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 <asm/unaligned.h>
38 #define BPF_R0 regs[BPF_REG_0]
39 #define BPF_R1 regs[BPF_REG_1]
40 #define BPF_R2 regs[BPF_REG_2]
41 #define BPF_R3 regs[BPF_REG_3]
42 #define BPF_R4 regs[BPF_REG_4]
43 #define BPF_R5 regs[BPF_REG_5]
44 #define BPF_R6 regs[BPF_REG_6]
45 #define BPF_R7 regs[BPF_REG_7]
46 #define BPF_R8 regs[BPF_REG_8]
47 #define BPF_R9 regs[BPF_REG_9]
48 #define BPF_R10 regs[BPF_REG_10]
51 #define DST regs[insn->dst_reg]
52 #define SRC regs[insn->src_reg]
53 #define FP regs[BPF_REG_FP]
54 #define AX regs[BPF_REG_AX]
55 #define ARG1 regs[BPF_REG_ARG1]
56 #define CTX regs[BPF_REG_CTX]
59 /* No hurry in this branch
61 * Exported for the bpf jit load helper.
63 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
68 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
69 else if (k >= SKF_LL_OFF)
70 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
72 if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
78 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
80 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
81 struct bpf_prog_aux *aux;
84 size = round_up(size, PAGE_SIZE);
85 fp = __vmalloc(size, gfp_flags);
89 aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
95 fp->pages = size / PAGE_SIZE;
98 fp->jit_requested = ebpf_jit_enabled();
100 INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
101 mutex_init(&fp->aux->used_maps_mutex);
102 mutex_init(&fp->aux->dst_mutex);
107 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
109 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
110 struct bpf_prog *prog;
113 prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
117 prog->aux->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
118 if (!prog->aux->stats) {
124 for_each_possible_cpu(cpu) {
125 struct bpf_prog_stats *pstats;
127 pstats = per_cpu_ptr(prog->aux->stats, cpu);
128 u64_stats_init(&pstats->syncp);
132 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
134 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
136 if (!prog->aux->nr_linfo || !prog->jit_requested)
139 prog->aux->jited_linfo = kcalloc(prog->aux->nr_linfo,
140 sizeof(*prog->aux->jited_linfo),
141 GFP_KERNEL | __GFP_NOWARN);
142 if (!prog->aux->jited_linfo)
148 void bpf_prog_free_jited_linfo(struct bpf_prog *prog)
150 kfree(prog->aux->jited_linfo);
151 prog->aux->jited_linfo = NULL;
154 void bpf_prog_free_unused_jited_linfo(struct bpf_prog *prog)
156 if (prog->aux->jited_linfo && !prog->aux->jited_linfo[0])
157 bpf_prog_free_jited_linfo(prog);
160 /* The jit engine is responsible to provide an array
161 * for insn_off to the jited_off mapping (insn_to_jit_off).
163 * The idx to this array is the insn_off. Hence, the insn_off
164 * here is relative to the prog itself instead of the main prog.
165 * This array has one entry for each xlated bpf insn.
167 * jited_off is the byte off to the last byte of the jited insn.
171 * The first bpf insn off of the prog. The insn off
172 * here is relative to the main prog.
173 * e.g. if prog is a subprog, insn_start > 0
175 * The prog's idx to prog->aux->linfo and jited_linfo
177 * jited_linfo[linfo_idx] = prog->bpf_func
181 * jited_linfo[i] = prog->bpf_func +
182 * insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
184 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
185 const u32 *insn_to_jit_off)
187 u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
188 const struct bpf_line_info *linfo;
191 if (!prog->aux->jited_linfo)
192 /* Userspace did not provide linfo */
195 linfo_idx = prog->aux->linfo_idx;
196 linfo = &prog->aux->linfo[linfo_idx];
197 insn_start = linfo[0].insn_off;
198 insn_end = insn_start + prog->len;
200 jited_linfo = &prog->aux->jited_linfo[linfo_idx];
201 jited_linfo[0] = prog->bpf_func;
203 nr_linfo = prog->aux->nr_linfo - linfo_idx;
205 for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
206 /* The verifier ensures that linfo[i].insn_off is
207 * strictly increasing
209 jited_linfo[i] = prog->bpf_func +
210 insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
213 void bpf_prog_free_linfo(struct bpf_prog *prog)
215 bpf_prog_free_jited_linfo(prog);
216 kvfree(prog->aux->linfo);
219 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
220 gfp_t gfp_extra_flags)
222 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
227 size = round_up(size, PAGE_SIZE);
228 pages = size / PAGE_SIZE;
229 if (pages <= fp_old->pages)
232 delta = pages - fp_old->pages;
233 ret = __bpf_prog_charge(fp_old->aux->user, delta);
237 fp = __vmalloc(size, gfp_flags);
239 __bpf_prog_uncharge(fp_old->aux->user, delta);
241 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
245 /* We keep fp->aux from fp_old around in the new
246 * reallocated structure.
249 __bpf_prog_free(fp_old);
255 void __bpf_prog_free(struct bpf_prog *fp)
258 mutex_destroy(&fp->aux->used_maps_mutex);
259 mutex_destroy(&fp->aux->dst_mutex);
260 free_percpu(fp->aux->stats);
261 kfree(fp->aux->poke_tab);
267 int bpf_prog_calc_tag(struct bpf_prog *fp)
269 const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
270 u32 raw_size = bpf_prog_tag_scratch_size(fp);
271 u32 digest[SHA1_DIGEST_WORDS];
272 u32 ws[SHA1_WORKSPACE_WORDS];
273 u32 i, bsize, psize, blocks;
274 struct bpf_insn *dst;
280 raw = vmalloc(raw_size);
285 memset(ws, 0, sizeof(ws));
287 /* We need to take out the map fd for the digest calculation
288 * since they are unstable from user space side.
291 for (i = 0, was_ld_map = false; i < fp->len; i++) {
292 dst[i] = fp->insnsi[i];
294 dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
295 (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
296 dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
299 } else if (was_ld_map &&
301 dst[i].dst_reg == 0 &&
302 dst[i].src_reg == 0 &&
311 psize = bpf_prog_insn_size(fp);
312 memset(&raw[psize], 0, raw_size - psize);
315 bsize = round_up(psize, SHA1_BLOCK_SIZE);
316 blocks = bsize / SHA1_BLOCK_SIZE;
318 if (bsize - psize >= sizeof(__be64)) {
319 bits = (__be64 *)(todo + bsize - sizeof(__be64));
321 bits = (__be64 *)(todo + bsize + bits_offset);
324 *bits = cpu_to_be64((psize - 1) << 3);
327 sha1_transform(digest, todo, ws);
328 todo += SHA1_BLOCK_SIZE;
331 result = (__force __be32 *)digest;
332 for (i = 0; i < SHA1_DIGEST_WORDS; i++)
333 result[i] = cpu_to_be32(digest[i]);
334 memcpy(fp->tag, result, sizeof(fp->tag));
340 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
341 s32 end_new, s32 curr, const bool probe_pass)
343 const s64 imm_min = S32_MIN, imm_max = S32_MAX;
344 s32 delta = end_new - end_old;
347 if (curr < pos && curr + imm + 1 >= end_old)
349 else if (curr >= end_new && curr + imm + 1 < end_new)
351 if (imm < imm_min || imm > imm_max)
358 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
359 s32 end_new, s32 curr, const bool probe_pass)
361 const s32 off_min = S16_MIN, off_max = S16_MAX;
362 s32 delta = end_new - end_old;
365 if (curr < pos && curr + off + 1 >= end_old)
367 else if (curr >= end_new && curr + off + 1 < end_new)
369 if (off < off_min || off > off_max)
376 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
377 s32 end_new, const bool probe_pass)
379 u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
380 struct bpf_insn *insn = prog->insnsi;
383 for (i = 0; i < insn_cnt; i++, insn++) {
386 /* In the probing pass we still operate on the original,
387 * unpatched image in order to check overflows before we
388 * do any other adjustments. Therefore skip the patchlet.
390 if (probe_pass && i == pos) {
392 insn = prog->insnsi + end_old;
395 if ((BPF_CLASS(code) != BPF_JMP &&
396 BPF_CLASS(code) != BPF_JMP32) ||
397 BPF_OP(code) == BPF_EXIT)
399 /* Adjust offset of jmps if we cross patch boundaries. */
400 if (BPF_OP(code) == BPF_CALL) {
401 if (insn->src_reg != BPF_PSEUDO_CALL)
403 ret = bpf_adj_delta_to_imm(insn, pos, end_old,
404 end_new, i, probe_pass);
406 ret = bpf_adj_delta_to_off(insn, pos, end_old,
407 end_new, i, probe_pass);
416 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
418 struct bpf_line_info *linfo;
421 nr_linfo = prog->aux->nr_linfo;
422 if (!nr_linfo || !delta)
425 linfo = prog->aux->linfo;
427 for (i = 0; i < nr_linfo; i++)
428 if (off < linfo[i].insn_off)
431 /* Push all off < linfo[i].insn_off by delta */
432 for (; i < nr_linfo; i++)
433 linfo[i].insn_off += delta;
436 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
437 const struct bpf_insn *patch, u32 len)
439 u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
440 const u32 cnt_max = S16_MAX;
441 struct bpf_prog *prog_adj;
444 /* Since our patchlet doesn't expand the image, we're done. */
445 if (insn_delta == 0) {
446 memcpy(prog->insnsi + off, patch, sizeof(*patch));
450 insn_adj_cnt = prog->len + insn_delta;
452 /* Reject anything that would potentially let the insn->off
453 * target overflow when we have excessive program expansions.
454 * We need to probe here before we do any reallocation where
455 * we afterwards may not fail anymore.
457 if (insn_adj_cnt > cnt_max &&
458 (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
461 /* Several new instructions need to be inserted. Make room
462 * for them. Likely, there's no need for a new allocation as
463 * last page could have large enough tailroom.
465 prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
468 return ERR_PTR(-ENOMEM);
470 prog_adj->len = insn_adj_cnt;
472 /* Patching happens in 3 steps:
474 * 1) Move over tail of insnsi from next instruction onwards,
475 * so we can patch the single target insn with one or more
476 * new ones (patching is always from 1 to n insns, n > 0).
477 * 2) Inject new instructions at the target location.
478 * 3) Adjust branch offsets if necessary.
480 insn_rest = insn_adj_cnt - off - len;
482 memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
483 sizeof(*patch) * insn_rest);
484 memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
486 /* We are guaranteed to not fail at this point, otherwise
487 * the ship has sailed to reverse to the original state. An
488 * overflow cannot happen at this point.
490 BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
492 bpf_adj_linfo(prog_adj, off, insn_delta);
497 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
499 /* Branch offsets can't overflow when program is shrinking, no need
500 * to call bpf_adj_branches(..., true) here
502 memmove(prog->insnsi + off, prog->insnsi + off + cnt,
503 sizeof(struct bpf_insn) * (prog->len - off - cnt));
506 return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
509 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
513 for (i = 0; i < fp->aux->func_cnt; i++)
514 bpf_prog_kallsyms_del(fp->aux->func[i]);
517 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
519 bpf_prog_kallsyms_del_subprogs(fp);
520 bpf_prog_kallsyms_del(fp);
523 #ifdef CONFIG_BPF_JIT
524 /* All BPF JIT sysctl knobs here. */
525 int bpf_jit_enable __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
526 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
527 int bpf_jit_harden __read_mostly;
528 long bpf_jit_limit __read_mostly;
531 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
533 const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
534 unsigned long addr = (unsigned long)hdr;
536 WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
538 prog->aux->ksym.start = (unsigned long) prog->bpf_func;
539 prog->aux->ksym.end = addr + hdr->pages * PAGE_SIZE;
543 bpf_prog_ksym_set_name(struct bpf_prog *prog)
545 char *sym = prog->aux->ksym.name;
546 const char *end = sym + KSYM_NAME_LEN;
547 const struct btf_type *type;
548 const char *func_name;
550 BUILD_BUG_ON(sizeof("bpf_prog_") +
551 sizeof(prog->tag) * 2 +
552 /* name has been null terminated.
553 * We should need +1 for the '_' preceding
554 * the name. However, the null character
555 * is double counted between the name and the
556 * sizeof("bpf_prog_") above, so we omit
559 sizeof(prog->aux->name) > KSYM_NAME_LEN);
561 sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
562 sym = bin2hex(sym, prog->tag, sizeof(prog->tag));
564 /* prog->aux->name will be ignored if full btf name is available */
565 if (prog->aux->func_info_cnt) {
566 type = btf_type_by_id(prog->aux->btf,
567 prog->aux->func_info[prog->aux->func_idx].type_id);
568 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
569 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
573 if (prog->aux->name[0])
574 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
579 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
581 return container_of(n, struct bpf_ksym, tnode)->start;
584 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
585 struct latch_tree_node *b)
587 return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
590 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
592 unsigned long val = (unsigned long)key;
593 const struct bpf_ksym *ksym;
595 ksym = container_of(n, struct bpf_ksym, tnode);
597 if (val < ksym->start)
599 if (val >= ksym->end)
605 static const struct latch_tree_ops bpf_tree_ops = {
606 .less = bpf_tree_less,
607 .comp = bpf_tree_comp,
610 static DEFINE_SPINLOCK(bpf_lock);
611 static LIST_HEAD(bpf_kallsyms);
612 static struct latch_tree_root bpf_tree __cacheline_aligned;
614 void bpf_ksym_add(struct bpf_ksym *ksym)
616 spin_lock_bh(&bpf_lock);
617 WARN_ON_ONCE(!list_empty(&ksym->lnode));
618 list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
619 latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
620 spin_unlock_bh(&bpf_lock);
623 static void __bpf_ksym_del(struct bpf_ksym *ksym)
625 if (list_empty(&ksym->lnode))
628 latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
629 list_del_rcu(&ksym->lnode);
632 void bpf_ksym_del(struct bpf_ksym *ksym)
634 spin_lock_bh(&bpf_lock);
635 __bpf_ksym_del(ksym);
636 spin_unlock_bh(&bpf_lock);
639 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
641 return fp->jited && !bpf_prog_was_classic(fp);
644 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
646 return list_empty(&fp->aux->ksym.lnode) ||
647 fp->aux->ksym.lnode.prev == LIST_POISON2;
650 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
652 if (!bpf_prog_kallsyms_candidate(fp) ||
656 bpf_prog_ksym_set_addr(fp);
657 bpf_prog_ksym_set_name(fp);
658 fp->aux->ksym.prog = true;
660 bpf_ksym_add(&fp->aux->ksym);
663 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
665 if (!bpf_prog_kallsyms_candidate(fp))
668 bpf_ksym_del(&fp->aux->ksym);
671 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
673 struct latch_tree_node *n;
675 n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
676 return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
679 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
680 unsigned long *off, char *sym)
682 struct bpf_ksym *ksym;
686 ksym = bpf_ksym_find(addr);
688 unsigned long symbol_start = ksym->start;
689 unsigned long symbol_end = ksym->end;
691 strncpy(sym, ksym->name, KSYM_NAME_LEN);
695 *size = symbol_end - symbol_start;
697 *off = addr - symbol_start;
704 bool is_bpf_text_address(unsigned long addr)
709 ret = bpf_ksym_find(addr) != NULL;
715 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
717 struct bpf_ksym *ksym = bpf_ksym_find(addr);
719 return ksym && ksym->prog ?
720 container_of(ksym, struct bpf_prog_aux, ksym)->prog :
724 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
726 const struct exception_table_entry *e = NULL;
727 struct bpf_prog *prog;
730 prog = bpf_prog_ksym_find(addr);
733 if (!prog->aux->num_exentries)
736 e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
742 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
745 struct bpf_ksym *ksym;
749 if (!bpf_jit_kallsyms_enabled())
753 list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
757 strncpy(sym, ksym->name, KSYM_NAME_LEN);
759 *value = ksym->start;
760 *type = BPF_SYM_ELF_TYPE;
770 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
771 struct bpf_jit_poke_descriptor *poke)
773 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
774 static const u32 poke_tab_max = 1024;
775 u32 slot = prog->aux->size_poke_tab;
778 if (size > poke_tab_max)
780 if (poke->tailcall_target || poke->tailcall_target_stable ||
781 poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
784 switch (poke->reason) {
785 case BPF_POKE_REASON_TAIL_CALL:
786 if (!poke->tail_call.map)
793 tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
797 memcpy(&tab[slot], poke, sizeof(*poke));
798 prog->aux->size_poke_tab = size;
799 prog->aux->poke_tab = tab;
804 static atomic_long_t bpf_jit_current;
806 /* Can be overridden by an arch's JIT compiler if it has a custom,
807 * dedicated BPF backend memory area, or if neither of the two
810 u64 __weak bpf_jit_alloc_exec_limit(void)
812 #if defined(MODULES_VADDR)
813 return MODULES_END - MODULES_VADDR;
815 return VMALLOC_END - VMALLOC_START;
819 static int __init bpf_jit_charge_init(void)
821 /* Only used as heuristic here to derive limit. */
822 bpf_jit_limit = min_t(u64, round_up(bpf_jit_alloc_exec_limit() >> 2,
823 PAGE_SIZE), LONG_MAX);
826 pure_initcall(bpf_jit_charge_init);
828 static int bpf_jit_charge_modmem(u32 pages)
830 if (atomic_long_add_return(pages, &bpf_jit_current) >
831 (bpf_jit_limit >> PAGE_SHIFT)) {
832 if (!capable(CAP_SYS_ADMIN)) {
833 atomic_long_sub(pages, &bpf_jit_current);
841 static void bpf_jit_uncharge_modmem(u32 pages)
843 atomic_long_sub(pages, &bpf_jit_current);
846 void *__weak bpf_jit_alloc_exec(unsigned long size)
848 return module_alloc(size);
851 void __weak bpf_jit_free_exec(void *addr)
853 module_memfree(addr);
856 struct bpf_binary_header *
857 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
858 unsigned int alignment,
859 bpf_jit_fill_hole_t bpf_fill_ill_insns)
861 struct bpf_binary_header *hdr;
862 u32 size, hole, start, pages;
864 WARN_ON_ONCE(!is_power_of_2(alignment) ||
865 alignment > BPF_IMAGE_ALIGNMENT);
867 /* Most of BPF filters are really small, but if some of them
868 * fill a page, allow at least 128 extra bytes to insert a
869 * random section of illegal instructions.
871 size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
872 pages = size / PAGE_SIZE;
874 if (bpf_jit_charge_modmem(pages))
876 hdr = bpf_jit_alloc_exec(size);
878 bpf_jit_uncharge_modmem(pages);
882 /* Fill space with illegal/arch-dep instructions. */
883 bpf_fill_ill_insns(hdr, size);
886 hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
887 PAGE_SIZE - sizeof(*hdr));
888 start = (get_random_int() % hole) & ~(alignment - 1);
890 /* Leave a random number of instructions before BPF code. */
891 *image_ptr = &hdr->image[start];
896 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
898 u32 pages = hdr->pages;
900 bpf_jit_free_exec(hdr);
901 bpf_jit_uncharge_modmem(pages);
904 /* This symbol is only overridden by archs that have different
905 * requirements than the usual eBPF JITs, f.e. when they only
906 * implement cBPF JIT, do not set images read-only, etc.
908 void __weak bpf_jit_free(struct bpf_prog *fp)
911 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
913 bpf_jit_binary_free(hdr);
915 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
918 bpf_prog_unlock_free(fp);
921 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
922 const struct bpf_insn *insn, bool extra_pass,
923 u64 *func_addr, bool *func_addr_fixed)
929 *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
930 if (!*func_addr_fixed) {
931 /* Place-holder address till the last pass has collected
932 * all addresses for JITed subprograms in which case we
933 * can pick them up from prog->aux.
937 else if (prog->aux->func &&
938 off >= 0 && off < prog->aux->func_cnt)
939 addr = (u8 *)prog->aux->func[off]->bpf_func;
943 /* Address of a BPF helper call. Since part of the core
944 * kernel, it's always at a fixed location. __bpf_call_base
945 * and the helper with imm relative to it are both in core
948 addr = (u8 *)__bpf_call_base + imm;
951 *func_addr = (unsigned long)addr;
955 static int bpf_jit_blind_insn(const struct bpf_insn *from,
956 const struct bpf_insn *aux,
957 struct bpf_insn *to_buff,
960 struct bpf_insn *to = to_buff;
961 u32 imm_rnd = get_random_int();
964 BUILD_BUG_ON(BPF_REG_AX + 1 != MAX_BPF_JIT_REG);
965 BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
967 /* Constraints on AX register:
969 * AX register is inaccessible from user space. It is mapped in
970 * all JITs, and used here for constant blinding rewrites. It is
971 * typically "stateless" meaning its contents are only valid within
972 * the executed instruction, but not across several instructions.
973 * There are a few exceptions however which are further detailed
976 * Constant blinding is only used by JITs, not in the interpreter.
977 * The interpreter uses AX in some occasions as a local temporary
978 * register e.g. in DIV or MOD instructions.
980 * In restricted circumstances, the verifier can also use the AX
981 * register for rewrites as long as they do not interfere with
984 if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
987 if (from->imm == 0 &&
988 (from->code == (BPF_ALU | BPF_MOV | BPF_K) ||
989 from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
990 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
994 switch (from->code) {
995 case BPF_ALU | BPF_ADD | BPF_K:
996 case BPF_ALU | BPF_SUB | BPF_K:
997 case BPF_ALU | BPF_AND | BPF_K:
998 case BPF_ALU | BPF_OR | BPF_K:
999 case BPF_ALU | BPF_XOR | BPF_K:
1000 case BPF_ALU | BPF_MUL | BPF_K:
1001 case BPF_ALU | BPF_MOV | BPF_K:
1002 case BPF_ALU | BPF_DIV | BPF_K:
1003 case BPF_ALU | BPF_MOD | BPF_K:
1004 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1005 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1006 *to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
1009 case BPF_ALU64 | BPF_ADD | BPF_K:
1010 case BPF_ALU64 | BPF_SUB | BPF_K:
1011 case BPF_ALU64 | BPF_AND | BPF_K:
1012 case BPF_ALU64 | BPF_OR | BPF_K:
1013 case BPF_ALU64 | BPF_XOR | BPF_K:
1014 case BPF_ALU64 | BPF_MUL | BPF_K:
1015 case BPF_ALU64 | BPF_MOV | BPF_K:
1016 case BPF_ALU64 | BPF_DIV | BPF_K:
1017 case BPF_ALU64 | BPF_MOD | BPF_K:
1018 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1019 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1020 *to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
1023 case BPF_JMP | BPF_JEQ | BPF_K:
1024 case BPF_JMP | BPF_JNE | BPF_K:
1025 case BPF_JMP | BPF_JGT | BPF_K:
1026 case BPF_JMP | BPF_JLT | BPF_K:
1027 case BPF_JMP | BPF_JGE | BPF_K:
1028 case BPF_JMP | BPF_JLE | BPF_K:
1029 case BPF_JMP | BPF_JSGT | BPF_K:
1030 case BPF_JMP | BPF_JSLT | BPF_K:
1031 case BPF_JMP | BPF_JSGE | BPF_K:
1032 case BPF_JMP | BPF_JSLE | BPF_K:
1033 case BPF_JMP | BPF_JSET | BPF_K:
1034 /* Accommodate for extra offset in case of a backjump. */
1038 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1039 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1040 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1043 case BPF_JMP32 | BPF_JEQ | BPF_K:
1044 case BPF_JMP32 | BPF_JNE | BPF_K:
1045 case BPF_JMP32 | BPF_JGT | BPF_K:
1046 case BPF_JMP32 | BPF_JLT | BPF_K:
1047 case BPF_JMP32 | BPF_JGE | BPF_K:
1048 case BPF_JMP32 | BPF_JLE | BPF_K:
1049 case BPF_JMP32 | BPF_JSGT | BPF_K:
1050 case BPF_JMP32 | BPF_JSLT | BPF_K:
1051 case BPF_JMP32 | BPF_JSGE | BPF_K:
1052 case BPF_JMP32 | BPF_JSLE | BPF_K:
1053 case BPF_JMP32 | BPF_JSET | BPF_K:
1054 /* Accommodate for extra offset in case of a backjump. */
1058 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1059 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1060 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1064 case BPF_LD | BPF_IMM | BPF_DW:
1065 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1066 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1067 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1068 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1070 case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1071 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1072 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1074 *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1075 *to++ = BPF_ALU64_REG(BPF_OR, aux[0].dst_reg, BPF_REG_AX);
1078 case BPF_ST | BPF_MEM | BPF_DW:
1079 case BPF_ST | BPF_MEM | BPF_W:
1080 case BPF_ST | BPF_MEM | BPF_H:
1081 case BPF_ST | BPF_MEM | BPF_B:
1082 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1083 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1084 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1088 return to - to_buff;
1091 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1092 gfp_t gfp_extra_flags)
1094 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1095 struct bpf_prog *fp;
1097 fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1099 /* aux->prog still points to the fp_other one, so
1100 * when promoting the clone to the real program,
1101 * this still needs to be adapted.
1103 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1109 static void bpf_prog_clone_free(struct bpf_prog *fp)
1111 /* aux was stolen by the other clone, so we cannot free
1112 * it from this path! It will be freed eventually by the
1113 * other program on release.
1115 * At this point, we don't need a deferred release since
1116 * clone is guaranteed to not be locked.
1119 __bpf_prog_free(fp);
1122 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1124 /* We have to repoint aux->prog to self, as we don't
1125 * know whether fp here is the clone or the original.
1128 bpf_prog_clone_free(fp_other);
1131 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1133 struct bpf_insn insn_buff[16], aux[2];
1134 struct bpf_prog *clone, *tmp;
1135 int insn_delta, insn_cnt;
1136 struct bpf_insn *insn;
1139 if (!bpf_jit_blinding_enabled(prog) || prog->blinded)
1142 clone = bpf_prog_clone_create(prog, GFP_USER);
1144 return ERR_PTR(-ENOMEM);
1146 insn_cnt = clone->len;
1147 insn = clone->insnsi;
1149 for (i = 0; i < insn_cnt; i++, insn++) {
1150 /* We temporarily need to hold the original ld64 insn
1151 * so that we can still access the first part in the
1152 * second blinding run.
1154 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1156 memcpy(aux, insn, sizeof(aux));
1158 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1159 clone->aux->verifier_zext);
1163 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1165 /* Patching may have repointed aux->prog during
1166 * realloc from the original one, so we need to
1167 * fix it up here on error.
1169 bpf_jit_prog_release_other(prog, clone);
1174 insn_delta = rewritten - 1;
1176 /* Walk new program and skip insns we just inserted. */
1177 insn = clone->insnsi + i + insn_delta;
1178 insn_cnt += insn_delta;
1185 #endif /* CONFIG_BPF_JIT */
1187 /* Base function for offset calculation. Needs to go into .text section,
1188 * therefore keeping it non-static as well; will also be used by JITs
1189 * anyway later on, so do not let the compiler omit it. This also needs
1190 * to go into kallsyms for correlation from e.g. bpftool, so naming
1193 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1197 EXPORT_SYMBOL_GPL(__bpf_call_base);
1199 /* All UAPI available opcodes. */
1200 #define BPF_INSN_MAP(INSN_2, INSN_3) \
1201 /* 32 bit ALU operations. */ \
1202 /* Register based. */ \
1203 INSN_3(ALU, ADD, X), \
1204 INSN_3(ALU, SUB, X), \
1205 INSN_3(ALU, AND, X), \
1206 INSN_3(ALU, OR, X), \
1207 INSN_3(ALU, LSH, X), \
1208 INSN_3(ALU, RSH, X), \
1209 INSN_3(ALU, XOR, X), \
1210 INSN_3(ALU, MUL, X), \
1211 INSN_3(ALU, MOV, X), \
1212 INSN_3(ALU, ARSH, X), \
1213 INSN_3(ALU, DIV, X), \
1214 INSN_3(ALU, MOD, X), \
1216 INSN_3(ALU, END, TO_BE), \
1217 INSN_3(ALU, END, TO_LE), \
1218 /* Immediate based. */ \
1219 INSN_3(ALU, ADD, K), \
1220 INSN_3(ALU, SUB, K), \
1221 INSN_3(ALU, AND, K), \
1222 INSN_3(ALU, OR, K), \
1223 INSN_3(ALU, LSH, K), \
1224 INSN_3(ALU, RSH, K), \
1225 INSN_3(ALU, XOR, K), \
1226 INSN_3(ALU, MUL, K), \
1227 INSN_3(ALU, MOV, K), \
1228 INSN_3(ALU, ARSH, K), \
1229 INSN_3(ALU, DIV, K), \
1230 INSN_3(ALU, MOD, K), \
1231 /* 64 bit ALU operations. */ \
1232 /* Register based. */ \
1233 INSN_3(ALU64, ADD, X), \
1234 INSN_3(ALU64, SUB, X), \
1235 INSN_3(ALU64, AND, X), \
1236 INSN_3(ALU64, OR, X), \
1237 INSN_3(ALU64, LSH, X), \
1238 INSN_3(ALU64, RSH, X), \
1239 INSN_3(ALU64, XOR, X), \
1240 INSN_3(ALU64, MUL, X), \
1241 INSN_3(ALU64, MOV, X), \
1242 INSN_3(ALU64, ARSH, X), \
1243 INSN_3(ALU64, DIV, X), \
1244 INSN_3(ALU64, MOD, X), \
1245 INSN_2(ALU64, NEG), \
1246 /* Immediate based. */ \
1247 INSN_3(ALU64, ADD, K), \
1248 INSN_3(ALU64, SUB, K), \
1249 INSN_3(ALU64, AND, K), \
1250 INSN_3(ALU64, OR, K), \
1251 INSN_3(ALU64, LSH, K), \
1252 INSN_3(ALU64, RSH, K), \
1253 INSN_3(ALU64, XOR, K), \
1254 INSN_3(ALU64, MUL, K), \
1255 INSN_3(ALU64, MOV, K), \
1256 INSN_3(ALU64, ARSH, K), \
1257 INSN_3(ALU64, DIV, K), \
1258 INSN_3(ALU64, MOD, K), \
1259 /* Call instruction. */ \
1260 INSN_2(JMP, CALL), \
1261 /* Exit instruction. */ \
1262 INSN_2(JMP, EXIT), \
1263 /* 32-bit Jump instructions. */ \
1264 /* Register based. */ \
1265 INSN_3(JMP32, JEQ, X), \
1266 INSN_3(JMP32, JNE, X), \
1267 INSN_3(JMP32, JGT, X), \
1268 INSN_3(JMP32, JLT, X), \
1269 INSN_3(JMP32, JGE, X), \
1270 INSN_3(JMP32, JLE, X), \
1271 INSN_3(JMP32, JSGT, X), \
1272 INSN_3(JMP32, JSLT, X), \
1273 INSN_3(JMP32, JSGE, X), \
1274 INSN_3(JMP32, JSLE, X), \
1275 INSN_3(JMP32, JSET, X), \
1276 /* Immediate based. */ \
1277 INSN_3(JMP32, JEQ, K), \
1278 INSN_3(JMP32, JNE, K), \
1279 INSN_3(JMP32, JGT, K), \
1280 INSN_3(JMP32, JLT, K), \
1281 INSN_3(JMP32, JGE, K), \
1282 INSN_3(JMP32, JLE, K), \
1283 INSN_3(JMP32, JSGT, K), \
1284 INSN_3(JMP32, JSLT, K), \
1285 INSN_3(JMP32, JSGE, K), \
1286 INSN_3(JMP32, JSLE, K), \
1287 INSN_3(JMP32, JSET, K), \
1288 /* Jump instructions. */ \
1289 /* Register based. */ \
1290 INSN_3(JMP, JEQ, X), \
1291 INSN_3(JMP, JNE, X), \
1292 INSN_3(JMP, JGT, X), \
1293 INSN_3(JMP, JLT, X), \
1294 INSN_3(JMP, JGE, X), \
1295 INSN_3(JMP, JLE, X), \
1296 INSN_3(JMP, JSGT, X), \
1297 INSN_3(JMP, JSLT, X), \
1298 INSN_3(JMP, JSGE, X), \
1299 INSN_3(JMP, JSLE, X), \
1300 INSN_3(JMP, JSET, X), \
1301 /* Immediate based. */ \
1302 INSN_3(JMP, JEQ, K), \
1303 INSN_3(JMP, JNE, K), \
1304 INSN_3(JMP, JGT, K), \
1305 INSN_3(JMP, JLT, K), \
1306 INSN_3(JMP, JGE, K), \
1307 INSN_3(JMP, JLE, K), \
1308 INSN_3(JMP, JSGT, K), \
1309 INSN_3(JMP, JSLT, K), \
1310 INSN_3(JMP, JSGE, K), \
1311 INSN_3(JMP, JSLE, K), \
1312 INSN_3(JMP, JSET, K), \
1314 /* Store instructions. */ \
1315 /* Register based. */ \
1316 INSN_3(STX, MEM, B), \
1317 INSN_3(STX, MEM, H), \
1318 INSN_3(STX, MEM, W), \
1319 INSN_3(STX, MEM, DW), \
1320 INSN_3(STX, XADD, W), \
1321 INSN_3(STX, XADD, DW), \
1322 /* Immediate based. */ \
1323 INSN_3(ST, MEM, B), \
1324 INSN_3(ST, MEM, H), \
1325 INSN_3(ST, MEM, W), \
1326 INSN_3(ST, MEM, DW), \
1327 /* Load instructions. */ \
1328 /* Register based. */ \
1329 INSN_3(LDX, MEM, B), \
1330 INSN_3(LDX, MEM, H), \
1331 INSN_3(LDX, MEM, W), \
1332 INSN_3(LDX, MEM, DW), \
1333 /* Immediate based. */ \
1336 bool bpf_opcode_in_insntable(u8 code)
1338 #define BPF_INSN_2_TBL(x, y) [BPF_##x | BPF_##y] = true
1339 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1340 static const bool public_insntable[256] = {
1341 [0 ... 255] = false,
1342 /* Now overwrite non-defaults ... */
1343 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1344 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1345 [BPF_LD | BPF_ABS | BPF_B] = true,
1346 [BPF_LD | BPF_ABS | BPF_H] = true,
1347 [BPF_LD | BPF_ABS | BPF_W] = true,
1348 [BPF_LD | BPF_IND | BPF_B] = true,
1349 [BPF_LD | BPF_IND | BPF_H] = true,
1350 [BPF_LD | BPF_IND | BPF_W] = true,
1352 #undef BPF_INSN_3_TBL
1353 #undef BPF_INSN_2_TBL
1354 return public_insntable[code];
1357 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1358 u64 __weak bpf_probe_read_kernel(void *dst, u32 size, const void *unsafe_ptr)
1360 memset(dst, 0, size);
1365 * __bpf_prog_run - run eBPF program on a given context
1366 * @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1367 * @insn: is the array of eBPF instructions
1368 * @stack: is the eBPF storage stack
1370 * Decode and execute eBPF instructions.
1372 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, u64 *stack)
1374 #define BPF_INSN_2_LBL(x, y) [BPF_##x | BPF_##y] = &&x##_##y
1375 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1376 static const void * const jumptable[256] __annotate_jump_table = {
1377 [0 ... 255] = &&default_label,
1378 /* Now overwrite non-defaults ... */
1379 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1380 /* Non-UAPI available opcodes. */
1381 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1382 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1383 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1384 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1385 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1386 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1388 #undef BPF_INSN_3_LBL
1389 #undef BPF_INSN_2_LBL
1390 u32 tail_call_cnt = 0;
1392 #define CONT ({ insn++; goto select_insn; })
1393 #define CONT_JMP ({ insn++; goto select_insn; })
1396 goto *jumptable[insn->code];
1399 #define ALU(OPCODE, OP) \
1400 ALU64_##OPCODE##_X: \
1404 DST = (u32) DST OP (u32) SRC; \
1406 ALU64_##OPCODE##_K: \
1410 DST = (u32) DST OP (u32) IMM; \
1441 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1445 DST = (u64) (u32) (((s32) DST) >> SRC);
1448 DST = (u64) (u32) (((s32) DST) >> IMM);
1451 (*(s64 *) &DST) >>= SRC;
1454 (*(s64 *) &DST) >>= IMM;
1457 div64_u64_rem(DST, SRC, &AX);
1462 DST = do_div(AX, (u32) SRC);
1465 div64_u64_rem(DST, IMM, &AX);
1470 DST = do_div(AX, (u32) IMM);
1473 DST = div64_u64(DST, SRC);
1477 do_div(AX, (u32) SRC);
1481 DST = div64_u64(DST, IMM);
1485 do_div(AX, (u32) IMM);
1491 DST = (__force u16) cpu_to_be16(DST);
1494 DST = (__force u32) cpu_to_be32(DST);
1497 DST = (__force u64) cpu_to_be64(DST);
1504 DST = (__force u16) cpu_to_le16(DST);
1507 DST = (__force u32) cpu_to_le32(DST);
1510 DST = (__force u64) cpu_to_le64(DST);
1517 /* Function call scratches BPF_R1-BPF_R5 registers,
1518 * preserves BPF_R6-BPF_R9, and stores return value
1521 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1526 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1529 insn + insn->off + 1);
1533 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1534 struct bpf_array *array = container_of(map, struct bpf_array, map);
1535 struct bpf_prog *prog;
1538 if (unlikely(index >= array->map.max_entries))
1540 if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1545 prog = READ_ONCE(array->ptrs[index]);
1549 /* ARG1 at this point is guaranteed to point to CTX from
1550 * the verifier side due to the fact that the tail call is
1551 * handled like a helper, that is, bpf_tail_call_proto,
1552 * where arg1_type is ARG_PTR_TO_CTX.
1554 insn = prog->insnsi;
1565 #define COND_JMP(SIGN, OPCODE, CMP_OP) \
1567 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) { \
1568 insn += insn->off; \
1572 JMP32_##OPCODE##_X: \
1573 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) { \
1574 insn += insn->off; \
1579 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) { \
1580 insn += insn->off; \
1584 JMP32_##OPCODE##_K: \
1585 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) { \
1586 insn += insn->off; \
1590 COND_JMP(u, JEQ, ==)
1591 COND_JMP(u, JNE, !=)
1594 COND_JMP(u, JGE, >=)
1595 COND_JMP(u, JLE, <=)
1596 COND_JMP(u, JSET, &)
1597 COND_JMP(s, JSGT, >)
1598 COND_JMP(s, JSLT, <)
1599 COND_JMP(s, JSGE, >=)
1600 COND_JMP(s, JSLE, <=)
1602 /* STX and ST and LDX*/
1603 #define LDST(SIZEOP, SIZE) \
1605 *(SIZE *)(unsigned long) (DST + insn->off) = SRC; \
1608 *(SIZE *)(unsigned long) (DST + insn->off) = IMM; \
1611 DST = *(SIZE *)(unsigned long) (SRC + insn->off); \
1619 #define LDX_PROBE(SIZEOP, SIZE) \
1620 LDX_PROBE_MEM_##SIZEOP: \
1621 bpf_probe_read_kernel(&DST, SIZE, (const void *)(long) (SRC + insn->off)); \
1629 STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
1630 atomic_add((u32) SRC, (atomic_t *)(unsigned long)
1633 STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
1634 atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
1639 /* If we ever reach this, we have a bug somewhere. Die hard here
1640 * instead of just returning 0; we could be somewhere in a subprog,
1641 * so execution could continue otherwise which we do /not/ want.
1643 * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
1645 pr_warn("BPF interpreter: unknown opcode %02x\n", insn->code);
1650 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
1651 #define DEFINE_BPF_PROG_RUN(stack_size) \
1652 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
1654 u64 stack[stack_size / sizeof(u64)]; \
1655 u64 regs[MAX_BPF_EXT_REG]; \
1657 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1658 ARG1 = (u64) (unsigned long) ctx; \
1659 return ___bpf_prog_run(regs, insn, stack); \
1662 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
1663 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
1664 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
1665 const struct bpf_insn *insn) \
1667 u64 stack[stack_size / sizeof(u64)]; \
1668 u64 regs[MAX_BPF_EXT_REG]; \
1670 FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1676 return ___bpf_prog_run(regs, insn, stack); \
1679 #define EVAL1(FN, X) FN(X)
1680 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
1681 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
1682 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
1683 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
1684 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
1686 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
1687 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
1688 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
1690 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
1691 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
1692 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
1694 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
1696 static unsigned int (*interpreters[])(const void *ctx,
1697 const struct bpf_insn *insn) = {
1698 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1699 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1700 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1702 #undef PROG_NAME_LIST
1703 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
1704 static u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
1705 const struct bpf_insn *insn) = {
1706 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1707 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1708 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1710 #undef PROG_NAME_LIST
1712 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
1714 stack_depth = max_t(u32, stack_depth, 1);
1715 insn->off = (s16) insn->imm;
1716 insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
1717 __bpf_call_base_args;
1718 insn->code = BPF_JMP | BPF_CALL_ARGS;
1722 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
1723 const struct bpf_insn *insn)
1725 /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
1726 * is not working properly, so warn about it!
1733 bool bpf_prog_array_compatible(struct bpf_array *array,
1734 const struct bpf_prog *fp)
1736 if (fp->kprobe_override)
1739 if (!array->aux->type) {
1740 /* There's no owner yet where we could check for
1743 array->aux->type = fp->type;
1744 array->aux->jited = fp->jited;
1748 return array->aux->type == fp->type &&
1749 array->aux->jited == fp->jited;
1752 static int bpf_check_tail_call(const struct bpf_prog *fp)
1754 struct bpf_prog_aux *aux = fp->aux;
1757 mutex_lock(&aux->used_maps_mutex);
1758 for (i = 0; i < aux->used_map_cnt; i++) {
1759 struct bpf_map *map = aux->used_maps[i];
1760 struct bpf_array *array;
1762 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1765 array = container_of(map, struct bpf_array, map);
1766 if (!bpf_prog_array_compatible(array, fp)) {
1773 mutex_unlock(&aux->used_maps_mutex);
1777 static void bpf_prog_select_func(struct bpf_prog *fp)
1779 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1780 u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
1782 fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
1784 fp->bpf_func = __bpf_prog_ret0_warn;
1789 * bpf_prog_select_runtime - select exec runtime for BPF program
1790 * @fp: bpf_prog populated with internal BPF program
1791 * @err: pointer to error variable
1793 * Try to JIT eBPF program, if JIT is not available, use interpreter.
1794 * The BPF program will be executed via BPF_PROG_RUN() macro.
1796 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1798 /* In case of BPF to BPF calls, verifier did all the prep
1799 * work with regards to JITing, etc.
1804 bpf_prog_select_func(fp);
1806 /* eBPF JITs can rewrite the program in case constant
1807 * blinding is active. However, in case of error during
1808 * blinding, bpf_int_jit_compile() must always return a
1809 * valid program, which in this case would simply not
1810 * be JITed, but falls back to the interpreter.
1812 if (!bpf_prog_is_dev_bound(fp->aux)) {
1813 *err = bpf_prog_alloc_jited_linfo(fp);
1817 fp = bpf_int_jit_compile(fp);
1819 bpf_prog_free_jited_linfo(fp);
1820 #ifdef CONFIG_BPF_JIT_ALWAYS_ON
1825 bpf_prog_free_unused_jited_linfo(fp);
1828 *err = bpf_prog_offload_compile(fp);
1834 bpf_prog_lock_ro(fp);
1836 /* The tail call compatibility check can only be done at
1837 * this late stage as we need to determine, if we deal
1838 * with JITed or non JITed program concatenations and not
1839 * all eBPF JITs might immediately support all features.
1841 *err = bpf_check_tail_call(fp);
1845 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1847 static unsigned int __bpf_prog_ret1(const void *ctx,
1848 const struct bpf_insn *insn)
1853 static struct bpf_prog_dummy {
1854 struct bpf_prog prog;
1855 } dummy_bpf_prog = {
1857 .bpf_func = __bpf_prog_ret1,
1861 /* to avoid allocating empty bpf_prog_array for cgroups that
1862 * don't have bpf program attached use one global 'empty_prog_array'
1863 * It will not be modified the caller of bpf_prog_array_alloc()
1864 * (since caller requested prog_cnt == 0)
1865 * that pointer should be 'freed' by bpf_prog_array_free()
1868 struct bpf_prog_array hdr;
1869 struct bpf_prog *null_prog;
1870 } empty_prog_array = {
1874 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
1877 return kzalloc(sizeof(struct bpf_prog_array) +
1878 sizeof(struct bpf_prog_array_item) *
1882 return &empty_prog_array.hdr;
1885 void bpf_prog_array_free(struct bpf_prog_array *progs)
1887 if (!progs || progs == &empty_prog_array.hdr)
1889 kfree_rcu(progs, rcu);
1892 int bpf_prog_array_length(struct bpf_prog_array *array)
1894 struct bpf_prog_array_item *item;
1897 for (item = array->items; item->prog; item++)
1898 if (item->prog != &dummy_bpf_prog.prog)
1903 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
1905 struct bpf_prog_array_item *item;
1907 for (item = array->items; item->prog; item++)
1908 if (item->prog != &dummy_bpf_prog.prog)
1913 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
1917 struct bpf_prog_array_item *item;
1920 for (item = array->items; item->prog; item++) {
1921 if (item->prog == &dummy_bpf_prog.prog)
1923 prog_ids[i] = item->prog->aux->id;
1924 if (++i == request_cnt) {
1930 return !!(item->prog);
1933 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
1934 __u32 __user *prog_ids, u32 cnt)
1936 unsigned long err = 0;
1940 /* users of this function are doing:
1941 * cnt = bpf_prog_array_length();
1943 * bpf_prog_array_copy_to_user(..., cnt);
1944 * so below kcalloc doesn't need extra cnt > 0 check.
1946 ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
1949 nospc = bpf_prog_array_copy_core(array, ids, cnt);
1950 err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
1959 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
1960 struct bpf_prog *old_prog)
1962 struct bpf_prog_array_item *item;
1964 for (item = array->items; item->prog; item++)
1965 if (item->prog == old_prog) {
1966 WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
1972 * bpf_prog_array_delete_safe_at() - Replaces the program at the given
1973 * index into the program array with
1974 * a dummy no-op program.
1975 * @array: a bpf_prog_array
1976 * @index: the index of the program to replace
1978 * Skips over dummy programs, by not counting them, when calculating
1979 * the position of the program to replace.
1983 * * -EINVAL - Invalid index value. Must be a non-negative integer.
1984 * * -ENOENT - Index out of range
1986 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
1988 return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
1992 * bpf_prog_array_update_at() - Updates the program at the given index
1993 * into the program array.
1994 * @array: a bpf_prog_array
1995 * @index: the index of the program to update
1996 * @prog: the program to insert into the array
1998 * Skips over dummy programs, by not counting them, when calculating
1999 * the position of the program to update.
2003 * * -EINVAL - Invalid index value. Must be a non-negative integer.
2004 * * -ENOENT - Index out of range
2006 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2007 struct bpf_prog *prog)
2009 struct bpf_prog_array_item *item;
2011 if (unlikely(index < 0))
2014 for (item = array->items; item->prog; item++) {
2015 if (item->prog == &dummy_bpf_prog.prog)
2018 WRITE_ONCE(item->prog, prog);
2026 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2027 struct bpf_prog *exclude_prog,
2028 struct bpf_prog *include_prog,
2029 struct bpf_prog_array **new_array)
2031 int new_prog_cnt, carry_prog_cnt = 0;
2032 struct bpf_prog_array_item *existing;
2033 struct bpf_prog_array *array;
2034 bool found_exclude = false;
2035 int new_prog_idx = 0;
2037 /* Figure out how many existing progs we need to carry over to
2041 existing = old_array->items;
2042 for (; existing->prog; existing++) {
2043 if (existing->prog == exclude_prog) {
2044 found_exclude = true;
2047 if (existing->prog != &dummy_bpf_prog.prog)
2049 if (existing->prog == include_prog)
2054 if (exclude_prog && !found_exclude)
2057 /* How many progs (not NULL) will be in the new array? */
2058 new_prog_cnt = carry_prog_cnt;
2062 /* Do we have any prog (not NULL) in the new array? */
2063 if (!new_prog_cnt) {
2068 /* +1 as the end of prog_array is marked with NULL */
2069 array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2073 /* Fill in the new prog array */
2074 if (carry_prog_cnt) {
2075 existing = old_array->items;
2076 for (; existing->prog; existing++)
2077 if (existing->prog != exclude_prog &&
2078 existing->prog != &dummy_bpf_prog.prog) {
2079 array->items[new_prog_idx++].prog =
2084 array->items[new_prog_idx++].prog = include_prog;
2085 array->items[new_prog_idx].prog = NULL;
2090 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2091 u32 *prog_ids, u32 request_cnt,
2097 cnt = bpf_prog_array_length(array);
2101 /* return early if user requested only program count or nothing to copy */
2102 if (!request_cnt || !cnt)
2105 /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2106 return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2110 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2111 struct bpf_map **used_maps, u32 len)
2113 struct bpf_map *map;
2116 for (i = 0; i < len; i++) {
2118 if (map->ops->map_poke_untrack)
2119 map->ops->map_poke_untrack(map, aux);
2124 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2126 __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2127 kfree(aux->used_maps);
2130 static void bpf_prog_free_deferred(struct work_struct *work)
2132 struct bpf_prog_aux *aux;
2135 aux = container_of(work, struct bpf_prog_aux, work);
2136 bpf_free_used_maps(aux);
2137 if (bpf_prog_is_dev_bound(aux))
2138 bpf_prog_offload_destroy(aux->prog);
2139 #ifdef CONFIG_PERF_EVENTS
2140 if (aux->prog->has_callchain_buf)
2141 put_callchain_buffers();
2143 if (aux->dst_trampoline)
2144 bpf_trampoline_put(aux->dst_trampoline);
2145 for (i = 0; i < aux->func_cnt; i++)
2146 bpf_jit_free(aux->func[i]);
2147 if (aux->func_cnt) {
2149 bpf_prog_unlock_free(aux->prog);
2151 bpf_jit_free(aux->prog);
2155 /* Free internal BPF program */
2156 void bpf_prog_free(struct bpf_prog *fp)
2158 struct bpf_prog_aux *aux = fp->aux;
2161 bpf_prog_put(aux->dst_prog);
2162 INIT_WORK(&aux->work, bpf_prog_free_deferred);
2163 schedule_work(&aux->work);
2165 EXPORT_SYMBOL_GPL(bpf_prog_free);
2167 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2168 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2170 void bpf_user_rnd_init_once(void)
2172 prandom_init_once(&bpf_user_rnd_state);
2175 BPF_CALL_0(bpf_user_rnd_u32)
2177 /* Should someone ever have the rather unwise idea to use some
2178 * of the registers passed into this function, then note that
2179 * this function is called from native eBPF and classic-to-eBPF
2180 * transformations. Register assignments from both sides are
2181 * different, f.e. classic always sets fn(ctx, A, X) here.
2183 struct rnd_state *state;
2186 state = &get_cpu_var(bpf_user_rnd_state);
2187 res = prandom_u32_state(state);
2188 put_cpu_var(bpf_user_rnd_state);
2193 BPF_CALL_0(bpf_get_raw_cpu_id)
2195 return raw_smp_processor_id();
2198 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2199 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2200 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2201 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2202 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2203 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2204 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2205 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2206 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2207 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2209 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2210 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2211 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2212 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2213 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2215 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2216 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2217 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2218 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2219 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2220 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2221 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2222 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2223 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2225 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2231 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2232 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2236 EXPORT_SYMBOL_GPL(bpf_event_output);
2238 /* Always built-in helper functions. */
2239 const struct bpf_func_proto bpf_tail_call_proto = {
2242 .ret_type = RET_VOID,
2243 .arg1_type = ARG_PTR_TO_CTX,
2244 .arg2_type = ARG_CONST_MAP_PTR,
2245 .arg3_type = ARG_ANYTHING,
2248 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2249 * It is encouraged to implement bpf_int_jit_compile() instead, so that
2250 * eBPF and implicitly also cBPF can get JITed!
2252 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2257 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2258 * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2260 void __weak bpf_jit_compile(struct bpf_prog *prog)
2264 bool __weak bpf_helper_changes_pkt_data(void *func)
2269 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2270 * analysis code and wants explicit zero extension inserted by verifier.
2271 * Otherwise, return FALSE.
2273 bool __weak bpf_jit_needs_zext(void)
2278 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2279 * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2281 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2287 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2288 void *addr1, void *addr2)
2293 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2294 EXPORT_SYMBOL(bpf_stats_enabled_key);
2296 /* All definitions of tracepoints related to BPF. */
2297 #define CREATE_TRACE_POINTS
2298 #include <linux/bpf_trace.h>
2300 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2301 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);