?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'net-mvpp2-misc-improvements'
[linux-2.6/btrfs-unstable.git] / net / core / filter.c
blob60ed6f343a6332737208f46e87f019bce5eab02f
1 /*
2 * Linux Socket Filter - Kernel level socket filtering
3 *
4 * Based on the design of the Berkeley Packet Filter. The new
5 * internal format has been designed by PLUMgrid:
6 *
7 * Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
8 *
9 * Authors:
10 *
11 * Jay Schulist <jschlst@samba.org>
12 * Alexei Starovoitov <ast@plumgrid.com>
13 * Daniel Borkmann <dborkman@redhat.com>
14 *
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.
19 *
20 * Andi Kleen - Fix a few bad bugs and races.
21 * Kris Katterjohn - Added many additional checks in bpf_check_classic()
22 */
23
24 #include <linux/module.h>
25 #include <linux/types.h>
26 #include <linux/mm.h>
27 #include <linux/fcntl.h>
28 #include <linux/socket.h>
29 #include <linux/sock_diag.h>
30 #include <linux/in.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>
36 #include <net/ip.h>
37 #include <net/protocol.h>
38 #include <net/netlink.h>
39 #include <linux/skbuff.h>
40 #include <net/sock.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>
54 #include <net/dst.h>
55 #include <net/sock_reuseport.h>
56 #include <net/busy_poll.h>
57
58 /**
59 * sk_filter_trim_cap - run a packet through a socket filter
60 * @sk: sock associated with &sk_buff
61 * @skb: buffer to filter
62 * @cap: limit on how short the eBPF program may trim the packet
63 *
64 * Run the eBPF program and then cut skb->data to correct size returned by
65 * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
66 * than pkt_len we keep whole skb->data. This is the socket level
67 * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
68 * be accepted or -EPERM if the packet should be tossed.
69 *
70 */
71 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
72 {
73 int err;
74 struct sk_filter *filter;
75
76 /*
77 * If the skb was allocated from pfmemalloc reserves, only
78 * allow SOCK_MEMALLOC sockets to use it as this socket is
79 * helping free memory
80 */
81 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC)) {
82 NET_INC_STATS(sock_net(sk), LINUX_MIB_PFMEMALLOCDROP);
83 return -ENOMEM;
84 }
85 err = BPF_CGROUP_RUN_PROG_INET_INGRESS(sk, skb);
86 if (err)
87 return err;
88
89 err = security_sock_rcv_skb(sk, skb);
90 if (err)
91 return err;
92
93 rcu_read_lock();
94 filter = rcu_dereference(sk->sk_filter);
95 if (filter) {
96 struct sock *save_sk = skb->sk;
97 unsigned int pkt_len;
98
99 skb->sk = sk;
100 pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
101 skb->sk = save_sk;
102 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
103 }
104 rcu_read_unlock();
105
106 return err;
107 }
108 EXPORT_SYMBOL(sk_filter_trim_cap);
109
110 BPF_CALL_1(__skb_get_pay_offset, struct sk_buff *, skb)
111 {
112 return skb_get_poff(skb);
113 }
114
115 BPF_CALL_3(__skb_get_nlattr, struct sk_buff *, skb, u32, a, u32, x)
116 {
117 struct nlattr *nla;
118
119 if (skb_is_nonlinear(skb))
120 return 0;
121
122 if (skb->len < sizeof(struct nlattr))
123 return 0;
124
125 if (a > skb->len - sizeof(struct nlattr))
126 return 0;
127
128 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
129 if (nla)
130 return (void *) nla - (void *) skb->data;
131
132 return 0;
133 }
134
135 BPF_CALL_3(__skb_get_nlattr_nest, struct sk_buff *, skb, u32, a, u32, x)
136 {
137 struct nlattr *nla;
138
139 if (skb_is_nonlinear(skb))
140 return 0;
141
142 if (skb->len < sizeof(struct nlattr))
143 return 0;
144
145 if (a > skb->len - sizeof(struct nlattr))
146 return 0;
147
148 nla = (struct nlattr *) &skb->data[a];
149 if (nla->nla_len > skb->len - a)
150 return 0;
151
152 nla = nla_find_nested(nla, x);
153 if (nla)
154 return (void *) nla - (void *) skb->data;
155
156 return 0;
157 }
158
159 BPF_CALL_0(__get_raw_cpu_id)
160 {
161 return raw_smp_processor_id();
162 }
163
164 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
165 .func = __get_raw_cpu_id,
166 .gpl_only = false,
167 .ret_type = RET_INTEGER,
168 };
169
170 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
171 struct bpf_insn *insn_buf)
172 {
173 struct bpf_insn *insn = insn_buf;
174
175 switch (skb_field) {
176 case SKF_AD_MARK:
177 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
178
179 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
180 offsetof(struct sk_buff, mark));
181 break;
182
183 case SKF_AD_PKTTYPE:
184 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
185 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
186 #ifdef __BIG_ENDIAN_BITFIELD
187 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
188 #endif
189 break;
190
191 case SKF_AD_QUEUE:
192 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
193
194 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
195 offsetof(struct sk_buff, queue_mapping));
196 break;
197
198 case SKF_AD_VLAN_TAG:
199 case SKF_AD_VLAN_TAG_PRESENT:
200 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
201 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
202
203 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
204 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
205 offsetof(struct sk_buff, vlan_tci));
206 if (skb_field == SKF_AD_VLAN_TAG) {
207 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
208 ~VLAN_TAG_PRESENT);
209 } else {
210 /* dst_reg >>= 12 */
211 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
212 /* dst_reg &= 1 */
213 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
214 }
215 break;
216 }
217
218 return insn - insn_buf;
219 }
220
221 static bool convert_bpf_extensions(struct sock_filter *fp,
222 struct bpf_insn **insnp)
223 {
224 struct bpf_insn *insn = *insnp;
225 u32 cnt;
226
227 switch (fp->k) {
228 case SKF_AD_OFF + SKF_AD_PROTOCOL:
229 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
230
231 /* A = *(u16 *) (CTX + offsetof(protocol)) */
232 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
233 offsetof(struct sk_buff, protocol));
234 /* A = ntohs(A) [emitting a nop or swap16] */
235 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
236 break;
237
238 case SKF_AD_OFF + SKF_AD_PKTTYPE:
239 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
240 insn += cnt - 1;
241 break;
242
243 case SKF_AD_OFF + SKF_AD_IFINDEX:
244 case SKF_AD_OFF + SKF_AD_HATYPE:
245 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
246 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
247
248 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
249 BPF_REG_TMP, BPF_REG_CTX,
250 offsetof(struct sk_buff, dev));
251 /* if (tmp != 0) goto pc + 1 */
252 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
253 *insn++ = BPF_EXIT_INSN();
254 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
255 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
256 offsetof(struct net_device, ifindex));
257 else
258 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
259 offsetof(struct net_device, type));
260 break;
261
262 case SKF_AD_OFF + SKF_AD_MARK:
263 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
264 insn += cnt - 1;
265 break;
266
267 case SKF_AD_OFF + SKF_AD_RXHASH:
268 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
269
270 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
271 offsetof(struct sk_buff, hash));
272 break;
273
274 case SKF_AD_OFF + SKF_AD_QUEUE:
275 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
276 insn += cnt - 1;
277 break;
278
279 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
280 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
281 BPF_REG_A, BPF_REG_CTX, insn);
282 insn += cnt - 1;
283 break;
284
285 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
286 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
287 BPF_REG_A, BPF_REG_CTX, insn);
288 insn += cnt - 1;
289 break;
290
291 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
292 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
293
294 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
295 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
296 offsetof(struct sk_buff, vlan_proto));
297 /* A = ntohs(A) [emitting a nop or swap16] */
298 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
299 break;
300
301 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
302 case SKF_AD_OFF + SKF_AD_NLATTR:
303 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
304 case SKF_AD_OFF + SKF_AD_CPU:
305 case SKF_AD_OFF + SKF_AD_RANDOM:
306 /* arg1 = CTX */
307 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
308 /* arg2 = A */
309 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
310 /* arg3 = X */
311 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
312 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
313 switch (fp->k) {
314 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
315 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
316 break;
317 case SKF_AD_OFF + SKF_AD_NLATTR:
318 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
319 break;
320 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
321 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
322 break;
323 case SKF_AD_OFF + SKF_AD_CPU:
324 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
325 break;
326 case SKF_AD_OFF + SKF_AD_RANDOM:
327 *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
328 bpf_user_rnd_init_once();
329 break;
330 }
331 break;
332
333 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
334 /* A ^= X */
335 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
336 break;
337
338 default:
339 /* This is just a dummy call to avoid letting the compiler
340 * evict __bpf_call_base() as an optimization. Placed here
341 * where no-one bothers.
342 */
343 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
344 return false;
345 }
346
347 *insnp = insn;
348 return true;
349 }
350
351 /**
352 * bpf_convert_filter - convert filter program
353 * @prog: the user passed filter program
354 * @len: the length of the user passed filter program
355 * @new_prog: allocated 'struct bpf_prog' or NULL
356 * @new_len: pointer to store length of converted program
357 *
358 * Remap 'sock_filter' style classic BPF (cBPF) instruction set to 'bpf_insn'
359 * style extended BPF (eBPF).
360 * Conversion workflow:
361 *
362 * 1) First pass for calculating the new program length:
363 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
364 *
365 * 2) 2nd pass to remap in two passes: 1st pass finds new
366 * jump offsets, 2nd pass remapping:
367 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
368 */
369 static int bpf_convert_filter(struct sock_filter *prog, int len,
370 struct bpf_prog *new_prog, int *new_len)
371 {
372 int new_flen = 0, pass = 0, target, i, stack_off;
373 struct bpf_insn *new_insn, *first_insn = NULL;
374 struct sock_filter *fp;
375 int *addrs = NULL;
376 u8 bpf_src;
377
378 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
379 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
380
381 if (len <= 0 || len > BPF_MAXINSNS)
382 return -EINVAL;
383
384 if (new_prog) {
385 first_insn = new_prog->insnsi;
386 addrs = kcalloc(len, sizeof(*addrs),
387 GFP_KERNEL | __GFP_NOWARN);
388 if (!addrs)
389 return -ENOMEM;
390 }
391
392 do_pass:
393 new_insn = first_insn;
394 fp = prog;
395
396 /* Classic BPF related prologue emission. */
397 if (new_prog) {
398 /* Classic BPF expects A and X to be reset first. These need
399 * to be guaranteed to be the first two instructions.
400 */
401 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
402 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
403
404 /* All programs must keep CTX in callee saved BPF_REG_CTX.
405 * In eBPF case it's done by the compiler, here we need to
406 * do this ourself. Initial CTX is present in BPF_REG_ARG1.
407 */
408 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
409 } else {
410 new_insn += 3;
411 }
412
413 for (i = 0; i < len; fp++, i++) {
414 struct bpf_insn tmp_insns[6] = { };
415 struct bpf_insn *insn = tmp_insns;
416
417 if (addrs)
418 addrs[i] = new_insn - first_insn;
419
420 switch (fp->code) {
421 /* All arithmetic insns and skb loads map as-is. */
422 case BPF_ALU | BPF_ADD | BPF_X:
423 case BPF_ALU | BPF_ADD | BPF_K:
424 case BPF_ALU | BPF_SUB | BPF_X:
425 case BPF_ALU | BPF_SUB | BPF_K:
426 case BPF_ALU | BPF_AND | BPF_X:
427 case BPF_ALU | BPF_AND | BPF_K:
428 case BPF_ALU | BPF_OR | BPF_X:
429 case BPF_ALU | BPF_OR | BPF_K:
430 case BPF_ALU | BPF_LSH | BPF_X:
431 case BPF_ALU | BPF_LSH | BPF_K:
432 case BPF_ALU | BPF_RSH | BPF_X:
433 case BPF_ALU | BPF_RSH | BPF_K:
434 case BPF_ALU | BPF_XOR | BPF_X:
435 case BPF_ALU | BPF_XOR | BPF_K:
436 case BPF_ALU | BPF_MUL | BPF_X:
437 case BPF_ALU | BPF_MUL | BPF_K:
438 case BPF_ALU | BPF_DIV | BPF_X:
439 case BPF_ALU | BPF_DIV | BPF_K:
440 case BPF_ALU | BPF_MOD | BPF_X:
441 case BPF_ALU | BPF_MOD | BPF_K:
442 case BPF_ALU | BPF_NEG:
443 case BPF_LD | BPF_ABS | BPF_W:
444 case BPF_LD | BPF_ABS | BPF_H:
445 case BPF_LD | BPF_ABS | BPF_B:
446 case BPF_LD | BPF_IND | BPF_W:
447 case BPF_LD | BPF_IND | BPF_H:
448 case BPF_LD | BPF_IND | BPF_B:
449 /* Check for overloaded BPF extension and
450 * directly convert it if found, otherwise
451 * just move on with mapping.
452 */
453 if (BPF_CLASS(fp->code) == BPF_LD &&
454 BPF_MODE(fp->code) == BPF_ABS &&
455 convert_bpf_extensions(fp, &insn))
456 break;
457
458 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
459 break;
460
461 /* Jump transformation cannot use BPF block macros
462 * everywhere as offset calculation and target updates
463 * require a bit more work than the rest, i.e. jump
464 * opcodes map as-is, but offsets need adjustment.
465 */
466
467 #define BPF_EMIT_JMP \
468 do { \
469 if (target >= len || target < 0) \
470 goto err; \
471 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
472 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
473 insn->off -= insn - tmp_insns; \
474 } while (0)
475
476 case BPF_JMP | BPF_JA:
477 target = i + fp->k + 1;
478 insn->code = fp->code;
479 BPF_EMIT_JMP;
480 break;
481
482 case BPF_JMP | BPF_JEQ | BPF_K:
483 case BPF_JMP | BPF_JEQ | BPF_X:
484 case BPF_JMP | BPF_JSET | BPF_K:
485 case BPF_JMP | BPF_JSET | BPF_X:
486 case BPF_JMP | BPF_JGT | BPF_K:
487 case BPF_JMP | BPF_JGT | BPF_X:
488 case BPF_JMP | BPF_JGE | BPF_K:
489 case BPF_JMP | BPF_JGE | BPF_X:
490 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
491 /* BPF immediates are signed, zero extend
492 * immediate into tmp register and use it
493 * in compare insn.
494 */
495 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
496
497 insn->dst_reg = BPF_REG_A;
498 insn->src_reg = BPF_REG_TMP;
499 bpf_src = BPF_X;
500 } else {
501 insn->dst_reg = BPF_REG_A;
502 insn->imm = fp->k;
503 bpf_src = BPF_SRC(fp->code);
504 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
505 }
506
507 /* Common case where 'jump_false' is next insn. */
508 if (fp->jf == 0) {
509 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
510 target = i + fp->jt + 1;
511 BPF_EMIT_JMP;
512 break;
513 }
514
515 /* Convert JEQ into JNE when 'jump_true' is next insn. */
516 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
517 insn->code = BPF_JMP | BPF_JNE | bpf_src;
518 target = i + fp->jf + 1;
519 BPF_EMIT_JMP;
520 break;
521 }
522
523 /* Other jumps are mapped into two insns: Jxx and JA. */
524 target = i + fp->jt + 1;
525 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
526 BPF_EMIT_JMP;
527 insn++;
528
529 insn->code = BPF_JMP | BPF_JA;
530 target = i + fp->jf + 1;
531 BPF_EMIT_JMP;
532 break;
533
534 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
535 case BPF_LDX | BPF_MSH | BPF_B:
536 /* tmp = A */
537 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
538 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
539 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
540 /* A &= 0xf */
541 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
542 /* A <<= 2 */
543 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
544 /* X = A */
545 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
546 /* A = tmp */
547 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
548 break;
549
550 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
551 * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
552 */
553 case BPF_RET | BPF_A:
554 case BPF_RET | BPF_K:
555 if (BPF_RVAL(fp->code) == BPF_K)
556 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
557 0, fp->k);
558 *insn = BPF_EXIT_INSN();
559 break;
560
561 /* Store to stack. */
562 case BPF_ST:
563 case BPF_STX:
564 stack_off = fp->k * 4 + 4;
565 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
566 BPF_ST ? BPF_REG_A : BPF_REG_X,
567 -stack_off);
568 /* check_load_and_stores() verifies that classic BPF can
569 * load from stack only after write, so tracking
570 * stack_depth for ST|STX insns is enough
571 */
572 if (new_prog && new_prog->aux->stack_depth < stack_off)
573 new_prog->aux->stack_depth = stack_off;
574 break;
575
576 /* Load from stack. */
577 case BPF_LD | BPF_MEM:
578 case BPF_LDX | BPF_MEM:
579 stack_off = fp->k * 4 + 4;
580 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
581 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
582 -stack_off);
583 break;
584
585 /* A = K or X = K */
586 case BPF_LD | BPF_IMM:
587 case BPF_LDX | BPF_IMM:
588 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
589 BPF_REG_A : BPF_REG_X, fp->k);
590 break;
591
592 /* X = A */
593 case BPF_MISC | BPF_TAX:
594 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
595 break;
596
597 /* A = X */
598 case BPF_MISC | BPF_TXA:
599 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
600 break;
601
602 /* A = skb->len or X = skb->len */
603 case BPF_LD | BPF_W | BPF_LEN:
604 case BPF_LDX | BPF_W | BPF_LEN:
605 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
606 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
607 offsetof(struct sk_buff, len));
608 break;
609
610 /* Access seccomp_data fields. */
611 case BPF_LDX | BPF_ABS | BPF_W:
612 /* A = *(u32 *) (ctx + K) */
613 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
614 break;
615
616 /* Unknown instruction. */
617 default:
618 goto err;
619 }
620
621 insn++;
622 if (new_prog)
623 memcpy(new_insn, tmp_insns,
624 sizeof(*insn) * (insn - tmp_insns));
625 new_insn += insn - tmp_insns;
626 }
627
628 if (!new_prog) {
629 /* Only calculating new length. */
630 *new_len = new_insn - first_insn;
631 return 0;
632 }
633
634 pass++;
635 if (new_flen != new_insn - first_insn) {
636 new_flen = new_insn - first_insn;
637 if (pass > 2)
638 goto err;
639 goto do_pass;
640 }
641
642 kfree(addrs);
643 BUG_ON(*new_len != new_flen);
644 return 0;
645 err:
646 kfree(addrs);
647 return -EINVAL;
648 }
649
650 /* Security:
651 *
652 * As we dont want to clear mem[] array for each packet going through
653 * __bpf_prog_run(), we check that filter loaded by user never try to read
654 * a cell if not previously written, and we check all branches to be sure
655 * a malicious user doesn't try to abuse us.
656 */
657 static int check_load_and_stores(const struct sock_filter *filter, int flen)
658 {
659 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
660 int pc, ret = 0;
661
662 BUILD_BUG_ON(BPF_MEMWORDS > 16);
663
664 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
665 if (!masks)
666 return -ENOMEM;
667
668 memset(masks, 0xff, flen * sizeof(*masks));
669
670 for (pc = 0; pc < flen; pc++) {
671 memvalid &= masks[pc];
672
673 switch (filter[pc].code) {
674 case BPF_ST:
675 case BPF_STX:
676 memvalid |= (1 << filter[pc].k);
677 break;
678 case BPF_LD | BPF_MEM:
679 case BPF_LDX | BPF_MEM:
680 if (!(memvalid & (1 << filter[pc].k))) {
681 ret = -EINVAL;
682 goto error;
683 }
684 break;
685 case BPF_JMP | BPF_JA:
686 /* A jump must set masks on target */
687 masks[pc + 1 + filter[pc].k] &= memvalid;
688 memvalid = ~0;
689 break;
690 case BPF_JMP | BPF_JEQ | BPF_K:
691 case BPF_JMP | BPF_JEQ | BPF_X:
692 case BPF_JMP | BPF_JGE | BPF_K:
693 case BPF_JMP | BPF_JGE | BPF_X:
694 case BPF_JMP | BPF_JGT | BPF_K:
695 case BPF_JMP | BPF_JGT | BPF_X:
696 case BPF_JMP | BPF_JSET | BPF_K:
697 case BPF_JMP | BPF_JSET | BPF_X:
698 /* A jump must set masks on targets */
699 masks[pc + 1 + filter[pc].jt] &= memvalid;
700 masks[pc + 1 + filter[pc].jf] &= memvalid;
701 memvalid = ~0;
702 break;
703 }
704 }
705 error:
706 kfree(masks);
707 return ret;
708 }
709
710 static bool chk_code_allowed(u16 code_to_probe)
711 {
712 static const bool codes[] = {
713 /* 32 bit ALU operations */
714 [BPF_ALU | BPF_ADD | BPF_K] = true,
715 [BPF_ALU | BPF_ADD | BPF_X] = true,
716 [BPF_ALU | BPF_SUB | BPF_K] = true,
717 [BPF_ALU | BPF_SUB | BPF_X] = true,
718 [BPF_ALU | BPF_MUL | BPF_K] = true,
719 [BPF_ALU | BPF_MUL | BPF_X] = true,
720 [BPF_ALU | BPF_DIV | BPF_K] = true,
721 [BPF_ALU | BPF_DIV | BPF_X] = true,
722 [BPF_ALU | BPF_MOD | BPF_K] = true,
723 [BPF_ALU | BPF_MOD | BPF_X] = true,
724 [BPF_ALU | BPF_AND | BPF_K] = true,
725 [BPF_ALU | BPF_AND | BPF_X] = true,
726 [BPF_ALU | BPF_OR | BPF_K] = true,
727 [BPF_ALU | BPF_OR | BPF_X] = true,
728 [BPF_ALU | BPF_XOR | BPF_K] = true,
729 [BPF_ALU | BPF_XOR | BPF_X] = true,
730 [BPF_ALU | BPF_LSH | BPF_K] = true,
731 [BPF_ALU | BPF_LSH | BPF_X] = true,
732 [BPF_ALU | BPF_RSH | BPF_K] = true,
733 [BPF_ALU | BPF_RSH | BPF_X] = true,
734 [BPF_ALU | BPF_NEG] = true,
735 /* Load instructions */
736 [BPF_LD | BPF_W | BPF_ABS] = true,
737 [BPF_LD | BPF_H | BPF_ABS] = true,
738 [BPF_LD | BPF_B | BPF_ABS] = true,
739 [BPF_LD | BPF_W | BPF_LEN] = true,
740 [BPF_LD | BPF_W | BPF_IND] = true,
741 [BPF_LD | BPF_H | BPF_IND] = true,
742 [BPF_LD | BPF_B | BPF_IND] = true,
743 [BPF_LD | BPF_IMM] = true,
744 [BPF_LD | BPF_MEM] = true,
745 [BPF_LDX | BPF_W | BPF_LEN] = true,
746 [BPF_LDX | BPF_B | BPF_MSH] = true,
747 [BPF_LDX | BPF_IMM] = true,
748 [BPF_LDX | BPF_MEM] = true,
749 /* Store instructions */
750 [BPF_ST] = true,
751 [BPF_STX] = true,
752 /* Misc instructions */
753 [BPF_MISC | BPF_TAX] = true,
754 [BPF_MISC | BPF_TXA] = true,
755 /* Return instructions */
756 [BPF_RET | BPF_K] = true,
757 [BPF_RET | BPF_A] = true,
758 /* Jump instructions */
759 [BPF_JMP | BPF_JA] = true,
760 [BPF_JMP | BPF_JEQ | BPF_K] = true,
761 [BPF_JMP | BPF_JEQ | BPF_X] = true,
762 [BPF_JMP | BPF_JGE | BPF_K] = true,
763 [BPF_JMP | BPF_JGE | BPF_X] = true,
764 [BPF_JMP | BPF_JGT | BPF_K] = true,
765 [BPF_JMP | BPF_JGT | BPF_X] = true,
766 [BPF_JMP | BPF_JSET | BPF_K] = true,
767 [BPF_JMP | BPF_JSET | BPF_X] = true,
768 };
769
770 if (code_to_probe >= ARRAY_SIZE(codes))
771 return false;
772
773 return codes[code_to_probe];
774 }
775
776 static bool bpf_check_basics_ok(const struct sock_filter *filter,
777 unsigned int flen)
778 {
779 if (filter == NULL)
780 return false;
781 if (flen == 0 || flen > BPF_MAXINSNS)
782 return false;
783
784 return true;
785 }
786
787 /**
788 * bpf_check_classic - verify socket filter code
789 * @filter: filter to verify
790 * @flen: length of filter
791 *
792 * Check the user's filter code. If we let some ugly
793 * filter code slip through kaboom! The filter must contain
794 * no references or jumps that are out of range, no illegal
795 * instructions, and must end with a RET instruction.
796 *
797 * All jumps are forward as they are not signed.
798 *
799 * Returns 0 if the rule set is legal or -EINVAL if not.
800 */
801 static int bpf_check_classic(const struct sock_filter *filter,
802 unsigned int flen)
803 {
804 bool anc_found;
805 int pc;
806
807 /* Check the filter code now */
808 for (pc = 0; pc < flen; pc++) {
809 const struct sock_filter *ftest = &filter[pc];
810
811 /* May we actually operate on this code? */
812 if (!chk_code_allowed(ftest->code))
813 return -EINVAL;
814
815 /* Some instructions need special checks */
816 switch (ftest->code) {
817 case BPF_ALU | BPF_DIV | BPF_K:
818 case BPF_ALU | BPF_MOD | BPF_K:
819 /* Check for division by zero */
820 if (ftest->k == 0)
821 return -EINVAL;
822 break;
823 case BPF_ALU | BPF_LSH | BPF_K:
824 case BPF_ALU | BPF_RSH | BPF_K:
825 if (ftest->k >= 32)
826 return -EINVAL;
827 break;
828 case BPF_LD | BPF_MEM:
829 case BPF_LDX | BPF_MEM:
830 case BPF_ST:
831 case BPF_STX:
832 /* Check for invalid memory addresses */
833 if (ftest->k >= BPF_MEMWORDS)
834 return -EINVAL;
835 break;
836 case BPF_JMP | BPF_JA:
837 /* Note, the large ftest->k might cause loops.
838 * Compare this with conditional jumps below,
839 * where offsets are limited. --ANK (981016)
840 */
841 if (ftest->k >= (unsigned int)(flen - pc - 1))
842 return -EINVAL;
843 break;
844 case BPF_JMP | BPF_JEQ | BPF_K:
845 case BPF_JMP | BPF_JEQ | BPF_X:
846 case BPF_JMP | BPF_JGE | BPF_K:
847 case BPF_JMP | BPF_JGE | BPF_X:
848 case BPF_JMP | BPF_JGT | BPF_K:
849 case BPF_JMP | BPF_JGT | BPF_X:
850 case BPF_JMP | BPF_JSET | BPF_K:
851 case BPF_JMP | BPF_JSET | BPF_X:
852 /* Both conditionals must be safe */
853 if (pc + ftest->jt + 1 >= flen ||
854 pc + ftest->jf + 1 >= flen)
855 return -EINVAL;
856 break;
857 case BPF_LD | BPF_W | BPF_ABS:
858 case BPF_LD | BPF_H | BPF_ABS:
859 case BPF_LD | BPF_B | BPF_ABS:
860 anc_found = false;
861 if (bpf_anc_helper(ftest) & BPF_ANC)
862 anc_found = true;
863 /* Ancillary operation unknown or unsupported */
864 if (anc_found == false && ftest->k >= SKF_AD_OFF)
865 return -EINVAL;
866 }
867 }
868
869 /* Last instruction must be a RET code */
870 switch (filter[flen - 1].code) {
871 case BPF_RET | BPF_K:
872 case BPF_RET | BPF_A:
873 return check_load_and_stores(filter, flen);
874 }
875
876 return -EINVAL;
877 }
878
879 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
880 const struct sock_fprog *fprog)
881 {
882 unsigned int fsize = bpf_classic_proglen(fprog);
883 struct sock_fprog_kern *fkprog;
884
885 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
886 if (!fp->orig_prog)
887 return -ENOMEM;
888
889 fkprog = fp->orig_prog;
890 fkprog->len = fprog->len;
891
892 fkprog->filter = kmemdup(fp->insns, fsize,
893 GFP_KERNEL | __GFP_NOWARN);
894 if (!fkprog->filter) {
895 kfree(fp->orig_prog);
896 return -ENOMEM;
897 }
898
899 return 0;
900 }
901
902 static void bpf_release_orig_filter(struct bpf_prog *fp)
903 {
904 struct sock_fprog_kern *fprog = fp->orig_prog;
905
906 if (fprog) {
907 kfree(fprog->filter);
908 kfree(fprog);
909 }
910 }
911
912 static void __bpf_prog_release(struct bpf_prog *prog)
913 {
914 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
915 bpf_prog_put(prog);
916 } else {
917 bpf_release_orig_filter(prog);
918 bpf_prog_free(prog);
919 }
920 }
921
922 static void __sk_filter_release(struct sk_filter *fp)
923 {
924 __bpf_prog_release(fp->prog);
925 kfree(fp);
926 }
927
928 /**
929 * sk_filter_release_rcu - Release a socket filter by rcu_head
930 * @rcu: rcu_head that contains the sk_filter to free
931 */
932 static void sk_filter_release_rcu(struct rcu_head *rcu)
933 {
934 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
935
936 __sk_filter_release(fp);
937 }
938
939 /**
940 * sk_filter_release - release a socket filter
941 * @fp: filter to remove
942 *
943 * Remove a filter from a socket and release its resources.
944 */
945 static void sk_filter_release(struct sk_filter *fp)
946 {
947 if (refcount_dec_and_test(&fp->refcnt))
948 call_rcu(&fp->rcu, sk_filter_release_rcu);
949 }
950
951 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
952 {
953 u32 filter_size = bpf_prog_size(fp->prog->len);
954
955 atomic_sub(filter_size, &sk->sk_omem_alloc);
956 sk_filter_release(fp);
957 }
958
959 /* try to charge the socket memory if there is space available
960 * return true on success
961 */
962 static bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
963 {
964 u32 filter_size = bpf_prog_size(fp->prog->len);
965
966 /* same check as in sock_kmalloc() */
967 if (filter_size <= sysctl_optmem_max &&
968 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
969 atomic_add(filter_size, &sk->sk_omem_alloc);
970 return true;
971 }
972 return false;
973 }
974
975 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
976 {
977 bool ret = __sk_filter_charge(sk, fp);
978 if (ret)
979 refcount_inc(&fp->refcnt);
980 return ret;
981 }
982
983 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
984 {
985 struct sock_filter *old_prog;
986 struct bpf_prog *old_fp;
987 int err, new_len, old_len = fp->len;
988
989 /* We are free to overwrite insns et al right here as it
990 * won't be used at this point in time anymore internally
991 * after the migration to the internal BPF instruction
992 * representation.
993 */
994 BUILD_BUG_ON(sizeof(struct sock_filter) !=
995 sizeof(struct bpf_insn));
996
997 /* Conversion cannot happen on overlapping memory areas,
998 * so we need to keep the user BPF around until the 2nd
999 * pass. At this time, the user BPF is stored in fp->insns.
1000 */
1001 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
1002 GFP_KERNEL | __GFP_NOWARN);
1003 if (!old_prog) {
1004 err = -ENOMEM;
1005 goto out_err;
1006 }
1007
1008 /* 1st pass: calculate the new program length. */
1009 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
1010 if (err)
1011 goto out_err_free;
1012
1013 /* Expand fp for appending the new filter representation. */
1014 old_fp = fp;
1015 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
1016 if (!fp) {
1017 /* The old_fp is still around in case we couldn't
1018 * allocate new memory, so uncharge on that one.
1019 */
1020 fp = old_fp;
1021 err = -ENOMEM;
1022 goto out_err_free;
1023 }
1024
1025 fp->len = new_len;
1026
1027 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1028 err = bpf_convert_filter(old_prog, old_len, fp, &new_len);
1029 if (err)
1030 /* 2nd bpf_convert_filter() can fail only if it fails
1031 * to allocate memory, remapping must succeed. Note,
1032 * that at this time old_fp has already been released
1033 * by krealloc().
1034 */
1035 goto out_err_free;
1036
1037 /* We are guaranteed to never error here with cBPF to eBPF
1038 * transitions, since there's no issue with type compatibility
1039 * checks on program arrays.
1040 */
1041 fp = bpf_prog_select_runtime(fp, &err);
1042
1043 kfree(old_prog);
1044 return fp;
1045
1046 out_err_free:
1047 kfree(old_prog);
1048 out_err:
1049 __bpf_prog_release(fp);
1050 return ERR_PTR(err);
1051 }
1052
1053 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1054 bpf_aux_classic_check_t trans)
1055 {
1056 int err;
1057
1058 fp->bpf_func = NULL;
1059 fp->jited = 0;
1060
1061 err = bpf_check_classic(fp->insns, fp->len);
1062 if (err) {
1063 __bpf_prog_release(fp);
1064 return ERR_PTR(err);
1065 }
1066
1067 /* There might be additional checks and transformations
1068 * needed on classic filters, f.e. in case of seccomp.
1069 */
1070 if (trans) {
1071 err = trans(fp->insns, fp->len);
1072 if (err) {
1073 __bpf_prog_release(fp);
1074 return ERR_PTR(err);
1075 }
1076 }
1077
1078 /* Probe if we can JIT compile the filter and if so, do
1079 * the compilation of the filter.
1080 */
1081 bpf_jit_compile(fp);
1082
1083 /* JIT compiler couldn't process this filter, so do the
1084 * internal BPF translation for the optimized interpreter.
1085 */
1086 if (!fp->jited)
1087 fp = bpf_migrate_filter(fp);
1088
1089 return fp;
1090 }
1091
1092 /**
1093 * bpf_prog_create - create an unattached filter
1094 * @pfp: the unattached filter that is created
1095 * @fprog: the filter program
1096 *
1097 * Create a filter independent of any socket. We first run some
1098 * sanity checks on it to make sure it does not explode on us later.
1099 * If an error occurs or there is insufficient memory for the filter
1100 * a negative errno code is returned. On success the return is zero.
1101 */
1102 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1103 {
1104 unsigned int fsize = bpf_classic_proglen(fprog);
1105 struct bpf_prog *fp;
1106
1107 /* Make sure new filter is there and in the right amounts. */
1108 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1109 return -EINVAL;
1110
1111 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1112 if (!fp)
1113 return -ENOMEM;
1114
1115 memcpy(fp->insns, fprog->filter, fsize);
1116
1117 fp->len = fprog->len;
1118 /* Since unattached filters are not copied back to user
1119 * space through sk_get_filter(), we do not need to hold
1120 * a copy here, and can spare us the work.
1121 */
1122 fp->orig_prog = NULL;
1123
1124 /* bpf_prepare_filter() already takes care of freeing
1125 * memory in case something goes wrong.
1126 */
1127 fp = bpf_prepare_filter(fp, NULL);
1128 if (IS_ERR(fp))
1129 return PTR_ERR(fp);
1130
1131 *pfp = fp;
1132 return 0;
1133 }
1134 EXPORT_SYMBOL_GPL(bpf_prog_create);
1135
1136 /**
1137 * bpf_prog_create_from_user - create an unattached filter from user buffer
1138 * @pfp: the unattached filter that is created
1139 * @fprog: the filter program
1140 * @trans: post-classic verifier transformation handler
1141 * @save_orig: save classic BPF program
1142 *
1143 * This function effectively does the same as bpf_prog_create(), only
1144 * that it builds up its insns buffer from user space provided buffer.
1145 * It also allows for passing a bpf_aux_classic_check_t handler.
1146 */
1147 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1148 bpf_aux_classic_check_t trans, bool save_orig)
1149 {
1150 unsigned int fsize = bpf_classic_proglen(fprog);
1151 struct bpf_prog *fp;
1152 int err;
1153
1154 /* Make sure new filter is there and in the right amounts. */
1155 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1156 return -EINVAL;
1157
1158 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1159 if (!fp)
1160 return -ENOMEM;
1161
1162 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1163 __bpf_prog_free(fp);
1164 return -EFAULT;
1165 }
1166
1167 fp->len = fprog->len;
1168 fp->orig_prog = NULL;
1169
1170 if (save_orig) {
1171 err = bpf_prog_store_orig_filter(fp, fprog);
1172 if (err) {
1173 __bpf_prog_free(fp);
1174 return -ENOMEM;
1175 }
1176 }
1177
1178 /* bpf_prepare_filter() already takes care of freeing
1179 * memory in case something goes wrong.
1180 */
1181 fp = bpf_prepare_filter(fp, trans);
1182 if (IS_ERR(fp))
1183 return PTR_ERR(fp);
1184
1185 *pfp = fp;
1186 return 0;
1187 }
1188 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1189
1190 void bpf_prog_destroy(struct bpf_prog *fp)
1191 {
1192 __bpf_prog_release(fp);
1193 }
1194 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1195
1196 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1197 {
1198 struct sk_filter *fp, *old_fp;
1199
1200 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1201 if (!fp)
1202 return -ENOMEM;
1203
1204 fp->prog = prog;
1205
1206 if (!__sk_filter_charge(sk, fp)) {
1207 kfree(fp);
1208 return -ENOMEM;
1209 }
1210 refcount_set(&fp->refcnt, 1);
1211
1212 old_fp = rcu_dereference_protected(sk->sk_filter,
1213 lockdep_sock_is_held(sk));
1214 rcu_assign_pointer(sk->sk_filter, fp);
1215
1216 if (old_fp)
1217 sk_filter_uncharge(sk, old_fp);
1218
1219 return 0;
1220 }
1221
1222 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1223 {
1224 struct bpf_prog *old_prog;
1225 int err;
1226
1227 if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1228 return -ENOMEM;
1229
1230 if (sk_unhashed(sk) && sk->sk_reuseport) {
1231 err = reuseport_alloc(sk);
1232 if (err)
1233 return err;
1234 } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1235 /* The socket wasn't bound with SO_REUSEPORT */
1236 return -EINVAL;
1237 }
1238
1239 old_prog = reuseport_attach_prog(sk, prog);
1240 if (old_prog)
1241 bpf_prog_destroy(old_prog);
1242
1243 return 0;
1244 }
1245
1246 static
1247 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1248 {
1249 unsigned int fsize = bpf_classic_proglen(fprog);
1250 struct bpf_prog *prog;
1251 int err;
1252
1253 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1254 return ERR_PTR(-EPERM);
1255
1256 /* Make sure new filter is there and in the right amounts. */
1257 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1258 return ERR_PTR(-EINVAL);
1259
1260 prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1261 if (!prog)
1262 return ERR_PTR(-ENOMEM);
1263
1264 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1265 __bpf_prog_free(prog);
1266 return ERR_PTR(-EFAULT);
1267 }
1268
1269 prog->len = fprog->len;
1270
1271 err = bpf_prog_store_orig_filter(prog, fprog);
1272 if (err) {
1273 __bpf_prog_free(prog);
1274 return ERR_PTR(-ENOMEM);
1275 }
1276
1277 /* bpf_prepare_filter() already takes care of freeing
1278 * memory in case something goes wrong.
1279 */
1280 return bpf_prepare_filter(prog, NULL);
1281 }
1282
1283 /**
1284 * sk_attach_filter - attach a socket filter
1285 * @fprog: the filter program
1286 * @sk: the socket to use
1287 *
1288 * Attach the user's filter code. We first run some sanity checks on
1289 * it to make sure it does not explode on us later. If an error
1290 * occurs or there is insufficient memory for the filter a negative
1291 * errno code is returned. On success the return is zero.
1292 */
1293 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1294 {
1295 struct bpf_prog *prog = __get_filter(fprog, sk);
1296 int err;
1297
1298 if (IS_ERR(prog))
1299 return PTR_ERR(prog);
1300
1301 err = __sk_attach_prog(prog, sk);
1302 if (err < 0) {
1303 __bpf_prog_release(prog);
1304 return err;
1305 }
1306
1307 return 0;
1308 }
1309 EXPORT_SYMBOL_GPL(sk_attach_filter);
1310
1311 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1312 {
1313 struct bpf_prog *prog = __get_filter(fprog, sk);
1314 int err;
1315
1316 if (IS_ERR(prog))
1317 return PTR_ERR(prog);
1318
1319 err = __reuseport_attach_prog(prog, sk);
1320 if (err < 0) {
1321 __bpf_prog_release(prog);
1322 return err;
1323 }
1324
1325 return 0;
1326 }
1327
1328 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1329 {
1330 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1331 return ERR_PTR(-EPERM);
1332
1333 return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1334 }
1335
1336 int sk_attach_bpf(u32 ufd, struct sock *sk)
1337 {
1338 struct bpf_prog *prog = __get_bpf(ufd, sk);
1339 int err;
1340
1341 if (IS_ERR(prog))
1342 return PTR_ERR(prog);
1343
1344 err = __sk_attach_prog(prog, sk);
1345 if (err < 0) {
1346 bpf_prog_put(prog);
1347 return err;
1348 }
1349
1350 return 0;
1351 }
1352
1353 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1354 {
1355 struct bpf_prog *prog = __get_bpf(ufd, sk);
1356 int err;
1357
1358 if (IS_ERR(prog))
1359 return PTR_ERR(prog);
1360
1361 err = __reuseport_attach_prog(prog, sk);
1362 if (err < 0) {
1363 bpf_prog_put(prog);
1364 return err;
1365 }
1366
1367 return 0;
1368 }
1369
1370 struct bpf_scratchpad {
1371 union {
1372 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1373 u8 buff[MAX_BPF_STACK];
1374 };
1375 };
1376
1377 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1378
1379 static inline int __bpf_try_make_writable(struct sk_buff *skb,
1380 unsigned int write_len)
1381 {
1382 return skb_ensure_writable(skb, write_len);
1383 }
1384
1385 static inline int bpf_try_make_writable(struct sk_buff *skb,
1386 unsigned int write_len)
1387 {
1388 int err = __bpf_try_make_writable(skb, write_len);
1389
1390 bpf_compute_data_end(skb);
1391 return err;
1392 }
1393
1394 static int bpf_try_make_head_writable(struct sk_buff *skb)
1395 {
1396 return bpf_try_make_writable(skb, skb_headlen(skb));
1397 }
1398
1399 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1400 {
1401 if (skb_at_tc_ingress(skb))
1402 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1403 }
1404
1405 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1406 {
1407 if (skb_at_tc_ingress(skb))
1408 skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1409 }
1410
1411 BPF_CALL_5(bpf_skb_store_bytes, struct sk_buff *, skb, u32, offset,
1412 const void *, from, u32, len, u64, flags)
1413 {
1414 void *ptr;
1415
1416 if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1417 return -EINVAL;
1418 if (unlikely(offset > 0xffff))
1419 return -EFAULT;
1420 if (unlikely(bpf_try_make_writable(skb, offset + len)))
1421 return -EFAULT;
1422
1423 ptr = skb->data + offset;
1424 if (flags & BPF_F_RECOMPUTE_CSUM)
1425 __skb_postpull_rcsum(skb, ptr, len, offset);
1426
1427 memcpy(ptr, from, len);
1428
1429 if (flags & BPF_F_RECOMPUTE_CSUM)
1430 __skb_postpush_rcsum(skb, ptr, len, offset);
1431 if (flags & BPF_F_INVALIDATE_HASH)
1432 skb_clear_hash(skb);
1433
1434 return 0;
1435 }
1436
1437 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1438 .func = bpf_skb_store_bytes,
1439 .gpl_only = false,
1440 .ret_type = RET_INTEGER,
1441 .arg1_type = ARG_PTR_TO_CTX,
1442 .arg2_type = ARG_ANYTHING,
1443 .arg3_type = ARG_PTR_TO_MEM,
1444 .arg4_type = ARG_CONST_SIZE,
1445 .arg5_type = ARG_ANYTHING,
1446 };
1447
1448 BPF_CALL_4(bpf_skb_load_bytes, const struct sk_buff *, skb, u32, offset,
1449 void *, to, u32, len)
1450 {
1451 void *ptr;
1452
1453 if (unlikely(offset > 0xffff))
1454 goto err_clear;
1455
1456 ptr = skb_header_pointer(skb, offset, len, to);
1457 if (unlikely(!ptr))
1458 goto err_clear;
1459 if (ptr != to)
1460 memcpy(to, ptr, len);
1461
1462 return 0;
1463 err_clear:
1464 memset(to, 0, len);
1465 return -EFAULT;
1466 }
1467
1468 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1469 .func = bpf_skb_load_bytes,
1470 .gpl_only = false,
1471 .ret_type = RET_INTEGER,
1472 .arg1_type = ARG_PTR_TO_CTX,
1473 .arg2_type = ARG_ANYTHING,
1474 .arg3_type = ARG_PTR_TO_UNINIT_MEM,
1475 .arg4_type = ARG_CONST_SIZE,
1476 };
1477
1478 BPF_CALL_2(bpf_skb_pull_data, struct sk_buff *, skb, u32, len)
1479 {
1480 /* Idea is the following: should the needed direct read/write
1481 * test fail during runtime, we can pull in more data and redo
1482 * again, since implicitly, we invalidate previous checks here.
1483 *
1484 * Or, since we know how much we need to make read/writeable,
1485 * this can be done once at the program beginning for direct
1486 * access case. By this we overcome limitations of only current
1487 * headroom being accessible.
1488 */
1489 return bpf_try_make_writable(skb, len ? : skb_headlen(skb));
1490 }
1491
1492 static const struct bpf_func_proto bpf_skb_pull_data_proto = {
1493 .func = bpf_skb_pull_data,
1494 .gpl_only = false,
1495 .ret_type = RET_INTEGER,
1496 .arg1_type = ARG_PTR_TO_CTX,
1497 .arg2_type = ARG_ANYTHING,
1498 };
1499
1500 BPF_CALL_5(bpf_l3_csum_replace, struct sk_buff *, skb, u32, offset,
1501 u64, from, u64, to, u64, flags)
1502 {
1503 __sum16 *ptr;
1504
1505 if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1506 return -EINVAL;
1507 if (unlikely(offset > 0xffff || offset & 1))
1508 return -EFAULT;
1509 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1510 return -EFAULT;
1511
1512 ptr = (__sum16 *)(skb->data + offset);
1513 switch (flags & BPF_F_HDR_FIELD_MASK) {
1514 case 0:
1515 if (unlikely(from != 0))
1516 return -EINVAL;
1517
1518 csum_replace_by_diff(ptr, to);
1519 break;
1520 case 2:
1521 csum_replace2(ptr, from, to);
1522 break;
1523 case 4:
1524 csum_replace4(ptr, from, to);
1525 break;
1526 default:
1527 return -EINVAL;
1528 }
1529
1530 return 0;
1531 }
1532
1533 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1534 .func = bpf_l3_csum_replace,
1535 .gpl_only = false,
1536 .ret_type = RET_INTEGER,
1537 .arg1_type = ARG_PTR_TO_CTX,
1538 .arg2_type = ARG_ANYTHING,
1539 .arg3_type = ARG_ANYTHING,
1540 .arg4_type = ARG_ANYTHING,
1541 .arg5_type = ARG_ANYTHING,
1542 };
1543
1544 BPF_CALL_5(bpf_l4_csum_replace, struct sk_buff *, skb, u32, offset,
1545 u64, from, u64, to, u64, flags)
1546 {
1547 bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1548 bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1549 bool do_mforce = flags & BPF_F_MARK_ENFORCE;
1550 __sum16 *ptr;
1551
1552 if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_MARK_ENFORCE |
1553 BPF_F_PSEUDO_HDR | BPF_F_HDR_FIELD_MASK)))
1554 return -EINVAL;
1555 if (unlikely(offset > 0xffff || offset & 1))
1556 return -EFAULT;
1557 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1558 return -EFAULT;
1559
1560 ptr = (__sum16 *)(skb->data + offset);
1561 if (is_mmzero && !do_mforce && !*ptr)
1562 return 0;
1563
1564 switch (flags & BPF_F_HDR_FIELD_MASK) {
1565 case 0:
1566 if (unlikely(from != 0))
1567 return -EINVAL;
1568
1569 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1570 break;
1571 case 2:
1572 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1573 break;
1574 case 4:
1575 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1576 break;
1577 default:
1578 return -EINVAL;
1579 }
1580
1581 if (is_mmzero && !*ptr)
1582 *ptr = CSUM_MANGLED_0;
1583 return 0;
1584 }
1585
1586 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1587 .func = bpf_l4_csum_replace,
1588 .gpl_only = false,
1589 .ret_type = RET_INTEGER,
1590 .arg1_type = ARG_PTR_TO_CTX,
1591 .arg2_type = ARG_ANYTHING,
1592 .arg3_type = ARG_ANYTHING,
1593 .arg4_type = ARG_ANYTHING,
1594 .arg5_type = ARG_ANYTHING,
1595 };
1596
1597 BPF_CALL_5(bpf_csum_diff, __be32 *, from, u32, from_size,
1598 __be32 *, to, u32, to_size, __wsum, seed)
1599 {
1600 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1601 u32 diff_size = from_size + to_size;
1602 int i, j = 0;
1603
1604 /* This is quite flexible, some examples:
1605 *
1606 * from_size == 0, to_size > 0, seed := csum --> pushing data
1607 * from_size > 0, to_size == 0, seed := csum --> pulling data
1608 * from_size > 0, to_size > 0, seed := 0 --> diffing data
1609 *
1610 * Even for diffing, from_size and to_size don't need to be equal.
1611 */
1612 if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1613 diff_size > sizeof(sp->diff)))
1614 return -EINVAL;
1615
1616 for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1617 sp->diff[j] = ~from[i];
1618 for (i = 0; i < to_size / sizeof(__be32); i++, j++)
1619 sp->diff[j] = to[i];
1620
1621 return csum_partial(sp->diff, diff_size, seed);
1622 }
1623
1624 static const struct bpf_func_proto bpf_csum_diff_proto = {
1625 .func = bpf_csum_diff,
1626 .gpl_only = false,
1627 .pkt_access = true,
1628 .ret_type = RET_INTEGER,
1629 .arg1_type = ARG_PTR_TO_MEM,
1630 .arg2_type = ARG_CONST_SIZE_OR_ZERO,
1631 .arg3_type = ARG_PTR_TO_MEM,
1632 .arg4_type = ARG_CONST_SIZE_OR_ZERO,
1633 .arg5_type = ARG_ANYTHING,
1634 };
1635
1636 BPF_CALL_2(bpf_csum_update, struct sk_buff *, skb, __wsum, csum)
1637 {
1638 /* The interface is to be used in combination with bpf_csum_diff()
1639 * for direct packet writes. csum rotation for alignment as well
1640 * as emulating csum_sub() can be done from the eBPF program.
1641 */
1642 if (skb->ip_summed == CHECKSUM_COMPLETE)
1643 return (skb->csum = csum_add(skb->csum, csum));
1644
1645 return -ENOTSUPP;
1646 }
1647
1648 static const struct bpf_func_proto bpf_csum_update_proto = {
1649 .func = bpf_csum_update,
1650 .gpl_only = false,
1651 .ret_type = RET_INTEGER,
1652 .arg1_type = ARG_PTR_TO_CTX,
1653 .arg2_type = ARG_ANYTHING,
1654 };
1655
1656 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1657 {
1658 return dev_forward_skb(dev, skb);
1659 }
1660
1661 static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
1662 struct sk_buff *skb)
1663 {
1664 int ret = ____dev_forward_skb(dev, skb);
1665
1666 if (likely(!ret)) {
1667 skb->dev = dev;
1668 ret = netif_rx(skb);
1669 }
1670
1671 return ret;
1672 }
1673
1674 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1675 {
1676 int ret;
1677
1678 if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1679 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1680 kfree_skb(skb);
1681 return -ENETDOWN;
1682 }
1683
1684 skb->dev = dev;
1685
1686 __this_cpu_inc(xmit_recursion);
1687 ret = dev_queue_xmit(skb);
1688 __this_cpu_dec(xmit_recursion);
1689
1690 return ret;
1691 }
1692
1693 static int __bpf_redirect_no_mac(struct sk_buff *skb, struct net_device *dev,
1694 u32 flags)
1695 {
1696 /* skb->mac_len is not set on normal egress */
1697 unsigned int mlen = skb->network_header - skb->mac_header;
1698
1699 __skb_pull(skb, mlen);
1700
1701 /* At ingress, the mac header has already been pulled once.
1702 * At egress, skb_pospull_rcsum has to be done in case that
1703 * the skb is originated from ingress (i.e. a forwarded skb)
1704 * to ensure that rcsum starts at net header.
1705 */
1706 if (!skb_at_tc_ingress(skb))
1707 skb_postpull_rcsum(skb, skb_mac_header(skb), mlen);
1708 skb_pop_mac_header(skb);
1709 skb_reset_mac_len(skb);
1710 return flags & BPF_F_INGRESS ?
1711 __bpf_rx_skb_no_mac(dev, skb) : __bpf_tx_skb(dev, skb);
1712 }
1713
1714 static int __bpf_redirect_common(struct sk_buff *skb, struct net_device *dev,
1715 u32 flags)
1716 {
1717 /* Verify that a link layer header is carried */
1718 if (unlikely(skb->mac_header >= skb->network_header)) {
1719 kfree_skb(skb);
1720 return -ERANGE;
1721 }
1722
1723 bpf_push_mac_rcsum(skb);
1724 return flags & BPF_F_INGRESS ?
1725 __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1726 }
1727
1728 static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
1729 u32 flags)
1730 {
1731 if (dev_is_mac_header_xmit(dev))
1732 return __bpf_redirect_common(skb, dev, flags);
1733 else
1734 return __bpf_redirect_no_mac(skb, dev, flags);
1735 }
1736
1737 BPF_CALL_3(bpf_clone_redirect, struct sk_buff *, skb, u32, ifindex, u64, flags)
1738 {
1739 struct net_device *dev;
1740 struct sk_buff *clone;
1741 int ret;
1742
1743 if (unlikely(flags & ~(BPF_F_INGRESS)))
1744 return -EINVAL;
1745
1746 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1747 if (unlikely(!dev))
1748 return -EINVAL;
1749
1750 clone = skb_clone(skb, GFP_ATOMIC);
1751 if (unlikely(!clone))
1752 return -ENOMEM;
1753
1754 /* For direct write, we need to keep the invariant that the skbs
1755 * we're dealing with need to be uncloned. Should uncloning fail
1756 * here, we need to free the just generated clone to unclone once
1757 * again.
1758 */
1759 ret = bpf_try_make_head_writable(skb);
1760 if (unlikely(ret)) {
1761 kfree_skb(clone);
1762 return -ENOMEM;
1763 }
1764
1765 return __bpf_redirect(clone, dev, flags);
1766 }
1767
1768 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1769 .func = bpf_clone_redirect,
1770 .gpl_only = false,
1771 .ret_type = RET_INTEGER,
1772 .arg1_type = ARG_PTR_TO_CTX,
1773 .arg2_type = ARG_ANYTHING,
1774 .arg3_type = ARG_ANYTHING,
1775 };
1776
1777 struct redirect_info {
1778 u32 ifindex;
1779 u32 flags;
1780 };
1781
1782 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1783
1784 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
1785 {
1786 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1787
1788 if (unlikely(flags & ~(BPF_F_INGRESS)))
1789 return TC_ACT_SHOT;
1790
1791 ri->ifindex = ifindex;
1792 ri->flags = flags;
1793
1794 return TC_ACT_REDIRECT;
1795 }
1796
1797 int skb_do_redirect(struct sk_buff *skb)
1798 {
1799 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1800 struct net_device *dev;
1801
1802 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1803 ri->ifindex = 0;
1804 if (unlikely(!dev)) {
1805 kfree_skb(skb);
1806 return -EINVAL;
1807 }
1808
1809 return __bpf_redirect(skb, dev, ri->flags);
1810 }
1811
1812 static const struct bpf_func_proto bpf_redirect_proto = {
1813 .func = bpf_redirect,
1814 .gpl_only = false,
1815 .ret_type = RET_INTEGER,
1816 .arg1_type = ARG_ANYTHING,
1817 .arg2_type = ARG_ANYTHING,
1818 };
1819
1820 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
1821 {
1822 return task_get_classid(skb);
1823 }
1824
1825 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1826 .func = bpf_get_cgroup_classid,
1827 .gpl_only = false,
1828 .ret_type = RET_INTEGER,
1829 .arg1_type = ARG_PTR_TO_CTX,
1830 };
1831
1832 BPF_CALL_1(bpf_get_route_realm, const struct sk_buff *, skb)
1833 {
1834 return dst_tclassid(skb);
1835 }
1836
1837 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1838 .func = bpf_get_route_realm,
1839 .gpl_only = false,
1840 .ret_type = RET_INTEGER,
1841 .arg1_type = ARG_PTR_TO_CTX,
1842 };
1843
1844 BPF_CALL_1(bpf_get_hash_recalc, struct sk_buff *, skb)
1845 {
1846 /* If skb_clear_hash() was called due to mangling, we can
1847 * trigger SW recalculation here. Later access to hash
1848 * can then use the inline skb->hash via context directly
1849 * instead of calling this helper again.
1850 */
1851 return skb_get_hash(skb);
1852 }
1853
1854 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1855 .func = bpf_get_hash_recalc,
1856 .gpl_only = false,
1857 .ret_type = RET_INTEGER,
1858 .arg1_type = ARG_PTR_TO_CTX,
1859 };
1860
1861 BPF_CALL_1(bpf_set_hash_invalid, struct sk_buff *, skb)
1862 {
1863 /* After all direct packet write, this can be used once for
1864 * triggering a lazy recalc on next skb_get_hash() invocation.
1865 */
1866 skb_clear_hash(skb);
1867 return 0;
1868 }
1869
1870 static const struct bpf_func_proto bpf_set_hash_invalid_proto = {
1871 .func = bpf_set_hash_invalid,
1872 .gpl_only = false,
1873 .ret_type = RET_INTEGER,
1874 .arg1_type = ARG_PTR_TO_CTX,
1875 };
1876
1877 BPF_CALL_2(bpf_set_hash, struct sk_buff *, skb, u32, hash)
1878 {
1879 /* Set user specified hash as L4(+), so that it gets returned
1880 * on skb_get_hash() call unless BPF prog later on triggers a
1881 * skb_clear_hash().
1882 */
1883 __skb_set_sw_hash(skb, hash, true);
1884 return 0;
1885 }
1886
1887 static const struct bpf_func_proto bpf_set_hash_proto = {
1888 .func = bpf_set_hash,
1889 .gpl_only = false,
1890 .ret_type = RET_INTEGER,
1891 .arg1_type = ARG_PTR_TO_CTX,
1892 .arg2_type = ARG_ANYTHING,
1893 };
1894
1895 BPF_CALL_3(bpf_skb_vlan_push, struct sk_buff *, skb, __be16, vlan_proto,
1896 u16, vlan_tci)
1897 {
1898 int ret;
1899
1900 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1901 vlan_proto != htons(ETH_P_8021AD)))
1902 vlan_proto = htons(ETH_P_8021Q);
1903
1904 bpf_push_mac_rcsum(skb);
1905 ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1906 bpf_pull_mac_rcsum(skb);
1907
1908 bpf_compute_data_end(skb);
1909 return ret;
1910 }
1911
1912 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1913 .func = bpf_skb_vlan_push,
1914 .gpl_only = false,
1915 .ret_type = RET_INTEGER,
1916 .arg1_type = ARG_PTR_TO_CTX,
1917 .arg2_type = ARG_ANYTHING,
1918 .arg3_type = ARG_ANYTHING,
1919 };
1920 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1921
1922 BPF_CALL_1(bpf_skb_vlan_pop, struct sk_buff *, skb)
1923 {
1924 int ret;
1925
1926 bpf_push_mac_rcsum(skb);
1927 ret = skb_vlan_pop(skb);
1928 bpf_pull_mac_rcsum(skb);
1929
1930 bpf_compute_data_end(skb);
1931 return ret;
1932 }
1933
1934 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1935 .func = bpf_skb_vlan_pop,
1936 .gpl_only = false,
1937 .ret_type = RET_INTEGER,
1938 .arg1_type = ARG_PTR_TO_CTX,
1939 };
1940 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1941
1942 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
1943 {
1944 /* Caller already did skb_cow() with len as headroom,
1945 * so no need to do it here.
1946 */
1947 skb_push(skb, len);
1948 memmove(skb->data, skb->data + len, off);
1949 memset(skb->data + off, 0, len);
1950
1951 /* No skb_postpush_rcsum(skb, skb->data + off, len)
1952 * needed here as it does not change the skb->csum
1953 * result for checksum complete when summing over
1954 * zeroed blocks.
1955 */
1956 return 0;
1957 }
1958
1959 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
1960 {
1961 /* skb_ensure_writable() is not needed here, as we're
1962 * already working on an uncloned skb.
1963 */
1964 if (unlikely(!pskb_may_pull(skb, off + len)))
1965 return -ENOMEM;
1966
1967 skb_postpull_rcsum(skb, skb->data + off, len);
1968 memmove(skb->data + len, skb->data, off);
1969 __skb_pull(skb, len);
1970
1971 return 0;
1972 }
1973
1974 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
1975 {
1976 bool trans_same = skb->transport_header == skb->network_header;
1977 int ret;
1978
1979 /* There's no need for __skb_push()/__skb_pull() pair to
1980 * get to the start of the mac header as we're guaranteed
1981 * to always start from here under eBPF.
1982 */
1983 ret = bpf_skb_generic_push(skb, off, len);
1984 if (likely(!ret)) {
1985 skb->mac_header -= len;
1986 skb->network_header -= len;
1987 if (trans_same)
1988 skb->transport_header = skb->network_header;
1989 }
1990
1991 return ret;
1992 }
1993
1994 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
1995 {
1996 bool trans_same = skb->transport_header == skb->network_header;
1997 int ret;
1998
1999 /* Same here, __skb_push()/__skb_pull() pair not needed. */
2000 ret = bpf_skb_generic_pop(skb, off, len);
2001 if (likely(!ret)) {
2002 skb->mac_header += len;
2003 skb->network_header += len;
2004 if (trans_same)
2005 skb->transport_header = skb->network_header;
2006 }
2007
2008 return ret;
2009 }
2010
2011 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
2012 {
2013 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2014 u32 off = skb->network_header - skb->mac_header;
2015 int ret;
2016
2017 ret = skb_cow(skb, len_diff);
2018 if (unlikely(ret < 0))
2019 return ret;
2020
2021 ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2022 if (unlikely(ret < 0))
2023 return ret;
2024
2025 if (skb_is_gso(skb)) {
2026 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV4 needs to
2027 * be changed into SKB_GSO_TCPV6.
2028 */
2029 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2030 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
2031 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV6;
2032 }
2033
2034 /* Due to IPv6 header, MSS needs to be downgraded. */
2035 skb_shinfo(skb)->gso_size -= len_diff;
2036 /* Header must be checked, and gso_segs recomputed. */
2037 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2038 skb_shinfo(skb)->gso_segs = 0;
2039 }
2040
2041 skb->protocol = htons(ETH_P_IPV6);
2042 skb_clear_hash(skb);
2043
2044 return 0;
2045 }
2046
2047 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
2048 {
2049 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2050 u32 off = skb->network_header - skb->mac_header;
2051 int ret;
2052
2053 ret = skb_unclone(skb, GFP_ATOMIC);
2054 if (unlikely(ret < 0))
2055 return ret;
2056
2057 ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2058 if (unlikely(ret < 0))
2059 return ret;
2060
2061 if (skb_is_gso(skb)) {
2062 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV6 needs to
2063 * be changed into SKB_GSO_TCPV4.
2064 */
2065 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
2066 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
2067 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
2068 }
2069
2070 /* Due to IPv4 header, MSS can be upgraded. */
2071 skb_shinfo(skb)->gso_size += len_diff;
2072 /* Header must be checked, and gso_segs recomputed. */
2073 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2074 skb_shinfo(skb)->gso_segs = 0;
2075 }
2076
2077 skb->protocol = htons(ETH_P_IP);
2078 skb_clear_hash(skb);
2079
2080 return 0;
2081 }
2082
2083 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
2084 {
2085 __be16 from_proto = skb->protocol;
2086
2087 if (from_proto == htons(ETH_P_IP) &&
2088 to_proto == htons(ETH_P_IPV6))
2089 return bpf_skb_proto_4_to_6(skb);
2090
2091 if (from_proto == htons(ETH_P_IPV6) &&
2092 to_proto == htons(ETH_P_IP))
2093 return bpf_skb_proto_6_to_4(skb);
2094
2095 return -ENOTSUPP;
2096 }
2097
2098 BPF_CALL_3(bpf_skb_change_proto, struct sk_buff *, skb, __be16, proto,
2099 u64, flags)
2100 {
2101 int ret;
2102
2103 if (unlikely(flags))
2104 return -EINVAL;
2105
2106 /* General idea is that this helper does the basic groundwork
2107 * needed for changing the protocol, and eBPF program fills the
2108 * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
2109 * and other helpers, rather than passing a raw buffer here.
2110 *
2111 * The rationale is to keep this minimal and without a need to
2112 * deal with raw packet data. F.e. even if we would pass buffers
2113 * here, the program still needs to call the bpf_lX_csum_replace()
2114 * helpers anyway. Plus, this way we keep also separation of
2115 * concerns, since f.e. bpf_skb_store_bytes() should only take
2116 * care of stores.
2117 *
2118 * Currently, additional options and extension header space are
2119 * not supported, but flags register is reserved so we can adapt
2120 * that. For offloads, we mark packet as dodgy, so that headers
2121 * need to be verified first.
2122 */
2123 ret = bpf_skb_proto_xlat(skb, proto);
2124 bpf_compute_data_end(skb);
2125 return ret;
2126 }
2127
2128 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
2129 .func = bpf_skb_change_proto,
2130 .gpl_only = false,
2131 .ret_type = RET_INTEGER,
2132 .arg1_type = ARG_PTR_TO_CTX,
2133 .arg2_type = ARG_ANYTHING,
2134 .arg3_type = ARG_ANYTHING,
2135 };
2136
2137 BPF_CALL_2(bpf_skb_change_type, struct sk_buff *, skb, u32, pkt_type)
2138 {
2139 /* We only allow a restricted subset to be changed for now. */
2140 if (unlikely(!skb_pkt_type_ok(skb->pkt_type) ||
2141 !skb_pkt_type_ok(pkt_type)))
2142 return -EINVAL;
2143
2144 skb->pkt_type = pkt_type;
2145 return 0;
2146 }
2147
2148 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2149 .func = bpf_skb_change_type,
2150 .gpl_only = false,
2151 .ret_type = RET_INTEGER,
2152 .arg1_type = ARG_PTR_TO_CTX,
2153 .arg2_type = ARG_ANYTHING,
2154 };
2155
2156 static u32 __bpf_skb_min_len(const struct sk_buff *skb)
2157 {
2158 u32 min_len = skb_network_offset(skb);
2159
2160 if (skb_transport_header_was_set(skb))
2161 min_len = skb_transport_offset(skb);
2162 if (skb->ip_summed == CHECKSUM_PARTIAL)
2163 min_len = skb_checksum_start_offset(skb) +
2164 skb->csum_offset + sizeof(__sum16);
2165 return min_len;
2166 }
2167
2168 static u32 __bpf_skb_max_len(const struct sk_buff *skb)
2169 {
2170 return skb->dev->mtu + skb->dev->hard_header_len;
2171 }
2172
2173 static int bpf_skb_grow_rcsum(struct sk_buff *skb, unsigned int new_len)
2174 {
2175 unsigned int old_len = skb->len;
2176 int ret;
2177
2178 ret = __skb_grow_rcsum(skb, new_len);
2179 if (!ret)
2180 memset(skb->data + old_len, 0, new_len - old_len);
2181 return ret;
2182 }
2183
2184 static int bpf_skb_trim_rcsum(struct sk_buff *skb, unsigned int new_len)
2185 {
2186 return __skb_trim_rcsum(skb, new_len);
2187 }
2188
2189 BPF_CALL_3(bpf_skb_change_tail, struct sk_buff *, skb, u32, new_len,
2190 u64, flags)
2191 {
2192 u32 max_len = __bpf_skb_max_len(skb);
2193 u32 min_len = __bpf_skb_min_len(skb);
2194 int ret;
2195
2196 if (unlikely(flags || new_len > max_len || new_len < min_len))
2197 return -EINVAL;
2198 if (skb->encapsulation)
2199 return -ENOTSUPP;
2200
2201 /* The basic idea of this helper is that it's performing the
2202 * needed work to either grow or trim an skb, and eBPF program
2203 * rewrites the rest via helpers like bpf_skb_store_bytes(),
2204 * bpf_lX_csum_replace() and others rather than passing a raw
2205 * buffer here. This one is a slow path helper and intended
2206 * for replies with control messages.
2207 *
2208 * Like in bpf_skb_change_proto(), we want to keep this rather
2209 * minimal and without protocol specifics so that we are able
2210 * to separate concerns as in bpf_skb_store_bytes() should only
2211 * be the one responsible for writing buffers.
2212 *
2213 * It's really expected to be a slow path operation here for
2214 * control message replies, so we're implicitly linearizing,
2215 * uncloning and drop offloads from the skb by this.
2216 */
2217 ret = __bpf_try_make_writable(skb, skb->len);
2218 if (!ret) {
2219 if (new_len > skb->len)
2220 ret = bpf_skb_grow_rcsum(skb, new_len);
2221 else if (new_len < skb->len)
2222 ret = bpf_skb_trim_rcsum(skb, new_len);
2223 if (!ret && skb_is_gso(skb))
2224 skb_gso_reset(skb);
2225 }
2226
2227 bpf_compute_data_end(skb);
2228 return ret;
2229 }
2230
2231 static const struct bpf_func_proto bpf_skb_change_tail_proto = {
2232 .func = bpf_skb_change_tail,
2233 .gpl_only = false,
2234 .ret_type = RET_INTEGER,
2235 .arg1_type = ARG_PTR_TO_CTX,
2236 .arg2_type = ARG_ANYTHING,
2237 .arg3_type = ARG_ANYTHING,
2238 };
2239
2240 BPF_CALL_3(bpf_skb_change_head, struct sk_buff *, skb, u32, head_room,
2241 u64, flags)
2242 {
2243 u32 max_len = __bpf_skb_max_len(skb);
2244 u32 new_len = skb->len + head_room;
2245 int ret;
2246
2247 if (unlikely(flags || (!skb_is_gso(skb) && new_len > max_len) ||
2248 new_len < skb->len))
2249 return -EINVAL;
2250
2251 ret = skb_cow(skb, head_room);
2252 if (likely(!ret)) {
2253 /* Idea for this helper is that we currently only
2254 * allow to expand on mac header. This means that
2255 * skb->protocol network header, etc, stay as is.
2256 * Compared to bpf_skb_change_tail(), we're more
2257 * flexible due to not needing to linearize or
2258 * reset GSO. Intention for this helper is to be
2259 * used by an L3 skb that needs to push mac header
2260 * for redirection into L2 device.
2261 */
2262 __skb_push(skb, head_room);
2263 memset(skb->data, 0, head_room);
2264 skb_reset_mac_header(skb);
2265 }
2266
2267 bpf_compute_data_end(skb);
2268 return 0;
2269 }
2270
2271 static const struct bpf_func_proto bpf_skb_change_head_proto = {
2272 .func = bpf_skb_change_head,
2273 .gpl_only = false,
2274 .ret_type = RET_INTEGER,
2275 .arg1_type = ARG_PTR_TO_CTX,
2276 .arg2_type = ARG_ANYTHING,
2277 .arg3_type = ARG_ANYTHING,
2278 };
2279
2280 BPF_CALL_2(bpf_xdp_adjust_head, struct xdp_buff *, xdp, int, offset)
2281 {
2282 void *data = xdp->data + offset;
2283
2284 if (unlikely(data < xdp->data_hard_start ||
2285 data > xdp->data_end - ETH_HLEN))
2286 return -EINVAL;
2287
2288 xdp->data = data;
2289
2290 return 0;
2291 }
2292
2293 static const struct bpf_func_proto bpf_xdp_adjust_head_proto = {
2294 .func = bpf_xdp_adjust_head,
2295 .gpl_only = false,
2296 .ret_type = RET_INTEGER,
2297 .arg1_type = ARG_PTR_TO_CTX,
2298 .arg2_type = ARG_ANYTHING,
2299 };
2300
2301 bool bpf_helper_changes_pkt_data(void *func)
2302 {
2303 if (func == bpf_skb_vlan_push ||
2304 func == bpf_skb_vlan_pop ||
2305 func == bpf_skb_store_bytes ||
2306 func == bpf_skb_change_proto ||
2307 func == bpf_skb_change_head ||
2308 func == bpf_skb_change_tail ||
2309 func == bpf_skb_pull_data ||
2310 func == bpf_clone_redirect ||
2311 func == bpf_l3_csum_replace ||
2312 func == bpf_l4_csum_replace ||
2313 func == bpf_xdp_adjust_head)
2314 return true;
2315
2316 return false;
2317 }
2318
2319 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2320 unsigned long off, unsigned long len)
2321 {
2322 void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2323
2324 if (unlikely(!ptr))
2325 return len;
2326 if (ptr != dst_buff)
2327 memcpy(dst_buff, ptr, len);
2328
2329 return 0;
2330 }
2331
2332 BPF_CALL_5(bpf_skb_event_output, struct sk_buff *, skb, struct bpf_map *, map,
2333 u64, flags, void *, meta, u64, meta_size)
2334 {
2335 u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2336
2337 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2338 return -EINVAL;
2339 if (unlikely(skb_size > skb->len))
2340 return -EFAULT;
2341
2342 return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2343 bpf_skb_copy);
2344 }
2345
2346 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2347 .func = bpf_skb_event_output,
2348 .gpl_only = true,
2349 .ret_type = RET_INTEGER,
2350 .arg1_type = ARG_PTR_TO_CTX,
2351 .arg2_type = ARG_CONST_MAP_PTR,
2352 .arg3_type = ARG_ANYTHING,
2353 .arg4_type = ARG_PTR_TO_MEM,
2354 .arg5_type = ARG_CONST_SIZE,
2355 };
2356
2357 static unsigned short bpf_tunnel_key_af(u64 flags)
2358 {
2359 return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2360 }
2361
2362 BPF_CALL_4(bpf_skb_get_tunnel_key, struct sk_buff *, skb, struct bpf_tunnel_key *, to,
2363 u32, size, u64, flags)
2364 {
2365 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2366 u8 compat[sizeof(struct bpf_tunnel_key)];
2367 void *to_orig = to;
2368 int err;
2369
2370 if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2371 err = -EINVAL;
2372 goto err_clear;
2373 }
2374 if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2375 err = -EPROTO;
2376 goto err_clear;
2377 }
2378 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2379 err = -EINVAL;
2380 switch (size) {
2381 case offsetof(struct bpf_tunnel_key, tunnel_label):
2382 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2383 goto set_compat;
2384 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2385 /* Fixup deprecated structure layouts here, so we have
2386 * a common path later on.
2387 */
2388 if (ip_tunnel_info_af(info) != AF_INET)
2389 goto err_clear;
2390 set_compat:
2391 to = (struct bpf_tunnel_key *)compat;
2392 break;
2393 default:
2394 goto err_clear;
2395 }
2396 }
2397
2398 to->tunnel_id = be64_to_cpu(info->key.tun_id);
2399 to->tunnel_tos = info->key.tos;
2400 to->tunnel_ttl = info->key.ttl;
2401
2402 if (flags & BPF_F_TUNINFO_IPV6) {
2403 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2404 sizeof(to->remote_ipv6));
2405 to->tunnel_label = be32_to_cpu(info->key.label);
2406 } else {
2407 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2408 }
2409
2410 if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2411 memcpy(to_orig, to, size);
2412
2413 return 0;
2414 err_clear:
2415 memset(to_orig, 0, size);
2416 return err;
2417 }
2418
2419 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2420 .func = bpf_skb_get_tunnel_key,
2421 .gpl_only = false,
2422 .ret_type = RET_INTEGER,
2423 .arg1_type = ARG_PTR_TO_CTX,
2424 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
2425 .arg3_type = ARG_CONST_SIZE,
2426 .arg4_type = ARG_ANYTHING,
2427 };
2428
2429 BPF_CALL_3(bpf_skb_get_tunnel_opt, struct sk_buff *, skb, u8 *, to, u32, size)
2430 {
2431 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2432 int err;
2433
2434 if (unlikely(!info ||
2435 !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2436 err = -ENOENT;
2437 goto err_clear;
2438 }
2439 if (unlikely(size < info->options_len)) {
2440 err = -ENOMEM;
2441 goto err_clear;
2442 }
2443
2444 ip_tunnel_info_opts_get(to, info);
2445 if (size > info->options_len)
2446 memset(to + info->options_len, 0, size - info->options_len);
2447
2448 return info->options_len;
2449 err_clear:
2450 memset(to, 0, size);
2451 return err;
2452 }
2453
2454 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2455 .func = bpf_skb_get_tunnel_opt,
2456 .gpl_only = false,
2457 .ret_type = RET_INTEGER,
2458 .arg1_type = ARG_PTR_TO_CTX,
2459 .arg2_type = ARG_PTR_TO_UNINIT_MEM,
2460 .arg3_type = ARG_CONST_SIZE,
2461 };
2462
2463 static struct metadata_dst __percpu *md_dst;
2464
2465 BPF_CALL_4(bpf_skb_set_tunnel_key, struct sk_buff *, skb,
2466 const struct bpf_tunnel_key *, from, u32, size, u64, flags)
2467 {
2468 struct metadata_dst *md = this_cpu_ptr(md_dst);
2469 u8 compat[sizeof(struct bpf_tunnel_key)];
2470 struct ip_tunnel_info *info;
2471
2472 if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2473 BPF_F_DONT_FRAGMENT)))
2474 return -EINVAL;
2475 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2476 switch (size) {
2477 case offsetof(struct bpf_tunnel_key, tunnel_label):
2478 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2479 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2480 /* Fixup deprecated structure layouts here, so we have
2481 * a common path later on.
2482 */
2483 memcpy(compat, from, size);
2484 memset(compat + size, 0, sizeof(compat) - size);
2485 from = (const struct bpf_tunnel_key *) compat;
2486 break;
2487 default:
2488 return -EINVAL;
2489 }
2490 }
2491 if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
2492 from->tunnel_ext))
2493 return -EINVAL;
2494
2495 skb_dst_drop(skb);
2496 dst_hold((struct dst_entry *) md);
2497 skb_dst_set(skb, (struct dst_entry *) md);
2498
2499 info = &md->u.tun_info;
2500 info->mode = IP_TUNNEL_INFO_TX;
2501
2502 info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
2503 if (flags & BPF_F_DONT_FRAGMENT)
2504 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
2505
2506 info->key.tun_id = cpu_to_be64(from->tunnel_id);
2507 info->key.tos = from->tunnel_tos;
2508 info->key.ttl = from->tunnel_ttl;
2509
2510 if (flags & BPF_F_TUNINFO_IPV6) {
2511 info->mode |= IP_TUNNEL_INFO_IPV6;
2512 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
2513 sizeof(from->remote_ipv6));
2514 info->key.label = cpu_to_be32(from->tunnel_label) &
2515 IPV6_FLOWLABEL_MASK;
2516 } else {
2517 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
2518 if (flags & BPF_F_ZERO_CSUM_TX)
2519 info->key.tun_flags &= ~TUNNEL_CSUM;
2520 }
2521
2522 return 0;
2523 }
2524
2525 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
2526 .func = bpf_skb_set_tunnel_key,
2527 .gpl_only = false,
2528 .ret_type = RET_INTEGER,
2529 .arg1_type = ARG_PTR_TO_CTX,
2530 .arg2_type = ARG_PTR_TO_MEM,
2531 .arg3_type = ARG_CONST_SIZE,
2532 .arg4_type = ARG_ANYTHING,
2533 };
2534
2535 BPF_CALL_3(bpf_skb_set_tunnel_opt, struct sk_buff *, skb,
2536 const u8 *, from, u32, size)
2537 {
2538 struct ip_tunnel_info *info = skb_tunnel_info(skb);
2539 const struct metadata_dst *md = this_cpu_ptr(md_dst);
2540
2541 if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
2542 return -EINVAL;
2543 if (unlikely(size > IP_TUNNEL_OPTS_MAX))
2544 return -ENOMEM;
2545
2546 ip_tunnel_info_opts_set(info, from, size);
2547
2548 return 0;
2549 }
2550
2551 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
2552 .func = bpf_skb_set_tunnel_opt,
2553 .gpl_only = false,
2554 .ret_type = RET_INTEGER,
2555 .arg1_type = ARG_PTR_TO_CTX,
2556 .arg2_type = ARG_PTR_TO_MEM,
2557 .arg3_type = ARG_CONST_SIZE,
2558 };
2559
2560 static const struct bpf_func_proto *
2561 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
2562 {
2563 if (!md_dst) {
2564 /* Race is not possible, since it's called from verifier
2565 * that is holding verifier mutex.
2566 */
2567 md_dst = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
2568 GFP_KERNEL);
2569 if (!md_dst)
2570 return NULL;
2571 }
2572
2573 switch (which) {
2574 case BPF_FUNC_skb_set_tunnel_key:
2575 return &bpf_skb_set_tunnel_key_proto;
2576 case BPF_FUNC_skb_set_tunnel_opt:
2577 return &bpf_skb_set_tunnel_opt_proto;
2578 default:
2579 return NULL;
2580 }
2581 }
2582
2583 BPF_CALL_3(bpf_skb_under_cgroup, struct sk_buff *, skb, struct bpf_map *, map,
2584 u32, idx)
2585 {
2586 struct bpf_array *array = container_of(map, struct bpf_array, map);
2587 struct cgroup *cgrp;
2588 struct sock *sk;
2589
2590 sk = skb_to_full_sk(skb);
2591 if (!sk || !sk_fullsock(sk))
2592 return -ENOENT;
2593 if (unlikely(idx >= array->map.max_entries))
2594 return -E2BIG;
2595
2596 cgrp = READ_ONCE(array->ptrs[idx]);
2597 if (unlikely(!cgrp))
2598 return -EAGAIN;
2599
2600 return sk_under_cgroup_hierarchy(sk, cgrp);
2601 }
2602
2603 static const struct bpf_func_proto bpf_skb_under_cgroup_proto = {
2604 .func = bpf_skb_under_cgroup,
2605 .gpl_only = false,
2606 .ret_type = RET_INTEGER,
2607 .arg1_type = ARG_PTR_TO_CTX,
2608 .arg2_type = ARG_CONST_MAP_PTR,
2609 .arg3_type = ARG_ANYTHING,
2610 };
2611
2612 static unsigned long bpf_xdp_copy(void *dst_buff, const void *src_buff,
2613 unsigned long off, unsigned long len)
2614 {
2615 memcpy(dst_buff, src_buff + off, len);
2616 return 0;
2617 }
2618
2619 BPF_CALL_5(bpf_xdp_event_output, struct xdp_buff *, xdp, struct bpf_map *, map,
2620 u64, flags, void *, meta, u64, meta_size)
2621 {
2622 u64 xdp_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2623
2624 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2625 return -EINVAL;
2626 if (unlikely(xdp_size > (unsigned long)(xdp->data_end - xdp->data)))
2627 return -EFAULT;
2628
2629 return bpf_event_output(map, flags, meta, meta_size, xdp->data,
2630 xdp_size, bpf_xdp_copy);
2631 }
2632
2633 static const struct bpf_func_proto bpf_xdp_event_output_proto = {
2634 .func = bpf_xdp_event_output,
2635 .gpl_only = true,
2636 .ret_type = RET_INTEGER,
2637 .arg1_type = ARG_PTR_TO_CTX,
2638 .arg2_type = ARG_CONST_MAP_PTR,
2639 .arg3_type = ARG_ANYTHING,
2640 .arg4_type = ARG_PTR_TO_MEM,
2641 .arg5_type = ARG_CONST_SIZE,
2642 };
2643
2644 BPF_CALL_1(bpf_get_socket_cookie, struct sk_buff *, skb)
2645 {
2646 return skb->sk ? sock_gen_cookie(skb->sk) : 0;
2647 }
2648
2649 static const struct bpf_func_proto bpf_get_socket_cookie_proto = {
2650 .func = bpf_get_socket_cookie,
2651 .gpl_only = false,
2652 .ret_type = RET_INTEGER,
2653 .arg1_type = ARG_PTR_TO_CTX,
2654 };
2655
2656 BPF_CALL_1(bpf_get_socket_uid, struct sk_buff *, skb)
2657 {
2658 struct sock *sk = sk_to_full_sk(skb->sk);
2659 kuid_t kuid;
2660
2661 if (!sk || !sk_fullsock(sk))
2662 return overflowuid;
2663 kuid = sock_net_uid(sock_net(sk), sk);
2664 return from_kuid_munged(sock_net(sk)->user_ns, kuid);
2665 }
2666
2667 static const struct bpf_func_proto bpf_get_socket_uid_proto = {
2668 .func = bpf_get_socket_uid,
2669 .gpl_only = false,
2670 .ret_type = RET_INTEGER,
2671 .arg1_type = ARG_PTR_TO_CTX,
2672 };
2673
2674 static const struct bpf_func_proto *
2675 bpf_base_func_proto(enum bpf_func_id func_id)
2676 {
2677 switch (func_id) {
2678 case BPF_FUNC_map_lookup_elem:
2679 return &bpf_map_lookup_elem_proto;
2680 case BPF_FUNC_map_update_elem:
2681 return &bpf_map_update_elem_proto;
2682 case BPF_FUNC_map_delete_elem:
2683 return &bpf_map_delete_elem_proto;
2684 case BPF_FUNC_get_prandom_u32:
2685 return &bpf_get_prandom_u32_proto;
2686 case BPF_FUNC_get_smp_processor_id:
2687 return &bpf_get_raw_smp_processor_id_proto;
2688 case BPF_FUNC_get_numa_node_id:
2689 return &bpf_get_numa_node_id_proto;
2690 case BPF_FUNC_tail_call:
2691 return &bpf_tail_call_proto;
2692 case BPF_FUNC_ktime_get_ns:
2693 return &bpf_ktime_get_ns_proto;
2694 case BPF_FUNC_trace_printk:
2695 if (capable(CAP_SYS_ADMIN))
2696 return bpf_get_trace_printk_proto();
2697 default:
2698 return NULL;
2699 }
2700 }
2701
2702 static const struct bpf_func_proto *
2703 sk_filter_func_proto(enum bpf_func_id func_id)
2704 {
2705 switch (func_id) {
2706 case BPF_FUNC_skb_load_bytes:
2707 return &bpf_skb_load_bytes_proto;
2708 case BPF_FUNC_get_socket_cookie:
2709 return &bpf_get_socket_cookie_proto;
2710 case BPF_FUNC_get_socket_uid:
2711 return &bpf_get_socket_uid_proto;
2712 default:
2713 return bpf_base_func_proto(func_id);
2714 }
2715 }
2716
2717 static const struct bpf_func_proto *
2718 tc_cls_act_func_proto(enum bpf_func_id func_id)
2719 {
2720 switch (func_id) {
2721 case BPF_FUNC_skb_store_bytes:
2722 return &bpf_skb_store_bytes_proto;
2723 case BPF_FUNC_skb_load_bytes:
2724 return &bpf_skb_load_bytes_proto;
2725 case BPF_FUNC_skb_pull_data:
2726 return &bpf_skb_pull_data_proto;
2727 case BPF_FUNC_csum_diff:
2728 return &bpf_csum_diff_proto;
2729 case BPF_FUNC_csum_update:
2730 return &bpf_csum_update_proto;
2731 case BPF_FUNC_l3_csum_replace:
2732 return &bpf_l3_csum_replace_proto;
2733 case BPF_FUNC_l4_csum_replace:
2734 return &bpf_l4_csum_replace_proto;
2735 case BPF_FUNC_clone_redirect:
2736 return &bpf_clone_redirect_proto;
2737 case BPF_FUNC_get_cgroup_classid:
2738 return &bpf_get_cgroup_classid_proto;
2739 case BPF_FUNC_skb_vlan_push:
2740 return &bpf_skb_vlan_push_proto;
2741 case BPF_FUNC_skb_vlan_pop:
2742 return &bpf_skb_vlan_pop_proto;
2743 case BPF_FUNC_skb_change_proto:
2744 return &bpf_skb_change_proto_proto;
2745 case BPF_FUNC_skb_change_type:
2746 return &bpf_skb_change_type_proto;
2747 case BPF_FUNC_skb_change_tail:
2748 return &bpf_skb_change_tail_proto;
2749 case BPF_FUNC_skb_get_tunnel_key:
2750 return &bpf_skb_get_tunnel_key_proto;
2751 case BPF_FUNC_skb_set_tunnel_key:
2752 return bpf_get_skb_set_tunnel_proto(func_id);
2753 case BPF_FUNC_skb_get_tunnel_opt:
2754 return &bpf_skb_get_tunnel_opt_proto;
2755 case BPF_FUNC_skb_set_tunnel_opt:
2756 return bpf_get_skb_set_tunnel_proto(func_id);
2757 case BPF_FUNC_redirect:
2758 return &bpf_redirect_proto;
2759 case BPF_FUNC_get_route_realm:
2760 return &bpf_get_route_realm_proto;
2761 case BPF_FUNC_get_hash_recalc:
2762 return &bpf_get_hash_recalc_proto;
2763 case BPF_FUNC_set_hash_invalid:
2764 return &bpf_set_hash_invalid_proto;
2765 case BPF_FUNC_set_hash:
2766 return &bpf_set_hash_proto;
2767 case BPF_FUNC_perf_event_output:
2768 return &bpf_skb_event_output_proto;
2769 case BPF_FUNC_get_smp_processor_id:
2770 return &bpf_get_smp_processor_id_proto;
2771 case BPF_FUNC_skb_under_cgroup:
2772 return &bpf_skb_under_cgroup_proto;
2773 case BPF_FUNC_get_socket_cookie:
2774 return &bpf_get_socket_cookie_proto;
2775 case BPF_FUNC_get_socket_uid:
2776 return &bpf_get_socket_uid_proto;
2777 default:
2778 return bpf_base_func_proto(func_id);
2779 }
2780 }
2781
2782 static const struct bpf_func_proto *
2783 xdp_func_proto(enum bpf_func_id func_id)
2784 {
2785 switch (func_id) {
2786 case BPF_FUNC_perf_event_output:
2787 return &bpf_xdp_event_output_proto;
2788 case BPF_FUNC_get_smp_processor_id:
2789 return &bpf_get_smp_processor_id_proto;
2790 case BPF_FUNC_xdp_adjust_head:
2791 return &bpf_xdp_adjust_head_proto;
2792 default:
2793 return bpf_base_func_proto(func_id);
2794 }
2795 }
2796
2797 static const struct bpf_func_proto *
2798 lwt_inout_func_proto(enum bpf_func_id func_id)
2799 {
2800 switch (func_id) {
2801 case BPF_FUNC_skb_load_bytes:
2802 return &bpf_skb_load_bytes_proto;
2803 case BPF_FUNC_skb_pull_data:
2804 return &bpf_skb_pull_data_proto;
2805 case BPF_FUNC_csum_diff:
2806 return &bpf_csum_diff_proto;
2807 case BPF_FUNC_get_cgroup_classid:
2808 return &bpf_get_cgroup_classid_proto;
2809 case BPF_FUNC_get_route_realm:
2810 return &bpf_get_route_realm_proto;
2811 case BPF_FUNC_get_hash_recalc:
2812 return &bpf_get_hash_recalc_proto;
2813 case BPF_FUNC_perf_event_output:
2814 return &bpf_skb_event_output_proto;
2815 case BPF_FUNC_get_smp_processor_id:
2816 return &bpf_get_smp_processor_id_proto;
2817 case BPF_FUNC_skb_under_cgroup:
2818 return &bpf_skb_under_cgroup_proto;
2819 default:
2820 return bpf_base_func_proto(func_id);
2821 }
2822 }
2823
2824 static const struct bpf_func_proto *
2825 lwt_xmit_func_proto(enum bpf_func_id func_id)
2826 {
2827 switch (func_id) {
2828 case BPF_FUNC_skb_get_tunnel_key:
2829 return &bpf_skb_get_tunnel_key_proto;
2830 case BPF_FUNC_skb_set_tunnel_key:
2831 return bpf_get_skb_set_tunnel_proto(func_id);
2832 case BPF_FUNC_skb_get_tunnel_opt:
2833 return &bpf_skb_get_tunnel_opt_proto;
2834 case BPF_FUNC_skb_set_tunnel_opt:
2835 return bpf_get_skb_set_tunnel_proto(func_id);
2836 case BPF_FUNC_redirect:
2837 return &bpf_redirect_proto;
2838 case BPF_FUNC_clone_redirect:
2839 return &bpf_clone_redirect_proto;
2840 case BPF_FUNC_skb_change_tail:
2841 return &bpf_skb_change_tail_proto;
2842 case BPF_FUNC_skb_change_head:
2843 return &bpf_skb_change_head_proto;
2844 case BPF_FUNC_skb_store_bytes:
2845 return &bpf_skb_store_bytes_proto;
2846 case BPF_FUNC_csum_update:
2847 return &bpf_csum_update_proto;
2848 case BPF_FUNC_l3_csum_replace:
2849 return &bpf_l3_csum_replace_proto;
2850 case BPF_FUNC_l4_csum_replace:
2851 return &bpf_l4_csum_replace_proto;
2852 case BPF_FUNC_set_hash_invalid:
2853 return &bpf_set_hash_invalid_proto;
2854 default:
2855 return lwt_inout_func_proto(func_id);
2856 }
2857 }
2858
2859 static bool __is_valid_access(int off, int size, enum bpf_access_type type,
2860 int *ctx_field_size)
2861 {
2862 if (off < 0 || off >= sizeof(struct __sk_buff))
2863 return false;
2864
2865 /* The verifier guarantees that size > 0. */
2866 if (off % size != 0)
2867 return false;
2868
2869 switch (off) {
2870 case offsetof(struct __sk_buff, cb[0]) ...
2871 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
2872 if (off + size >
2873 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32))
2874 return false;
2875 break;
2876 case offsetof(struct __sk_buff, data) ...
2877 offsetof(struct __sk_buff, data) + sizeof(__u32) - 1:
2878 case offsetof(struct __sk_buff, data_end) ...
2879 offsetof(struct __sk_buff, data_end) + sizeof(__u32) - 1:
2880 if (size != sizeof(__u32))
2881 return false;
2882 break;
2883 default:
2884 /* permit narrower load for not cb/data/data_end fields */
2885 *ctx_field_size = 4;
2886 if (type == BPF_WRITE) {
2887 if (size != sizeof(__u32))
2888 return false;
2889 } else {
2890 if (size != sizeof(__u32))
2891 #ifdef __LITTLE_ENDIAN
2892 return (off & 0x3) == 0 && (size == 1 || size == 2);
2893 #else
2894 return (off & 0x3) + size == 4 && (size == 1 || size == 2);
2895 #endif
2896 }
2897 }
2898
2899 return true;
2900 }
2901
2902 static bool sk_filter_is_valid_access(int off, int size,
2903 enum bpf_access_type type,
2904 enum bpf_reg_type *reg_type,
2905 int *ctx_field_size)
2906 {
2907 switch (off) {
2908 case offsetof(struct __sk_buff, tc_classid) ...
2909 offsetof(struct __sk_buff, tc_classid) + sizeof(__u32) - 1:
2910 case offsetof(struct __sk_buff, data) ...
2911 offsetof(struct __sk_buff, data) + sizeof(__u32) - 1:
2912 case offsetof(struct __sk_buff, data_end) ...
2913 offsetof(struct __sk_buff, data_end) + sizeof(__u32) - 1:
2914 return false;
2915 }
2916
2917 if (type == BPF_WRITE) {
2918 switch (off) {
2919 case offsetof(struct __sk_buff, cb[0]) ...
2920 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
2921 break;
2922 default:
2923 return false;
2924 }
2925 }
2926
2927 return __is_valid_access(off, size, type, ctx_field_size);
2928 }
2929
2930 static bool lwt_is_valid_access(int off, int size,
2931 enum bpf_access_type type,
2932 enum bpf_reg_type *reg_type,
2933 int *ctx_field_size)
2934 {
2935 switch (off) {
2936 case offsetof(struct __sk_buff, tc_classid) ...
2937 offsetof(struct __sk_buff, tc_classid) + sizeof(__u32) - 1:
2938 return false;
2939 }
2940
2941 if (type == BPF_WRITE) {
2942 switch (off) {
2943 case offsetof(struct __sk_buff, mark):
2944 case offsetof(struct __sk_buff, priority):
2945 case offsetof(struct __sk_buff, cb[0]) ...
2946 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
2947 break;
2948 default:
2949 return false;
2950 }
2951 }
2952
2953 switch (off) {
2954 case offsetof(struct __sk_buff, data):
2955 *reg_type = PTR_TO_PACKET;
2956 break;
2957 case offsetof(struct __sk_buff, data_end):
2958 *reg_type = PTR_TO_PACKET_END;
2959 break;
2960 }
2961
2962 return __is_valid_access(off, size, type, ctx_field_size);
2963 }
2964
2965 static bool sock_filter_is_valid_access(int off, int size,
2966 enum bpf_access_type type,
2967 enum bpf_reg_type *reg_type,
2968 int *ctx_field_size)
2969 {
2970 if (type == BPF_WRITE) {
2971 switch (off) {
2972 case offsetof(struct bpf_sock, bound_dev_if):
2973 break;
2974 default:
2975 return false;
2976 }
2977 }
2978
2979 if (off < 0 || off + size > sizeof(struct bpf_sock))
2980 return false;
2981 /* The verifier guarantees that size > 0. */
2982 if (off % size != 0)
2983 return false;
2984 if (size != sizeof(__u32))
2985 return false;
2986
2987 return true;
2988 }
2989
2990 static int tc_cls_act_prologue(struct bpf_insn *insn_buf, bool direct_write,
2991 const struct bpf_prog *prog)
2992 {
2993 struct bpf_insn *insn = insn_buf;
2994
2995 if (!direct_write)
2996 return 0;
2997
2998 /* if (!skb->cloned)
2999 * goto start;
3000 *
3001 * (Fast-path, otherwise approximation that we might be
3002 * a clone, do the rest in helper.)
3003 */
3004 *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_6, BPF_REG_1, CLONED_OFFSET());
3005 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_6, CLONED_MASK);
3006 *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_6, 0, 7);
3007
3008 /* ret = bpf_skb_pull_data(skb, 0); */
3009 *insn++ = BPF_MOV64_REG(BPF_REG_6, BPF_REG_1);
3010 *insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_2, BPF_REG_2);
3011 *insn++ = BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
3012 BPF_FUNC_skb_pull_data);
3013 /* if (!ret)
3014 * goto restore;
3015 * return TC_ACT_SHOT;
3016 */
3017 *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2);
3018 *insn++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_0, TC_ACT_SHOT);
3019 *insn++ = BPF_EXIT_INSN();
3020
3021 /* restore: */
3022 *insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_6);
3023 /* start: */
3024 *insn++ = prog->insnsi[0];
3025
3026 return insn - insn_buf;
3027 }
3028
3029 static bool tc_cls_act_is_valid_access(int off, int size,
3030 enum bpf_access_type type,
3031 enum bpf_reg_type *reg_type,
3032 int *ctx_field_size)
3033 {
3034 if (type == BPF_WRITE) {
3035 switch (off) {
3036 case offsetof(struct __sk_buff, mark):
3037 case offsetof(struct __sk_buff, tc_index):
3038 case offsetof(struct __sk_buff, priority):
3039 case offsetof(struct __sk_buff, cb[0]) ...
3040 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
3041 case offsetof(struct __sk_buff, tc_classid):
3042 break;
3043 default:
3044 return false;
3045 }
3046 }
3047
3048 switch (off) {
3049 case offsetof(struct __sk_buff, data):
3050 *reg_type = PTR_TO_PACKET;
3051 break;
3052 case offsetof(struct __sk_buff, data_end):
3053 *reg_type = PTR_TO_PACKET_END;
3054 break;
3055 }
3056
3057 return __is_valid_access(off, size, type, ctx_field_size);
3058 }
3059
3060 static bool __is_valid_xdp_access(int off, int size)
3061 {
3062 if (off < 0 || off >= sizeof(struct xdp_md))
3063 return false;
3064 if (off % size != 0)
3065 return false;
3066 if (size != sizeof(__u32))
3067 return false;
3068
3069 return true;
3070 }
3071
3072 static bool xdp_is_valid_access(int off, int size,
3073 enum bpf_access_type type,
3074 enum bpf_reg_type *reg_type,
3075 int *ctx_field_size)
3076 {
3077 if (type == BPF_WRITE)
3078 return false;
3079
3080 switch (off) {
3081 case offsetof(struct xdp_md, data):
3082 *reg_type = PTR_TO_PACKET;
3083 break;
3084 case offsetof(struct xdp_md, data_end):
3085 *reg_type = PTR_TO_PACKET_END;
3086 break;
3087 }
3088
3089 return __is_valid_xdp_access(off, size);
3090 }
3091
3092 void bpf_warn_invalid_xdp_action(u32 act)
3093 {
3094 WARN_ONCE(1, "Illegal XDP return value %u, expect packet loss\n", act);
3095 }
3096 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
3097
3098 static u32 bpf_convert_ctx_access(enum bpf_access_type type,
3099 const struct bpf_insn *si,
3100 struct bpf_insn *insn_buf,
3101 struct bpf_prog *prog)
3102 {
3103 struct bpf_insn *insn = insn_buf;
3104 int off;
3105
3106 switch (si->off) {
3107 case offsetof(struct __sk_buff, len):
3108 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
3109
3110 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3111 offsetof(struct sk_buff, len));
3112 break;
3113
3114 case offsetof(struct __sk_buff, protocol):
3115 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
3116
3117 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3118 offsetof(struct sk_buff, protocol));
3119 break;
3120
3121 case offsetof(struct __sk_buff, vlan_proto):
3122 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
3123
3124 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3125 offsetof(struct sk_buff, vlan_proto));
3126 break;
3127
3128 case offsetof(struct __sk_buff, priority):
3129 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
3130
3131 if (type == BPF_WRITE)
3132 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3133 offsetof(struct sk_buff, priority));
3134 else
3135 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3136 offsetof(struct sk_buff, priority));
3137 break;
3138
3139 case offsetof(struct __sk_buff, ingress_ifindex):
3140 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
3141
3142 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3143 offsetof(struct sk_buff, skb_iif));
3144 break;
3145
3146 case offsetof(struct __sk_buff, ifindex):
3147 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
3148
3149 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
3150 si->dst_reg, si->src_reg,
3151 offsetof(struct sk_buff, dev));
3152 *insn++ = BPF_JMP_IMM(BPF_JEQ, si->dst_reg, 0, 1);
3153 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3154 offsetof(struct net_device, ifindex));
3155 break;
3156
3157 case offsetof(struct __sk_buff, hash):
3158 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
3159
3160 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3161 offsetof(struct sk_buff, hash));
3162 break;
3163
3164 case offsetof(struct __sk_buff, mark):
3165 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
3166
3167 if (type == BPF_WRITE)
3168 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3169 offsetof(struct sk_buff, mark));
3170 else
3171 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3172 offsetof(struct sk_buff, mark));
3173 break;
3174
3175 case offsetof(struct __sk_buff, pkt_type):
3176 return convert_skb_access(SKF_AD_PKTTYPE, si->dst_reg,
3177 si->src_reg, insn);
3178
3179 case offsetof(struct __sk_buff, queue_mapping):
3180 return convert_skb_access(SKF_AD_QUEUE, si->dst_reg,
3181 si->src_reg, insn);
3182
3183 case offsetof(struct __sk_buff, vlan_present):
3184 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
3185 si->dst_reg, si->src_reg, insn);
3186
3187 case offsetof(struct __sk_buff, vlan_tci):
3188 return convert_skb_access(SKF_AD_VLAN_TAG,
3189 si->dst_reg, si->src_reg, insn);
3190
3191 case offsetof(struct __sk_buff, cb[0]) ...
3192 offsetof(struct __sk_buff, cb[4]) + sizeof(__u32) - 1:
3193 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
3194 BUILD_BUG_ON((offsetof(struct sk_buff, cb) +
3195 offsetof(struct qdisc_skb_cb, data)) %
3196 sizeof(__u64));
3197
3198 prog->cb_access = 1;
3199 off = si->off;
3200 off -= offsetof(struct __sk_buff, cb[0]);
3201 off += offsetof(struct sk_buff, cb);
3202 off += offsetof(struct qdisc_skb_cb, data);
3203 if (type == BPF_WRITE)
3204 *insn++ = BPF_STX_MEM(BPF_SIZE(si->code), si->dst_reg,
3205 si->src_reg, off);
3206 else
3207 *insn++ = BPF_LDX_MEM(BPF_SIZE(si->code), si->dst_reg,
3208 si->src_reg, off);
3209 break;
3210
3211 case offsetof(struct __sk_buff, tc_classid):
3212 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, tc_classid) != 2);
3213
3214 off = si->off;
3215 off -= offsetof(struct __sk_buff, tc_classid);
3216 off += offsetof(struct sk_buff, cb);
3217 off += offsetof(struct qdisc_skb_cb, tc_classid);
3218 if (type == BPF_WRITE)
3219 *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg,
3220 si->src_reg, off);
3221 else
3222 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg,
3223 si->src_reg, off);
3224 break;
3225
3226 case offsetof(struct __sk_buff, data):
3227 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, data),
3228 si->dst_reg, si->src_reg,
3229 offsetof(struct sk_buff, data));
3230 break;
3231
3232 case offsetof(struct __sk_buff, data_end):
3233 off = si->off;
3234 off -= offsetof(struct __sk_buff, data_end);
3235 off += offsetof(struct sk_buff, cb);
3236 off += offsetof(struct bpf_skb_data_end, data_end);
3237 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
3238 si->src_reg, off);
3239 break;
3240
3241 case offsetof(struct __sk_buff, tc_index):
3242 #ifdef CONFIG_NET_SCHED
3243 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
3244
3245 if (type == BPF_WRITE)
3246 *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg, si->src_reg,
3247 offsetof(struct sk_buff, tc_index));
3248 else
3249 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3250 offsetof(struct sk_buff, tc_index));
3251 #else
3252 if (type == BPF_WRITE)
3253 *insn++ = BPF_MOV64_REG(si->dst_reg, si->dst_reg);
3254 else
3255 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3256 #endif
3257 break;
3258
3259 case offsetof(struct __sk_buff, napi_id):
3260 #if defined(CONFIG_NET_RX_BUSY_POLL)
3261 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, napi_id) != 4);
3262
3263 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3264 offsetof(struct sk_buff, napi_id));
3265 *insn++ = BPF_JMP_IMM(BPF_JGE, si->dst_reg, MIN_NAPI_ID, 1);
3266 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3267 #else
3268 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
3269 #endif
3270 break;
3271 }
3272
3273 return insn - insn_buf;
3274 }
3275
3276 static u32 sock_filter_convert_ctx_access(enum bpf_access_type type,
3277 const struct bpf_insn *si,
3278 struct bpf_insn *insn_buf,
3279 struct bpf_prog *prog)
3280 {
3281 struct bpf_insn *insn = insn_buf;
3282
3283 switch (si->off) {
3284 case offsetof(struct bpf_sock, bound_dev_if):
3285 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_bound_dev_if) != 4);
3286
3287 if (type == BPF_WRITE)
3288 *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3289 offsetof(struct sock, sk_bound_dev_if));
3290 else
3291 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3292 offsetof(struct sock, sk_bound_dev_if));
3293 break;
3294
3295 case offsetof(struct bpf_sock, family):
3296 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_family) != 2);
3297
3298 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3299 offsetof(struct sock, sk_family));
3300 break;
3301
3302 case offsetof(struct bpf_sock, type):
3303 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3304 offsetof(struct sock, __sk_flags_offset));
3305 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_TYPE_MASK);
3306 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_TYPE_SHIFT);
3307 break;
3308
3309 case offsetof(struct bpf_sock, protocol):
3310 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3311 offsetof(struct sock, __sk_flags_offset));
3312 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_PROTO_MASK);
3313 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_PROTO_SHIFT);
3314 break;
3315 }
3316
3317 return insn - insn_buf;
3318 }
3319
3320 static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type,
3321 const struct bpf_insn *si,
3322 struct bpf_insn *insn_buf,
3323 struct bpf_prog *prog)
3324 {
3325 struct bpf_insn *insn = insn_buf;
3326
3327 switch (si->off) {
3328 case offsetof(struct __sk_buff, ifindex):
3329 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
3330
3331 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
3332 si->dst_reg, si->src_reg,
3333 offsetof(struct sk_buff, dev));
3334 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3335 offsetof(struct net_device, ifindex));
3336 break;
3337 default:
3338 return bpf_convert_ctx_access(type, si, insn_buf, prog);
3339 }
3340
3341 return insn - insn_buf;
3342 }
3343
3344 static u32 xdp_convert_ctx_access(enum bpf_access_type type,
3345 const struct bpf_insn *si,
3346 struct bpf_insn *insn_buf,
3347 struct bpf_prog *prog)
3348 {
3349 struct bpf_insn *insn = insn_buf;
3350
3351 switch (si->off) {
3352 case offsetof(struct xdp_md, data):
3353 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data),
3354 si->dst_reg, si->src_reg,
3355 offsetof(struct xdp_buff, data));
3356 break;
3357 case offsetof(struct xdp_md, data_end):
3358 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_end),
3359 si->dst_reg, si->src_reg,
3360 offsetof(struct xdp_buff, data_end));
3361 break;
3362 }
3363
3364 return insn - insn_buf;
3365 }
3366
3367 const struct bpf_verifier_ops sk_filter_prog_ops = {
3368 .get_func_proto = sk_filter_func_proto,
3369 .is_valid_access = sk_filter_is_valid_access,
3370 .convert_ctx_access = bpf_convert_ctx_access,
3371 };
3372
3373 const struct bpf_verifier_ops tc_cls_act_prog_ops = {
3374 .get_func_proto = tc_cls_act_func_proto,
3375 .is_valid_access = tc_cls_act_is_valid_access,
3376 .convert_ctx_access = tc_cls_act_convert_ctx_access,
3377 .gen_prologue = tc_cls_act_prologue,
3378 .test_run = bpf_prog_test_run_skb,
3379 };
3380
3381 const struct bpf_verifier_ops xdp_prog_ops = {
3382 .get_func_proto = xdp_func_proto,
3383 .is_valid_access = xdp_is_valid_access,
3384 .convert_ctx_access = xdp_convert_ctx_access,
3385 .test_run = bpf_prog_test_run_xdp,
3386 };
3387
3388 const struct bpf_verifier_ops cg_skb_prog_ops = {
3389 .get_func_proto = sk_filter_func_proto,
3390 .is_valid_access = sk_filter_is_valid_access,
3391 .convert_ctx_access = bpf_convert_ctx_access,
3392 .test_run = bpf_prog_test_run_skb,
3393 };
3394
3395 const struct bpf_verifier_ops lwt_inout_prog_ops = {
3396 .get_func_proto = lwt_inout_func_proto,
3397 .is_valid_access = lwt_is_valid_access,
3398 .convert_ctx_access = bpf_convert_ctx_access,
3399 .test_run = bpf_prog_test_run_skb,
3400 };
3401
3402 const struct bpf_verifier_ops lwt_xmit_prog_ops = {
3403 .get_func_proto = lwt_xmit_func_proto,
3404 .is_valid_access = lwt_is_valid_access,
3405 .convert_ctx_access = bpf_convert_ctx_access,
3406 .gen_prologue = tc_cls_act_prologue,
3407 .test_run = bpf_prog_test_run_skb,
3408 };
3409
3410 const struct bpf_verifier_ops cg_sock_prog_ops = {
3411 .get_func_proto = bpf_base_func_proto,
3412 .is_valid_access = sock_filter_is_valid_access,
3413 .convert_ctx_access = sock_filter_convert_ctx_access,
3414 };
3415
3416 int sk_detach_filter(struct sock *sk)
3417 {
3418 int ret = -ENOENT;
3419 struct sk_filter *filter;
3420
3421 if (sock_flag(sk, SOCK_FILTER_LOCKED))
3422 return -EPERM;
3423
3424 filter = rcu_dereference_protected(sk->sk_filter,
3425 lockdep_sock_is_held(sk));
3426 if (filter) {
3427 RCU_INIT_POINTER(sk->sk_filter, NULL);
3428 sk_filter_uncharge(sk, filter);
3429 ret = 0;
3430 }
3431
3432 return ret;
3433 }
3434 EXPORT_SYMBOL_GPL(sk_detach_filter);
3435
3436 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
3437 unsigned int len)
3438 {
3439 struct sock_fprog_kern *fprog;
3440 struct sk_filter *filter;
3441 int ret = 0;
3442
3443 lock_sock(sk);
3444 filter = rcu_dereference_protected(sk->sk_filter,
3445 lockdep_sock_is_held(sk));
3446 if (!filter)
3447 goto out;
3448
3449 /* We're copying the filter that has been originally attached,
3450 * so no conversion/decode needed anymore. eBPF programs that
3451 * have no original program cannot be dumped through this.
3452 */
3453 ret = -EACCES;
3454 fprog = filter->prog->orig_prog;
3455 if (!fprog)
3456 goto out;
3457
3458 ret = fprog->len;
3459 if (!len)
3460 /* User space only enquires number of filter blocks. */
3461 goto out;
3462
3463 ret = -EINVAL;
3464 if (len < fprog->len)
3465 goto out;
3466
3467 ret = -EFAULT;
3468 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
3469 goto out;
3470
3471 /* Instead of bytes, the API requests to return the number
3472 * of filter blocks.
3473 */
3474 ret = fprog->len;
3475 out:
3476 release_sock(sk);
3477 return ret;
3478 }
(deprecated) Btrfs devel tree