2 * Linux Socket Filter - Kernel level socket filtering
4 * Based on the design of the Berkeley Packet Filter. The new
5 * internal format has been designed by PLUMgrid:
7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
11 * Jay Schulist <jschlst@samba.org>
12 * Alexei Starovoitov <ast@plumgrid.com>
13 * Daniel Borkmann <dborkman@redhat.com>
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
20 * Andi Kleen - Fix a few bad bugs and races.
21 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
24 #include <linux/module.h>
25 #include <linux/types.h>
27 #include <linux/fcntl.h>
28 #include <linux/socket.h>
29 #include <linux/sock_diag.h>
31 #include <linux/inet.h>
32 #include <linux/netdevice.h>
33 #include <linux/if_packet.h>
34 #include <linux/if_arp.h>
35 #include <linux/gfp.h>
37 #include <net/protocol.h>
38 #include <net/netlink.h>
39 #include <linux/skbuff.h>
41 #include <net/flow_dissector.h>
42 #include <linux/errno.h>
43 #include <linux/timer.h>
44 #include <linux/uaccess.h>
45 #include <asm/unaligned.h>
46 #include <linux/filter.h>
47 #include <linux/ratelimit.h>
48 #include <linux/seccomp.h>
49 #include <linux/if_vlan.h>
50 #include <linux/bpf.h>
51 #include <net/sch_generic.h>
52 #include <net/cls_cgroup.h>
53 #include <net/dst_metadata.h>
55 #include <net/sock_reuseport.h>
56 #include <net/busy_poll.h>
59 * sk_filter_trim_cap - run a packet through a socket filter
60 * @sk: sock associated with &sk_buff
61 * @skb: buffer to filter
62 * @cap: limit on how short the eBPF program may trim the packet
64 * Run the eBPF program and then cut skb->data to correct size returned by
65 * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
66 * than pkt_len we keep whole skb->data. This is the socket level
67 * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
68 * be accepted or -EPERM if the packet should be tossed.
71 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
74 struct sk_filter *filter;
77 * If the skb was allocated from pfmemalloc reserves, only
78 * allow SOCK_MEMALLOC sockets to use it as this socket is
81 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC)) {
82 NET_INC_STATS(sock_net(sk), LINUX_MIB_PFMEMALLOCDROP);
85 err = BPF_CGROUP_RUN_PROG_INET_INGRESS(sk, skb);
89 err = security_sock_rcv_skb(sk, skb);
94 filter = rcu_dereference(sk->sk_filter);
96 struct sock *save_sk = skb->sk;
100 pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
102 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
108 EXPORT_SYMBOL(sk_filter_trim_cap);
110 BPF_CALL_1(__skb_get_pay_offset, struct sk_buff *, skb)
112 return skb_get_poff(skb);
115 BPF_CALL_3(__skb_get_nlattr, struct sk_buff *, skb, u32, a, u32, x)
119 if (skb_is_nonlinear(skb))
122 if (skb->len < sizeof(struct nlattr))
125 if (a > skb->len - sizeof(struct nlattr))
128 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
130 return (void *) nla - (void *) skb->data;
135 BPF_CALL_3(__skb_get_nlattr_nest, struct sk_buff *, skb, u32, a, u32, x)
139 if (skb_is_nonlinear(skb))
142 if (skb->len < sizeof(struct nlattr))
145 if (a > skb->len - sizeof(struct nlattr))
148 nla = (struct nlattr *) &skb->data[a];
149 if (nla->nla_len > skb->len - a)
152 nla = nla_find_nested(nla, x);
154 return (void *) nla - (void *) skb->data;
159 BPF_CALL_0(__get_raw_cpu_id)
161 return raw_smp_processor_id();
164 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
165 .func = __get_raw_cpu_id,
167 .ret_type = RET_INTEGER,
170 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
171 struct bpf_insn *insn_buf)
173 struct bpf_insn *insn = insn_buf;
177 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
179 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
180 offsetof(struct sk_buff, mark));
184 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
185 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
186 #ifdef __BIG_ENDIAN_BITFIELD
187 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
192 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
194 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
195 offsetof(struct sk_buff, queue_mapping));
198 case SKF_AD_VLAN_TAG:
199 case SKF_AD_VLAN_TAG_PRESENT:
200 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
201 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
203 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
204 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
205 offsetof(struct sk_buff, vlan_tci));
206 if (skb_field == SKF_AD_VLAN_TAG) {
207 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
211 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
213 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
218 return insn - insn_buf;
221 static bool convert_bpf_extensions(struct sock_filter *fp,
222 struct bpf_insn **insnp)
224 struct bpf_insn *insn = *insnp;
228 case SKF_AD_OFF + SKF_AD_PROTOCOL:
229 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
231 /* A = *(u16 *) (CTX + offsetof(protocol)) */
232 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
233 offsetof(struct sk_buff, protocol));
234 /* A = ntohs(A) [emitting a nop or swap16] */
235 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
238 case SKF_AD_OFF + SKF_AD_PKTTYPE:
239 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
243 case SKF_AD_OFF + SKF_AD_IFINDEX:
244 case SKF_AD_OFF + SKF_AD_HATYPE:
245 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
246 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
248 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
249 BPF_REG_TMP, BPF_REG_CTX,
250 offsetof(struct sk_buff, dev));
251 /* if (tmp != 0) goto pc + 1 */
252 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
253 *insn++ = BPF_EXIT_INSN();
254 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
255 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
256 offsetof(struct net_device, ifindex));
258 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
259 offsetof(struct net_device, type));
262 case SKF_AD_OFF + SKF_AD_MARK:
263 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
267 case SKF_AD_OFF + SKF_AD_RXHASH:
268 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
270 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
271 offsetof(struct sk_buff, hash));
274 case SKF_AD_OFF + SKF_AD_QUEUE:
275 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
279 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
280 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
281 BPF_REG_A, BPF_REG_CTX, insn);
285 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
286 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
287 BPF_REG_A, BPF_REG_CTX, insn);
291 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
292 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
294 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
295 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
296 offsetof(struct sk_buff, vlan_proto));
297 /* A = ntohs(A) [emitting a nop or swap16] */
298 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
301 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
302 case SKF_AD_OFF + SKF_AD_NLATTR:
303 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
304 case SKF_AD_OFF + SKF_AD_CPU:
305 case SKF_AD_OFF + SKF_AD_RANDOM:
307 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
309 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
311 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
312 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
314 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
315 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
317 case SKF_AD_OFF + SKF_AD_NLATTR:
318 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
320 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
321 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
323 case SKF_AD_OFF + SKF_AD_CPU:
324 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
326 case SKF_AD_OFF + SKF_AD_RANDOM:
327 *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
328 bpf_user_rnd_init_once();
333 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
335 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
339 /* This is just a dummy call to avoid letting the compiler
340 * evict __bpf_call_base() as an optimization. Placed here
341 * where no-one bothers.
343 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
352 * bpf_convert_filter - convert filter program
353 * @prog: the user passed filter program
354 * @len: the length of the user passed filter program
355 * @new_prog: allocated 'struct bpf_prog' or NULL
356 * @new_len: pointer to store length of converted program
358 * Remap 'sock_filter' style classic BPF (cBPF) instruction set to 'bpf_insn'
359 * style extended BPF (eBPF).
360 * Conversion workflow:
362 * 1) First pass for calculating the new program length:
363 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
365 * 2) 2nd pass to remap in two passes: 1st pass finds new
366 * jump offsets, 2nd pass remapping:
367 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
369 static int bpf_convert_filter(struct sock_filter *prog, int len,
370 struct bpf_prog *new_prog, int *new_len)
372 int new_flen = 0, pass = 0, target, i, stack_off;
373 struct bpf_insn *new_insn, *first_insn = NULL;
374 struct sock_filter *fp;
378 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
379 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
381 if (len <= 0 || len > BPF_MAXINSNS)
385 first_insn = new_prog->insnsi;
386 addrs = kcalloc(len, sizeof(*addrs),
387 GFP_KERNEL | __GFP_NOWARN);
393 new_insn = first_insn;
396 /* Classic BPF related prologue emission. */
398 /* Classic BPF expects A and X to be reset first. These need
399 * to be guaranteed to be the first two instructions.
401 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
402 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
404 /* All programs must keep CTX in callee saved BPF_REG_CTX.
405 * In eBPF case it's done by the compiler, here we need to
406 * do this ourself. Initial CTX is present in BPF_REG_ARG1.
408 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
413 for (i = 0; i < len; fp++, i++) {
414 struct bpf_insn tmp_insns[6] = { };
415 struct bpf_insn *insn = tmp_insns;
418 addrs[i] = new_insn - first_insn;
421 /* All arithmetic insns and skb loads map as-is. */
422 case BPF_ALU | BPF_ADD | BPF_X:
423 case BPF_ALU | BPF_ADD | BPF_K:
424 case BPF_ALU | BPF_SUB | BPF_X:
425 case BPF_ALU | BPF_SUB | BPF_K:
426 case BPF_ALU | BPF_AND | BPF_X:
427 case BPF_ALU | BPF_AND | BPF_K:
428 case BPF_ALU | BPF_OR | BPF_X:
429 case BPF_ALU | BPF_OR | BPF_K:
430 case BPF_ALU | BPF_LSH | BPF_X:
431 case BPF_ALU | BPF_LSH | BPF_K:
432 case BPF_ALU | BPF_RSH | BPF_X:
433 case BPF_ALU | BPF_RSH | BPF_K:
434 case BPF_ALU | BPF_XOR | BPF_X:
435 case BPF_ALU | BPF_XOR | BPF_K:
436 case BPF_ALU | BPF_MUL | BPF_X:
437 case BPF_ALU | BPF_MUL | BPF_K:
438 case BPF_ALU | BPF_DIV | BPF_X:
439 case BPF_ALU | BPF_DIV | BPF_K:
440 case BPF_ALU | BPF_MOD | BPF_X:
441 case BPF_ALU | BPF_MOD | BPF_K:
442 case BPF_ALU | BPF_NEG:
443 case BPF_LD | BPF_ABS | BPF_W:
444 case BPF_LD | BPF_ABS | BPF_H:
445 case BPF_LD | BPF_ABS | BPF_B:
446 case BPF_LD | BPF_IND | BPF_W:
447 case BPF_LD | BPF_IND | BPF_H:
448 case BPF_LD | BPF_IND | BPF_B:
449 /* Check for overloaded BPF extension and
450 * directly convert it if found, otherwise
451 * just move on with mapping.
453 if (BPF_CLASS(fp->code) == BPF_LD &&
454 BPF_MODE(fp->code) == BPF_ABS &&
455 convert_bpf_extensions(fp, &insn))
458 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
461 /* Jump transformation cannot use BPF block macros
462 * everywhere as offset calculation and target updates
463 * require a bit more work than the rest, i.e. jump
464 * opcodes map as-is, but offsets need adjustment.
467 #define BPF_EMIT_JMP \
469 if (target >= len || target < 0) \
471 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
472 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
473 insn->off -= insn - tmp_insns; \
476 case BPF_JMP | BPF_JA:
477 target = i + fp->k + 1;
478 insn->code = fp->code;
482 case BPF_JMP | BPF_JEQ | BPF_K:
483 case BPF_JMP | BPF_JEQ | BPF_X:
484 case BPF_JMP | BPF_JSET | BPF_K:
485 case BPF_JMP | BPF_JSET | BPF_X:
486 case BPF_JMP | BPF_JGT | BPF_K:
487 case BPF_JMP | BPF_JGT | BPF_X:
488 case BPF_JMP | BPF_JGE | BPF_K:
489 case BPF_JMP | BPF_JGE | BPF_X:
490 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
491 /* BPF immediates are signed, zero extend
492 * immediate into tmp register and use it
495 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
497 insn->dst_reg = BPF_REG_A;
498 insn->src_reg = BPF_REG_TMP;
501 insn->dst_reg = BPF_REG_A;
503 bpf_src = BPF_SRC(fp->code);
504 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
507 /* Common case where 'jump_false' is next insn. */
509 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
510 target = i + fp->jt + 1;
515 /* Convert JEQ into JNE when 'jump_true' is next insn. */
516 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
517 insn->code = BPF_JMP | BPF_JNE | bpf_src;
518 target = i + fp->jf + 1;
523 /* Other jumps are mapped into two insns: Jxx and JA. */
524 target = i + fp->jt + 1;
525 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
529 insn->code = BPF_JMP | BPF_JA;
530 target = i + fp->jf + 1;
534 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
535 case BPF_LDX | BPF_MSH | BPF_B:
537 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
538 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
539 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
541 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
543 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
545 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
547 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
550 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
551 * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
553 case BPF_RET | BPF_A:
554 case BPF_RET | BPF_K:
555 if (BPF_RVAL(fp->code) == BPF_K)
556 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
558 *insn = BPF_EXIT_INSN();
561 /* Store to stack. */
564 stack_off = fp->k * 4 + 4;
565 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
566 BPF_ST ? BPF_REG_A : BPF_REG_X,
568 /* check_load_and_stores() verifies that classic BPF can
569 * load from stack only after write, so tracking
570 * stack_depth for ST|STX insns is enough
572 if (new_prog && new_prog->aux->stack_depth < stack_off)
573 new_prog->aux->stack_depth = stack_off;
576 /* Load from stack. */
577 case BPF_LD | BPF_MEM:
578 case BPF_LDX | BPF_MEM:
579 stack_off = fp->k * 4 + 4;
580 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
581 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
586 case BPF_LD | BPF_IMM:
587 case BPF_LDX | BPF_IMM:
588 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
589 BPF_REG_A : BPF_REG_X, fp->k);
593 case BPF_MISC | BPF_TAX:
594 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
598 case BPF_MISC | BPF_TXA:
599 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
602 /* A = skb->len or X = skb->len */
603 case BPF_LD | BPF_W | BPF_LEN:
604 case BPF_LDX | BPF_W | BPF_LEN:
605 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
606 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
607 offsetof(struct sk_buff, len));
610 /* Access seccomp_data fields. */
611 case BPF_LDX | BPF_ABS | BPF_W:
612 /* A = *(u32 *) (ctx + K) */
613 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
616 /* Unknown instruction. */
623 memcpy(new_insn, tmp_insns,
624 sizeof(*insn) * (insn - tmp_insns));
625 new_insn += insn - tmp_insns;
629 /* Only calculating new length. */
630 *new_len = new_insn - first_insn;
635 if (new_flen != new_insn - first_insn) {
636 new_flen = new_insn - first_insn;
643 BUG_ON(*new_len != new_flen);
652 * As we dont want to clear mem[] array for each packet going through
653 * __bpf_prog_run(), we check that filter loaded by user never try to read
654 * a cell if not previously written, and we check all branches to be sure
655 * a malicious user doesn't try to abuse us.
657 static int check_load_and_stores(const struct sock_filter *filter, int flen)
659 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
662 BUILD_BUG_ON(BPF_MEMWORDS > 16);
664 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
668 memset(masks, 0xff, flen * sizeof(*masks));
670 for (pc = 0; pc < flen; pc++) {
671 memvalid &= masks[pc];
673 switch (filter[pc].code) {
676 memvalid |= (1 << filter[pc].k);
678 case BPF_LD | BPF_MEM:
679 case BPF_LDX | BPF_MEM:
680 if (!(memvalid & (1 << filter[pc].k))) {
685 case BPF_JMP | BPF_JA:
686 /* A jump must set masks on target */
687 masks[pc + 1 + filter[pc].k] &= memvalid;
690 case BPF_JMP | BPF_JEQ | BPF_K:
691 case BPF_JMP | BPF_JEQ | BPF_X:
692 case BPF_JMP | BPF_JGE | BPF_K:
693 case BPF_JMP | BPF_JGE | BPF_X:
694 case BPF_JMP | BPF_JGT | BPF_K:
695 case BPF_JMP | BPF_JGT | BPF_X:
696 case BPF_JMP | BPF_JSET | BPF_K:
697 case BPF_JMP | BPF_JSET | BPF_X:
698 /* A jump must set masks on targets */
699 masks[pc + 1 + filter[pc].jt] &= memvalid;
700 masks[pc + 1 + filter[pc].jf] &= memvalid;
710 static bool chk_code_allowed(u16 code_to_probe)
712 static const bool codes[] = {
713 /* 32 bit ALU operations */
714 [BPF_ALU | BPF_ADD | BPF_K] = true,
715 [BPF_ALU | BPF_ADD | BPF_X] = true,
716 [BPF_ALU | BPF_SUB | BPF_K] = true,
717 [BPF_ALU | BPF_SUB | BPF_X] = true,
718 [BPF_ALU | BPF_MUL | BPF_K] = true,
719 [BPF_ALU | BPF_MUL | BPF_X] = true,
720 [BPF_ALU | BPF_DIV | BPF_K] = true,
721 [BPF_ALU | BPF_DIV | BPF_X] = true,
722 [BPF_ALU | BPF_MOD | BPF_K] = true,
723 [BPF_ALU | BPF_MOD | BPF_X] = true,
724 [BPF_ALU | BPF_AND | BPF_K] = true,
725 [BPF_ALU | BPF_AND | BPF_X] = true,
726 [BPF_ALU | BPF_OR | BPF_K] = true,
727 [BPF_ALU | BPF_OR | BPF_X] = true,
728 [BPF_ALU | BPF_XOR | BPF_K] = true,
729 [BPF_ALU | BPF_XOR | BPF_X] = true,
730 [BPF_ALU | BPF_LSH | BPF_K] = true,
731 [BPF_ALU | BPF_LSH | BPF_X] = true,
732 [BPF_ALU | BPF_RSH | BPF_K] = true,
733 [BPF_ALU | BPF_RSH | BPF_X] = true,
734 [BPF_ALU | BPF_NEG] = true,
735 /* Load instructions */
736 [BPF_LD | BPF_W | BPF_ABS] = true,
737 [BPF_LD | BPF_H | BPF_ABS] = true,
738 [BPF_LD | BPF_B | BPF_ABS] = true,
739 [BPF_LD | BPF_W | BPF_LEN] = true,
740 [BPF_LD | BPF_W | BPF_IND] = true,
741 [BPF_LD | BPF_H | BPF_IND] = true,
742 [BPF_LD | BPF_B | BPF_IND] = true,
743 [BPF_LD | BPF_IMM] = true,
744 [BPF_LD | BPF_MEM] = true,
745 [BPF_LDX | BPF_W | BPF_LEN] = true,
746 [BPF_LDX | BPF_B | BPF_MSH] = true,
747 [BPF_LDX | BPF_IMM] = true,
748 [BPF_LDX | BPF_MEM] = true,
749 /* Store instructions */
752 /* Misc instructions */
753 [BPF_MISC | BPF_TAX] = true,
754 [BPF_MISC | BPF_TXA] = true,
755 /* Return instructions */
756 [BPF_RET | BPF_K] = true,
757 [BPF_RET | BPF_A] = true,
758 /* Jump instructions */
759 [BPF_JMP | BPF_JA] = true,
760 [BPF_JMP | BPF_JEQ | BPF_K] = true,
761 [BPF_JMP | BPF_JEQ | BPF_X] = true,
762 [BPF_JMP | BPF_JGE | BPF_K] = true,
763 [BPF_JMP | BPF_JGE | BPF_X] = true,
764 [BPF_JMP | BPF_JGT | BPF_K] = true,
765 [BPF_JMP | BPF_JGT | BPF_X] = true,
766 [BPF_JMP | BPF_JSET | BPF_K] = true,
767 [BPF_JMP | BPF_JSET | BPF_X] = true,
770 if (code_to_probe >= ARRAY_SIZE(codes))
773 return codes[code_to_probe];
776 static bool bpf_check_basics_ok(const struct sock_filter *filter,
781 if (flen == 0 || flen > BPF_MAXINSNS)
788 * bpf_check_classic - verify socket filter code
789 * @filter: filter to verify
790 * @flen: length of filter
792 * Check the user's filter code. If we let some ugly
793 * filter code slip through kaboom! The filter must contain
794 * no references or jumps that are out of range, no illegal
795 * instructions, and must end with a RET instruction.
797 * All jumps are forward as they are not signed.
799 * Returns 0 if the rule set is legal or -EINVAL if not.
801 static int bpf_check_classic(const struct sock_filter *filter,
807 /* Check the filter code now */
808 for (pc = 0; pc < flen; pc++) {
809 const struct sock_filter *ftest = &filter[pc];
811 /* May we actually operate on this code? */
812 if (!chk_code_allowed(ftest->code))
815 /* Some instructions need special checks */
816 switch (ftest->code) {
817 case BPF_ALU | BPF_DIV | BPF_K:
818 case BPF_ALU | BPF_MOD | BPF_K:
819 /* Check for division by zero */
823 case BPF_ALU | BPF_LSH | BPF_K:
824 case BPF_ALU | BPF_RSH | BPF_K:
828 case BPF_LD | BPF_MEM:
829 case BPF_LDX | BPF_MEM:
832 /* Check for invalid memory addresses */
833 if (ftest->k >= BPF_MEMWORDS)
836 case BPF_JMP | BPF_JA:
837 /* Note, the large ftest->k might cause loops.
838 * Compare this with conditional jumps below,
839 * where offsets are limited. --ANK (981016)
841 if (ftest->k >= (unsigned int)(flen - pc - 1))
844 case BPF_JMP | BPF_JEQ | BPF_K:
845 case BPF_JMP | BPF_JEQ | BPF_X:
846 case BPF_JMP | BPF_JGE | BPF_K:
847 case BPF_JMP | BPF_JGE | BPF_X:
848 case BPF_JMP | BPF_JGT | BPF_K:
849 case BPF_JMP | BPF_JGT | BPF_X:
850 case BPF_JMP | BPF_JSET | BPF_K:
851 case BPF_JMP | BPF_JSET | BPF_X:
852 /* Both conditionals must be safe */
853 if (pc + ftest->jt + 1 >= flen ||
854 pc + ftest->jf + 1 >= flen)
857 case BPF_LD | BPF_W | BPF_ABS:
858 case BPF_LD | BPF_H | BPF_ABS:
859 case BPF_LD | BPF_B | BPF_ABS:
861 if (bpf_anc_helper(ftest) & BPF_ANC)
863 /* Ancillary operation unknown or unsupported */
864 if (anc_found == false && ftest->k >= SKF_AD_OFF)
869 /* Last instruction must be a RET code */
870 switch (filter[flen - 1].code) {
871 case BPF_RET | BPF_K:
872 case BPF_RET | BPF_A:
873 return check_load_and_stores(filter, flen);
879 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
880 const struct sock_fprog *fprog)
882 unsigned int fsize = bpf_classic_proglen(fprog);
883 struct sock_fprog_kern *fkprog;
885 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
889 fkprog = fp->orig_prog;
890 fkprog->len = fprog->len;
892 fkprog->filter = kmemdup(fp->insns, fsize,
893 GFP_KERNEL | __GFP_NOWARN);
894 if (!fkprog->filter) {
895 kfree(fp->orig_prog);
902 static void bpf_release_orig_filter(struct bpf_prog *fp)
904 struct sock_fprog_kern *fprog = fp->orig_prog;
907 kfree(fprog->filter);
912 static void __bpf_prog_release(struct bpf_prog *prog)
914 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
917 bpf_release_orig_filter(prog);
922 static void __sk_filter_release(struct sk_filter *fp)
924 __bpf_prog_release(fp->prog);
929 * sk_filter_release_rcu - Release a socket filter by rcu_head
930 * @rcu: rcu_head that contains the sk_filter to free
932 static void sk_filter_release_rcu(struct rcu_head *rcu)
934 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
936 __sk_filter_release(fp);
940 * sk_filter_release - release a socket filter
941 * @fp: filter to remove
943 * Remove a filter from a socket and release its resources.
945 static void sk_filter_release(struct sk_filter *fp)
947 if (refcount_dec_and_test(&fp->refcnt))
948 call_rcu(&fp->rcu, sk_filter_release_rcu);
951 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
953 u32 filter_size = bpf_prog_size(fp->prog->len);
955 atomic_sub(filter_size, &sk->sk_omem_alloc);
956 sk_filter_release(fp);
959 /* try to charge the socket memory if there is space available
960 * return true on success
962 static bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
964 u32 filter_size = bpf_prog_size(fp->prog->len);
966 /* same check as in sock_kmalloc() */
967 if (filter_size <= sysctl_optmem_max &&
968 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
969 atomic_add(filter_size, &sk->sk_omem_alloc);
975 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
977 bool ret = __sk_filter_charge(sk, fp);
979 refcount_inc(&fp->refcnt);
983 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
985 struct sock_filter *old_prog;
986 struct bpf_prog *old_fp;
987 int err, new_len, old_len = fp->len;
989 /* We are free to overwrite insns et al right here as it
990 * won't be used at this point in time anymore internally
991 * after the migration to the internal BPF instruction
994 BUILD_BUG_ON(sizeof(struct sock_filter) !=
995 sizeof(struct bpf_insn));
997 /* Conversion cannot happen on overlapping memory areas,
998 * so we need to keep the user BPF around until the 2nd
999 * pass. At this time, the user BPF is stored in fp->insns.
1001 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
1002 GFP_KERNEL | __GFP_NOWARN);
1008 /* 1st pass: calculate the new program length. */
1009 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
1013 /* Expand fp for appending the new filter representation. */
1015 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
1017 /* The old_fp is still around in case we couldn't
1018 * allocate new memory, so uncharge on that one.
1027 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1028 err = bpf_convert_filter(old_prog, old_len, fp, &new_len);
1030 /* 2nd bpf_convert_filter() can fail only if it fails
1031 * to allocate memory, remapping must succeed. Note,
1032 * that at this time old_fp has already been released
1037 /* We are guaranteed to never error here with cBPF to eBPF
1038 * transitions, since there's no issue with type compatibility
1039 * checks on program arrays.
1041 fp = bpf_prog_select_runtime(fp, &err);
1049 __bpf_prog_release(fp);
1050 return ERR_PTR(err);
1053 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1054 bpf_aux_classic_check_t trans)
1058 fp->bpf_func = NULL;
1061 err = bpf_check_classic(fp->insns, fp->len);
1063 __bpf_prog_release(fp);
1064 return ERR_PTR(err);
1067 /* There might be additional checks and transformations
1068 * needed on classic filters, f.e. in case of seccomp.
1071 err = trans(fp->insns, fp->len);
1073 __bpf_prog_release(fp);
1074 return ERR_PTR(err);
1078 /* Probe if we can JIT compile the filter and if so, do
1079 * the compilation of the filter.
1081 bpf_jit_compile(fp);
1083 /* JIT compiler couldn't process this filter, so do the
1084 * internal BPF translation for the optimized interpreter.
1087 fp = bpf_migrate_filter(fp);
1093 * bpf_prog_create - create an unattached filter
1094 * @pfp: the unattached filter that is created
1095 * @fprog: the filter program
1097 * Create a filter independent of any socket. We first run some
1098 * sanity checks on it to make sure it does not explode on us later.
1099 * If an error occurs or there is insufficient memory for the filter
1100 * a negative errno code is returned. On success the return is zero.
1102 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1104 unsigned int fsize = bpf_classic_proglen(fprog);
1105 struct bpf_prog *fp;
1107 /* Make sure new filter is there and in the right amounts. */
1108 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1111 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1115 memcpy(fp->insns, fprog->filter, fsize);
1117 fp->len = fprog->len;
1118 /* Since unattached filters are not copied back to user
1119 * space through sk_get_filter(), we do not need to hold
1120 * a copy here, and can spare us the work.
1122 fp->orig_prog = NULL;
1124 /* bpf_prepare_filter() already takes care of freeing
1125 * memory in case something goes wrong.
1127 fp = bpf_prepare_filter(fp, NULL);
1134 EXPORT_SYMBOL_GPL(bpf_prog_create);
1137 * bpf_prog_create_from_user - create an unattached filter from user buffer
1138 * @pfp: the unattached filter that is created
1139 * @fprog: the filter program
1140 * @trans: post-classic verifier transformation handler
1141 * @save_orig: save classic BPF program
1143 * This function effectively does the same as bpf_prog_create(), only
1144 * that it builds up its insns buffer from user space provided buffer.
1145 * It also allows for passing a bpf_aux_classic_check_t handler.
1147 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1148 bpf_aux_classic_check_t trans, bool save_orig)
1150 unsigned int fsize = bpf_classic_proglen(fprog);
1151 struct bpf_prog *fp;
1154 /* Make sure new filter is there and in the right amounts. */
1155 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1158 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1162 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1163 __bpf_prog_free(fp);
1167 fp->len = fprog->len;
1168 fp->orig_prog = NULL;
1171 err = bpf_prog_store_orig_filter(fp, fprog);
1173 __bpf_prog_free(fp);
1178 /* bpf_prepare_filter() already takes care of freeing
1179 * memory in case something goes wrong.
1181 fp = bpf_prepare_filter(fp, trans);
1188 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1190 void bpf_prog_destroy(struct bpf_prog *fp)
1192 __bpf_prog_release(fp);
1194 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1196 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1198 struct sk_filter *fp, *old_fp;
1200 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1206 if (!__sk_filter_charge(sk, fp)) {
1210 refcount_set(&fp->refcnt, 1);
1212 old_fp = rcu_dereference_protected(sk->sk_filter,
1213 lockdep_sock_is_held(sk));
1214 rcu_assign_pointer(sk->sk_filter, fp);
1217 sk_filter_uncharge(sk, old_fp);
1222 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1224 struct bpf_prog *old_prog;
1227 if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1230 if (sk_unhashed(sk) && sk->sk_reuseport) {
1231 err = reuseport_alloc(sk);
1234 } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1235 /* The socket wasn't bound with SO_REUSEPORT */
1239 old_prog = reuseport_attach_prog(sk, prog);
1241 bpf_prog_destroy(old_prog);
1247 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1249 unsigned int fsize = bpf_classic_proglen(fprog);
1250 struct bpf_prog *prog;
1253 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1254 return ERR_PTR(-EPERM);
1256 /* Make sure new filter is there and in the right amounts. */
1257 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1258 return ERR_PTR(-EINVAL);
1260 prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1262 return ERR_PTR(-ENOMEM);
1264 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1265 __bpf_prog_free(prog);
1266 return ERR_PTR(-EFAULT);
1269 prog->len = fprog->len;
1271 err = bpf_prog_store_orig_filter(prog, fprog);
1273 __bpf_prog_free(prog);
1274 return ERR_PTR(-ENOMEM);
1277 /* bpf_prepare_filter() already takes care of freeing
1278 * memory in case something goes wrong.
1280 return bpf_prepare_filter(prog, NULL);
1284 * sk_attach_filter - attach a socket filter
1285 * @fprog: the filter program
1286 * @sk: the socket to use
1288 * Attach the user's filter code. We first run some sanity checks on
1289 * it to make sure it does not explode on us later. If an error
1290 * occurs or there is insufficient memory for the filter a negative
1291 * errno code is returned. On success the return is zero.
1293 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1295 struct bpf_prog *prog = __get_filter(fprog, sk);
1299 return PTR_ERR(prog);
1301 err = __sk_attach_prog(prog, sk);
1303 __bpf_prog_release(prog);
1309 EXPORT_SYMBOL_GPL(sk_attach_filter);
1311 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1313 struct bpf_prog *prog = __get_filter(fprog, sk);
1317 return PTR_ERR(prog);
1319 err = __reuseport_attach_prog(prog, sk);
1321 __bpf_prog_release(prog);
1328 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1330 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1331 return ERR_PTR(-EPERM);
1333 return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1336 int sk_attach_bpf(u32 ufd, struct sock *sk)
1338 struct bpf_prog *prog = __get_bpf(ufd, sk);
1342 return PTR_ERR(prog);
1344 err = __sk_attach_prog(prog, sk);
1353 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1355 struct bpf_prog *prog = __get_bpf(ufd, sk);
1359 return PTR_ERR(prog);
1361 err = __reuseport_attach_prog(prog, sk);
1370 struct bpf_scratchpad {
1372 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1373 u8 buff[MAX_BPF_STACK];
1377 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1379 static inline int __bpf_try_make_writable(struct sk_buff *skb,
1380 unsigned int write_len)
1382 return skb_ensure_writable(skb, write_len);
1385 static inline int bpf_try_make_writable(struct sk_buff *skb,
1386 unsigned int write_len)
1388 int err = __bpf_try_make_writable(skb, write_len);
1390 bpf_compute_data_end(skb);
1394 static int bpf_try_make_head_writable(struct sk_buff *skb)
1396 return bpf_try_make_writable(skb, skb_headlen(skb));
1399 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1401 if (skb_at_tc_ingress(skb))
1402 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1405 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1407 if (skb_at_tc_ingress(skb))
1408 skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1411 BPF_CALL_5(bpf_skb_store_bytes, struct sk_buff *, skb, u32, offset,
1412 const void *, from, u32, len, u64, flags)
1416 if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1418 if (unlikely(offset > 0xffff))
1420 if (unlikely(bpf_try_make_writable(skb, offset + len)))
1423 ptr = skb->data + offset;
1424 if (flags & BPF_F_RECOMPUTE_CSUM)
1425 __skb_postpull_rcsum(skb, ptr, len, offset);
1427 memcpy(ptr, from, len);
1429 if (flags & BPF_F_RECOMPUTE_CSUM)
1430 __skb_postpush_rcsum(skb, ptr, len, offset);
1431 if (flags & BPF_F_INVALIDATE_HASH)
1432 skb_clear_hash(skb);
1437 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1438 .func = bpf_skb_store_bytes,
1440 .ret_type = RET_INTEGER,
1441 .arg1_type = ARG_PTR_TO_CTX,
1442 .arg2_type = ARG_ANYTHING,
1443 .arg3_type = ARG_PTR_TO_MEM,
1444 .arg4_type = ARG_CONST_SIZE,
1445 .arg5_type = ARG_ANYTHING,
1448 BPF_CALL_4(bpf_skb_load_bytes, const struct sk_buff *, skb, u32, offset,
1449 void *, to, u32, len)
1453 if (unlikely(offset > 0xffff))
1456 ptr = skb_header_pointer(skb, offset, len, to);
1460 memcpy(to, ptr, len);
1468 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1469 .func = bpf_skb_load_bytes,
1471 .ret_type = RET_INTEGER,
1472 .arg1_type = ARG_PTR_TO_CTX,
1473 .arg2_type = ARG_ANYTHING,
1474 .arg3_type = ARG_PTR_TO_UNINIT_MEM,
1475 .arg4_type = ARG_CONST_SIZE,
1478 BPF_CALL_2(bpf_skb_pull_data, struct sk_buff *, skb, u32, len)
1480 /* Idea is the following: should the needed direct read/write
1481 * test fail during runtime, we can pull in more data and redo
1482 * again, since implicitly, we invalidate previous checks here.
1484 * Or, since we know how much we need to make read/writeable,
1485 * this can be done once at the program beginning for direct
1486 * access case. By this we overcome limitations of only current
1487 * headroom being accessible.
1489 return bpf_try_make_writable(skb, len ? : skb_headlen(skb));
1492 static const struct bpf_func_proto bpf_skb_pull_data_proto = {
1493 .func = bpf_skb_pull_data,
1495 .ret_type = RET_INTEGER,
1496 .arg1_type = ARG_PTR_TO_CTX,
1497 .arg2_type = ARG_ANYTHING,
1500 BPF_CALL_5(bpf_l3_csum_replace, struct sk_buff *, skb, u32, offset,
1501 u64, from, u64, to, u64, flags)
1505 if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1507 if (unlikely(offset > 0xffff || offset & 1))
1509 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1512 ptr = (__sum16 *)(skb->data + offset);
1513 switch (flags & BPF_F_HDR_FIELD_MASK) {
1515 if (unlikely(from != 0))
1518 csum_replace_by_diff(ptr, to);
1521 csum_replace2(ptr, from, to);
1524 csum_replace4(ptr, from, to);
1533 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1534 .func = bpf_l3_csum_replace,
1536 .ret_type = RET_INTEGER,
1537 .arg1_type = ARG_PTR_TO_CTX,
1538 .arg2_type = ARG_ANYTHING,
1539 .arg3_type = ARG_ANYTHING,
1540 .arg4_type = ARG_ANYTHING,
1541 .arg5_type = ARG_ANYTHING,
1544 BPF_CALL_5(bpf_l4_csum_replace, struct sk_buff *, skb, u32, offset,
1545 u64, from, u64, to, u64, flags)
1547 bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1548 bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1549 bool do_mforce = flags & BPF_F_MARK_ENFORCE;
1552 if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_MARK_ENFORCE |
1553 BPF_F_PSEUDO_HDR | BPF_F_HDR_FIELD_MASK)))
1555 if (unlikely(offset > 0xffff || offset & 1))
1557 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1560 ptr = (__sum16 *)(skb->data + offset);
1561 if (is_mmzero && !do_mforce && !*ptr)
1564 switch (flags & BPF_F_HDR_FIELD_MASK) {
1566 if (unlikely(from != 0))
1569 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1572 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1575 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1581 if (is_mmzero && !*ptr)
1582 *ptr = CSUM_MANGLED_0;
1586 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1587 .func = bpf_l4_csum_replace,
1589 .ret_type = RET_INTEGER,
1590 .arg1_type = ARG_PTR_TO_CTX,
1591 .arg2_type = ARG_ANYTHING,
1592 .arg3_type = ARG_ANYTHING,
1593 .arg4_type = ARG_ANYTHING,
1594 .arg5_type = ARG_ANYTHING,
1597 BPF_CALL_5(bpf_csum_diff, __be32 *, from, u32, from_size,
1598 __be32 *, to, u32, to_size, __wsum, seed)
1600 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1601 u32 diff_size = from_size + to_size;
1604 /* This is quite flexible, some examples:
1606 * from_size == 0, to_size > 0, seed := csum --> pushing data
1607 * from_size > 0, to_size == 0, seed := csum --> pulling data
1608 * from_size > 0, to_size > 0, seed := 0 --> diffing data
1610 * Even for diffing, from_size and to_size don't need to be equal.
1612 if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1613 diff_size > sizeof(sp->diff)))
1616 for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1617 sp->diff[j] = ~from[i];
1618 for (i = 0; i < to_size / sizeof(__be32); i++, j++)
1619 sp->diff[j] = to[i];
1621 return csum_partial(sp->diff, diff_size, seed);
1624 static const struct bpf_func_proto bpf_csum_diff_proto = {
1625 .func = bpf_csum_diff,
1628 .ret_type = RET_INTEGER,
1629 .arg1_type = ARG_PTR_TO_MEM,
1630 .arg2_type = ARG_CONST_SIZE_OR_ZERO,
1631 .arg3_type = ARG_PTR_TO_MEM,
1632 .arg4_type = ARG_CONST_SIZE_OR_ZERO,
1633 .arg5_type = ARG_ANYTHING,
1636 BPF_CALL_2(bpf_csum_update, struct sk_buff *, skb, __wsum, csum)
1638 /* The interface is to be used in combination with bpf_csum_diff()
1639 * for direct packet writes. csum rotation for alignment as well
1640 * as emulating csum_sub() can be done from the eBPF program.
1642 if (skb->ip_summed == CHECKSUM_COMPLETE)
1643 return (skb->csum = csum_add(skb->csum, csum));
1648 static const struct bpf_func_proto bpf_csum_update_proto = {
1649 .func = bpf_csum_update,
1651 .ret_type = RET_INTEGER,
1652 .arg1_type = ARG_PTR_TO_CTX,
1653 .arg2_type = ARG_ANYTHING,
1656 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1658 return dev_forward_skb(dev, skb);
1661 static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
1662 struct sk_buff *skb)
1664 int ret = ____dev_forward_skb(dev, skb);
1668 ret = netif_rx(skb);
1674 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1678 if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1679 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1686 __this_cpu_inc(xmit_recursion);
1687 ret = dev_queue_xmit(skb);
1688 __this_cpu_dec(xmit_recursion);
1693 static int __bpf_redirect_no_mac(struct sk_buff *skb, struct net_device *dev,
1696 /* skb->mac_len is not set on normal egress */
1697 unsigned int mlen = skb->network_header - skb->mac_header;
1699 __skb_pull(skb, mlen);
1701 /* At ingress, the mac header has already been pulled once.
1702 * At egress, skb_pospull_rcsum has to be done in case that
1703 * the skb is originated from ingress (i.e. a forwarded skb)
1704 * to ensure that rcsum starts at net header.
1706 if (!skb_at_tc_ingress(skb))
1707 skb_postpull_rcsum(skb, skb_mac_header(skb), mlen);
1708 skb_pop_mac_header(skb);
1709 skb_reset_mac_len(skb);
1710 return flags & BPF_F_INGRESS ?
1711 __bpf_rx_skb_no_mac(dev, skb) : __bpf_tx_skb(dev, skb);
1714 static int __bpf_redirect_common(struct sk_buff *skb, struct net_device *dev,
1717 /* Verify that a link layer header is carried */
1718 if (unlikely(skb->mac_header >= skb->network_header)) {
1723 bpf_push_mac_rcsum(skb);
1724 return flags & BPF_F_INGRESS ?
1725 __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1728 static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
1731 if (dev_is_mac_header_xmit(dev))
1732 return __bpf_redirect_common(skb, dev, flags);
1734 return __bpf_redirect_no_mac(skb, dev, flags);
1737 BPF_CALL_3(bpf_clone_redirect, struct sk_buff *, skb, u32, ifindex, u64, flags)
1739 struct net_device *dev;
1740 struct sk_buff *clone;
1743 if (unlikely(flags & ~(BPF_F_INGRESS)))
1746 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1750 clone = skb_clone(skb, GFP_ATOMIC);
1751 if (unlikely(!clone))
1754 /* For direct write, we need to keep the invariant that the skbs
1755 * we're dealing with need to be uncloned. Should uncloning fail
1756 * here, we need to free the just generated clone to unclone once
1759 ret = bpf_try_make_head_writable(skb);
1760 if (unlikely(ret)) {
1765 return __bpf_redirect(clone, dev, flags);
1768 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1769 .func = bpf_clone_redirect,
1771 .ret_type = RET_INTEGER,
1772 .arg1_type = ARG_PTR_TO_CTX,
1773 .arg2_type = ARG_ANYTHING,
1774 .arg3_type = ARG_ANYTHING,
1777 struct redirect_info {
1782 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1784 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
1786 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1788 if (unlikely(flags & ~(BPF_F_INGRESS)))
1791 ri->ifindex = ifindex;
1794 return TC_ACT_REDIRECT;
1797 int skb_do_redirect(struct sk_buff *skb)
1799 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1800 struct net_device *dev;
1802 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1804 if (unlikely(!dev)) {
1809 return __bpf_redirect(skb, dev, ri->flags);
1812 static const struct bpf_func_proto bpf_redirect_proto = {
1813 .func = bpf_redirect,
1815 .ret_type = RET_INTEGER,
1816 .arg1_type = ARG_ANYTHING,
1817 .arg2_type = ARG_ANYTHING,
1820 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
1822 return task_get_classid(skb);
1825 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1826 .func = bpf_get_cgroup_classid,
1828 .ret_type = RET_INTEGER,
1829 .arg1_type = ARG_PTR_TO_CTX,
1832 BPF_CALL_1(bpf_get_route_realm, const struct sk_buff *, skb)
1834 return dst_tclassid(skb);
1837 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1838 .func = bpf_get_route_realm,
1840 .ret_type = RET_INTEGER,
1841 .arg1_type = ARG_PTR_TO_CTX,
1844 BPF_CALL_1(bpf_get_hash_recalc, struct sk_buff *, skb)
1846 /* If skb_clear_hash() was called due to mangling, we can
1847 * trigger SW recalculation here. Later access to hash
1848 * can then use the inline skb->hash via context directly
1849 * instead of calling this helper again.
1851 return skb_get_hash(skb);
1854 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1855 .func = bpf_get_hash_recalc,
1857 .ret_type = RET_INTEGER,
1858 .arg1_type = ARG_PTR_TO_CTX,
1861 BPF_CALL_1(bpf_set_hash_invalid, struct sk_buff *, skb)
1863 /* After all direct packet write, this can be used once for
1864 * triggering a lazy recalc on next skb_get_hash() invocation.
1866 skb_clear_hash(skb);
1870 static const struct bpf_func_proto bpf_set_hash_invalid_proto = {
1871 .func = bpf_set_hash_invalid,
1873 .ret_type = RET_INTEGER,
1874 .arg1_type = ARG_PTR_TO_CTX,
1877 BPF_CALL_2(bpf_set_hash, struct sk_buff *, skb, u32, hash)
1879 /* Set user specified hash as L4(+), so that it gets returned
1880 * on skb_get_hash() call unless BPF prog later on triggers a
1883 __skb_set_sw_hash(skb, hash, true);
1887 static const struct bpf_func_proto bpf_set_hash_proto = {
1888 .func = bpf_set_hash,
1890 .ret_type = RET_INTEGER,
1891 .arg1_type = ARG_PTR_TO_CTX,
1892 .arg2_type = ARG_ANYTHING,
1895 BPF_CALL_3(bpf_skb_vlan_push, struct sk_buff *, skb, __be16, vlan_proto,
1900 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1901 vlan_proto != htons(ETH_P_8021AD)))
1902 vlan_proto = htons(ETH_P_8021Q);
1904 bpf_push_mac_rcsum(skb);
1905 ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1906 bpf_pull_mac_rcsum(skb);
1908 bpf_compute_data_end(skb);
1912 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1913 .func = bpf_skb_vlan_push,
1915 .ret_type = RET_INTEGER,
1916 .arg1_type = ARG_PTR_TO_CTX,
1917 .arg2_type = ARG_ANYTHING,
1918 .arg3_type = ARG_ANYTHING,
1920 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1922 BPF_CALL_1(bpf_skb_vlan_pop, struct sk_buff *, skb)
1926 bpf_push_mac_rcsum(skb);
1927 ret = skb_vlan_pop(skb);
1928 bpf_pull_mac_rcsum(skb);
1930 bpf_compute_data_end(skb);
1934 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1935 .func = bpf_skb_vlan_pop,
1937 .ret_type = RET_INTEGER,
1938 .arg1_type = ARG_PTR_TO_CTX,
1940 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1942 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
1944 /* Caller already did skb_cow() with len as headroom,
1945 * so no need to do it here.
1948 memmove(skb->data, skb->data + len, off);
1949 memset(skb->data + off, 0, len);
1951 /* No skb_postpush_rcsum(skb, skb->data + off, len)
1952 * needed here as it does not change the skb->csum
1953 * result for checksum complete when summing over
1959 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
1961 /* skb_ensure_writable() is not needed here, as we're
1962 * already working on an uncloned skb.
1964 if (unlikely(!pskb_may_pull(skb, off + len)))
1967 skb_postpull_rcsum(skb, skb->data + off, len);
1968 memmove(skb->data + len, skb->data, off);
1969 __skb_pull(skb, len);
1974 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
1976 bool trans_same = skb->transport_header == skb->network_header;
1979 /* There's no need for __skb_push()/__skb_pull() pair to
1980 * get to the start of the mac header as we're guaranteed
1981 * to always start from here under eBPF.
1983 ret = bpf_skb_generic_push(skb, off, len);
1985 skb->mac_header -= len;
1986 skb->network_header -= len;
1988 skb->transport_header = skb->network_header;
1994 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
1996 bool trans_same = skb->transport_header == skb->network_header;
1999 /* Same here, __skb_push()/__skb_pull() pair not needed. */
2000 ret = bpf_skb_generic_pop(skb, off, len);
2002 skb->mac_header += len;
2003 skb->network_header += len;
2005 skb->transport_header = skb->network_header;
2011 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
2013 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2014 u32 off = skb->network_header - skb->mac_header;
2017 ret = skb_cow(skb, len_diff);
2018 if (unlikely(ret < 0))
2021 ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2022 if (unlikely(ret < 0))
2025 if (skb_is_gso(skb)) {
2026 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV4 needs to
2027 * be changed into SKB_GSO_TCPV6.
2029 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2030 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
2031 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV6;
2034 /* Due to IPv6 header, MSS needs to be downgraded. */
2035 skb_shinfo(skb)->gso_size -= len_diff;
2036 /* Header must be checked, and gso_segs recomputed. */
2037 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2038 skb_shinfo(skb)->gso_segs = 0;
2041 skb->protocol = htons(ETH_P_IPV6);
2042 skb_clear_hash(skb);
2047 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
2049 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2050 u32 off = skb->network_header - skb->mac_header;
2053 ret = skb_unclone(skb, GFP_ATOMIC);
2054 if (unlikely(ret < 0))
2057 ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2058 if (unlikely(ret < 0))
2061 if (skb_is_gso(skb)) {
2062 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV6 needs to
2063 * be changed into SKB_GSO_TCPV4.
2065 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
2066 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
2067 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
2070 /* Due to IPv4 header, MSS can be upgraded. */
2071 skb_shinfo(skb)->gso_size += len_diff;
2072 /* Header must be checked, and gso_segs recomputed. */
2073 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2074 skb_shinfo(skb)->gso_segs = 0;
2077 skb->protocol = htons(ETH_P_IP);
2078 skb_clear_hash(skb);
2083 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
2085 __be16 from_proto = skb->protocol;
2087 if (from_proto == htons(ETH_P_IP) &&
2088 to_proto == htons(ETH_P_IPV6))
2089 return bpf_skb_proto_4_to_6(skb);
2091 if (from_proto == htons(ETH_P_IPV6) &&
2092 to_proto == htons(ETH_P_IP))
2093 return bpf_skb_proto_6_to_4(skb);
2098 BPF_CALL_3(bpf_skb_change_proto, struct sk_buff *, skb, __be16, proto,
2103 if (unlikely(flags))
2106 /* General idea is that this helper does the basic groundwork
2107 * needed for changing the protocol, and eBPF program fills the
2108 * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
2109 * and other helpers, rather than passing a raw buffer here.
2111 * The rationale is to keep this minimal and without a need to
2112 * deal with raw packet data. F.e. even if we would pass buffers
2113 * here, the program still needs to call the bpf_lX_csum_replace()
2114 * helpers anyway. Plus, this way we keep also separation of
2115 * concerns, since f.e. bpf_skb_store_bytes() should only take
2118 * Currently, additional options and extension header space are
2119 * not supported, but flags register is reserved so we can adapt
2120 * that. For offloads, we mark packet as dodgy, so that headers
2121 * need to be verified first.
2123 ret = bpf_skb_proto_xlat(skb, proto);
2124 bpf_compute_data_end(skb);
2128 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
2129 .func = bpf_skb_change_proto,
2131 .ret_type = RET_INTEGER,
2132 .arg1_type = ARG_PTR_TO_CTX,
2133 .arg2_type = ARG_ANYTHING,
2134 .arg3_type = ARG_ANYTHING,
2137 BPF_CALL_2(bpf_skb_change_type, struct sk_buff *, skb, u32, pkt_type)
2139 /* We only allow a restricted subset to be changed for now. */
2140 if (unlikely(!skb_pkt_type_ok(skb->pkt_type) ||
2141 !skb_pkt_type_ok(pkt_type)))
2144 skb->pkt_type = pkt_type;
2148 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2149 .func = bpf_skb_change_type,
2151 .ret_type = RET_INTEGER,
2152 .arg1_type = ARG_PTR_TO_CTX,
2153 .arg2_type = ARG_ANYTHING,
2156 static u32 __bpf_skb_min_len(const struct sk_buff *skb)
2158 u32 min_len = skb_network_offset(skb);
2160 if (skb_transport_header_was_set(skb))
2161 min_len = skb_transport_offset(skb);
2162 if (skb->ip_summed == CHECKSUM_PARTIAL)
2163 min_len = skb_checksum_start_offset(skb) +
2164 skb->csum_offset + sizeof(__sum16);
2168 static u32 __bpf_skb_max_len(const struct sk_buff *skb)
2170 return skb->dev->mtu + skb->dev->hard_header_len;
2173 static int bpf_skb_grow_rcsum(struct sk_buff *skb, unsigned int new_len)
2175 unsigned int old_len = skb->len;
2178 ret = __skb_grow_rcsum(skb, new_len);
2180 memset(skb->data + old_len, 0, new_len - old_len);
2184 static int bpf_skb_trim_rcsum(struct sk_buff *skb, unsigned int new_len)
2186 return __skb_trim_rcsum(skb, new_len);
2189 BPF_CALL_3(bpf_skb_change_tail, struct sk_buff *, skb, u32, new_len,
2192 u32 max_len = __bpf_skb_max_len(skb);
2193 u32 min_len = __bpf_skb_min_len(skb);
2196 if (unlikely(flags || new_len > max_len || new_len < min_len))
2198 if (skb->encapsulation)
2201 /* The basic idea of this helper is that it's performing the
2202 * needed work to either grow or trim an skb, and eBPF program
2203 * rewrites the rest via helpers like bpf_skb_store_bytes(),
2204 * bpf_lX_csum_replace() and others rather than passing a raw
2205 * buffer here. This one is a slow path helper and intended
2206 * for replies with control messages.
2208 * Like in bpf_skb_change_proto(), we want to keep this rather
2209 * minimal and without protocol specifics so that we are able
2210 * to separate concerns as in bpf_skb_store_bytes() should only
2211 * be the one responsible for writing buffers.
2213 * It's really expected to be a slow path operation here for
2214 * control message replies, so we're implicitly linearizing,
2215 * uncloning and drop offloads from the skb by this.
2217 ret = __bpf_try_make_writable(skb, skb->len);
2219 if (new_len > skb->len)
2220 ret = bpf_skb_grow_rcsum(skb, new_len);
2221 else if (new_len < skb->len)
2222 ret = bpf_skb_trim_rcsum(skb, new_len);
2223 if (!ret && skb_is_gso(skb))
2227 bpf_compute_data_end(skb);
2231 static const struct bpf_func_proto bpf_skb_change_tail_proto = {
2232 .func = bpf_skb_change_tail,
2234 .ret_type = RET_INTEGER,
2235 .arg1_type = ARG_PTR_TO_CTX,
2236 .arg2_type = ARG_ANYTHING,
2237 .arg3_type = ARG_ANYTHING,
2240 BPF_CALL_3(bpf_skb_change_head, struct sk_buff *, skb, u32, head_room,
2243 u32 max_len = __bpf_skb_max_len(skb);
2244 u32 new_len = skb->len + head_room;
2247 if (unlikely(flags || (!skb_is_gso(skb) && new_len > max_len) ||
2248 new_len < skb->len))
2251 ret = skb_cow(skb, head_room);
2253 /* Idea for this helper is that we currently only
2254 * allow to expand on mac header. This means that
2255 * skb->protocol network header, etc, stay as is.
2256 * Compared to bpf_skb_change_tail(), we're more
2257 * flexible due to not needing to linearize or
2258 * reset GSO. Intention for this helper is to be
2259 * used by an L3 skb that needs to push mac header
2260 * for redirection into L2 device.
2262 __skb_push(skb, head_room);
2263 memset(skb->data, 0, head_room);
2264 skb_reset_mac_header(skb);
2267 bpf_compute_data_end(skb);
2271 static const struct bpf_func_proto bpf_skb_change_head_proto = {
2272 .func = bpf_skb_change_head,
2274 .ret_type = RET_INTEGER,
2275 .arg1_type = ARG_PTR_TO_CTX,
2276 .arg2_type = ARG_ANYTHING,
2277 .arg3_type = ARG_ANYTHING,
2280 BPF_CALL_2(bpf_xdp_adjust_head, struct xdp_buff *, xdp, int, offset)
2282 void *data = xdp->data + offset;
2284 if (unlikely(data < xdp->data_hard_start ||
2285 data > xdp->data_end - ETH_HLEN))
2293 static const struct bpf_func_proto bpf_xdp_adjust_head_proto = {
2294 .func = bpf_xdp_adjust_head,
2296 .ret_type = RET_INTEGER,
2297 .arg1_type = ARG_PTR_TO_CTX,
2298 .arg2_type = ARG_ANYTHING,
2301 bool bpf_helper_changes_pkt_data(void *func)
2303 if (func == bpf_skb_vlan_push ||
2304 func == bpf_skb_vlan_pop ||
2305 func == bpf_skb_store_bytes ||
2306 func == bpf_skb_change_proto ||
2307 func == bpf_skb_change_head ||
2308 func == bpf_skb_change_tail ||
2309 func == bpf_skb_pull_data ||
2310 func == bpf_clone_redirect ||
2311 func == bpf_l3_csum_replace ||
2312 func == bpf_l4_csum_replace ||
2313 func == bpf_xdp_adjust_head)
2319 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2320 unsigned long off, unsigned long len)
2322 void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2326 if (ptr != dst_buff)
2327 memcpy(dst_buff, ptr, len);
2332 BPF_CALL_5(bpf_skb_event_output, struct sk_buff *, skb, struct bpf_map *, map,
2333 u64, flags, void *, meta, u64, meta_size)
2335 u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2337 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2339 if (unlikely(skb_size > skb->len))
2342 return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2346 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2347 .func = bpf_skb_event_output,
2349 .ret_type = RET_INTEGER,
2350 .arg1_type = ARG_PTR_TO_CTX,
2351 .arg2_type = ARG_CONST_MAP_PTR,
2352 .arg3_type = ARG_ANYTHING,
2353 .arg4_type = ARG_PTR_TO_MEM,
2354 .arg5_type = ARG_CONST_SIZE,
2357 static unsigned short bpf_tunnel_key_af(u64 flags)
2359 return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2362 BPF_CALL_4(bpf_skb_get_tunnel_key, struct sk_buff *, skb, struct bpf_tunnel_key *, to,
2363 u32, size, u64, flags)
2365 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2366 u8 compat[sizeof(struct bpf_tunnel_key)];
2370 if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2374 if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2378 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2381 case offsetof(struct bpf_tunnel_key, tunnel_label):
2382 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2384 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2385 /* Fixup deprecated structure layouts here, so we have
2386 * a common path later on.
2388 if (ip_tunnel_info_af(info) != AF_INET)
2391 to = (struct bpf_tunnel_key *)compat;
2398 to->tunnel_id = be64_to_cpu(info->key.tun_id);
2399 to->tunnel_tos = info->key.tos;
2400 to->tunnel_ttl = info->key.ttl;
2402 if (flags & BPF_F_TUNINFO_IPV6) {
2403 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2404 sizeof(to->remote_ipv6));
2405 to->tunnel_label = be32_to_cpu(info->key.label);
2407 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2410 if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2411 memcpy(to_orig, to, size);
2415 memset(to_orig, 0, size);
2419 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2420 .func = bpf_skb_get_tunnel_key,
2422 .ret_type = RET_INTEGER,
2423 .arg1_type = ARG_PTR_TO_CTX,
2424 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
2425 .arg3_type = ARG_CONST_SIZE,
2426 .arg4_type = ARG_ANYTHING,
2429 BPF_CALL_3(bpf_skb_get_tunnel_opt, struct sk_buff *, skb, u8 *, to, u32, size)
2431 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2434 if (unlikely(!info ||
2435 !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2439 if (unlikely(size < info->options_len)) {
2444 ip_tunnel_info_opts_get(to, info);
2445 if (size > info->options_len)
2446 memset(to + info->options_len, 0, size - info->options_len);
2448 return info->options_len;
2450 memset(to, 0, size);
2454 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2455 .func = bpf_skb_get_tunnel_opt,
2457 .ret_type = RET_INTEGER,
2458 .arg1_type = ARG_PTR_TO_CTX,
2459 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
2460 .arg3_type = ARG_CONST_SIZE,
2463 static struct metadata_dst __percpu *md_dst;
2465 BPF_CALL_4(bpf_skb_set_tunnel_key, struct sk_buff *, skb,
2466 const struct bpf_tunnel_key *, from, u32, size, u64, flags)
2468 struct metadata_dst *md = this_cpu_ptr(md_dst);
2469 u8 compat[sizeof(struct bpf_tunnel_key)];
2470 struct ip_tunnel_info *info;
2472 if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2473 BPF_F_DONT_FRAGMENT)))
2475 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2477 case offsetof(struct bpf_tunnel_key, tunnel_label):
2478 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2479 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2480 /* Fixup deprecated structure layouts here, so we have
2481 * a common path later on.
2483 memcpy(compat, from, size);
2484 memset(compat + size, 0, sizeof(compat) - size);
2485 from = (const struct bpf_tunnel_key *) compat;
2491 if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
2496 dst_hold((struct dst_entry *) md);
2497 skb_dst_set(skb, (struct dst_entry *) md);
2499 info = &md->u.tun_info;
2500 info->mode = IP_TUNNEL_INFO_TX;
2502 info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
2503 if (flags & BPF_F_DONT_FRAGMENT)
2504 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
2506 info->key.tun_id = cpu_to_be64(from->tunnel_id);
2507 info->key.tos = from->tunnel_tos;
2508 info->key.ttl = from->tunnel_ttl;
2510 if (flags & BPF_F_TUNINFO_IPV6) {
2511 info->mode |= IP_TUNNEL_INFO_IPV6;
2512 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
2513 sizeof(from->remote_ipv6));
2514 info->key.label = cpu_to_be32(from->tunnel_label) &
2515 IPV6_FLOWLABEL_MASK;
2517 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
2518 if (flags & BPF_F_ZERO_CSUM_TX)
2519 info->key.tun_flags &= ~TUNNEL_CSUM;
2525 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
2526 .func = bpf_skb_set_tunnel_key,
2528 .ret_type = RET_INTEGER,
2529 .arg1_type = ARG_PTR_TO_CTX,
2530 .arg2_type = ARG_PTR_TO_MEM,
2531 .arg3_type = ARG_CONST_SIZE,
2532 .arg4_type = ARG_ANYTHING,
2535 BPF_CALL_3(bpf_skb_set_tunnel_opt, struct sk_buff *, skb,
2536 const u8 *, from, u32, size)
2538 struct ip_tunnel_info *info = skb_tunnel_info(skb);
2539 const struct metadata_dst *md = this_cpu_ptr(md_dst);
2541 if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
2543 if (unlikely(size > IP_TUNNEL_OPTS_MAX))
2546 ip_tunnel_info_opts_set(info, from, size);
2551 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
2552 .func = bpf_skb_set_tunnel_opt,
2554 .ret_type = RET_INTEGER,
2555 .arg1_type = ARG_PTR_TO_CTX,
2556 .arg2_type = ARG_PTR_TO_MEM,
2557 .arg3_type = ARG_CONST_SIZE,
2560 static const struct bpf_func_proto *
2561 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
2564 /* Race is not possible, since it's called from verifier
2565 * that is holding verifier mutex.
2567 md_dst = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
2575 case BPF_FUNC_skb_set_tunnel_key:
2576 return &bpf_skb_set_tunnel_key_proto;
2577 case BPF_FUNC_skb_set_tunnel_opt:
2578 return &bpf_skb_set_tunnel_opt_proto;
2584 BPF_CALL_3(bpf_skb_under_cgroup, struct sk_buff *, skb, struct bpf_map *, map,
2587 struct bpf_array *array = container_of(map, struct bpf_array, map);
2588 struct cgroup *cgrp;
2591 sk = skb_to_full_sk(skb);
2592 if (!sk || !sk_fullsock(sk))
2594 if (unlikely(idx >= array->map.max_entries))
2597 cgrp = READ_ONCE(array->ptrs[idx]);
2598 if (unlikely(!cgrp))
2601 return sk_under_cgroup_hierarchy(sk, cgrp);
2604 static const struct bpf_func_proto bpf_skb_under_cgroup_proto = {
2605 .func = bpf_skb_under_cgroup,
2607 .ret_type = RET_INTEGER,
2608 .arg1_type = ARG_PTR_TO_CTX,
2609 .arg2_type = ARG_CONST_MAP_PTR,
2610 .arg3_type = ARG_ANYTHING,
2613 static unsigned long bpf_xdp_copy(void *dst_buff, const void *src_buff,
2614 unsigned long off, unsigned long len)
2616 memcpy(dst_buff, src_buff + off, len);
2620 BPF_CALL_5(bpf_xdp_event_output, struct xdp_buff *, xdp, struct bpf_map *, map,
2621 u64, flags, void *, meta, u64, meta_size)
2623 u64 xdp_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2625 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2627 if (unlikely(xdp_size > (unsigned long)(xdp->data_end - xdp->data)))
2630 return bpf_event_output(map, flags, meta, meta_size, xdp->data,
2631 xdp_size, bpf_xdp_copy);
2634 static const struct bpf_func_proto bpf_xdp_event_output_proto = {
2635 .func = bpf_xdp_event_output,
2637 .ret_type = RET_INTEGER,
2638 .arg1_type = ARG_PTR_TO_CTX,
2639 .arg2_type = ARG_CONST_MAP_PTR,
2640 .arg3_type = ARG_ANYTHING,
2641 .arg4_type = ARG_PTR_TO_MEM,
2642 .arg5_type = ARG_CONST_SIZE,
2645 BPF_CALL_1(bpf_get_socket_cookie, struct sk_buff *, skb)
2647 return skb->sk ? sock_gen_cookie(skb->sk) : 0;
2650 static const struct bpf_func_proto bpf_get_socket_cookie_proto = {
2651 .func = bpf_get_socket_cookie,
2653 .ret_type = RET_INTEGER,
2654 .arg1_type = ARG_PTR_TO_CTX,
2657 BPF_CALL_1(bpf_get_socket_uid, struct sk_buff *, skb)
2659 struct sock *sk = sk_to_full_sk(skb->sk);
2662 if (!sk || !sk_fullsock(sk))
2664 kuid = sock_net_uid(sock_net(sk), sk);
2665 return from_kuid_munged(sock_net(sk)->user_ns, kuid);
2668 static const struct bpf_func_proto bpf_get_socket_uid_proto = {
2669 .func = bpf_get_socket_uid,
2671 .ret_type = RET_INTEGER,
2672 .arg1_type = ARG_PTR_TO_CTX,
2675 static const struct bpf_func_proto *
2676 bpf_base_func_proto(enum bpf_func_id func_id)
2679 case BPF_FUNC_map_lookup_elem:
2680 return &bpf_map_lookup_elem_proto;
2681 case BPF_FUNC_map_update_elem:
2682 return &bpf_map_update_elem_proto;
2683 case BPF_FUNC_map_delete_elem:
2684 return &bpf_map_delete_elem_proto;
2685 case BPF_FUNC_get_prandom_u32:
2686 return &bpf_get_prandom_u32_proto;
2687 case BPF_FUNC_get_smp_processor_id:
2688 return &bpf_get_raw_smp_processor_id_proto;
2689 case BPF_FUNC_get_numa_node_id:
2690 return &bpf_get_numa_node_id_proto;
2691 case BPF_FUNC_tail_call:
2692 return &bpf_tail_call_proto;
2693 case BPF_FUNC_ktime_get_ns:
2694 return &bpf_ktime_get_ns_proto;
2695 case BPF_FUNC_trace_printk:
2696 if (capable(CAP_SYS_ADMIN))
2697 return bpf_get_trace_printk_proto();
2703 static const struct bpf_func_proto *
2704 sk_filter_func_proto(enum bpf_func_id func_id)
2707 case BPF_FUNC_skb_load_bytes:
2708 return &bpf_skb_load_bytes_proto;
2709 case BPF_FUNC_get_socket_cookie:
2710 return &bpf_get_socket_cookie_proto;
2711 case BPF_FUNC_get_socket_uid:
2712 return &bpf_get_socket_uid_proto;
2714 return bpf_base_func_proto(func_id);
2718 static const struct bpf_func_proto *
2719 tc_cls_act_func_proto(enum bpf_func_id func_id)
2722 case BPF_FUNC_skb_store_bytes:
2723 return &bpf_skb_store_bytes_proto;
2724 case BPF_FUNC_skb_load_bytes:
2725 return &bpf_skb_load_bytes_proto;
2726 case BPF_FUNC_skb_pull_data:
2727 return &bpf_skb_pull_data_proto;
2728 case BPF_FUNC_csum_diff:
2729 return &bpf_csum_diff_proto;
2730 case BPF_FUNC_csum_update:
2731 return &bpf_csum_update_proto;
2732 case BPF_FUNC_l3_csum_replace:
2733 return &bpf_l3_csum_replace_proto;
2734 case BPF_FUNC_l4_csum_replace:
2735 return &bpf_l4_csum_replace_proto;
2736 case BPF_FUNC_clone_redirect:
2737 return &bpf_clone_redirect_proto;
2738 case BPF_FUNC_get_cgroup_classid:
2739 return &bpf_get_cgroup_classid_proto;
2740 case BPF_FUNC_skb_vlan_push:
2741 return &bpf_skb_vlan_push_proto;
2742 case BPF_FUNC_skb_vlan_pop:
2743 return &bpf_skb_vlan_pop_proto;
2744 case BPF_FUNC_skb_change_proto:
2745 return &bpf_skb_change_proto_proto;
2746 case BPF_FUNC_skb_change_type:
2747 return &bpf_skb_change_type_proto;
2748 case BPF_FUNC_skb_change_tail:
2749 return &bpf_skb_change_tail_proto;
2750 case BPF_FUNC_skb_get_tunnel_key:
2751 return &bpf_skb_get_tunnel_key_proto;
2752 case BPF_FUNC_skb_set_tunnel_key:
2753 return bpf_get_skb_set_tunnel_proto(func_id);
2754 case BPF_FUNC_skb_get_tunnel_opt:
2755 return &bpf_skb_get_tunnel_opt_proto;
2756 case BPF_FUNC_skb_set_tunnel_opt:
2757 return bpf_get_skb_set_tunnel_proto(func_id);
2758 case BPF_FUNC_redirect:
2759 return &bpf_redirect_proto;
2760 case BPF_FUNC_get_route_realm:
2761 return &bpf_get_route_realm_proto;
2762 case BPF_FUNC_get_hash_recalc:
2763 return &bpf_get_hash_recalc_proto;
2764 case BPF_FUNC_set_hash_invalid:
2765 return &bpf_set_hash_invalid_proto;
2766 case BPF_FUNC_set_hash:
2767 return &bpf_set_hash_proto;
2768 case BPF_FUNC_perf_event_output:
2769 return &bpf_skb_event_output_proto;
2770 case BPF_FUNC_get_smp_processor_id:
2771 return &bpf_get_smp_processor_id_proto;
2772 case BPF_FUNC_skb_under_cgroup:
2773 return &bpf_skb_under_cgroup_proto;
2774 case BPF_FUNC_get_socket_cookie:
2775 return &bpf_get_socket_cookie_proto;
2776 case BPF_FUNC_get_socket_uid:
2777 return &bpf_get_socket_uid_proto;
2779 return bpf_base_func_proto(func_id);
2783 static const struct bpf_func_proto *
2784 xdp_func_proto(enum bpf_func_id func_id)
2787 case BPF_FUNC_perf_event_output:
2788 return &bpf_xdp_event_output_proto;
2789 case BPF_FUNC_get_smp_processor_id:
2790 return &bpf_get_smp_processor_id_proto;
2791 case BPF_FUNC_xdp_adjust_head:
2792 return &bpf_xdp_adjust_head_proto;
2794 return bpf_base_func_proto(func_id);
2798 static const struct bpf_func_proto *
2799 lwt_inout_func_proto(enum bpf_func_id func_id)
2802 case BPF_FUNC_skb_load_bytes:
2803 return &bpf_skb_load_bytes_proto;
2804 case BPF_FUNC_skb_pull_data:
2805 return &bpf_skb_pull_data_proto;
2806 case BPF_FUNC_csum_diff:
2807 return &bpf_csum_diff_proto;
2808 case BPF_FUNC_get_cgroup_classid:
2809 return &bpf_get_cgroup_classid_proto;
2810 case BPF_FUNC_get_route_realm:
2811 return &bpf_get_route_realm_proto;
2812 case BPF_FUNC_get_hash_recalc:
2813 return &bpf_get_hash_recalc_proto;
2814 case BPF_FUNC_perf_event_output:
2815 return &bpf_skb_event_output_proto;
2816 case BPF_FUNC_get_smp_processor_id:
2817 return &bpf_get_smp_processor_id_proto;
2818 case BPF_FUNC_skb_under_cgroup:
2819 return &bpf_skb_under_cgroup_proto;
2821 return bpf_base_func_proto(func_id);
2825 static const struct bpf_func_proto *
2826 lwt_xmit_func_proto(enum bpf_func_id func_id)
2829 case BPF_FUNC_skb_get_tunnel_key:
2830 return &bpf_skb_get_tunnel_key_proto;
2831 case BPF_FUNC_skb_set_tunnel_key:
2832 return bpf_get_skb_set_tunnel_proto(func_id);
2833 case BPF_FUNC_skb_get_tunnel_opt:
2834 return &bpf_skb_get_tunnel_opt_proto;
2835 case BPF_FUNC_skb_set_tunnel_opt:
2836 return bpf_get_skb_set_tunnel_proto(func_id);
2837 case BPF_FUNC_redirect:
2838 return &bpf_redirect_proto;
2839 case BPF_FUNC_clone_redirect:
2840 return &bpf_clone_redirect_proto;
2841 case BPF_FUNC_skb_change_tail:
2842 return &bpf_skb_change_tail_proto;
2843 case BPF_FUNC_skb_change_head:
2844 return &bpf_skb_change_head_proto;
2845 case BPF_FUNC_skb_store_bytes:
2846 return &bpf_skb_store_bytes_proto;
2847 case BPF_FUNC_csum_update:
2848 return &bpf_csum_update_proto;
2849 case BPF_FUNC_l3_csum_replace:
2850 return &bpf_l3_csum_replace_proto;
2851 case BPF_FUNC_l4_csum_replace:
2852 return &bpf_l4_csum_replace_proto;
2853 case BPF_FUNC_set_hash_invalid:
2854 return &bpf_set_hash_invalid_proto;
2856 return lwt_inout_func_proto(func_id);
2860 static void __set_access_aux_info(int off, struct bpf_insn_access_aux *info)
2862 info->ctx_field_size = 4;
2864 case offsetof(struct __sk_buff, pkt_type) ...
2865 offsetof(struct __sk_buff, pkt_type) + sizeof(__u32) - 1:
2866 case offsetof(struct __sk_buff, vlan_present) ...
2867 offsetof(struct __sk_buff, vlan_present) + sizeof(__u32) - 1:
2868 info->converted_op_size = 1;
2870 case offsetof(struct __sk_buff, queue_mapping) ...
2871 offsetof(struct __sk_buff, queue_mapping) + sizeof(__u32) - 1:
2872 case offsetof(struct __sk_buff, protocol) ...
2873 offsetof(struct __sk_buff, protocol) + sizeof(__u32) - 1:
2874 case offsetof(struct __sk_buff, vlan_tci) ...
2875 offsetof(struct __sk_buff, vlan_tci) + sizeof(__u32) - 1:
2876 case offsetof(struct __sk_buff, vlan_proto) ...
2877 offsetof(struct __sk_buff, vlan_proto) + sizeof(__u32) - 1:
2878 case offsetof(struct __sk_buff, tc_index) ...
2879 offsetof(struct __sk_buff, tc_index) + sizeof(__u32) - 1:
2880 case offsetof(struct __sk_buff, tc_classid) ...
2881 offsetof(struct __sk_buff, tc_classid) + sizeof(__u32) - 1:
2882 info->converted_op_size = 2;
2885 info->converted_op_size = 4;
2889 static bool __is_valid_access(int off, int size, enum bpf_access_type type,
2890 struct bpf_insn_access_aux *info)
2892 if (off < 0 || off >= sizeof(struct __sk_buff))
2895 /* The verifier guarantees that size > 0. */
2896 if (off % size != 0)
2900 case offsetof(struct __sk_buff, cb[0]) ...
2901 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
2903 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32))
2906 case offsetof(struct __sk_buff, data) ...
2907 offsetof(struct __sk_buff, data) + sizeof(__u32) - 1:
2908 if (size != sizeof(__u32))
2910 info->reg_type = PTR_TO_PACKET;
2912 case offsetof(struct __sk_buff, data_end) ...
2913 offsetof(struct __sk_buff, data_end) + sizeof(__u32) - 1:
2914 if (size != sizeof(__u32))
2916 info->reg_type = PTR_TO_PACKET_END;
2919 if (type == BPF_WRITE) {
2920 if (size != sizeof(__u32))
2925 /* permit narrower load for not cb/data/data_end fields */
2926 #ifdef __LITTLE_ENDIAN
2927 allowed = (off & 0x3) == 0 && size <= 4 && (size & (size - 1)) == 0;
2929 allowed = (off & 0x3) + size == 4 && size <= 4 && (size & (size - 1)) == 0;
2933 __set_access_aux_info(off, info);
2940 static bool sk_filter_is_valid_access(int off, int size,
2941 enum bpf_access_type type,
2942 struct bpf_insn_access_aux *info)
2945 case offsetof(struct __sk_buff, tc_classid) ...
2946 offsetof(struct __sk_buff, tc_classid) + sizeof(__u32) - 1:
2947 case offsetof(struct __sk_buff, data) ...
2948 offsetof(struct __sk_buff, data) + sizeof(__u32) - 1:
2949 case offsetof(struct __sk_buff, data_end) ...
2950 offsetof(struct __sk_buff, data_end) + sizeof(__u32) - 1:
2954 if (type == BPF_WRITE) {
2956 case offsetof(struct __sk_buff, cb[0]) ...
2957 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
2964 return __is_valid_access(off, size, type, info);
2967 static bool lwt_is_valid_access(int off, int size,
2968 enum bpf_access_type type,
2969 struct bpf_insn_access_aux *info)
2972 case offsetof(struct __sk_buff, tc_classid) ...
2973 offsetof(struct __sk_buff, tc_classid) + sizeof(__u32) - 1:
2977 if (type == BPF_WRITE) {
2979 case offsetof(struct __sk_buff, mark):
2980 case offsetof(struct __sk_buff, priority):
2981 case offsetof(struct __sk_buff, cb[0]) ...
2982 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
2989 return __is_valid_access(off, size, type, info);
2992 static bool sock_filter_is_valid_access(int off, int size,
2993 enum bpf_access_type type,
2994 struct bpf_insn_access_aux *info)
2996 if (type == BPF_WRITE) {
2998 case offsetof(struct bpf_sock, bound_dev_if):
3005 if (off < 0 || off + size > sizeof(struct bpf_sock))
3007 /* The verifier guarantees that size > 0. */
3008 if (off % size != 0)
3010 if (size != sizeof(__u32))
3016 static int tc_cls_act_prologue(struct bpf_insn *insn_buf, bool direct_write,
3017 const struct bpf_prog *prog)
3019 struct bpf_insn *insn = insn_buf;
3024 /* if (!skb->cloned)
3027 * (Fast-path, otherwise approximation that we might be
3028 * a clone, do the rest in helper.)
3030 *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_6, BPF_REG_1, CLONED_OFFSET());
3031 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_6, CLONED_MASK);
3032 *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_6, 0, 7);
3034 /* ret = bpf_skb_pull_data(skb, 0); */
3035 *insn++ = BPF_MOV64_REG(BPF_REG_6, BPF_REG_1);
3036 *insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_2, BPF_REG_2);
3037 *insn++ = BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
3038 BPF_FUNC_skb_pull_data);
3041 * return TC_ACT_SHOT;
3043 *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2);
3044 *insn++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_0, TC_ACT_SHOT);
3045 *insn++ = BPF_EXIT_INSN();
3048 *insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_6);
3050 *insn++ = prog->insnsi[0];
3052 return insn - insn_buf;
3055 static bool tc_cls_act_is_valid_access(int off, int size,
3056 enum bpf_access_type type,
3057 struct bpf_insn_access_aux *info)
3059 if (type == BPF_WRITE) {
3061 case offsetof(struct __sk_buff, mark):
3062 case offsetof(struct __sk_buff, tc_index):
3063 case offsetof(struct __sk_buff, priority):
3064 case offsetof(struct __sk_buff, cb[0]) ...
3065 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
3066 case offsetof(struct __sk_buff, tc_classid):
3073 return __is_valid_access(off, size, type, info);
3076 static bool __is_valid_xdp_access(int off, int size)
3078 if (off < 0 || off >= sizeof(struct xdp_md))
3080 if (off % size != 0)
3082 if (size != sizeof(__u32))
3088 static bool xdp_is_valid_access(int off, int size,
3089 enum bpf_access_type type,
3090 struct bpf_insn_access_aux *info)
3092 if (type == BPF_WRITE)
3096 case offsetof(struct xdp_md, data):
3097 info->reg_type = PTR_TO_PACKET;
3099 case offsetof(struct xdp_md, data_end):
3100 info->reg_type = PTR_TO_PACKET_END;
3104 return __is_valid_xdp_access(off, size);
3107 void bpf_warn_invalid_xdp_action(u32 act)
3109 WARN_ONCE(1, "Illegal XDP return value %u, expect packet loss\n", act);
3111 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
3113 static u32 bpf_convert_ctx_access(enum bpf_access_type type,
3114 const struct bpf_insn *si,
3115 struct bpf_insn *insn_buf,
3116 struct bpf_prog *prog)
3118 struct bpf_insn *insn = insn_buf;
3122 case offsetof(struct __sk_buff, len):
3123 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
3125 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3126 offsetof(struct sk_buff, len));
3129 case offsetof(struct __sk_buff, protocol):
3130 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
3132 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3133 offsetof(struct sk_buff, protocol));
3136 case offsetof(struct __sk_buff, vlan_proto):
3137 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
3139 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3140 offsetof(struct sk_buff, vlan_proto));
3143 case offsetof(struct __sk_buff, priority):
3144 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
3146 if (type == BPF_WRITE)
3147 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3148 offsetof(struct sk_buff, priority));
3150 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3151 offsetof(struct sk_buff, priority));
3154 case offsetof(struct __sk_buff, ingress_ifindex):
3155 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
3157 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3158 offsetof(struct sk_buff, skb_iif));
3161 case offsetof(struct __sk_buff, ifindex):
3162 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
3164 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
3165 si->dst_reg, si->src_reg,
3166 offsetof(struct sk_buff, dev));
3167 *insn++ = BPF_JMP_IMM(BPF_JEQ, si->dst_reg, 0, 1);
3168 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3169 offsetof(struct net_device, ifindex));
3172 case offsetof(struct __sk_buff, hash):
3173 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
3175 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3176 offsetof(struct sk_buff, hash));
3179 case offsetof(struct __sk_buff, mark):
3180 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
3182 if (type == BPF_WRITE)
3183 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3184 offsetof(struct sk_buff, mark));
3186 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3187 offsetof(struct sk_buff, mark));
3190 case offsetof(struct __sk_buff, pkt_type):
3191 return convert_skb_access(SKF_AD_PKTTYPE, si->dst_reg,
3194 case offsetof(struct __sk_buff, queue_mapping):
3195 return convert_skb_access(SKF_AD_QUEUE, si->dst_reg,
3198 case offsetof(struct __sk_buff, vlan_present):
3199 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
3200 si->dst_reg, si->src_reg, insn);
3202 case offsetof(struct __sk_buff, vlan_tci):
3203 return convert_skb_access(SKF_AD_VLAN_TAG,
3204 si->dst_reg, si->src_reg, insn);
3206 case offsetof(struct __sk_buff, cb[0]) ...
3207 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
3208 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
3209 BUILD_BUG_ON((offsetof(struct sk_buff, cb) +
3210 offsetof(struct qdisc_skb_cb, data)) %
3213 prog->cb_access = 1;
3215 off -= offsetof(struct __sk_buff, cb[0]);
3216 off += offsetof(struct sk_buff, cb);
3217 off += offsetof(struct qdisc_skb_cb, data);
3218 if (type == BPF_WRITE)
3219 *insn++ = BPF_STX_MEM(BPF_SIZE(si->code), si->dst_reg,
3222 *insn++ = BPF_LDX_MEM(BPF_SIZE(si->code), si->dst_reg,
3226 case offsetof(struct __sk_buff, tc_classid):
3227 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, tc_classid) != 2);
3230 off -= offsetof(struct __sk_buff, tc_classid);
3231 off += offsetof(struct sk_buff, cb);
3232 off += offsetof(struct qdisc_skb_cb, tc_classid);
3233 if (type == BPF_WRITE)
3234 *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg,
3237 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg,
3241 case offsetof(struct __sk_buff, data):
3242 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, data),
3243 si->dst_reg, si->src_reg,
3244 offsetof(struct sk_buff, data));
3247 case offsetof(struct __sk_buff, data_end):
3249 off -= offsetof(struct __sk_buff, data_end);
3250 off += offsetof(struct sk_buff, cb);
3251 off += offsetof(struct bpf_skb_data_end, data_end);
3252 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
3256 case offsetof(struct __sk_buff, tc_index):
3257 #ifdef CONFIG_NET_SCHED
3258 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
3260 if (type == BPF_WRITE)
3261 *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg, si->src_reg,
3262 offsetof(struct sk_buff, tc_index));
3264 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3265 offsetof(struct sk_buff, tc_index));
3267 if (type == BPF_WRITE)
3268 *insn++ = BPF_MOV64_REG(si->dst_reg, si->dst_reg);
3270 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3274 case offsetof(struct __sk_buff, napi_id):
3275 #if defined(CONFIG_NET_RX_BUSY_POLL)
3276 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, napi_id) != 4);
3278 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3279 offsetof(struct sk_buff, napi_id));
3280 *insn++ = BPF_JMP_IMM(BPF_JGE, si->dst_reg, MIN_NAPI_ID, 1);
3281 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3283 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3288 return insn - insn_buf;
3291 static u32 sock_filter_convert_ctx_access(enum bpf_access_type type,
3292 const struct bpf_insn *si,
3293 struct bpf_insn *insn_buf,
3294 struct bpf_prog *prog)
3296 struct bpf_insn *insn = insn_buf;
3299 case offsetof(struct bpf_sock, bound_dev_if):
3300 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_bound_dev_if) != 4);
3302 if (type == BPF_WRITE)
3303 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3304 offsetof(struct sock, sk_bound_dev_if));
3306 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3307 offsetof(struct sock, sk_bound_dev_if));
3310 case offsetof(struct bpf_sock, family):
3311 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_family) != 2);
3313 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3314 offsetof(struct sock, sk_family));
3317 case offsetof(struct bpf_sock, type):
3318 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3319 offsetof(struct sock, __sk_flags_offset));
3320 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_TYPE_MASK);
3321 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_TYPE_SHIFT);
3324 case offsetof(struct bpf_sock, protocol):
3325 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3326 offsetof(struct sock, __sk_flags_offset));
3327 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_PROTO_MASK);
3328 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_PROTO_SHIFT);
3332 return insn - insn_buf;
3335 static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type,
3336 const struct bpf_insn *si,
3337 struct bpf_insn *insn_buf,
3338 struct bpf_prog *prog)
3340 struct bpf_insn *insn = insn_buf;
3343 case offsetof(struct __sk_buff, ifindex):
3344 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
3346 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
3347 si->dst_reg, si->src_reg,
3348 offsetof(struct sk_buff, dev));
3349 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3350 offsetof(struct net_device, ifindex));
3353 return bpf_convert_ctx_access(type, si, insn_buf, prog);
3356 return insn - insn_buf;
3359 static u32 xdp_convert_ctx_access(enum bpf_access_type type,
3360 const struct bpf_insn *si,
3361 struct bpf_insn *insn_buf,
3362 struct bpf_prog *prog)
3364 struct bpf_insn *insn = insn_buf;
3367 case offsetof(struct xdp_md, data):
3368 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data),
3369 si->dst_reg, si->src_reg,
3370 offsetof(struct xdp_buff, data));
3372 case offsetof(struct xdp_md, data_end):
3373 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_end),
3374 si->dst_reg, si->src_reg,
3375 offsetof(struct xdp_buff, data_end));
3379 return insn - insn_buf;
3382 const struct bpf_verifier_ops sk_filter_prog_ops = {
3383 .get_func_proto = sk_filter_func_proto,
3384 .is_valid_access = sk_filter_is_valid_access,
3385 .convert_ctx_access = bpf_convert_ctx_access,
3388 const struct bpf_verifier_ops tc_cls_act_prog_ops = {
3389 .get_func_proto = tc_cls_act_func_proto,
3390 .is_valid_access = tc_cls_act_is_valid_access,
3391 .convert_ctx_access = tc_cls_act_convert_ctx_access,
3392 .gen_prologue = tc_cls_act_prologue,
3393 .test_run = bpf_prog_test_run_skb,
3396 const struct bpf_verifier_ops xdp_prog_ops = {
3397 .get_func_proto = xdp_func_proto,
3398 .is_valid_access = xdp_is_valid_access,
3399 .convert_ctx_access = xdp_convert_ctx_access,
3400 .test_run = bpf_prog_test_run_xdp,
3403 const struct bpf_verifier_ops cg_skb_prog_ops = {
3404 .get_func_proto = sk_filter_func_proto,
3405 .is_valid_access = sk_filter_is_valid_access,
3406 .convert_ctx_access = bpf_convert_ctx_access,
3407 .test_run = bpf_prog_test_run_skb,
3410 const struct bpf_verifier_ops lwt_inout_prog_ops = {
3411 .get_func_proto = lwt_inout_func_proto,
3412 .is_valid_access = lwt_is_valid_access,
3413 .convert_ctx_access = bpf_convert_ctx_access,
3414 .test_run = bpf_prog_test_run_skb,
3417 const struct bpf_verifier_ops lwt_xmit_prog_ops = {
3418 .get_func_proto = lwt_xmit_func_proto,
3419 .is_valid_access = lwt_is_valid_access,
3420 .convert_ctx_access = bpf_convert_ctx_access,
3421 .gen_prologue = tc_cls_act_prologue,
3422 .test_run = bpf_prog_test_run_skb,
3425 const struct bpf_verifier_ops cg_sock_prog_ops = {
3426 .get_func_proto = bpf_base_func_proto,
3427 .is_valid_access = sock_filter_is_valid_access,
3428 .convert_ctx_access = sock_filter_convert_ctx_access,
3431 int sk_detach_filter(struct sock *sk)
3434 struct sk_filter *filter;
3436 if (sock_flag(sk, SOCK_FILTER_LOCKED))
3439 filter = rcu_dereference_protected(sk->sk_filter,
3440 lockdep_sock_is_held(sk));
3442 RCU_INIT_POINTER(sk->sk_filter, NULL);
3443 sk_filter_uncharge(sk, filter);
3449 EXPORT_SYMBOL_GPL(sk_detach_filter);
3451 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
3454 struct sock_fprog_kern *fprog;
3455 struct sk_filter *filter;
3459 filter = rcu_dereference_protected(sk->sk_filter,
3460 lockdep_sock_is_held(sk));
3464 /* We're copying the filter that has been originally attached,
3465 * so no conversion/decode needed anymore. eBPF programs that
3466 * have no original program cannot be dumped through this.
3469 fprog = filter->prog->orig_prog;
3475 /* User space only enquires number of filter blocks. */
3479 if (len < fprog->len)
3483 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
3486 /* Instead of bytes, the API requests to return the number