1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
3
4 #include <linux/bpf.h>
5 #include <linux/filter.h>
6 #include <linux/errno.h>
7 #include <linux/file.h>
8 #include <linux/net.h>
9 #include <linux/workqueue.h>
10 #include <linux/skmsg.h>
11 #include <linux/list.h>
12 #include <linux/jhash.h>
13 #include <linux/sock_diag.h>
14 #include <net/udp.h>
15
16 struct bpf_stab {
17 struct bpf_map map;
18 struct sock **sks;
19 struct sk_psock_progs progs;
20 raw_spinlock_t lock;
21 };
22
23 #define SOCK_CREATE_FLAG_MASK \
24 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
25
26 static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
27 {
28 struct bpf_stab *stab;
29 u64 cost;
30 int err;
31
32 if (!capable(CAP_NET_ADMIN))
33 return ERR_PTR(-EPERM);
34 if (attr->max_entries == 0 ||
35 attr->key_size != 4 ||
36 (attr->value_size != sizeof(u32) &&
37 attr->value_size != sizeof(u64)) ||
38 attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
39 return ERR_PTR(-EINVAL);
40
41 stab = kzalloc(sizeof(*stab), GFP_USER);
42 if (!stab)
43 return ERR_PTR(-ENOMEM);
44
45 bpf_map_init_from_attr(&stab->map, attr);
46 raw_spin_lock_init(&stab->lock);
47
48 /* Make sure page count doesn't overflow. */
49 cost = (u64) stab->map.max_entries * sizeof(struct sock *);
50 err = bpf_map_charge_init(&stab->map.memory, cost);
51 if (err)
52 goto free_stab;
53
54 stab->sks = bpf_map_area_alloc(stab->map.max_entries *
55 sizeof(struct sock *),
56 stab->map.numa_node);
57 if (stab->sks)
58 return &stab->map;
59 err = -ENOMEM;
60 bpf_map_charge_finish(&stab->map.memory);
61 free_stab:
62 kfree(stab);
63 return ERR_PTR(err);
64 }
65
66 int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog *prog)
67 {
68 u32 ufd = attr->target_fd;
69 struct bpf_map *map;
70 struct fd f;
71 int ret;
72
73 if (attr->attach_flags || attr->replace_bpf_fd)
74 return -EINVAL;
75
76 f = fdget(ufd);
77 map = __bpf_map_get(f);
78 if (IS_ERR(map))
79 return PTR_ERR(map);
80 ret = sock_map_prog_update(map, prog, NULL, attr->attach_type);
81 fdput(f);
82 return ret;
83 }
84
85 int sock_map_prog_detach(const union bpf_attr *attr, enum bpf_prog_type ptype)
86 {
87 u32 ufd = attr->target_fd;
88 struct bpf_prog *prog;
89 struct bpf_map *map;
90 struct fd f;
91 int ret;
92
93 if (attr->attach_flags || attr->replace_bpf_fd)
94 return -EINVAL;
95
96 f = fdget(ufd);
97 map = __bpf_map_get(f);
98 if (IS_ERR(map))
99 return PTR_ERR(map);
100
101 prog = bpf_prog_get(attr->attach_bpf_fd);
102 if (IS_ERR(prog)) {
103 ret = PTR_ERR(prog);
104 goto put_map;
105 }
106
107 if (prog->type != ptype) {
108 ret = -EINVAL;
109 goto put_prog;
110 }
111
112 ret = sock_map_prog_update(map, NULL, prog, attr->attach_type);
113 put_prog:
114 bpf_prog_put(prog);
115 put_map:
116 fdput(f);
117 return ret;
118 }
119
120 static void sock_map_sk_acquire(struct sock *sk)
121 __acquires(&sk->sk_lock.slock)
122 {
123 lock_sock(sk);
124 preempt_disable();
125 rcu_read_lock();
126 }
127
128 static void sock_map_sk_release(struct sock *sk)
129 __releases(&sk->sk_lock.slock)
130 {
131 rcu_read_unlock();
132 preempt_enable();
133 release_sock(sk);
134 }
135
136 static void sock_map_add_link(struct sk_psock *psock,
137 struct sk_psock_link *link,
138 struct bpf_map *map, void *link_raw)
139 {
140 link->link_raw = link_raw;
141 link->map = map;
142 spin_lock_bh(&psock->link_lock);
143 list_add_tail(&link->list, &psock->link);
144 spin_unlock_bh(&psock->link_lock);
145 }
146
147 static void sock_map_del_link(struct sock *sk,
148 struct sk_psock *psock, void *link_raw)
149 {
150 struct sk_psock_link *link, *tmp;
151 bool strp_stop = false;
152
153 spin_lock_bh(&psock->link_lock);
154 list_for_each_entry_safe(link, tmp, &psock->link, list) {
155 if (link->link_raw == link_raw) {
156 struct bpf_map *map = link->map;
157 struct bpf_stab *stab = container_of(map, struct bpf_stab,
158 map);
159 if (psock->parser.enabled && stab->progs.skb_parser)
160 strp_stop = true;
161 list_del(&link->list);
162 sk_psock_free_link(link);
163 }
164 }
165 spin_unlock_bh(&psock->link_lock);
166 if (strp_stop) {
167 write_lock_bh(&sk->sk_callback_lock);
168 sk_psock_stop_strp(sk, psock);
169 write_unlock_bh(&sk->sk_callback_lock);
170 }
171 }
172
173 static void sock_map_unref(struct sock *sk, void *link_raw)
174 {
175 struct sk_psock *psock = sk_psock(sk);
176
177 if (likely(psock)) {
178 sock_map_del_link(sk, psock, link_raw);
179 sk_psock_put(sk, psock);
180 }
181 }
182
183 static int sock_map_init_proto(struct sock *sk, struct sk_psock *psock)
184 {
185 struct proto *prot;
186
187 sock_owned_by_me(sk);
188
189 switch (sk->sk_type) {
190 case SOCK_STREAM:
191 prot = tcp_bpf_get_proto(sk, psock);
192 break;
193
194 case SOCK_DGRAM:
195 prot = udp_bpf_get_proto(sk, psock);
196 break;
197
198 default:
199 return -EINVAL;
200 }
201
202 if (IS_ERR(prot))
203 return PTR_ERR(prot);
204
205 sk_psock_update_proto(sk, psock, prot);
206 return 0;
207 }
208
209 static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
210 {
211 struct sk_psock *psock;
212
213 rcu_read_lock();
214 psock = sk_psock(sk);
215 if (psock) {
216 if (sk->sk_prot->close != sock_map_close) {
217 psock = ERR_PTR(-EBUSY);
218 goto out;
219 }
220
221 if (!refcount_inc_not_zero(&psock->refcnt))
222 psock = ERR_PTR(-EBUSY);
223 }
224 out:
225 rcu_read_unlock();
226 return psock;
227 }
228
229 static int sock_map_link(struct bpf_map *map, struct sk_psock_progs *progs,
230 struct sock *sk)
231 {
232 struct bpf_prog *msg_parser, *skb_parser, *skb_verdict;
233 struct sk_psock *psock;
234 bool skb_progs;
235 int ret;
236
237 skb_verdict = READ_ONCE(progs->skb_verdict);
238 skb_parser = READ_ONCE(progs->skb_parser);
239 skb_progs = skb_parser && skb_verdict;
240 if (skb_progs) {
241 skb_verdict = bpf_prog_inc_not_zero(skb_verdict);
242 if (IS_ERR(skb_verdict))
243 return PTR_ERR(skb_verdict);
244 skb_parser = bpf_prog_inc_not_zero(skb_parser);
245 if (IS_ERR(skb_parser)) {
246 bpf_prog_put(skb_verdict);
247 return PTR_ERR(skb_parser);
248 }
249 }
250
251 msg_parser = READ_ONCE(progs->msg_parser);
252 if (msg_parser) {
253 msg_parser = bpf_prog_inc_not_zero(msg_parser);
254 if (IS_ERR(msg_parser)) {
255 ret = PTR_ERR(msg_parser);
256 goto out;
257 }
258 }
259
260 psock = sock_map_psock_get_checked(sk);
261 if (IS_ERR(psock)) {
262 ret = PTR_ERR(psock);
263 goto out_progs;
264 }
265
266 if (psock) {
267 if ((msg_parser && READ_ONCE(psock->progs.msg_parser)) ||
268 (skb_progs && READ_ONCE(psock->progs.skb_parser))) {
269 sk_psock_put(sk, psock);
270 ret = -EBUSY;
271 goto out_progs;
272 }
273 } else {
274 psock = sk_psock_init(sk, map->numa_node);
275 if (!psock) {
276 ret = -ENOMEM;
277 goto out_progs;
278 }
279 }
280
281 if (msg_parser)
282 psock_set_prog(&psock->progs.msg_parser, msg_parser);
283
284 ret = sock_map_init_proto(sk, psock);
285 if (ret < 0)
286 goto out_drop;
287
288 write_lock_bh(&sk->sk_callback_lock);
289 if (skb_progs && !psock->parser.enabled) {
290 ret = sk_psock_init_strp(sk, psock);
291 if (ret) {
292 write_unlock_bh(&sk->sk_callback_lock);
293 goto out_drop;
294 }
295 psock_set_prog(&psock->progs.skb_verdict, skb_verdict);
296 psock_set_prog(&psock->progs.skb_parser, skb_parser);
297 sk_psock_start_strp(sk, psock);
298 }
299 write_unlock_bh(&sk->sk_callback_lock);
300 return 0;
301 out_drop:
302 sk_psock_put(sk, psock);
303 out_progs:
304 if (msg_parser)
305 bpf_prog_put(msg_parser);
306 out:
307 if (skb_progs) {
308 bpf_prog_put(skb_verdict);
309 bpf_prog_put(skb_parser);
310 }
311 return ret;
312 }
313
314 static int sock_map_link_no_progs(struct bpf_map *map, struct sock *sk)
315 {
316 struct sk_psock *psock;
317 int ret;
318
319 psock = sock_map_psock_get_checked(sk);
320 if (IS_ERR(psock))
321 return PTR_ERR(psock);
322
323 if (!psock) {
324 psock = sk_psock_init(sk, map->numa_node);
325 if (!psock)
326 return -ENOMEM;
327 }
328
329 ret = sock_map_init_proto(sk, psock);
330 if (ret < 0)
331 sk_psock_put(sk, psock);
332 return ret;
333 }
334
335 static void sock_map_free(struct bpf_map *map)
336 {
337 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
338 int i;
339
340 /* After the sync no updates or deletes will be in-flight so it
341 * is safe to walk map and remove entries without risking a race
342 * in EEXIST update case.
343 */
344 synchronize_rcu();
345 for (i = 0; i < stab->map.max_entries; i++) {
346 struct sock **psk = &stab->sks[i];
347 struct sock *sk;
348
349 sk = xchg(psk, NULL);
350 if (sk) {
351 lock_sock(sk);
352 rcu_read_lock();
353 sock_map_unref(sk, psk);
354 rcu_read_unlock();
355 release_sock(sk);
356 }
357 }
358
359 /* wait for psock readers accessing its map link */
360 synchronize_rcu();
361
362 bpf_map_area_free(stab->sks);
363 kfree(stab);
364 }
365
366 static void sock_map_release_progs(struct bpf_map *map)
367 {
368 psock_progs_drop(&container_of(map, struct bpf_stab, map)->progs);
369 }
370
371 static struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
372 {
373 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
374
375 WARN_ON_ONCE(!rcu_read_lock_held());
376
377 if (unlikely(key >= map->max_entries))
378 return NULL;
379 return READ_ONCE(stab->sks[key]);
380 }
381
382 static void *sock_map_lookup(struct bpf_map *map, void *key)
383 {
384 return __sock_map_lookup_elem(map, *(u32 *)key);
385 }
386
387 static void *sock_map_lookup_sys(struct bpf_map *map, void *key)
388 {
389 struct sock *sk;
390
391 if (map->value_size != sizeof(u64))
392 return ERR_PTR(-ENOSPC);
393
394 sk = __sock_map_lookup_elem(map, *(u32 *)key);
395 if (!sk)
396 return ERR_PTR(-ENOENT);
397
398 sock_gen_cookie(sk);
399 return &sk->sk_cookie;
400 }
401
402 static int __sock_map_delete(struct bpf_stab *stab, struct sock *sk_test,
403 struct sock **psk)
404 {
405 struct sock *sk;
406 int err = 0;
407
408 raw_spin_lock_bh(&stab->lock);
409 sk = *psk;
410 if (!sk_test || sk_test == sk)
411 sk = xchg(psk, NULL);
412
413 if (likely(sk))
414 sock_map_unref(sk, psk);
415 else
416 err = -EINVAL;
417
418 raw_spin_unlock_bh(&stab->lock);
419 return err;
420 }
421
422 static void sock_map_delete_from_link(struct bpf_map *map, struct sock *sk,
423 void *link_raw)
424 {
425 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
426
427 __sock_map_delete(stab, sk, link_raw);
428 }
429
430 static int sock_map_delete_elem(struct bpf_map *map, void *key)
431 {
432 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
433 u32 i = *(u32 *)key;
434 struct sock **psk;
435
436 if (unlikely(i >= map->max_entries))
437 return -EINVAL;
438
439 psk = &stab->sks[i];
440 return __sock_map_delete(stab, NULL, psk);
441 }
442
443 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next)
444 {
445 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
446 u32 i = key ? *(u32 *)key : U32_MAX;
447 u32 *key_next = next;
448
449 if (i == stab->map.max_entries - 1)
450 return -ENOENT;
451 if (i >= stab->map.max_entries)
452 *key_next = 0;
453 else
454 *key_next = i + 1;
455 return 0;
456 }
457
458 static bool sock_map_redirect_allowed(const struct sock *sk);
459
460 static int sock_map_update_common(struct bpf_map *map, u32 idx,
461 struct sock *sk, u64 flags)
462 {
463 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
464 struct sk_psock_link *link;
465 struct sk_psock *psock;
466 struct sock *osk;
467 int ret;
468
469 WARN_ON_ONCE(!rcu_read_lock_held());
470 if (unlikely(flags > BPF_EXIST))
471 return -EINVAL;
472 if (unlikely(idx >= map->max_entries))
473 return -E2BIG;
474 if (inet_csk_has_ulp(sk))
475 return -EINVAL;
476
477 link = sk_psock_init_link();
478 if (!link)
479 return -ENOMEM;
480
481 /* Only sockets we can redirect into/from in BPF need to hold
482 * refs to parser/verdict progs and have their sk_data_ready
483 * and sk_write_space callbacks overridden.
484 */
485 if (sock_map_redirect_allowed(sk))
486 ret = sock_map_link(map, &stab->progs, sk);
487 else
488 ret = sock_map_link_no_progs(map, sk);
489 if (ret < 0)
490 goto out_free;
491
492 psock = sk_psock(sk);
493 WARN_ON_ONCE(!psock);
494
495 raw_spin_lock_bh(&stab->lock);
496 osk = stab->sks[idx];
497 if (osk && flags == BPF_NOEXIST) {
498 ret = -EEXIST;
499 goto out_unlock;
500 } else if (!osk && flags == BPF_EXIST) {
501 ret = -ENOENT;
502 goto out_unlock;
503 }
504
505 sock_map_add_link(psock, link, map, &stab->sks[idx]);
506 stab->sks[idx] = sk;
507 if (osk)
508 sock_map_unref(osk, &stab->sks[idx]);
509 raw_spin_unlock_bh(&stab->lock);
510 return 0;
511 out_unlock:
512 raw_spin_unlock_bh(&stab->lock);
513 if (psock)
514 sk_psock_put(sk, psock);
515 out_free:
516 sk_psock_free_link(link);
517 return ret;
518 }
519
520 static bool sock_map_op_okay(const struct bpf_sock_ops_kern *ops)
521 {
522 return ops->op == BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB ||
523 ops->op == BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB ||
524 ops->op == BPF_SOCK_OPS_TCP_LISTEN_CB;
525 }
526
527 static bool sk_is_tcp(const struct sock *sk)
528 {
529 return sk->sk_type == SOCK_STREAM &&
530 sk->sk_protocol == IPPROTO_TCP;
531 }
532
533 static bool sk_is_udp(const struct sock *sk)
534 {
535 return sk->sk_type == SOCK_DGRAM &&
536 sk->sk_protocol == IPPROTO_UDP;
537 }
538
539 static bool sock_map_redirect_allowed(const struct sock *sk)
540 {
541 return sk_is_tcp(sk) && sk->sk_state != TCP_LISTEN;
542 }
543
544 static bool sock_map_sk_is_suitable(const struct sock *sk)
545 {
546 return sk_is_tcp(sk) || sk_is_udp(sk);
547 }
548
549 static bool sock_map_sk_state_allowed(const struct sock *sk)
550 {
551 if (sk_is_tcp(sk))
552 return (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_LISTEN);
553 else if (sk_is_udp(sk))
554 return sk_hashed(sk);
555
556 return false;
557 }
558
559 static int sock_map_update_elem(struct bpf_map *map, void *key,
560 void *value, u64 flags)
561 {
562 u32 idx = *(u32 *)key;
563 struct socket *sock;
564 struct sock *sk;
565 int ret;
566 u64 ufd;
567
568 if (map->value_size == sizeof(u64))
569 ufd = *(u64 *)value;
570 else
571 ufd = *(u32 *)value;
572 if (ufd > S32_MAX)
573 return -EINVAL;
574
575 sock = sockfd_lookup(ufd, &ret);
576 if (!sock)
577 return ret;
578 sk = sock->sk;
579 if (!sk) {
580 ret = -EINVAL;
581 goto out;
582 }
583 if (!sock_map_sk_is_suitable(sk)) {
584 ret = -EOPNOTSUPP;
585 goto out;
586 }
587
588 sock_map_sk_acquire(sk);
589 if (!sock_map_sk_state_allowed(sk))
590 ret = -EOPNOTSUPP;
591 else
592 ret = sock_map_update_common(map, idx, sk, flags);
593 sock_map_sk_release(sk);
594 out:
595 fput(sock->file);
596 return ret;
597 }
598
599 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, sops,
600 struct bpf_map *, map, void *, key, u64, flags)
601 {
602 WARN_ON_ONCE(!rcu_read_lock_held());
603
604 if (likely(sock_map_sk_is_suitable(sops->sk) &&
605 sock_map_op_okay(sops)))
606 return sock_map_update_common(map, *(u32 *)key, sops->sk,
607 flags);
608 return -EOPNOTSUPP;
609 }
610
611 const struct bpf_func_proto bpf_sock_map_update_proto = {
612 .func = bpf_sock_map_update,
613 .gpl_only = false,
614 .pkt_access = true,
615 .ret_type = RET_INTEGER,
616 .arg1_type = ARG_PTR_TO_CTX,
617 .arg2_type = ARG_CONST_MAP_PTR,
618 .arg3_type = ARG_PTR_TO_MAP_KEY,
619 .arg4_type = ARG_ANYTHING,
620 };
621
622 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
623 struct bpf_map *, map, u32, key, u64, flags)
624 {
625 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
626 struct sock *sk;
627
628 if (unlikely(flags & ~(BPF_F_INGRESS)))
629 return SK_DROP;
630
631 sk = __sock_map_lookup_elem(map, key);
632 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
633 return SK_DROP;
634
635 tcb->bpf.flags = flags;
636 tcb->bpf.sk_redir = sk;
637 return SK_PASS;
638 }
639
640 const struct bpf_func_proto bpf_sk_redirect_map_proto = {
641 .func = bpf_sk_redirect_map,
642 .gpl_only = false,
643 .ret_type = RET_INTEGER,
644 .arg1_type = ARG_PTR_TO_CTX,
645 .arg2_type = ARG_CONST_MAP_PTR,
646 .arg3_type = ARG_ANYTHING,
647 .arg4_type = ARG_ANYTHING,
648 };
649
650 BPF_CALL_4(bpf_msg_redirect_map, struct sk_msg *, msg,
651 struct bpf_map *, map, u32, key, u64, flags)
652 {
653 struct sock *sk;
654
655 if (unlikely(flags & ~(BPF_F_INGRESS)))
656 return SK_DROP;
657
658 sk = __sock_map_lookup_elem(map, key);
659 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
660 return SK_DROP;
661
662 msg->flags = flags;
663 msg->sk_redir = sk;
664 return SK_PASS;
665 }
666
667 const struct bpf_func_proto bpf_msg_redirect_map_proto = {
668 .func = bpf_msg_redirect_map,
669 .gpl_only = false,
670 .ret_type = RET_INTEGER,
671 .arg1_type = ARG_PTR_TO_CTX,
672 .arg2_type = ARG_CONST_MAP_PTR,
673 .arg3_type = ARG_ANYTHING,
674 .arg4_type = ARG_ANYTHING,
675 };
676
677 const struct bpf_map_ops sock_map_ops = {
678 .map_alloc = sock_map_alloc,
679 .map_free = sock_map_free,
680 .map_get_next_key = sock_map_get_next_key,
681 .map_lookup_elem_sys_only = sock_map_lookup_sys,
682 .map_update_elem = sock_map_update_elem,
683 .map_delete_elem = sock_map_delete_elem,
684 .map_lookup_elem = sock_map_lookup,
685 .map_release_uref = sock_map_release_progs,
686 .map_check_btf = map_check_no_btf,
687 };
688
689 struct bpf_htab_elem {
690 struct rcu_head rcu;
691 u32 hash;
692 struct sock *sk;
693 struct hlist_node node;
694 u8 key[];
695 };
696
697 struct bpf_htab_bucket {
698 struct hlist_head head;
699 raw_spinlock_t lock;
700 };
701
702 struct bpf_htab {
703 struct bpf_map map;
704 struct bpf_htab_bucket *buckets;
705 u32 buckets_num;
706 u32 elem_size;
707 struct sk_psock_progs progs;
708 atomic_t count;
709 };
710
711 static inline u32 sock_hash_bucket_hash(const void *key, u32 len)
712 {
713 return jhash(key, len, 0);
714 }
715
716 static struct bpf_htab_bucket *sock_hash_select_bucket(struct bpf_htab *htab,
717 u32 hash)
718 {
719 return &htab->buckets[hash & (htab->buckets_num - 1)];
720 }
721
722 static struct bpf_htab_elem *
723 sock_hash_lookup_elem_raw(struct hlist_head *head, u32 hash, void *key,
724 u32 key_size)
725 {
726 struct bpf_htab_elem *elem;
727
728 hlist_for_each_entry_rcu(elem, head, node) {
729 if (elem->hash == hash &&
730 !memcmp(&elem->key, key, key_size))
731 return elem;
732 }
733
734 return NULL;
735 }
736
737 static struct sock *__sock_hash_lookup_elem(struct bpf_map *map, void *key)
738 {
739 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
740 u32 key_size = map->key_size, hash;
741 struct bpf_htab_bucket *bucket;
742 struct bpf_htab_elem *elem;
743
744 WARN_ON_ONCE(!rcu_read_lock_held());
745
746 hash = sock_hash_bucket_hash(key, key_size);
747 bucket = sock_hash_select_bucket(htab, hash);
748 elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
749
750 return elem ? elem->sk : NULL;
751 }
752
753 static void sock_hash_free_elem(struct bpf_htab *htab,
754 struct bpf_htab_elem *elem)
755 {
756 atomic_dec(&htab->count);
757 kfree_rcu(elem, rcu);
758 }
759
760 static void sock_hash_delete_from_link(struct bpf_map *map, struct sock *sk,
761 void *link_raw)
762 {
763 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
764 struct bpf_htab_elem *elem_probe, *elem = link_raw;
765 struct bpf_htab_bucket *bucket;
766
767 WARN_ON_ONCE(!rcu_read_lock_held());
768 bucket = sock_hash_select_bucket(htab, elem->hash);
769
770 /* elem may be deleted in parallel from the map, but access here
771 * is okay since it's going away only after RCU grace period.
772 * However, we need to check whether it's still present.
773 */
774 raw_spin_lock_bh(&bucket->lock);
775 elem_probe = sock_hash_lookup_elem_raw(&bucket->head, elem->hash,
776 elem->key, map->key_size);
777 if (elem_probe && elem_probe == elem) {
778 hlist_del_rcu(&elem->node);
779 sock_map_unref(elem->sk, elem);
780 sock_hash_free_elem(htab, elem);
781 }
782 raw_spin_unlock_bh(&bucket->lock);
783 }
784
785 static int sock_hash_delete_elem(struct bpf_map *map, void *key)
786 {
787 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
788 u32 hash, key_size = map->key_size;
789 struct bpf_htab_bucket *bucket;
790 struct bpf_htab_elem *elem;
791 int ret = -ENOENT;
792
793 hash = sock_hash_bucket_hash(key, key_size);
794 bucket = sock_hash_select_bucket(htab, hash);
795
796 raw_spin_lock_bh(&bucket->lock);
797 elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
798 if (elem) {
799 hlist_del_rcu(&elem->node);
800 sock_map_unref(elem->sk, elem);
801 sock_hash_free_elem(htab, elem);
802 ret = 0;
803 }
804 raw_spin_unlock_bh(&bucket->lock);
805 return ret;
806 }
807
808 static struct bpf_htab_elem *sock_hash_alloc_elem(struct bpf_htab *htab,
809 void *key, u32 key_size,
810 u32 hash, struct sock *sk,
811 struct bpf_htab_elem *old)
812 {
813 struct bpf_htab_elem *new;
814
815 if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
816 if (!old) {
817 atomic_dec(&htab->count);
818 return ERR_PTR(-E2BIG);
819 }
820 }
821
822 new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
823 htab->map.numa_node);
824 if (!new) {
825 atomic_dec(&htab->count);
826 return ERR_PTR(-ENOMEM);
827 }
828 memcpy(new->key, key, key_size);
829 new->sk = sk;
830 new->hash = hash;
831 return new;
832 }
833
834 static int sock_hash_update_common(struct bpf_map *map, void *key,
835 struct sock *sk, u64 flags)
836 {
837 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
838 u32 key_size = map->key_size, hash;
839 struct bpf_htab_elem *elem, *elem_new;
840 struct bpf_htab_bucket *bucket;
841 struct sk_psock_link *link;
842 struct sk_psock *psock;
843 int ret;
844
845 WARN_ON_ONCE(!rcu_read_lock_held());
846 if (unlikely(flags > BPF_EXIST))
847 return -EINVAL;
848 if (inet_csk_has_ulp(sk))
849 return -EINVAL;
850
851 link = sk_psock_init_link();
852 if (!link)
853 return -ENOMEM;
854
855 /* Only sockets we can redirect into/from in BPF need to hold
856 * refs to parser/verdict progs and have their sk_data_ready
857 * and sk_write_space callbacks overridden.
858 */
859 if (sock_map_redirect_allowed(sk))
860 ret = sock_map_link(map, &htab->progs, sk);
861 else
862 ret = sock_map_link_no_progs(map, sk);
863 if (ret < 0)
864 goto out_free;
865
866 psock = sk_psock(sk);
867 WARN_ON_ONCE(!psock);
868
869 hash = sock_hash_bucket_hash(key, key_size);
870 bucket = sock_hash_select_bucket(htab, hash);
871
872 raw_spin_lock_bh(&bucket->lock);
873 elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
874 if (elem && flags == BPF_NOEXIST) {
875 ret = -EEXIST;
876 goto out_unlock;
877 } else if (!elem && flags == BPF_EXIST) {
878 ret = -ENOENT;
879 goto out_unlock;
880 }
881
882 elem_new = sock_hash_alloc_elem(htab, key, key_size, hash, sk, elem);
883 if (IS_ERR(elem_new)) {
884 ret = PTR_ERR(elem_new);
885 goto out_unlock;
886 }
887
888 sock_map_add_link(psock, link, map, elem_new);
889 /* Add new element to the head of the list, so that
890 * concurrent search will find it before old elem.
891 */
892 hlist_add_head_rcu(&elem_new->node, &bucket->head);
893 if (elem) {
894 hlist_del_rcu(&elem->node);
895 sock_map_unref(elem->sk, elem);
896 sock_hash_free_elem(htab, elem);
897 }
898 raw_spin_unlock_bh(&bucket->lock);
899 return 0;
900 out_unlock:
901 raw_spin_unlock_bh(&bucket->lock);
902 sk_psock_put(sk, psock);
903 out_free:
904 sk_psock_free_link(link);
905 return ret;
906 }
907
908 static int sock_hash_update_elem(struct bpf_map *map, void *key,
909 void *value, u64 flags)
910 {
911 struct socket *sock;
912 struct sock *sk;
913 int ret;
914 u64 ufd;
915
916 if (map->value_size == sizeof(u64))
917 ufd = *(u64 *)value;
918 else
919 ufd = *(u32 *)value;
920 if (ufd > S32_MAX)
921 return -EINVAL;
922
923 sock = sockfd_lookup(ufd, &ret);
924 if (!sock)
925 return ret;
926 sk = sock->sk;
927 if (!sk) {
928 ret = -EINVAL;
929 goto out;
930 }
931 if (!sock_map_sk_is_suitable(sk)) {
932 ret = -EOPNOTSUPP;
933 goto out;
934 }
935
936 sock_map_sk_acquire(sk);
937 if (!sock_map_sk_state_allowed(sk))
938 ret = -EOPNOTSUPP;
939 else
940 ret = sock_hash_update_common(map, key, sk, flags);
941 sock_map_sk_release(sk);
942 out:
943 fput(sock->file);
944 return ret;
945 }
946
947 static int sock_hash_get_next_key(struct bpf_map *map, void *key,
948 void *key_next)
949 {
950 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
951 struct bpf_htab_elem *elem, *elem_next;
952 u32 hash, key_size = map->key_size;
953 struct hlist_head *head;
954 int i = 0;
955
956 if (!key)
957 goto find_first_elem;
958 hash = sock_hash_bucket_hash(key, key_size);
959 head = &sock_hash_select_bucket(htab, hash)->head;
960 elem = sock_hash_lookup_elem_raw(head, hash, key, key_size);
961 if (!elem)
962 goto find_first_elem;
963
964 elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&elem->node)),
965 struct bpf_htab_elem, node);
966 if (elem_next) {
967 memcpy(key_next, elem_next->key, key_size);
968 return 0;
969 }
970
971 i = hash & (htab->buckets_num - 1);
972 i++;
973 find_first_elem:
974 for (; i < htab->buckets_num; i++) {
975 head = &sock_hash_select_bucket(htab, i)->head;
976 elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
977 struct bpf_htab_elem, node);
978 if (elem_next) {
979 memcpy(key_next, elem_next->key, key_size);
980 return 0;
981 }
982 }
983
984 return -ENOENT;
985 }
986
987 static struct bpf_map *sock_hash_alloc(union bpf_attr *attr)
988 {
989 struct bpf_htab *htab;
990 int i, err;
991 u64 cost;
992
993 if (!capable(CAP_NET_ADMIN))
994 return ERR_PTR(-EPERM);
995 if (attr->max_entries == 0 ||
996 attr->key_size == 0 ||
997 (attr->value_size != sizeof(u32) &&
998 attr->value_size != sizeof(u64)) ||
999 attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
1000 return ERR_PTR(-EINVAL);
1001 if (attr->key_size > MAX_BPF_STACK)
1002 return ERR_PTR(-E2BIG);
1003
1004 htab = kzalloc(sizeof(*htab), GFP_USER);
1005 if (!htab)
1006 return ERR_PTR(-ENOMEM);
1007
1008 bpf_map_init_from_attr(&htab->map, attr);
1009
1010 htab->buckets_num = roundup_pow_of_two(htab->map.max_entries);
1011 htab->elem_size = sizeof(struct bpf_htab_elem) +
1012 round_up(htab->map.key_size, 8);
1013 if (htab->buckets_num == 0 ||
1014 htab->buckets_num > U32_MAX / sizeof(struct bpf_htab_bucket)) {
1015 err = -EINVAL;
1016 goto free_htab;
1017 }
1018
1019 cost = (u64) htab->buckets_num * sizeof(struct bpf_htab_bucket) +
1020 (u64) htab->elem_size * htab->map.max_entries;
1021 if (cost >= U32_MAX - PAGE_SIZE) {
1022 err = -EINVAL;
1023 goto free_htab;
1024 }
1025 err = bpf_map_charge_init(&htab->map.memory, cost);
1026 if (err)
1027 goto free_htab;
1028
1029 htab->buckets = bpf_map_area_alloc(htab->buckets_num *
1030 sizeof(struct bpf_htab_bucket),
1031 htab->map.numa_node);
1032 if (!htab->buckets) {
1033 bpf_map_charge_finish(&htab->map.memory);
1034 err = -ENOMEM;
1035 goto free_htab;
1036 }
1037
1038 for (i = 0; i < htab->buckets_num; i++) {
1039 INIT_HLIST_HEAD(&htab->buckets[i].head);
1040 raw_spin_lock_init(&htab->buckets[i].lock);
1041 }
1042
1043 return &htab->map;
1044 free_htab:
1045 kfree(htab);
1046 return ERR_PTR(err);
1047 }
1048
1049 static void sock_hash_free(struct bpf_map *map)
1050 {
1051 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1052 struct bpf_htab_bucket *bucket;
1053 struct hlist_head unlink_list;
1054 struct bpf_htab_elem *elem;
1055 struct hlist_node *node;
1056 int i;
1057
1058 /* After the sync no updates or deletes will be in-flight so it
1059 * is safe to walk map and remove entries without risking a race
1060 * in EEXIST update case.
1061 */
1062 synchronize_rcu();
1063 for (i = 0; i < htab->buckets_num; i++) {
1064 bucket = sock_hash_select_bucket(htab, i);
1065
1066 /* We are racing with sock_hash_delete_from_link to
1067 * enter the spin-lock critical section. Every socket on
1068 * the list is still linked to sockhash. Since link
1069 * exists, psock exists and holds a ref to socket. That
1070 * lets us to grab a socket ref too.
1071 */
1072 raw_spin_lock_bh(&bucket->lock);
1073 hlist_for_each_entry(elem, &bucket->head, node)
1074 sock_hold(elem->sk);
1075 hlist_move_list(&bucket->head, &unlink_list);
1076 raw_spin_unlock_bh(&bucket->lock);
1077
1078 /* Process removed entries out of atomic context to
1079 * block for socket lock before deleting the psock's
1080 * link to sockhash.
1081 */
1082 hlist_for_each_entry_safe(elem, node, &unlink_list, node) {
1083 hlist_del(&elem->node);
1084 lock_sock(elem->sk);
1085 rcu_read_lock();
1086 sock_map_unref(elem->sk, elem);
1087 rcu_read_unlock();
1088 release_sock(elem->sk);
1089 sock_put(elem->sk);
1090 sock_hash_free_elem(htab, elem);
1091 }
1092 }
1093
1094 /* wait for psock readers accessing its map link */
1095 synchronize_rcu();
1096
1097 bpf_map_area_free(htab->buckets);
1098 kfree(htab);
1099 }
1100
1101 static void *sock_hash_lookup_sys(struct bpf_map *map, void *key)
1102 {
1103 struct sock *sk;
1104
1105 if (map->value_size != sizeof(u64))
1106 return ERR_PTR(-ENOSPC);
1107
1108 sk = __sock_hash_lookup_elem(map, key);
1109 if (!sk)
1110 return ERR_PTR(-ENOENT);
1111
1112 sock_gen_cookie(sk);
1113 return &sk->sk_cookie;
1114 }
1115
1116 static void *sock_hash_lookup(struct bpf_map *map, void *key)
1117 {
1118 return __sock_hash_lookup_elem(map, key);
1119 }
1120
1121 static void sock_hash_release_progs(struct bpf_map *map)
1122 {
1123 psock_progs_drop(&container_of(map, struct bpf_htab, map)->progs);
1124 }
1125
1126 BPF_CALL_4(bpf_sock_hash_update, struct bpf_sock_ops_kern *, sops,
1127 struct bpf_map *, map, void *, key, u64, flags)
1128 {
1129 WARN_ON_ONCE(!rcu_read_lock_held());
1130
1131 if (likely(sock_map_sk_is_suitable(sops->sk) &&
1132 sock_map_op_okay(sops)))
1133 return sock_hash_update_common(map, key, sops->sk, flags);
1134 return -EOPNOTSUPP;
1135 }
1136
1137 const struct bpf_func_proto bpf_sock_hash_update_proto = {
1138 .func = bpf_sock_hash_update,
1139 .gpl_only = false,
1140 .pkt_access = true,
1141 .ret_type = RET_INTEGER,
1142 .arg1_type = ARG_PTR_TO_CTX,
1143 .arg2_type = ARG_CONST_MAP_PTR,
1144 .arg3_type = ARG_PTR_TO_MAP_KEY,
1145 .arg4_type = ARG_ANYTHING,
1146 };
1147
1148 BPF_CALL_4(bpf_sk_redirect_hash, struct sk_buff *, skb,
1149 struct bpf_map *, map, void *, key, u64, flags)
1150 {
1151 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1152 struct sock *sk;
1153
1154 if (unlikely(flags & ~(BPF_F_INGRESS)))
1155 return SK_DROP;
1156
1157 sk = __sock_hash_lookup_elem(map, key);
1158 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1159 return SK_DROP;
1160
1161 tcb->bpf.flags = flags;
1162 tcb->bpf.sk_redir = sk;
1163 return SK_PASS;
1164 }
1165
1166 const struct bpf_func_proto bpf_sk_redirect_hash_proto = {
1167 .func = bpf_sk_redirect_hash,
1168 .gpl_only = false,
1169 .ret_type = RET_INTEGER,
1170 .arg1_type = ARG_PTR_TO_CTX,
1171 .arg2_type = ARG_CONST_MAP_PTR,
1172 .arg3_type = ARG_PTR_TO_MAP_KEY,
1173 .arg4_type = ARG_ANYTHING,
1174 };
1175
1176 BPF_CALL_4(bpf_msg_redirect_hash, struct sk_msg *, msg,
1177 struct bpf_map *, map, void *, key, u64, flags)
1178 {
1179 struct sock *sk;
1180
1181 if (unlikely(flags & ~(BPF_F_INGRESS)))
1182 return SK_DROP;
1183
1184 sk = __sock_hash_lookup_elem(map, key);
1185 if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1186 return SK_DROP;
1187
1188 msg->flags = flags;
1189 msg->sk_redir = sk;
1190 return SK_PASS;
1191 }
1192
1193 const struct bpf_func_proto bpf_msg_redirect_hash_proto = {
1194 .func = bpf_msg_redirect_hash,
1195 .gpl_only = false,
1196 .ret_type = RET_INTEGER,
1197 .arg1_type = ARG_PTR_TO_CTX,
1198 .arg2_type = ARG_CONST_MAP_PTR,
1199 .arg3_type = ARG_PTR_TO_MAP_KEY,
1200 .arg4_type = ARG_ANYTHING,
1201 };
1202
1203 const struct bpf_map_ops sock_hash_ops = {
1204 .map_alloc = sock_hash_alloc,
1205 .map_free = sock_hash_free,
1206 .map_get_next_key = sock_hash_get_next_key,
1207 .map_update_elem = sock_hash_update_elem,
1208 .map_delete_elem = sock_hash_delete_elem,
1209 .map_lookup_elem = sock_hash_lookup,
1210 .map_lookup_elem_sys_only = sock_hash_lookup_sys,
1211 .map_release_uref = sock_hash_release_progs,
1212 .map_check_btf = map_check_no_btf,
1213 };
1214
1215 static struct sk_psock_progs *sock_map_progs(struct bpf_map *map)
1216 {
1217 switch (map->map_type) {
1218 case BPF_MAP_TYPE_SOCKMAP:
1219 return &container_of(map, struct bpf_stab, map)->progs;
1220 case BPF_MAP_TYPE_SOCKHASH:
1221 return &container_of(map, struct bpf_htab, map)->progs;
1222 default:
1223 break;
1224 }
1225
1226 return NULL;
1227 }
1228
1229 int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
1230 struct bpf_prog *old, u32 which)
1231 {
1232 struct sk_psock_progs *progs = sock_map_progs(map);
1233 struct bpf_prog **pprog;
1234
1235 if (!progs)
1236 return -EOPNOTSUPP;
1237
1238 switch (which) {
1239 case BPF_SK_MSG_VERDICT:
1240 pprog = &progs->msg_parser;
1241 break;
1242 case BPF_SK_SKB_STREAM_PARSER:
1243 pprog = &progs->skb_parser;
1244 break;
1245 case BPF_SK_SKB_STREAM_VERDICT:
1246 pprog = &progs->skb_verdict;
1247 break;
1248 default:
1249 return -EOPNOTSUPP;
1250 }
1251
1252 if (old)
1253 return psock_replace_prog(pprog, prog, old);
1254
1255 psock_set_prog(pprog, prog);
1256 return 0;
1257 }
1258
1259 static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
1260 {
1261 switch (link->map->map_type) {
1262 case BPF_MAP_TYPE_SOCKMAP:
1263 return sock_map_delete_from_link(link->map, sk,
1264 link->link_raw);
1265 case BPF_MAP_TYPE_SOCKHASH:
1266 return sock_hash_delete_from_link(link->map, sk,
1267 link->link_raw);
1268 default:
1269 break;
1270 }
1271 }
1272
1273 static void sock_map_remove_links(struct sock *sk, struct sk_psock *psock)
1274 {
1275 struct sk_psock_link *link;
1276
1277 while ((link = sk_psock_link_pop(psock))) {
1278 sock_map_unlink(sk, link);
1279 sk_psock_free_link(link);
1280 }
1281 }
1282
1283 void sock_map_unhash(struct sock *sk)
1284 {
1285 void (*saved_unhash)(struct sock *sk);
1286 struct sk_psock *psock;
1287
1288 rcu_read_lock();
1289 psock = sk_psock(sk);
1290 if (unlikely(!psock)) {
1291 rcu_read_unlock();
1292 if (sk->sk_prot->unhash)
1293 sk->sk_prot->unhash(sk);
1294 return;
1295 }
1296
1297 saved_unhash = psock->saved_unhash;
1298 sock_map_remove_links(sk, psock);
1299 rcu_read_unlock();
1300 saved_unhash(sk);
1301 }
1302
1303 void sock_map_close(struct sock *sk, long timeout)
1304 {
1305 void (*saved_close)(struct sock *sk, long timeout);
1306 struct sk_psock *psock;
1307
1308 lock_sock(sk);
1309 rcu_read_lock();
1310 psock = sk_psock(sk);
1311 if (unlikely(!psock)) {
1312 rcu_read_unlock();
1313 release_sock(sk);
1314 return sk->sk_prot->close(sk, timeout);
1315 }
1316
1317 saved_close = psock->saved_close;
1318 sock_map_remove_links(sk, psock);
1319 rcu_read_unlock();
1320 release_sock(sk);
1321 saved_close(sk, timeout);
1322 }
1323
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.