projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'bpf-csum-complete'
[deliverable/linux.git] / net / core / filter.c
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/in.h>
30 #include <linux/inet.h>
31 #include <linux/netdevice.h>
32 #include <linux/if_packet.h>
33 #include <linux/gfp.h>
34 #include <net/ip.h>
35 #include <net/protocol.h>
36 #include <net/netlink.h>
37 #include <linux/skbuff.h>
38 #include <net/sock.h>
39 #include <net/flow_dissector.h>
40 #include <linux/errno.h>
41 #include <linux/timer.h>
42 #include <asm/uaccess.h>
43 #include <asm/unaligned.h>
44 #include <linux/filter.h>
45 #include <linux/ratelimit.h>
46 #include <linux/seccomp.h>
47 #include <linux/if_vlan.h>
48 #include <linux/bpf.h>
49 #include <net/sch_generic.h>
50 #include <net/cls_cgroup.h>
51 #include <net/dst_metadata.h>
52 #include <net/dst.h>
53 #include <net/sock_reuseport.h>
54
55 /**
56 * sk_filter_trim_cap - run a packet through a socket filter
57 * @sk: sock associated with &sk_buff
58 * @skb: buffer to filter
59 * @cap: limit on how short the eBPF program may trim the packet
60 *
61 * Run the eBPF program and then cut skb->data to correct size returned by
62 * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
63 * than pkt_len we keep whole skb->data. This is the socket level
64 * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
65 * be accepted or -EPERM if the packet should be tossed.
66 *
67 */
68 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
69 {
70 int err;
71 struct sk_filter *filter;
72
73 /*
74 * If the skb was allocated from pfmemalloc reserves, only
75 * allow SOCK_MEMALLOC sockets to use it as this socket is
76 * helping free memory
77 */
78 if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC))
79 return -ENOMEM;
80
81 err = security_sock_rcv_skb(sk, skb);
82 if (err)
83 return err;
84
85 rcu_read_lock();
86 filter = rcu_dereference(sk->sk_filter);
87 if (filter) {
88 unsigned int pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
89 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
90 }
91 rcu_read_unlock();
92
93 return err;
94 }
95 EXPORT_SYMBOL(sk_filter_trim_cap);
96
97 static u64 __skb_get_pay_offset(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
98 {
99 return skb_get_poff((struct sk_buff *)(unsigned long) ctx);
100 }
101
102 static u64 __skb_get_nlattr(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
103 {
104 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
105 struct nlattr *nla;
106
107 if (skb_is_nonlinear(skb))
108 return 0;
109
110 if (skb->len < sizeof(struct nlattr))
111 return 0;
112
113 if (a > skb->len - sizeof(struct nlattr))
114 return 0;
115
116 nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
117 if (nla)
118 return (void *) nla - (void *) skb->data;
119
120 return 0;
121 }
122
123 static u64 __skb_get_nlattr_nest(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
124 {
125 struct sk_buff *skb = (struct sk_buff *)(unsigned long) ctx;
126 struct nlattr *nla;
127
128 if (skb_is_nonlinear(skb))
129 return 0;
130
131 if (skb->len < sizeof(struct nlattr))
132 return 0;
133
134 if (a > skb->len - sizeof(struct nlattr))
135 return 0;
136
137 nla = (struct nlattr *) &skb->data[a];
138 if (nla->nla_len > skb->len - a)
139 return 0;
140
141 nla = nla_find_nested(nla, x);
142 if (nla)
143 return (void *) nla - (void *) skb->data;
144
145 return 0;
146 }
147
148 static u64 __get_raw_cpu_id(u64 ctx, u64 a, u64 x, u64 r4, u64 r5)
149 {
150 return raw_smp_processor_id();
151 }
152
153 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
154 .func = __get_raw_cpu_id,
155 .gpl_only = false,
156 .ret_type = RET_INTEGER,
157 };
158
159 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
160 struct bpf_insn *insn_buf)
161 {
162 struct bpf_insn *insn = insn_buf;
163
164 switch (skb_field) {
165 case SKF_AD_MARK:
166 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
167
168 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
169 offsetof(struct sk_buff, mark));
170 break;
171
172 case SKF_AD_PKTTYPE:
173 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
174 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
175 #ifdef __BIG_ENDIAN_BITFIELD
176 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
177 #endif
178 break;
179
180 case SKF_AD_QUEUE:
181 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
182
183 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
184 offsetof(struct sk_buff, queue_mapping));
185 break;
186
187 case SKF_AD_VLAN_TAG:
188 case SKF_AD_VLAN_TAG_PRESENT:
189 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
190 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
191
192 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
193 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
194 offsetof(struct sk_buff, vlan_tci));
195 if (skb_field == SKF_AD_VLAN_TAG) {
196 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
197 ~VLAN_TAG_PRESENT);
198 } else {
199 /* dst_reg >>= 12 */
200 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
201 /* dst_reg &= 1 */
202 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
203 }
204 break;
205 }
206
207 return insn - insn_buf;
208 }
209
210 static bool convert_bpf_extensions(struct sock_filter *fp,
211 struct bpf_insn **insnp)
212 {
213 struct bpf_insn *insn = *insnp;
214 u32 cnt;
215
216 switch (fp->k) {
217 case SKF_AD_OFF + SKF_AD_PROTOCOL:
218 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
219
220 /* A = *(u16 *) (CTX + offsetof(protocol)) */
221 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
222 offsetof(struct sk_buff, protocol));
223 /* A = ntohs(A) [emitting a nop or swap16] */
224 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
225 break;
226
227 case SKF_AD_OFF + SKF_AD_PKTTYPE:
228 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
229 insn += cnt - 1;
230 break;
231
232 case SKF_AD_OFF + SKF_AD_IFINDEX:
233 case SKF_AD_OFF + SKF_AD_HATYPE:
234 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
235 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
236 BUILD_BUG_ON(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)) < 0);
237
238 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
239 BPF_REG_TMP, BPF_REG_CTX,
240 offsetof(struct sk_buff, dev));
241 /* if (tmp != 0) goto pc + 1 */
242 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
243 *insn++ = BPF_EXIT_INSN();
244 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
245 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
246 offsetof(struct net_device, ifindex));
247 else
248 *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
249 offsetof(struct net_device, type));
250 break;
251
252 case SKF_AD_OFF + SKF_AD_MARK:
253 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
254 insn += cnt - 1;
255 break;
256
257 case SKF_AD_OFF + SKF_AD_RXHASH:
258 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
259
260 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
261 offsetof(struct sk_buff, hash));
262 break;
263
264 case SKF_AD_OFF + SKF_AD_QUEUE:
265 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
266 insn += cnt - 1;
267 break;
268
269 case SKF_AD_OFF + SKF_AD_VLAN_TAG:
270 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
271 BPF_REG_A, BPF_REG_CTX, insn);
272 insn += cnt - 1;
273 break;
274
275 case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
276 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
277 BPF_REG_A, BPF_REG_CTX, insn);
278 insn += cnt - 1;
279 break;
280
281 case SKF_AD_OFF + SKF_AD_VLAN_TPID:
282 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
283
284 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
285 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
286 offsetof(struct sk_buff, vlan_proto));
287 /* A = ntohs(A) [emitting a nop or swap16] */
288 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
289 break;
290
291 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
292 case SKF_AD_OFF + SKF_AD_NLATTR:
293 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
294 case SKF_AD_OFF + SKF_AD_CPU:
295 case SKF_AD_OFF + SKF_AD_RANDOM:
296 /* arg1 = CTX */
297 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
298 /* arg2 = A */
299 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
300 /* arg3 = X */
301 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
302 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
303 switch (fp->k) {
304 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
305 *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
306 break;
307 case SKF_AD_OFF + SKF_AD_NLATTR:
308 *insn = BPF_EMIT_CALL(__skb_get_nlattr);
309 break;
310 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
311 *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
312 break;
313 case SKF_AD_OFF + SKF_AD_CPU:
314 *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
315 break;
316 case SKF_AD_OFF + SKF_AD_RANDOM:
317 *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
318 bpf_user_rnd_init_once();
319 break;
320 }
321 break;
322
323 case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
324 /* A ^= X */
325 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
326 break;
327
328 default:
329 /* This is just a dummy call to avoid letting the compiler
330 * evict __bpf_call_base() as an optimization. Placed here
331 * where no-one bothers.
332 */
333 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
334 return false;
335 }
336
337 *insnp = insn;
338 return true;
339 }
340
341 /**
342 * bpf_convert_filter - convert filter program
343 * @prog: the user passed filter program
344 * @len: the length of the user passed filter program
345 * @new_prog: buffer where converted program will be stored
346 * @new_len: pointer to store length of converted program
347 *
348 * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style.
349 * Conversion workflow:
350 *
351 * 1) First pass for calculating the new program length:
352 * bpf_convert_filter(old_prog, old_len, NULL, &new_len)
353 *
354 * 2) 2nd pass to remap in two passes: 1st pass finds new
355 * jump offsets, 2nd pass remapping:
356 * new_prog = kmalloc(sizeof(struct bpf_insn) * new_len);
357 * bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
358 */
359 static int bpf_convert_filter(struct sock_filter *prog, int len,
360 struct bpf_insn *new_prog, int *new_len)
361 {
362 int new_flen = 0, pass = 0, target, i;
363 struct bpf_insn *new_insn;
364 struct sock_filter *fp;
365 int *addrs = NULL;
366 u8 bpf_src;
367
368 BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
369 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
370
371 if (len <= 0 || len > BPF_MAXINSNS)
372 return -EINVAL;
373
374 if (new_prog) {
375 addrs = kcalloc(len, sizeof(*addrs),
376 GFP_KERNEL | __GFP_NOWARN);
377 if (!addrs)
378 return -ENOMEM;
379 }
380
381 do_pass:
382 new_insn = new_prog;
383 fp = prog;
384
385 /* Classic BPF related prologue emission. */
386 if (new_insn) {
387 /* Classic BPF expects A and X to be reset first. These need
388 * to be guaranteed to be the first two instructions.
389 */
390 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
391 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
392
393 /* All programs must keep CTX in callee saved BPF_REG_CTX.
394 * In eBPF case it's done by the compiler, here we need to
395 * do this ourself. Initial CTX is present in BPF_REG_ARG1.
396 */
397 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
398 } else {
399 new_insn += 3;
400 }
401
402 for (i = 0; i < len; fp++, i++) {
403 struct bpf_insn tmp_insns[6] = { };
404 struct bpf_insn *insn = tmp_insns;
405
406 if (addrs)
407 addrs[i] = new_insn - new_prog;
408
409 switch (fp->code) {
410 /* All arithmetic insns and skb loads map as-is. */
411 case BPF_ALU | BPF_ADD | BPF_X:
412 case BPF_ALU | BPF_ADD | BPF_K:
413 case BPF_ALU | BPF_SUB | BPF_X:
414 case BPF_ALU | BPF_SUB | BPF_K:
415 case BPF_ALU | BPF_AND | BPF_X:
416 case BPF_ALU | BPF_AND | BPF_K:
417 case BPF_ALU | BPF_OR | BPF_X:
418 case BPF_ALU | BPF_OR | BPF_K:
419 case BPF_ALU | BPF_LSH | BPF_X:
420 case BPF_ALU | BPF_LSH | BPF_K:
421 case BPF_ALU | BPF_RSH | BPF_X:
422 case BPF_ALU | BPF_RSH | BPF_K:
423 case BPF_ALU | BPF_XOR | BPF_X:
424 case BPF_ALU | BPF_XOR | BPF_K:
425 case BPF_ALU | BPF_MUL | BPF_X:
426 case BPF_ALU | BPF_MUL | BPF_K:
427 case BPF_ALU | BPF_DIV | BPF_X:
428 case BPF_ALU | BPF_DIV | BPF_K:
429 case BPF_ALU | BPF_MOD | BPF_X:
430 case BPF_ALU | BPF_MOD | BPF_K:
431 case BPF_ALU | BPF_NEG:
432 case BPF_LD | BPF_ABS | BPF_W:
433 case BPF_LD | BPF_ABS | BPF_H:
434 case BPF_LD | BPF_ABS | BPF_B:
435 case BPF_LD | BPF_IND | BPF_W:
436 case BPF_LD | BPF_IND | BPF_H:
437 case BPF_LD | BPF_IND | BPF_B:
438 /* Check for overloaded BPF extension and
439 * directly convert it if found, otherwise
440 * just move on with mapping.
441 */
442 if (BPF_CLASS(fp->code) == BPF_LD &&
443 BPF_MODE(fp->code) == BPF_ABS &&
444 convert_bpf_extensions(fp, &insn))
445 break;
446
447 *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
448 break;
449
450 /* Jump transformation cannot use BPF block macros
451 * everywhere as offset calculation and target updates
452 * require a bit more work than the rest, i.e. jump
453 * opcodes map as-is, but offsets need adjustment.
454 */
455
456 #define BPF_EMIT_JMP \
457 do { \
458 if (target >= len || target < 0) \
459 goto err; \
460 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0; \
461 /* Adjust pc relative offset for 2nd or 3rd insn. */ \
462 insn->off -= insn - tmp_insns; \
463 } while (0)
464
465 case BPF_JMP | BPF_JA:
466 target = i + fp->k + 1;
467 insn->code = fp->code;
468 BPF_EMIT_JMP;
469 break;
470
471 case BPF_JMP | BPF_JEQ | BPF_K:
472 case BPF_JMP | BPF_JEQ | BPF_X:
473 case BPF_JMP | BPF_JSET | BPF_K:
474 case BPF_JMP | BPF_JSET | BPF_X:
475 case BPF_JMP | BPF_JGT | BPF_K:
476 case BPF_JMP | BPF_JGT | BPF_X:
477 case BPF_JMP | BPF_JGE | BPF_K:
478 case BPF_JMP | BPF_JGE | BPF_X:
479 if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
480 /* BPF immediates are signed, zero extend
481 * immediate into tmp register and use it
482 * in compare insn.
483 */
484 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
485
486 insn->dst_reg = BPF_REG_A;
487 insn->src_reg = BPF_REG_TMP;
488 bpf_src = BPF_X;
489 } else {
490 insn->dst_reg = BPF_REG_A;
491 insn->imm = fp->k;
492 bpf_src = BPF_SRC(fp->code);
493 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
494 }
495
496 /* Common case where 'jump_false' is next insn. */
497 if (fp->jf == 0) {
498 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
499 target = i + fp->jt + 1;
500 BPF_EMIT_JMP;
501 break;
502 }
503
504 /* Convert JEQ into JNE when 'jump_true' is next insn. */
505 if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
506 insn->code = BPF_JMP | BPF_JNE | bpf_src;
507 target = i + fp->jf + 1;
508 BPF_EMIT_JMP;
509 break;
510 }
511
512 /* Other jumps are mapped into two insns: Jxx and JA. */
513 target = i + fp->jt + 1;
514 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
515 BPF_EMIT_JMP;
516 insn++;
517
518 insn->code = BPF_JMP | BPF_JA;
519 target = i + fp->jf + 1;
520 BPF_EMIT_JMP;
521 break;
522
523 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
524 case BPF_LDX | BPF_MSH | BPF_B:
525 /* tmp = A */
526 *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
527 /* A = BPF_R0 = *(u8 *) (skb->data + K) */
528 *insn++ = BPF_LD_ABS(BPF_B, fp->k);
529 /* A &= 0xf */
530 *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
531 /* A <<= 2 */
532 *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
533 /* X = A */
534 *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
535 /* A = tmp */
536 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
537 break;
538
539 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
540 * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
541 */
542 case BPF_RET | BPF_A:
543 case BPF_RET | BPF_K:
544 if (BPF_RVAL(fp->code) == BPF_K)
545 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
546 0, fp->k);
547 *insn = BPF_EXIT_INSN();
548 break;
549
550 /* Store to stack. */
551 case BPF_ST:
552 case BPF_STX:
553 *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
554 BPF_ST ? BPF_REG_A : BPF_REG_X,
555 -(BPF_MEMWORDS - fp->k) * 4);
556 break;
557
558 /* Load from stack. */
559 case BPF_LD | BPF_MEM:
560 case BPF_LDX | BPF_MEM:
561 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
562 BPF_REG_A : BPF_REG_X, BPF_REG_FP,
563 -(BPF_MEMWORDS - fp->k) * 4);
564 break;
565
566 /* A = K or X = K */
567 case BPF_LD | BPF_IMM:
568 case BPF_LDX | BPF_IMM:
569 *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
570 BPF_REG_A : BPF_REG_X, fp->k);
571 break;
572
573 /* X = A */
574 case BPF_MISC | BPF_TAX:
575 *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
576 break;
577
578 /* A = X */
579 case BPF_MISC | BPF_TXA:
580 *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
581 break;
582
583 /* A = skb->len or X = skb->len */
584 case BPF_LD | BPF_W | BPF_LEN:
585 case BPF_LDX | BPF_W | BPF_LEN:
586 *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
587 BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
588 offsetof(struct sk_buff, len));
589 break;
590
591 /* Access seccomp_data fields. */
592 case BPF_LDX | BPF_ABS | BPF_W:
593 /* A = *(u32 *) (ctx + K) */
594 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
595 break;
596
597 /* Unknown instruction. */
598 default:
599 goto err;
600 }
601
602 insn++;
603 if (new_prog)
604 memcpy(new_insn, tmp_insns,
605 sizeof(*insn) * (insn - tmp_insns));
606 new_insn += insn - tmp_insns;
607 }
608
609 if (!new_prog) {
610 /* Only calculating new length. */
611 *new_len = new_insn - new_prog;
612 return 0;
613 }
614
615 pass++;
616 if (new_flen != new_insn - new_prog) {
617 new_flen = new_insn - new_prog;
618 if (pass > 2)
619 goto err;
620 goto do_pass;
621 }
622
623 kfree(addrs);
624 BUG_ON(*new_len != new_flen);
625 return 0;
626 err:
627 kfree(addrs);
628 return -EINVAL;
629 }
630
631 /* Security:
632 *
633 * As we dont want to clear mem[] array for each packet going through
634 * __bpf_prog_run(), we check that filter loaded by user never try to read
635 * a cell if not previously written, and we check all branches to be sure
636 * a malicious user doesn't try to abuse us.
637 */
638 static int check_load_and_stores(const struct sock_filter *filter, int flen)
639 {
640 u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
641 int pc, ret = 0;
642
643 BUILD_BUG_ON(BPF_MEMWORDS > 16);
644
645 masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
646 if (!masks)
647 return -ENOMEM;
648
649 memset(masks, 0xff, flen * sizeof(*masks));
650
651 for (pc = 0; pc < flen; pc++) {
652 memvalid &= masks[pc];
653
654 switch (filter[pc].code) {
655 case BPF_ST:
656 case BPF_STX:
657 memvalid |= (1 << filter[pc].k);
658 break;
659 case BPF_LD | BPF_MEM:
660 case BPF_LDX | BPF_MEM:
661 if (!(memvalid & (1 << filter[pc].k))) {
662 ret = -EINVAL;
663 goto error;
664 }
665 break;
666 case BPF_JMP | BPF_JA:
667 /* A jump must set masks on target */
668 masks[pc + 1 + filter[pc].k] &= memvalid;
669 memvalid = ~0;
670 break;
671 case BPF_JMP | BPF_JEQ | BPF_K:
672 case BPF_JMP | BPF_JEQ | BPF_X:
673 case BPF_JMP | BPF_JGE | BPF_K:
674 case BPF_JMP | BPF_JGE | BPF_X:
675 case BPF_JMP | BPF_JGT | BPF_K:
676 case BPF_JMP | BPF_JGT | BPF_X:
677 case BPF_JMP | BPF_JSET | BPF_K:
678 case BPF_JMP | BPF_JSET | BPF_X:
679 /* A jump must set masks on targets */
680 masks[pc + 1 + filter[pc].jt] &= memvalid;
681 masks[pc + 1 + filter[pc].jf] &= memvalid;
682 memvalid = ~0;
683 break;
684 }
685 }
686 error:
687 kfree(masks);
688 return ret;
689 }
690
691 static bool chk_code_allowed(u16 code_to_probe)
692 {
693 static const bool codes[] = {
694 /* 32 bit ALU operations */
695 [BPF_ALU | BPF_ADD | BPF_K] = true,
696 [BPF_ALU | BPF_ADD | BPF_X] = true,
697 [BPF_ALU | BPF_SUB | BPF_K] = true,
698 [BPF_ALU | BPF_SUB | BPF_X] = true,
699 [BPF_ALU | BPF_MUL | BPF_K] = true,
700 [BPF_ALU | BPF_MUL | BPF_X] = true,
701 [BPF_ALU | BPF_DIV | BPF_K] = true,
702 [BPF_ALU | BPF_DIV | BPF_X] = true,
703 [BPF_ALU | BPF_MOD | BPF_K] = true,
704 [BPF_ALU | BPF_MOD | BPF_X] = true,
705 [BPF_ALU | BPF_AND | BPF_K] = true,
706 [BPF_ALU | BPF_AND | BPF_X] = true,
707 [BPF_ALU | BPF_OR | BPF_K] = true,
708 [BPF_ALU | BPF_OR | BPF_X] = true,
709 [BPF_ALU | BPF_XOR | BPF_K] = true,
710 [BPF_ALU | BPF_XOR | BPF_X] = true,
711 [BPF_ALU | BPF_LSH | BPF_K] = true,
712 [BPF_ALU | BPF_LSH | BPF_X] = true,
713 [BPF_ALU | BPF_RSH | BPF_K] = true,
714 [BPF_ALU | BPF_RSH | BPF_X] = true,
715 [BPF_ALU | BPF_NEG] = true,
716 /* Load instructions */
717 [BPF_LD | BPF_W | BPF_ABS] = true,
718 [BPF_LD | BPF_H | BPF_ABS] = true,
719 [BPF_LD | BPF_B | BPF_ABS] = true,
720 [BPF_LD | BPF_W | BPF_LEN] = true,
721 [BPF_LD | BPF_W | BPF_IND] = true,
722 [BPF_LD | BPF_H | BPF_IND] = true,
723 [BPF_LD | BPF_B | BPF_IND] = true,
724 [BPF_LD | BPF_IMM] = true,
725 [BPF_LD | BPF_MEM] = true,
726 [BPF_LDX | BPF_W | BPF_LEN] = true,
727 [BPF_LDX | BPF_B | BPF_MSH] = true,
728 [BPF_LDX | BPF_IMM] = true,
729 [BPF_LDX | BPF_MEM] = true,
730 /* Store instructions */
731 [BPF_ST] = true,
732 [BPF_STX] = true,
733 /* Misc instructions */
734 [BPF_MISC | BPF_TAX] = true,
735 [BPF_MISC | BPF_TXA] = true,
736 /* Return instructions */
737 [BPF_RET | BPF_K] = true,
738 [BPF_RET | BPF_A] = true,
739 /* Jump instructions */
740 [BPF_JMP | BPF_JA] = true,
741 [BPF_JMP | BPF_JEQ | BPF_K] = true,
742 [BPF_JMP | BPF_JEQ | BPF_X] = true,
743 [BPF_JMP | BPF_JGE | BPF_K] = true,
744 [BPF_JMP | BPF_JGE | BPF_X] = true,
745 [BPF_JMP | BPF_JGT | BPF_K] = true,
746 [BPF_JMP | BPF_JGT | BPF_X] = true,
747 [BPF_JMP | BPF_JSET | BPF_K] = true,
748 [BPF_JMP | BPF_JSET | BPF_X] = true,
749 };
750
751 if (code_to_probe >= ARRAY_SIZE(codes))
752 return false;
753
754 return codes[code_to_probe];
755 }
756
757 static bool bpf_check_basics_ok(const struct sock_filter *filter,
758 unsigned int flen)
759 {
760 if (filter == NULL)
761 return false;
762 if (flen == 0 || flen > BPF_MAXINSNS)
763 return false;
764
765 return true;
766 }
767
768 /**
769 * bpf_check_classic - verify socket filter code
770 * @filter: filter to verify
771 * @flen: length of filter
772 *
773 * Check the user's filter code. If we let some ugly
774 * filter code slip through kaboom! The filter must contain
775 * no references or jumps that are out of range, no illegal
776 * instructions, and must end with a RET instruction.
777 *
778 * All jumps are forward as they are not signed.
779 *
780 * Returns 0 if the rule set is legal or -EINVAL if not.
781 */
782 static int bpf_check_classic(const struct sock_filter *filter,
783 unsigned int flen)
784 {
785 bool anc_found;
786 int pc;
787
788 /* Check the filter code now */
789 for (pc = 0; pc < flen; pc++) {
790 const struct sock_filter *ftest = &filter[pc];
791
792 /* May we actually operate on this code? */
793 if (!chk_code_allowed(ftest->code))
794 return -EINVAL;
795
796 /* Some instructions need special checks */
797 switch (ftest->code) {
798 case BPF_ALU | BPF_DIV | BPF_K:
799 case BPF_ALU | BPF_MOD | BPF_K:
800 /* Check for division by zero */
801 if (ftest->k == 0)
802 return -EINVAL;
803 break;
804 case BPF_ALU | BPF_LSH | BPF_K:
805 case BPF_ALU | BPF_RSH | BPF_K:
806 if (ftest->k >= 32)
807 return -EINVAL;
808 break;
809 case BPF_LD | BPF_MEM:
810 case BPF_LDX | BPF_MEM:
811 case BPF_ST:
812 case BPF_STX:
813 /* Check for invalid memory addresses */
814 if (ftest->k >= BPF_MEMWORDS)
815 return -EINVAL;
816 break;
817 case BPF_JMP | BPF_JA:
818 /* Note, the large ftest->k might cause loops.
819 * Compare this with conditional jumps below,
820 * where offsets are limited. --ANK (981016)
821 */
822 if (ftest->k >= (unsigned int)(flen - pc - 1))
823 return -EINVAL;
824 break;
825 case BPF_JMP | BPF_JEQ | BPF_K:
826 case BPF_JMP | BPF_JEQ | BPF_X:
827 case BPF_JMP | BPF_JGE | BPF_K:
828 case BPF_JMP | BPF_JGE | BPF_X:
829 case BPF_JMP | BPF_JGT | BPF_K:
830 case BPF_JMP | BPF_JGT | BPF_X:
831 case BPF_JMP | BPF_JSET | BPF_K:
832 case BPF_JMP | BPF_JSET | BPF_X:
833 /* Both conditionals must be safe */
834 if (pc + ftest->jt + 1 >= flen ||
835 pc + ftest->jf + 1 >= flen)
836 return -EINVAL;
837 break;
838 case BPF_LD | BPF_W | BPF_ABS:
839 case BPF_LD | BPF_H | BPF_ABS:
840 case BPF_LD | BPF_B | BPF_ABS:
841 anc_found = false;
842 if (bpf_anc_helper(ftest) & BPF_ANC)
843 anc_found = true;
844 /* Ancillary operation unknown or unsupported */
845 if (anc_found == false && ftest->k >= SKF_AD_OFF)
846 return -EINVAL;
847 }
848 }
849
850 /* Last instruction must be a RET code */
851 switch (filter[flen - 1].code) {
852 case BPF_RET | BPF_K:
853 case BPF_RET | BPF_A:
854 return check_load_and_stores(filter, flen);
855 }
856
857 return -EINVAL;
858 }
859
860 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
861 const struct sock_fprog *fprog)
862 {
863 unsigned int fsize = bpf_classic_proglen(fprog);
864 struct sock_fprog_kern *fkprog;
865
866 fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
867 if (!fp->orig_prog)
868 return -ENOMEM;
869
870 fkprog = fp->orig_prog;
871 fkprog->len = fprog->len;
872
873 fkprog->filter = kmemdup(fp->insns, fsize,
874 GFP_KERNEL | __GFP_NOWARN);
875 if (!fkprog->filter) {
876 kfree(fp->orig_prog);
877 return -ENOMEM;
878 }
879
880 return 0;
881 }
882
883 static void bpf_release_orig_filter(struct bpf_prog *fp)
884 {
885 struct sock_fprog_kern *fprog = fp->orig_prog;
886
887 if (fprog) {
888 kfree(fprog->filter);
889 kfree(fprog);
890 }
891 }
892
893 static void __bpf_prog_release(struct bpf_prog *prog)
894 {
895 if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
896 bpf_prog_put(prog);
897 } else {
898 bpf_release_orig_filter(prog);
899 bpf_prog_free(prog);
900 }
901 }
902
903 static void __sk_filter_release(struct sk_filter *fp)
904 {
905 __bpf_prog_release(fp->prog);
906 kfree(fp);
907 }
908
909 /**
910 * sk_filter_release_rcu - Release a socket filter by rcu_head
911 * @rcu: rcu_head that contains the sk_filter to free
912 */
913 static void sk_filter_release_rcu(struct rcu_head *rcu)
914 {
915 struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
916
917 __sk_filter_release(fp);
918 }
919
920 /**
921 * sk_filter_release - release a socket filter
922 * @fp: filter to remove
923 *
924 * Remove a filter from a socket and release its resources.
925 */
926 static void sk_filter_release(struct sk_filter *fp)
927 {
928 if (atomic_dec_and_test(&fp->refcnt))
929 call_rcu(&fp->rcu, sk_filter_release_rcu);
930 }
931
932 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
933 {
934 u32 filter_size = bpf_prog_size(fp->prog->len);
935
936 atomic_sub(filter_size, &sk->sk_omem_alloc);
937 sk_filter_release(fp);
938 }
939
940 /* try to charge the socket memory if there is space available
941 * return true on success
942 */
943 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
944 {
945 u32 filter_size = bpf_prog_size(fp->prog->len);
946
947 /* same check as in sock_kmalloc() */
948 if (filter_size <= sysctl_optmem_max &&
949 atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
950 atomic_inc(&fp->refcnt);
951 atomic_add(filter_size, &sk->sk_omem_alloc);
952 return true;
953 }
954 return false;
955 }
956
957 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
958 {
959 struct sock_filter *old_prog;
960 struct bpf_prog *old_fp;
961 int err, new_len, old_len = fp->len;
962
963 /* We are free to overwrite insns et al right here as it
964 * won't be used at this point in time anymore internally
965 * after the migration to the internal BPF instruction
966 * representation.
967 */
968 BUILD_BUG_ON(sizeof(struct sock_filter) !=
969 sizeof(struct bpf_insn));
970
971 /* Conversion cannot happen on overlapping memory areas,
972 * so we need to keep the user BPF around until the 2nd
973 * pass. At this time, the user BPF is stored in fp->insns.
974 */
975 old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
976 GFP_KERNEL | __GFP_NOWARN);
977 if (!old_prog) {
978 err = -ENOMEM;
979 goto out_err;
980 }
981
982 /* 1st pass: calculate the new program length. */
983 err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
984 if (err)
985 goto out_err_free;
986
987 /* Expand fp for appending the new filter representation. */
988 old_fp = fp;
989 fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
990 if (!fp) {
991 /* The old_fp is still around in case we couldn't
992 * allocate new memory, so uncharge on that one.
993 */
994 fp = old_fp;
995 err = -ENOMEM;
996 goto out_err_free;
997 }
998
999 fp->len = new_len;
1000
1001 /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1002 err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
1003 if (err)
1004 /* 2nd bpf_convert_filter() can fail only if it fails
1005 * to allocate memory, remapping must succeed. Note,
1006 * that at this time old_fp has already been released
1007 * by krealloc().
1008 */
1009 goto out_err_free;
1010
1011 /* We are guaranteed to never error here with cBPF to eBPF
1012 * transitions, since there's no issue with type compatibility
1013 * checks on program arrays.
1014 */
1015 fp = bpf_prog_select_runtime(fp, &err);
1016
1017 kfree(old_prog);
1018 return fp;
1019
1020 out_err_free:
1021 kfree(old_prog);
1022 out_err:
1023 __bpf_prog_release(fp);
1024 return ERR_PTR(err);
1025 }
1026
1027 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1028 bpf_aux_classic_check_t trans)
1029 {
1030 int err;
1031
1032 fp->bpf_func = NULL;
1033 fp->jited = 0;
1034
1035 err = bpf_check_classic(fp->insns, fp->len);
1036 if (err) {
1037 __bpf_prog_release(fp);
1038 return ERR_PTR(err);
1039 }
1040
1041 /* There might be additional checks and transformations
1042 * needed on classic filters, f.e. in case of seccomp.
1043 */
1044 if (trans) {
1045 err = trans(fp->insns, fp->len);
1046 if (err) {
1047 __bpf_prog_release(fp);
1048 return ERR_PTR(err);
1049 }
1050 }
1051
1052 /* Probe if we can JIT compile the filter and if so, do
1053 * the compilation of the filter.
1054 */
1055 bpf_jit_compile(fp);
1056
1057 /* JIT compiler couldn't process this filter, so do the
1058 * internal BPF translation for the optimized interpreter.
1059 */
1060 if (!fp->jited)
1061 fp = bpf_migrate_filter(fp);
1062
1063 return fp;
1064 }
1065
1066 /**
1067 * bpf_prog_create - create an unattached filter
1068 * @pfp: the unattached filter that is created
1069 * @fprog: the filter program
1070 *
1071 * Create a filter independent of any socket. We first run some
1072 * sanity checks on it to make sure it does not explode on us later.
1073 * If an error occurs or there is insufficient memory for the filter
1074 * a negative errno code is returned. On success the return is zero.
1075 */
1076 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1077 {
1078 unsigned int fsize = bpf_classic_proglen(fprog);
1079 struct bpf_prog *fp;
1080
1081 /* Make sure new filter is there and in the right amounts. */
1082 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1083 return -EINVAL;
1084
1085 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1086 if (!fp)
1087 return -ENOMEM;
1088
1089 memcpy(fp->insns, fprog->filter, fsize);
1090
1091 fp->len = fprog->len;
1092 /* Since unattached filters are not copied back to user
1093 * space through sk_get_filter(), we do not need to hold
1094 * a copy here, and can spare us the work.
1095 */
1096 fp->orig_prog = NULL;
1097
1098 /* bpf_prepare_filter() already takes care of freeing
1099 * memory in case something goes wrong.
1100 */
1101 fp = bpf_prepare_filter(fp, NULL);
1102 if (IS_ERR(fp))
1103 return PTR_ERR(fp);
1104
1105 *pfp = fp;
1106 return 0;
1107 }
1108 EXPORT_SYMBOL_GPL(bpf_prog_create);
1109
1110 /**
1111 * bpf_prog_create_from_user - create an unattached filter from user buffer
1112 * @pfp: the unattached filter that is created
1113 * @fprog: the filter program
1114 * @trans: post-classic verifier transformation handler
1115 * @save_orig: save classic BPF program
1116 *
1117 * This function effectively does the same as bpf_prog_create(), only
1118 * that it builds up its insns buffer from user space provided buffer.
1119 * It also allows for passing a bpf_aux_classic_check_t handler.
1120 */
1121 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1122 bpf_aux_classic_check_t trans, bool save_orig)
1123 {
1124 unsigned int fsize = bpf_classic_proglen(fprog);
1125 struct bpf_prog *fp;
1126 int err;
1127
1128 /* Make sure new filter is there and in the right amounts. */
1129 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1130 return -EINVAL;
1131
1132 fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1133 if (!fp)
1134 return -ENOMEM;
1135
1136 if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1137 __bpf_prog_free(fp);
1138 return -EFAULT;
1139 }
1140
1141 fp->len = fprog->len;
1142 fp->orig_prog = NULL;
1143
1144 if (save_orig) {
1145 err = bpf_prog_store_orig_filter(fp, fprog);
1146 if (err) {
1147 __bpf_prog_free(fp);
1148 return -ENOMEM;
1149 }
1150 }
1151
1152 /* bpf_prepare_filter() already takes care of freeing
1153 * memory in case something goes wrong.
1154 */
1155 fp = bpf_prepare_filter(fp, trans);
1156 if (IS_ERR(fp))
1157 return PTR_ERR(fp);
1158
1159 *pfp = fp;
1160 return 0;
1161 }
1162 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1163
1164 void bpf_prog_destroy(struct bpf_prog *fp)
1165 {
1166 __bpf_prog_release(fp);
1167 }
1168 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1169
1170 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1171 {
1172 struct sk_filter *fp, *old_fp;
1173
1174 fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1175 if (!fp)
1176 return -ENOMEM;
1177
1178 fp->prog = prog;
1179 atomic_set(&fp->refcnt, 0);
1180
1181 if (!sk_filter_charge(sk, fp)) {
1182 kfree(fp);
1183 return -ENOMEM;
1184 }
1185
1186 old_fp = rcu_dereference_protected(sk->sk_filter,
1187 lockdep_sock_is_held(sk));
1188 rcu_assign_pointer(sk->sk_filter, fp);
1189
1190 if (old_fp)
1191 sk_filter_uncharge(sk, old_fp);
1192
1193 return 0;
1194 }
1195
1196 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1197 {
1198 struct bpf_prog *old_prog;
1199 int err;
1200
1201 if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1202 return -ENOMEM;
1203
1204 if (sk_unhashed(sk) && sk->sk_reuseport) {
1205 err = reuseport_alloc(sk);
1206 if (err)
1207 return err;
1208 } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1209 /* The socket wasn't bound with SO_REUSEPORT */
1210 return -EINVAL;
1211 }
1212
1213 old_prog = reuseport_attach_prog(sk, prog);
1214 if (old_prog)
1215 bpf_prog_destroy(old_prog);
1216
1217 return 0;
1218 }
1219
1220 static
1221 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1222 {
1223 unsigned int fsize = bpf_classic_proglen(fprog);
1224 struct bpf_prog *prog;
1225 int err;
1226
1227 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1228 return ERR_PTR(-EPERM);
1229
1230 /* Make sure new filter is there and in the right amounts. */
1231 if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1232 return ERR_PTR(-EINVAL);
1233
1234 prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1235 if (!prog)
1236 return ERR_PTR(-ENOMEM);
1237
1238 if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1239 __bpf_prog_free(prog);
1240 return ERR_PTR(-EFAULT);
1241 }
1242
1243 prog->len = fprog->len;
1244
1245 err = bpf_prog_store_orig_filter(prog, fprog);
1246 if (err) {
1247 __bpf_prog_free(prog);
1248 return ERR_PTR(-ENOMEM);
1249 }
1250
1251 /* bpf_prepare_filter() already takes care of freeing
1252 * memory in case something goes wrong.
1253 */
1254 return bpf_prepare_filter(prog, NULL);
1255 }
1256
1257 /**
1258 * sk_attach_filter - attach a socket filter
1259 * @fprog: the filter program
1260 * @sk: the socket to use
1261 *
1262 * Attach the user's filter code. We first run some sanity checks on
1263 * it to make sure it does not explode on us later. If an error
1264 * occurs or there is insufficient memory for the filter a negative
1265 * errno code is returned. On success the return is zero.
1266 */
1267 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1268 {
1269 struct bpf_prog *prog = __get_filter(fprog, sk);
1270 int err;
1271
1272 if (IS_ERR(prog))
1273 return PTR_ERR(prog);
1274
1275 err = __sk_attach_prog(prog, sk);
1276 if (err < 0) {
1277 __bpf_prog_release(prog);
1278 return err;
1279 }
1280
1281 return 0;
1282 }
1283 EXPORT_SYMBOL_GPL(sk_attach_filter);
1284
1285 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1286 {
1287 struct bpf_prog *prog = __get_filter(fprog, sk);
1288 int err;
1289
1290 if (IS_ERR(prog))
1291 return PTR_ERR(prog);
1292
1293 err = __reuseport_attach_prog(prog, sk);
1294 if (err < 0) {
1295 __bpf_prog_release(prog);
1296 return err;
1297 }
1298
1299 return 0;
1300 }
1301
1302 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1303 {
1304 if (sock_flag(sk, SOCK_FILTER_LOCKED))
1305 return ERR_PTR(-EPERM);
1306
1307 return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1308 }
1309
1310 int sk_attach_bpf(u32 ufd, struct sock *sk)
1311 {
1312 struct bpf_prog *prog = __get_bpf(ufd, sk);
1313 int err;
1314
1315 if (IS_ERR(prog))
1316 return PTR_ERR(prog);
1317
1318 err = __sk_attach_prog(prog, sk);
1319 if (err < 0) {
1320 bpf_prog_put(prog);
1321 return err;
1322 }
1323
1324 return 0;
1325 }
1326
1327 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1328 {
1329 struct bpf_prog *prog = __get_bpf(ufd, sk);
1330 int err;
1331
1332 if (IS_ERR(prog))
1333 return PTR_ERR(prog);
1334
1335 err = __reuseport_attach_prog(prog, sk);
1336 if (err < 0) {
1337 bpf_prog_put(prog);
1338 return err;
1339 }
1340
1341 return 0;
1342 }
1343
1344 struct bpf_scratchpad {
1345 union {
1346 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1347 u8 buff[MAX_BPF_STACK];
1348 };
1349 };
1350
1351 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1352
1353 static inline int bpf_try_make_writable(struct sk_buff *skb,
1354 unsigned int write_len)
1355 {
1356 int err;
1357
1358 if (!skb_cloned(skb))
1359 return 0;
1360 if (skb_clone_writable(skb, write_len))
1361 return 0;
1362 err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
1363 if (!err)
1364 bpf_compute_data_end(skb);
1365 return err;
1366 }
1367
1368 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1369 {
1370 if (skb_at_tc_ingress(skb))
1371 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1372 }
1373
1374 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1375 {
1376 if (skb_at_tc_ingress(skb))
1377 skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1378 }
1379
1380 static u64 bpf_skb_store_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 flags)
1381 {
1382 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1383 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1384 int offset = (int) r2;
1385 void *from = (void *) (long) r3;
1386 unsigned int len = (unsigned int) r4;
1387 void *ptr;
1388
1389 if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1390 return -EINVAL;
1391
1392 /* bpf verifier guarantees that:
1393 * 'from' pointer points to bpf program stack
1394 * 'len' bytes of it were initialized
1395 * 'len' > 0
1396 * 'skb' is a valid pointer to 'struct sk_buff'
1397 *
1398 * so check for invalid 'offset' and too large 'len'
1399 */
1400 if (unlikely((u32) offset > 0xffff || len > sizeof(sp->buff)))
1401 return -EFAULT;
1402 if (unlikely(bpf_try_make_writable(skb, offset + len)))
1403 return -EFAULT;
1404
1405 ptr = skb_header_pointer(skb, offset, len, sp->buff);
1406 if (unlikely(!ptr))
1407 return -EFAULT;
1408
1409 if (flags & BPF_F_RECOMPUTE_CSUM)
1410 __skb_postpull_rcsum(skb, ptr, len, offset);
1411
1412 memcpy(ptr, from, len);
1413
1414 if (ptr == sp->buff)
1415 /* skb_store_bits cannot return -EFAULT here */
1416 skb_store_bits(skb, offset, ptr, len);
1417
1418 if (flags & BPF_F_RECOMPUTE_CSUM)
1419 __skb_postpush_rcsum(skb, ptr, len, offset);
1420 if (flags & BPF_F_INVALIDATE_HASH)
1421 skb_clear_hash(skb);
1422
1423 return 0;
1424 }
1425
1426 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1427 .func = bpf_skb_store_bytes,
1428 .gpl_only = false,
1429 .ret_type = RET_INTEGER,
1430 .arg1_type = ARG_PTR_TO_CTX,
1431 .arg2_type = ARG_ANYTHING,
1432 .arg3_type = ARG_PTR_TO_STACK,
1433 .arg4_type = ARG_CONST_STACK_SIZE,
1434 .arg5_type = ARG_ANYTHING,
1435 };
1436
1437 static u64 bpf_skb_load_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1438 {
1439 const struct sk_buff *skb = (const struct sk_buff *)(unsigned long) r1;
1440 int offset = (int) r2;
1441 void *to = (void *)(unsigned long) r3;
1442 unsigned int len = (unsigned int) r4;
1443 void *ptr;
1444
1445 if (unlikely((u32) offset > 0xffff))
1446 goto err_clear;
1447
1448 ptr = skb_header_pointer(skb, offset, len, to);
1449 if (unlikely(!ptr))
1450 goto err_clear;
1451 if (ptr != to)
1452 memcpy(to, ptr, len);
1453
1454 return 0;
1455 err_clear:
1456 memset(to, 0, len);
1457 return -EFAULT;
1458 }
1459
1460 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1461 .func = bpf_skb_load_bytes,
1462 .gpl_only = false,
1463 .ret_type = RET_INTEGER,
1464 .arg1_type = ARG_PTR_TO_CTX,
1465 .arg2_type = ARG_ANYTHING,
1466 .arg3_type = ARG_PTR_TO_RAW_STACK,
1467 .arg4_type = ARG_CONST_STACK_SIZE,
1468 };
1469
1470 static u64 bpf_l3_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1471 {
1472 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1473 int offset = (int) r2;
1474 __sum16 sum, *ptr;
1475
1476 if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1477 return -EINVAL;
1478 if (unlikely((u32) offset > 0xffff))
1479 return -EFAULT;
1480 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(sum))))
1481 return -EFAULT;
1482
1483 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1484 if (unlikely(!ptr))
1485 return -EFAULT;
1486
1487 switch (flags & BPF_F_HDR_FIELD_MASK) {
1488 case 0:
1489 if (unlikely(from != 0))
1490 return -EINVAL;
1491
1492 csum_replace_by_diff(ptr, to);
1493 break;
1494 case 2:
1495 csum_replace2(ptr, from, to);
1496 break;
1497 case 4:
1498 csum_replace4(ptr, from, to);
1499 break;
1500 default:
1501 return -EINVAL;
1502 }
1503
1504 if (ptr == &sum)
1505 /* skb_store_bits guaranteed to not return -EFAULT here */
1506 skb_store_bits(skb, offset, ptr, sizeof(sum));
1507
1508 return 0;
1509 }
1510
1511 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1512 .func = bpf_l3_csum_replace,
1513 .gpl_only = false,
1514 .ret_type = RET_INTEGER,
1515 .arg1_type = ARG_PTR_TO_CTX,
1516 .arg2_type = ARG_ANYTHING,
1517 .arg3_type = ARG_ANYTHING,
1518 .arg4_type = ARG_ANYTHING,
1519 .arg5_type = ARG_ANYTHING,
1520 };
1521
1522 static u64 bpf_l4_csum_replace(u64 r1, u64 r2, u64 from, u64 to, u64 flags)
1523 {
1524 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1525 bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1526 bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1527 int offset = (int) r2;
1528 __sum16 sum, *ptr;
1529
1530 if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_PSEUDO_HDR |
1531 BPF_F_HDR_FIELD_MASK)))
1532 return -EINVAL;
1533 if (unlikely((u32) offset > 0xffff))
1534 return -EFAULT;
1535 if (unlikely(bpf_try_make_writable(skb, offset + sizeof(sum))))
1536 return -EFAULT;
1537
1538 ptr = skb_header_pointer(skb, offset, sizeof(sum), &sum);
1539 if (unlikely(!ptr))
1540 return -EFAULT;
1541 if (is_mmzero && !*ptr)
1542 return 0;
1543
1544 switch (flags & BPF_F_HDR_FIELD_MASK) {
1545 case 0:
1546 if (unlikely(from != 0))
1547 return -EINVAL;
1548
1549 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1550 break;
1551 case 2:
1552 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1553 break;
1554 case 4:
1555 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1556 break;
1557 default:
1558 return -EINVAL;
1559 }
1560
1561 if (is_mmzero && !*ptr)
1562 *ptr = CSUM_MANGLED_0;
1563 if (ptr == &sum)
1564 /* skb_store_bits guaranteed to not return -EFAULT here */
1565 skb_store_bits(skb, offset, ptr, sizeof(sum));
1566
1567 return 0;
1568 }
1569
1570 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1571 .func = bpf_l4_csum_replace,
1572 .gpl_only = false,
1573 .ret_type = RET_INTEGER,
1574 .arg1_type = ARG_PTR_TO_CTX,
1575 .arg2_type = ARG_ANYTHING,
1576 .arg3_type = ARG_ANYTHING,
1577 .arg4_type = ARG_ANYTHING,
1578 .arg5_type = ARG_ANYTHING,
1579 };
1580
1581 static u64 bpf_csum_diff(u64 r1, u64 from_size, u64 r3, u64 to_size, u64 seed)
1582 {
1583 struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1584 u64 diff_size = from_size + to_size;
1585 __be32 *from = (__be32 *) (long) r1;
1586 __be32 *to = (__be32 *) (long) r3;
1587 int i, j = 0;
1588
1589 /* This is quite flexible, some examples:
1590 *
1591 * from_size == 0, to_size > 0, seed := csum --> pushing data
1592 * from_size > 0, to_size == 0, seed := csum --> pulling data
1593 * from_size > 0, to_size > 0, seed := 0 --> diffing data
1594 *
1595 * Even for diffing, from_size and to_size don't need to be equal.
1596 */
1597 if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1598 diff_size > sizeof(sp->diff)))
1599 return -EINVAL;
1600
1601 for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1602 sp->diff[j] = ~from[i];
1603 for (i = 0; i < to_size / sizeof(__be32); i++, j++)
1604 sp->diff[j] = to[i];
1605
1606 return csum_partial(sp->diff, diff_size, seed);
1607 }
1608
1609 static const struct bpf_func_proto bpf_csum_diff_proto = {
1610 .func = bpf_csum_diff,
1611 .gpl_only = false,
1612 .ret_type = RET_INTEGER,
1613 .arg1_type = ARG_PTR_TO_STACK,
1614 .arg2_type = ARG_CONST_STACK_SIZE_OR_ZERO,
1615 .arg3_type = ARG_PTR_TO_STACK,
1616 .arg4_type = ARG_CONST_STACK_SIZE_OR_ZERO,
1617 .arg5_type = ARG_ANYTHING,
1618 };
1619
1620 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1621 {
1622 return dev_forward_skb(dev, skb);
1623 }
1624
1625 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1626 {
1627 int ret;
1628
1629 if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1630 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1631 kfree_skb(skb);
1632 return -ENETDOWN;
1633 }
1634
1635 skb->dev = dev;
1636
1637 __this_cpu_inc(xmit_recursion);
1638 ret = dev_queue_xmit(skb);
1639 __this_cpu_dec(xmit_recursion);
1640
1641 return ret;
1642 }
1643
1644 static u64 bpf_clone_redirect(u64 r1, u64 ifindex, u64 flags, u64 r4, u64 r5)
1645 {
1646 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1647 struct net_device *dev;
1648
1649 if (unlikely(flags & ~(BPF_F_INGRESS)))
1650 return -EINVAL;
1651
1652 dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1653 if (unlikely(!dev))
1654 return -EINVAL;
1655
1656 skb = skb_clone(skb, GFP_ATOMIC);
1657 if (unlikely(!skb))
1658 return -ENOMEM;
1659
1660 bpf_push_mac_rcsum(skb);
1661
1662 return flags & BPF_F_INGRESS ?
1663 __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1664 }
1665
1666 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1667 .func = bpf_clone_redirect,
1668 .gpl_only = false,
1669 .ret_type = RET_INTEGER,
1670 .arg1_type = ARG_PTR_TO_CTX,
1671 .arg2_type = ARG_ANYTHING,
1672 .arg3_type = ARG_ANYTHING,
1673 };
1674
1675 struct redirect_info {
1676 u32 ifindex;
1677 u32 flags;
1678 };
1679
1680 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1681
1682 static u64 bpf_redirect(u64 ifindex, u64 flags, u64 r3, u64 r4, u64 r5)
1683 {
1684 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1685
1686 if (unlikely(flags & ~(BPF_F_INGRESS)))
1687 return TC_ACT_SHOT;
1688
1689 ri->ifindex = ifindex;
1690 ri->flags = flags;
1691
1692 return TC_ACT_REDIRECT;
1693 }
1694
1695 int skb_do_redirect(struct sk_buff *skb)
1696 {
1697 struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1698 struct net_device *dev;
1699
1700 dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1701 ri->ifindex = 0;
1702 if (unlikely(!dev)) {
1703 kfree_skb(skb);
1704 return -EINVAL;
1705 }
1706
1707 bpf_push_mac_rcsum(skb);
1708
1709 return ri->flags & BPF_F_INGRESS ?
1710 __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1711 }
1712
1713 static const struct bpf_func_proto bpf_redirect_proto = {
1714 .func = bpf_redirect,
1715 .gpl_only = false,
1716 .ret_type = RET_INTEGER,
1717 .arg1_type = ARG_ANYTHING,
1718 .arg2_type = ARG_ANYTHING,
1719 };
1720
1721 static u64 bpf_get_cgroup_classid(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1722 {
1723 return task_get_classid((struct sk_buff *) (unsigned long) r1);
1724 }
1725
1726 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1727 .func = bpf_get_cgroup_classid,
1728 .gpl_only = false,
1729 .ret_type = RET_INTEGER,
1730 .arg1_type = ARG_PTR_TO_CTX,
1731 };
1732
1733 static u64 bpf_get_route_realm(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1734 {
1735 return dst_tclassid((struct sk_buff *) (unsigned long) r1);
1736 }
1737
1738 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1739 .func = bpf_get_route_realm,
1740 .gpl_only = false,
1741 .ret_type = RET_INTEGER,
1742 .arg1_type = ARG_PTR_TO_CTX,
1743 };
1744
1745 static u64 bpf_get_hash_recalc(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1746 {
1747 /* If skb_clear_hash() was called due to mangling, we can
1748 * trigger SW recalculation here. Later access to hash
1749 * can then use the inline skb->hash via context directly
1750 * instead of calling this helper again.
1751 */
1752 return skb_get_hash((struct sk_buff *) (unsigned long) r1);
1753 }
1754
1755 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1756 .func = bpf_get_hash_recalc,
1757 .gpl_only = false,
1758 .ret_type = RET_INTEGER,
1759 .arg1_type = ARG_PTR_TO_CTX,
1760 };
1761
1762 static u64 bpf_skb_vlan_push(u64 r1, u64 r2, u64 vlan_tci, u64 r4, u64 r5)
1763 {
1764 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1765 __be16 vlan_proto = (__force __be16) r2;
1766 int ret;
1767
1768 if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1769 vlan_proto != htons(ETH_P_8021AD)))
1770 vlan_proto = htons(ETH_P_8021Q);
1771
1772 bpf_push_mac_rcsum(skb);
1773 ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1774 bpf_pull_mac_rcsum(skb);
1775
1776 bpf_compute_data_end(skb);
1777 return ret;
1778 }
1779
1780 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1781 .func = bpf_skb_vlan_push,
1782 .gpl_only = false,
1783 .ret_type = RET_INTEGER,
1784 .arg1_type = ARG_PTR_TO_CTX,
1785 .arg2_type = ARG_ANYTHING,
1786 .arg3_type = ARG_ANYTHING,
1787 };
1788 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1789
1790 static u64 bpf_skb_vlan_pop(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1791 {
1792 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1793 int ret;
1794
1795 bpf_push_mac_rcsum(skb);
1796 ret = skb_vlan_pop(skb);
1797 bpf_pull_mac_rcsum(skb);
1798
1799 bpf_compute_data_end(skb);
1800 return ret;
1801 }
1802
1803 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1804 .func = bpf_skb_vlan_pop,
1805 .gpl_only = false,
1806 .ret_type = RET_INTEGER,
1807 .arg1_type = ARG_PTR_TO_CTX,
1808 };
1809 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1810
1811 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
1812 {
1813 /* Caller already did skb_cow() with len as headroom,
1814 * so no need to do it here.
1815 */
1816 skb_push(skb, len);
1817 memmove(skb->data, skb->data + len, off);
1818 memset(skb->data + off, 0, len);
1819
1820 /* No skb_postpush_rcsum(skb, skb->data + off, len)
1821 * needed here as it does not change the skb->csum
1822 * result for checksum complete when summing over
1823 * zeroed blocks.
1824 */
1825 return 0;
1826 }
1827
1828 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
1829 {
1830 /* skb_ensure_writable() is not needed here, as we're
1831 * already working on an uncloned skb.
1832 */
1833 if (unlikely(!pskb_may_pull(skb, off + len)))
1834 return -ENOMEM;
1835
1836 skb_postpull_rcsum(skb, skb->data + off, len);
1837 memmove(skb->data + len, skb->data, off);
1838 __skb_pull(skb, len);
1839
1840 return 0;
1841 }
1842
1843 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
1844 {
1845 bool trans_same = skb->transport_header == skb->network_header;
1846 int ret;
1847
1848 /* There's no need for __skb_push()/__skb_pull() pair to
1849 * get to the start of the mac header as we're guaranteed
1850 * to always start from here under eBPF.
1851 */
1852 ret = bpf_skb_generic_push(skb, off, len);
1853 if (likely(!ret)) {
1854 skb->mac_header -= len;
1855 skb->network_header -= len;
1856 if (trans_same)
1857 skb->transport_header = skb->network_header;
1858 }
1859
1860 return ret;
1861 }
1862
1863 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
1864 {
1865 bool trans_same = skb->transport_header == skb->network_header;
1866 int ret;
1867
1868 /* Same here, __skb_push()/__skb_pull() pair not needed. */
1869 ret = bpf_skb_generic_pop(skb, off, len);
1870 if (likely(!ret)) {
1871 skb->mac_header += len;
1872 skb->network_header += len;
1873 if (trans_same)
1874 skb->transport_header = skb->network_header;
1875 }
1876
1877 return ret;
1878 }
1879
1880 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
1881 {
1882 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
1883 u32 off = skb->network_header - skb->mac_header;
1884 int ret;
1885
1886 ret = skb_cow(skb, len_diff);
1887 if (unlikely(ret < 0))
1888 return ret;
1889
1890 ret = bpf_skb_net_hdr_push(skb, off, len_diff);
1891 if (unlikely(ret < 0))
1892 return ret;
1893
1894 if (skb_is_gso(skb)) {
1895 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV4 needs to
1896 * be changed into SKB_GSO_TCPV6.
1897 */
1898 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
1899 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
1900 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV6;
1901 }
1902
1903 /* Due to IPv6 header, MSS needs to be downgraded. */
1904 skb_shinfo(skb)->gso_size -= len_diff;
1905 /* Header must be checked, and gso_segs recomputed. */
1906 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1907 skb_shinfo(skb)->gso_segs = 0;
1908 }
1909
1910 skb->protocol = htons(ETH_P_IPV6);
1911 skb_clear_hash(skb);
1912
1913 return 0;
1914 }
1915
1916 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
1917 {
1918 const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
1919 u32 off = skb->network_header - skb->mac_header;
1920 int ret;
1921
1922 ret = skb_unclone(skb, GFP_ATOMIC);
1923 if (unlikely(ret < 0))
1924 return ret;
1925
1926 ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
1927 if (unlikely(ret < 0))
1928 return ret;
1929
1930 if (skb_is_gso(skb)) {
1931 /* SKB_GSO_UDP stays as is. SKB_GSO_TCPV6 needs to
1932 * be changed into SKB_GSO_TCPV4.
1933 */
1934 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
1935 skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
1936 skb_shinfo(skb)->gso_type |= SKB_GSO_TCPV4;
1937 }
1938
1939 /* Due to IPv4 header, MSS can be upgraded. */
1940 skb_shinfo(skb)->gso_size += len_diff;
1941 /* Header must be checked, and gso_segs recomputed. */
1942 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1943 skb_shinfo(skb)->gso_segs = 0;
1944 }
1945
1946 skb->protocol = htons(ETH_P_IP);
1947 skb_clear_hash(skb);
1948
1949 return 0;
1950 }
1951
1952 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
1953 {
1954 __be16 from_proto = skb->protocol;
1955
1956 if (from_proto == htons(ETH_P_IP) &&
1957 to_proto == htons(ETH_P_IPV6))
1958 return bpf_skb_proto_4_to_6(skb);
1959
1960 if (from_proto == htons(ETH_P_IPV6) &&
1961 to_proto == htons(ETH_P_IP))
1962 return bpf_skb_proto_6_to_4(skb);
1963
1964 return -ENOTSUPP;
1965 }
1966
1967 static u64 bpf_skb_change_proto(u64 r1, u64 r2, u64 flags, u64 r4, u64 r5)
1968 {
1969 struct sk_buff *skb = (struct sk_buff *) (long) r1;
1970 __be16 proto = (__force __be16) r2;
1971 int ret;
1972
1973 if (unlikely(flags))
1974 return -EINVAL;
1975
1976 /* General idea is that this helper does the basic groundwork
1977 * needed for changing the protocol, and eBPF program fills the
1978 * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
1979 * and other helpers, rather than passing a raw buffer here.
1980 *
1981 * The rationale is to keep this minimal and without a need to
1982 * deal with raw packet data. F.e. even if we would pass buffers
1983 * here, the program still needs to call the bpf_lX_csum_replace()
1984 * helpers anyway. Plus, this way we keep also separation of
1985 * concerns, since f.e. bpf_skb_store_bytes() should only take
1986 * care of stores.
1987 *
1988 * Currently, additional options and extension header space are
1989 * not supported, but flags register is reserved so we can adapt
1990 * that. For offloads, we mark packet as dodgy, so that headers
1991 * need to be verified first.
1992 */
1993 ret = bpf_skb_proto_xlat(skb, proto);
1994 bpf_compute_data_end(skb);
1995 return ret;
1996 }
1997
1998 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
1999 .func = bpf_skb_change_proto,
2000 .gpl_only = false,
2001 .ret_type = RET_INTEGER,
2002 .arg1_type = ARG_PTR_TO_CTX,
2003 .arg2_type = ARG_ANYTHING,
2004 .arg3_type = ARG_ANYTHING,
2005 };
2006
2007 static u64 bpf_skb_change_type(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
2008 {
2009 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2010 u32 pkt_type = r2;
2011
2012 /* We only allow a restricted subset to be changed for now. */
2013 if (unlikely(skb->pkt_type > PACKET_OTHERHOST ||
2014 pkt_type > PACKET_OTHERHOST))
2015 return -EINVAL;
2016
2017 skb->pkt_type = pkt_type;
2018 return 0;
2019 }
2020
2021 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2022 .func = bpf_skb_change_type,
2023 .gpl_only = false,
2024 .ret_type = RET_INTEGER,
2025 .arg1_type = ARG_PTR_TO_CTX,
2026 .arg2_type = ARG_ANYTHING,
2027 };
2028
2029 bool bpf_helper_changes_skb_data(void *func)
2030 {
2031 if (func == bpf_skb_vlan_push)
2032 return true;
2033 if (func == bpf_skb_vlan_pop)
2034 return true;
2035 if (func == bpf_skb_store_bytes)
2036 return true;
2037 if (func == bpf_skb_change_proto)
2038 return true;
2039 if (func == bpf_l3_csum_replace)
2040 return true;
2041 if (func == bpf_l4_csum_replace)
2042 return true;
2043
2044 return false;
2045 }
2046
2047 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2048 unsigned long off, unsigned long len)
2049 {
2050 void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2051
2052 if (unlikely(!ptr))
2053 return len;
2054 if (ptr != dst_buff)
2055 memcpy(dst_buff, ptr, len);
2056
2057 return 0;
2058 }
2059
2060 static u64 bpf_skb_event_output(u64 r1, u64 r2, u64 flags, u64 r4,
2061 u64 meta_size)
2062 {
2063 struct sk_buff *skb = (struct sk_buff *)(long) r1;
2064 struct bpf_map *map = (struct bpf_map *)(long) r2;
2065 u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2066 void *meta = (void *)(long) r4;
2067
2068 if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2069 return -EINVAL;
2070 if (unlikely(skb_size > skb->len))
2071 return -EFAULT;
2072
2073 return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2074 bpf_skb_copy);
2075 }
2076
2077 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2078 .func = bpf_skb_event_output,
2079 .gpl_only = true,
2080 .ret_type = RET_INTEGER,
2081 .arg1_type = ARG_PTR_TO_CTX,
2082 .arg2_type = ARG_CONST_MAP_PTR,
2083 .arg3_type = ARG_ANYTHING,
2084 .arg4_type = ARG_PTR_TO_STACK,
2085 .arg5_type = ARG_CONST_STACK_SIZE,
2086 };
2087
2088 static unsigned short bpf_tunnel_key_af(u64 flags)
2089 {
2090 return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2091 }
2092
2093 static u64 bpf_skb_get_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
2094 {
2095 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2096 struct bpf_tunnel_key *to = (struct bpf_tunnel_key *) (long) r2;
2097 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2098 u8 compat[sizeof(struct bpf_tunnel_key)];
2099 void *to_orig = to;
2100 int err;
2101
2102 if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2103 err = -EINVAL;
2104 goto err_clear;
2105 }
2106 if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2107 err = -EPROTO;
2108 goto err_clear;
2109 }
2110 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2111 err = -EINVAL;
2112 switch (size) {
2113 case offsetof(struct bpf_tunnel_key, tunnel_label):
2114 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2115 goto set_compat;
2116 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2117 /* Fixup deprecated structure layouts here, so we have
2118 * a common path later on.
2119 */
2120 if (ip_tunnel_info_af(info) != AF_INET)
2121 goto err_clear;
2122 set_compat:
2123 to = (struct bpf_tunnel_key *)compat;
2124 break;
2125 default:
2126 goto err_clear;
2127 }
2128 }
2129
2130 to->tunnel_id = be64_to_cpu(info->key.tun_id);
2131 to->tunnel_tos = info->key.tos;
2132 to->tunnel_ttl = info->key.ttl;
2133
2134 if (flags & BPF_F_TUNINFO_IPV6) {
2135 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2136 sizeof(to->remote_ipv6));
2137 to->tunnel_label = be32_to_cpu(info->key.label);
2138 } else {
2139 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2140 }
2141
2142 if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2143 memcpy(to_orig, to, size);
2144
2145 return 0;
2146 err_clear:
2147 memset(to_orig, 0, size);
2148 return err;
2149 }
2150
2151 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2152 .func = bpf_skb_get_tunnel_key,
2153 .gpl_only = false,
2154 .ret_type = RET_INTEGER,
2155 .arg1_type = ARG_PTR_TO_CTX,
2156 .arg2_type = ARG_PTR_TO_RAW_STACK,
2157 .arg3_type = ARG_CONST_STACK_SIZE,
2158 .arg4_type = ARG_ANYTHING,
2159 };
2160
2161 static u64 bpf_skb_get_tunnel_opt(u64 r1, u64 r2, u64 size, u64 r4, u64 r5)
2162 {
2163 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2164 u8 *to = (u8 *) (long) r2;
2165 const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2166 int err;
2167
2168 if (unlikely(!info ||
2169 !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2170 err = -ENOENT;
2171 goto err_clear;
2172 }
2173 if (unlikely(size < info->options_len)) {
2174 err = -ENOMEM;
2175 goto err_clear;
2176 }
2177
2178 ip_tunnel_info_opts_get(to, info);
2179 if (size > info->options_len)
2180 memset(to + info->options_len, 0, size - info->options_len);
2181
2182 return info->options_len;
2183 err_clear:
2184 memset(to, 0, size);
2185 return err;
2186 }
2187
2188 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2189 .func = bpf_skb_get_tunnel_opt,
2190 .gpl_only = false,
2191 .ret_type = RET_INTEGER,
2192 .arg1_type = ARG_PTR_TO_CTX,
2193 .arg2_type = ARG_PTR_TO_RAW_STACK,
2194 .arg3_type = ARG_CONST_STACK_SIZE,
2195 };
2196
2197 static struct metadata_dst __percpu *md_dst;
2198
2199 static u64 bpf_skb_set_tunnel_key(u64 r1, u64 r2, u64 size, u64 flags, u64 r5)
2200 {
2201 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2202 struct bpf_tunnel_key *from = (struct bpf_tunnel_key *) (long) r2;
2203 struct metadata_dst *md = this_cpu_ptr(md_dst);
2204 u8 compat[sizeof(struct bpf_tunnel_key)];
2205 struct ip_tunnel_info *info;
2206
2207 if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2208 BPF_F_DONT_FRAGMENT)))
2209 return -EINVAL;
2210 if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2211 switch (size) {
2212 case offsetof(struct bpf_tunnel_key, tunnel_label):
2213 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2214 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2215 /* Fixup deprecated structure layouts here, so we have
2216 * a common path later on.
2217 */
2218 memcpy(compat, from, size);
2219 memset(compat + size, 0, sizeof(compat) - size);
2220 from = (struct bpf_tunnel_key *)compat;
2221 break;
2222 default:
2223 return -EINVAL;
2224 }
2225 }
2226 if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
2227 from->tunnel_ext))
2228 return -EINVAL;
2229
2230 skb_dst_drop(skb);
2231 dst_hold((struct dst_entry *) md);
2232 skb_dst_set(skb, (struct dst_entry *) md);
2233
2234 info = &md->u.tun_info;
2235 info->mode = IP_TUNNEL_INFO_TX;
2236
2237 info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
2238 if (flags & BPF_F_DONT_FRAGMENT)
2239 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
2240
2241 info->key.tun_id = cpu_to_be64(from->tunnel_id);
2242 info->key.tos = from->tunnel_tos;
2243 info->key.ttl = from->tunnel_ttl;
2244
2245 if (flags & BPF_F_TUNINFO_IPV6) {
2246 info->mode |= IP_TUNNEL_INFO_IPV6;
2247 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
2248 sizeof(from->remote_ipv6));
2249 info->key.label = cpu_to_be32(from->tunnel_label) &
2250 IPV6_FLOWLABEL_MASK;
2251 } else {
2252 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
2253 if (flags & BPF_F_ZERO_CSUM_TX)
2254 info->key.tun_flags &= ~TUNNEL_CSUM;
2255 }
2256
2257 return 0;
2258 }
2259
2260 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
2261 .func = bpf_skb_set_tunnel_key,
2262 .gpl_only = false,
2263 .ret_type = RET_INTEGER,
2264 .arg1_type = ARG_PTR_TO_CTX,
2265 .arg2_type = ARG_PTR_TO_STACK,
2266 .arg3_type = ARG_CONST_STACK_SIZE,
2267 .arg4_type = ARG_ANYTHING,
2268 };
2269
2270 static u64 bpf_skb_set_tunnel_opt(u64 r1, u64 r2, u64 size, u64 r4, u64 r5)
2271 {
2272 struct sk_buff *skb = (struct sk_buff *) (long) r1;
2273 u8 *from = (u8 *) (long) r2;
2274 struct ip_tunnel_info *info = skb_tunnel_info(skb);
2275 const struct metadata_dst *md = this_cpu_ptr(md_dst);
2276
2277 if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
2278 return -EINVAL;
2279 if (unlikely(size > IP_TUNNEL_OPTS_MAX))
2280 return -ENOMEM;
2281
2282 ip_tunnel_info_opts_set(info, from, size);
2283
2284 return 0;
2285 }
2286
2287 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
2288 .func = bpf_skb_set_tunnel_opt,
2289 .gpl_only = false,
2290 .ret_type = RET_INTEGER,
2291 .arg1_type = ARG_PTR_TO_CTX,
2292 .arg2_type = ARG_PTR_TO_STACK,
2293 .arg3_type = ARG_CONST_STACK_SIZE,
2294 };
2295
2296 static const struct bpf_func_proto *
2297 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
2298 {
2299 if (!md_dst) {
2300 /* Race is not possible, since it's called from verifier
2301 * that is holding verifier mutex.
2302 */
2303 md_dst = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
2304 GFP_KERNEL);
2305 if (!md_dst)
2306 return NULL;
2307 }
2308
2309 switch (which) {
2310 case BPF_FUNC_skb_set_tunnel_key:
2311 return &bpf_skb_set_tunnel_key_proto;
2312 case BPF_FUNC_skb_set_tunnel_opt:
2313 return &bpf_skb_set_tunnel_opt_proto;
2314 default:
2315 return NULL;
2316 }
2317 }
2318
2319 #ifdef CONFIG_SOCK_CGROUP_DATA
2320 static u64 bpf_skb_in_cgroup(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
2321 {
2322 struct sk_buff *skb = (struct sk_buff *)(long)r1;
2323 struct bpf_map *map = (struct bpf_map *)(long)r2;
2324 struct bpf_array *array = container_of(map, struct bpf_array, map);
2325 struct cgroup *cgrp;
2326 struct sock *sk;
2327 u32 i = (u32)r3;
2328
2329 sk = skb->sk;
2330 if (!sk || !sk_fullsock(sk))
2331 return -ENOENT;
2332
2333 if (unlikely(i >= array->map.max_entries))
2334 return -E2BIG;
2335
2336 cgrp = READ_ONCE(array->ptrs[i]);
2337 if (unlikely(!cgrp))
2338 return -EAGAIN;
2339
2340 return cgroup_is_descendant(sock_cgroup_ptr(&sk->sk_cgrp_data), cgrp);
2341 }
2342
2343 static const struct bpf_func_proto bpf_skb_in_cgroup_proto = {
2344 .func = bpf_skb_in_cgroup,
2345 .gpl_only = false,
2346 .ret_type = RET_INTEGER,
2347 .arg1_type = ARG_PTR_TO_CTX,
2348 .arg2_type = ARG_CONST_MAP_PTR,
2349 .arg3_type = ARG_ANYTHING,
2350 };
2351 #endif
2352
2353 static const struct bpf_func_proto *
2354 sk_filter_func_proto(enum bpf_func_id func_id)
2355 {
2356 switch (func_id) {
2357 case BPF_FUNC_map_lookup_elem:
2358 return &bpf_map_lookup_elem_proto;
2359 case BPF_FUNC_map_update_elem:
2360 return &bpf_map_update_elem_proto;
2361 case BPF_FUNC_map_delete_elem:
2362 return &bpf_map_delete_elem_proto;
2363 case BPF_FUNC_get_prandom_u32:
2364 return &bpf_get_prandom_u32_proto;
2365 case BPF_FUNC_get_smp_processor_id:
2366 return &bpf_get_raw_smp_processor_id_proto;
2367 case BPF_FUNC_tail_call:
2368 return &bpf_tail_call_proto;
2369 case BPF_FUNC_ktime_get_ns:
2370 return &bpf_ktime_get_ns_proto;
2371 case BPF_FUNC_trace_printk:
2372 if (capable(CAP_SYS_ADMIN))
2373 return bpf_get_trace_printk_proto();
2374 default:
2375 return NULL;
2376 }
2377 }
2378
2379 static const struct bpf_func_proto *
2380 tc_cls_act_func_proto(enum bpf_func_id func_id)
2381 {
2382 switch (func_id) {
2383 case BPF_FUNC_skb_store_bytes:
2384 return &bpf_skb_store_bytes_proto;
2385 case BPF_FUNC_skb_load_bytes:
2386 return &bpf_skb_load_bytes_proto;
2387 case BPF_FUNC_csum_diff:
2388 return &bpf_csum_diff_proto;
2389 case BPF_FUNC_l3_csum_replace:
2390 return &bpf_l3_csum_replace_proto;
2391 case BPF_FUNC_l4_csum_replace:
2392 return &bpf_l4_csum_replace_proto;
2393 case BPF_FUNC_clone_redirect:
2394 return &bpf_clone_redirect_proto;
2395 case BPF_FUNC_get_cgroup_classid:
2396 return &bpf_get_cgroup_classid_proto;
2397 case BPF_FUNC_skb_vlan_push:
2398 return &bpf_skb_vlan_push_proto;
2399 case BPF_FUNC_skb_vlan_pop:
2400 return &bpf_skb_vlan_pop_proto;
2401 case BPF_FUNC_skb_change_proto:
2402 return &bpf_skb_change_proto_proto;
2403 case BPF_FUNC_skb_change_type:
2404 return &bpf_skb_change_type_proto;
2405 case BPF_FUNC_skb_get_tunnel_key:
2406 return &bpf_skb_get_tunnel_key_proto;
2407 case BPF_FUNC_skb_set_tunnel_key:
2408 return bpf_get_skb_set_tunnel_proto(func_id);
2409 case BPF_FUNC_skb_get_tunnel_opt:
2410 return &bpf_skb_get_tunnel_opt_proto;
2411 case BPF_FUNC_skb_set_tunnel_opt:
2412 return bpf_get_skb_set_tunnel_proto(func_id);
2413 case BPF_FUNC_redirect:
2414 return &bpf_redirect_proto;
2415 case BPF_FUNC_get_route_realm:
2416 return &bpf_get_route_realm_proto;
2417 case BPF_FUNC_get_hash_recalc:
2418 return &bpf_get_hash_recalc_proto;
2419 case BPF_FUNC_perf_event_output:
2420 return &bpf_skb_event_output_proto;
2421 case BPF_FUNC_get_smp_processor_id:
2422 return &bpf_get_smp_processor_id_proto;
2423 #ifdef CONFIG_SOCK_CGROUP_DATA
2424 case BPF_FUNC_skb_in_cgroup:
2425 return &bpf_skb_in_cgroup_proto;
2426 #endif
2427 default:
2428 return sk_filter_func_proto(func_id);
2429 }
2430 }
2431
2432 static const struct bpf_func_proto *
2433 xdp_func_proto(enum bpf_func_id func_id)
2434 {
2435 return sk_filter_func_proto(func_id);
2436 }
2437
2438 static bool __is_valid_access(int off, int size, enum bpf_access_type type)
2439 {
2440 if (off < 0 || off >= sizeof(struct __sk_buff))
2441 return false;
2442 /* The verifier guarantees that size > 0. */
2443 if (off % size != 0)
2444 return false;
2445 if (size != sizeof(__u32))
2446 return false;
2447
2448 return true;
2449 }
2450
2451 static bool sk_filter_is_valid_access(int off, int size,
2452 enum bpf_access_type type,
2453 enum bpf_reg_type *reg_type)
2454 {
2455 switch (off) {
2456 case offsetof(struct __sk_buff, tc_classid):
2457 case offsetof(struct __sk_buff, data):
2458 case offsetof(struct __sk_buff, data_end):
2459 return false;
2460 }
2461
2462 if (type == BPF_WRITE) {
2463 switch (off) {
2464 case offsetof(struct __sk_buff, cb[0]) ...
2465 offsetof(struct __sk_buff, cb[4]):
2466 break;
2467 default:
2468 return false;
2469 }
2470 }
2471
2472 return __is_valid_access(off, size, type);
2473 }
2474
2475 static bool tc_cls_act_is_valid_access(int off, int size,
2476 enum bpf_access_type type,
2477 enum bpf_reg_type *reg_type)
2478 {
2479 if (type == BPF_WRITE) {
2480 switch (off) {
2481 case offsetof(struct __sk_buff, mark):
2482 case offsetof(struct __sk_buff, tc_index):
2483 case offsetof(struct __sk_buff, priority):
2484 case offsetof(struct __sk_buff, cb[0]) ...
2485 offsetof(struct __sk_buff, cb[4]):
2486 case offsetof(struct __sk_buff, tc_classid):
2487 break;
2488 default:
2489 return false;
2490 }
2491 }
2492
2493 switch (off) {
2494 case offsetof(struct __sk_buff, data):
2495 *reg_type = PTR_TO_PACKET;
2496 break;
2497 case offsetof(struct __sk_buff, data_end):
2498 *reg_type = PTR_TO_PACKET_END;
2499 break;
2500 }
2501
2502 return __is_valid_access(off, size, type);
2503 }
2504
2505 static bool __is_valid_xdp_access(int off, int size,
2506 enum bpf_access_type type)
2507 {
2508 if (off < 0 || off >= sizeof(struct xdp_md))
2509 return false;
2510 if (off % size != 0)
2511 return false;
2512 if (size != 4)
2513 return false;
2514
2515 return true;
2516 }
2517
2518 static bool xdp_is_valid_access(int off, int size,
2519 enum bpf_access_type type,
2520 enum bpf_reg_type *reg_type)
2521 {
2522 if (type == BPF_WRITE)
2523 return false;
2524
2525 switch (off) {
2526 case offsetof(struct xdp_md, data):
2527 *reg_type = PTR_TO_PACKET;
2528 break;
2529 case offsetof(struct xdp_md, data_end):
2530 *reg_type = PTR_TO_PACKET_END;
2531 break;
2532 }
2533
2534 return __is_valid_xdp_access(off, size, type);
2535 }
2536
2537 void bpf_warn_invalid_xdp_action(u32 act)
2538 {
2539 WARN_ONCE(1, "Illegal XDP return value %u, expect packet loss\n", act);
2540 }
2541 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
2542
2543 static u32 bpf_net_convert_ctx_access(enum bpf_access_type type, int dst_reg,
2544 int src_reg, int ctx_off,
2545 struct bpf_insn *insn_buf,
2546 struct bpf_prog *prog)
2547 {
2548 struct bpf_insn *insn = insn_buf;
2549
2550 switch (ctx_off) {
2551 case offsetof(struct __sk_buff, len):
2552 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
2553
2554 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2555 offsetof(struct sk_buff, len));
2556 break;
2557
2558 case offsetof(struct __sk_buff, protocol):
2559 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
2560
2561 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2562 offsetof(struct sk_buff, protocol));
2563 break;
2564
2565 case offsetof(struct __sk_buff, vlan_proto):
2566 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
2567
2568 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2569 offsetof(struct sk_buff, vlan_proto));
2570 break;
2571
2572 case offsetof(struct __sk_buff, priority):
2573 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
2574
2575 if (type == BPF_WRITE)
2576 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2577 offsetof(struct sk_buff, priority));
2578 else
2579 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2580 offsetof(struct sk_buff, priority));
2581 break;
2582
2583 case offsetof(struct __sk_buff, ingress_ifindex):
2584 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
2585
2586 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2587 offsetof(struct sk_buff, skb_iif));
2588 break;
2589
2590 case offsetof(struct __sk_buff, ifindex):
2591 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
2592
2593 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, dev)),
2594 dst_reg, src_reg,
2595 offsetof(struct sk_buff, dev));
2596 *insn++ = BPF_JMP_IMM(BPF_JEQ, dst_reg, 0, 1);
2597 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
2598 offsetof(struct net_device, ifindex));
2599 break;
2600
2601 case offsetof(struct __sk_buff, hash):
2602 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
2603
2604 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2605 offsetof(struct sk_buff, hash));
2606 break;
2607
2608 case offsetof(struct __sk_buff, mark):
2609 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
2610
2611 if (type == BPF_WRITE)
2612 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2613 offsetof(struct sk_buff, mark));
2614 else
2615 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2616 offsetof(struct sk_buff, mark));
2617 break;
2618
2619 case offsetof(struct __sk_buff, pkt_type):
2620 return convert_skb_access(SKF_AD_PKTTYPE, dst_reg, src_reg, insn);
2621
2622 case offsetof(struct __sk_buff, queue_mapping):
2623 return convert_skb_access(SKF_AD_QUEUE, dst_reg, src_reg, insn);
2624
2625 case offsetof(struct __sk_buff, vlan_present):
2626 return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
2627 dst_reg, src_reg, insn);
2628
2629 case offsetof(struct __sk_buff, vlan_tci):
2630 return convert_skb_access(SKF_AD_VLAN_TAG,
2631 dst_reg, src_reg, insn);
2632
2633 case offsetof(struct __sk_buff, cb[0]) ...
2634 offsetof(struct __sk_buff, cb[4]):
2635 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
2636
2637 prog->cb_access = 1;
2638 ctx_off -= offsetof(struct __sk_buff, cb[0]);
2639 ctx_off += offsetof(struct sk_buff, cb);
2640 ctx_off += offsetof(struct qdisc_skb_cb, data);
2641 if (type == BPF_WRITE)
2642 *insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2643 else
2644 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2645 break;
2646
2647 case offsetof(struct __sk_buff, tc_classid):
2648 ctx_off -= offsetof(struct __sk_buff, tc_classid);
2649 ctx_off += offsetof(struct sk_buff, cb);
2650 ctx_off += offsetof(struct qdisc_skb_cb, tc_classid);
2651 if (type == BPF_WRITE)
2652 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2653 else
2654 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2655 break;
2656
2657 case offsetof(struct __sk_buff, data):
2658 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct sk_buff, data)),
2659 dst_reg, src_reg,
2660 offsetof(struct sk_buff, data));
2661 break;
2662
2663 case offsetof(struct __sk_buff, data_end):
2664 ctx_off -= offsetof(struct __sk_buff, data_end);
2665 ctx_off += offsetof(struct sk_buff, cb);
2666 ctx_off += offsetof(struct bpf_skb_data_end, data_end);
2667 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(sizeof(void *)),
2668 dst_reg, src_reg, ctx_off);
2669 break;
2670
2671 case offsetof(struct __sk_buff, tc_index):
2672 #ifdef CONFIG_NET_SCHED
2673 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
2674
2675 if (type == BPF_WRITE)
2676 *insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg,
2677 offsetof(struct sk_buff, tc_index));
2678 else
2679 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2680 offsetof(struct sk_buff, tc_index));
2681 break;
2682 #else
2683 if (type == BPF_WRITE)
2684 *insn++ = BPF_MOV64_REG(dst_reg, dst_reg);
2685 else
2686 *insn++ = BPF_MOV64_IMM(dst_reg, 0);
2687 break;
2688 #endif
2689 }
2690
2691 return insn - insn_buf;
2692 }
2693
2694 static u32 xdp_convert_ctx_access(enum bpf_access_type type, int dst_reg,
2695 int src_reg, int ctx_off,
2696 struct bpf_insn *insn_buf,
2697 struct bpf_prog *prog)
2698 {
2699 struct bpf_insn *insn = insn_buf;
2700
2701 switch (ctx_off) {
2702 case offsetof(struct xdp_md, data):
2703 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct xdp_buff, data)),
2704 dst_reg, src_reg,
2705 offsetof(struct xdp_buff, data));
2706 break;
2707 case offsetof(struct xdp_md, data_end):
2708 *insn++ = BPF_LDX_MEM(bytes_to_bpf_size(FIELD_SIZEOF(struct xdp_buff, data_end)),
2709 dst_reg, src_reg,
2710 offsetof(struct xdp_buff, data_end));
2711 break;
2712 }
2713
2714 return insn - insn_buf;
2715 }
2716
2717 static const struct bpf_verifier_ops sk_filter_ops = {
2718 .get_func_proto = sk_filter_func_proto,
2719 .is_valid_access = sk_filter_is_valid_access,
2720 .convert_ctx_access = bpf_net_convert_ctx_access,
2721 };
2722
2723 static const struct bpf_verifier_ops tc_cls_act_ops = {
2724 .get_func_proto = tc_cls_act_func_proto,
2725 .is_valid_access = tc_cls_act_is_valid_access,
2726 .convert_ctx_access = bpf_net_convert_ctx_access,
2727 };
2728
2729 static const struct bpf_verifier_ops xdp_ops = {
2730 .get_func_proto = xdp_func_proto,
2731 .is_valid_access = xdp_is_valid_access,
2732 .convert_ctx_access = xdp_convert_ctx_access,
2733 };
2734
2735 static struct bpf_prog_type_list sk_filter_type __read_mostly = {
2736 .ops = &sk_filter_ops,
2737 .type = BPF_PROG_TYPE_SOCKET_FILTER,
2738 };
2739
2740 static struct bpf_prog_type_list sched_cls_type __read_mostly = {
2741 .ops = &tc_cls_act_ops,
2742 .type = BPF_PROG_TYPE_SCHED_CLS,
2743 };
2744
2745 static struct bpf_prog_type_list sched_act_type __read_mostly = {
2746 .ops = &tc_cls_act_ops,
2747 .type = BPF_PROG_TYPE_SCHED_ACT,
2748 };
2749
2750 static struct bpf_prog_type_list xdp_type __read_mostly = {
2751 .ops = &xdp_ops,
2752 .type = BPF_PROG_TYPE_XDP,
2753 };
2754
2755 static int __init register_sk_filter_ops(void)
2756 {
2757 bpf_register_prog_type(&sk_filter_type);
2758 bpf_register_prog_type(&sched_cls_type);
2759 bpf_register_prog_type(&sched_act_type);
2760 bpf_register_prog_type(&xdp_type);
2761
2762 return 0;
2763 }
2764 late_initcall(register_sk_filter_ops);
2765
2766 int sk_detach_filter(struct sock *sk)
2767 {
2768 int ret = -ENOENT;
2769 struct sk_filter *filter;
2770
2771 if (sock_flag(sk, SOCK_FILTER_LOCKED))
2772 return -EPERM;
2773
2774 filter = rcu_dereference_protected(sk->sk_filter,
2775 lockdep_sock_is_held(sk));
2776 if (filter) {
2777 RCU_INIT_POINTER(sk->sk_filter, NULL);
2778 sk_filter_uncharge(sk, filter);
2779 ret = 0;
2780 }
2781
2782 return ret;
2783 }
2784 EXPORT_SYMBOL_GPL(sk_detach_filter);
2785
2786 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
2787 unsigned int len)
2788 {
2789 struct sock_fprog_kern *fprog;
2790 struct sk_filter *filter;
2791 int ret = 0;
2792
2793 lock_sock(sk);
2794 filter = rcu_dereference_protected(sk->sk_filter,
2795 lockdep_sock_is_held(sk));
2796 if (!filter)
2797 goto out;
2798
2799 /* We're copying the filter that has been originally attached,
2800 * so no conversion/decode needed anymore. eBPF programs that
2801 * have no original program cannot be dumped through this.
2802 */
2803 ret = -EACCES;
2804 fprog = filter->prog->orig_prog;
2805 if (!fprog)
2806 goto out;
2807
2808 ret = fprog->len;
2809 if (!len)
2810 /* User space only enquires number of filter blocks. */
2811 goto out;
2812
2813 ret = -EINVAL;
2814 if (len < fprog->len)
2815 goto out;
2816
2817 ret = -EFAULT;
2818 if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
2819 goto out;
2820
2821 /* Instead of bytes, the API requests to return the number
2822 * of filter blocks.
2823 */
2824 ret = fprog->len;
2825 out:
2826 release_sock(sk);
2827 return ret;
2828 }
Linux kernel - Deliverables
This page took 0.172158 seconds and 6 git commands to generate.