->off),
2021 (u64) BPF_R0, (u64) SRC);
2030 /* If we ever reach this, we have a bug somewhere. Die hard here
2031 * instead of just returning 0; we could be somewhere in a subprog,
2032 * so execution could continue otherwise which we do /not/ want.
2034 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
2036 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
2037 insn->code, insn->imm);
2042 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2043 #define DEFINE_BPF_PROG_RUN(stack_size) \
2044 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2046 u64 stack[stack_size / sizeof(u64)]; \
2047 u64 regs[MAX_BPF_EXT_REG]; \
2049 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2050 ARG1 = (u64) (unsigned long) ctx; \
2051 return ___bpf_prog_run(regs, insn); \
2054 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2055 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2056 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2057 const struct bpf_insn *insn) \
2059 u64 stack[stack_size / sizeof(u64)]; \
2060 u64 regs[MAX_BPF_EXT_REG]; \
2062 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2068 return ___bpf_prog_run(regs, insn); \
2071 #define EVAL1(FN, X) FN(X)
2072 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2073 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2074 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2075 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2076 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2078 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2079 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2080 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2082 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2083 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2084 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2086 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2088 static unsigned int (*interpreters[])(const void *ctx,
2089 const struct bpf_insn *insn) = {
2090 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2091 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2092 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2094 #undef PROG_NAME_LIST
2095 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2096 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2097 const struct bpf_insn *insn) = {
2098 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2099 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2100 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2102 #undef PROG_NAME_LIST
2104 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2106 stack_depth = max_t(u32, stack_depth, 1);
2107 insn->off = (s16) insn->imm;
2108 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2109 __bpf_call_base_args;
2110 insn->code = BPF_JMP | BPF_CALL_ARGS;
2114 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2115 const struct bpf_insn *insn)
2117 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2118 * is not working properly, so warn about it!
2125 bool bpf_prog_map_compatible(struct bpf_map *map,
2126 const struct bpf_prog *fp)
2130 if (fp->kprobe_override)
2133 spin_lock(&map->owner.lock);
2134 if (!map->owner.type) {
2135 /* There's no owner yet where we could check for
2138 map->owner.type = fp->type;
2139 map->owner.jited = fp->jited;
2140 map->owner.xdp_has_frags = fp->aux->xdp_has_frags;
2143 ret = map->owner.type == fp->type &&
2144 map->owner.jited == fp->jited &&
2145 map->owner.xdp_has_frags == fp->aux->xdp_has_frags;
2147 spin_unlock(&map->owner.lock);
2152 static int bpf_check_tail_call(const struct bpf_prog *fp)
2154 struct bpf_prog_aux *aux = fp->aux;
2157 mutex_lock(&aux->used_maps_mutex);
2158 for (i = 0; i < aux->used_map_cnt; i++) {
2159 struct bpf_map *map = aux->used_maps[i];
2161 if (!map_type_contains_progs(map))
2164 if (!bpf_prog_map_compatible(map, fp)) {
2171 mutex_unlock(&aux->used_maps_mutex);
2175 static void bpf_prog_select_func(struct bpf_prog *fp)
2177 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2178 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2180 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2182 fp->bpf_func = __bpf_prog_ret0_warn;
2187 * bpf_prog_select_runtime - select exec runtime for BPF program
2188 * @fp: bpf_prog populated with BPF program
2189 * @err: pointer to error variable
2191 * Try to JIT eBPF program, if JIT is not available, use interpreter.
2192 * The BPF program will be executed via bpf_prog_run() function.
2194 * Return: the &fp argument along with &err set to 0 for success or
2195 * a negative errno code on failure
2197 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2199 /* In case of BPF to BPF calls, verifier did all the prep
2200 * work with regards to JITing, etc.
2202 bool jit_needed = false;
2207 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2208 bpf_prog_has_kfunc_call(fp))
2211 bpf_prog_select_func(fp);
2213 /* eBPF JITs can rewrite the program in case constant
2214 * blinding is active. However, in case of error during
2215 * blinding, bpf_int_jit_compile() must always return a
2216 * valid program, which in this case would simply not
2217 * be JITed, but falls back to the interpreter.
2219 if (!bpf_prog_is_dev_bound(fp->aux)) {
2220 *err = bpf_prog_alloc_jited_linfo(fp);
2224 fp = bpf_int_jit_compile(fp);
2225 bpf_prog_jit_attempt_done(fp);
2226 if (!fp->jited && jit_needed) {
2231 *err = bpf_prog_offload_compile(fp);
2237 bpf_prog_lock_ro(fp);
2239 /* The tail call compatibility check can only be done at
2240 * this late stage as we need to determine, if we deal
2241 * with JITed or non JITed program concatenations and not
2242 * all eBPF JITs might immediately support all features.
2244 *err = bpf_check_tail_call(fp);
2248 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2250 static unsigned int __bpf_prog_ret1(const void *ctx,
2251 const struct bpf_insn *insn)
2256 static struct bpf_prog_dummy {
2257 struct bpf_prog prog;
2258 } dummy_bpf_prog = {
2260 .bpf_func = __bpf_prog_ret1,
2264 struct bpf_empty_prog_array bpf_empty_prog_array = {
2267 EXPORT_SYMBOL(bpf_empty_prog_array);
2269 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2272 return kzalloc(sizeof(struct bpf_prog_array) +
2273 sizeof(struct bpf_prog_array_item) *
2277 return &bpf_empty_prog_array.hdr;
2280 void bpf_prog_array_free(struct bpf_prog_array *progs)
2282 if (!progs || progs == &bpf_empty_prog_array.hdr)
2284 kfree_rcu(progs, rcu);
2287 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2289 struct bpf_prog_array *progs;
2291 progs = container_of(rcu, struct bpf_prog_array, rcu);
2292 kfree_rcu(progs, rcu);
2295 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2297 if (!progs || progs == &bpf_empty_prog_array.hdr)
2299 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2302 int bpf_prog_array_length(struct bpf_prog_array *array)
2304 struct bpf_prog_array_item *item;
2307 for (item = array->items; item->prog; item++)
2308 if (item->prog != &dummy_bpf_prog.prog)
2313 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2315 struct bpf_prog_array_item *item;
2317 for (item = array->items; item->prog; item++)
2318 if (item->prog != &dummy_bpf_prog.prog)
2323 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2327 struct bpf_prog_array_item *item;
2330 for (item = array->items; item->prog; item++) {
2331 if (item->prog == &dummy_bpf_prog.prog)
2333 prog_ids[i] = item->prog->aux->id;
2334 if (++i == request_cnt) {
2340 return !!(item->prog);
2343 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2344 __u32 __user *prog_ids, u32 cnt)
2346 unsigned long err = 0;
2350 /* users of this function are doing:
2351 * cnt = bpf_prog_array_length();
2353 * bpf_prog_array_copy_to_user(..., cnt);
2354 * so below kcalloc doesn't need extra cnt > 0 check.
2356 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2359 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2360 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2369 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2370 struct bpf_prog *old_prog)
2372 struct bpf_prog_array_item *item;
2374 for (item = array->items; item->prog; item++)
2375 if (item->prog == old_prog) {
2376 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2382 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2383 * index into the program array with
2384 * a dummy no-op program.
2385 * @array: a bpf_prog_array
2386 * @index: the index of the program to replace
2388 * Skips over dummy programs, by not counting them, when calculating
2389 * the position of the program to replace.
2393 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2394 * * -ENOENT - Index out of range
2396 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2398 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2402 * bpf_prog_array_update_at() - Updates the program at the given index
2403 * into the program array.
2404 * @array: a bpf_prog_array
2405 * @index: the index of the program to update
2406 * @prog: the program to insert into the array
2408 * Skips over dummy programs, by not counting them, when calculating
2409 * the position of the program to update.
2413 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2414 * * -ENOENT - Index out of range
2416 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2417 struct bpf_prog *prog)
2419 struct bpf_prog_array_item *item;
2421 if (unlikely(index < 0))
2424 for (item = array->items; item->prog; item++) {
2425 if (item->prog == &dummy_bpf_prog.prog)
2428 WRITE_ONCE(item->prog, prog);
2436 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2437 struct bpf_prog *exclude_prog,
2438 struct bpf_prog *include_prog,
2440 struct bpf_prog_array **new_array)
2442 int new_prog_cnt, carry_prog_cnt = 0;
2443 struct bpf_prog_array_item *existing, *new;
2444 struct bpf_prog_array *array;
2445 bool found_exclude = false;
2447 /* Figure out how many existing progs we need to carry over to
2451 existing = old_array->items;
2452 for (; existing->prog; existing++) {
2453 if (existing->prog == exclude_prog) {
2454 found_exclude = true;
2457 if (existing->prog != &dummy_bpf_prog.prog)
2459 if (existing->prog == include_prog)
2464 if (exclude_prog && !found_exclude)
2467 /* How many progs (not NULL) will be in the new array? */
2468 new_prog_cnt = carry_prog_cnt;
2472 /* Do we have any prog (not NULL) in the new array? */
2473 if (!new_prog_cnt) {
2478 /* +1 as the end of prog_array is marked with NULL */
2479 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2484 /* Fill in the new prog array */
2485 if (carry_prog_cnt) {
2486 existing = old_array->items;
2487 for (; existing->prog; existing++) {
2488 if (existing->prog == exclude_prog ||
2489 existing->prog == &dummy_bpf_prog.prog)
2492 new->prog = existing->prog;
2493 new->bpf_cookie = existing->bpf_cookie;
2498 new->prog = include_prog;
2499 new->bpf_cookie = bpf_cookie;
2507 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2508 u32 *prog_ids, u32 request_cnt,
2514 cnt = bpf_prog_array_length(array);
2518 /* return early if user requested only program count or nothing to copy */
2519 if (!request_cnt || !cnt)
2522 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2523 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2527 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2528 struct bpf_map **used_maps, u32 len)
2530 struct bpf_map *map;
2533 for (i = 0; i < len; i++) {
2535 if (map->ops->map_poke_untrack)
2536 map->ops->map_poke_untrack(map, aux);
2541 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2543 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2544 kfree(aux->used_maps);
2547 void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2548 struct btf_mod_pair *used_btfs, u32 len)
2550 #ifdef CONFIG_BPF_SYSCALL
2551 struct btf_mod_pair *btf_mod;
2554 for (i = 0; i < len; i++) {
2555 btf_mod = &used_btfs[i];
2556 if (btf_mod->module)
2557 module_put(btf_mod->module);
2558 btf_put(btf_mod->btf);
2563 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2565 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2566 kfree(aux->used_btfs);
2569 static void bpf_prog_free_deferred(struct work_struct *work)
2571 struct bpf_prog_aux *aux;
2574 aux = container_of(work, struct bpf_prog_aux, work);
2575 #ifdef CONFIG_BPF_SYSCALL
2576 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2578 #ifdef CONFIG_CGROUP_BPF
2579 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2580 bpf_cgroup_atype_put(aux->cgroup_atype);
2582 bpf_free_used_maps(aux);
2583 bpf_free_used_btfs(aux);
2584 if (bpf_prog_is_dev_bound(aux))
2585 bpf_prog_offload_destroy(aux->prog);
2586 #ifdef CONFIG_PERF_EVENTS
2587 if (aux->prog->has_callchain_buf)
2588 put_callchain_buffers();
2590 if (aux->dst_trampoline)
2591 bpf_trampoline_put(aux->dst_trampoline);
2592 for (i = 0; i < aux->func_cnt; i++) {
2593 /* We can just unlink the subprog poke descriptor table as
2594 * it was originally linked to the main program and is also
2595 * released along with it.
2597 aux->func[i]->aux->poke_tab = NULL;
2598 bpf_jit_free(aux->func[i]);
2600 if (aux->func_cnt) {
2602 bpf_prog_unlock_free(aux->prog);
2604 bpf_jit_free(aux->prog);
2608 void bpf_prog_free(struct bpf_prog *fp)
2610 struct bpf_prog_aux *aux = fp->aux;
2613 bpf_prog_put(aux->dst_prog);
2614 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2615 schedule_work(&aux->work);
2617 EXPORT_SYMBOL_GPL(bpf_prog_free);
2619 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2620 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2622 void bpf_user_rnd_init_once(void)
2624 prandom_init_once(&bpf_user_rnd_state);
2627 BPF_CALL_0(bpf_user_rnd_u32)
2629 /* Should someone ever have the rather unwise idea to use some
2630 * of the registers passed into this function, then note that
2631 * this function is called from native eBPF and classic-to-eBPF
2632 * transformations. Register assignments from both sides are
2633 * different, f.e. classic always sets fn(ctx, A, X) here.
2635 struct rnd_state *state;
2638 state = &get_cpu_var(bpf_user_rnd_state);
2639 res = prandom_u32_state(state);
2640 put_cpu_var(bpf_user_rnd_state);
2645 BPF_CALL_0(bpf_get_raw_cpu_id)
2647 return raw_smp_processor_id();
2650 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2651 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2652 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2653 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2654 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2655 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2656 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2657 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2658 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2659 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2660 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2662 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2663 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2664 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2665 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2666 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2667 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2669 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2670 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2671 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2672 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2673 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2674 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2675 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2676 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2677 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2678 const struct bpf_func_proto bpf_set_retval_proto __weak;
2679 const struct bpf_func_proto bpf_get_retval_proto __weak;
2681 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2686 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2692 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2693 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2697 EXPORT_SYMBOL_GPL(bpf_event_output);
2699 /* Always built-in helper functions. */
2700 const struct bpf_func_proto bpf_tail_call_proto = {
2703 .ret_type = RET_VOID,
2704 .arg1_type = ARG_PTR_TO_CTX,
2705 .arg2_type = ARG_CONST_MAP_PTR,
2706 .arg3_type = ARG_ANYTHING,
2709 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2710 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2711 * eBPF and implicitly also cBPF can get JITed!
2713 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2718 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2719 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2721 void __weak bpf_jit_compile(struct bpf_prog *prog)
2725 bool __weak bpf_helper_changes_pkt_data(void *func)
2730 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2731 * analysis code and wants explicit zero extension inserted by verifier.
2732 * Otherwise, return FALSE.
2734 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2735 * you don't override this. JITs that don't want these extra insns can detect
2736 * them using insn_is_zext.
2738 bool __weak bpf_jit_needs_zext(void)
2743 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2744 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2749 bool __weak bpf_jit_supports_kfunc_call(void)
2754 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2755 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2757 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2763 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2764 void *addr1, void *addr2)
2769 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
2771 return ERR_PTR(-ENOTSUPP);
2774 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
2779 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2780 EXPORT_SYMBOL(bpf_stats_enabled_key);
2782 /* All definitions of tracepoints related to BPF. */
2783 #define CREATE_TRACE_POINTS
2784 #include <linux/bpf_trace.h>
2786 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2787 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);