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>
30 #include <linux/inet.h>
31 #include <linux/netdevice.h>
32 #include <linux/if_packet.h>
33 #include <linux/gfp.h>
35 #include <net/protocol.h>
36 #include <net/netlink.h>
37 #include <linux/skbuff.h>
39 #include <net/flow_dissector.h>
40 #include <linux/errno.h>
41 #include <linux/timer.h>
42 #include <asm/uaccess.h>
43 #include <asm/unaligned.h>
44 #include <linux/filter.h>
45 #include <linux/ratelimit.h>
46 #include <linux/seccomp.h>
47 #include <linux/if_vlan.h>
48 #include <linux/bpf.h>
49 #include <net/sch_generic.h>
52 * sk_filter - run a packet through a socket filter
53 * @sk: sock associated with &sk_buff
54 * @skb: buffer to filter
56 * Run the filter code and then cut skb->data to correct size returned by
57 * SK_RUN_FILTER. If pkt_len is 0 we toss packet. If skb->len is smaller
58 * than pkt_len we keep whole skb->data. This is the socket level
59 * wrapper to SK_RUN_FILTER. It returns 0 if the packet should
60 * be accepted or -EPERM if the packet should be tossed.
63 int sk_filter(struct sock *sk, struct sk_buff *skb)
66 struct sk_filter *filter;
69 * If the skb was allocated from pfmemalloc reserves, only
70 * allow SOCK_MEMALLOC sockets to use it as this socket is
73 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC))
76 err = security_sock_rcv_skb(sk, skb);
81 filter = rcu_dereference(sk->sk_filter);
83 unsigned int pkt_len = SK_RUN_FILTER(filter, skb);
85 err = pkt_len ? pskb_trim(skb, pkt_len) : -EPERM;
91 EXPORT_SYMBOL(sk_filter);
93 static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
95 return skb_get_poff((struct sk_buff *)(unsigned long) ctx);
98 static u64 __skb_get_nlattr(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
100 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
103 if (skb_is_nonlinear(skb))
106 if (skb->len < sizeof(struct nlattr))
109 if (a > skb->len - sizeof(struct nlattr))
112 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
114 return (void *) nla - (void *) skb->data;
119 static u64 __skb_get_nlattr_nest(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
121 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
124 if (skb_is_nonlinear(skb))
127 if (skb->len < sizeof(struct nlattr))
130 if (a > skb->len - sizeof(struct nlattr))
133 nla = (struct nlattr *) &skb->data[a];
134 if (nla->nla_len > skb->len - a)
137 nla = nla_find_nested(nla, x);
139 return (void *) nla - (void *) skb->data;
144 static u64 __get_raw_cpu_id(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
146 return raw_smp_processor_id();
149 /* note that this only generates 32-bit random numbers */
150 static u64 __get_random_u32(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
152 return prandom_u32();
155 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
156 struct bpf_insn *insn_buf)
158 struct bpf_insn *insn = insn_buf;
162 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
164 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
165 offsetof(struct sk_buff, mark));
169 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
170 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
171 #ifdef __BIG_ENDIAN_BITFIELD
172 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
177 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
179 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
180 offsetof(struct sk_buff, queue_mapping));
183 case SKF_AD_VLAN_TAG:
184 case SKF_AD_VLAN_TAG_PRESENT:
185 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
186 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
188 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
189 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
190 offsetof(struct sk_buff, vlan_tci));
191 if (skb_field == SKF_AD_VLAN_TAG) {
192 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
196 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
198 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
203 return insn - insn_buf;
206 static bool convert_bpf_extensions(struct sock_filter *fp,
207 struct bpf_insn **insnp)
209 struct bpf_insn *insn = *insnp;
213 case SKF_AD_OFF + SKF_AD_PROTOCOL:
214 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
216 /* A = *(u16 *) (CTX + offsetof(protocol)) */
217 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
218 offsetof(struct sk_buff, protocol));
219 /* A = ntohs(A) [emitting a nop or swap16] */
220 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
223 case SKF_AD_OFF + SKF_AD_PKTTYPE:
224 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
228 case SKF_AD_OFF + SKF_AD_IFINDEX:
229 case SKF_AD_OFF + SKF_AD_HATYPE:
230 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
231 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
232 BUILD_BUG_ON(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)) < 0);
234 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
235 BPF_REG_TMP, BPF_REG_CTX,
236 offsetof(struct sk_buff, dev));
237 /* if (tmp != 0) goto pc + 1 */
238 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
239 *insn++ = BPF_EXIT_INSN();
240 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
241 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
242 offsetof(struct net_device, ifindex));
244 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
245 offsetof(struct net_device, type));
248 case SKF_AD_OFF + SKF_AD_MARK:
249 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
253 case SKF_AD_OFF + SKF_AD_RXHASH:
254 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
256 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
257 offsetof(struct sk_buff, hash));
260 case SKF_AD_OFF + SKF_AD_QUEUE:
261 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
265 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
266 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
267 BPF_REG_A, BPF_REG_CTX, insn);
271 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
272 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
273 BPF_REG_A, BPF_REG_CTX, insn);
277 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
278 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
280 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
281 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
282 offsetof(struct sk_buff, vlan_proto));
283 /* A = ntohs(A) [emitting a nop or swap16] */
284 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
287 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
288 case SKF_AD_OFF + SKF_AD_NLATTR:
289 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
290 case SKF_AD_OFF + SKF_AD_CPU:
291 case SKF_AD_OFF + SKF_AD_RANDOM:
293 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
295 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
297 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
298 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
300 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
301 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
303 case SKF_AD_OFF + SKF_AD_NLATTR:
304 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
306 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
307 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
309 case SKF_AD_OFF + SKF_AD_CPU:
310 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
312 case SKF_AD_OFF + SKF_AD_RANDOM:
313 *insn = BPF_EMIT_CALL(__get_random_u32);
318 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
320 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
324 /* This is just a dummy call to avoid letting the compiler
325 * evict __bpf_call_base() as an optimization. Placed here
326 * where no-one bothers.
328 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
337 * bpf_convert_filter - convert filter program
338 * @prog: the user passed filter program
339 * @len: the length of the user passed filter program
340 * @new_prog: buffer where converted program will be stored
341 * @new_len: pointer to store length of converted program
343 * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style.
344 * Conversion workflow:
346 * 1) First pass for calculating the new program length:
347 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
349 * 2) 2nd pass to remap in two passes: 1st pass finds new
350 * jump offsets, 2nd pass remapping:
351 * new_prog = kmalloc(sizeof(struct bpf_insn) * new_len);
352 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
354 * User BPF's register A is mapped to our BPF register 6, user BPF
355 * register X is mapped to BPF register 7; frame pointer is always
356 * register 10; Context 'void *ctx' is stored in register 1, that is,
357 * for socket filters: ctx == 'struct sk_buff *', for seccomp:
358 * ctx == 'struct seccomp_data *'.
360 static int bpf_convert_filter(struct sock_filter *prog, int len,
361 struct bpf_insn *new_prog, int *new_len)
363 int new_flen = 0, pass = 0, target, i;
364 struct bpf_insn *new_insn;
365 struct sock_filter *fp;
369 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
370 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
372 if (len <= 0 || len > BPF_MAXINSNS)
376 addrs = kcalloc(len, sizeof(*addrs),
377 GFP_KERNEL | __GFP_NOWARN);
387 *new_insn = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
390 for (i = 0; i < len; fp++, i++) {
391 struct bpf_insn tmp_insns[6] = { };
392 struct bpf_insn *insn = tmp_insns;
395 addrs[i] = new_insn - new_prog;
398 /* All arithmetic insns and skb loads map as-is. */
399 case BPF_ALU | BPF_ADD | BPF_X:
400 case BPF_ALU | BPF_ADD | BPF_K:
401 case BPF_ALU | BPF_SUB | BPF_X:
402 case BPF_ALU | BPF_SUB | BPF_K:
403 case BPF_ALU | BPF_AND | BPF_X:
404 case BPF_ALU | BPF_AND | BPF_K:
405 case BPF_ALU | BPF_OR | BPF_X:
406 case BPF_ALU | BPF_OR | BPF_K:
407 case BPF_ALU | BPF_LSH | BPF_X:
408 case BPF_ALU | BPF_LSH | BPF_K:
409 case BPF_ALU | BPF_RSH | BPF_X:
410 case BPF_ALU | BPF_RSH | BPF_K:
411 case BPF_ALU | BPF_XOR | BPF_X:
412 case BPF_ALU | BPF_XOR | BPF_K:
413 case BPF_ALU | BPF_MUL | BPF_X:
414 case BPF_ALU | BPF_MUL | BPF_K:
415 case BPF_ALU | BPF_DIV | BPF_X:
416 case BPF_ALU | BPF_DIV | BPF_K:
417 case BPF_ALU | BPF_MOD | BPF_X:
418 case BPF_ALU | BPF_MOD | BPF_K:
419 case BPF_ALU | BPF_NEG:
420 case BPF_LD | BPF_ABS | BPF_W:
421 case BPF_LD | BPF_ABS | BPF_H:
422 case BPF_LD | BPF_ABS | BPF_B:
423 case BPF_LD | BPF_IND | BPF_W:
424 case BPF_LD | BPF_IND | BPF_H:
425 case BPF_LD | BPF_IND | BPF_B:
426 /* Check for overloaded BPF extension and
427 * directly convert it if found, otherwise
428 * just move on with mapping.
430 if (BPF_CLASS(fp->code) == BPF_LD &&
431 BPF_MODE(fp->code) == BPF_ABS &&
432 convert_bpf_extensions(fp, &insn))
435 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
438 /* Jump transformation cannot use BPF block macros
439 * everywhere as offset calculation and target updates
440 * require a bit more work than the rest, i.e. jump
441 * opcodes map as-is, but offsets need adjustment.
444 #define BPF_EMIT_JMP \
446 if (target >= len || target < 0) \
448 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
449 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
450 insn->off -= insn - tmp_insns; \
453 case BPF_JMP | BPF_JA:
454 target = i + fp->k + 1;
455 insn->code = fp->code;
459 case BPF_JMP | BPF_JEQ | BPF_K:
460 case BPF_JMP | BPF_JEQ | BPF_X:
461 case BPF_JMP | BPF_JSET | BPF_K:
462 case BPF_JMP | BPF_JSET | BPF_X:
463 case BPF_JMP | BPF_JGT | BPF_K:
464 case BPF_JMP | BPF_JGT | BPF_X:
465 case BPF_JMP | BPF_JGE | BPF_K:
466 case BPF_JMP | BPF_JGE | BPF_X:
467 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
468 /* BPF immediates are signed, zero extend
469 * immediate into tmp register and use it
472 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
474 insn->dst_reg = BPF_REG_A;
475 insn->src_reg = BPF_REG_TMP;
478 insn->dst_reg = BPF_REG_A;
479 insn->src_reg = BPF_REG_X;
481 bpf_src = BPF_SRC(fp->code);
484 /* Common case where 'jump_false' is next insn. */
486 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
487 target = i + fp->jt + 1;
492 /* Convert JEQ into JNE when 'jump_true' is next insn. */
493 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
494 insn->code = BPF_JMP | BPF_JNE | bpf_src;
495 target = i + fp->jf + 1;
500 /* Other jumps are mapped into two insns: Jxx and JA. */
501 target = i + fp->jt + 1;
502 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
506 insn->code = BPF_JMP | BPF_JA;
507 target = i + fp->jf + 1;
511 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
512 case BPF_LDX | BPF_MSH | BPF_B:
514 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
515 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
516 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
518 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
520 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
522 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
524 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
527 /* RET_K, RET_A are remaped into 2 insns. */
528 case BPF_RET | BPF_A:
529 case BPF_RET | BPF_K:
530 *insn++ = BPF_MOV32_RAW(BPF_RVAL(fp->code) == BPF_K ?
531 BPF_K : BPF_X, BPF_REG_0,
533 *insn = BPF_EXIT_INSN();
536 /* Store to stack. */
539 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
540 BPF_ST ? BPF_REG_A : BPF_REG_X,
541 -(BPF_MEMWORDS - fp->k) * 4);
544 /* Load from stack. */
545 case BPF_LD | BPF_MEM:
546 case BPF_LDX | BPF_MEM:
547 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
548 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
549 -(BPF_MEMWORDS - fp->k) * 4);
553 case BPF_LD | BPF_IMM:
554 case BPF_LDX | BPF_IMM:
555 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
556 BPF_REG_A : BPF_REG_X, fp->k);
560 case BPF_MISC | BPF_TAX:
561 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
565 case BPF_MISC | BPF_TXA:
566 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
569 /* A = skb->len or X = skb->len */
570 case BPF_LD | BPF_W | BPF_LEN:
571 case BPF_LDX | BPF_W | BPF_LEN:
572 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
573 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
574 offsetof(struct sk_buff, len));
577 /* Access seccomp_data fields. */
578 case BPF_LDX | BPF_ABS | BPF_W:
579 /* A = *(u32 *) (ctx + K) */
580 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
583 /* Unknown instruction. */
590 memcpy(new_insn, tmp_insns,
591 sizeof(*insn) * (insn - tmp_insns));
592 new_insn += insn - tmp_insns;
596 /* Only calculating new length. */
597 *new_len = new_insn - new_prog;
602 if (new_flen != new_insn - new_prog) {
603 new_flen = new_insn - new_prog;
610 BUG_ON(*new_len != new_flen);
619 * As we dont want to clear mem[] array for each packet going through
620 * __bpf_prog_run(), we check that filter loaded by user never try to read
621 * a cell if not previously written, and we check all branches to be sure
622 * a malicious user doesn't try to abuse us.
624 static int check_load_and_stores(const struct sock_filter *filter, int flen)
626 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
629 BUILD_BUG_ON(BPF_MEMWORDS > 16);
631 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
635 memset(masks, 0xff, flen * sizeof(*masks));
637 for (pc = 0; pc < flen; pc++) {
638 memvalid &= masks[pc];
640 switch (filter[pc].code) {
643 memvalid |= (1 << filter[pc].k);
645 case BPF_LD | BPF_MEM:
646 case BPF_LDX | BPF_MEM:
647 if (!(memvalid & (1 << filter[pc].k))) {
652 case BPF_JMP | BPF_JA:
653 /* A jump must set masks on target */
654 masks[pc + 1 + filter[pc].k] &= memvalid;
657 case BPF_JMP | BPF_JEQ | BPF_K:
658 case BPF_JMP | BPF_JEQ | BPF_X:
659 case BPF_JMP | BPF_JGE | BPF_K:
660 case BPF_JMP | BPF_JGE | BPF_X:
661 case BPF_JMP | BPF_JGT | BPF_K:
662 case BPF_JMP | BPF_JGT | BPF_X:
663 case BPF_JMP | BPF_JSET | BPF_K:
664 case BPF_JMP | BPF_JSET | BPF_X:
665 /* A jump must set masks on targets */
666 masks[pc + 1 + filter[pc].jt] &= memvalid;
667 masks[pc + 1 + filter[pc].jf] &= memvalid;
677 static bool chk_code_allowed(u16 code_to_probe)
679 static const bool codes[] = {
680 /* 32 bit ALU operations */
681 [BPF_ALU | BPF_ADD | BPF_K] = true,
682 [BPF_ALU | BPF_ADD | BPF_X] = true,
683 [BPF_ALU | BPF_SUB | BPF_K] = true,
684 [BPF_ALU | BPF_SUB | BPF_X] = true,
685 [BPF_ALU | BPF_MUL | BPF_K] = true,
686 [BPF_ALU | BPF_MUL | BPF_X] = true,
687 [BPF_ALU | BPF_DIV | BPF_K] = true,
688 [BPF_ALU | BPF_DIV | BPF_X] = true,
689 [BPF_ALU | BPF_MOD | BPF_K] = true,
690 [BPF_ALU | BPF_MOD | BPF_X] = true,
691 [BPF_ALU | BPF_AND | BPF_K] = true,
692 [BPF_ALU | BPF_AND | BPF_X] = true,
693 [BPF_ALU | BPF_OR | BPF_K] = true,
694 [BPF_ALU | BPF_OR | BPF_X] = true,
695 [BPF_ALU | BPF_XOR | BPF_K] = true,
696 [BPF_ALU | BPF_XOR | BPF_X] = true,
697 [BPF_ALU | BPF_LSH | BPF_K] = true,
698 [BPF_ALU | BPF_LSH | BPF_X] = true,
699 [BPF_ALU | BPF_RSH | BPF_K] = true,
700 [BPF_ALU | BPF_RSH | BPF_X] = true,
701 [BPF_ALU | BPF_NEG] = true,
702 /* Load instructions */
703 [BPF_LD | BPF_W | BPF_ABS] = true,
704 [BPF_LD | BPF_H | BPF_ABS] = true,
705 [BPF_LD | BPF_B | BPF_ABS] = true,
706 [BPF_LD | BPF_W | BPF_LEN] = true,
707 [BPF_LD | BPF_W | BPF_IND] = true,
708 [BPF_LD | BPF_H | BPF_IND] = true,
709 [BPF_LD | BPF_B | BPF_IND] = true,
710 [BPF_LD | BPF_IMM] = true,
711 [BPF_LD | BPF_MEM] = true,
712 [BPF_LDX | BPF_W | BPF_LEN] = true,
713 [BPF_LDX | BPF_B | BPF_MSH] = true,
714 [BPF_LDX | BPF_IMM] = true,
715 [BPF_LDX | BPF_MEM] = true,
716 /* Store instructions */
719 /* Misc instructions */
720 [BPF_MISC | BPF_TAX] = true,
721 [BPF_MISC | BPF_TXA] = true,
722 /* Return instructions */
723 [BPF_RET | BPF_K] = true,
724 [BPF_RET | BPF_A] = true,
725 /* Jump instructions */
726 [BPF_JMP | BPF_JA] = true,
727 [BPF_JMP | BPF_JEQ | BPF_K] = true,
728 [BPF_JMP | BPF_JEQ | BPF_X] = true,
729 [BPF_JMP | BPF_JGE | BPF_K] = true,
730 [BPF_JMP | BPF_JGE | BPF_X] = true,
731 [BPF_JMP | BPF_JGT | BPF_K] = true,
732 [BPF_JMP | BPF_JGT | BPF_X] = true,
733 [BPF_JMP | BPF_JSET | BPF_K] = true,
734 [BPF_JMP | BPF_JSET | BPF_X] = true,
737 if (code_to_probe >= ARRAY_SIZE(codes))
740 return codes[code_to_probe];
744 * bpf_check_classic - verify socket filter code
745 * @filter: filter to verify
746 * @flen: length of filter
748 * Check the user's filter code. If we let some ugly
749 * filter code slip through kaboom! The filter must contain
750 * no references or jumps that are out of range, no illegal
751 * instructions, and must end with a RET instruction.
753 * All jumps are forward as they are not signed.
755 * Returns 0 if the rule set is legal or -EINVAL if not.
757 static int bpf_check_classic(const struct sock_filter *filter,
763 if (flen == 0 || flen > BPF_MAXINSNS)
766 /* Check the filter code now */
767 for (pc = 0; pc < flen; pc++) {
768 const struct sock_filter *ftest = &filter[pc];
770 /* May we actually operate on this code? */
771 if (!chk_code_allowed(ftest->code))
774 /* Some instructions need special checks */
775 switch (ftest->code) {
776 case BPF_ALU | BPF_DIV | BPF_K:
777 case BPF_ALU | BPF_MOD | BPF_K:
778 /* Check for division by zero */
782 case BPF_LD | BPF_MEM:
783 case BPF_LDX | BPF_MEM:
786 /* Check for invalid memory addresses */
787 if (ftest->k >= BPF_MEMWORDS)
790 case BPF_JMP | BPF_JA:
791 /* Note, the large ftest->k might cause loops.
792 * Compare this with conditional jumps below,
793 * where offsets are limited. --ANK (981016)
795 if (ftest->k >= (unsigned int)(flen - pc - 1))
798 case BPF_JMP | BPF_JEQ | BPF_K:
799 case BPF_JMP | BPF_JEQ | BPF_X:
800 case BPF_JMP | BPF_JGE | BPF_K:
801 case BPF_JMP | BPF_JGE | BPF_X:
802 case BPF_JMP | BPF_JGT | BPF_K:
803 case BPF_JMP | BPF_JGT | BPF_X:
804 case BPF_JMP | BPF_JSET | BPF_K:
805 case BPF_JMP | BPF_JSET | BPF_X:
806 /* Both conditionals must be safe */
807 if (pc + ftest->jt + 1 >= flen ||
808 pc + ftest->jf + 1 >= flen)
811 case BPF_LD | BPF_W | BPF_ABS:
812 case BPF_LD | BPF_H | BPF_ABS:
813 case BPF_LD | BPF_B | BPF_ABS:
815 if (bpf_anc_helper(ftest) & BPF_ANC)
817 /* Ancillary operation unknown or unsupported */
818 if (anc_found == false && ftest->k >= SKF_AD_OFF)
823 /* Last instruction must be a RET code */
824 switch (filter[flen - 1].code) {
825 case BPF_RET | BPF_K:
826 case BPF_RET | BPF_A:
827 return check_load_and_stores(filter, flen);
833 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
834 const struct sock_fprog *fprog)
836 unsigned int fsize = bpf_classic_proglen(fprog);
837 struct sock_fprog_kern *fkprog;
839 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
843 fkprog = fp->orig_prog;
844 fkprog->len = fprog->len;
846 fkprog->filter = kmemdup(fp->insns, fsize,
847 GFP_KERNEL | __GFP_NOWARN);
848 if (!fkprog->filter) {
849 kfree(fp->orig_prog);
856 static void bpf_release_orig_filter(struct bpf_prog *fp)
858 struct sock_fprog_kern *fprog = fp->orig_prog;
861 kfree(fprog->filter);
866 static void __bpf_prog_release(struct bpf_prog *prog)
868 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
871 bpf_release_orig_filter(prog);
876 static void __sk_filter_release(struct sk_filter *fp)
878 __bpf_prog_release(fp->prog);
883 * sk_filter_release_rcu - Release a socket filter by rcu_head
884 * @rcu: rcu_head that contains the sk_filter to free
886 static void sk_filter_release_rcu(struct rcu_head *rcu)
888 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
890 __sk_filter_release(fp);
894 * sk_filter_release - release a socket filter
895 * @fp: filter to remove
897 * Remove a filter from a socket and release its resources.
899 static void sk_filter_release(struct sk_filter *fp)
901 if (atomic_dec_and_test(&fp->refcnt))
902 call_rcu(&fp->rcu, sk_filter_release_rcu);
905 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
907 u32 filter_size = bpf_prog_size(fp->prog->len);
909 atomic_sub(filter_size, &sk->sk_omem_alloc);
910 sk_filter_release(fp);
913 /* try to charge the socket memory if there is space available
914 * return true on success
916 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
918 u32 filter_size = bpf_prog_size(fp->prog->len);
920 /* same check as in sock_kmalloc() */
921 if (filter_size <= sysctl_optmem_max &&
922 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
923 atomic_inc(&fp->refcnt);
924 atomic_add(filter_size, &sk->sk_omem_alloc);
930 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
932 struct sock_filter *old_prog;
933 struct bpf_prog *old_fp;
934 int err, new_len, old_len = fp->len;
936 /* We are free to overwrite insns et al right here as it
937 * won't be used at this point in time anymore internally
938 * after the migration to the internal BPF instruction
941 BUILD_BUG_ON(sizeof(struct sock_filter) !=
942 sizeof(struct bpf_insn));
944 /* Conversion cannot happen on overlapping memory areas,
945 * so we need to keep the user BPF around until the 2nd
946 * pass. At this time, the user BPF is stored in fp->insns.
948 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
949 GFP_KERNEL | __GFP_NOWARN);
955 /* 1st pass: calculate the new program length. */
956 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
960 /* Expand fp for appending the new filter representation. */
962 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
964 /* The old_fp is still around in case we couldn't
965 * allocate new memory, so uncharge on that one.
974 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
975 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
977 /* 2nd bpf_convert_filter() can fail only if it fails
978 * to allocate memory, remapping must succeed. Note,
979 * that at this time old_fp has already been released
984 bpf_prog_select_runtime(fp);
992 __bpf_prog_release(fp);
996 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
997 bpf_aux_classic_check_t trans)
1001 fp->bpf_func = NULL;
1004 err = bpf_check_classic(fp->insns, fp->len);
1006 __bpf_prog_release(fp);
1007 return ERR_PTR(err);
1010 /* There might be additional checks and transformations
1011 * needed on classic filters, f.e. in case of seccomp.
1014 err = trans(fp->insns, fp->len);
1016 __bpf_prog_release(fp);
1017 return ERR_PTR(err);
1021 /* Probe if we can JIT compile the filter and if so, do
1022 * the compilation of the filter.
1024 bpf_jit_compile(fp);
1026 /* JIT compiler couldn't process this filter, so do the
1027 * internal BPF translation for the optimized interpreter.
1030 fp = bpf_migrate_filter(fp);
1036 * bpf_prog_create - create an unattached filter
1037 * @pfp: the unattached filter that is created
1038 * @fprog: the filter program
1040 * Create a filter independent of any socket. We first run some
1041 * sanity checks on it to make sure it does not explode on us later.
1042 * If an error occurs or there is insufficient memory for the filter
1043 * a negative errno code is returned. On success the return is zero.
1045 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1047 unsigned int fsize = bpf_classic_proglen(fprog);
1048 struct bpf_prog *fp;
1050 /* Make sure new filter is there and in the right amounts. */
1051 if (fprog->filter == NULL)
1054 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1058 memcpy(fp->insns, fprog->filter, fsize);
1060 fp->len = fprog->len;
1061 /* Since unattached filters are not copied back to user
1062 * space through sk_get_filter(), we do not need to hold
1063 * a copy here, and can spare us the work.
1065 fp->orig_prog = NULL;
1067 /* bpf_prepare_filter() already takes care of freeing
1068 * memory in case something goes wrong.
1070 fp = bpf_prepare_filter(fp, NULL);
1077 EXPORT_SYMBOL_GPL(bpf_prog_create);
1080 * bpf_prog_create_from_user - create an unattached filter from user buffer
1081 * @pfp: the unattached filter that is created
1082 * @fprog: the filter program
1083 * @trans: post-classic verifier transformation handler
1085 * This function effectively does the same as bpf_prog_create(), only
1086 * that it builds up its insns buffer from user space provided buffer.
1087 * It also allows for passing a bpf_aux_classic_check_t handler.
1089 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1090 bpf_aux_classic_check_t trans)
1092 unsigned int fsize = bpf_classic_proglen(fprog);
1093 struct bpf_prog *fp;
1095 /* Make sure new filter is there and in the right amounts. */
1096 if (fprog->filter == NULL)
1099 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1103 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1104 __bpf_prog_free(fp);
1108 fp->len = fprog->len;
1109 /* Since unattached filters are not copied back to user
1110 * space through sk_get_filter(), we do not need to hold
1111 * a copy here, and can spare us the work.
1113 fp->orig_prog = NULL;
1115 /* bpf_prepare_filter() already takes care of freeing
1116 * memory in case something goes wrong.
1118 fp = bpf_prepare_filter(fp, trans);
1126 void bpf_prog_destroy(struct bpf_prog *fp)
1128 __bpf_prog_release(fp);
1130 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1132 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1134 struct sk_filter *fp, *old_fp;
1136 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1141 atomic_set(&fp->refcnt, 0);
1143 if (!sk_filter_charge(sk, fp)) {
1148 old_fp = rcu_dereference_protected(sk->sk_filter,
1149 sock_owned_by_user(sk));
1150 rcu_assign_pointer(sk->sk_filter, fp);
1153 sk_filter_uncharge(sk, old_fp);
1159 * sk_attach_filter - attach a socket filter
1160 * @fprog: the filter program
1161 * @sk: the socket to use
1163 * Attach the user's filter code. We first run some sanity checks on
1164 * it to make sure it does not explode on us later. If an error
1165 * occurs or there is insufficient memory for the filter a negative
1166 * errno code is returned. On success the return is zero.
1168 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1170 unsigned int fsize = bpf_classic_proglen(fprog);
1171 unsigned int bpf_fsize = bpf_prog_size(fprog->len);
1172 struct bpf_prog *prog;
1175 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1178 /* Make sure new filter is there and in the right amounts. */
1179 if (fprog->filter == NULL)
1182 prog = bpf_prog_alloc(bpf_fsize, 0);
1186 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1187 __bpf_prog_free(prog);
1191 prog->len = fprog->len;
1193 err = bpf_prog_store_orig_filter(prog, fprog);
1195 __bpf_prog_free(prog);
1199 /* bpf_prepare_filter() already takes care of freeing
1200 * memory in case something goes wrong.
1202 prog = bpf_prepare_filter(prog, NULL);
1204 return PTR_ERR(prog);
1206 err = __sk_attach_prog(prog, sk);
1208 __bpf_prog_release(prog);
1214 EXPORT_SYMBOL_GPL(sk_attach_filter);
1216 int sk_attach_bpf(u32 ufd, struct sock *sk)
1218 struct bpf_prog *prog;
1221 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1224 prog = bpf_prog_get(ufd);
1226 return PTR_ERR(prog);
1228 if (prog->type != BPF_PROG_TYPE_SOCKET_FILTER) {
1233 err = __sk_attach_prog(prog, sk);
1242 #define BPF_RECOMPUTE_CSUM(flags) ((flags) & 1)
1244 static u64 bpf_skb_store_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 flags)
1246 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1247 int offset = (int) r2;
1248 void *from = (void *) (long) r3;
1249 unsigned int len = (unsigned int) r4;
1253 /* bpf verifier guarantees that:
1254 * 'from' pointer points to bpf program stack
1255 * 'len' bytes of it were initialized
1257 * 'skb' is a valid pointer to 'struct sk_buff'
1259 * so check for invalid 'offset' and too large 'len'
1261 if (unlikely((u32) offset > 0xffff || len > sizeof(buf)))
1264 if (unlikely(skb_cloned(skb) &&
1265 !skb_clone_writable(skb, offset + len)))
1268 ptr = skb_header_pointer(skb, offset, len, buf);
1272 if (BPF_RECOMPUTE_CSUM(flags))
1273 skb_postpull_rcsum(skb, ptr, len);
1275 memcpy(ptr, from, len);
1278 /* skb_store_bits cannot return -EFAULT here */
1279 skb_store_bits(skb, offset, ptr, len);
1281 if (BPF_RECOMPUTE_CSUM(flags) && skb->ip_summed == CHECKSUM_COMPLETE)
1282 skb->csum = csum_add(skb->csum, csum_partial(ptr, len, 0));
1286 const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1287 .func = bpf_skb_store_bytes,
1289 .ret_type = RET_INTEGER,
1290 .arg1_type = ARG_PTR_TO_CTX,
1291 .arg2_type = ARG_ANYTHING,
1292 .arg3_type = ARG_PTR_TO_STACK,
1293 .arg4_type = ARG_CONST_STACK_SIZE,
1294 .arg5_type = ARG_ANYTHING,
1297 #define BPF_HEADER_FIELD_SIZE(flags) ((flags) & 0x0f)
1298 #define BPF_IS_PSEUDO_HEADER(flags) ((flags) & 0x10)
1300 static u64 bpf_l3_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1302 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1303 int offset = (int) r2;
1306 if (unlikely((u32) offset > 0xffff))
1309 if (unlikely(skb_cloned(skb) &&
1310 !skb_clone_writable(skb, offset + sizeof(sum))))
1313 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1317 switch (BPF_HEADER_FIELD_SIZE(flags)) {
1319 csum_replace2(ptr, from, to);
1322 csum_replace4(ptr, from, to);
1329 /* skb_store_bits guaranteed to not return -EFAULT here */
1330 skb_store_bits(skb, offset, ptr, sizeof(sum));
1335 const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1336 .func = bpf_l3_csum_replace,
1338 .ret_type = RET_INTEGER,
1339 .arg1_type = ARG_PTR_TO_CTX,
1340 .arg2_type = ARG_ANYTHING,
1341 .arg3_type = ARG_ANYTHING,
1342 .arg4_type = ARG_ANYTHING,
1343 .arg5_type = ARG_ANYTHING,
1346 static u64 bpf_l4_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1348 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1349 u32 is_pseudo = BPF_IS_PSEUDO_HEADER(flags);
1350 int offset = (int) r2;
1353 if (unlikely((u32) offset > 0xffff))
1356 if (unlikely(skb_cloned(skb) &&
1357 !skb_clone_writable(skb, offset + sizeof(sum))))
1360 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1364 switch (BPF_HEADER_FIELD_SIZE(flags)) {
1366 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1369 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1376 /* skb_store_bits guaranteed to not return -EFAULT here */
1377 skb_store_bits(skb, offset, ptr, sizeof(sum));
1382 const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1383 .func = bpf_l4_csum_replace,
1385 .ret_type = RET_INTEGER,
1386 .arg1_type = ARG_PTR_TO_CTX,
1387 .arg2_type = ARG_ANYTHING,
1388 .arg3_type = ARG_ANYTHING,
1389 .arg4_type = ARG_ANYTHING,
1390 .arg5_type = ARG_ANYTHING,
1393 #define BPF_IS_REDIRECT_INGRESS(flags) ((flags) & 1)
1395 static u64 bpf_clone_redirect(u64 r1, u64 ifindex, u64 flags, u64 r4, u64 r5)
1397 struct sk_buff *skb = (struct sk_buff *) (long) r1, *skb2;
1398 struct net_device *dev;
1400 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1404 if (unlikely(!(dev->flags & IFF_UP)))
1407 skb2 = skb_clone(skb, GFP_ATOMIC);
1408 if (unlikely(!skb2))
1411 if (BPF_IS_REDIRECT_INGRESS(flags))
1412 return dev_forward_skb(dev, skb2);
1415 return dev_queue_xmit(skb2);
1418 const struct bpf_func_proto bpf_clone_redirect_proto = {
1419 .func = bpf_clone_redirect,
1421 .ret_type = RET_INTEGER,
1422 .arg1_type = ARG_PTR_TO_CTX,
1423 .arg2_type = ARG_ANYTHING,
1424 .arg3_type = ARG_ANYTHING,
1427 static const struct bpf_func_proto *
1428 sk_filter_func_proto(enum bpf_func_id func_id)
1431 case BPF_FUNC_map_lookup_elem:
1432 return &bpf_map_lookup_elem_proto;
1433 case BPF_FUNC_map_update_elem:
1434 return &bpf_map_update_elem_proto;
1435 case BPF_FUNC_map_delete_elem:
1436 return &bpf_map_delete_elem_proto;
1437 case BPF_FUNC_get_prandom_u32:
1438 return &bpf_get_prandom_u32_proto;
1439 case BPF_FUNC_get_smp_processor_id:
1440 return &bpf_get_smp_processor_id_proto;
1441 case BPF_FUNC_tail_call:
1442 return &bpf_tail_call_proto;
1443 case BPF_FUNC_ktime_get_ns:
1444 return &bpf_ktime_get_ns_proto;
1445 case BPF_FUNC_trace_printk:
1446 return bpf_get_trace_printk_proto();
1452 static const struct bpf_func_proto *
1453 tc_cls_act_func_proto(enum bpf_func_id func_id)
1456 case BPF_FUNC_skb_store_bytes:
1457 return &bpf_skb_store_bytes_proto;
1458 case BPF_FUNC_l3_csum_replace:
1459 return &bpf_l3_csum_replace_proto;
1460 case BPF_FUNC_l4_csum_replace:
1461 return &bpf_l4_csum_replace_proto;
1462 case BPF_FUNC_clone_redirect:
1463 return &bpf_clone_redirect_proto;
1465 return sk_filter_func_proto(func_id);
1469 static bool __is_valid_access(int off, int size, enum bpf_access_type type)
1472 if (off < 0 || off >= sizeof(struct __sk_buff))
1475 /* disallow misaligned access */
1476 if (off % size != 0)
1479 /* all __sk_buff fields are __u32 */
1486 static bool sk_filter_is_valid_access(int off, int size,
1487 enum bpf_access_type type)
1489 if (type == BPF_WRITE) {
1491 case offsetof(struct __sk_buff, cb[0]) ...
1492 offsetof(struct __sk_buff, cb[4]):
1499 return __is_valid_access(off, size, type);
1502 static bool tc_cls_act_is_valid_access(int off, int size,
1503 enum bpf_access_type type)
1505 if (type == BPF_WRITE) {
1507 case offsetof(struct __sk_buff, mark):
1508 case offsetof(struct __sk_buff, tc_index):
1509 case offsetof(struct __sk_buff, cb[0]) ...
1510 offsetof(struct __sk_buff, cb[4]):
1516 return __is_valid_access(off, size, type);
1519 static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
1520 int src_reg, int ctx_off,
1521 struct bpf_insn *insn_buf)
1523 struct bpf_insn *insn = insn_buf;
1526 case offsetof(struct __sk_buff, len):
1527 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
1529 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1530 offsetof(struct sk_buff, len));
1533 case offsetof(struct __sk_buff, protocol):
1534 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
1536 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1537 offsetof(struct sk_buff, protocol));
1540 case offsetof(struct __sk_buff, vlan_proto):
1541 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
1543 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1544 offsetof(struct sk_buff, vlan_proto));
1547 case offsetof(struct __sk_buff, priority):
1548 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
1550 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1551 offsetof(struct sk_buff, priority));
1554 case offsetof(struct __sk_buff, ingress_ifindex):
1555 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
1557 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1558 offsetof(struct sk_buff, skb_iif));
1561 case offsetof(struct __sk_buff, ifindex):
1562 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
1564 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
1566 offsetof(struct sk_buff, dev));
1567 *insn++ = BPF_JMP_IMM(BPF_JEQ, dst_reg, 0, 1);
1568 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
1569 offsetof(struct net_device, ifindex));
1572 case offsetof(struct __sk_buff, mark):
1573 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
1575 if (type == BPF_WRITE)
1576 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
1577 offsetof(struct sk_buff, mark));
1579 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
1580 offsetof(struct sk_buff, mark));
1583 case offsetof(struct __sk_buff, pkt_type):
1584 return convert_skb_access(SKF_AD_PKTTYPE, dst_reg, src_reg, insn);
1586 case offsetof(struct __sk_buff, queue_mapping):
1587 return convert_skb_access(SKF_AD_QUEUE, dst_reg, src_reg, insn);
1589 case offsetof(struct __sk_buff, vlan_present):
1590 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
1591 dst_reg, src_reg, insn);
1593 case offsetof(struct __sk_buff, vlan_tci):
1594 return convert_skb_access(SKF_AD_VLAN_TAG,
1595 dst_reg, src_reg, insn);
1597 case offsetof(struct __sk_buff, cb[0]) ...
1598 offsetof(struct __sk_buff, cb[4]):
1599 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
1601 ctx_off -= offsetof(struct __sk_buff, cb[0]);
1602 ctx_off += offsetof(struct sk_buff, cb);
1603 ctx_off += offsetof(struct qdisc_skb_cb, data);
1604 if (type == BPF_WRITE)
1605 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
1607 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
1610 case offsetof(struct __sk_buff, tc_index):
1611 #ifdef CONFIG_NET_SCHED
1612 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
1614 if (type == BPF_WRITE)
1615 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg,
1616 offsetof(struct sk_buff, tc_index));
1618 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
1619 offsetof(struct sk_buff, tc_index));
1622 if (type == BPF_WRITE)
1623 *insn++ = BPF_MOV64_REG(dst_reg, dst_reg);
1625 *insn++ = BPF_MOV64_IMM(dst_reg, 0);
1630 return insn - insn_buf;
1633 static const struct bpf_verifier_ops sk_filter_ops = {
1634 .get_func_proto = sk_filter_func_proto,
1635 .is_valid_access = sk_filter_is_valid_access,
1636 .convert_ctx_access = bpf_net_convert_ctx_access,
1639 static const struct bpf_verifier_ops tc_cls_act_ops = {
1640 .get_func_proto = tc_cls_act_func_proto,
1641 .is_valid_access = tc_cls_act_is_valid_access,
1642 .convert_ctx_access = bpf_net_convert_ctx_access,
1645 static struct bpf_prog_type_list sk_filter_type __read_mostly = {
1646 .ops = &sk_filter_ops,
1647 .type = BPF_PROG_TYPE_SOCKET_FILTER,
1650 static struct bpf_prog_type_list sched_cls_type __read_mostly = {
1651 .ops = &tc_cls_act_ops,
1652 .type = BPF_PROG_TYPE_SCHED_CLS,
1655 static struct bpf_prog_type_list sched_act_type __read_mostly = {
1656 .ops = &tc_cls_act_ops,
1657 .type = BPF_PROG_TYPE_SCHED_ACT,
1660 static int __init register_sk_filter_ops(void)
1662 bpf_register_prog_type(&sk_filter_type);
1663 bpf_register_prog_type(&sched_cls_type);
1664 bpf_register_prog_type(&sched_act_type);
1668 late_initcall(register_sk_filter_ops);
1670 int sk_detach_filter(struct sock *sk)
1673 struct sk_filter *filter;
1675 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1678 filter = rcu_dereference_protected(sk->sk_filter,
1679 sock_owned_by_user(sk));
1681 RCU_INIT_POINTER(sk->sk_filter, NULL);
1682 sk_filter_uncharge(sk, filter);
1688 EXPORT_SYMBOL_GPL(sk_detach_filter);
1690 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
1693 struct sock_fprog_kern *fprog;
1694 struct sk_filter *filter;
1698 filter = rcu_dereference_protected(sk->sk_filter,
1699 sock_owned_by_user(sk));
1703 /* We're copying the filter that has been originally attached,
1704 * so no conversion/decode needed anymore.
1706 fprog = filter->prog->orig_prog;
1710 /* User space only enquires number of filter blocks. */
1714 if (len < fprog->len)
1718 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
1721 /* Instead of bytes, the API requests to return the number