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 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
657 if (!bpf_prog_kallsyms_candidate(fp) ||
661 bpf_prog_ksym_set_addr(fp);
662 bpf_prog_ksym_set_name(fp);
663 fp->aux->ksym.prog = true;
665 bpf_ksym_add(&fp->aux->ksym);
668 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
670 if (!bpf_prog_kallsyms_candidate(fp))
673 bpf_ksym_del(&fp->aux->ksym);
676 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
678 struct latch_tree_node *n;
680 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
681 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
684 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
685 unsigned long *off, char *sym)
687 struct bpf_ksym *ksym;
691 ksym = bpf_ksym_find(addr);
693 unsigned long symbol_start = ksym->start;
694 unsigned long symbol_end = ksym->end;
696 strncpy(sym, ksym->name, KSYM_NAME_LEN);
700 *size = symbol_end - symbol_start;
702 *off = addr - symbol_start;
709 bool is_bpf_text_address(unsigned long addr)
714 ret = bpf_ksym_find(addr) != NULL;
720 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
722 struct bpf_ksym *ksym = bpf_ksym_find(addr);
724 return ksym && ksym->prog ?
725 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
729 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
731 const struct exception_table_entry *e = NULL;
732 struct bpf_prog *prog;
735 prog = bpf_prog_ksym_find(addr);
738 if (!prog->aux->num_exentries)
741 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
747 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
750 struct bpf_ksym *ksym;
754 if (!bpf_jit_kallsyms_enabled())
758 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
762 strncpy(sym, ksym->name, KSYM_NAME_LEN);
764 *value = ksym->start;
765 *type = BPF_SYM_ELF_TYPE;
775 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
776 struct bpf_jit_poke_descriptor *poke)
778 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
779 static const u32 poke_tab_max = 1024;
780 u32 slot = prog->aux->size_poke_tab;
783 if (size > poke_tab_max)
785 if (poke->tailcall_target || poke->tailcall_target_stable ||
786 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
789 switch (poke->reason) {
790 case BPF_POKE_REASON_TAIL_CALL:
791 if (!poke->tail_call.map)
798 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
802 memcpy(&tab[slot], poke, sizeof(*poke));
803 prog->aux->size_poke_tab = size;
804 prog->aux->poke_tab = tab;
810 * BPF program pack allocator.
812 * Most BPF programs are pretty small. Allocating a hole page for each
813 * program is sometime a waste. Many small bpf program also adds pressure
814 * to instruction TLB. To solve this issue, we introduce a BPF program pack
815 * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
816 * to host BPF programs.
818 #define BPF_PROG_CHUNK_SHIFT 6
819 #define BPF_PROG_CHUNK_SIZE (1 << BPF_PROG_CHUNK_SHIFT)
820 #define BPF_PROG_CHUNK_MASK (~(BPF_PROG_CHUNK_SIZE - 1))
822 struct bpf_prog_pack {
823 struct list_head list;
825 unsigned long bitmap[];
828 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
830 static DEFINE_MUTEX(pack_mutex);
831 static LIST_HEAD(pack_list);
833 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
834 * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
837 #define BPF_PROG_PACK_SIZE (PMD_SIZE * num_possible_nodes())
839 #define BPF_PROG_PACK_SIZE PAGE_SIZE
842 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
844 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
846 struct bpf_prog_pack *pack;
848 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
852 pack->ptr = module_alloc(BPF_PROG_PACK_SIZE);
857 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
858 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
859 list_add_tail(&pack->list, &pack_list);
861 set_vm_flush_reset_perms(pack->ptr);
862 set_memory_ro((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
863 set_memory_x((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
867 static void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
869 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
870 struct bpf_prog_pack *pack;
874 mutex_lock(&pack_mutex);
875 if (size > BPF_PROG_PACK_SIZE) {
876 size = round_up(size, PAGE_SIZE);
877 ptr = module_alloc(size);
879 bpf_fill_ill_insns(ptr, size);
880 set_vm_flush_reset_perms(ptr);
881 set_memory_ro((unsigned long)ptr, size / PAGE_SIZE);
882 set_memory_x((unsigned long)ptr, size / PAGE_SIZE);
886 list_for_each_entry(pack, &pack_list, list) {
887 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
889 if (pos < BPF_PROG_CHUNK_COUNT)
890 goto found_free_area;
893 pack = alloc_new_pack(bpf_fill_ill_insns);
900 bitmap_set(pack->bitmap, pos, nbits);
901 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
904 mutex_unlock(&pack_mutex);
908 static void bpf_prog_pack_free(struct bpf_binary_header *hdr)
910 struct bpf_prog_pack *pack = NULL, *tmp;
914 mutex_lock(&pack_mutex);
915 if (hdr->size > BPF_PROG_PACK_SIZE) {
920 list_for_each_entry(tmp, &pack_list, list) {
921 if ((void *)hdr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > (void *)hdr) {
927 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
930 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size);
931 pos = ((unsigned long)hdr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
933 WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size),
934 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
936 bitmap_clear(pack->bitmap, pos, nbits);
937 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
938 BPF_PROG_CHUNK_COUNT, 0) == 0) {
939 list_del(&pack->list);
940 module_memfree(pack->ptr);
944 mutex_unlock(&pack_mutex);
947 static atomic_long_t bpf_jit_current;
949 /* Can be overridden by an arch's JIT compiler if it has a custom,
950 * dedicated BPF backend memory area, or if neither of the two
953 u64 __weak bpf_jit_alloc_exec_limit(void)
955 #if defined(MODULES_VADDR)
956 return MODULES_END - MODULES_VADDR;
958 return VMALLOC_END - VMALLOC_START;
962 static int __init bpf_jit_charge_init(void)
964 /* Only used as heuristic here to derive limit. */
965 bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
966 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2,
967 PAGE_SIZE), LONG_MAX);
970 pure_initcall(bpf_jit_charge_init);
972 int bpf_jit_charge_modmem(u32 size)
974 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
975 if (!bpf_capable()) {
976 atomic_long_sub(size, &bpf_jit_current);
984 void bpf_jit_uncharge_modmem(u32 size)
986 atomic_long_sub(size, &bpf_jit_current);
989 void *__weak bpf_jit_alloc_exec(unsigned long size)
991 return module_alloc(size);
994 void __weak bpf_jit_free_exec(void *addr)
996 module_memfree(addr);
999 struct bpf_binary_header *
1000 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1001 unsigned int alignment,
1002 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1004 struct bpf_binary_header *hdr;
1005 u32 size, hole, start;
1007 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1008 alignment > BPF_IMAGE_ALIGNMENT);
1010 /* Most of BPF filters are really small, but if some of them
1011 * fill a page, allow at least 128 extra bytes to insert a
1012 * random section of illegal instructions.
1014 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1016 if (bpf_jit_charge_modmem(size))
1018 hdr = bpf_jit_alloc_exec(size);
1020 bpf_jit_uncharge_modmem(size);
1024 /* Fill space with illegal/arch-dep instructions. */
1025 bpf_fill_ill_insns(hdr, size);
1028 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1029 PAGE_SIZE - sizeof(*hdr));
1030 start = (get_random_int() % hole) & ~(alignment - 1);
1032 /* Leave a random number of instructions before BPF code. */
1033 *image_ptr = &hdr->image[start];
1038 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1040 u32 size = hdr->size;
1042 bpf_jit_free_exec(hdr);
1043 bpf_jit_uncharge_modmem(size);
1046 /* Allocate jit binary from bpf_prog_pack allocator.
1047 * Since the allocated memory is RO+X, the JIT engine cannot write directly
1048 * to the memory. To solve this problem, a RW buffer is also allocated at
1049 * as the same time. The JIT engine should calculate offsets based on the
1050 * RO memory address, but write JITed program to the RW buffer. Once the
1051 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1052 * the JITed program to the RO memory.
1054 struct bpf_binary_header *
1055 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1056 unsigned int alignment,
1057 struct bpf_binary_header **rw_header,
1059 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1061 struct bpf_binary_header *ro_header;
1062 u32 size, hole, start;
1064 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1065 alignment > BPF_IMAGE_ALIGNMENT);
1067 /* add 16 bytes for a random section of illegal instructions */
1068 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1070 if (bpf_jit_charge_modmem(size))
1072 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1074 bpf_jit_uncharge_modmem(size);
1078 *rw_header = kvmalloc(size, GFP_KERNEL);
1080 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size));
1081 bpf_prog_pack_free(ro_header);
1082 bpf_jit_uncharge_modmem(size);
1086 /* Fill space with illegal/arch-dep instructions. */
1087 bpf_fill_ill_insns(*rw_header, size);
1088 (*rw_header)->size = size;
1090 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1091 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1092 start = (get_random_int() % hole) & ~(alignment - 1);
1094 *image_ptr = &ro_header->image[start];
1095 *rw_image = &(*rw_header)->image[start];
1100 /* Copy JITed text from rw_header to its final location, the ro_header. */
1101 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog,
1102 struct bpf_binary_header *ro_header,
1103 struct bpf_binary_header *rw_header)
1107 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1112 bpf_prog_pack_free(ro_header);
1113 return PTR_ERR(ptr);
1118 /* bpf_jit_binary_pack_free is called in two different scenarios:
1119 * 1) when the program is freed after;
1120 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1121 * For case 2), we need to free both the RO memory and the RW buffer.
1123 * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1124 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1125 * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1126 * bpf_arch_text_copy (when jit fails).
1128 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1129 struct bpf_binary_header *rw_header)
1131 u32 size = ro_header->size;
1133 bpf_prog_pack_free(ro_header);
1135 bpf_jit_uncharge_modmem(size);
1138 struct bpf_binary_header *
1139 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1141 unsigned long real_start = (unsigned long)fp->bpf_func;
1144 addr = real_start & BPF_PROG_CHUNK_MASK;
1145 return (void *)addr;
1148 static inline struct bpf_binary_header *
1149 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1151 unsigned long real_start = (unsigned long)fp->bpf_func;
1154 addr = real_start & PAGE_MASK;
1155 return (void *)addr;
1158 /* This symbol is only overridden by archs that have different
1159 * requirements than the usual eBPF JITs, f.e. when they only
1160 * implement cBPF JIT, do not set images read-only, etc.
1162 void __weak bpf_jit_free(struct bpf_prog *fp)
1165 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1167 bpf_jit_binary_free(hdr);
1168 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1171 bpf_prog_unlock_free(fp);
1174 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1175 const struct bpf_insn *insn, bool extra_pass,
1176 u64 *func_addr, bool *func_addr_fixed)
1178 s16 off = insn->off;
1179 s32 imm = insn->imm;
1182 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1183 if (!*func_addr_fixed) {
1184 /* Place-holder address till the last pass has collected
1185 * all addresses for JITed subprograms in which case we
1186 * can pick them up from prog->aux.
1190 else if (prog->aux->func &&
1191 off >= 0 && off < prog->aux->func_cnt)
1192 addr = (u8 *)prog->aux->func[off]->bpf_func;
1196 /* Address of a BPF helper call. Since part of the core
1197 * kernel, it's always at a fixed location. __bpf_call_base
1198 * and the helper with imm relative to it are both in core
1201 addr = (u8 *)__bpf_call_base + imm;
1204 *func_addr = (unsigned long)addr;
1208 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1209 const struct bpf_insn *aux,
1210 struct bpf_insn *to_buff,
1213 struct bpf_insn *to = to_buff;
1214 u32 imm_rnd = get_random_int();
1217 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
1218 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1220 /* Constraints on AX register:
1222 * AX register is inaccessible from user space. It is mapped in
1223 * all JITs, and used here for constant blinding rewrites. It is
1224 * typically "stateless" meaning its contents are only valid within
1225 * the executed instruction, but not across several instructions.
1226 * There are a few exceptions however which are further detailed
1229 * Constant blinding is only used by JITs, not in the interpreter.
1230 * The interpreter uses AX in some occasions as a local temporary
1231 * register e.g. in DIV or MOD instructions.
1233 * In restricted circumstances, the verifier can also use the AX
1234 * register for rewrites as long as they do not interfere with
1237 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1240 if (from->imm == 0 &&
1241 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
1242 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1243 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1247 switch (from->code) {
1248 case BPF_ALU | BPF_ADD | BPF_K:
1249 case BPF_ALU | BPF_SUB | BPF_K:
1250 case BPF_ALU | BPF_AND | BPF_K:
1251 case BPF_ALU | BPF_OR | BPF_K:
1252 case BPF_ALU | BPF_XOR | BPF_K:
1253 case BPF_ALU | BPF_MUL | BPF_K:
1254 case BPF_ALU | BPF_MOV | BPF_K:
1255 case BPF_ALU | BPF_DIV | BPF_K:
1256 case BPF_ALU | BPF_MOD | BPF_K:
1257 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1258 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1259 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1262 case BPF_ALU64 | BPF_ADD | BPF_K:
1263 case BPF_ALU64 | BPF_SUB | BPF_K:
1264 case BPF_ALU64 | BPF_AND | BPF_K:
1265 case BPF_ALU64 | BPF_OR | BPF_K:
1266 case BPF_ALU64 | BPF_XOR | BPF_K:
1267 case BPF_ALU64 | BPF_MUL | BPF_K:
1268 case BPF_ALU64 | BPF_MOV | BPF_K:
1269 case BPF_ALU64 | BPF_DIV | BPF_K:
1270 case BPF_ALU64 | BPF_MOD | BPF_K:
1271 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1272 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1273 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1276 case BPF_JMP | BPF_JEQ | BPF_K:
1277 case BPF_JMP | BPF_JNE | BPF_K:
1278 case BPF_JMP | BPF_JGT | BPF_K:
1279 case BPF_JMP | BPF_JLT | BPF_K:
1280 case BPF_JMP | BPF_JGE | BPF_K:
1281 case BPF_JMP | BPF_JLE | BPF_K:
1282 case BPF_JMP | BPF_JSGT | BPF_K:
1283 case BPF_JMP | BPF_JSLT | BPF_K:
1284 case BPF_JMP | BPF_JSGE | BPF_K:
1285 case BPF_JMP | BPF_JSLE | BPF_K:
1286 case BPF_JMP | BPF_JSET | BPF_K:
1287 /* Accommodate for extra offset in case of a backjump. */
1291 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1292 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1293 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1296 case BPF_JMP32 | BPF_JEQ | BPF_K:
1297 case BPF_JMP32 | BPF_JNE | BPF_K:
1298 case BPF_JMP32 | BPF_JGT | BPF_K:
1299 case BPF_JMP32 | BPF_JLT | BPF_K:
1300 case BPF_JMP32 | BPF_JGE | BPF_K:
1301 case BPF_JMP32 | BPF_JLE | BPF_K:
1302 case BPF_JMP32 | BPF_JSGT | BPF_K:
1303 case BPF_JMP32 | BPF_JSLT | BPF_K:
1304 case BPF_JMP32 | BPF_JSGE | BPF_K:
1305 case BPF_JMP32 | BPF_JSLE | BPF_K:
1306 case BPF_JMP32 | BPF_JSET | BPF_K:
1307 /* Accommodate for extra offset in case of a backjump. */
1311 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1312 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1313 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1317 case BPF_LD | BPF_IMM | BPF_DW:
1318 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1319 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1320 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1321 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1323 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1324 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1325 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1327 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1328 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1331 case BPF_ST | BPF_MEM | BPF_DW:
1332 case BPF_ST | BPF_MEM | BPF_W:
1333 case BPF_ST | BPF_MEM | BPF_H:
1334 case BPF_ST | BPF_MEM | BPF_B:
1335 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1336 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1337 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1341 return to - to_buff;
1344 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1345 gfp_t gfp_extra_flags)
1347 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1348 struct bpf_prog *fp;
1350 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1352 /* aux->prog still points to the fp_other one, so
1353 * when promoting the clone to the real program,
1354 * this still needs to be adapted.
1356 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1362 static void bpf_prog_clone_free(struct bpf_prog *fp)
1364 /* aux was stolen by the other clone, so we cannot free
1365 * it from this path! It will be freed eventually by the
1366 * other program on release.
1368 * At this point, we don't need a deferred release since
1369 * clone is guaranteed to not be locked.
1374 __bpf_prog_free(fp);
1377 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1379 /* We have to repoint aux->prog to self, as we don't
1380 * know whether fp here is the clone or the original.
1383 bpf_prog_clone_free(fp_other);
1386 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1388 struct bpf_insn insn_buff[16], aux[2];
1389 struct bpf_prog *clone, *tmp;
1390 int insn_delta, insn_cnt;
1391 struct bpf_insn *insn;
1394 if (!prog->blinding_requested || prog->blinded)
1397 clone = bpf_prog_clone_create(prog, GFP_USER);
1399 return ERR_PTR(-ENOMEM);
1401 insn_cnt = clone->len;
1402 insn = clone->insnsi;
1404 for (i = 0; i < insn_cnt; i++, insn++) {
1405 if (bpf_pseudo_func(insn)) {
1406 /* ld_imm64 with an address of bpf subprog is not
1407 * a user controlled constant. Don't randomize it,
1408 * since it will conflict with jit_subprogs() logic.
1415 /* We temporarily need to hold the original ld64 insn
1416 * so that we can still access the first part in the
1417 * second blinding run.
1419 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1421 memcpy(aux, insn, sizeof(aux));
1423 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1424 clone->aux->verifier_zext);
1428 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1430 /* Patching may have repointed aux->prog during
1431 * realloc from the original one, so we need to
1432 * fix it up here on error.
1434 bpf_jit_prog_release_other(prog, clone);
1439 insn_delta = rewritten - 1;
1441 /* Walk new program and skip insns we just inserted. */
1442 insn = clone->insnsi + i + insn_delta;
1443 insn_cnt += insn_delta;
1450 #endif /* CONFIG_BPF_JIT */
1452 /* Base function for offset calculation. Needs to go into .text section,
1453 * therefore keeping it non-static as well; will also be used by JITs
1454 * anyway later on, so do not let the compiler omit it. This also needs
1455 * to go into kallsyms for correlation from e.g. bpftool, so naming
1458 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1462 EXPORT_SYMBOL_GPL(__bpf_call_base);
1464 /* All UAPI available opcodes. */
1465 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1466 /* 32 bit ALU operations. */ \
1467 /* Register based. */ \
1468 INSN_3(ALU, ADD, X), \
1469 INSN_3(ALU, SUB, X), \
1470 INSN_3(ALU, AND, X), \
1471 INSN_3(ALU, OR, X), \
1472 INSN_3(ALU, LSH, X), \
1473 INSN_3(ALU, RSH, X), \
1474 INSN_3(ALU, XOR, X), \
1475 INSN_3(ALU, MUL, X), \
1476 INSN_3(ALU, MOV, X), \
1477 INSN_3(ALU, ARSH, X), \
1478 INSN_3(ALU, DIV, X), \
1479 INSN_3(ALU, MOD, X), \
1481 INSN_3(ALU, END, TO_BE), \
1482 INSN_3(ALU, END, TO_LE), \
1483 /* Immediate based. */ \
1484 INSN_3(ALU, ADD, K), \
1485 INSN_3(ALU, SUB, K), \
1486 INSN_3(ALU, AND, K), \
1487 INSN_3(ALU, OR, K), \
1488 INSN_3(ALU, LSH, K), \
1489 INSN_3(ALU, RSH, K), \
1490 INSN_3(ALU, XOR, K), \
1491 INSN_3(ALU, MUL, K), \
1492 INSN_3(ALU, MOV, K), \
1493 INSN_3(ALU, ARSH, K), \
1494 INSN_3(ALU, DIV, K), \
1495 INSN_3(ALU, MOD, K), \
1496 /* 64 bit ALU operations. */ \
1497 /* Register based. */ \
1498 INSN_3(ALU64, ADD, X), \
1499 INSN_3(ALU64, SUB, X), \
1500 INSN_3(ALU64, AND, X), \
1501 INSN_3(ALU64, OR, X), \
1502 INSN_3(ALU64, LSH, X), \
1503 INSN_3(ALU64, RSH, X), \
1504 INSN_3(ALU64, XOR, X), \
1505 INSN_3(ALU64, MUL, X), \
1506 INSN_3(ALU64, MOV, X), \
1507 INSN_3(ALU64, ARSH, X), \
1508 INSN_3(ALU64, DIV, X), \
1509 INSN_3(ALU64, MOD, X), \
1510 INSN_2(ALU64, NEG), \
1511 /* Immediate based. */ \
1512 INSN_3(ALU64, ADD, K), \
1513 INSN_3(ALU64, SUB, K), \
1514 INSN_3(ALU64, AND, K), \
1515 INSN_3(ALU64, OR, K), \
1516 INSN_3(ALU64, LSH, K), \
1517 INSN_3(ALU64, RSH, K), \
1518 INSN_3(ALU64, XOR, K), \
1519 INSN_3(ALU64, MUL, K), \
1520 INSN_3(ALU64, MOV, K), \
1521 INSN_3(ALU64, ARSH, K), \
1522 INSN_3(ALU64, DIV, K), \
1523 INSN_3(ALU64, MOD, K), \
1524 /* Call instruction. */ \
1525 INSN_2(JMP, CALL), \
1526 /* Exit instruction. */ \
1527 INSN_2(JMP, EXIT), \
1528 /* 32-bit Jump instructions. */ \
1529 /* Register based. */ \
1530 INSN_3(JMP32, JEQ, X), \
1531 INSN_3(JMP32, JNE, X), \
1532 INSN_3(JMP32, JGT, X), \
1533 INSN_3(JMP32, JLT, X), \
1534 INSN_3(JMP32, JGE, X), \
1535 INSN_3(JMP32, JLE, X), \
1536 INSN_3(JMP32, JSGT, X), \
1537 INSN_3(JMP32, JSLT, X), \
1538 INSN_3(JMP32, JSGE, X), \
1539 INSN_3(JMP32, JSLE, X), \
1540 INSN_3(JMP32, JSET, X), \
1541 /* Immediate based. */ \
1542 INSN_3(JMP32, JEQ, K), \
1543 INSN_3(JMP32, JNE, K), \
1544 INSN_3(JMP32, JGT, K), \
1545 INSN_3(JMP32, JLT, K), \
1546 INSN_3(JMP32, JGE, K), \
1547 INSN_3(JMP32, JLE, K), \
1548 INSN_3(JMP32, JSGT, K), \
1549 INSN_3(JMP32, JSLT, K), \
1550 INSN_3(JMP32, JSGE, K), \
1551 INSN_3(JMP32, JSLE, K), \
1552 INSN_3(JMP32, JSET, K), \
1553 /* Jump instructions. */ \
1554 /* Register based. */ \
1555 INSN_3(JMP, JEQ, X), \
1556 INSN_3(JMP, JNE, X), \
1557 INSN_3(JMP, JGT, X), \
1558 INSN_3(JMP, JLT, X), \
1559 INSN_3(JMP, JGE, X), \
1560 INSN_3(JMP, JLE, X), \
1561 INSN_3(JMP, JSGT, X), \
1562 INSN_3(JMP, JSLT, X), \
1563 INSN_3(JMP, JSGE, X), \
1564 INSN_3(JMP, JSLE, X), \
1565 INSN_3(JMP, JSET, X), \
1566 /* Immediate based. */ \
1567 INSN_3(JMP, JEQ, K), \
1568 INSN_3(JMP, JNE, K), \
1569 INSN_3(JMP, JGT, K), \
1570 INSN_3(JMP, JLT, K), \
1571 INSN_3(JMP, JGE, K), \
1572 INSN_3(JMP, JLE, K), \
1573 INSN_3(JMP, JSGT, K), \
1574 INSN_3(JMP, JSLT, K), \
1575 INSN_3(JMP, JSGE, K), \
1576 INSN_3(JMP, JSLE, K), \
1577 INSN_3(JMP, JSET, K), \
1579 /* Store instructions. */ \
1580 /* Register based. */ \
1581 INSN_3(STX, MEM, B), \
1582 INSN_3(STX, MEM, H), \
1583 INSN_3(STX, MEM, W), \
1584 INSN_3(STX, MEM, DW), \
1585 INSN_3(STX, ATOMIC, W), \
1586 INSN_3(STX, ATOMIC, DW), \
1587 /* Immediate based. */ \
1588 INSN_3(ST, MEM, B), \
1589 INSN_3(ST, MEM, H), \
1590 INSN_3(ST, MEM, W), \
1591 INSN_3(ST, MEM, DW), \
1592 /* Load instructions. */ \
1593 /* Register based. */ \
1594 INSN_3(LDX, MEM, B), \
1595 INSN_3(LDX, MEM, H), \
1596 INSN_3(LDX, MEM, W), \
1597 INSN_3(LDX, MEM, DW), \
1598 /* Immediate based. */ \
1601 bool bpf_opcode_in_insntable(u8 code)
1603 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1604 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1605 static const bool public_insntable[256] = {
1606 [0 ... 255] = false,
1607 /* Now overwrite non-defaults ... */
1608 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1609 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1610 [BPF_LD | BPF_ABS | BPF_B] = true,
1611 [BPF_LD | BPF_ABS | BPF_H] = true,
1612 [BPF_LD | BPF_ABS | BPF_W] = true,
1613 [BPF_LD | BPF_IND | BPF_B] = true,
1614 [BPF_LD | BPF_IND | BPF_H] = true,
1615 [BPF_LD | BPF_IND | BPF_W] = true,
1617 #undef BPF_INSN_3_TBL
1618 #undef BPF_INSN_2_TBL
1619 return public_insntable[code];
1622 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1623 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1625 memset(dst, 0, size);
1630 * ___bpf_prog_run - run eBPF program on a given context
1631 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1632 * @insn: is the array of eBPF instructions
1634 * Decode and execute eBPF instructions.
1636 * Return: whatever value is in %BPF_R0 at program exit
1638 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1640 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1641 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1642 static const void * const jumptable[256] __annotate_jump_table = {
1643 [0 ... 255] = &&default_label,
1644 /* Now overwrite non-defaults ... */
1645 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1646 /* Non-UAPI available opcodes. */
1647 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1648 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1649 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1650 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1651 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1652 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1653 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1655 #undef BPF_INSN_3_LBL
1656 #undef BPF_INSN_2_LBL
1657 u32 tail_call_cnt = 0;
1659 #define CONT ({ insn++; goto select_insn; })
1660 #define CONT_JMP ({ insn++; goto select_insn; })
1663 goto *jumptable[insn->code];
1665 /* Explicitly mask the register-based shift amounts with 63 or 31
1666 * to avoid undefined behavior. Normally this won't affect the
1667 * generated code, for example, in case of native 64 bit archs such
1668 * as x86-64 or arm64, the compiler is optimizing the AND away for
1669 * the interpreter. In case of JITs, each of the JIT backends compiles
1670 * the BPF shift operations to machine instructions which produce
1671 * implementation-defined results in such a case; the resulting
1672 * contents of the register may be arbitrary, but program behaviour
1673 * as a whole remains defined. In other words, in case of JIT backends,
1674 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1677 #define SHT(OPCODE, OP) \
1678 ALU64_##OPCODE##_X: \
1679 DST = DST OP (SRC & 63); \
1682 DST = (u32) DST OP ((u32) SRC & 31); \
1684 ALU64_##OPCODE##_K: \
1688 DST = (u32) DST OP (u32) IMM; \
1691 #define ALU(OPCODE, OP) \
1692 ALU64_##OPCODE##_X: \
1696 DST = (u32) DST OP (u32) SRC; \
1698 ALU64_##OPCODE##_K: \
1702 DST = (u32) DST OP (u32) IMM; \
1733 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1737 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1740 DST = (u64) (u32) (((s32) DST) >> IMM);
1743 (*(s64 *) &DST) >>= (SRC & 63);
1746 (*(s64 *) &DST) >>= IMM;
1749 div64_u64_rem(DST, SRC, &AX);
1754 DST = do_div(AX, (u32) SRC);
1757 div64_u64_rem(DST, IMM, &AX);
1762 DST = do_div(AX, (u32) IMM);
1765 DST = div64_u64(DST, SRC);
1769 do_div(AX, (u32) SRC);
1773 DST = div64_u64(DST, IMM);
1777 do_div(AX, (u32) IMM);
1783 DST = (__force u16) cpu_to_be16(DST);
1786 DST = (__force u32) cpu_to_be32(DST);
1789 DST = (__force u64) cpu_to_be64(DST);
1796 DST = (__force u16) cpu_to_le16(DST);
1799 DST = (__force u32) cpu_to_le32(DST);
1802 DST = (__force u64) cpu_to_le64(DST);
1809 /* Function call scratches BPF_R1-BPF_R5 registers,
1810 * preserves BPF_R6-BPF_R9, and stores return value
1813 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1818 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1821 insn + insn->off + 1);
1825 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1826 struct bpf_array *array = container_of(map, struct bpf_array, map);
1827 struct bpf_prog *prog;
1830 if (unlikely(index >= array->map.max_entries))
1833 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
1838 prog = READ_ONCE(array->ptrs[index]);
1842 /* ARG1 at this point is guaranteed to point to CTX from
1843 * the verifier side due to the fact that the tail call is
1844 * handled like a helper, that is, bpf_tail_call_proto,
1845 * where arg1_type is ARG_PTR_TO_CTX.
1847 insn = prog->insnsi;
1858 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1860 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1861 insn += insn->off; \
1865 JMP32_##OPCODE##_X: \
1866 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1867 insn += insn->off; \
1872 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1873 insn += insn->off; \
1877 JMP32_##OPCODE##_K: \
1878 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1879 insn += insn->off; \
1883 COND_JMP(u, JEQ, ==)
1884 COND_JMP(u, JNE, !=)
1887 COND_JMP(u, JGE, >=)
1888 COND_JMP(u, JLE, <=)
1889 COND_JMP(u, JSET, &)
1890 COND_JMP(s, JSGT, >)
1891 COND_JMP(s, JSLT, <)
1892 COND_JMP(s, JSGE, >=)
1893 COND_JMP(s, JSLE, <=)
1895 /* ST, STX and LDX*/
1897 /* Speculation barrier for mitigating Speculative Store Bypass.
1898 * In case of arm64, we rely on the firmware mitigation as
1899 * controlled via the ssbd kernel parameter. Whenever the
1900 * mitigation is enabled, it works for all of the kernel code
1901 * with no need to provide any additional instructions here.
1902 * In case of x86, we use 'lfence' insn for mitigation. We
1903 * reuse preexisting logic from Spectre v1 mitigation that
1904 * happens to produce the required code on x86 for v4 as well.
1910 #define LDST(SIZEOP, SIZE) \
1912 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1915 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1918 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1920 LDX_PROBE_MEM_##SIZEOP: \
1921 bpf_probe_read_kernel(&DST, sizeof(SIZE), \
1922 (const void *)(long) (SRC + insn->off)); \
1923 DST = *((SIZE *)&DST); \
1932 #define ATOMIC_ALU_OP(BOP, KOP) \
1934 if (BPF_SIZE(insn->code) == BPF_W) \
1935 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
1936 (DST + insn->off)); \
1938 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
1939 (DST + insn->off)); \
1941 case BOP | BPF_FETCH: \
1942 if (BPF_SIZE(insn->code) == BPF_W) \
1943 SRC = (u32) atomic_fetch_##KOP( \
1945 (atomic_t *)(unsigned long) (DST + insn->off)); \
1947 SRC = (u64) atomic64_fetch_##KOP( \
1949 (atomic64_t *)(unsigned long) (DST + insn->off)); \
1955 ATOMIC_ALU_OP(BPF_ADD, add)
1956 ATOMIC_ALU_OP(BPF_AND, and)
1957 ATOMIC_ALU_OP(BPF_OR, or)
1958 ATOMIC_ALU_OP(BPF_XOR, xor)
1959 #undef ATOMIC_ALU_OP
1962 if (BPF_SIZE(insn->code) == BPF_W)
1963 SRC = (u32) atomic_xchg(
1964 (atomic_t *)(unsigned long) (DST + insn->off),
1967 SRC = (u64) atomic64_xchg(
1968 (atomic64_t *)(unsigned long) (DST + insn->off),
1972 if (BPF_SIZE(insn->code) == BPF_W)
1973 BPF_R0 = (u32) atomic_cmpxchg(
1974 (atomic_t *)(unsigned long) (DST + insn->off),
1975 (u32) BPF_R0, (u32) SRC);
1977 BPF_R0 = (u64) atomic64_cmpxchg(
1978 (atomic64_t *)(unsigned long) (DST + insn->off),
1979 (u64) BPF_R0, (u64) SRC);
1988 /* If we ever reach this, we have a bug somewhere. Die hard here
1989 * instead of just returning 0; we could be somewhere in a subprog,
1990 * so execution could continue otherwise which we do /not/ want.
1992 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
1994 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
1995 insn->code, insn->imm);
2000 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2001 #define DEFINE_BPF_PROG_RUN(stack_size) \
2002 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2004 u64 stack[stack_size / sizeof(u64)]; \
2005 u64 regs[MAX_BPF_EXT_REG]; \
2007 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2008 ARG1 = (u64) (unsigned long) ctx; \
2009 return ___bpf_prog_run(regs, insn); \
2012 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2013 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2014 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2015 const struct bpf_insn *insn) \
2017 u64 stack[stack_size / sizeof(u64)]; \
2018 u64 regs[MAX_BPF_EXT_REG]; \
2020 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2026 return ___bpf_prog_run(regs, insn); \
2029 #define EVAL1(FN, X) FN(X)
2030 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2031 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2032 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2033 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2034 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2036 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2037 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2038 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2040 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2041 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2042 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2044 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2046 static unsigned int (*interpreters[])(const void *ctx,
2047 const struct bpf_insn *insn) = {
2048 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2049 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2050 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2052 #undef PROG_NAME_LIST
2053 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2054 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2055 const struct bpf_insn *insn) = {
2056 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2057 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2058 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2060 #undef PROG_NAME_LIST
2062 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2064 stack_depth = max_t(u32, stack_depth, 1);
2065 insn->off = (s16) insn->imm;
2066 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2067 __bpf_call_base_args;
2068 insn->code = BPF_JMP | BPF_CALL_ARGS;
2072 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2073 const struct bpf_insn *insn)
2075 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2076 * is not working properly, so warn about it!
2083 bool bpf_prog_map_compatible(struct bpf_map *map,
2084 const struct bpf_prog *fp)
2088 if (fp->kprobe_override)
2091 spin_lock(&map->owner.lock);
2092 if (!map->owner.type) {
2093 /* There's no owner yet where we could check for
2096 map->owner.type = fp->type;
2097 map->owner.jited = fp->jited;
2098 map->owner.xdp_has_frags = fp->aux->xdp_has_frags;
2101 ret = map->owner.type == fp->type &&
2102 map->owner.jited == fp->jited &&
2103 map->owner.xdp_has_frags == fp->aux->xdp_has_frags;
2105 spin_unlock(&map->owner.lock);
2110 static int bpf_check_tail_call(const struct bpf_prog *fp)
2112 struct bpf_prog_aux *aux = fp->aux;
2115 mutex_lock(&aux->used_maps_mutex);
2116 for (i = 0; i < aux->used_map_cnt; i++) {
2117 struct bpf_map *map = aux->used_maps[i];
2119 if (!map_type_contains_progs(map))
2122 if (!bpf_prog_map_compatible(map, fp)) {
2129 mutex_unlock(&aux->used_maps_mutex);
2133 static void bpf_prog_select_func(struct bpf_prog *fp)
2135 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2136 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2138 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2140 fp->bpf_func = __bpf_prog_ret0_warn;
2145 * bpf_prog_select_runtime - select exec runtime for BPF program
2146 * @fp: bpf_prog populated with BPF program
2147 * @err: pointer to error variable
2149 * Try to JIT eBPF program, if JIT is not available, use interpreter.
2150 * The BPF program will be executed via bpf_prog_run() function.
2152 * Return: the &fp argument along with &err set to 0 for success or
2153 * a negative errno code on failure
2155 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2157 /* In case of BPF to BPF calls, verifier did all the prep
2158 * work with regards to JITing, etc.
2160 bool jit_needed = false;
2165 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2166 bpf_prog_has_kfunc_call(fp))
2169 bpf_prog_select_func(fp);
2171 /* eBPF JITs can rewrite the program in case constant
2172 * blinding is active. However, in case of error during
2173 * blinding, bpf_int_jit_compile() must always return a
2174 * valid program, which in this case would simply not
2175 * be JITed, but falls back to the interpreter.
2177 if (!bpf_prog_is_dev_bound(fp->aux)) {
2178 *err = bpf_prog_alloc_jited_linfo(fp);
2182 fp = bpf_int_jit_compile(fp);
2183 bpf_prog_jit_attempt_done(fp);
2184 if (!fp->jited && jit_needed) {
2189 *err = bpf_prog_offload_compile(fp);
2195 bpf_prog_lock_ro(fp);
2197 /* The tail call compatibility check can only be done at
2198 * this late stage as we need to determine, if we deal
2199 * with JITed or non JITed program concatenations and not
2200 * all eBPF JITs might immediately support all features.
2202 *err = bpf_check_tail_call(fp);
2206 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2208 static unsigned int __bpf_prog_ret1(const void *ctx,
2209 const struct bpf_insn *insn)
2214 static struct bpf_prog_dummy {
2215 struct bpf_prog prog;
2216 } dummy_bpf_prog = {
2218 .bpf_func = __bpf_prog_ret1,
2222 struct bpf_empty_prog_array bpf_empty_prog_array = {
2225 EXPORT_SYMBOL(bpf_empty_prog_array);
2227 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2230 return kzalloc(sizeof(struct bpf_prog_array) +
2231 sizeof(struct bpf_prog_array_item) *
2235 return &bpf_empty_prog_array.hdr;
2238 void bpf_prog_array_free(struct bpf_prog_array *progs)
2240 if (!progs || progs == &bpf_empty_prog_array.hdr)
2242 kfree_rcu(progs, rcu);
2245 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2247 struct bpf_prog_array *progs;
2249 progs = container_of(rcu, struct bpf_prog_array, rcu);
2250 kfree_rcu(progs, rcu);
2253 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2255 if (!progs || progs == &bpf_empty_prog_array.hdr)
2257 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2260 int bpf_prog_array_length(struct bpf_prog_array *array)
2262 struct bpf_prog_array_item *item;
2265 for (item = array->items; item->prog; item++)
2266 if (item->prog != &dummy_bpf_prog.prog)
2271 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2273 struct bpf_prog_array_item *item;
2275 for (item = array->items; item->prog; item++)
2276 if (item->prog != &dummy_bpf_prog.prog)
2281 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2285 struct bpf_prog_array_item *item;
2288 for (item = array->items; item->prog; item++) {
2289 if (item->prog == &dummy_bpf_prog.prog)
2291 prog_ids[i] = item->prog->aux->id;
2292 if (++i == request_cnt) {
2298 return !!(item->prog);
2301 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2302 __u32 __user *prog_ids, u32 cnt)
2304 unsigned long err = 0;
2308 /* users of this function are doing:
2309 * cnt = bpf_prog_array_length();
2311 * bpf_prog_array_copy_to_user(..., cnt);
2312 * so below kcalloc doesn't need extra cnt > 0 check.
2314 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2317 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2318 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2327 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2328 struct bpf_prog *old_prog)
2330 struct bpf_prog_array_item *item;
2332 for (item = array->items; item->prog; item++)
2333 if (item->prog == old_prog) {
2334 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2340 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2341 * index into the program array with
2342 * a dummy no-op program.
2343 * @array: a bpf_prog_array
2344 * @index: the index of the program to replace
2346 * Skips over dummy programs, by not counting them, when calculating
2347 * the position of the program to replace.
2351 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2352 * * -ENOENT - Index out of range
2354 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2356 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2360 * bpf_prog_array_update_at() - Updates the program at the given index
2361 * into the program array.
2362 * @array: a bpf_prog_array
2363 * @index: the index of the program to update
2364 * @prog: the program to insert into the array
2366 * Skips over dummy programs, by not counting them, when calculating
2367 * the position of the program to update.
2371 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2372 * * -ENOENT - Index out of range
2374 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2375 struct bpf_prog *prog)
2377 struct bpf_prog_array_item *item;
2379 if (unlikely(index < 0))
2382 for (item = array->items; item->prog; item++) {
2383 if (item->prog == &dummy_bpf_prog.prog)
2386 WRITE_ONCE(item->prog, prog);
2394 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2395 struct bpf_prog *exclude_prog,
2396 struct bpf_prog *include_prog,
2398 struct bpf_prog_array **new_array)
2400 int new_prog_cnt, carry_prog_cnt = 0;
2401 struct bpf_prog_array_item *existing, *new;
2402 struct bpf_prog_array *array;
2403 bool found_exclude = false;
2405 /* Figure out how many existing progs we need to carry over to
2409 existing = old_array->items;
2410 for (; existing->prog; existing++) {
2411 if (existing->prog == exclude_prog) {
2412 found_exclude = true;
2415 if (existing->prog != &dummy_bpf_prog.prog)
2417 if (existing->prog == include_prog)
2422 if (exclude_prog && !found_exclude)
2425 /* How many progs (not NULL) will be in the new array? */
2426 new_prog_cnt = carry_prog_cnt;
2430 /* Do we have any prog (not NULL) in the new array? */
2431 if (!new_prog_cnt) {
2436 /* +1 as the end of prog_array is marked with NULL */
2437 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2442 /* Fill in the new prog array */
2443 if (carry_prog_cnt) {
2444 existing = old_array->items;
2445 for (; existing->prog; existing++) {
2446 if (existing->prog == exclude_prog ||
2447 existing->prog == &dummy_bpf_prog.prog)
2450 new->prog = existing->prog;
2451 new->bpf_cookie = existing->bpf_cookie;
2456 new->prog = include_prog;
2457 new->bpf_cookie = bpf_cookie;
2465 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2466 u32 *prog_ids, u32 request_cnt,
2472 cnt = bpf_prog_array_length(array);
2476 /* return early if user requested only program count or nothing to copy */
2477 if (!request_cnt || !cnt)
2480 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2481 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2485 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2486 struct bpf_map **used_maps, u32 len)
2488 struct bpf_map *map;
2491 for (i = 0; i < len; i++) {
2493 if (map->ops->map_poke_untrack)
2494 map->ops->map_poke_untrack(map, aux);
2499 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2501 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2502 kfree(aux->used_maps);
2505 void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2506 struct btf_mod_pair *used_btfs, u32 len)
2508 #ifdef CONFIG_BPF_SYSCALL
2509 struct btf_mod_pair *btf_mod;
2512 for (i = 0; i < len; i++) {
2513 btf_mod = &used_btfs[i];
2514 if (btf_mod->module)
2515 module_put(btf_mod->module);
2516 btf_put(btf_mod->btf);
2521 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2523 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2524 kfree(aux->used_btfs);
2527 static void bpf_prog_free_deferred(struct work_struct *work)
2529 struct bpf_prog_aux *aux;
2532 aux = container_of(work, struct bpf_prog_aux, work);
2533 #ifdef CONFIG_BPF_SYSCALL
2534 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2536 #ifdef CONFIG_CGROUP_BPF
2537 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2538 bpf_cgroup_atype_put(aux->cgroup_atype);
2540 bpf_free_used_maps(aux);
2541 bpf_free_used_btfs(aux);
2542 if (bpf_prog_is_dev_bound(aux))
2543 bpf_prog_offload_destroy(aux->prog);
2544 #ifdef CONFIG_PERF_EVENTS
2545 if (aux->prog->has_callchain_buf)
2546 put_callchain_buffers();
2548 if (aux->dst_trampoline)
2549 bpf_trampoline_put(aux->dst_trampoline);
2550 for (i = 0; i < aux->func_cnt; i++) {
2551 /* We can just unlink the subprog poke descriptor table as
2552 * it was originally linked to the main program and is also
2553 * released along with it.
2555 aux->func[i]->aux->poke_tab = NULL;
2556 bpf_jit_free(aux->func[i]);
2558 if (aux->func_cnt) {
2560 bpf_prog_unlock_free(aux->prog);
2562 bpf_jit_free(aux->prog);
2566 void bpf_prog_free(struct bpf_prog *fp)
2568 struct bpf_prog_aux *aux = fp->aux;
2571 bpf_prog_put(aux->dst_prog);
2572 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2573 schedule_work(&aux->work);
2575 EXPORT_SYMBOL_GPL(bpf_prog_free);
2577 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2578 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2580 void bpf_user_rnd_init_once(void)
2582 prandom_init_once(&bpf_user_rnd_state);
2585 BPF_CALL_0(bpf_user_rnd_u32)
2587 /* Should someone ever have the rather unwise idea to use some
2588 * of the registers passed into this function, then note that
2589 * this function is called from native eBPF and classic-to-eBPF
2590 * transformations. Register assignments from both sides are
2591 * different, f.e. classic always sets fn(ctx, A, X) here.
2593 struct rnd_state *state;
2596 state = &get_cpu_var(bpf_user_rnd_state);
2597 res = prandom_u32_state(state);
2598 put_cpu_var(bpf_user_rnd_state);
2603 BPF_CALL_0(bpf_get_raw_cpu_id)
2605 return raw_smp_processor_id();
2608 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2609 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2610 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2611 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2612 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2613 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2614 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2615 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2616 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2617 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2618 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2620 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2621 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2622 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2623 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2624 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2625 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2627 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2628 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2629 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2630 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2631 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2632 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2633 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2634 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2635 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2636 const struct bpf_func_proto bpf_set_retval_proto __weak;
2637 const struct bpf_func_proto bpf_get_retval_proto __weak;
2639 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2644 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2650 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2651 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2655 EXPORT_SYMBOL_GPL(bpf_event_output);
2657 /* Always built-in helper functions. */
2658 const struct bpf_func_proto bpf_tail_call_proto = {
2661 .ret_type = RET_VOID,
2662 .arg1_type = ARG_PTR_TO_CTX,
2663 .arg2_type = ARG_CONST_MAP_PTR,
2664 .arg3_type = ARG_ANYTHING,
2667 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2668 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2669 * eBPF and implicitly also cBPF can get JITed!
2671 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2676 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2677 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2679 void __weak bpf_jit_compile(struct bpf_prog *prog)
2683 bool __weak bpf_helper_changes_pkt_data(void *func)
2688 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2689 * analysis code and wants explicit zero extension inserted by verifier.
2690 * Otherwise, return FALSE.
2692 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2693 * you don't override this. JITs that don't want these extra insns can detect
2694 * them using insn_is_zext.
2696 bool __weak bpf_jit_needs_zext(void)
2701 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2702 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2707 bool __weak bpf_jit_supports_kfunc_call(void)
2712 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2713 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2715 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2721 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2722 void *addr1, void *addr2)
2727 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
2729 return ERR_PTR(-ENOTSUPP);
2732 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
2737 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2738 EXPORT_SYMBOL(bpf_stats_enabled_key);
2740 /* All definitions of tracepoints related to BPF. */
2741 #define CREATE_TRACE_POINTS
2742 #include <linux/bpf_trace.h>
2744 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2745 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);