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>
58 #include <linux/bpf_trace.h>
61 * sk_filter_trim_cap - run a packet through a socket filter
62 * @sk: sock associated with &sk_buff
63 * @skb: buffer to filter
64 * @cap: limit on how short the eBPF program may trim the packet
66 * Run the eBPF program and then cut skb->data to correct size returned by
67 * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
68 * than pkt_len we keep whole skb->data. This is the socket level
69 * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
70 * be accepted or -EPERM if the packet should be tossed.
73 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
76 struct sk_filter *filter;
79 * If the skb was allocated from pfmemalloc reserves, only
80 * allow SOCK_MEMALLOC sockets to use it as this socket is
83 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC)) {
84 NET_INC_STATS(sock_net(sk), LINUX_MIB_PFMEMALLOCDROP);
87 err = BPF_CGROUP_RUN_PROG_INET_INGRESS(sk, skb);
91 err = security_sock_rcv_skb(sk, skb);
96 filter = rcu_dereference(sk->sk_filter);
98 struct sock *save_sk = skb->sk;
102 pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
104 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
110 EXPORT_SYMBOL(sk_filter_trim_cap);
112 BPF_CALL_1(__skb_get_pay_offset, struct sk_buff *, skb)
114 return skb_get_poff(skb);
117 BPF_CALL_3(__skb_get_nlattr, struct sk_buff *, skb, u32, a, u32, x)
121 if (skb_is_nonlinear(skb))
124 if (skb->len < sizeof(struct nlattr))
127 if (a > skb->len - sizeof(struct nlattr))
130 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
132 return (void *) nla - (void *) skb->data;
137 BPF_CALL_3(__skb_get_nlattr_nest, struct sk_buff *, skb, u32, a, u32, x)
141 if (skb_is_nonlinear(skb))
144 if (skb->len < sizeof(struct nlattr))
147 if (a > skb->len - sizeof(struct nlattr))
150 nla = (struct nlattr *) &skb->data[a];
151 if (nla->nla_len > skb->len - a)
154 nla = nla_find_nested(nla, x);
156 return (void *) nla - (void *) skb->data;
161 BPF_CALL_0(__get_raw_cpu_id)
163 return raw_smp_processor_id();
166 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
167 .func = __get_raw_cpu_id,
169 .ret_type = RET_INTEGER,
172 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
173 struct bpf_insn *insn_buf)
175 struct bpf_insn *insn = insn_buf;
179 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
181 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
182 offsetof(struct sk_buff, mark));
186 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
187 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
188 #ifdef __BIG_ENDIAN_BITFIELD
189 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
194 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
196 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
197 offsetof(struct sk_buff, queue_mapping));
200 case SKF_AD_VLAN_TAG:
201 case SKF_AD_VLAN_TAG_PRESENT:
202 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
203 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
205 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
206 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
207 offsetof(struct sk_buff, vlan_tci));
208 if (skb_field == SKF_AD_VLAN_TAG) {
209 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
213 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
215 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
220 return insn - insn_buf;
223 static bool convert_bpf_extensions(struct sock_filter *fp,
224 struct bpf_insn **insnp)
226 struct bpf_insn *insn = *insnp;
230 case SKF_AD_OFF + SKF_AD_PROTOCOL:
231 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
233 /* A = *(u16 *) (CTX + offsetof(protocol)) */
234 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
235 offsetof(struct sk_buff, protocol));
236 /* A = ntohs(A) [emitting a nop or swap16] */
237 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
240 case SKF_AD_OFF + SKF_AD_PKTTYPE:
241 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
245 case SKF_AD_OFF + SKF_AD_IFINDEX:
246 case SKF_AD_OFF + SKF_AD_HATYPE:
247 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
248 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
250 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
251 BPF_REG_TMP, BPF_REG_CTX,
252 offsetof(struct sk_buff, dev));
253 /* if (tmp != 0) goto pc + 1 */
254 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
255 *insn++ = BPF_EXIT_INSN();
256 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
257 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
258 offsetof(struct net_device, ifindex));
260 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
261 offsetof(struct net_device, type));
264 case SKF_AD_OFF + SKF_AD_MARK:
265 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
269 case SKF_AD_OFF + SKF_AD_RXHASH:
270 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
272 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
273 offsetof(struct sk_buff, hash));
276 case SKF_AD_OFF + SKF_AD_QUEUE:
277 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
281 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
282 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
283 BPF_REG_A, BPF_REG_CTX, insn);
287 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
288 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
289 BPF_REG_A, BPF_REG_CTX, insn);
293 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
294 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
296 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
297 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
298 offsetof(struct sk_buff, vlan_proto));
299 /* A = ntohs(A) [emitting a nop or swap16] */
300 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
303 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
304 case SKF_AD_OFF + SKF_AD_NLATTR:
305 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
306 case SKF_AD_OFF + SKF_AD_CPU:
307 case SKF_AD_OFF + SKF_AD_RANDOM:
309 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
311 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
313 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
314 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
316 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
317 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
319 case SKF_AD_OFF + SKF_AD_NLATTR:
320 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
322 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
323 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
325 case SKF_AD_OFF + SKF_AD_CPU:
326 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
328 case SKF_AD_OFF + SKF_AD_RANDOM:
329 *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
330 bpf_user_rnd_init_once();
335 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
337 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
341 /* This is just a dummy call to avoid letting the compiler
342 * evict __bpf_call_base() as an optimization. Placed here
343 * where no-one bothers.
345 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
354 * bpf_convert_filter - convert filter program
355 * @prog: the user passed filter program
356 * @len: the length of the user passed filter program
357 * @new_prog: allocated 'struct bpf_prog' or NULL
358 * @new_len: pointer to store length of converted program
360 * Remap 'sock_filter' style classic BPF (cBPF) instruction set to 'bpf_insn'
361 * style extended BPF (eBPF).
362 * Conversion workflow:
364 * 1) First pass for calculating the new program length:
365 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
367 * 2) 2nd pass to remap in two passes: 1st pass finds new
368 * jump offsets, 2nd pass remapping:
369 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
371 static int bpf_convert_filter(struct sock_filter *prog, int len,
372 struct bpf_prog *new_prog, int *new_len)
374 int new_flen = 0, pass = 0, target, i, stack_off;
375 struct bpf_insn *new_insn, *first_insn = NULL;
376 struct sock_filter *fp;
380 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
381 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
383 if (len <= 0 || len > BPF_MAXINSNS)
387 first_insn = new_prog->insnsi;
388 addrs = kcalloc(len, sizeof(*addrs),
389 GFP_KERNEL | __GFP_NOWARN);
395 new_insn = first_insn;
398 /* Classic BPF related prologue emission. */
400 /* Classic BPF expects A and X to be reset first. These need
401 * to be guaranteed to be the first two instructions.
403 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
404 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
406 /* All programs must keep CTX in callee saved BPF_REG_CTX.
407 * In eBPF case it's done by the compiler, here we need to
408 * do this ourself. Initial CTX is present in BPF_REG_ARG1.
410 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
415 for (i = 0; i < len; fp++, i++) {
416 struct bpf_insn tmp_insns[6] = { };
417 struct bpf_insn *insn = tmp_insns;
420 addrs[i] = new_insn - first_insn;
423 /* All arithmetic insns and skb loads map as-is. */
424 case BPF_ALU | BPF_ADD | BPF_X:
425 case BPF_ALU | BPF_ADD | BPF_K:
426 case BPF_ALU | BPF_SUB | BPF_X:
427 case BPF_ALU | BPF_SUB | BPF_K:
428 case BPF_ALU | BPF_AND | BPF_X:
429 case BPF_ALU | BPF_AND | BPF_K:
430 case BPF_ALU | BPF_OR | BPF_X:
431 case BPF_ALU | BPF_OR | BPF_K:
432 case BPF_ALU | BPF_LSH | BPF_X:
433 case BPF_ALU | BPF_LSH | BPF_K:
434 case BPF_ALU | BPF_RSH | BPF_X:
435 case BPF_ALU | BPF_RSH | BPF_K:
436 case BPF_ALU | BPF_XOR | BPF_X:
437 case BPF_ALU | BPF_XOR | BPF_K:
438 case BPF_ALU | BPF_MUL | BPF_X:
439 case BPF_ALU | BPF_MUL | BPF_K:
440 case BPF_ALU | BPF_DIV | BPF_X:
441 case BPF_ALU | BPF_DIV | BPF_K:
442 case BPF_ALU | BPF_MOD | BPF_X:
443 case BPF_ALU | BPF_MOD | BPF_K:
444 case BPF_ALU | BPF_NEG:
445 case BPF_LD | BPF_ABS | BPF_W:
446 case BPF_LD | BPF_ABS | BPF_H:
447 case BPF_LD | BPF_ABS | BPF_B:
448 case BPF_LD | BPF_IND | BPF_W:
449 case BPF_LD | BPF_IND | BPF_H:
450 case BPF_LD | BPF_IND | BPF_B:
451 /* Check for overloaded BPF extension and
452 * directly convert it if found, otherwise
453 * just move on with mapping.
455 if (BPF_CLASS(fp->code) == BPF_LD &&
456 BPF_MODE(fp->code) == BPF_ABS &&
457 convert_bpf_extensions(fp, &insn))
460 if (fp->code == (BPF_ALU | BPF_DIV | BPF_X) ||
461 fp->code == (BPF_ALU | BPF_MOD | BPF_X))
462 *insn++ = BPF_MOV32_REG(BPF_REG_X, BPF_REG_X);
464 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
467 /* Jump transformation cannot use BPF block macros
468 * everywhere as offset calculation and target updates
469 * require a bit more work than the rest, i.e. jump
470 * opcodes map as-is, but offsets need adjustment.
473 #define BPF_EMIT_JMP \
475 const s32 off_min = S16_MIN, off_max = S16_MAX; \
478 if (target >= len || target < 0) \
480 off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
481 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
482 off -= insn - tmp_insns; \
483 /* Reject anything not fitting into insn->off. */ \
484 if (off < off_min || off > off_max) \
489 case BPF_JMP | BPF_JA:
490 target = i + fp->k + 1;
491 insn->code = fp->code;
495 case BPF_JMP | BPF_JEQ | BPF_K:
496 case BPF_JMP | BPF_JEQ | BPF_X:
497 case BPF_JMP | BPF_JSET | BPF_K:
498 case BPF_JMP | BPF_JSET | BPF_X:
499 case BPF_JMP | BPF_JGT | BPF_K:
500 case BPF_JMP | BPF_JGT | BPF_X:
501 case BPF_JMP | BPF_JGE | BPF_K:
502 case BPF_JMP | BPF_JGE | BPF_X:
503 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
504 /* BPF immediates are signed, zero extend
505 * immediate into tmp register and use it
508 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
510 insn->dst_reg = BPF_REG_A;
511 insn->src_reg = BPF_REG_TMP;
514 insn->dst_reg = BPF_REG_A;
516 bpf_src = BPF_SRC(fp->code);
517 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
520 /* Common case where 'jump_false' is next insn. */
522 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
523 target = i + fp->jt + 1;
528 /* Convert some jumps when 'jump_true' is next insn. */
530 switch (BPF_OP(fp->code)) {
532 insn->code = BPF_JMP | BPF_JNE | bpf_src;
535 insn->code = BPF_JMP | BPF_JLE | bpf_src;
538 insn->code = BPF_JMP | BPF_JLT | bpf_src;
544 target = i + fp->jf + 1;
549 /* Other jumps are mapped into two insns: Jxx and JA. */
550 target = i + fp->jt + 1;
551 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
555 insn->code = BPF_JMP | BPF_JA;
556 target = i + fp->jf + 1;
560 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
561 case BPF_LDX | BPF_MSH | BPF_B:
563 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
564 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
565 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
567 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
569 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
571 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
573 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
576 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
577 * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
579 case BPF_RET | BPF_A:
580 case BPF_RET | BPF_K:
581 if (BPF_RVAL(fp->code) == BPF_K)
582 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
584 *insn = BPF_EXIT_INSN();
587 /* Store to stack. */
590 stack_off = fp->k * 4 + 4;
591 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
592 BPF_ST ? BPF_REG_A : BPF_REG_X,
594 /* check_load_and_stores() verifies that classic BPF can
595 * load from stack only after write, so tracking
596 * stack_depth for ST|STX insns is enough
598 if (new_prog && new_prog->aux->stack_depth < stack_off)
599 new_prog->aux->stack_depth = stack_off;
602 /* Load from stack. */
603 case BPF_LD | BPF_MEM:
604 case BPF_LDX | BPF_MEM:
605 stack_off = fp->k * 4 + 4;
606 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
607 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
612 case BPF_LD | BPF_IMM:
613 case BPF_LDX | BPF_IMM:
614 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
615 BPF_REG_A : BPF_REG_X, fp->k);
619 case BPF_MISC | BPF_TAX:
620 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
624 case BPF_MISC | BPF_TXA:
625 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
628 /* A = skb->len or X = skb->len */
629 case BPF_LD | BPF_W | BPF_LEN:
630 case BPF_LDX | BPF_W | BPF_LEN:
631 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
632 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
633 offsetof(struct sk_buff, len));
636 /* Access seccomp_data fields. */
637 case BPF_LDX | BPF_ABS | BPF_W:
638 /* A = *(u32 *) (ctx + K) */
639 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
642 /* Unknown instruction. */
649 memcpy(new_insn, tmp_insns,
650 sizeof(*insn) * (insn - tmp_insns));
651 new_insn += insn - tmp_insns;
655 /* Only calculating new length. */
656 *new_len = new_insn - first_insn;
661 if (new_flen != new_insn - first_insn) {
662 new_flen = new_insn - first_insn;
669 BUG_ON(*new_len != new_flen);
678 * As we dont want to clear mem[] array for each packet going through
679 * __bpf_prog_run(), we check that filter loaded by user never try to read
680 * a cell if not previously written, and we check all branches to be sure
681 * a malicious user doesn't try to abuse us.
683 static int check_load_and_stores(const struct sock_filter *filter, int flen)
685 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
688 BUILD_BUG_ON(BPF_MEMWORDS > 16);
690 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
694 memset(masks, 0xff, flen * sizeof(*masks));
696 for (pc = 0; pc < flen; pc++) {
697 memvalid &= masks[pc];
699 switch (filter[pc].code) {
702 memvalid |= (1 << filter[pc].k);
704 case BPF_LD | BPF_MEM:
705 case BPF_LDX | BPF_MEM:
706 if (!(memvalid & (1 << filter[pc].k))) {
711 case BPF_JMP | BPF_JA:
712 /* A jump must set masks on target */
713 masks[pc + 1 + filter[pc].k] &= memvalid;
716 case BPF_JMP | BPF_JEQ | BPF_K:
717 case BPF_JMP | BPF_JEQ | BPF_X:
718 case BPF_JMP | BPF_JGE | BPF_K:
719 case BPF_JMP | BPF_JGE | BPF_X:
720 case BPF_JMP | BPF_JGT | BPF_K:
721 case BPF_JMP | BPF_JGT | BPF_X:
722 case BPF_JMP | BPF_JSET | BPF_K:
723 case BPF_JMP | BPF_JSET | BPF_X:
724 /* A jump must set masks on targets */
725 masks[pc + 1 + filter[pc].jt] &= memvalid;
726 masks[pc + 1 + filter[pc].jf] &= memvalid;
736 static bool chk_code_allowed(u16 code_to_probe)
738 static const bool codes[] = {
739 /* 32 bit ALU operations */
740 [BPF_ALU | BPF_ADD | BPF_K] = true,
741 [BPF_ALU | BPF_ADD | BPF_X] = true,
742 [BPF_ALU | BPF_SUB | BPF_K] = true,
743 [BPF_ALU | BPF_SUB | BPF_X] = true,
744 [BPF_ALU | BPF_MUL | BPF_K] = true,
745 [BPF_ALU | BPF_MUL | BPF_X] = true,
746 [BPF_ALU | BPF_DIV | BPF_K] = true,
747 [BPF_ALU | BPF_DIV | BPF_X] = true,
748 [BPF_ALU | BPF_MOD | BPF_K] = true,
749 [BPF_ALU | BPF_MOD | BPF_X] = true,
750 [BPF_ALU | BPF_AND | BPF_K] = true,
751 [BPF_ALU | BPF_AND | BPF_X] = true,
752 [BPF_ALU | BPF_OR | BPF_K] = true,
753 [BPF_ALU | BPF_OR | BPF_X] = true,
754 [BPF_ALU | BPF_XOR | BPF_K] = true,
755 [BPF_ALU | BPF_XOR | BPF_X] = true,
756 [BPF_ALU | BPF_LSH | BPF_K] = true,
757 [BPF_ALU | BPF_LSH | BPF_X] = true,
758 [BPF_ALU | BPF_RSH | BPF_K] = true,
759 [BPF_ALU | BPF_RSH | BPF_X] = true,
760 [BPF_ALU | BPF_NEG] = true,
761 /* Load instructions */
762 [BPF_LD | BPF_W | BPF_ABS] = true,
763 [BPF_LD | BPF_H | BPF_ABS] = true,
764 [BPF_LD | BPF_B | BPF_ABS] = true,
765 [BPF_LD | BPF_W | BPF_LEN] = true,
766 [BPF_LD | BPF_W | BPF_IND] = true,
767 [BPF_LD | BPF_H | BPF_IND] = true,
768 [BPF_LD | BPF_B | BPF_IND] = true,
769 [BPF_LD | BPF_IMM] = true,
770 [BPF_LD | BPF_MEM] = true,
771 [BPF_LDX | BPF_W | BPF_LEN] = true,
772 [BPF_LDX | BPF_B | BPF_MSH] = true,
773 [BPF_LDX | BPF_IMM] = true,
774 [BPF_LDX | BPF_MEM] = true,
775 /* Store instructions */
778 /* Misc instructions */
779 [BPF_MISC | BPF_TAX] = true,
780 [BPF_MISC | BPF_TXA] = true,
781 /* Return instructions */
782 [BPF_RET | BPF_K] = true,
783 [BPF_RET | BPF_A] = true,
784 /* Jump instructions */
785 [BPF_JMP | BPF_JA] = true,
786 [BPF_JMP | BPF_JEQ | BPF_K] = true,
787 [BPF_JMP | BPF_JEQ | BPF_X] = true,
788 [BPF_JMP | BPF_JGE | BPF_K] = true,
789 [BPF_JMP | BPF_JGE | BPF_X] = true,
790 [BPF_JMP | BPF_JGT | BPF_K] = true,
791 [BPF_JMP | BPF_JGT | BPF_X] = true,
792 [BPF_JMP | BPF_JSET | BPF_K] = true,
793 [BPF_JMP | BPF_JSET | BPF_X] = true,
796 if (code_to_probe >= ARRAY_SIZE(codes))
799 return codes[code_to_probe];
802 static bool bpf_check_basics_ok(const struct sock_filter *filter,
807 if (flen == 0 || flen > BPF_MAXINSNS)
814 * bpf_check_classic - verify socket filter code
815 * @filter: filter to verify
816 * @flen: length of filter
818 * Check the user's filter code. If we let some ugly
819 * filter code slip through kaboom! The filter must contain
820 * no references or jumps that are out of range, no illegal
821 * instructions, and must end with a RET instruction.
823 * All jumps are forward as they are not signed.
825 * Returns 0 if the rule set is legal or -EINVAL if not.
827 static int bpf_check_classic(const struct sock_filter *filter,
833 /* Check the filter code now */
834 for (pc = 0; pc < flen; pc++) {
835 const struct sock_filter *ftest = &filter[pc];
837 /* May we actually operate on this code? */
838 if (!chk_code_allowed(ftest->code))
841 /* Some instructions need special checks */
842 switch (ftest->code) {
843 case BPF_ALU | BPF_DIV | BPF_K:
844 case BPF_ALU | BPF_MOD | BPF_K:
845 /* Check for division by zero */
849 case BPF_ALU | BPF_LSH | BPF_K:
850 case BPF_ALU | BPF_RSH | BPF_K:
854 case BPF_LD | BPF_MEM:
855 case BPF_LDX | BPF_MEM:
858 /* Check for invalid memory addresses */
859 if (ftest->k >= BPF_MEMWORDS)
862 case BPF_JMP | BPF_JA:
863 /* Note, the large ftest->k might cause loops.
864 * Compare this with conditional jumps below,
865 * where offsets are limited. --ANK (981016)
867 if (ftest->k >= (unsigned int)(flen - pc - 1))
870 case BPF_JMP | BPF_JEQ | BPF_K:
871 case BPF_JMP | BPF_JEQ | BPF_X:
872 case BPF_JMP | BPF_JGE | BPF_K:
873 case BPF_JMP | BPF_JGE | BPF_X:
874 case BPF_JMP | BPF_JGT | BPF_K:
875 case BPF_JMP | BPF_JGT | BPF_X:
876 case BPF_JMP | BPF_JSET | BPF_K:
877 case BPF_JMP | BPF_JSET | BPF_X:
878 /* Both conditionals must be safe */
879 if (pc + ftest->jt + 1 >= flen ||
880 pc + ftest->jf + 1 >= flen)
883 case BPF_LD | BPF_W | BPF_ABS:
884 case BPF_LD | BPF_H | BPF_ABS:
885 case BPF_LD | BPF_B | BPF_ABS:
887 if (bpf_anc_helper(ftest) & BPF_ANC)
889 /* Ancillary operation unknown or unsupported */
890 if (anc_found == false && ftest->k >= SKF_AD_OFF)
895 /* Last instruction must be a RET code */
896 switch (filter[flen - 1].code) {
897 case BPF_RET | BPF_K:
898 case BPF_RET | BPF_A:
899 return check_load_and_stores(filter, flen);
905 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
906 const struct sock_fprog *fprog)
908 unsigned int fsize = bpf_classic_proglen(fprog);
909 struct sock_fprog_kern *fkprog;
911 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
915 fkprog = fp->orig_prog;
916 fkprog->len = fprog->len;
918 fkprog->filter = kmemdup(fp->insns, fsize,
919 GFP_KERNEL | __GFP_NOWARN);
920 if (!fkprog->filter) {
921 kfree(fp->orig_prog);
928 static void bpf_release_orig_filter(struct bpf_prog *fp)
930 struct sock_fprog_kern *fprog = fp->orig_prog;
933 kfree(fprog->filter);
938 static void __bpf_prog_release(struct bpf_prog *prog)
940 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
943 bpf_release_orig_filter(prog);
948 static void __sk_filter_release(struct sk_filter *fp)
950 __bpf_prog_release(fp->prog);
955 * sk_filter_release_rcu - Release a socket filter by rcu_head
956 * @rcu: rcu_head that contains the sk_filter to free
958 static void sk_filter_release_rcu(struct rcu_head *rcu)
960 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
962 __sk_filter_release(fp);
966 * sk_filter_release - release a socket filter
967 * @fp: filter to remove
969 * Remove a filter from a socket and release its resources.
971 static void sk_filter_release(struct sk_filter *fp)
973 if (refcount_dec_and_test(&fp->refcnt))
974 call_rcu(&fp->rcu, sk_filter_release_rcu);
977 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
979 u32 filter_size = bpf_prog_size(fp->prog->len);
981 atomic_sub(filter_size, &sk->sk_omem_alloc);
982 sk_filter_release(fp);
985 /* try to charge the socket memory if there is space available
986 * return true on success
988 static bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
990 u32 filter_size = bpf_prog_size(fp->prog->len);
992 /* same check as in sock_kmalloc() */
993 if (filter_size <= sysctl_optmem_max &&
994 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
995 atomic_add(filter_size, &sk->sk_omem_alloc);
1001 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
1003 if (!refcount_inc_not_zero(&fp->refcnt))
1006 if (!__sk_filter_charge(sk, fp)) {
1007 sk_filter_release(fp);
1013 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
1015 struct sock_filter *old_prog;
1016 struct bpf_prog *old_fp;
1017 int err, new_len, old_len = fp->len;
1019 /* We are free to overwrite insns et al right here as it
1020 * won't be used at this point in time anymore internally
1021 * after the migration to the internal BPF instruction
1024 BUILD_BUG_ON(sizeof(struct sock_filter) !=
1025 sizeof(struct bpf_insn));
1027 /* Conversion cannot happen on overlapping memory areas,
1028 * so we need to keep the user BPF around until the 2nd
1029 * pass. At this time, the user BPF is stored in fp->insns.
1031 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
1032 GFP_KERNEL | __GFP_NOWARN);
1038 /* 1st pass: calculate the new program length. */
1039 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
1043 /* Expand fp for appending the new filter representation. */
1045 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
1047 /* The old_fp is still around in case we couldn't
1048 * allocate new memory, so uncharge on that one.
1057 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1058 err = bpf_convert_filter(old_prog, old_len, fp, &new_len);
1060 /* 2nd bpf_convert_filter() can fail only if it fails
1061 * to allocate memory, remapping must succeed. Note,
1062 * that at this time old_fp has already been released
1067 fp = bpf_prog_select_runtime(fp, &err);
1077 __bpf_prog_release(fp);
1078 return ERR_PTR(err);
1081 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1082 bpf_aux_classic_check_t trans)
1086 fp->bpf_func = NULL;
1089 err = bpf_check_classic(fp->insns, fp->len);
1091 __bpf_prog_release(fp);
1092 return ERR_PTR(err);
1095 /* There might be additional checks and transformations
1096 * needed on classic filters, f.e. in case of seccomp.
1099 err = trans(fp->insns, fp->len);
1101 __bpf_prog_release(fp);
1102 return ERR_PTR(err);
1106 /* Probe if we can JIT compile the filter and if so, do
1107 * the compilation of the filter.
1109 bpf_jit_compile(fp);
1111 /* JIT compiler couldn't process this filter, so do the
1112 * internal BPF translation for the optimized interpreter.
1115 fp = bpf_migrate_filter(fp);
1121 * bpf_prog_create - create an unattached filter
1122 * @pfp: the unattached filter that is created
1123 * @fprog: the filter program
1125 * Create a filter independent of any socket. We first run some
1126 * sanity checks on it to make sure it does not explode on us later.
1127 * If an error occurs or there is insufficient memory for the filter
1128 * a negative errno code is returned. On success the return is zero.
1130 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1132 unsigned int fsize = bpf_classic_proglen(fprog);
1133 struct bpf_prog *fp;
1135 /* Make sure new filter is there and in the right amounts. */
1136 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1139 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1143 memcpy(fp->insns, fprog->filter, fsize);
1145 fp->len = fprog->len;
1146 /* Since unattached filters are not copied back to user
1147 * space through sk_get_filter(), we do not need to hold
1148 * a copy here, and can spare us the work.
1150 fp->orig_prog = NULL;
1152 /* bpf_prepare_filter() already takes care of freeing
1153 * memory in case something goes wrong.
1155 fp = bpf_prepare_filter(fp, NULL);
1162 EXPORT_SYMBOL_GPL(bpf_prog_create);
1165 * bpf_prog_create_from_user - create an unattached filter from user buffer
1166 * @pfp: the unattached filter that is created
1167 * @fprog: the filter program
1168 * @trans: post-classic verifier transformation handler
1169 * @save_orig: save classic BPF program
1171 * This function effectively does the same as bpf_prog_create(), only
1172 * that it builds up its insns buffer from user space provided buffer.
1173 * It also allows for passing a bpf_aux_classic_check_t handler.
1175 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1176 bpf_aux_classic_check_t trans, bool save_orig)
1178 unsigned int fsize = bpf_classic_proglen(fprog);
1179 struct bpf_prog *fp;
1182 /* Make sure new filter is there and in the right amounts. */
1183 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1186 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1190 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1191 __bpf_prog_free(fp);
1195 fp->len = fprog->len;
1196 fp->orig_prog = NULL;
1199 err = bpf_prog_store_orig_filter(fp, fprog);
1201 __bpf_prog_free(fp);
1206 /* bpf_prepare_filter() already takes care of freeing
1207 * memory in case something goes wrong.
1209 fp = bpf_prepare_filter(fp, trans);
1216 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1218 void bpf_prog_destroy(struct bpf_prog *fp)
1220 __bpf_prog_release(fp);
1222 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1224 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1226 struct sk_filter *fp, *old_fp;
1228 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1234 if (!__sk_filter_charge(sk, fp)) {
1238 refcount_set(&fp->refcnt, 1);
1240 old_fp = rcu_dereference_protected(sk->sk_filter,
1241 lockdep_sock_is_held(sk));
1242 rcu_assign_pointer(sk->sk_filter, fp);
1245 sk_filter_uncharge(sk, old_fp);
1250 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1252 struct bpf_prog *old_prog;
1255 if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1258 if (sk_unhashed(sk) && sk->sk_reuseport) {
1259 err = reuseport_alloc(sk);
1262 } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1263 /* The socket wasn't bound with SO_REUSEPORT */
1267 old_prog = reuseport_attach_prog(sk, prog);
1269 bpf_prog_destroy(old_prog);
1275 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1277 unsigned int fsize = bpf_classic_proglen(fprog);
1278 struct bpf_prog *prog;
1281 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1282 return ERR_PTR(-EPERM);
1284 /* Make sure new filter is there and in the right amounts. */
1285 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1286 return ERR_PTR(-EINVAL);
1288 prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1290 return ERR_PTR(-ENOMEM);
1292 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1293 __bpf_prog_free(prog);
1294 return ERR_PTR(-EFAULT);
1297 prog->len = fprog->len;
1299 err = bpf_prog_store_orig_filter(prog, fprog);
1301 __bpf_prog_free(prog);
1302 return ERR_PTR(-ENOMEM);
1305 /* bpf_prepare_filter() already takes care of freeing
1306 * memory in case something goes wrong.
1308 return bpf_prepare_filter(prog, NULL);
1312 * sk_attach_filter - attach a socket filter
1313 * @fprog: the filter program
1314 * @sk: the socket to use
1316 * Attach the user's filter code. We first run some sanity checks on
1317 * it to make sure it does not explode on us later. If an error
1318 * occurs or there is insufficient memory for the filter a negative
1319 * errno code is returned. On success the return is zero.
1321 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1323 struct bpf_prog *prog = __get_filter(fprog, sk);
1327 return PTR_ERR(prog);
1329 err = __sk_attach_prog(prog, sk);
1331 __bpf_prog_release(prog);
1337 EXPORT_SYMBOL_GPL(sk_attach_filter);
1339 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1341 struct bpf_prog *prog = __get_filter(fprog, sk);
1345 return PTR_ERR(prog);
1347 err = __reuseport_attach_prog(prog, sk);
1349 __bpf_prog_release(prog);
1356 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1358 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1359 return ERR_PTR(-EPERM);
1361 return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1364 int sk_attach_bpf(u32 ufd, struct sock *sk)
1366 struct bpf_prog *prog = __get_bpf(ufd, sk);
1370 return PTR_ERR(prog);
1372 err = __sk_attach_prog(prog, sk);
1381 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1383 struct bpf_prog *prog = __get_bpf(ufd, sk);
1387 return PTR_ERR(prog);
1389 err = __reuseport_attach_prog(prog, sk);
1398 struct bpf_scratchpad {
1400 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1401 u8 buff[MAX_BPF_STACK];
1405 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1407 static inline int __bpf_try_make_writable(struct sk_buff *skb,
1408 unsigned int write_len)
1410 return skb_ensure_writable(skb, write_len);
1413 static inline int bpf_try_make_writable(struct sk_buff *skb,
1414 unsigned int write_len)
1416 int err = __bpf_try_make_writable(skb, write_len);
1418 bpf_compute_data_end(skb);
1422 static int bpf_try_make_head_writable(struct sk_buff *skb)
1424 return bpf_try_make_writable(skb, skb_headlen(skb));
1427 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1429 if (skb_at_tc_ingress(skb))
1430 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1433 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1435 if (skb_at_tc_ingress(skb))
1436 skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1439 BPF_CALL_5(bpf_skb_store_bytes, struct sk_buff *, skb, u32, offset,
1440 const void *, from, u32, len, u64, flags)
1444 if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1446 if (unlikely(offset > INT_MAX))
1448 if (unlikely(bpf_try_make_writable(skb, offset + len)))
1451 ptr = skb->data + offset;
1452 if (flags & BPF_F_RECOMPUTE_CSUM)
1453 __skb_postpull_rcsum(skb, ptr, len, offset);
1455 memcpy(ptr, from, len);
1457 if (flags & BPF_F_RECOMPUTE_CSUM)
1458 __skb_postpush_rcsum(skb, ptr, len, offset);
1459 if (flags & BPF_F_INVALIDATE_HASH)
1460 skb_clear_hash(skb);
1465 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1466 .func = bpf_skb_store_bytes,
1468 .ret_type = RET_INTEGER,
1469 .arg1_type = ARG_PTR_TO_CTX,
1470 .arg2_type = ARG_ANYTHING,
1471 .arg3_type = ARG_PTR_TO_MEM,
1472 .arg4_type = ARG_CONST_SIZE,
1473 .arg5_type = ARG_ANYTHING,
1476 BPF_CALL_4(bpf_skb_load_bytes, const struct sk_buff *, skb, u32, offset,
1477 void *, to, u32, len)
1481 if (unlikely(offset > INT_MAX))
1484 ptr = skb_header_pointer(skb, offset, len, to);
1488 memcpy(to, ptr, len);
1496 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1497 .func = bpf_skb_load_bytes,
1499 .ret_type = RET_INTEGER,
1500 .arg1_type = ARG_PTR_TO_CTX,
1501 .arg2_type = ARG_ANYTHING,
1502 .arg3_type = ARG_PTR_TO_UNINIT_MEM,
1503 .arg4_type = ARG_CONST_SIZE,
1506 BPF_CALL_2(bpf_skb_pull_data, struct sk_buff *, skb, u32, len)
1508 /* Idea is the following: should the needed direct read/write
1509 * test fail during runtime, we can pull in more data and redo
1510 * again, since implicitly, we invalidate previous checks here.
1512 * Or, since we know how much we need to make read/writeable,
1513 * this can be done once at the program beginning for direct
1514 * access case. By this we overcome limitations of only current
1515 * headroom being accessible.
1517 return bpf_try_make_writable(skb, len ? : skb_headlen(skb));
1520 static const struct bpf_func_proto bpf_skb_pull_data_proto = {
1521 .func = bpf_skb_pull_data,
1523 .ret_type = RET_INTEGER,
1524 .arg1_type = ARG_PTR_TO_CTX,
1525 .arg2_type = ARG_ANYTHING,
1528 BPF_CALL_5(bpf_l3_csum_replace, struct sk_buff *, skb, u32, offset,
1529 u64, from, u64, to, u64, flags)
1533 if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1535 if (unlikely(offset > 0xffff || offset & 1))
1537 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1540 ptr = (__sum16 *)(skb->data + offset);
1541 switch (flags & BPF_F_HDR_FIELD_MASK) {
1543 if (unlikely(from != 0))
1546 csum_replace_by_diff(ptr, to);
1549 csum_replace2(ptr, from, to);
1552 csum_replace4(ptr, from, to);
1561 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1562 .func = bpf_l3_csum_replace,
1564 .ret_type = RET_INTEGER,
1565 .arg1_type = ARG_PTR_TO_CTX,
1566 .arg2_type = ARG_ANYTHING,
1567 .arg3_type = ARG_ANYTHING,
1568 .arg4_type = ARG_ANYTHING,
1569 .arg5_type = ARG_ANYTHING,
1572 BPF_CALL_5(bpf_l4_csum_replace, struct sk_buff *, skb, u32, offset,
1573 u64, from, u64, to, u64, flags)
1575 bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1576 bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1577 bool do_mforce = flags & BPF_F_MARK_ENFORCE;
1580 if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_MARK_ENFORCE |
1581 BPF_F_PSEUDO_HDR | BPF_F_HDR_FIELD_MASK)))
1583 if (unlikely(offset > 0xffff || offset & 1))
1585 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1588 ptr = (__sum16 *)(skb->data + offset);
1589 if (is_mmzero && !do_mforce && !*ptr)
1592 switch (flags & BPF_F_HDR_FIELD_MASK) {
1594 if (unlikely(from != 0))
1597 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1600 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1603 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1609 if (is_mmzero && !*ptr)
1610 *ptr = CSUM_MANGLED_0;
1614 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1615 .func = bpf_l4_csum_replace,
1617 .ret_type = RET_INTEGER,
1618 .arg1_type = ARG_PTR_TO_CTX,
1619 .arg2_type = ARG_ANYTHING,
1620 .arg3_type = ARG_ANYTHING,
1621 .arg4_type = ARG_ANYTHING,
1622 .arg5_type = ARG_ANYTHING,
1625 BPF_CALL_5(bpf_csum_diff, __be32 *, from, u32, from_size,
1626 __be32 *, to, u32, to_size, __wsum, seed)
1628 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1629 u32 diff_size = from_size + to_size;
1632 /* This is quite flexible, some examples:
1634 * from_size == 0, to_size > 0, seed := csum --> pushing data
1635 * from_size > 0, to_size == 0, seed := csum --> pulling data
1636 * from_size > 0, to_size > 0, seed := 0 --> diffing data
1638 * Even for diffing, from_size and to_size don't need to be equal.
1640 if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1641 diff_size > sizeof(sp->diff)))
1644 for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1645 sp->diff[j] = ~from[i];
1646 for (i = 0; i < to_size / sizeof(__be32); i++, j++)
1647 sp->diff[j] = to[i];
1649 return csum_partial(sp->diff, diff_size, seed);
1652 static const struct bpf_func_proto bpf_csum_diff_proto = {
1653 .func = bpf_csum_diff,
1656 .ret_type = RET_INTEGER,
1657 .arg1_type = ARG_PTR_TO_MEM,
1658 .arg2_type = ARG_CONST_SIZE_OR_ZERO,
1659 .arg3_type = ARG_PTR_TO_MEM,
1660 .arg4_type = ARG_CONST_SIZE_OR_ZERO,
1661 .arg5_type = ARG_ANYTHING,
1664 BPF_CALL_2(bpf_csum_update, struct sk_buff *, skb, __wsum, csum)
1666 /* The interface is to be used in combination with bpf_csum_diff()
1667 * for direct packet writes. csum rotation for alignment as well
1668 * as emulating csum_sub() can be done from the eBPF program.
1670 if (skb->ip_summed == CHECKSUM_COMPLETE)
1671 return (skb->csum = csum_add(skb->csum, csum));
1676 static const struct bpf_func_proto bpf_csum_update_proto = {
1677 .func = bpf_csum_update,
1679 .ret_type = RET_INTEGER,
1680 .arg1_type = ARG_PTR_TO_CTX,
1681 .arg2_type = ARG_ANYTHING,
1684 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1686 return dev_forward_skb(dev, skb);
1689 static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
1690 struct sk_buff *skb)
1692 int ret = ____dev_forward_skb(dev, skb);
1696 ret = netif_rx(skb);
1702 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1706 if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1707 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1714 __this_cpu_inc(xmit_recursion);
1715 ret = dev_queue_xmit(skb);
1716 __this_cpu_dec(xmit_recursion);
1721 static int __bpf_redirect_no_mac(struct sk_buff *skb, struct net_device *dev,
1724 unsigned int mlen = skb_network_offset(skb);
1727 __skb_pull(skb, mlen);
1728 if (unlikely(!skb->len)) {
1733 /* At ingress, the mac header has already been pulled once.
1734 * At egress, skb_pospull_rcsum has to be done in case that
1735 * the skb is originated from ingress (i.e. a forwarded skb)
1736 * to ensure that rcsum starts at net header.
1738 if (!skb_at_tc_ingress(skb))
1739 skb_postpull_rcsum(skb, skb_mac_header(skb), mlen);
1741 skb_pop_mac_header(skb);
1742 skb_reset_mac_len(skb);
1743 return flags & BPF_F_INGRESS ?
1744 __bpf_rx_skb_no_mac(dev, skb) : __bpf_tx_skb(dev, skb);
1747 static int __bpf_redirect_common(struct sk_buff *skb, struct net_device *dev,
1750 /* Verify that a link layer header is carried */
1751 if (unlikely(skb->mac_header >= skb->network_header)) {
1756 bpf_push_mac_rcsum(skb);
1757 return flags & BPF_F_INGRESS ?
1758 __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1761 static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
1764 if (dev_is_mac_header_xmit(dev))
1765 return __bpf_redirect_common(skb, dev, flags);
1767 return __bpf_redirect_no_mac(skb, dev, flags);
1770 BPF_CALL_3(bpf_clone_redirect, struct sk_buff *, skb, u32, ifindex, u64, flags)
1772 struct net_device *dev;
1773 struct sk_buff *clone;
1776 if (unlikely(flags & ~(BPF_F_INGRESS)))
1779 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1783 clone = skb_clone(skb, GFP_ATOMIC);
1784 if (unlikely(!clone))
1787 /* For direct write, we need to keep the invariant that the skbs
1788 * we're dealing with need to be uncloned. Should uncloning fail
1789 * here, we need to free the just generated clone to unclone once
1792 ret = bpf_try_make_head_writable(skb);
1793 if (unlikely(ret)) {
1798 return __bpf_redirect(clone, dev, flags);
1801 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1802 .func = bpf_clone_redirect,
1804 .ret_type = RET_INTEGER,
1805 .arg1_type = ARG_PTR_TO_CTX,
1806 .arg2_type = ARG_ANYTHING,
1807 .arg3_type = ARG_ANYTHING,
1810 struct redirect_info {
1813 struct bpf_map *map;
1814 struct bpf_map *map_to_flush;
1815 unsigned long map_owner;
1818 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1820 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
1822 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1824 if (unlikely(flags & ~(BPF_F_INGRESS)))
1827 ri->ifindex = ifindex;
1830 return TC_ACT_REDIRECT;
1833 int skb_do_redirect(struct sk_buff *skb)
1835 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1836 struct net_device *dev;
1838 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1840 if (unlikely(!dev)) {
1845 return __bpf_redirect(skb, dev, ri->flags);
1848 static const struct bpf_func_proto bpf_redirect_proto = {
1849 .func = bpf_redirect,
1851 .ret_type = RET_INTEGER,
1852 .arg1_type = ARG_ANYTHING,
1853 .arg2_type = ARG_ANYTHING,
1856 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
1857 struct bpf_map *, map, u32, key, u64, flags)
1859 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1861 /* If user passes invalid input drop the packet. */
1862 if (unlikely(flags))
1866 tcb->bpf.flags = flags;
1872 struct sock *do_sk_redirect_map(struct sk_buff *skb)
1874 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1875 struct sock *sk = NULL;
1878 sk = __sock_map_lookup_elem(tcb->bpf.map, tcb->bpf.key);
1881 tcb->bpf.map = NULL;
1887 static const struct bpf_func_proto bpf_sk_redirect_map_proto = {
1888 .func = bpf_sk_redirect_map,
1890 .ret_type = RET_INTEGER,
1891 .arg1_type = ARG_PTR_TO_CTX,
1892 .arg2_type = ARG_CONST_MAP_PTR,
1893 .arg3_type = ARG_ANYTHING,
1894 .arg4_type = ARG_ANYTHING,
1897 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
1899 return task_get_classid(skb);
1902 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1903 .func = bpf_get_cgroup_classid,
1905 .ret_type = RET_INTEGER,
1906 .arg1_type = ARG_PTR_TO_CTX,
1909 BPF_CALL_1(bpf_get_route_realm, const struct sk_buff *, skb)
1911 return dst_tclassid(skb);
1914 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1915 .func = bpf_get_route_realm,
1917 .ret_type = RET_INTEGER,
1918 .arg1_type = ARG_PTR_TO_CTX,
1921 BPF_CALL_1(bpf_get_hash_recalc, struct sk_buff *, skb)
1923 /* If skb_clear_hash() was called due to mangling, we can
1924 * trigger SW recalculation here. Later access to hash
1925 * can then use the inline skb->hash via context directly
1926 * instead of calling this helper again.
1928 return skb_get_hash(skb);
1931 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1932 .func = bpf_get_hash_recalc,
1934 .ret_type = RET_INTEGER,
1935 .arg1_type = ARG_PTR_TO_CTX,
1938 BPF_CALL_1(bpf_set_hash_invalid, struct sk_buff *, skb)
1940 /* After all direct packet write, this can be used once for
1941 * triggering a lazy recalc on next skb_get_hash() invocation.
1943 skb_clear_hash(skb);
1947 static const struct bpf_func_proto bpf_set_hash_invalid_proto = {
1948 .func = bpf_set_hash_invalid,
1950 .ret_type = RET_INTEGER,
1951 .arg1_type = ARG_PTR_TO_CTX,
1954 BPF_CALL_2(bpf_set_hash, struct sk_buff *, skb, u32, hash)
1956 /* Set user specified hash as L4(+), so that it gets returned
1957 * on skb_get_hash() call unless BPF prog later on triggers a
1960 __skb_set_sw_hash(skb, hash, true);
1964 static const struct bpf_func_proto bpf_set_hash_proto = {
1965 .func = bpf_set_hash,
1967 .ret_type = RET_INTEGER,
1968 .arg1_type = ARG_PTR_TO_CTX,
1969 .arg2_type = ARG_ANYTHING,
1972 BPF_CALL_3(bpf_skb_vlan_push, struct sk_buff *, skb, __be16, vlan_proto,
1977 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1978 vlan_proto != htons(ETH_P_8021AD)))
1979 vlan_proto = htons(ETH_P_8021Q);
1981 bpf_push_mac_rcsum(skb);
1982 ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1983 bpf_pull_mac_rcsum(skb);
1985 bpf_compute_data_end(skb);
1989 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1990 .func = bpf_skb_vlan_push,
1992 .ret_type = RET_INTEGER,
1993 .arg1_type = ARG_PTR_TO_CTX,
1994 .arg2_type = ARG_ANYTHING,
1995 .arg3_type = ARG_ANYTHING,
1997 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1999 BPF_CALL_1(bpf_skb_vlan_pop, struct sk_buff *, skb)
2003 bpf_push_mac_rcsum(skb);
2004 ret = skb_vlan_pop(skb);
2005 bpf_pull_mac_rcsum(skb);
2007 bpf_compute_data_end(skb);
2011 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
2012 .func = bpf_skb_vlan_pop,
2014 .ret_type = RET_INTEGER,
2015 .arg1_type = ARG_PTR_TO_CTX,
2017 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
2019 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
2021 /* Caller already did skb_cow() with len as headroom,
2022 * so no need to do it here.
2025 memmove(skb->data, skb->data + len, off);
2026 memset(skb->data + off, 0, len);
2028 /* No skb_postpush_rcsum(skb, skb->data + off, len)
2029 * needed here as it does not change the skb->csum
2030 * result for checksum complete when summing over
2036 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
2040 /* skb_ensure_writable() is not needed here, as we're
2041 * already working on an uncloned skb.
2043 if (unlikely(!pskb_may_pull(skb, off + len)))
2046 old_data = skb->data;
2047 __skb_pull(skb, len);
2048 skb_postpull_rcsum(skb, old_data + off, len);
2049 memmove(skb->data, old_data, off);
2054 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
2056 bool trans_same = skb->transport_header == skb->network_header;
2059 /* There's no need for __skb_push()/__skb_pull() pair to
2060 * get to the start of the mac header as we're guaranteed
2061 * to always start from here under eBPF.
2063 ret = bpf_skb_generic_push(skb, off, len);
2065 skb->mac_header -= len;
2066 skb->network_header -= len;
2068 skb->transport_header = skb->network_header;
2074 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
2076 bool trans_same = skb->transport_header == skb->network_header;
2079 /* Same here, __skb_push()/__skb_pull() pair not needed. */
2080 ret = bpf_skb_generic_pop(skb, off, len);
2082 skb->mac_header += len;
2083 skb->network_header += len;
2085 skb->transport_header = skb->network_header;
2091 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
2093 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2094 u32 off = skb_mac_header_len(skb);
2097 ret = skb_cow(skb, len_diff);
2098 if (unlikely(ret < 0))
2101 ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2102 if (unlikely(ret < 0))
2105 if (skb_is_gso(skb)) {
2106 /* SKB_GSO_TCPV4 needs to be changed into
2109 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2110 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
2111 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV6;
2114 /* Due to IPv6 header, MSS needs to be downgraded. */
2115 skb_shinfo(skb)->gso_size -= len_diff;
2116 /* Header must be checked, and gso_segs recomputed. */
2117 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2118 skb_shinfo(skb)->gso_segs = 0;
2121 skb->protocol = htons(ETH_P_IPV6);
2122 skb_clear_hash(skb);
2127 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
2129 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2130 u32 off = skb_mac_header_len(skb);
2133 ret = skb_unclone(skb, GFP_ATOMIC);
2134 if (unlikely(ret < 0))
2137 ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2138 if (unlikely(ret < 0))
2141 if (skb_is_gso(skb)) {
2142 /* SKB_GSO_TCPV6 needs to be changed into
2145 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
2146 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
2147 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
2150 /* Due to IPv4 header, MSS can be upgraded. */
2151 skb_shinfo(skb)->gso_size += len_diff;
2152 /* Header must be checked, and gso_segs recomputed. */
2153 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2154 skb_shinfo(skb)->gso_segs = 0;
2157 skb->protocol = htons(ETH_P_IP);
2158 skb_clear_hash(skb);
2163 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
2165 __be16 from_proto = skb->protocol;
2167 if (from_proto == htons(ETH_P_IP) &&
2168 to_proto == htons(ETH_P_IPV6))
2169 return bpf_skb_proto_4_to_6(skb);
2171 if (from_proto == htons(ETH_P_IPV6) &&
2172 to_proto == htons(ETH_P_IP))
2173 return bpf_skb_proto_6_to_4(skb);
2178 BPF_CALL_3(bpf_skb_change_proto, struct sk_buff *, skb, __be16, proto,
2183 if (unlikely(flags))
2186 /* General idea is that this helper does the basic groundwork
2187 * needed for changing the protocol, and eBPF program fills the
2188 * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
2189 * and other helpers, rather than passing a raw buffer here.
2191 * The rationale is to keep this minimal and without a need to
2192 * deal with raw packet data. F.e. even if we would pass buffers
2193 * here, the program still needs to call the bpf_lX_csum_replace()
2194 * helpers anyway. Plus, this way we keep also separation of
2195 * concerns, since f.e. bpf_skb_store_bytes() should only take
2198 * Currently, additional options and extension header space are
2199 * not supported, but flags register is reserved so we can adapt
2200 * that. For offloads, we mark packet as dodgy, so that headers
2201 * need to be verified first.
2203 ret = bpf_skb_proto_xlat(skb, proto);
2204 bpf_compute_data_end(skb);
2208 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
2209 .func = bpf_skb_change_proto,
2211 .ret_type = RET_INTEGER,
2212 .arg1_type = ARG_PTR_TO_CTX,
2213 .arg2_type = ARG_ANYTHING,
2214 .arg3_type = ARG_ANYTHING,
2217 BPF_CALL_2(bpf_skb_change_type, struct sk_buff *, skb, u32, pkt_type)
2219 /* We only allow a restricted subset to be changed for now. */
2220 if (unlikely(!skb_pkt_type_ok(skb->pkt_type) ||
2221 !skb_pkt_type_ok(pkt_type)))
2224 skb->pkt_type = pkt_type;
2228 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2229 .func = bpf_skb_change_type,
2231 .ret_type = RET_INTEGER,
2232 .arg1_type = ARG_PTR_TO_CTX,
2233 .arg2_type = ARG_ANYTHING,
2236 static u32 bpf_skb_net_base_len(const struct sk_buff *skb)
2238 switch (skb->protocol) {
2239 case htons(ETH_P_IP):
2240 return sizeof(struct iphdr);
2241 case htons(ETH_P_IPV6):
2242 return sizeof(struct ipv6hdr);
2248 static int bpf_skb_net_grow(struct sk_buff *skb, u32 len_diff)
2250 u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2253 ret = skb_cow(skb, len_diff);
2254 if (unlikely(ret < 0))
2257 ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2258 if (unlikely(ret < 0))
2261 if (skb_is_gso(skb)) {
2262 /* Due to header grow, MSS needs to be downgraded. */
2263 skb_shinfo(skb)->gso_size -= len_diff;
2264 /* Header must be checked, and gso_segs recomputed. */
2265 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2266 skb_shinfo(skb)->gso_segs = 0;
2272 static int bpf_skb_net_shrink(struct sk_buff *skb, u32 len_diff)
2274 u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2277 ret = skb_unclone(skb, GFP_ATOMIC);
2278 if (unlikely(ret < 0))
2281 ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2282 if (unlikely(ret < 0))
2285 if (skb_is_gso(skb)) {
2286 /* Due to header shrink, MSS can be upgraded. */
2287 skb_shinfo(skb)->gso_size += len_diff;
2288 /* Header must be checked, and gso_segs recomputed. */
2289 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2290 skb_shinfo(skb)->gso_segs = 0;
2296 #define BPF_SKB_MAX_LEN SKB_MAX_ALLOC
2298 static int bpf_skb_adjust_net(struct sk_buff *skb, s32 len_diff)
2300 bool trans_same = skb->transport_header == skb->network_header;
2301 u32 len_cur, len_diff_abs = abs(len_diff);
2302 u32 len_min = bpf_skb_net_base_len(skb);
2303 u32 len_max = BPF_SKB_MAX_LEN;
2304 __be16 proto = skb->protocol;
2305 bool shrink = len_diff < 0;
2308 if (unlikely(len_diff_abs > 0xfffU))
2310 if (unlikely(proto != htons(ETH_P_IP) &&
2311 proto != htons(ETH_P_IPV6)))
2314 len_cur = skb->len - skb_network_offset(skb);
2315 if (skb_transport_header_was_set(skb) && !trans_same)
2316 len_cur = skb_network_header_len(skb);
2317 if ((shrink && (len_diff_abs >= len_cur ||
2318 len_cur - len_diff_abs < len_min)) ||
2319 (!shrink && (skb->len + len_diff_abs > len_max &&
2323 ret = shrink ? bpf_skb_net_shrink(skb, len_diff_abs) :
2324 bpf_skb_net_grow(skb, len_diff_abs);
2326 bpf_compute_data_end(skb);
2330 BPF_CALL_4(bpf_skb_adjust_room, struct sk_buff *, skb, s32, len_diff,
2331 u32, mode, u64, flags)
2333 if (unlikely(flags))
2335 if (likely(mode == BPF_ADJ_ROOM_NET))
2336 return bpf_skb_adjust_net(skb, len_diff);
2341 static const struct bpf_func_proto bpf_skb_adjust_room_proto = {
2342 .func = bpf_skb_adjust_room,
2344 .ret_type = RET_INTEGER,
2345 .arg1_type = ARG_PTR_TO_CTX,
2346 .arg2_type = ARG_ANYTHING,
2347 .arg3_type = ARG_ANYTHING,
2348 .arg4_type = ARG_ANYTHING,
2351 static u32 __bpf_skb_min_len(const struct sk_buff *skb)
2353 u32 min_len = skb_network_offset(skb);
2355 if (skb_transport_header_was_set(skb))
2356 min_len = skb_transport_offset(skb);
2357 if (skb->ip_summed == CHECKSUM_PARTIAL)
2358 min_len = skb_checksum_start_offset(skb) +
2359 skb->csum_offset + sizeof(__sum16);
2363 static int bpf_skb_grow_rcsum(struct sk_buff *skb, unsigned int new_len)
2365 unsigned int old_len = skb->len;
2368 ret = __skb_grow_rcsum(skb, new_len);
2370 memset(skb->data + old_len, 0, new_len - old_len);
2374 static int bpf_skb_trim_rcsum(struct sk_buff *skb, unsigned int new_len)
2376 return __skb_trim_rcsum(skb, new_len);
2379 BPF_CALL_3(bpf_skb_change_tail, struct sk_buff *, skb, u32, new_len,
2382 u32 max_len = BPF_SKB_MAX_LEN;
2383 u32 min_len = __bpf_skb_min_len(skb);
2386 if (unlikely(flags || new_len > max_len || new_len < min_len))
2388 if (skb->encapsulation)
2391 /* The basic idea of this helper is that it's performing the
2392 * needed work to either grow or trim an skb, and eBPF program
2393 * rewrites the rest via helpers like bpf_skb_store_bytes(),
2394 * bpf_lX_csum_replace() and others rather than passing a raw
2395 * buffer here. This one is a slow path helper and intended
2396 * for replies with control messages.
2398 * Like in bpf_skb_change_proto(), we want to keep this rather
2399 * minimal and without protocol specifics so that we are able
2400 * to separate concerns as in bpf_skb_store_bytes() should only
2401 * be the one responsible for writing buffers.
2403 * It's really expected to be a slow path operation here for
2404 * control message replies, so we're implicitly linearizing,
2405 * uncloning and drop offloads from the skb by this.
2407 ret = __bpf_try_make_writable(skb, skb->len);
2409 if (new_len > skb->len)
2410 ret = bpf_skb_grow_rcsum(skb, new_len);
2411 else if (new_len < skb->len)
2412 ret = bpf_skb_trim_rcsum(skb, new_len);
2413 if (!ret && skb_is_gso(skb))
2417 bpf_compute_data_end(skb);
2421 static const struct bpf_func_proto bpf_skb_change_tail_proto = {
2422 .func = bpf_skb_change_tail,
2424 .ret_type = RET_INTEGER,
2425 .arg1_type = ARG_PTR_TO_CTX,
2426 .arg2_type = ARG_ANYTHING,
2427 .arg3_type = ARG_ANYTHING,
2430 BPF_CALL_3(bpf_skb_change_head, struct sk_buff *, skb, u32, head_room,
2433 u32 max_len = BPF_SKB_MAX_LEN;
2434 u32 new_len = skb->len + head_room;
2437 if (unlikely(flags || (!skb_is_gso(skb) && new_len > max_len) ||
2438 new_len < skb->len))
2441 ret = skb_cow(skb, head_room);
2443 /* Idea for this helper is that we currently only
2444 * allow to expand on mac header. This means that
2445 * skb->protocol network header, etc, stay as is.
2446 * Compared to bpf_skb_change_tail(), we're more
2447 * flexible due to not needing to linearize or
2448 * reset GSO. Intention for this helper is to be
2449 * used by an L3 skb that needs to push mac header
2450 * for redirection into L2 device.
2452 __skb_push(skb, head_room);
2453 memset(skb->data, 0, head_room);
2454 skb_reset_mac_header(skb);
2455 skb_reset_mac_len(skb);
2458 bpf_compute_data_end(skb);
2462 static const struct bpf_func_proto bpf_skb_change_head_proto = {
2463 .func = bpf_skb_change_head,
2465 .ret_type = RET_INTEGER,
2466 .arg1_type = ARG_PTR_TO_CTX,
2467 .arg2_type = ARG_ANYTHING,
2468 .arg3_type = ARG_ANYTHING,
2471 BPF_CALL_2(bpf_xdp_adjust_head, struct xdp_buff *, xdp, int, offset)
2473 void *data = xdp->data + offset;
2475 if (unlikely(data < xdp->data_hard_start ||
2476 data > xdp->data_end - ETH_HLEN))
2484 static const struct bpf_func_proto bpf_xdp_adjust_head_proto = {
2485 .func = bpf_xdp_adjust_head,
2487 .ret_type = RET_INTEGER,
2488 .arg1_type = ARG_PTR_TO_CTX,
2489 .arg2_type = ARG_ANYTHING,
2492 static int __bpf_tx_xdp(struct net_device *dev,
2493 struct bpf_map *map,
2494 struct xdp_buff *xdp,
2499 if (!dev->netdev_ops->ndo_xdp_xmit) {
2503 err = dev->netdev_ops->ndo_xdp_xmit(dev, xdp);
2507 __dev_map_insert_ctx(map, index);
2509 dev->netdev_ops->ndo_xdp_flush(dev);
2513 void xdp_do_flush_map(void)
2515 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2516 struct bpf_map *map = ri->map_to_flush;
2518 ri->map_to_flush = NULL;
2520 __dev_map_flush(map);
2522 EXPORT_SYMBOL_GPL(xdp_do_flush_map);
2524 static inline bool xdp_map_invalid(const struct bpf_prog *xdp_prog,
2527 return (unsigned long)xdp_prog->aux != aux;
2530 static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
2531 struct bpf_prog *xdp_prog)
2533 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2534 unsigned long map_owner = ri->map_owner;
2535 struct bpf_map *map = ri->map;
2536 struct net_device *fwd = NULL;
2537 u32 index = ri->ifindex;
2544 if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2550 fwd = __dev_map_lookup_elem(map, index);
2555 if (ri->map_to_flush && ri->map_to_flush != map)
2558 err = __bpf_tx_xdp(fwd, map, xdp, index);
2562 ri->map_to_flush = map;
2563 _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
2566 _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
2570 int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
2571 struct bpf_prog *xdp_prog)
2573 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2574 struct net_device *fwd;
2575 u32 index = ri->ifindex;
2579 return xdp_do_redirect_map(dev, xdp, xdp_prog);
2581 fwd = dev_get_by_index_rcu(dev_net(dev), index);
2583 if (unlikely(!fwd)) {
2588 err = __bpf_tx_xdp(fwd, NULL, xdp, 0);
2592 _trace_xdp_redirect(dev, xdp_prog, index);
2595 _trace_xdp_redirect_err(dev, xdp_prog, index, err);
2598 EXPORT_SYMBOL_GPL(xdp_do_redirect);
2600 int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
2601 struct bpf_prog *xdp_prog)
2603 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2604 unsigned long map_owner = ri->map_owner;
2605 struct bpf_map *map = ri->map;
2606 struct net_device *fwd = NULL;
2607 u32 index = ri->ifindex;
2616 if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2621 fwd = __dev_map_lookup_elem(map, index);
2623 fwd = dev_get_by_index_rcu(dev_net(dev), index);
2625 if (unlikely(!fwd)) {
2630 if (unlikely(!(fwd->flags & IFF_UP))) {
2635 len = fwd->mtu + fwd->hard_header_len + VLAN_HLEN;
2636 if (skb->len > len) {
2642 map ? _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index)
2643 : _trace_xdp_redirect(dev, xdp_prog, index);
2646 map ? _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err)
2647 : _trace_xdp_redirect_err(dev, xdp_prog, index, err);
2650 EXPORT_SYMBOL_GPL(xdp_do_generic_redirect);
2652 BPF_CALL_2(bpf_xdp_redirect, u32, ifindex, u64, flags)
2654 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2656 if (unlikely(flags))
2659 ri->ifindex = ifindex;
2664 return XDP_REDIRECT;
2667 static const struct bpf_func_proto bpf_xdp_redirect_proto = {
2668 .func = bpf_xdp_redirect,
2670 .ret_type = RET_INTEGER,
2671 .arg1_type = ARG_ANYTHING,
2672 .arg2_type = ARG_ANYTHING,
2675 BPF_CALL_4(bpf_xdp_redirect_map, struct bpf_map *, map, u32, ifindex, u64, flags,
2676 unsigned long, map_owner)
2678 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2680 if (unlikely(flags))
2683 ri->ifindex = ifindex;
2686 ri->map_owner = map_owner;
2688 return XDP_REDIRECT;
2691 /* Note, arg4 is hidden from users and populated by the verifier
2692 * with the right pointer.
2694 static const struct bpf_func_proto bpf_xdp_redirect_map_proto = {
2695 .func = bpf_xdp_redirect_map,
2697 .ret_type = RET_INTEGER,
2698 .arg1_type = ARG_CONST_MAP_PTR,
2699 .arg2_type = ARG_ANYTHING,
2700 .arg3_type = ARG_ANYTHING,
2703 bool bpf_helper_changes_pkt_data(void *func)
2705 if (func == bpf_skb_vlan_push ||
2706 func == bpf_skb_vlan_pop ||
2707 func == bpf_skb_store_bytes ||
2708 func == bpf_skb_change_proto ||
2709 func == bpf_skb_change_head ||
2710 func == bpf_skb_change_tail ||
2711 func == bpf_skb_adjust_room ||
2712 func == bpf_skb_pull_data ||
2713 func == bpf_clone_redirect ||
2714 func == bpf_l3_csum_replace ||
2715 func == bpf_l4_csum_replace ||
2716 func == bpf_xdp_adjust_head)
2722 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2723 unsigned long off, unsigned long len)
2725 void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2729 if (ptr != dst_buff)
2730 memcpy(dst_buff, ptr, len);
2735 BPF_CALL_5(bpf_skb_event_output, struct sk_buff *, skb, struct bpf_map *, map,
2736 u64, flags, void *, meta, u64, meta_size)
2738 u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2740 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2742 if (unlikely(skb_size > skb->len))
2745 return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2749 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2750 .func = bpf_skb_event_output,
2752 .ret_type = RET_INTEGER,
2753 .arg1_type = ARG_PTR_TO_CTX,
2754 .arg2_type = ARG_CONST_MAP_PTR,
2755 .arg3_type = ARG_ANYTHING,
2756 .arg4_type = ARG_PTR_TO_MEM,
2757 .arg5_type = ARG_CONST_SIZE,
2760 static unsigned short bpf_tunnel_key_af(u64 flags)
2762 return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2765 BPF_CALL_4(bpf_skb_get_tunnel_key, struct sk_buff *, skb, struct bpf_tunnel_key *, to,
2766 u32, size, u64, flags)
2768 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2769 u8 compat[sizeof(struct bpf_tunnel_key)];
2773 if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2777 if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2781 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2784 case offsetof(struct bpf_tunnel_key, tunnel_label):
2785 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2787 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2788 /* Fixup deprecated structure layouts here, so we have
2789 * a common path later on.
2791 if (ip_tunnel_info_af(info) != AF_INET)
2794 to = (struct bpf_tunnel_key *)compat;
2801 to->tunnel_id = be64_to_cpu(info->key.tun_id);
2802 to->tunnel_tos = info->key.tos;
2803 to->tunnel_ttl = info->key.ttl;
2805 if (flags & BPF_F_TUNINFO_IPV6) {
2806 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2807 sizeof(to->remote_ipv6));
2808 to->tunnel_label = be32_to_cpu(info->key.label);
2810 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2813 if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2814 memcpy(to_orig, to, size);
2818 memset(to_orig, 0, size);
2822 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2823 .func = bpf_skb_get_tunnel_key,
2825 .ret_type = RET_INTEGER,
2826 .arg1_type = ARG_PTR_TO_CTX,
2827 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
2828 .arg3_type = ARG_CONST_SIZE,
2829 .arg4_type = ARG_ANYTHING,
2832 BPF_CALL_3(bpf_skb_get_tunnel_opt, struct sk_buff *, skb, u8 *, to, u32, size)
2834 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2837 if (unlikely(!info ||
2838 !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2842 if (unlikely(size < info->options_len)) {
2847 ip_tunnel_info_opts_get(to, info);
2848 if (size > info->options_len)
2849 memset(to + info->options_len, 0, size - info->options_len);
2851 return info->options_len;
2853 memset(to, 0, size);
2857 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2858 .func = bpf_skb_get_tunnel_opt,
2860 .ret_type = RET_INTEGER,
2861 .arg1_type = ARG_PTR_TO_CTX,
2862 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
2863 .arg3_type = ARG_CONST_SIZE,
2866 static struct metadata_dst __percpu *md_dst;
2868 BPF_CALL_4(bpf_skb_set_tunnel_key, struct sk_buff *, skb,
2869 const struct bpf_tunnel_key *, from, u32, size, u64, flags)
2871 struct metadata_dst *md = this_cpu_ptr(md_dst);
2872 u8 compat[sizeof(struct bpf_tunnel_key)];
2873 struct ip_tunnel_info *info;
2875 if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2876 BPF_F_DONT_FRAGMENT)))
2878 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2880 case offsetof(struct bpf_tunnel_key, tunnel_label):
2881 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2882 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2883 /* Fixup deprecated structure layouts here, so we have
2884 * a common path later on.
2886 memcpy(compat, from, size);
2887 memset(compat + size, 0, sizeof(compat) - size);
2888 from = (const struct bpf_tunnel_key *) compat;
2894 if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
2899 dst_hold((struct dst_entry *) md);
2900 skb_dst_set(skb, (struct dst_entry *) md);
2902 info = &md->u.tun_info;
2903 info->mode = IP_TUNNEL_INFO_TX;
2905 info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
2906 if (flags & BPF_F_DONT_FRAGMENT)
2907 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
2909 info->key.tun_id = cpu_to_be64(from->tunnel_id);
2910 info->key.tos = from->tunnel_tos;
2911 info->key.ttl = from->tunnel_ttl;
2913 if (flags & BPF_F_TUNINFO_IPV6) {
2914 info->mode |= IP_TUNNEL_INFO_IPV6;
2915 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
2916 sizeof(from->remote_ipv6));
2917 info->key.label = cpu_to_be32(from->tunnel_label) &
2918 IPV6_FLOWLABEL_MASK;
2920 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
2921 if (flags & BPF_F_ZERO_CSUM_TX)
2922 info->key.tun_flags &= ~TUNNEL_CSUM;
2928 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
2929 .func = bpf_skb_set_tunnel_key,
2931 .ret_type = RET_INTEGER,
2932 .arg1_type = ARG_PTR_TO_CTX,
2933 .arg2_type = ARG_PTR_TO_MEM,
2934 .arg3_type = ARG_CONST_SIZE,
2935 .arg4_type = ARG_ANYTHING,
2938 BPF_CALL_3(bpf_skb_set_tunnel_opt, struct sk_buff *, skb,
2939 const u8 *, from, u32, size)
2941 struct ip_tunnel_info *info = skb_tunnel_info(skb);
2942 const struct metadata_dst *md = this_cpu_ptr(md_dst);
2944 if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
2946 if (unlikely(size > IP_TUNNEL_OPTS_MAX))
2949 ip_tunnel_info_opts_set(info, from, size);
2954 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
2955 .func = bpf_skb_set_tunnel_opt,
2957 .ret_type = RET_INTEGER,
2958 .arg1_type = ARG_PTR_TO_CTX,
2959 .arg2_type = ARG_PTR_TO_MEM,
2960 .arg3_type = ARG_CONST_SIZE,
2963 static const struct bpf_func_proto *
2964 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
2967 /* Race is not possible, since it's called from verifier
2968 * that is holding verifier mutex.
2970 md_dst = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
2978 case BPF_FUNC_skb_set_tunnel_key:
2979 return &bpf_skb_set_tunnel_key_proto;
2980 case BPF_FUNC_skb_set_tunnel_opt:
2981 return &bpf_skb_set_tunnel_opt_proto;
2987 BPF_CALL_3(bpf_skb_under_cgroup, struct sk_buff *, skb, struct bpf_map *, map,
2990 struct bpf_array *array = container_of(map, struct bpf_array, map);
2991 struct cgroup *cgrp;
2994 sk = skb_to_full_sk(skb);
2995 if (!sk || !sk_fullsock(sk))
2997 if (unlikely(idx >= array->map.max_entries))
3000 cgrp = READ_ONCE(array->ptrs[idx]);
3001 if (unlikely(!cgrp))
3004 return sk_under_cgroup_hierarchy(sk, cgrp);
3007 static const struct bpf_func_proto bpf_skb_under_cgroup_proto = {
3008 .func = bpf_skb_under_cgroup,
3010 .ret_type = RET_INTEGER,
3011 .arg1_type = ARG_PTR_TO_CTX,
3012 .arg2_type = ARG_CONST_MAP_PTR,
3013 .arg3_type = ARG_ANYTHING,
3016 static unsigned long bpf_xdp_copy(void *dst_buff, const void *src_buff,
3017 unsigned long off, unsigned long len)
3019 memcpy(dst_buff, src_buff + off, len);
3023 BPF_CALL_5(bpf_xdp_event_output, struct xdp_buff *, xdp, struct bpf_map *, map,
3024 u64, flags, void *, meta, u64, meta_size)
3026 u64 xdp_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
3028 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
3030 if (unlikely(xdp_size > (unsigned long)(xdp->data_end - xdp->data)))
3033 return bpf_event_output(map, flags, meta, meta_size, xdp->data,
3034 xdp_size, bpf_xdp_copy);
3037 static const struct bpf_func_proto bpf_xdp_event_output_proto = {
3038 .func = bpf_xdp_event_output,
3040 .ret_type = RET_INTEGER,
3041 .arg1_type = ARG_PTR_TO_CTX,
3042 .arg2_type = ARG_CONST_MAP_PTR,
3043 .arg3_type = ARG_ANYTHING,
3044 .arg4_type = ARG_PTR_TO_MEM,
3045 .arg5_type = ARG_CONST_SIZE,
3048 BPF_CALL_1(bpf_get_socket_cookie, struct sk_buff *, skb)
3050 return skb->sk ? sock_gen_cookie(skb->sk) : 0;
3053 static const struct bpf_func_proto bpf_get_socket_cookie_proto = {
3054 .func = bpf_get_socket_cookie,
3056 .ret_type = RET_INTEGER,
3057 .arg1_type = ARG_PTR_TO_CTX,
3060 BPF_CALL_1(bpf_get_socket_uid, struct sk_buff *, skb)
3062 struct sock *sk = sk_to_full_sk(skb->sk);
3065 if (!sk || !sk_fullsock(sk))
3067 kuid = sock_net_uid(sock_net(sk), sk);
3068 return from_kuid_munged(sock_net(sk)->user_ns, kuid);
3071 static const struct bpf_func_proto bpf_get_socket_uid_proto = {
3072 .func = bpf_get_socket_uid,
3074 .ret_type = RET_INTEGER,
3075 .arg1_type = ARG_PTR_TO_CTX,
3078 BPF_CALL_5(bpf_setsockopt, struct bpf_sock_ops_kern *, bpf_sock,
3079 int, level, int, optname, char *, optval, int, optlen)
3081 struct sock *sk = bpf_sock->sk;
3085 if (!sk_fullsock(sk))
3088 if (level == SOL_SOCKET) {
3089 if (optlen != sizeof(int))
3091 val = *((int *)optval);
3093 /* Only some socketops are supported */
3096 val = min_t(u32, val, sysctl_rmem_max);
3097 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
3098 sk->sk_rcvbuf = max_t(int, val * 2, SOCK_MIN_RCVBUF);
3101 val = min_t(u32, val, sysctl_wmem_max);
3102 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
3103 sk->sk_sndbuf = max_t(int, val * 2, SOCK_MIN_SNDBUF);
3105 case SO_MAX_PACING_RATE:
3106 sk->sk_max_pacing_rate = val;
3107 sk->sk_pacing_rate = min(sk->sk_pacing_rate,
3108 sk->sk_max_pacing_rate);
3111 sk->sk_priority = val;
3116 sk->sk_rcvlowat = val ? : 1;
3119 if (sk->sk_mark != val) {
3128 } else if (level == SOL_TCP &&
3129 sk->sk_prot->setsockopt == tcp_setsockopt) {
3130 if (optname == TCP_CONGESTION) {
3131 char name[TCP_CA_NAME_MAX];
3132 bool reinit = bpf_sock->op > BPF_SOCK_OPS_NEEDS_ECN;
3134 strncpy(name, optval, min_t(long, optlen,
3135 TCP_CA_NAME_MAX-1));
3136 name[TCP_CA_NAME_MAX-1] = 0;
3137 ret = tcp_set_congestion_control(sk, name, false,
3140 struct tcp_sock *tp = tcp_sk(sk);
3142 if (optlen != sizeof(int))
3145 val = *((int *)optval);
3146 /* Only some options are supported */
3149 if (val <= 0 || tp->data_segs_out > tp->syn_data)
3154 case TCP_BPF_SNDCWND_CLAMP:
3158 tp->snd_cwnd_clamp = val;
3159 tp->snd_ssthresh = val;
3173 static const struct bpf_func_proto bpf_setsockopt_proto = {
3174 .func = bpf_setsockopt,
3176 .ret_type = RET_INTEGER,
3177 .arg1_type = ARG_PTR_TO_CTX,
3178 .arg2_type = ARG_ANYTHING,
3179 .arg3_type = ARG_ANYTHING,
3180 .arg4_type = ARG_PTR_TO_MEM,
3181 .arg5_type = ARG_CONST_SIZE,
3184 static const struct bpf_func_proto *
3185 bpf_base_func_proto(enum bpf_func_id func_id)
3188 case BPF_FUNC_map_lookup_elem:
3189 return &bpf_map_lookup_elem_proto;
3190 case BPF_FUNC_map_update_elem:
3191 return &bpf_map_update_elem_proto;
3192 case BPF_FUNC_map_delete_elem:
3193 return &bpf_map_delete_elem_proto;
3194 case BPF_FUNC_get_prandom_u32:
3195 return &bpf_get_prandom_u32_proto;
3196 case BPF_FUNC_get_smp_processor_id:
3197 return &bpf_get_raw_smp_processor_id_proto;
3198 case BPF_FUNC_get_numa_node_id:
3199 return &bpf_get_numa_node_id_proto;
3200 case BPF_FUNC_tail_call:
3201 return &bpf_tail_call_proto;
3202 case BPF_FUNC_ktime_get_ns:
3203 return &bpf_ktime_get_ns_proto;
3204 case BPF_FUNC_trace_printk:
3205 if (capable(CAP_SYS_ADMIN))
3206 return bpf_get_trace_printk_proto();
3212 static const struct bpf_func_proto *
3213 sock_filter_func_proto(enum bpf_func_id func_id)
3216 /* inet and inet6 sockets are created in a process
3217 * context so there is always a valid uid/gid
3219 case BPF_FUNC_get_current_uid_gid:
3220 return &bpf_get_current_uid_gid_proto;
3222 return bpf_base_func_proto(func_id);
3226 static const struct bpf_func_proto *
3227 sk_filter_func_proto(enum bpf_func_id func_id)
3230 case BPF_FUNC_skb_load_bytes:
3231 return &bpf_skb_load_bytes_proto;
3232 case BPF_FUNC_get_socket_cookie:
3233 return &bpf_get_socket_cookie_proto;
3234 case BPF_FUNC_get_socket_uid:
3235 return &bpf_get_socket_uid_proto;
3237 return bpf_base_func_proto(func_id);
3241 static const struct bpf_func_proto *
3242 tc_cls_act_func_proto(enum bpf_func_id func_id)
3245 case BPF_FUNC_skb_store_bytes:
3246 return &bpf_skb_store_bytes_proto;
3247 case BPF_FUNC_skb_load_bytes:
3248 return &bpf_skb_load_bytes_proto;
3249 case BPF_FUNC_skb_pull_data:
3250 return &bpf_skb_pull_data_proto;
3251 case BPF_FUNC_csum_diff:
3252 return &bpf_csum_diff_proto;
3253 case BPF_FUNC_csum_update:
3254 return &bpf_csum_update_proto;
3255 case BPF_FUNC_l3_csum_replace:
3256 return &bpf_l3_csum_replace_proto;
3257 case BPF_FUNC_l4_csum_replace:
3258 return &bpf_l4_csum_replace_proto;
3259 case BPF_FUNC_clone_redirect:
3260 return &bpf_clone_redirect_proto;
3261 case BPF_FUNC_get_cgroup_classid:
3262 return &bpf_get_cgroup_classid_proto;
3263 case BPF_FUNC_skb_vlan_push:
3264 return &bpf_skb_vlan_push_proto;
3265 case BPF_FUNC_skb_vlan_pop:
3266 return &bpf_skb_vlan_pop_proto;
3267 case BPF_FUNC_skb_change_proto:
3268 return &bpf_skb_change_proto_proto;
3269 case BPF_FUNC_skb_change_type:
3270 return &bpf_skb_change_type_proto;
3271 case BPF_FUNC_skb_adjust_room:
3272 return &bpf_skb_adjust_room_proto;
3273 case BPF_FUNC_skb_change_tail:
3274 return &bpf_skb_change_tail_proto;
3275 case BPF_FUNC_skb_get_tunnel_key:
3276 return &bpf_skb_get_tunnel_key_proto;
3277 case BPF_FUNC_skb_set_tunnel_key:
3278 return bpf_get_skb_set_tunnel_proto(func_id);
3279 case BPF_FUNC_skb_get_tunnel_opt:
3280 return &bpf_skb_get_tunnel_opt_proto;
3281 case BPF_FUNC_skb_set_tunnel_opt:
3282 return bpf_get_skb_set_tunnel_proto(func_id);
3283 case BPF_FUNC_redirect:
3284 return &bpf_redirect_proto;
3285 case BPF_FUNC_get_route_realm:
3286 return &bpf_get_route_realm_proto;
3287 case BPF_FUNC_get_hash_recalc:
3288 return &bpf_get_hash_recalc_proto;
3289 case BPF_FUNC_set_hash_invalid:
3290 return &bpf_set_hash_invalid_proto;
3291 case BPF_FUNC_set_hash:
3292 return &bpf_set_hash_proto;
3293 case BPF_FUNC_perf_event_output:
3294 return &bpf_skb_event_output_proto;
3295 case BPF_FUNC_get_smp_processor_id:
3296 return &bpf_get_smp_processor_id_proto;
3297 case BPF_FUNC_skb_under_cgroup:
3298 return &bpf_skb_under_cgroup_proto;
3299 case BPF_FUNC_get_socket_cookie:
3300 return &bpf_get_socket_cookie_proto;
3301 case BPF_FUNC_get_socket_uid:
3302 return &bpf_get_socket_uid_proto;
3304 return bpf_base_func_proto(func_id);
3308 static const struct bpf_func_proto *
3309 xdp_func_proto(enum bpf_func_id func_id)
3312 case BPF_FUNC_perf_event_output:
3313 return &bpf_xdp_event_output_proto;
3314 case BPF_FUNC_get_smp_processor_id:
3315 return &bpf_get_smp_processor_id_proto;
3316 case BPF_FUNC_xdp_adjust_head:
3317 return &bpf_xdp_adjust_head_proto;
3318 case BPF_FUNC_redirect:
3319 return &bpf_xdp_redirect_proto;
3320 case BPF_FUNC_redirect_map:
3321 return &bpf_xdp_redirect_map_proto;
3323 return bpf_base_func_proto(func_id);
3327 static const struct bpf_func_proto *
3328 lwt_inout_func_proto(enum bpf_func_id func_id)
3331 case BPF_FUNC_skb_load_bytes:
3332 return &bpf_skb_load_bytes_proto;
3333 case BPF_FUNC_skb_pull_data:
3334 return &bpf_skb_pull_data_proto;
3335 case BPF_FUNC_csum_diff:
3336 return &bpf_csum_diff_proto;
3337 case BPF_FUNC_get_cgroup_classid:
3338 return &bpf_get_cgroup_classid_proto;
3339 case BPF_FUNC_get_route_realm:
3340 return &bpf_get_route_realm_proto;
3341 case BPF_FUNC_get_hash_recalc:
3342 return &bpf_get_hash_recalc_proto;
3343 case BPF_FUNC_perf_event_output:
3344 return &bpf_skb_event_output_proto;
3345 case BPF_FUNC_get_smp_processor_id:
3346 return &bpf_get_smp_processor_id_proto;
3347 case BPF_FUNC_skb_under_cgroup:
3348 return &bpf_skb_under_cgroup_proto;
3350 return bpf_base_func_proto(func_id);
3354 static const struct bpf_func_proto *
3355 sock_ops_func_proto(enum bpf_func_id func_id)
3358 case BPF_FUNC_setsockopt:
3359 return &bpf_setsockopt_proto;
3360 case BPF_FUNC_sock_map_update:
3361 return &bpf_sock_map_update_proto;
3363 return bpf_base_func_proto(func_id);
3367 static const struct bpf_func_proto *sk_skb_func_proto(enum bpf_func_id func_id)
3370 case BPF_FUNC_skb_store_bytes:
3371 return &bpf_skb_store_bytes_proto;
3372 case BPF_FUNC_skb_load_bytes:
3373 return &bpf_skb_load_bytes_proto;
3374 case BPF_FUNC_skb_pull_data:
3375 return &bpf_skb_pull_data_proto;
3376 case BPF_FUNC_skb_change_tail:
3377 return &bpf_skb_change_tail_proto;
3378 case BPF_FUNC_skb_change_head:
3379 return &bpf_skb_change_head_proto;
3380 case BPF_FUNC_get_socket_cookie:
3381 return &bpf_get_socket_cookie_proto;
3382 case BPF_FUNC_get_socket_uid:
3383 return &bpf_get_socket_uid_proto;
3384 case BPF_FUNC_sk_redirect_map:
3385 return &bpf_sk_redirect_map_proto;
3387 return bpf_base_func_proto(func_id);
3391 static const struct bpf_func_proto *
3392 lwt_xmit_func_proto(enum bpf_func_id func_id)
3395 case BPF_FUNC_skb_get_tunnel_key:
3396 return &bpf_skb_get_tunnel_key_proto;
3397 case BPF_FUNC_skb_set_tunnel_key:
3398 return bpf_get_skb_set_tunnel_proto(func_id);
3399 case BPF_FUNC_skb_get_tunnel_opt:
3400 return &bpf_skb_get_tunnel_opt_proto;
3401 case BPF_FUNC_skb_set_tunnel_opt:
3402 return bpf_get_skb_set_tunnel_proto(func_id);
3403 case BPF_FUNC_redirect:
3404 return &bpf_redirect_proto;
3405 case BPF_FUNC_clone_redirect:
3406 return &bpf_clone_redirect_proto;
3407 case BPF_FUNC_skb_change_tail:
3408 return &bpf_skb_change_tail_proto;
3409 case BPF_FUNC_skb_change_head:
3410 return &bpf_skb_change_head_proto;
3411 case BPF_FUNC_skb_store_bytes:
3412 return &bpf_skb_store_bytes_proto;
3413 case BPF_FUNC_csum_update:
3414 return &bpf_csum_update_proto;
3415 case BPF_FUNC_l3_csum_replace:
3416 return &bpf_l3_csum_replace_proto;
3417 case BPF_FUNC_l4_csum_replace:
3418 return &bpf_l4_csum_replace_proto;
3419 case BPF_FUNC_set_hash_invalid:
3420 return &bpf_set_hash_invalid_proto;
3422 return lwt_inout_func_proto(func_id);
3426 static bool bpf_skb_is_valid_access(int off, int size, enum bpf_access_type type,
3427 struct bpf_insn_access_aux *info)
3429 const int size_default = sizeof(__u32);
3431 if (off < 0 || off >= sizeof(struct __sk_buff))
3434 /* The verifier guarantees that size > 0. */
3435 if (off % size != 0)
3439 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3440 if (off + size > offsetofend(struct __sk_buff, cb[4]))
3443 case bpf_ctx_range_till(struct __sk_buff, remote_ip6[0], remote_ip6[3]):
3444 case bpf_ctx_range_till(struct __sk_buff, local_ip6[0], local_ip6[3]):
3445 case bpf_ctx_range_till(struct __sk_buff, remote_ip4, remote_ip4):
3446 case bpf_ctx_range_till(struct __sk_buff, local_ip4, local_ip4):
3447 case bpf_ctx_range(struct __sk_buff, data):
3448 case bpf_ctx_range(struct __sk_buff, data_end):
3449 if (size != size_default)
3453 /* Only narrow read access allowed for now. */
3454 if (type == BPF_WRITE) {
3455 if (size != size_default)
3458 bpf_ctx_record_field_size(info, size_default);
3459 if (!bpf_ctx_narrow_access_ok(off, size, size_default))
3467 static bool sk_filter_is_valid_access(int off, int size,
3468 enum bpf_access_type type,
3469 struct bpf_insn_access_aux *info)
3472 case bpf_ctx_range(struct __sk_buff, tc_classid):
3473 case bpf_ctx_range(struct __sk_buff, data):
3474 case bpf_ctx_range(struct __sk_buff, data_end):
3475 case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3479 if (type == BPF_WRITE) {
3481 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3488 return bpf_skb_is_valid_access(off, size, type, info);
3491 static bool lwt_is_valid_access(int off, int size,
3492 enum bpf_access_type type,
3493 struct bpf_insn_access_aux *info)
3496 case bpf_ctx_range(struct __sk_buff, tc_classid):
3497 case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3501 if (type == BPF_WRITE) {
3503 case bpf_ctx_range(struct __sk_buff, mark):
3504 case bpf_ctx_range(struct __sk_buff, priority):
3505 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3513 case bpf_ctx_range(struct __sk_buff, data):
3514 info->reg_type = PTR_TO_PACKET;
3516 case bpf_ctx_range(struct __sk_buff, data_end):
3517 info->reg_type = PTR_TO_PACKET_END;
3521 return bpf_skb_is_valid_access(off, size, type, info);
3524 static bool sock_filter_is_valid_access(int off, int size,
3525 enum bpf_access_type type,
3526 struct bpf_insn_access_aux *info)
3528 if (type == BPF_WRITE) {
3530 case offsetof(struct bpf_sock, bound_dev_if):
3531 case offsetof(struct bpf_sock, mark):
3532 case offsetof(struct bpf_sock, priority):
3539 if (off < 0 || off + size > sizeof(struct bpf_sock))
3541 /* The verifier guarantees that size > 0. */
3542 if (off % size != 0)
3544 if (size != sizeof(__u32))
3550 static int bpf_unclone_prologue(struct bpf_insn *insn_buf, bool direct_write,
3551 const struct bpf_prog *prog, int drop_verdict)
3553 struct bpf_insn *insn = insn_buf;
3558 /* if (!skb->cloned)
3561 * (Fast-path, otherwise approximation that we might be
3562 * a clone, do the rest in helper.)
3564 *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_6, BPF_REG_1, CLONED_OFFSET());
3565 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_6, CLONED_MASK);
3566 *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_6, 0, 7);
3568 /* ret = bpf_skb_pull_data(skb, 0); */
3569 *insn++ = BPF_MOV64_REG(BPF_REG_6, BPF_REG_1);
3570 *insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_2, BPF_REG_2);
3571 *insn++ = BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
3572 BPF_FUNC_skb_pull_data);
3575 * return TC_ACT_SHOT;
3577 *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2);
3578 *insn++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_0, drop_verdict);
3579 *insn++ = BPF_EXIT_INSN();
3582 *insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_6);
3584 *insn++ = prog->insnsi[0];
3586 return insn - insn_buf;
3589 static int tc_cls_act_prologue(struct bpf_insn *insn_buf, bool direct_write,
3590 const struct bpf_prog *prog)
3592 return bpf_unclone_prologue(insn_buf, direct_write, prog, TC_ACT_SHOT);
3595 static bool tc_cls_act_is_valid_access(int off, int size,
3596 enum bpf_access_type type,
3597 struct bpf_insn_access_aux *info)
3599 if (type == BPF_WRITE) {
3601 case bpf_ctx_range(struct __sk_buff, mark):
3602 case bpf_ctx_range(struct __sk_buff, tc_index):
3603 case bpf_ctx_range(struct __sk_buff, priority):
3604 case bpf_ctx_range(struct __sk_buff, tc_classid):
3605 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3613 case bpf_ctx_range(struct __sk_buff, data):
3614 info->reg_type = PTR_TO_PACKET;
3616 case bpf_ctx_range(struct __sk_buff, data_end):
3617 info->reg_type = PTR_TO_PACKET_END;
3619 case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3623 return bpf_skb_is_valid_access(off, size, type, info);
3626 static bool __is_valid_xdp_access(int off, int size)
3628 if (off < 0 || off >= sizeof(struct xdp_md))
3630 if (off % size != 0)
3632 if (size != sizeof(__u32))
3638 static bool xdp_is_valid_access(int off, int size,
3639 enum bpf_access_type type,
3640 struct bpf_insn_access_aux *info)
3642 if (type == BPF_WRITE)
3646 case offsetof(struct xdp_md, data):
3647 info->reg_type = PTR_TO_PACKET;
3649 case offsetof(struct xdp_md, data_end):
3650 info->reg_type = PTR_TO_PACKET_END;
3654 return __is_valid_xdp_access(off, size);
3657 void bpf_warn_invalid_xdp_action(u32 act)
3659 const u32 act_max = XDP_REDIRECT;
3661 pr_warn_once("%s XDP return value %u, expect packet loss!\n",
3662 act > act_max ? "Illegal" : "Driver unsupported",
3665 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
3667 static bool __is_valid_sock_ops_access(int off, int size)
3669 if (off < 0 || off >= sizeof(struct bpf_sock_ops))
3671 /* The verifier guarantees that size > 0. */
3672 if (off % size != 0)
3674 if (size != sizeof(__u32))
3680 static bool sock_ops_is_valid_access(int off, int size,
3681 enum bpf_access_type type,
3682 struct bpf_insn_access_aux *info)
3684 if (type == BPF_WRITE) {
3686 case offsetof(struct bpf_sock_ops, op) ...
3687 offsetof(struct bpf_sock_ops, replylong[3]):
3694 return __is_valid_sock_ops_access(off, size);
3697 static int sk_skb_prologue(struct bpf_insn *insn_buf, bool direct_write,
3698 const struct bpf_prog *prog)
3700 return bpf_unclone_prologue(insn_buf, direct_write, prog, SK_DROP);
3703 static bool sk_skb_is_valid_access(int off, int size,
3704 enum bpf_access_type type,
3705 struct bpf_insn_access_aux *info)
3707 if (type == BPF_WRITE) {
3709 case bpf_ctx_range(struct __sk_buff, tc_index):
3710 case bpf_ctx_range(struct __sk_buff, priority):
3718 case bpf_ctx_range(struct __sk_buff, mark):
3719 case bpf_ctx_range(struct __sk_buff, tc_classid):
3721 case bpf_ctx_range(struct __sk_buff, data):
3722 info->reg_type = PTR_TO_PACKET;
3724 case bpf_ctx_range(struct __sk_buff, data_end):
3725 info->reg_type = PTR_TO_PACKET_END;
3729 return bpf_skb_is_valid_access(off, size, type, info);
3732 static u32 bpf_convert_ctx_access(enum bpf_access_type type,
3733 const struct bpf_insn *si,
3734 struct bpf_insn *insn_buf,
3735 struct bpf_prog *prog, u32 *target_size)
3737 struct bpf_insn *insn = insn_buf;
3741 case offsetof(struct __sk_buff, len):
3742 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3743 bpf_target_off(struct sk_buff, len, 4,
3747 case offsetof(struct __sk_buff, protocol):
3748 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3749 bpf_target_off(struct sk_buff, protocol, 2,
3753 case offsetof(struct __sk_buff, vlan_proto):
3754 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3755 bpf_target_off(struct sk_buff, vlan_proto, 2,
3759 case offsetof(struct __sk_buff, priority):
3760 if (type == BPF_WRITE)
3761 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3762 bpf_target_off(struct sk_buff, priority, 4,
3765 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3766 bpf_target_off(struct sk_buff, priority, 4,
3770 case offsetof(struct __sk_buff, ingress_ifindex):
3771 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3772 bpf_target_off(struct sk_buff, skb_iif, 4,
3776 case offsetof(struct __sk_buff, ifindex):
3777 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
3778 si->dst_reg, si->src_reg,
3779 offsetof(struct sk_buff, dev));
3780 *insn++ = BPF_JMP_IMM(BPF_JEQ, si->dst_reg, 0, 1);
3781 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3782 bpf_target_off(struct net_device, ifindex, 4,
3786 case offsetof(struct __sk_buff, hash):
3787 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3788 bpf_target_off(struct sk_buff, hash, 4,
3792 case offsetof(struct __sk_buff, mark):
3793 if (type == BPF_WRITE)
3794 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3795 bpf_target_off(struct sk_buff, mark, 4,
3798 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3799 bpf_target_off(struct sk_buff, mark, 4,
3803 case offsetof(struct __sk_buff, pkt_type):
3805 *insn++ = BPF_LDX_MEM(BPF_B, si->dst_reg, si->src_reg,
3807 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, PKT_TYPE_MAX);
3808 #ifdef __BIG_ENDIAN_BITFIELD
3809 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 5);
3813 case offsetof(struct __sk_buff, queue_mapping):
3814 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3815 bpf_target_off(struct sk_buff, queue_mapping, 2,
3819 case offsetof(struct __sk_buff, vlan_present):
3820 case offsetof(struct __sk_buff, vlan_tci):
3821 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
3823 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3824 bpf_target_off(struct sk_buff, vlan_tci, 2,
3826 if (si->off == offsetof(struct __sk_buff, vlan_tci)) {
3827 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg,
3830 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 12);
3831 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, 1);
3835 case offsetof(struct __sk_buff, cb[0]) ...
3836 offsetofend(struct __sk_buff, cb[4]) - 1:
3837 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
3838 BUILD_BUG_ON((offsetof(struct sk_buff, cb) +
3839 offsetof(struct qdisc_skb_cb, data)) %
3842 prog->cb_access = 1;
3844 off -= offsetof(struct __sk_buff, cb[0]);
3845 off += offsetof(struct sk_buff, cb);
3846 off += offsetof(struct qdisc_skb_cb, data);
3847 if (type == BPF_WRITE)
3848 *insn++ = BPF_STX_MEM(BPF_SIZE(si->code), si->dst_reg,
3851 *insn++ = BPF_LDX_MEM(BPF_SIZE(si->code), si->dst_reg,
3855 case offsetof(struct __sk_buff, tc_classid):
3856 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, tc_classid) != 2);
3859 off -= offsetof(struct __sk_buff, tc_classid);
3860 off += offsetof(struct sk_buff, cb);
3861 off += offsetof(struct qdisc_skb_cb, tc_classid);
3863 if (type == BPF_WRITE)
3864 *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg,
3867 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg,
3871 case offsetof(struct __sk_buff, data):
3872 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, data),
3873 si->dst_reg, si->src_reg,
3874 offsetof(struct sk_buff, data));
3877 case offsetof(struct __sk_buff, data_end):
3879 off -= offsetof(struct __sk_buff, data_end);
3880 off += offsetof(struct sk_buff, cb);
3881 off += offsetof(struct bpf_skb_data_end, data_end);
3882 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
3886 case offsetof(struct __sk_buff, tc_index):
3887 #ifdef CONFIG_NET_SCHED
3888 if (type == BPF_WRITE)
3889 *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg, si->src_reg,
3890 bpf_target_off(struct sk_buff, tc_index, 2,
3893 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3894 bpf_target_off(struct sk_buff, tc_index, 2,
3898 if (type == BPF_WRITE)
3899 *insn++ = BPF_MOV64_REG(si->dst_reg, si->dst_reg);
3901 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3905 case offsetof(struct __sk_buff, napi_id):
3906 #if defined(CONFIG_NET_RX_BUSY_POLL)
3907 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3908 bpf_target_off(struct sk_buff, napi_id, 4,
3910 *insn++ = BPF_JMP_IMM(BPF_JGE, si->dst_reg, MIN_NAPI_ID, 1);
3911 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3914 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3917 case offsetof(struct __sk_buff, family):
3918 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
3920 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
3921 si->dst_reg, si->src_reg,
3922 offsetof(struct sk_buff, sk));
3923 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
3924 bpf_target_off(struct sock_common,
3928 case offsetof(struct __sk_buff, remote_ip4):
3929 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
3931 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
3932 si->dst_reg, si->src_reg,
3933 offsetof(struct sk_buff, sk));
3934 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3935 bpf_target_off(struct sock_common,
3939 case offsetof(struct __sk_buff, local_ip4):
3940 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
3941 skc_rcv_saddr) != 4);
3943 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
3944 si->dst_reg, si->src_reg,
3945 offsetof(struct sk_buff, sk));
3946 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3947 bpf_target_off(struct sock_common,
3951 case offsetof(struct __sk_buff, remote_ip6[0]) ...
3952 offsetof(struct __sk_buff, remote_ip6[3]):
3953 #if IS_ENABLED(CONFIG_IPV6)
3954 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
3955 skc_v6_daddr.s6_addr32[0]) != 4);
3958 off -= offsetof(struct __sk_buff, remote_ip6[0]);
3960 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
3961 si->dst_reg, si->src_reg,
3962 offsetof(struct sk_buff, sk));
3963 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3964 offsetof(struct sock_common,
3965 skc_v6_daddr.s6_addr32[0]) +
3968 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
3971 case offsetof(struct __sk_buff, local_ip6[0]) ...
3972 offsetof(struct __sk_buff, local_ip6[3]):
3973 #if IS_ENABLED(CONFIG_IPV6)
3974 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
3975 skc_v6_rcv_saddr.s6_addr32[0]) != 4);
3978 off -= offsetof(struct __sk_buff, local_ip6[0]);
3980 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
3981 si->dst_reg, si->src_reg,
3982 offsetof(struct sk_buff, sk));
3983 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3984 offsetof(struct sock_common,
3985 skc_v6_rcv_saddr.s6_addr32[0]) +
3988 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
3992 case offsetof(struct __sk_buff, remote_port):
3993 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
3995 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
3996 si->dst_reg, si->src_reg,
3997 offsetof(struct sk_buff, sk));
3998 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
3999 bpf_target_off(struct sock_common,
4002 #ifndef __BIG_ENDIAN_BITFIELD
4003 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
4007 case offsetof(struct __sk_buff, local_port):
4008 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
4010 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4011 si->dst_reg, si->src_reg,
4012 offsetof(struct sk_buff, sk));
4013 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4014 bpf_target_off(struct sock_common,
4015 skc_num, 2, target_size));
4019 return insn - insn_buf;
4022 static u32 sock_filter_convert_ctx_access(enum bpf_access_type type,
4023 const struct bpf_insn *si,
4024 struct bpf_insn *insn_buf,
4025 struct bpf_prog *prog, u32 *target_size)
4027 struct bpf_insn *insn = insn_buf;
4030 case offsetof(struct bpf_sock, bound_dev_if):
4031 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_bound_dev_if) != 4);
4033 if (type == BPF_WRITE)
4034 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4035 offsetof(struct sock, sk_bound_dev_if));
4037 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4038 offsetof(struct sock, sk_bound_dev_if));
4041 case offsetof(struct bpf_sock, mark):
4042 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_mark) != 4);
4044 if (type == BPF_WRITE)
4045 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4046 offsetof(struct sock, sk_mark));
4048 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4049 offsetof(struct sock, sk_mark));
4052 case offsetof(struct bpf_sock, priority):
4053 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_priority) != 4);
4055 if (type == BPF_WRITE)
4056 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4057 offsetof(struct sock, sk_priority));
4059 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4060 offsetof(struct sock, sk_priority));
4063 case offsetof(struct bpf_sock, family):
4064 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_family) != 2);
4066 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4067 offsetof(struct sock, sk_family));
4070 case offsetof(struct bpf_sock, type):
4071 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4072 offsetof(struct sock, __sk_flags_offset));
4073 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_TYPE_MASK);
4074 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_TYPE_SHIFT);
4077 case offsetof(struct bpf_sock, protocol):
4078 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4079 offsetof(struct sock, __sk_flags_offset));
4080 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_PROTO_MASK);
4081 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_PROTO_SHIFT);
4085 return insn - insn_buf;
4088 static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type,
4089 const struct bpf_insn *si,
4090 struct bpf_insn *insn_buf,
4091 struct bpf_prog *prog, u32 *target_size)
4093 struct bpf_insn *insn = insn_buf;
4096 case offsetof(struct __sk_buff, ifindex):
4097 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
4098 si->dst_reg, si->src_reg,
4099 offsetof(struct sk_buff, dev));
4100 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4101 bpf_target_off(struct net_device, ifindex, 4,
4105 return bpf_convert_ctx_access(type, si, insn_buf, prog,
4109 return insn - insn_buf;
4112 static u32 xdp_convert_ctx_access(enum bpf_access_type type,
4113 const struct bpf_insn *si,
4114 struct bpf_insn *insn_buf,
4115 struct bpf_prog *prog, u32 *target_size)
4117 struct bpf_insn *insn = insn_buf;
4120 case offsetof(struct xdp_md, data):
4121 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data),
4122 si->dst_reg, si->src_reg,
4123 offsetof(struct xdp_buff, data));
4125 case offsetof(struct xdp_md, data_end):
4126 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_end),
4127 si->dst_reg, si->src_reg,
4128 offsetof(struct xdp_buff, data_end));
4132 return insn - insn_buf;
4135 static u32 sock_ops_convert_ctx_access(enum bpf_access_type type,
4136 const struct bpf_insn *si,
4137 struct bpf_insn *insn_buf,
4138 struct bpf_prog *prog,
4141 struct bpf_insn *insn = insn_buf;
4145 case offsetof(struct bpf_sock_ops, op) ...
4146 offsetof(struct bpf_sock_ops, replylong[3]):
4147 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, op) !=
4148 FIELD_SIZEOF(struct bpf_sock_ops_kern, op));
4149 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, reply) !=
4150 FIELD_SIZEOF(struct bpf_sock_ops_kern, reply));
4151 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, replylong) !=
4152 FIELD_SIZEOF(struct bpf_sock_ops_kern, replylong));
4154 off -= offsetof(struct bpf_sock_ops, op);
4155 off += offsetof(struct bpf_sock_ops_kern, op);
4156 if (type == BPF_WRITE)
4157 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4160 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4164 case offsetof(struct bpf_sock_ops, family):
4165 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
4167 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4168 struct bpf_sock_ops_kern, sk),
4169 si->dst_reg, si->src_reg,
4170 offsetof(struct bpf_sock_ops_kern, sk));
4171 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4172 offsetof(struct sock_common, skc_family));
4175 case offsetof(struct bpf_sock_ops, remote_ip4):
4176 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
4178 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4179 struct bpf_sock_ops_kern, sk),
4180 si->dst_reg, si->src_reg,
4181 offsetof(struct bpf_sock_ops_kern, sk));
4182 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4183 offsetof(struct sock_common, skc_daddr));
4186 case offsetof(struct bpf_sock_ops, local_ip4):
4187 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_rcv_saddr) != 4);
4189 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4190 struct bpf_sock_ops_kern, sk),
4191 si->dst_reg, si->src_reg,
4192 offsetof(struct bpf_sock_ops_kern, sk));
4193 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4194 offsetof(struct sock_common,
4198 case offsetof(struct bpf_sock_ops, remote_ip6[0]) ...
4199 offsetof(struct bpf_sock_ops, remote_ip6[3]):
4200 #if IS_ENABLED(CONFIG_IPV6)
4201 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4202 skc_v6_daddr.s6_addr32[0]) != 4);
4205 off -= offsetof(struct bpf_sock_ops, remote_ip6[0]);
4206 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4207 struct bpf_sock_ops_kern, sk),
4208 si->dst_reg, si->src_reg,
4209 offsetof(struct bpf_sock_ops_kern, sk));
4210 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4211 offsetof(struct sock_common,
4212 skc_v6_daddr.s6_addr32[0]) +
4215 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4219 case offsetof(struct bpf_sock_ops, local_ip6[0]) ...
4220 offsetof(struct bpf_sock_ops, local_ip6[3]):
4221 #if IS_ENABLED(CONFIG_IPV6)
4222 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4223 skc_v6_rcv_saddr.s6_addr32[0]) != 4);
4226 off -= offsetof(struct bpf_sock_ops, local_ip6[0]);
4227 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4228 struct bpf_sock_ops_kern, sk),
4229 si->dst_reg, si->src_reg,
4230 offsetof(struct bpf_sock_ops_kern, sk));
4231 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4232 offsetof(struct sock_common,
4233 skc_v6_rcv_saddr.s6_addr32[0]) +
4236 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4240 case offsetof(struct bpf_sock_ops, remote_port):
4241 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
4243 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4244 struct bpf_sock_ops_kern, sk),
4245 si->dst_reg, si->src_reg,
4246 offsetof(struct bpf_sock_ops_kern, sk));
4247 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4248 offsetof(struct sock_common, skc_dport));
4249 #ifndef __BIG_ENDIAN_BITFIELD
4250 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
4254 case offsetof(struct bpf_sock_ops, local_port):
4255 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
4257 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4258 struct bpf_sock_ops_kern, sk),
4259 si->dst_reg, si->src_reg,
4260 offsetof(struct bpf_sock_ops_kern, sk));
4261 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4262 offsetof(struct sock_common, skc_num));
4265 return insn - insn_buf;
4268 static u32 sk_skb_convert_ctx_access(enum bpf_access_type type,
4269 const struct bpf_insn *si,
4270 struct bpf_insn *insn_buf,
4271 struct bpf_prog *prog, u32 *target_size)
4273 struct bpf_insn *insn = insn_buf;
4277 case offsetof(struct __sk_buff, data_end):
4279 off -= offsetof(struct __sk_buff, data_end);
4280 off += offsetof(struct sk_buff, cb);
4281 off += offsetof(struct tcp_skb_cb, bpf.data_end);
4282 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4286 return bpf_convert_ctx_access(type, si, insn_buf, prog,
4290 return insn - insn_buf;
4293 const struct bpf_verifier_ops sk_filter_prog_ops = {
4294 .get_func_proto = sk_filter_func_proto,
4295 .is_valid_access = sk_filter_is_valid_access,
4296 .convert_ctx_access = bpf_convert_ctx_access,
4299 const struct bpf_verifier_ops tc_cls_act_prog_ops = {
4300 .get_func_proto = tc_cls_act_func_proto,
4301 .is_valid_access = tc_cls_act_is_valid_access,
4302 .convert_ctx_access = tc_cls_act_convert_ctx_access,
4303 .gen_prologue = tc_cls_act_prologue,
4304 .test_run = bpf_prog_test_run_skb,
4307 const struct bpf_verifier_ops xdp_prog_ops = {
4308 .get_func_proto = xdp_func_proto,
4309 .is_valid_access = xdp_is_valid_access,
4310 .convert_ctx_access = xdp_convert_ctx_access,
4311 .test_run = bpf_prog_test_run_xdp,
4314 const struct bpf_verifier_ops cg_skb_prog_ops = {
4315 .get_func_proto = sk_filter_func_proto,
4316 .is_valid_access = sk_filter_is_valid_access,
4317 .convert_ctx_access = bpf_convert_ctx_access,
4318 .test_run = bpf_prog_test_run_skb,
4321 const struct bpf_verifier_ops lwt_inout_prog_ops = {
4322 .get_func_proto = lwt_inout_func_proto,
4323 .is_valid_access = lwt_is_valid_access,
4324 .convert_ctx_access = bpf_convert_ctx_access,
4325 .test_run = bpf_prog_test_run_skb,
4328 const struct bpf_verifier_ops lwt_xmit_prog_ops = {
4329 .get_func_proto = lwt_xmit_func_proto,
4330 .is_valid_access = lwt_is_valid_access,
4331 .convert_ctx_access = bpf_convert_ctx_access,
4332 .gen_prologue = tc_cls_act_prologue,
4333 .test_run = bpf_prog_test_run_skb,
4336 const struct bpf_verifier_ops cg_sock_prog_ops = {
4337 .get_func_proto = sock_filter_func_proto,
4338 .is_valid_access = sock_filter_is_valid_access,
4339 .convert_ctx_access = sock_filter_convert_ctx_access,
4342 const struct bpf_verifier_ops sock_ops_prog_ops = {
4343 .get_func_proto = sock_ops_func_proto,
4344 .is_valid_access = sock_ops_is_valid_access,
4345 .convert_ctx_access = sock_ops_convert_ctx_access,
4348 const struct bpf_verifier_ops sk_skb_prog_ops = {
4349 .get_func_proto = sk_skb_func_proto,
4350 .is_valid_access = sk_skb_is_valid_access,
4351 .convert_ctx_access = sk_skb_convert_ctx_access,
4352 .gen_prologue = sk_skb_prologue,
4355 int sk_detach_filter(struct sock *sk)
4358 struct sk_filter *filter;
4360 if (sock_flag(sk, SOCK_FILTER_LOCKED))
4363 filter = rcu_dereference_protected(sk->sk_filter,
4364 lockdep_sock_is_held(sk));
4366 RCU_INIT_POINTER(sk->sk_filter, NULL);
4367 sk_filter_uncharge(sk, filter);
4373 EXPORT_SYMBOL_GPL(sk_detach_filter);
4375 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
4378 struct sock_fprog_kern *fprog;
4379 struct sk_filter *filter;
4383 filter = rcu_dereference_protected(sk->sk_filter,
4384 lockdep_sock_is_held(sk));
4388 /* We're copying the filter that has been originally attached,
4389 * so no conversion/decode needed anymore. eBPF programs that
4390 * have no original program cannot be dumped through this.
4393 fprog = filter->prog->orig_prog;
4399 /* User space only enquires number of filter blocks. */
4403 if (len < fprog->len)
4407 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
4410 /* Instead of bytes, the API requests to return the number