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 <linux/errno.h>
40 #include <linux/timer.h>
41 #include <asm/uaccess.h>
42 #include <asm/unaligned.h>
43 #include <linux/filter.h>
44 #include <linux/ratelimit.h>
45 #include <linux/seccomp.h>
46 #include <linux/if_vlan.h>
47 #include <linux/bpf.h>
50 * sk_filter - run a packet through a socket filter
51 * @sk: sock associated with &sk_buff
52 * @skb: buffer to filter
54 * Run the filter code and then cut skb->data to correct size returned by
55 * SK_RUN_FILTER. If pkt_len is 0 we toss packet. If skb->len is smaller
56 * than pkt_len we keep whole skb->data. This is the socket level
57 * wrapper to SK_RUN_FILTER. It returns 0 if the packet should
58 * be accepted or -EPERM if the packet should be tossed.
61 int sk_filter(struct sock *sk, struct sk_buff *skb)
64 struct sk_filter *filter;
67 * If the skb was allocated from pfmemalloc reserves, only
68 * allow SOCK_MEMALLOC sockets to use it as this socket is
71 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC))
74 err = security_sock_rcv_skb(sk, skb);
79 filter = rcu_dereference(sk->sk_filter);
81 unsigned int pkt_len = SK_RUN_FILTER(filter, skb);
83 err = pkt_len ? pskb_trim(skb, pkt_len) : -EPERM;
89 EXPORT_SYMBOL(sk_filter);
91 static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
93 return skb_get_poff((struct sk_buff *)(unsigned long) ctx);
96 static u64 __skb_get_nlattr(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
98 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
101 if (skb_is_nonlinear(skb))
104 if (skb->len < sizeof(struct nlattr))
107 if (a > skb->len - sizeof(struct nlattr))
110 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
112 return (void *) nla - (void *) skb->data;
117 static u64 __skb_get_nlattr_nest(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
119 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
122 if (skb_is_nonlinear(skb))
125 if (skb->len < sizeof(struct nlattr))
128 if (a > skb->len - sizeof(struct nlattr))
131 nla = (struct nlattr *) &skb->data[a];
132 if (nla->nla_len > skb->len - a)
135 nla = nla_find_nested(nla, x);
137 return (void *) nla - (void *) skb->data;
142 static u64 __get_raw_cpu_id(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
144 return raw_smp_processor_id();
147 /* note that this only generates 32-bit random numbers */
148 static u64 __get_random_u32(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
150 return prandom_u32();
153 static bool convert_bpf_extensions(struct sock_filter *fp,
154 struct bpf_insn **insnp)
156 struct bpf_insn *insn = *insnp;
159 case SKF_AD_OFF + SKF_AD_PROTOCOL:
160 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
162 /* A = *(u16 *) (CTX + offsetof(protocol)) */
163 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
164 offsetof(struct sk_buff, protocol));
165 /* A = ntohs(A) [emitting a nop or swap16] */
166 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
169 case SKF_AD_OFF + SKF_AD_PKTTYPE:
170 *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_A, BPF_REG_CTX,
172 *insn = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, PKT_TYPE_MAX);
173 #ifdef __BIG_ENDIAN_BITFIELD
175 *insn = BPF_ALU32_IMM(BPF_RSH, BPF_REG_A, 5);
179 case SKF_AD_OFF + SKF_AD_IFINDEX:
180 case SKF_AD_OFF + SKF_AD_HATYPE:
181 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
182 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
183 BUILD_BUG_ON(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)) < 0);
185 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
186 BPF_REG_TMP, BPF_REG_CTX,
187 offsetof(struct sk_buff, dev));
188 /* if (tmp != 0) goto pc + 1 */
189 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
190 *insn++ = BPF_EXIT_INSN();
191 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
192 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
193 offsetof(struct net_device, ifindex));
195 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
196 offsetof(struct net_device, type));
199 case SKF_AD_OFF + SKF_AD_MARK:
200 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
202 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
203 offsetof(struct sk_buff, mark));
206 case SKF_AD_OFF + SKF_AD_RXHASH:
207 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
209 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
210 offsetof(struct sk_buff, hash));
213 case SKF_AD_OFF + SKF_AD_QUEUE:
214 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
216 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
217 offsetof(struct sk_buff, queue_mapping));
220 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
221 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
222 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
223 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
225 /* A = *(u16 *) (CTX + offsetof(vlan_tci)) */
226 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
227 offsetof(struct sk_buff, vlan_tci));
228 if (fp->k == SKF_AD_OFF + SKF_AD_VLAN_TAG) {
229 *insn = BPF_ALU32_IMM(BPF_AND, BPF_REG_A,
233 *insn++ = BPF_ALU32_IMM(BPF_RSH, BPF_REG_A, 12);
235 *insn = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 1);
239 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
240 case SKF_AD_OFF + SKF_AD_NLATTR:
241 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
242 case SKF_AD_OFF + SKF_AD_CPU:
243 case SKF_AD_OFF + SKF_AD_RANDOM:
245 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
247 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
249 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
250 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
252 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
253 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
255 case SKF_AD_OFF + SKF_AD_NLATTR:
256 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
258 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
259 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
261 case SKF_AD_OFF + SKF_AD_CPU:
262 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
264 case SKF_AD_OFF + SKF_AD_RANDOM:
265 *insn = BPF_EMIT_CALL(__get_random_u32);
270 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
272 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
276 /* This is just a dummy call to avoid letting the compiler
277 * evict __bpf_call_base() as an optimization. Placed here
278 * where no-one bothers.
280 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
289 * bpf_convert_filter - convert filter program
290 * @prog: the user passed filter program
291 * @len: the length of the user passed filter program
292 * @new_prog: buffer where converted program will be stored
293 * @new_len: pointer to store length of converted program
295 * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style.
296 * Conversion workflow:
298 * 1) First pass for calculating the new program length:
299 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
301 * 2) 2nd pass to remap in two passes: 1st pass finds new
302 * jump offsets, 2nd pass remapping:
303 * new_prog = kmalloc(sizeof(struct bpf_insn) * new_len);
304 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
306 * User BPF's register A is mapped to our BPF register 6, user BPF
307 * register X is mapped to BPF register 7; frame pointer is always
308 * register 10; Context 'void *ctx' is stored in register 1, that is,
309 * for socket filters: ctx == 'struct sk_buff *', for seccomp:
310 * ctx == 'struct seccomp_data *'.
312 int bpf_convert_filter(struct sock_filter *prog, int len,
313 struct bpf_insn *new_prog, int *new_len)
315 int new_flen = 0, pass = 0, target, i;
316 struct bpf_insn *new_insn;
317 struct sock_filter *fp;
321 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
322 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
324 if (len <= 0 || len > BPF_MAXINSNS)
328 addrs = kcalloc(len, sizeof(*addrs), GFP_KERNEL);
338 *new_insn = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
341 for (i = 0; i < len; fp++, i++) {
342 struct bpf_insn tmp_insns[6] = { };
343 struct bpf_insn *insn = tmp_insns;
346 addrs[i] = new_insn - new_prog;
349 /* All arithmetic insns and skb loads map as-is. */
350 case BPF_ALU | BPF_ADD | BPF_X:
351 case BPF_ALU | BPF_ADD | BPF_K:
352 case BPF_ALU | BPF_SUB | BPF_X:
353 case BPF_ALU | BPF_SUB | BPF_K:
354 case BPF_ALU | BPF_AND | BPF_X:
355 case BPF_ALU | BPF_AND | BPF_K:
356 case BPF_ALU | BPF_OR | BPF_X:
357 case BPF_ALU | BPF_OR | BPF_K:
358 case BPF_ALU | BPF_LSH | BPF_X:
359 case BPF_ALU | BPF_LSH | BPF_K:
360 case BPF_ALU | BPF_RSH | BPF_X:
361 case BPF_ALU | BPF_RSH | BPF_K:
362 case BPF_ALU | BPF_XOR | BPF_X:
363 case BPF_ALU | BPF_XOR | BPF_K:
364 case BPF_ALU | BPF_MUL | BPF_X:
365 case BPF_ALU | BPF_MUL | BPF_K:
366 case BPF_ALU | BPF_DIV | BPF_X:
367 case BPF_ALU | BPF_DIV | BPF_K:
368 case BPF_ALU | BPF_MOD | BPF_X:
369 case BPF_ALU | BPF_MOD | BPF_K:
370 case BPF_ALU | BPF_NEG:
371 case BPF_LD | BPF_ABS | BPF_W:
372 case BPF_LD | BPF_ABS | BPF_H:
373 case BPF_LD | BPF_ABS | BPF_B:
374 case BPF_LD | BPF_IND | BPF_W:
375 case BPF_LD | BPF_IND | BPF_H:
376 case BPF_LD | BPF_IND | BPF_B:
377 /* Check for overloaded BPF extension and
378 * directly convert it if found, otherwise
379 * just move on with mapping.
381 if (BPF_CLASS(fp->code) == BPF_LD &&
382 BPF_MODE(fp->code) == BPF_ABS &&
383 convert_bpf_extensions(fp, &insn))
386 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
389 /* Jump transformation cannot use BPF block macros
390 * everywhere as offset calculation and target updates
391 * require a bit more work than the rest, i.e. jump
392 * opcodes map as-is, but offsets need adjustment.
395 #define BPF_EMIT_JMP \
397 if (target >= len || target < 0) \
399 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
400 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
401 insn->off -= insn - tmp_insns; \
404 case BPF_JMP | BPF_JA:
405 target = i + fp->k + 1;
406 insn->code = fp->code;
410 case BPF_JMP | BPF_JEQ | BPF_K:
411 case BPF_JMP | BPF_JEQ | BPF_X:
412 case BPF_JMP | BPF_JSET | BPF_K:
413 case BPF_JMP | BPF_JSET | BPF_X:
414 case BPF_JMP | BPF_JGT | BPF_K:
415 case BPF_JMP | BPF_JGT | BPF_X:
416 case BPF_JMP | BPF_JGE | BPF_K:
417 case BPF_JMP | BPF_JGE | BPF_X:
418 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
419 /* BPF immediates are signed, zero extend
420 * immediate into tmp register and use it
423 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
425 insn->dst_reg = BPF_REG_A;
426 insn->src_reg = BPF_REG_TMP;
429 insn->dst_reg = BPF_REG_A;
430 insn->src_reg = BPF_REG_X;
432 bpf_src = BPF_SRC(fp->code);
435 /* Common case where 'jump_false' is next insn. */
437 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
438 target = i + fp->jt + 1;
443 /* Convert JEQ into JNE when 'jump_true' is next insn. */
444 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
445 insn->code = BPF_JMP | BPF_JNE | bpf_src;
446 target = i + fp->jf + 1;
451 /* Other jumps are mapped into two insns: Jxx and JA. */
452 target = i + fp->jt + 1;
453 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
457 insn->code = BPF_JMP | BPF_JA;
458 target = i + fp->jf + 1;
462 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
463 case BPF_LDX | BPF_MSH | BPF_B:
465 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
466 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
467 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
469 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
471 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
473 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
475 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
478 /* RET_K, RET_A are remaped into 2 insns. */
479 case BPF_RET | BPF_A:
480 case BPF_RET | BPF_K:
481 *insn++ = BPF_MOV32_RAW(BPF_RVAL(fp->code) == BPF_K ?
482 BPF_K : BPF_X, BPF_REG_0,
484 *insn = BPF_EXIT_INSN();
487 /* Store to stack. */
490 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
491 BPF_ST ? BPF_REG_A : BPF_REG_X,
492 -(BPF_MEMWORDS - fp->k) * 4);
495 /* Load from stack. */
496 case BPF_LD | BPF_MEM:
497 case BPF_LDX | BPF_MEM:
498 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
499 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
500 -(BPF_MEMWORDS - fp->k) * 4);
504 case BPF_LD | BPF_IMM:
505 case BPF_LDX | BPF_IMM:
506 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
507 BPF_REG_A : BPF_REG_X, fp->k);
511 case BPF_MISC | BPF_TAX:
512 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
516 case BPF_MISC | BPF_TXA:
517 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
520 /* A = skb->len or X = skb->len */
521 case BPF_LD | BPF_W | BPF_LEN:
522 case BPF_LDX | BPF_W | BPF_LEN:
523 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
524 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
525 offsetof(struct sk_buff, len));
528 /* Access seccomp_data fields. */
529 case BPF_LDX | BPF_ABS | BPF_W:
530 /* A = *(u32 *) (ctx + K) */
531 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
534 /* Unkown instruction. */
541 memcpy(new_insn, tmp_insns,
542 sizeof(*insn) * (insn - tmp_insns));
543 new_insn += insn - tmp_insns;
547 /* Only calculating new length. */
548 *new_len = new_insn - new_prog;
553 if (new_flen != new_insn - new_prog) {
554 new_flen = new_insn - new_prog;
561 BUG_ON(*new_len != new_flen);
570 * As we dont want to clear mem[] array for each packet going through
571 * __bpf_prog_run(), we check that filter loaded by user never try to read
572 * a cell if not previously written, and we check all branches to be sure
573 * a malicious user doesn't try to abuse us.
575 static int check_load_and_stores(const struct sock_filter *filter, int flen)
577 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
580 BUILD_BUG_ON(BPF_MEMWORDS > 16);
582 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
586 memset(masks, 0xff, flen * sizeof(*masks));
588 for (pc = 0; pc < flen; pc++) {
589 memvalid &= masks[pc];
591 switch (filter[pc].code) {
594 memvalid |= (1 << filter[pc].k);
596 case BPF_LD | BPF_MEM:
597 case BPF_LDX | BPF_MEM:
598 if (!(memvalid & (1 << filter[pc].k))) {
603 case BPF_JMP | BPF_JA:
604 /* A jump must set masks on target */
605 masks[pc + 1 + filter[pc].k] &= memvalid;
608 case BPF_JMP | BPF_JEQ | BPF_K:
609 case BPF_JMP | BPF_JEQ | BPF_X:
610 case BPF_JMP | BPF_JGE | BPF_K:
611 case BPF_JMP | BPF_JGE | BPF_X:
612 case BPF_JMP | BPF_JGT | BPF_K:
613 case BPF_JMP | BPF_JGT | BPF_X:
614 case BPF_JMP | BPF_JSET | BPF_K:
615 case BPF_JMP | BPF_JSET | BPF_X:
616 /* A jump must set masks on targets */
617 masks[pc + 1 + filter[pc].jt] &= memvalid;
618 masks[pc + 1 + filter[pc].jf] &= memvalid;
628 static bool chk_code_allowed(u16 code_to_probe)
630 static const bool codes[] = {
631 /* 32 bit ALU operations */
632 [BPF_ALU | BPF_ADD | BPF_K] = true,
633 [BPF_ALU | BPF_ADD | BPF_X] = true,
634 [BPF_ALU | BPF_SUB | BPF_K] = true,
635 [BPF_ALU | BPF_SUB | BPF_X] = true,
636 [BPF_ALU | BPF_MUL | BPF_K] = true,
637 [BPF_ALU | BPF_MUL | BPF_X] = true,
638 [BPF_ALU | BPF_DIV | BPF_K] = true,
639 [BPF_ALU | BPF_DIV | BPF_X] = true,
640 [BPF_ALU | BPF_MOD | BPF_K] = true,
641 [BPF_ALU | BPF_MOD | BPF_X] = true,
642 [BPF_ALU | BPF_AND | BPF_K] = true,
643 [BPF_ALU | BPF_AND | BPF_X] = true,
644 [BPF_ALU | BPF_OR | BPF_K] = true,
645 [BPF_ALU | BPF_OR | BPF_X] = true,
646 [BPF_ALU | BPF_XOR | BPF_K] = true,
647 [BPF_ALU | BPF_XOR | BPF_X] = true,
648 [BPF_ALU | BPF_LSH | BPF_K] = true,
649 [BPF_ALU | BPF_LSH | BPF_X] = true,
650 [BPF_ALU | BPF_RSH | BPF_K] = true,
651 [BPF_ALU | BPF_RSH | BPF_X] = true,
652 [BPF_ALU | BPF_NEG] = true,
653 /* Load instructions */
654 [BPF_LD | BPF_W | BPF_ABS] = true,
655 [BPF_LD | BPF_H | BPF_ABS] = true,
656 [BPF_LD | BPF_B | BPF_ABS] = true,
657 [BPF_LD | BPF_W | BPF_LEN] = true,
658 [BPF_LD | BPF_W | BPF_IND] = true,
659 [BPF_LD | BPF_H | BPF_IND] = true,
660 [BPF_LD | BPF_B | BPF_IND] = true,
661 [BPF_LD | BPF_IMM] = true,
662 [BPF_LD | BPF_MEM] = true,
663 [BPF_LDX | BPF_W | BPF_LEN] = true,
664 [BPF_LDX | BPF_B | BPF_MSH] = true,
665 [BPF_LDX | BPF_IMM] = true,
666 [BPF_LDX | BPF_MEM] = true,
667 /* Store instructions */
670 /* Misc instructions */
671 [BPF_MISC | BPF_TAX] = true,
672 [BPF_MISC | BPF_TXA] = true,
673 /* Return instructions */
674 [BPF_RET | BPF_K] = true,
675 [BPF_RET | BPF_A] = true,
676 /* Jump instructions */
677 [BPF_JMP | BPF_JA] = true,
678 [BPF_JMP | BPF_JEQ | BPF_K] = true,
679 [BPF_JMP | BPF_JEQ | BPF_X] = true,
680 [BPF_JMP | BPF_JGE | BPF_K] = true,
681 [BPF_JMP | BPF_JGE | BPF_X] = true,
682 [BPF_JMP | BPF_JGT | BPF_K] = true,
683 [BPF_JMP | BPF_JGT | BPF_X] = true,
684 [BPF_JMP | BPF_JSET | BPF_K] = true,
685 [BPF_JMP | BPF_JSET | BPF_X] = true,
688 if (code_to_probe >= ARRAY_SIZE(codes))
691 return codes[code_to_probe];
695 * bpf_check_classic - verify socket filter code
696 * @filter: filter to verify
697 * @flen: length of filter
699 * Check the user's filter code. If we let some ugly
700 * filter code slip through kaboom! The filter must contain
701 * no references or jumps that are out of range, no illegal
702 * instructions, and must end with a RET instruction.
704 * All jumps are forward as they are not signed.
706 * Returns 0 if the rule set is legal or -EINVAL if not.
708 int bpf_check_classic(const struct sock_filter *filter, unsigned int flen)
713 if (flen == 0 || flen > BPF_MAXINSNS)
716 /* Check the filter code now */
717 for (pc = 0; pc < flen; pc++) {
718 const struct sock_filter *ftest = &filter[pc];
720 /* May we actually operate on this code? */
721 if (!chk_code_allowed(ftest->code))
724 /* Some instructions need special checks */
725 switch (ftest->code) {
726 case BPF_ALU | BPF_DIV | BPF_K:
727 case BPF_ALU | BPF_MOD | BPF_K:
728 /* Check for division by zero */
732 case BPF_LD | BPF_MEM:
733 case BPF_LDX | BPF_MEM:
736 /* Check for invalid memory addresses */
737 if (ftest->k >= BPF_MEMWORDS)
740 case BPF_JMP | BPF_JA:
741 /* Note, the large ftest->k might cause loops.
742 * Compare this with conditional jumps below,
743 * where offsets are limited. --ANK (981016)
745 if (ftest->k >= (unsigned int)(flen - pc - 1))
748 case BPF_JMP | BPF_JEQ | BPF_K:
749 case BPF_JMP | BPF_JEQ | BPF_X:
750 case BPF_JMP | BPF_JGE | BPF_K:
751 case BPF_JMP | BPF_JGE | BPF_X:
752 case BPF_JMP | BPF_JGT | BPF_K:
753 case BPF_JMP | BPF_JGT | BPF_X:
754 case BPF_JMP | BPF_JSET | BPF_K:
755 case BPF_JMP | BPF_JSET | BPF_X:
756 /* Both conditionals must be safe */
757 if (pc + ftest->jt + 1 >= flen ||
758 pc + ftest->jf + 1 >= flen)
761 case BPF_LD | BPF_W | BPF_ABS:
762 case BPF_LD | BPF_H | BPF_ABS:
763 case BPF_LD | BPF_B | BPF_ABS:
765 if (bpf_anc_helper(ftest) & BPF_ANC)
767 /* Ancillary operation unknown or unsupported */
768 if (anc_found == false && ftest->k >= SKF_AD_OFF)
773 /* Last instruction must be a RET code */
774 switch (filter[flen - 1].code) {
775 case BPF_RET | BPF_K:
776 case BPF_RET | BPF_A:
777 return check_load_and_stores(filter, flen);
782 EXPORT_SYMBOL(bpf_check_classic);
784 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
785 const struct sock_fprog *fprog)
787 unsigned int fsize = bpf_classic_proglen(fprog);
788 struct sock_fprog_kern *fkprog;
790 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
794 fkprog = fp->orig_prog;
795 fkprog->len = fprog->len;
796 fkprog->filter = kmemdup(fp->insns, fsize, GFP_KERNEL);
797 if (!fkprog->filter) {
798 kfree(fp->orig_prog);
805 static void bpf_release_orig_filter(struct bpf_prog *fp)
807 struct sock_fprog_kern *fprog = fp->orig_prog;
810 kfree(fprog->filter);
815 static void __bpf_prog_release(struct bpf_prog *prog)
817 if (prog->aux->prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
820 bpf_release_orig_filter(prog);
825 static void __sk_filter_release(struct sk_filter *fp)
827 __bpf_prog_release(fp->prog);
832 * sk_filter_release_rcu - Release a socket filter by rcu_head
833 * @rcu: rcu_head that contains the sk_filter to free
835 static void sk_filter_release_rcu(struct rcu_head *rcu)
837 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
839 __sk_filter_release(fp);
843 * sk_filter_release - release a socket filter
844 * @fp: filter to remove
846 * Remove a filter from a socket and release its resources.
848 static void sk_filter_release(struct sk_filter *fp)
850 if (atomic_dec_and_test(&fp->refcnt))
851 call_rcu(&fp->rcu, sk_filter_release_rcu);
854 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
856 u32 filter_size = bpf_prog_size(fp->prog->len);
858 atomic_sub(filter_size, &sk->sk_omem_alloc);
859 sk_filter_release(fp);
862 /* try to charge the socket memory if there is space available
863 * return true on success
865 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
867 u32 filter_size = bpf_prog_size(fp->prog->len);
869 /* same check as in sock_kmalloc() */
870 if (filter_size <= sysctl_optmem_max &&
871 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
872 atomic_inc(&fp->refcnt);
873 atomic_add(filter_size, &sk->sk_omem_alloc);
879 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
881 struct sock_filter *old_prog;
882 struct bpf_prog *old_fp;
883 int err, new_len, old_len = fp->len;
885 /* We are free to overwrite insns et al right here as it
886 * won't be used at this point in time anymore internally
887 * after the migration to the internal BPF instruction
890 BUILD_BUG_ON(sizeof(struct sock_filter) !=
891 sizeof(struct bpf_insn));
893 /* Conversion cannot happen on overlapping memory areas,
894 * so we need to keep the user BPF around until the 2nd
895 * pass. At this time, the user BPF is stored in fp->insns.
897 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
904 /* 1st pass: calculate the new program length. */
905 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
909 /* Expand fp for appending the new filter representation. */
911 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
913 /* The old_fp is still around in case we couldn't
914 * allocate new memory, so uncharge on that one.
923 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
924 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
926 /* 2nd bpf_convert_filter() can fail only if it fails
927 * to allocate memory, remapping must succeed. Note,
928 * that at this time old_fp has already been released
933 bpf_prog_select_runtime(fp);
941 __bpf_prog_release(fp);
945 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp)
952 err = bpf_check_classic(fp->insns, fp->len);
954 __bpf_prog_release(fp);
958 /* Probe if we can JIT compile the filter and if so, do
959 * the compilation of the filter.
963 /* JIT compiler couldn't process this filter, so do the
964 * internal BPF translation for the optimized interpreter.
967 fp = bpf_migrate_filter(fp);
973 * bpf_prog_create - create an unattached filter
974 * @pfp: the unattached filter that is created
975 * @fprog: the filter program
977 * Create a filter independent of any socket. We first run some
978 * sanity checks on it to make sure it does not explode on us later.
979 * If an error occurs or there is insufficient memory for the filter
980 * a negative errno code is returned. On success the return is zero.
982 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
984 unsigned int fsize = bpf_classic_proglen(fprog);
987 /* Make sure new filter is there and in the right amounts. */
988 if (fprog->filter == NULL)
991 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
995 memcpy(fp->insns, fprog->filter, fsize);
997 fp->len = fprog->len;
998 /* Since unattached filters are not copied back to user
999 * space through sk_get_filter(), we do not need to hold
1000 * a copy here, and can spare us the work.
1002 fp->orig_prog = NULL;
1004 /* bpf_prepare_filter() already takes care of freeing
1005 * memory in case something goes wrong.
1007 fp = bpf_prepare_filter(fp);
1014 EXPORT_SYMBOL_GPL(bpf_prog_create);
1016 void bpf_prog_destroy(struct bpf_prog *fp)
1018 __bpf_prog_release(fp);
1020 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1023 * sk_attach_filter - attach a socket filter
1024 * @fprog: the filter program
1025 * @sk: the socket to use
1027 * Attach the user's filter code. We first run some sanity checks on
1028 * it to make sure it does not explode on us later. If an error
1029 * occurs or there is insufficient memory for the filter a negative
1030 * errno code is returned. On success the return is zero.
1032 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1034 struct sk_filter *fp, *old_fp;
1035 unsigned int fsize = bpf_classic_proglen(fprog);
1036 unsigned int bpf_fsize = bpf_prog_size(fprog->len);
1037 struct bpf_prog *prog;
1040 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1043 /* Make sure new filter is there and in the right amounts. */
1044 if (fprog->filter == NULL)
1047 prog = bpf_prog_alloc(bpf_fsize, 0);
1051 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1052 __bpf_prog_free(prog);
1056 prog->len = fprog->len;
1058 err = bpf_prog_store_orig_filter(prog, fprog);
1060 __bpf_prog_free(prog);
1064 /* bpf_prepare_filter() already takes care of freeing
1065 * memory in case something goes wrong.
1067 prog = bpf_prepare_filter(prog);
1069 return PTR_ERR(prog);
1071 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1073 __bpf_prog_release(prog);
1078 atomic_set(&fp->refcnt, 0);
1080 if (!sk_filter_charge(sk, fp)) {
1081 __sk_filter_release(fp);
1085 old_fp = rcu_dereference_protected(sk->sk_filter,
1086 sock_owned_by_user(sk));
1087 rcu_assign_pointer(sk->sk_filter, fp);
1090 sk_filter_uncharge(sk, old_fp);
1094 EXPORT_SYMBOL_GPL(sk_attach_filter);
1096 #ifdef CONFIG_BPF_SYSCALL
1097 int sk_attach_bpf(u32 ufd, struct sock *sk)
1099 struct sk_filter *fp, *old_fp;
1100 struct bpf_prog *prog;
1102 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1105 prog = bpf_prog_get(ufd);
1107 return PTR_ERR(prog);
1109 if (prog->aux->prog_type != BPF_PROG_TYPE_SOCKET_FILTER) {
1110 /* valid fd, but invalid program type */
1115 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1122 atomic_set(&fp->refcnt, 0);
1124 if (!sk_filter_charge(sk, fp)) {
1125 __sk_filter_release(fp);
1129 old_fp = rcu_dereference_protected(sk->sk_filter,
1130 sock_owned_by_user(sk));
1131 rcu_assign_pointer(sk->sk_filter, fp);
1134 sk_filter_uncharge(sk, old_fp);
1139 /* allow socket filters to call
1140 * bpf_map_lookup_elem(), bpf_map_update_elem(), bpf_map_delete_elem()
1142 static const struct bpf_func_proto *sock_filter_func_proto(enum bpf_func_id func_id)
1145 case BPF_FUNC_map_lookup_elem:
1146 return &bpf_map_lookup_elem_proto;
1147 case BPF_FUNC_map_update_elem:
1148 return &bpf_map_update_elem_proto;
1149 case BPF_FUNC_map_delete_elem:
1150 return &bpf_map_delete_elem_proto;
1156 static bool sock_filter_is_valid_access(int off, int size, enum bpf_access_type type)
1158 /* skb fields cannot be accessed yet */
1162 static struct bpf_verifier_ops sock_filter_ops = {
1163 .get_func_proto = sock_filter_func_proto,
1164 .is_valid_access = sock_filter_is_valid_access,
1167 static struct bpf_prog_type_list tl = {
1168 .ops = &sock_filter_ops,
1169 .type = BPF_PROG_TYPE_SOCKET_FILTER,
1172 static int __init register_sock_filter_ops(void)
1174 bpf_register_prog_type(&tl);
1177 late_initcall(register_sock_filter_ops);
1179 int sk_attach_bpf(u32 ufd, struct sock *sk)
1184 int sk_detach_filter(struct sock *sk)
1187 struct sk_filter *filter;
1189 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1192 filter = rcu_dereference_protected(sk->sk_filter,
1193 sock_owned_by_user(sk));
1195 RCU_INIT_POINTER(sk->sk_filter, NULL);
1196 sk_filter_uncharge(sk, filter);
1202 EXPORT_SYMBOL_GPL(sk_detach_filter);
1204 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
1207 struct sock_fprog_kern *fprog;
1208 struct sk_filter *filter;
1212 filter = rcu_dereference_protected(sk->sk_filter,
1213 sock_owned_by_user(sk));
1217 /* We're copying the filter that has been originally attached,
1218 * so no conversion/decode needed anymore.
1220 fprog = filter->prog->orig_prog;
1224 /* User space only enquires number of filter blocks. */
1228 if (len < fprog->len)
1232 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
1235 /* Instead of bytes, the API requests to return the number