1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Linux Socket Filter - Kernel level socket filtering
5 * Based on the design of the Berkeley Packet Filter. The new
6 * internal format has been designed by PLUMgrid:
8 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
16 * Andi Kleen - Fix a few bad bugs and races.
17 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
20 #include <uapi/linux/btf.h>
21 #include <linux/filter.h>
22 #include <linux/skbuff.h>
23 #include <linux/vmalloc.h>
24 #include <linux/random.h>
25 #include <linux/moduleloader.h>
26 #include <linux/bpf.h>
27 #include <linux/btf.h>
28 #include <linux/objtool.h>
29 #include <linux/rbtree_latch.h>
30 #include <linux/kallsyms.h>
31 #include <linux/rcupdate.h>
32 #include <linux/perf_event.h>
33 #include <linux/extable.h>
34 #include <linux/log2.h>
35 #include <linux/bpf_verifier.h>
36 #include <linux/nodemask.h>
37 #include <linux/nospec.h>
38 #include <linux/bpf_mem_alloc.h>
39 #include <linux/memcontrol.h>
41 #include <asm/barrier.h>
42 #include <asm/unaligned.h>
45 #define BPF_R0 regs[BPF_REG_0]
46 #define BPF_R1 regs[BPF_REG_1]
47 #define BPF_R2 regs[BPF_REG_2]
48 #define BPF_R3 regs[BPF_REG_3]
49 #define BPF_R4 regs[BPF_REG_4]
50 #define BPF_R5 regs[BPF_REG_5]
51 #define BPF_R6 regs[BPF_REG_6]
52 #define BPF_R7 regs[BPF_REG_7]
53 #define BPF_R8 regs[BPF_REG_8]
54 #define BPF_R9 regs[BPF_REG_9]
55 #define BPF_R10 regs[BPF_REG_10]
58 #define DST regs[insn->dst_reg]
59 #define SRC regs[insn->src_reg]
60 #define FP regs[BPF_REG_FP]
61 #define AX regs[BPF_REG_AX]
62 #define ARG1 regs[BPF_REG_ARG1]
63 #define CTX regs[BPF_REG_CTX]
66 struct bpf_mem_alloc bpf_global_ma;
67 bool bpf_global_ma_set;
69 /* No hurry in this branch
71 * Exported for the bpf jit load helper.
73 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
77 if (k >= SKF_NET_OFF) {
78 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
79 } else if (k >= SKF_LL_OFF) {
80 if (unlikely(!skb_mac_header_was_set(skb)))
82 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
84 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
90 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
92 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
93 struct bpf_prog_aux *aux;
96 size = round_up(size, PAGE_SIZE);
97 fp = __vmalloc(size, gfp_flags);
101 aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
106 fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
113 fp->pages = size / PAGE_SIZE;
116 fp->jit_requested = ebpf_jit_enabled();
117 fp->blinding_requested = bpf_jit_blinding_enabled(fp);
118 #ifdef CONFIG_CGROUP_BPF
119 aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
122 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
123 mutex_init(&fp->aux->used_maps_mutex);
124 mutex_init(&fp->aux->dst_mutex);
129 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
131 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
132 struct bpf_prog *prog;
135 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
139 prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
141 free_percpu(prog->active);
147 for_each_possible_cpu(cpu) {
148 struct bpf_prog_stats *pstats;
150 pstats = per_cpu_ptr(prog->stats, cpu);
151 u64_stats_init(&pstats->syncp);
155 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
157 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
159 if (!prog->aux->nr_linfo || !prog->jit_requested)
162 prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
163 sizeof(*prog->aux->jited_linfo),
164 bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN));
165 if (!prog->aux->jited_linfo)
171 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
173 if (prog->aux->jited_linfo &&
174 (!prog->jited || !prog->aux->jited_linfo[0])) {
175 kvfree(prog->aux->jited_linfo);
176 prog->aux->jited_linfo = NULL;
179 kfree(prog->aux->kfunc_tab);
180 prog->aux->kfunc_tab = NULL;
183 /* The jit engine is responsible to provide an array
184 * for insn_off to the jited_off mapping (insn_to_jit_off).
186 * The idx to this array is the insn_off. Hence, the insn_off
187 * here is relative to the prog itself instead of the main prog.
188 * This array has one entry for each xlated bpf insn.
190 * jited_off is the byte off to the end of the jited insn.
194 * The first bpf insn off of the prog. The insn off
195 * here is relative to the main prog.
196 * e.g. if prog is a subprog, insn_start > 0
198 * The prog's idx to prog->aux->linfo and jited_linfo
200 * jited_linfo[linfo_idx] = prog->bpf_func
204 * jited_linfo[i] = prog->bpf_func +
205 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
207 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
208 const u32 *insn_to_jit_off)
210 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
211 const struct bpf_line_info *linfo;
214 if (!prog->aux->jited_linfo)
215 /* Userspace did not provide linfo */
218 linfo_idx = prog->aux->linfo_idx;
219 linfo = &prog->aux->linfo[linfo_idx];
220 insn_start = linfo[0].insn_off;
221 insn_end = insn_start + prog->len;
223 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
224 jited_linfo[0] = prog->bpf_func;
226 nr_linfo = prog->aux->nr_linfo - linfo_idx;
228 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
229 /* The verifier ensures that linfo[i].insn_off is
230 * strictly increasing
232 jited_linfo[i] = prog->bpf_func +
233 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
236 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
237 gfp_t gfp_extra_flags)
239 gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
243 size = round_up(size, PAGE_SIZE);
244 pages = size / PAGE_SIZE;
245 if (pages <= fp_old->pages)
248 fp = __vmalloc(size, gfp_flags);
250 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
254 /* We keep fp->aux from fp_old around in the new
255 * reallocated structure.
258 fp_old->stats = NULL;
259 fp_old->active = NULL;
260 __bpf_prog_free(fp_old);
266 void __bpf_prog_free(struct bpf_prog *fp)
269 mutex_destroy(&fp->aux->used_maps_mutex);
270 mutex_destroy(&fp->aux->dst_mutex);
271 kfree(fp->aux->poke_tab);
274 free_percpu(fp->stats);
275 free_percpu(fp->active);
279 int bpf_prog_calc_tag(struct bpf_prog *fp)
281 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
282 u32 raw_size = bpf_prog_tag_scratch_size(fp);
283 u32 digest[SHA1_DIGEST_WORDS];
284 u32 ws[SHA1_WORKSPACE_WORDS];
285 u32 i, bsize, psize, blocks;
286 struct bpf_insn *dst;
292 raw = vmalloc(raw_size);
297 memset(ws, 0, sizeof(ws));
299 /* We need to take out the map fd for the digest calculation
300 * since they are unstable from user space side.
303 for (i = 0, was_ld_map = false; i < fp->len; i++) {
304 dst[i] = fp->insnsi[i];
306 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
307 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
308 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
311 } else if (was_ld_map &&
313 dst[i].dst_reg == 0 &&
314 dst[i].src_reg == 0 &&
323 psize = bpf_prog_insn_size(fp);
324 memset(&raw[psize], 0, raw_size - psize);
327 bsize = round_up(psize, SHA1_BLOCK_SIZE);
328 blocks = bsize / SHA1_BLOCK_SIZE;
330 if (bsize - psize >= sizeof(__be64)) {
331 bits = (__be64 *)(todo + bsize - sizeof(__be64));
333 bits = (__be64 *)(todo + bsize + bits_offset);
336 *bits = cpu_to_be64((psize - 1) << 3);
339 sha1_transform(digest, todo, ws);
340 todo += SHA1_BLOCK_SIZE;
343 result = (__force __be32 *)digest;
344 for (i = 0; i < SHA1_DIGEST_WORDS; i++)
345 result[i] = cpu_to_be32(digest[i]);
346 memcpy(fp->tag, result, sizeof(fp->tag));
352 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
353 s32 end_new, s32 curr, const bool probe_pass)
355 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
356 s32 delta = end_new - end_old;
359 if (curr < pos && curr + imm + 1 >= end_old)
361 else if (curr >= end_new && curr + imm + 1 < end_new)
363 if (imm < imm_min || imm > imm_max)
370 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
371 s32 end_new, s32 curr, const bool probe_pass)
373 const s32 off_min = S16_MIN, off_max = S16_MAX;
374 s32 delta = end_new - end_old;
377 if (curr < pos && curr + off + 1 >= end_old)
379 else if (curr >= end_new && curr + off + 1 < end_new)
381 if (off < off_min || off > off_max)
388 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
389 s32 end_new, const bool probe_pass)
391 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
392 struct bpf_insn *insn = prog->insnsi;
395 for (i = 0; i < insn_cnt; i++, insn++) {
398 /* In the probing pass we still operate on the original,
399 * unpatched image in order to check overflows before we
400 * do any other adjustments. Therefore skip the patchlet.
402 if (probe_pass && i == pos) {
404 insn = prog->insnsi + end_old;
406 if (bpf_pseudo_func(insn)) {
407 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
408 end_new, i, probe_pass);
414 if ((BPF_CLASS(code) != BPF_JMP &&
415 BPF_CLASS(code) != BPF_JMP32) ||
416 BPF_OP(code) == BPF_EXIT)
418 /* Adjust offset of jmps if we cross patch boundaries. */
419 if (BPF_OP(code) == BPF_CALL) {
420 if (insn->src_reg != BPF_PSEUDO_CALL)
422 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
423 end_new, i, probe_pass);
425 ret = bpf_adj_delta_to_off(insn, pos, end_old,
426 end_new, i, probe_pass);
435 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
437 struct bpf_line_info *linfo;
440 nr_linfo = prog->aux->nr_linfo;
441 if (!nr_linfo || !delta)
444 linfo = prog->aux->linfo;
446 for (i = 0; i < nr_linfo; i++)
447 if (off < linfo[i].insn_off)
450 /* Push all off < linfo[i].insn_off by delta */
451 for (; i < nr_linfo; i++)
452 linfo[i].insn_off += delta;
455 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
456 const struct bpf_insn *patch, u32 len)
458 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
459 const u32 cnt_max = S16_MAX;
460 struct bpf_prog *prog_adj;
463 /* Since our patchlet doesn't expand the image, we're done. */
464 if (insn_delta == 0) {
465 memcpy(prog->insnsi + off, patch, sizeof(*patch));
469 insn_adj_cnt = prog->len + insn_delta;
471 /* Reject anything that would potentially let the insn->off
472 * target overflow when we have excessive program expansions.
473 * We need to probe here before we do any reallocation where
474 * we afterwards may not fail anymore.
476 if (insn_adj_cnt > cnt_max &&
477 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
480 /* Several new instructions need to be inserted. Make room
481 * for them. Likely, there's no need for a new allocation as
482 * last page could have large enough tailroom.
484 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
487 return ERR_PTR(-ENOMEM);
489 prog_adj->len = insn_adj_cnt;
491 /* Patching happens in 3 steps:
493 * 1) Move over tail of insnsi from next instruction onwards,
494 * so we can patch the single target insn with one or more
495 * new ones (patching is always from 1 to n insns, n > 0).
496 * 2) Inject new instructions at the target location.
497 * 3) Adjust branch offsets if necessary.
499 insn_rest = insn_adj_cnt - off - len;
501 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
502 sizeof(*patch) * insn_rest);
503 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
505 /* We are guaranteed to not fail at this point, otherwise
506 * the ship has sailed to reverse to the original state. An
507 * overflow cannot happen at this point.
509 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
511 bpf_adj_linfo(prog_adj, off, insn_delta);
516 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
518 /* Branch offsets can't overflow when program is shrinking, no need
519 * to call bpf_adj_branches(..., true) here
521 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
522 sizeof(struct bpf_insn) * (prog->len - off - cnt));
525 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
528 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
532 for (i = 0; i < fp->aux->func_cnt; i++)
533 bpf_prog_kallsyms_del(fp->aux->func[i]);
536 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
538 bpf_prog_kallsyms_del_subprogs(fp);
539 bpf_prog_kallsyms_del(fp);
542 #ifdef CONFIG_BPF_JIT
543 /* All BPF JIT sysctl knobs here. */
544 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
545 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
546 int bpf_jit_harden __read_mostly;
547 long bpf_jit_limit __read_mostly;
548 long bpf_jit_limit_max __read_mostly;
551 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
553 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
555 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
556 prog->aux->ksym.end = prog->aux->ksym.start + prog->jited_len;
560 bpf_prog_ksym_set_name(struct bpf_prog *prog)
562 char *sym = prog->aux->ksym.name;
563 const char *end = sym + KSYM_NAME_LEN;
564 const struct btf_type *type;
565 const char *func_name;
567 BUILD_BUG_ON(sizeof("bpf_prog_") +
568 sizeof(prog->tag) * 2 +
569 /* name has been null terminated.
570 * We should need +1 for the '_' preceding
571 * the name. However, the null character
572 * is double counted between the name and the
573 * sizeof("bpf_prog_") above, so we omit
576 sizeof(prog->aux->name) > KSYM_NAME_LEN);
578 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
579 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
581 /* prog->aux->name will be ignored if full btf name is available */
582 if (prog->aux->func_info_cnt) {
583 type = btf_type_by_id(prog->aux->btf,
584 prog->aux->func_info[prog->aux->func_idx].type_id);
585 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
586 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
590 if (prog->aux->name[0])
591 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
596 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
598 return container_of(n, struct bpf_ksym, tnode)->start;
601 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
602 struct latch_tree_node *b)
604 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
607 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
609 unsigned long val = (unsigned long)key;
610 const struct bpf_ksym *ksym;
612 ksym = container_of(n, struct bpf_ksym, tnode);
614 if (val < ksym->start)
616 if (val >= ksym->end)
622 static const struct latch_tree_ops bpf_tree_ops = {
623 .less = bpf_tree_less,
624 .comp = bpf_tree_comp,
627 static DEFINE_SPINLOCK(bpf_lock);
628 static LIST_HEAD(bpf_kallsyms);
629 static struct latch_tree_root bpf_tree __cacheline_aligned;
631 void bpf_ksym_add(struct bpf_ksym *ksym)
633 spin_lock_bh(&bpf_lock);
634 WARN_ON_ONCE(!list_empty(&ksym->lnode));
635 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
636 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
637 spin_unlock_bh(&bpf_lock);
640 static void __bpf_ksym_del(struct bpf_ksym *ksym)
642 if (list_empty(&ksym->lnode))
645 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
646 list_del_rcu(&ksym->lnode);
649 void bpf_ksym_del(struct bpf_ksym *ksym)
651 spin_lock_bh(&bpf_lock);
652 __bpf_ksym_del(ksym);
653 spin_unlock_bh(&bpf_lock);
656 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
658 return fp->jited && !bpf_prog_was_classic(fp);
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 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size)
836 memset(area, 0, size);
839 #define BPF_PROG_SIZE_TO_NBITS(size) (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
841 static DEFINE_MUTEX(pack_mutex);
842 static LIST_HEAD(pack_list);
844 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
845 * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
848 #define BPF_PROG_PACK_SIZE (PMD_SIZE * num_possible_nodes())
850 #define BPF_PROG_PACK_SIZE PAGE_SIZE
853 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
855 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
857 struct bpf_prog_pack *pack;
859 pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
863 pack->ptr = module_alloc(BPF_PROG_PACK_SIZE);
868 bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
869 bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
870 list_add_tail(&pack->list, &pack_list);
872 set_vm_flush_reset_perms(pack->ptr);
873 set_memory_rox((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
877 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
879 unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
880 struct bpf_prog_pack *pack;
884 mutex_lock(&pack_mutex);
885 if (size > BPF_PROG_PACK_SIZE) {
886 size = round_up(size, PAGE_SIZE);
887 ptr = module_alloc(size);
889 bpf_fill_ill_insns(ptr, size);
890 set_vm_flush_reset_perms(ptr);
891 set_memory_rox((unsigned long)ptr, size / PAGE_SIZE);
895 list_for_each_entry(pack, &pack_list, list) {
896 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
898 if (pos < BPF_PROG_CHUNK_COUNT)
899 goto found_free_area;
902 pack = alloc_new_pack(bpf_fill_ill_insns);
909 bitmap_set(pack->bitmap, pos, nbits);
910 ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
913 mutex_unlock(&pack_mutex);
917 void bpf_prog_pack_free(struct bpf_binary_header *hdr)
919 struct bpf_prog_pack *pack = NULL, *tmp;
923 mutex_lock(&pack_mutex);
924 if (hdr->size > BPF_PROG_PACK_SIZE) {
929 list_for_each_entry(tmp, &pack_list, list) {
930 if ((void *)hdr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > (void *)hdr) {
936 if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
939 nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size);
940 pos = ((unsigned long)hdr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
942 WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size),
943 "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
945 bitmap_clear(pack->bitmap, pos, nbits);
946 if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
947 BPF_PROG_CHUNK_COUNT, 0) == 0) {
948 list_del(&pack->list);
949 module_memfree(pack->ptr);
953 mutex_unlock(&pack_mutex);
956 static atomic_long_t bpf_jit_current;
958 /* Can be overridden by an arch's JIT compiler if it has a custom,
959 * dedicated BPF backend memory area, or if neither of the two
962 u64 __weak bpf_jit_alloc_exec_limit(void)
964 #if defined(MODULES_VADDR)
965 return MODULES_END - MODULES_VADDR;
967 return VMALLOC_END - VMALLOC_START;
971 static int __init bpf_jit_charge_init(void)
973 /* Only used as heuristic here to derive limit. */
974 bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
975 bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 2,
976 PAGE_SIZE), LONG_MAX);
979 pure_initcall(bpf_jit_charge_init);
981 int bpf_jit_charge_modmem(u32 size)
983 if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
984 if (!bpf_capable()) {
985 atomic_long_sub(size, &bpf_jit_current);
993 void bpf_jit_uncharge_modmem(u32 size)
995 atomic_long_sub(size, &bpf_jit_current);
998 void *__weak bpf_jit_alloc_exec(unsigned long size)
1000 return module_alloc(size);
1003 void __weak bpf_jit_free_exec(void *addr)
1005 module_memfree(addr);
1008 struct bpf_binary_header *
1009 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1010 unsigned int alignment,
1011 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1013 struct bpf_binary_header *hdr;
1014 u32 size, hole, start;
1016 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1017 alignment > BPF_IMAGE_ALIGNMENT);
1019 /* Most of BPF filters are really small, but if some of them
1020 * fill a page, allow at least 128 extra bytes to insert a
1021 * random section of illegal instructions.
1023 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1025 if (bpf_jit_charge_modmem(size))
1027 hdr = bpf_jit_alloc_exec(size);
1029 bpf_jit_uncharge_modmem(size);
1033 /* Fill space with illegal/arch-dep instructions. */
1034 bpf_fill_ill_insns(hdr, size);
1037 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1038 PAGE_SIZE - sizeof(*hdr));
1039 start = get_random_u32_below(hole) & ~(alignment - 1);
1041 /* Leave a random number of instructions before BPF code. */
1042 *image_ptr = &hdr->image[start];
1047 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1049 u32 size = hdr->size;
1051 bpf_jit_free_exec(hdr);
1052 bpf_jit_uncharge_modmem(size);
1055 /* Allocate jit binary from bpf_prog_pack allocator.
1056 * Since the allocated memory is RO+X, the JIT engine cannot write directly
1057 * to the memory. To solve this problem, a RW buffer is also allocated at
1058 * as the same time. The JIT engine should calculate offsets based on the
1059 * RO memory address, but write JITed program to the RW buffer. Once the
1060 * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1061 * the JITed program to the RO memory.
1063 struct bpf_binary_header *
1064 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1065 unsigned int alignment,
1066 struct bpf_binary_header **rw_header,
1068 bpf_jit_fill_hole_t bpf_fill_ill_insns)
1070 struct bpf_binary_header *ro_header;
1071 u32 size, hole, start;
1073 WARN_ON_ONCE(!is_power_of_2(alignment) ||
1074 alignment > BPF_IMAGE_ALIGNMENT);
1076 /* add 16 bytes for a random section of illegal instructions */
1077 size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1079 if (bpf_jit_charge_modmem(size))
1081 ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1083 bpf_jit_uncharge_modmem(size);
1087 *rw_header = kvmalloc(size, GFP_KERNEL);
1089 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size));
1090 bpf_prog_pack_free(ro_header);
1091 bpf_jit_uncharge_modmem(size);
1095 /* Fill space with illegal/arch-dep instructions. */
1096 bpf_fill_ill_insns(*rw_header, size);
1097 (*rw_header)->size = size;
1099 hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1100 BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1101 start = get_random_u32_below(hole) & ~(alignment - 1);
1103 *image_ptr = &ro_header->image[start];
1104 *rw_image = &(*rw_header)->image[start];
1109 /* Copy JITed text from rw_header to its final location, the ro_header. */
1110 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog,
1111 struct bpf_binary_header *ro_header,
1112 struct bpf_binary_header *rw_header)
1116 ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1121 bpf_prog_pack_free(ro_header);
1122 return PTR_ERR(ptr);
1127 /* bpf_jit_binary_pack_free is called in two different scenarios:
1128 * 1) when the program is freed after;
1129 * 2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1130 * For case 2), we need to free both the RO memory and the RW buffer.
1132 * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1133 * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1134 * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1135 * bpf_arch_text_copy (when jit fails).
1137 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1138 struct bpf_binary_header *rw_header)
1140 u32 size = ro_header->size;
1142 bpf_prog_pack_free(ro_header);
1144 bpf_jit_uncharge_modmem(size);
1147 struct bpf_binary_header *
1148 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1150 unsigned long real_start = (unsigned long)fp->bpf_func;
1153 addr = real_start & BPF_PROG_CHUNK_MASK;
1154 return (void *)addr;
1157 static inline struct bpf_binary_header *
1158 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1160 unsigned long real_start = (unsigned long)fp->bpf_func;
1163 addr = real_start & PAGE_MASK;
1164 return (void *)addr;
1167 /* This symbol is only overridden by archs that have different
1168 * requirements than the usual eBPF JITs, f.e. when they only
1169 * implement cBPF JIT, do not set images read-only, etc.
1171 void __weak bpf_jit_free(struct bpf_prog *fp)
1174 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1176 bpf_jit_binary_free(hdr);
1177 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1180 bpf_prog_unlock_free(fp);
1183 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1184 const struct bpf_insn *insn, bool extra_pass,
1185 u64 *func_addr, bool *func_addr_fixed)
1187 s16 off = insn->off;
1188 s32 imm = insn->imm;
1191 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1192 if (!*func_addr_fixed) {
1193 /* Place-holder address till the last pass has collected
1194 * all addresses for JITed subprograms in which case we
1195 * can pick them up from prog->aux.
1199 else if (prog->aux->func &&
1200 off >= 0 && off < prog->aux->func_cnt)
1201 addr = (u8 *)prog->aux->func[off]->bpf_func;
1205 /* Address of a BPF helper call. Since part of the core
1206 * kernel, it's always at a fixed location. __bpf_call_base
1207 * and the helper with imm relative to it are both in core
1210 addr = (u8 *)__bpf_call_base + imm;
1213 *func_addr = (unsigned long)addr;
1217 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1218 const struct bpf_insn *aux,
1219 struct bpf_insn *to_buff,
1222 struct bpf_insn *to = to_buff;
1223 u32 imm_rnd = get_random_u32();
1226 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
1227 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1229 /* Constraints on AX register:
1231 * AX register is inaccessible from user space. It is mapped in
1232 * all JITs, and used here for constant blinding rewrites. It is
1233 * typically "stateless" meaning its contents are only valid within
1234 * the executed instruction, but not across several instructions.
1235 * There are a few exceptions however which are further detailed
1238 * Constant blinding is only used by JITs, not in the interpreter.
1239 * The interpreter uses AX in some occasions as a local temporary
1240 * register e.g. in DIV or MOD instructions.
1242 * In restricted circumstances, the verifier can also use the AX
1243 * register for rewrites as long as they do not interfere with
1246 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1249 if (from->imm == 0 &&
1250 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
1251 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1252 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1256 switch (from->code) {
1257 case BPF_ALU | BPF_ADD | BPF_K:
1258 case BPF_ALU | BPF_SUB | BPF_K:
1259 case BPF_ALU | BPF_AND | BPF_K:
1260 case BPF_ALU | BPF_OR | BPF_K:
1261 case BPF_ALU | BPF_XOR | BPF_K:
1262 case BPF_ALU | BPF_MUL | BPF_K:
1263 case BPF_ALU | BPF_MOV | BPF_K:
1264 case BPF_ALU | BPF_DIV | BPF_K:
1265 case BPF_ALU | BPF_MOD | BPF_K:
1266 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1267 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1268 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1271 case BPF_ALU64 | BPF_ADD | BPF_K:
1272 case BPF_ALU64 | BPF_SUB | BPF_K:
1273 case BPF_ALU64 | BPF_AND | BPF_K:
1274 case BPF_ALU64 | BPF_OR | BPF_K:
1275 case BPF_ALU64 | BPF_XOR | BPF_K:
1276 case BPF_ALU64 | BPF_MUL | BPF_K:
1277 case BPF_ALU64 | BPF_MOV | BPF_K:
1278 case BPF_ALU64 | BPF_DIV | BPF_K:
1279 case BPF_ALU64 | BPF_MOD | BPF_K:
1280 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1281 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1282 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1285 case BPF_JMP | BPF_JEQ | BPF_K:
1286 case BPF_JMP | BPF_JNE | BPF_K:
1287 case BPF_JMP | BPF_JGT | BPF_K:
1288 case BPF_JMP | BPF_JLT | BPF_K:
1289 case BPF_JMP | BPF_JGE | BPF_K:
1290 case BPF_JMP | BPF_JLE | BPF_K:
1291 case BPF_JMP | BPF_JSGT | BPF_K:
1292 case BPF_JMP | BPF_JSLT | BPF_K:
1293 case BPF_JMP | BPF_JSGE | BPF_K:
1294 case BPF_JMP | BPF_JSLE | BPF_K:
1295 case BPF_JMP | BPF_JSET | BPF_K:
1296 /* Accommodate for extra offset in case of a backjump. */
1300 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1301 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1302 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1305 case BPF_JMP32 | BPF_JEQ | BPF_K:
1306 case BPF_JMP32 | BPF_JNE | BPF_K:
1307 case BPF_JMP32 | BPF_JGT | BPF_K:
1308 case BPF_JMP32 | BPF_JLT | BPF_K:
1309 case BPF_JMP32 | BPF_JGE | BPF_K:
1310 case BPF_JMP32 | BPF_JLE | BPF_K:
1311 case BPF_JMP32 | BPF_JSGT | BPF_K:
1312 case BPF_JMP32 | BPF_JSLT | BPF_K:
1313 case BPF_JMP32 | BPF_JSGE | BPF_K:
1314 case BPF_JMP32 | BPF_JSLE | BPF_K:
1315 case BPF_JMP32 | BPF_JSET | BPF_K:
1316 /* Accommodate for extra offset in case of a backjump. */
1320 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1321 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1322 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1326 case BPF_LD | BPF_IMM | BPF_DW:
1327 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1328 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1329 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1330 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1332 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1333 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1334 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1336 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1337 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1340 case BPF_ST | BPF_MEM | BPF_DW:
1341 case BPF_ST | BPF_MEM | BPF_W:
1342 case BPF_ST | BPF_MEM | BPF_H:
1343 case BPF_ST | BPF_MEM | BPF_B:
1344 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1345 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1346 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1350 return to - to_buff;
1353 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1354 gfp_t gfp_extra_flags)
1356 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1357 struct bpf_prog *fp;
1359 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1361 /* aux->prog still points to the fp_other one, so
1362 * when promoting the clone to the real program,
1363 * this still needs to be adapted.
1365 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1371 static void bpf_prog_clone_free(struct bpf_prog *fp)
1373 /* aux was stolen by the other clone, so we cannot free
1374 * it from this path! It will be freed eventually by the
1375 * other program on release.
1377 * At this point, we don't need a deferred release since
1378 * clone is guaranteed to not be locked.
1383 __bpf_prog_free(fp);
1386 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1388 /* We have to repoint aux->prog to self, as we don't
1389 * know whether fp here is the clone or the original.
1392 bpf_prog_clone_free(fp_other);
1395 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1397 struct bpf_insn insn_buff[16], aux[2];
1398 struct bpf_prog *clone, *tmp;
1399 int insn_delta, insn_cnt;
1400 struct bpf_insn *insn;
1403 if (!prog->blinding_requested || prog->blinded)
1406 clone = bpf_prog_clone_create(prog, GFP_USER);
1408 return ERR_PTR(-ENOMEM);
1410 insn_cnt = clone->len;
1411 insn = clone->insnsi;
1413 for (i = 0; i < insn_cnt; i++, insn++) {
1414 if (bpf_pseudo_func(insn)) {
1415 /* ld_imm64 with an address of bpf subprog is not
1416 * a user controlled constant. Don't randomize it,
1417 * since it will conflict with jit_subprogs() logic.
1424 /* We temporarily need to hold the original ld64 insn
1425 * so that we can still access the first part in the
1426 * second blinding run.
1428 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1430 memcpy(aux, insn, sizeof(aux));
1432 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1433 clone->aux->verifier_zext);
1437 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1439 /* Patching may have repointed aux->prog during
1440 * realloc from the original one, so we need to
1441 * fix it up here on error.
1443 bpf_jit_prog_release_other(prog, clone);
1448 insn_delta = rewritten - 1;
1450 /* Walk new program and skip insns we just inserted. */
1451 insn = clone->insnsi + i + insn_delta;
1452 insn_cnt += insn_delta;
1459 #endif /* CONFIG_BPF_JIT */
1461 /* Base function for offset calculation. Needs to go into .text section,
1462 * therefore keeping it non-static as well; will also be used by JITs
1463 * anyway later on, so do not let the compiler omit it. This also needs
1464 * to go into kallsyms for correlation from e.g. bpftool, so naming
1467 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1471 EXPORT_SYMBOL_GPL(__bpf_call_base);
1473 /* All UAPI available opcodes. */
1474 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1475 /* 32 bit ALU operations. */ \
1476 /* Register based. */ \
1477 INSN_3(ALU, ADD, X), \
1478 INSN_3(ALU, SUB, X), \
1479 INSN_3(ALU, AND, X), \
1480 INSN_3(ALU, OR, X), \
1481 INSN_3(ALU, LSH, X), \
1482 INSN_3(ALU, RSH, X), \
1483 INSN_3(ALU, XOR, X), \
1484 INSN_3(ALU, MUL, X), \
1485 INSN_3(ALU, MOV, X), \
1486 INSN_3(ALU, ARSH, X), \
1487 INSN_3(ALU, DIV, X), \
1488 INSN_3(ALU, MOD, X), \
1490 INSN_3(ALU, END, TO_BE), \
1491 INSN_3(ALU, END, TO_LE), \
1492 /* Immediate based. */ \
1493 INSN_3(ALU, ADD, K), \
1494 INSN_3(ALU, SUB, K), \
1495 INSN_3(ALU, AND, K), \
1496 INSN_3(ALU, OR, K), \
1497 INSN_3(ALU, LSH, K), \
1498 INSN_3(ALU, RSH, K), \
1499 INSN_3(ALU, XOR, K), \
1500 INSN_3(ALU, MUL, K), \
1501 INSN_3(ALU, MOV, K), \
1502 INSN_3(ALU, ARSH, K), \
1503 INSN_3(ALU, DIV, K), \
1504 INSN_3(ALU, MOD, K), \
1505 /* 64 bit ALU operations. */ \
1506 /* Register based. */ \
1507 INSN_3(ALU64, ADD, X), \
1508 INSN_3(ALU64, SUB, X), \
1509 INSN_3(ALU64, AND, X), \
1510 INSN_3(ALU64, OR, X), \
1511 INSN_3(ALU64, LSH, X), \
1512 INSN_3(ALU64, RSH, X), \
1513 INSN_3(ALU64, XOR, X), \
1514 INSN_3(ALU64, MUL, X), \
1515 INSN_3(ALU64, MOV, X), \
1516 INSN_3(ALU64, ARSH, X), \
1517 INSN_3(ALU64, DIV, X), \
1518 INSN_3(ALU64, MOD, X), \
1519 INSN_2(ALU64, NEG), \
1520 /* Immediate based. */ \
1521 INSN_3(ALU64, ADD, K), \
1522 INSN_3(ALU64, SUB, K), \
1523 INSN_3(ALU64, AND, K), \
1524 INSN_3(ALU64, OR, K), \
1525 INSN_3(ALU64, LSH, K), \
1526 INSN_3(ALU64, RSH, K), \
1527 INSN_3(ALU64, XOR, K), \
1528 INSN_3(ALU64, MUL, K), \
1529 INSN_3(ALU64, MOV, K), \
1530 INSN_3(ALU64, ARSH, K), \
1531 INSN_3(ALU64, DIV, K), \
1532 INSN_3(ALU64, MOD, K), \
1533 /* Call instruction. */ \
1534 INSN_2(JMP, CALL), \
1535 /* Exit instruction. */ \
1536 INSN_2(JMP, EXIT), \
1537 /* 32-bit Jump instructions. */ \
1538 /* Register based. */ \
1539 INSN_3(JMP32, JEQ, X), \
1540 INSN_3(JMP32, JNE, X), \
1541 INSN_3(JMP32, JGT, X), \
1542 INSN_3(JMP32, JLT, X), \
1543 INSN_3(JMP32, JGE, X), \
1544 INSN_3(JMP32, JLE, X), \
1545 INSN_3(JMP32, JSGT, X), \
1546 INSN_3(JMP32, JSLT, X), \
1547 INSN_3(JMP32, JSGE, X), \
1548 INSN_3(JMP32, JSLE, X), \
1549 INSN_3(JMP32, JSET, X), \
1550 /* Immediate based. */ \
1551 INSN_3(JMP32, JEQ, K), \
1552 INSN_3(JMP32, JNE, K), \
1553 INSN_3(JMP32, JGT, K), \
1554 INSN_3(JMP32, JLT, K), \
1555 INSN_3(JMP32, JGE, K), \
1556 INSN_3(JMP32, JLE, K), \
1557 INSN_3(JMP32, JSGT, K), \
1558 INSN_3(JMP32, JSLT, K), \
1559 INSN_3(JMP32, JSGE, K), \
1560 INSN_3(JMP32, JSLE, K), \
1561 INSN_3(JMP32, JSET, K), \
1562 /* Jump instructions. */ \
1563 /* Register based. */ \
1564 INSN_3(JMP, JEQ, X), \
1565 INSN_3(JMP, JNE, X), \
1566 INSN_3(JMP, JGT, X), \
1567 INSN_3(JMP, JLT, X), \
1568 INSN_3(JMP, JGE, X), \
1569 INSN_3(JMP, JLE, X), \
1570 INSN_3(JMP, JSGT, X), \
1571 INSN_3(JMP, JSLT, X), \
1572 INSN_3(JMP, JSGE, X), \
1573 INSN_3(JMP, JSLE, X), \
1574 INSN_3(JMP, JSET, X), \
1575 /* Immediate based. */ \
1576 INSN_3(JMP, JEQ, K), \
1577 INSN_3(JMP, JNE, K), \
1578 INSN_3(JMP, JGT, K), \
1579 INSN_3(JMP, JLT, K), \
1580 INSN_3(JMP, JGE, K), \
1581 INSN_3(JMP, JLE, K), \
1582 INSN_3(JMP, JSGT, K), \
1583 INSN_3(JMP, JSLT, K), \
1584 INSN_3(JMP, JSGE, K), \
1585 INSN_3(JMP, JSLE, K), \
1586 INSN_3(JMP, JSET, K), \
1588 /* Store instructions. */ \
1589 /* Register based. */ \
1590 INSN_3(STX, MEM, B), \
1591 INSN_3(STX, MEM, H), \
1592 INSN_3(STX, MEM, W), \
1593 INSN_3(STX, MEM, DW), \
1594 INSN_3(STX, ATOMIC, W), \
1595 INSN_3(STX, ATOMIC, DW), \
1596 /* Immediate based. */ \
1597 INSN_3(ST, MEM, B), \
1598 INSN_3(ST, MEM, H), \
1599 INSN_3(ST, MEM, W), \
1600 INSN_3(ST, MEM, DW), \
1601 /* Load instructions. */ \
1602 /* Register based. */ \
1603 INSN_3(LDX, MEM, B), \
1604 INSN_3(LDX, MEM, H), \
1605 INSN_3(LDX, MEM, W), \
1606 INSN_3(LDX, MEM, DW), \
1607 /* Immediate based. */ \
1610 bool bpf_opcode_in_insntable(u8 code)
1612 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1613 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1614 static const bool public_insntable[256] = {
1615 [0 ... 255] = false,
1616 /* Now overwrite non-defaults ... */
1617 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1618 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1619 [BPF_LD | BPF_ABS | BPF_B] = true,
1620 [BPF_LD | BPF_ABS | BPF_H] = true,
1621 [BPF_LD | BPF_ABS | BPF_W] = true,
1622 [BPF_LD | BPF_IND | BPF_B] = true,
1623 [BPF_LD | BPF_IND | BPF_H] = true,
1624 [BPF_LD | BPF_IND | BPF_W] = true,
1626 #undef BPF_INSN_3_TBL
1627 #undef BPF_INSN_2_TBL
1628 return public_insntable[code];
1631 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1632 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1634 memset(dst, 0, size);
1639 * ___bpf_prog_run - run eBPF program on a given context
1640 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1641 * @insn: is the array of eBPF instructions
1643 * Decode and execute eBPF instructions.
1645 * Return: whatever value is in %BPF_R0 at program exit
1647 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1649 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1650 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1651 static const void * const jumptable[256] __annotate_jump_table = {
1652 [0 ... 255] = &&default_label,
1653 /* Now overwrite non-defaults ... */
1654 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1655 /* Non-UAPI available opcodes. */
1656 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1657 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1658 [BPF_ST | BPF_NOSPEC] = &&ST_NOSPEC,
1659 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1660 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1661 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1662 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1664 #undef BPF_INSN_3_LBL
1665 #undef BPF_INSN_2_LBL
1666 u32 tail_call_cnt = 0;
1668 #define CONT ({ insn++; goto select_insn; })
1669 #define CONT_JMP ({ insn++; goto select_insn; })
1672 goto *jumptable[insn->code];
1674 /* Explicitly mask the register-based shift amounts with 63 or 31
1675 * to avoid undefined behavior. Normally this won't affect the
1676 * generated code, for example, in case of native 64 bit archs such
1677 * as x86-64 or arm64, the compiler is optimizing the AND away for
1678 * the interpreter. In case of JITs, each of the JIT backends compiles
1679 * the BPF shift operations to machine instructions which produce
1680 * implementation-defined results in such a case; the resulting
1681 * contents of the register may be arbitrary, but program behaviour
1682 * as a whole remains defined. In other words, in case of JIT backends,
1683 * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1686 #define SHT(OPCODE, OP) \
1687 ALU64_##OPCODE##_X: \
1688 DST = DST OP (SRC & 63); \
1691 DST = (u32) DST OP ((u32) SRC & 31); \
1693 ALU64_##OPCODE##_K: \
1697 DST = (u32) DST OP (u32) IMM; \
1700 #define ALU(OPCODE, OP) \
1701 ALU64_##OPCODE##_X: \
1705 DST = (u32) DST OP (u32) SRC; \
1707 ALU64_##OPCODE##_K: \
1711 DST = (u32) DST OP (u32) IMM; \
1742 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1746 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1749 DST = (u64) (u32) (((s32) DST) >> IMM);
1752 (*(s64 *) &DST) >>= (SRC & 63);
1755 (*(s64 *) &DST) >>= IMM;
1758 div64_u64_rem(DST, SRC, &AX);
1763 DST = do_div(AX, (u32) SRC);
1766 div64_u64_rem(DST, IMM, &AX);
1771 DST = do_div(AX, (u32) IMM);
1774 DST = div64_u64(DST, SRC);
1778 do_div(AX, (u32) SRC);
1782 DST = div64_u64(DST, IMM);
1786 do_div(AX, (u32) IMM);
1792 DST = (__force u16) cpu_to_be16(DST);
1795 DST = (__force u32) cpu_to_be32(DST);
1798 DST = (__force u64) cpu_to_be64(DST);
1805 DST = (__force u16) cpu_to_le16(DST);
1808 DST = (__force u32) cpu_to_le32(DST);
1811 DST = (__force u64) cpu_to_le64(DST);
1818 /* Function call scratches BPF_R1-BPF_R5 registers,
1819 * preserves BPF_R6-BPF_R9, and stores return value
1822 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1827 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1830 insn + insn->off + 1);
1834 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1835 struct bpf_array *array = container_of(map, struct bpf_array, map);
1836 struct bpf_prog *prog;
1839 if (unlikely(index >= array->map.max_entries))
1842 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
1847 prog = READ_ONCE(array->ptrs[index]);
1851 /* ARG1 at this point is guaranteed to point to CTX from
1852 * the verifier side due to the fact that the tail call is
1853 * handled like a helper, that is, bpf_tail_call_proto,
1854 * where arg1_type is ARG_PTR_TO_CTX.
1856 insn = prog->insnsi;
1867 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1869 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1870 insn += insn->off; \
1874 JMP32_##OPCODE##_X: \
1875 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1876 insn += insn->off; \
1881 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1882 insn += insn->off; \
1886 JMP32_##OPCODE##_K: \
1887 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1888 insn += insn->off; \
1892 COND_JMP(u, JEQ, ==)
1893 COND_JMP(u, JNE, !=)
1896 COND_JMP(u, JGE, >=)
1897 COND_JMP(u, JLE, <=)
1898 COND_JMP(u, JSET, &)
1899 COND_JMP(s, JSGT, >)
1900 COND_JMP(s, JSLT, <)
1901 COND_JMP(s, JSGE, >=)
1902 COND_JMP(s, JSLE, <=)
1904 /* ST, STX and LDX*/
1906 /* Speculation barrier for mitigating Speculative Store Bypass.
1907 * In case of arm64, we rely on the firmware mitigation as
1908 * controlled via the ssbd kernel parameter. Whenever the
1909 * mitigation is enabled, it works for all of the kernel code
1910 * with no need to provide any additional instructions here.
1911 * In case of x86, we use 'lfence' insn for mitigation. We
1912 * reuse preexisting logic from Spectre v1 mitigation that
1913 * happens to produce the required code on x86 for v4 as well.
1917 #define LDST(SIZEOP, SIZE) \
1919 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1922 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1925 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1927 LDX_PROBE_MEM_##SIZEOP: \
1928 bpf_probe_read_kernel(&DST, sizeof(SIZE), \
1929 (const void *)(long) (SRC + insn->off)); \
1930 DST = *((SIZE *)&DST); \
1939 #define ATOMIC_ALU_OP(BOP, KOP) \
1941 if (BPF_SIZE(insn->code) == BPF_W) \
1942 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
1943 (DST + insn->off)); \
1945 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
1946 (DST + insn->off)); \
1948 case BOP | BPF_FETCH: \
1949 if (BPF_SIZE(insn->code) == BPF_W) \
1950 SRC = (u32) atomic_fetch_##KOP( \
1952 (atomic_t *)(unsigned long) (DST + insn->off)); \
1954 SRC = (u64) atomic64_fetch_##KOP( \
1956 (atomic64_t *)(unsigned long) (DST + insn->off)); \
1962 ATOMIC_ALU_OP(BPF_ADD, add)
1963 ATOMIC_ALU_OP(BPF_AND, and)
1964 ATOMIC_ALU_OP(BPF_OR, or)
1965 ATOMIC_ALU_OP(BPF_XOR, xor)
1966 #undef ATOMIC_ALU_OP
1969 if (BPF_SIZE(insn->code) == BPF_W)
1970 SRC = (u32) atomic_xchg(
1971 (atomic_t *)(unsigned long) (DST + insn->off),
1974 SRC = (u64) atomic64_xchg(
1975 (atomic64_t *)(unsigned long) (DST + insn->off),
1979 if (BPF_SIZE(insn->code) == BPF_W)
1980 BPF_R0 = (u32) atomic_cmpxchg(
1981 (atomic_t *)(unsigned long) (DST + insn->off),
1982 (u32) BPF_R0, (u32) SRC);
1984 BPF_R0 = (u64) atomic64_cmpxchg(
1985 (atomic64_t *)(unsigned long) (DST + insn->off),
1986 (u64) BPF_R0, (u64) SRC);
1995 /* If we ever reach this, we have a bug somewhere. Die hard here
1996 * instead of just returning 0; we could be somewhere in a subprog,
1997 * so execution could continue otherwise which we do /not/ want.
1999 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
2001 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
2002 insn->code, insn->imm);
2007 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2008 #define DEFINE_BPF_PROG_RUN(stack_size) \
2009 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2011 u64 stack[stack_size / sizeof(u64)]; \
2012 u64 regs[MAX_BPF_EXT_REG] = {}; \
2014 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2015 ARG1 = (u64) (unsigned long) ctx; \
2016 return ___bpf_prog_run(regs, insn); \
2019 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2020 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2021 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2022 const struct bpf_insn *insn) \
2024 u64 stack[stack_size / sizeof(u64)]; \
2025 u64 regs[MAX_BPF_EXT_REG]; \
2027 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2033 return ___bpf_prog_run(regs, insn); \
2036 #define EVAL1(FN, X) FN(X)
2037 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2038 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2039 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2040 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2041 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2043 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2044 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2045 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2047 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2048 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2049 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2051 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2053 static unsigned int (*interpreters[])(const void *ctx,
2054 const struct bpf_insn *insn) = {
2055 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2056 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2057 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2059 #undef PROG_NAME_LIST
2060 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2061 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2062 const struct bpf_insn *insn) = {
2063 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2064 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2065 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2067 #undef PROG_NAME_LIST
2069 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2071 stack_depth = max_t(u32, stack_depth, 1);
2072 insn->off = (s16) insn->imm;
2073 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2074 __bpf_call_base_args;
2075 insn->code = BPF_JMP | BPF_CALL_ARGS;
2079 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2080 const struct bpf_insn *insn)
2082 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2083 * is not working properly, so warn about it!
2090 bool bpf_prog_map_compatible(struct bpf_map *map,
2091 const struct bpf_prog *fp)
2093 enum bpf_prog_type prog_type = resolve_prog_type(fp);
2096 if (fp->kprobe_override)
2099 /* XDP programs inserted into maps are not guaranteed to run on
2100 * a particular netdev (and can run outside driver context entirely
2101 * in the case of devmap and cpumap). Until device checks
2102 * are implemented, prohibit adding dev-bound programs to program maps.
2104 if (bpf_prog_is_dev_bound(fp->aux))
2107 spin_lock(&map->owner.lock);
2108 if (!map->owner.type) {
2109 /* There's no owner yet where we could check for
2112 map->owner.type = prog_type;
2113 map->owner.jited = fp->jited;
2114 map->owner.xdp_has_frags = fp->aux->xdp_has_frags;
2117 ret = map->owner.type == prog_type &&
2118 map->owner.jited == fp->jited &&
2119 map->owner.xdp_has_frags == fp->aux->xdp_has_frags;
2121 spin_unlock(&map->owner.lock);
2126 static int bpf_check_tail_call(const struct bpf_prog *fp)
2128 struct bpf_prog_aux *aux = fp->aux;
2131 mutex_lock(&aux->used_maps_mutex);
2132 for (i = 0; i < aux->used_map_cnt; i++) {
2133 struct bpf_map *map = aux->used_maps[i];
2135 if (!map_type_contains_progs(map))
2138 if (!bpf_prog_map_compatible(map, fp)) {
2145 mutex_unlock(&aux->used_maps_mutex);
2149 static void bpf_prog_select_func(struct bpf_prog *fp)
2151 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2152 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2154 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2156 fp->bpf_func = __bpf_prog_ret0_warn;
2161 * bpf_prog_select_runtime - select exec runtime for BPF program
2162 * @fp: bpf_prog populated with BPF program
2163 * @err: pointer to error variable
2165 * Try to JIT eBPF program, if JIT is not available, use interpreter.
2166 * The BPF program will be executed via bpf_prog_run() function.
2168 * Return: the &fp argument along with &err set to 0 for success or
2169 * a negative errno code on failure
2171 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2173 /* In case of BPF to BPF calls, verifier did all the prep
2174 * work with regards to JITing, etc.
2176 bool jit_needed = false;
2181 if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2182 bpf_prog_has_kfunc_call(fp))
2185 bpf_prog_select_func(fp);
2187 /* eBPF JITs can rewrite the program in case constant
2188 * blinding is active. However, in case of error during
2189 * blinding, bpf_int_jit_compile() must always return a
2190 * valid program, which in this case would simply not
2191 * be JITed, but falls back to the interpreter.
2193 if (!bpf_prog_is_offloaded(fp->aux)) {
2194 *err = bpf_prog_alloc_jited_linfo(fp);
2198 fp = bpf_int_jit_compile(fp);
2199 bpf_prog_jit_attempt_done(fp);
2200 if (!fp->jited && jit_needed) {
2205 *err = bpf_prog_offload_compile(fp);
2211 bpf_prog_lock_ro(fp);
2213 /* The tail call compatibility check can only be done at
2214 * this late stage as we need to determine, if we deal
2215 * with JITed or non JITed program concatenations and not
2216 * all eBPF JITs might immediately support all features.
2218 *err = bpf_check_tail_call(fp);
2222 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2224 static unsigned int __bpf_prog_ret1(const void *ctx,
2225 const struct bpf_insn *insn)
2230 static struct bpf_prog_dummy {
2231 struct bpf_prog prog;
2232 } dummy_bpf_prog = {
2234 .bpf_func = __bpf_prog_ret1,
2238 struct bpf_empty_prog_array bpf_empty_prog_array = {
2241 EXPORT_SYMBOL(bpf_empty_prog_array);
2243 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2246 return kzalloc(sizeof(struct bpf_prog_array) +
2247 sizeof(struct bpf_prog_array_item) *
2251 return &bpf_empty_prog_array.hdr;
2254 void bpf_prog_array_free(struct bpf_prog_array *progs)
2256 if (!progs || progs == &bpf_empty_prog_array.hdr)
2258 kfree_rcu(progs, rcu);
2261 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2263 struct bpf_prog_array *progs;
2265 /* If RCU Tasks Trace grace period implies RCU grace period, there is
2266 * no need to call kfree_rcu(), just call kfree() directly.
2268 progs = container_of(rcu, struct bpf_prog_array, rcu);
2269 if (rcu_trace_implies_rcu_gp())
2272 kfree_rcu(progs, rcu);
2275 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2277 if (!progs || progs == &bpf_empty_prog_array.hdr)
2279 call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2282 int bpf_prog_array_length(struct bpf_prog_array *array)
2284 struct bpf_prog_array_item *item;
2287 for (item = array->items; item->prog; item++)
2288 if (item->prog != &dummy_bpf_prog.prog)
2293 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2295 struct bpf_prog_array_item *item;
2297 for (item = array->items; item->prog; item++)
2298 if (item->prog != &dummy_bpf_prog.prog)
2303 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2307 struct bpf_prog_array_item *item;
2310 for (item = array->items; item->prog; item++) {
2311 if (item->prog == &dummy_bpf_prog.prog)
2313 prog_ids[i] = item->prog->aux->id;
2314 if (++i == request_cnt) {
2320 return !!(item->prog);
2323 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2324 __u32 __user *prog_ids, u32 cnt)
2326 unsigned long err = 0;
2330 /* users of this function are doing:
2331 * cnt = bpf_prog_array_length();
2333 * bpf_prog_array_copy_to_user(..., cnt);
2334 * so below kcalloc doesn't need extra cnt > 0 check.
2336 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2339 nospc = bpf_prog_array_copy_core(array, ids, cnt);
2340 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2349 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2350 struct bpf_prog *old_prog)
2352 struct bpf_prog_array_item *item;
2354 for (item = array->items; item->prog; item++)
2355 if (item->prog == old_prog) {
2356 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2362 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2363 * index into the program array with
2364 * a dummy no-op program.
2365 * @array: a bpf_prog_array
2366 * @index: the index of the program to replace
2368 * Skips over dummy programs, by not counting them, when calculating
2369 * the position of the program to replace.
2373 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2374 * * -ENOENT - Index out of range
2376 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2378 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2382 * bpf_prog_array_update_at() - Updates the program at the given index
2383 * into the program array.
2384 * @array: a bpf_prog_array
2385 * @index: the index of the program to update
2386 * @prog: the program to insert into the array
2388 * Skips over dummy programs, by not counting them, when calculating
2389 * the position of the program to update.
2393 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2394 * * -ENOENT - Index out of range
2396 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2397 struct bpf_prog *prog)
2399 struct bpf_prog_array_item *item;
2401 if (unlikely(index < 0))
2404 for (item = array->items; item->prog; item++) {
2405 if (item->prog == &dummy_bpf_prog.prog)
2408 WRITE_ONCE(item->prog, prog);
2416 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2417 struct bpf_prog *exclude_prog,
2418 struct bpf_prog *include_prog,
2420 struct bpf_prog_array **new_array)
2422 int new_prog_cnt, carry_prog_cnt = 0;
2423 struct bpf_prog_array_item *existing, *new;
2424 struct bpf_prog_array *array;
2425 bool found_exclude = false;
2427 /* Figure out how many existing progs we need to carry over to
2431 existing = old_array->items;
2432 for (; existing->prog; existing++) {
2433 if (existing->prog == exclude_prog) {
2434 found_exclude = true;
2437 if (existing->prog != &dummy_bpf_prog.prog)
2439 if (existing->prog == include_prog)
2444 if (exclude_prog && !found_exclude)
2447 /* How many progs (not NULL) will be in the new array? */
2448 new_prog_cnt = carry_prog_cnt;
2452 /* Do we have any prog (not NULL) in the new array? */
2453 if (!new_prog_cnt) {
2458 /* +1 as the end of prog_array is marked with NULL */
2459 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2464 /* Fill in the new prog array */
2465 if (carry_prog_cnt) {
2466 existing = old_array->items;
2467 for (; existing->prog; existing++) {
2468 if (existing->prog == exclude_prog ||
2469 existing->prog == &dummy_bpf_prog.prog)
2472 new->prog = existing->prog;
2473 new->bpf_cookie = existing->bpf_cookie;
2478 new->prog = include_prog;
2479 new->bpf_cookie = bpf_cookie;
2487 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2488 u32 *prog_ids, u32 request_cnt,
2494 cnt = bpf_prog_array_length(array);
2498 /* return early if user requested only program count or nothing to copy */
2499 if (!request_cnt || !cnt)
2502 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2503 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2507 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2508 struct bpf_map **used_maps, u32 len)
2510 struct bpf_map *map;
2513 for (i = 0; i < len; i++) {
2515 if (map->ops->map_poke_untrack)
2516 map->ops->map_poke_untrack(map, aux);
2521 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2523 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2524 kfree(aux->used_maps);
2527 void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2528 struct btf_mod_pair *used_btfs, u32 len)
2530 #ifdef CONFIG_BPF_SYSCALL
2531 struct btf_mod_pair *btf_mod;
2534 for (i = 0; i < len; i++) {
2535 btf_mod = &used_btfs[i];
2536 if (btf_mod->module)
2537 module_put(btf_mod->module);
2538 btf_put(btf_mod->btf);
2543 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2545 __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2546 kfree(aux->used_btfs);
2549 static void bpf_prog_free_deferred(struct work_struct *work)
2551 struct bpf_prog_aux *aux;
2554 aux = container_of(work, struct bpf_prog_aux, work);
2555 #ifdef CONFIG_BPF_SYSCALL
2556 bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2558 #ifdef CONFIG_CGROUP_BPF
2559 if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2560 bpf_cgroup_atype_put(aux->cgroup_atype);
2562 bpf_free_used_maps(aux);
2563 bpf_free_used_btfs(aux);
2564 if (bpf_prog_is_dev_bound(aux))
2565 bpf_prog_dev_bound_destroy(aux->prog);
2566 #ifdef CONFIG_PERF_EVENTS
2567 if (aux->prog->has_callchain_buf)
2568 put_callchain_buffers();
2570 if (aux->dst_trampoline)
2571 bpf_trampoline_put(aux->dst_trampoline);
2572 for (i = 0; i < aux->func_cnt; i++) {
2573 /* We can just unlink the subprog poke descriptor table as
2574 * it was originally linked to the main program and is also
2575 * released along with it.
2577 aux->func[i]->aux->poke_tab = NULL;
2578 bpf_jit_free(aux->func[i]);
2580 if (aux->func_cnt) {
2582 bpf_prog_unlock_free(aux->prog);
2584 bpf_jit_free(aux->prog);
2588 void bpf_prog_free(struct bpf_prog *fp)
2590 struct bpf_prog_aux *aux = fp->aux;
2593 bpf_prog_put(aux->dst_prog);
2594 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2595 schedule_work(&aux->work);
2597 EXPORT_SYMBOL_GPL(bpf_prog_free);
2599 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2600 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2602 void bpf_user_rnd_init_once(void)
2604 prandom_init_once(&bpf_user_rnd_state);
2607 BPF_CALL_0(bpf_user_rnd_u32)
2609 /* Should someone ever have the rather unwise idea to use some
2610 * of the registers passed into this function, then note that
2611 * this function is called from native eBPF and classic-to-eBPF
2612 * transformations. Register assignments from both sides are
2613 * different, f.e. classic always sets fn(ctx, A, X) here.
2615 struct rnd_state *state;
2618 state = &get_cpu_var(bpf_user_rnd_state);
2619 res = prandom_u32_state(state);
2620 put_cpu_var(bpf_user_rnd_state);
2625 BPF_CALL_0(bpf_get_raw_cpu_id)
2627 return raw_smp_processor_id();
2630 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2631 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2632 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2633 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2634 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2635 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2636 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2637 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2638 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2639 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2640 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2642 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2643 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2644 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2645 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2646 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2647 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2648 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak;
2650 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2651 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2652 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2653 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2654 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2655 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2656 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2657 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2658 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2659 const struct bpf_func_proto bpf_set_retval_proto __weak;
2660 const struct bpf_func_proto bpf_get_retval_proto __weak;
2662 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2667 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2673 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2674 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2678 EXPORT_SYMBOL_GPL(bpf_event_output);
2680 /* Always built-in helper functions. */
2681 const struct bpf_func_proto bpf_tail_call_proto = {
2684 .ret_type = RET_VOID,
2685 .arg1_type = ARG_PTR_TO_CTX,
2686 .arg2_type = ARG_CONST_MAP_PTR,
2687 .arg3_type = ARG_ANYTHING,
2690 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2691 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2692 * eBPF and implicitly also cBPF can get JITed!
2694 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2699 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2700 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2702 void __weak bpf_jit_compile(struct bpf_prog *prog)
2706 bool __weak bpf_helper_changes_pkt_data(void *func)
2711 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2712 * analysis code and wants explicit zero extension inserted by verifier.
2713 * Otherwise, return FALSE.
2715 * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2716 * you don't override this. JITs that don't want these extra insns can detect
2717 * them using insn_is_zext.
2719 bool __weak bpf_jit_needs_zext(void)
2724 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2725 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2730 bool __weak bpf_jit_supports_kfunc_call(void)
2735 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2736 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2738 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2744 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2745 void *addr1, void *addr2)
2750 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
2752 return ERR_PTR(-ENOTSUPP);
2755 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
2760 #ifdef CONFIG_BPF_SYSCALL
2761 static int __init bpf_global_ma_init(void)
2765 ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false);
2766 bpf_global_ma_set = !ret;
2769 late_initcall(bpf_global_ma_init);
2772 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2773 EXPORT_SYMBOL(bpf_stats_enabled_key);
2775 /* All definitions of tracepoints related to BPF. */
2776 #define CREATE_TRACE_POINTS
2777 #include <linux/bpf_trace.h>
2779 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2780 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);