1 /* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
3 * This program is free software; you can redistribute it and/or
4 * modify it under the terms of version 2 of the GNU General Public
5 * License as published by the Free Software Foundation.
7 * This program is distributed in the hope that it will be useful, but
8 * WITHOUT ANY WARRANTY; without even the implied warranty of
9 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
10 * General Public License for more details.
13 /* A BPF sock_map is used to store sock objects. This is primarly used
14 * for doing socket redirect with BPF helper routines.
16 * A sock map may have BPF programs attached to it, currently a program
17 * used to parse packets and a program to provide a verdict and redirect
18 * decision on the packet are supported. Any programs attached to a sock
19 * map are inherited by sock objects when they are added to the map. If
20 * no BPF programs are attached the sock object may only be used for sock
23 * A sock object may be in multiple maps, but can only inherit a single
24 * parse or verdict program. If adding a sock object to a map would result
25 * in having multiple parsing programs the update will return an EBUSY error.
27 * For reference this program is similar to devmap used in XDP context
28 * reviewing these together may be useful. For an example please review
29 * ./samples/bpf/sockmap/.
31 #include <linux/bpf.h>
33 #include <linux/filter.h>
34 #include <linux/errno.h>
35 #include <linux/file.h>
36 #include <linux/kernel.h>
37 #include <linux/net.h>
38 #include <linux/skbuff.h>
39 #include <linux/workqueue.h>
40 #include <linux/list.h>
41 #include <net/strparser.h>
44 #define SOCK_CREATE_FLAG_MASK \
45 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
49 struct sock **sock_map;
50 struct bpf_prog *bpf_parse;
51 struct bpf_prog *bpf_verdict;
54 enum smap_psock_state {
58 struct smap_psock_map_entry {
59 struct list_head list;
65 /* refcnt is used inside sk_callback_lock */
68 /* datapath variables */
69 struct sk_buff_head rxqueue;
72 /* datapath error path cache across tx work invocations */
75 struct sk_buff *save_skb;
77 struct strparser strp;
78 struct bpf_prog *bpf_parse;
79 struct bpf_prog *bpf_verdict;
80 struct list_head maps;
82 /* Back reference used when sock callback trigger sockmap operations */
86 struct work_struct tx_work;
87 struct work_struct gc_work;
89 struct proto *sk_proto;
90 void (*save_close)(struct sock *sk, long timeout);
91 void (*save_data_ready)(struct sock *sk);
92 void (*save_write_space)(struct sock *sk);
95 static inline struct smap_psock *smap_psock_sk(const struct sock *sk)
97 return rcu_dereference_sk_user_data(sk);
100 static struct proto tcp_bpf_proto;
101 static int bpf_tcp_init(struct sock *sk)
103 struct smap_psock *psock;
106 psock = smap_psock_sk(sk);
107 if (unlikely(!psock)) {
112 if (unlikely(psock->sk_proto)) {
117 psock->save_close = sk->sk_prot->close;
118 psock->sk_proto = sk->sk_prot;
119 sk->sk_prot = &tcp_bpf_proto;
124 static void bpf_tcp_release(struct sock *sk)
126 struct smap_psock *psock;
129 psock = smap_psock_sk(sk);
132 sk->sk_prot = psock->sk_proto;
133 psock->sk_proto = NULL;
138 static void smap_release_sock(struct smap_psock *psock, struct sock *sock);
140 static void bpf_tcp_close(struct sock *sk, long timeout)
142 void (*close_fun)(struct sock *sk, long timeout);
143 struct smap_psock_map_entry *e, *tmp;
144 struct smap_psock *psock;
148 psock = smap_psock_sk(sk);
149 if (unlikely(!psock)) {
151 return sk->sk_prot->close(sk, timeout);
154 /* The psock may be destroyed anytime after exiting the RCU critial
155 * section so by the time we use close_fun the psock may no longer
156 * be valid. However, bpf_tcp_close is called with the sock lock
157 * held so the close hook and sk are still valid.
159 close_fun = psock->save_close;
161 write_lock_bh(&sk->sk_callback_lock);
162 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
163 osk = cmpxchg(e->entry, sk, NULL);
166 smap_release_sock(psock, sk);
169 write_unlock_bh(&sk->sk_callback_lock);
171 close_fun(sk, timeout);
180 static struct tcp_ulp_ops bpf_tcp_ulp_ops __read_mostly = {
183 .user_visible = false,
185 .init = bpf_tcp_init,
186 .release = bpf_tcp_release,
189 static int bpf_tcp_ulp_register(void)
191 tcp_bpf_proto = tcp_prot;
192 tcp_bpf_proto.close = bpf_tcp_close;
193 return tcp_register_ulp(&bpf_tcp_ulp_ops);
196 static int smap_verdict_func(struct smap_psock *psock, struct sk_buff *skb)
198 struct bpf_prog *prog = READ_ONCE(psock->bpf_verdict);
205 /* We need to ensure that BPF metadata for maps is also cleared
206 * when we orphan the skb so that we don't have the possibility
207 * to reference a stale map.
209 TCP_SKB_CB(skb)->bpf.map = NULL;
210 skb->sk = psock->sock;
211 bpf_compute_data_pointers(skb);
213 rc = (*prog->bpf_func)(skb, prog->insnsi);
217 /* Moving return codes from UAPI namespace into internal namespace */
218 return rc == SK_PASS ?
219 (TCP_SKB_CB(skb)->bpf.map ? __SK_REDIRECT : __SK_PASS) :
223 static void smap_do_verdict(struct smap_psock *psock, struct sk_buff *skb)
228 rc = smap_verdict_func(psock, skb);
231 sk = do_sk_redirect_map(skb);
233 struct smap_psock *peer = smap_psock_sk(sk);
236 test_bit(SMAP_TX_RUNNING, &peer->state) &&
237 !sock_flag(sk, SOCK_DEAD) &&
238 sock_writeable(sk))) {
239 skb_set_owner_w(skb, sk);
240 skb_queue_tail(&peer->rxqueue, skb);
241 schedule_work(&peer->tx_work);
245 /* Fall through and free skb otherwise */
252 static void smap_report_sk_error(struct smap_psock *psock, int err)
254 struct sock *sk = psock->sock;
257 sk->sk_error_report(sk);
260 static void smap_read_sock_strparser(struct strparser *strp,
263 struct smap_psock *psock;
266 psock = container_of(strp, struct smap_psock, strp);
267 smap_do_verdict(psock, skb);
271 /* Called with lock held on socket */
272 static void smap_data_ready(struct sock *sk)
274 struct smap_psock *psock;
277 psock = smap_psock_sk(sk);
279 write_lock_bh(&sk->sk_callback_lock);
280 strp_data_ready(&psock->strp);
281 write_unlock_bh(&sk->sk_callback_lock);
286 static void smap_tx_work(struct work_struct *w)
288 struct smap_psock *psock;
292 psock = container_of(w, struct smap_psock, tx_work);
294 /* lock sock to avoid losing sk_socket at some point during loop */
295 lock_sock(psock->sock);
296 if (psock->save_skb) {
297 skb = psock->save_skb;
298 rem = psock->save_rem;
299 off = psock->save_off;
300 psock->save_skb = NULL;
304 while ((skb = skb_dequeue(&psock->rxqueue))) {
309 if (likely(psock->sock->sk_socket))
310 n = skb_send_sock_locked(psock->sock,
316 /* Retry when space is available */
317 psock->save_skb = skb;
318 psock->save_rem = rem;
319 psock->save_off = off;
322 /* Hard errors break pipe and stop xmit */
323 smap_report_sk_error(psock, n ? -n : EPIPE);
324 clear_bit(SMAP_TX_RUNNING, &psock->state);
334 release_sock(psock->sock);
337 static void smap_write_space(struct sock *sk)
339 struct smap_psock *psock;
342 psock = smap_psock_sk(sk);
343 if (likely(psock && test_bit(SMAP_TX_RUNNING, &psock->state)))
344 schedule_work(&psock->tx_work);
348 static void smap_stop_sock(struct smap_psock *psock, struct sock *sk)
350 if (!psock->strp_enabled)
352 sk->sk_data_ready = psock->save_data_ready;
353 sk->sk_write_space = psock->save_write_space;
354 psock->save_data_ready = NULL;
355 psock->save_write_space = NULL;
356 strp_stop(&psock->strp);
357 psock->strp_enabled = false;
360 static void smap_destroy_psock(struct rcu_head *rcu)
362 struct smap_psock *psock = container_of(rcu,
363 struct smap_psock, rcu);
365 /* Now that a grace period has passed there is no longer
366 * any reference to this sock in the sockmap so we can
367 * destroy the psock, strparser, and bpf programs. But,
368 * because we use workqueue sync operations we can not
369 * do it in rcu context
371 schedule_work(&psock->gc_work);
374 static void smap_release_sock(struct smap_psock *psock, struct sock *sock)
380 tcp_cleanup_ulp(sock);
381 smap_stop_sock(psock, sock);
382 clear_bit(SMAP_TX_RUNNING, &psock->state);
383 rcu_assign_sk_user_data(sock, NULL);
384 call_rcu_sched(&psock->rcu, smap_destroy_psock);
387 static int smap_parse_func_strparser(struct strparser *strp,
390 struct smap_psock *psock;
391 struct bpf_prog *prog;
395 psock = container_of(strp, struct smap_psock, strp);
396 prog = READ_ONCE(psock->bpf_parse);
398 if (unlikely(!prog)) {
403 /* Attach socket for bpf program to use if needed we can do this
404 * because strparser clones the skb before handing it to a upper
405 * layer, meaning skb_orphan has been called. We NULL sk on the
406 * way out to ensure we don't trigger a BUG_ON in skb/sk operations
407 * later and because we are not charging the memory of this skb to
410 skb->sk = psock->sock;
411 bpf_compute_data_pointers(skb);
412 rc = (*prog->bpf_func)(skb, prog->insnsi);
419 static int smap_read_sock_done(struct strparser *strp, int err)
424 static int smap_init_sock(struct smap_psock *psock,
427 static const struct strp_callbacks cb = {
428 .rcv_msg = smap_read_sock_strparser,
429 .parse_msg = smap_parse_func_strparser,
430 .read_sock_done = smap_read_sock_done,
433 return strp_init(&psock->strp, sk, &cb);
436 static void smap_init_progs(struct smap_psock *psock,
437 struct bpf_stab *stab,
438 struct bpf_prog *verdict,
439 struct bpf_prog *parse)
441 struct bpf_prog *orig_parse, *orig_verdict;
443 orig_parse = xchg(&psock->bpf_parse, parse);
444 orig_verdict = xchg(&psock->bpf_verdict, verdict);
447 bpf_prog_put(orig_verdict);
449 bpf_prog_put(orig_parse);
452 static void smap_start_sock(struct smap_psock *psock, struct sock *sk)
454 if (sk->sk_data_ready == smap_data_ready)
456 psock->save_data_ready = sk->sk_data_ready;
457 psock->save_write_space = sk->sk_write_space;
458 sk->sk_data_ready = smap_data_ready;
459 sk->sk_write_space = smap_write_space;
460 psock->strp_enabled = true;
463 static void sock_map_remove_complete(struct bpf_stab *stab)
465 bpf_map_area_free(stab->sock_map);
469 static void smap_gc_work(struct work_struct *w)
471 struct smap_psock_map_entry *e, *tmp;
472 struct smap_psock *psock;
474 psock = container_of(w, struct smap_psock, gc_work);
476 /* no callback lock needed because we already detached sockmap ops */
477 if (psock->strp_enabled)
478 strp_done(&psock->strp);
480 cancel_work_sync(&psock->tx_work);
481 __skb_queue_purge(&psock->rxqueue);
483 /* At this point all strparser and xmit work must be complete */
484 if (psock->bpf_parse)
485 bpf_prog_put(psock->bpf_parse);
486 if (psock->bpf_verdict)
487 bpf_prog_put(psock->bpf_verdict);
489 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
494 sock_put(psock->sock);
498 static struct smap_psock *smap_init_psock(struct sock *sock,
499 struct bpf_stab *stab)
501 struct smap_psock *psock;
503 psock = kzalloc_node(sizeof(struct smap_psock),
504 GFP_ATOMIC | __GFP_NOWARN,
505 stab->map.numa_node);
507 return ERR_PTR(-ENOMEM);
510 skb_queue_head_init(&psock->rxqueue);
511 INIT_WORK(&psock->tx_work, smap_tx_work);
512 INIT_WORK(&psock->gc_work, smap_gc_work);
513 INIT_LIST_HEAD(&psock->maps);
516 rcu_assign_sk_user_data(sock, psock);
521 static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
523 struct bpf_stab *stab;
527 if (!capable(CAP_NET_ADMIN))
528 return ERR_PTR(-EPERM);
530 /* check sanity of attributes */
531 if (attr->max_entries == 0 || attr->key_size != 4 ||
532 attr->value_size != 4 || attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
533 return ERR_PTR(-EINVAL);
535 if (attr->value_size > KMALLOC_MAX_SIZE)
536 return ERR_PTR(-E2BIG);
538 err = bpf_tcp_ulp_register();
539 if (err && err != -EEXIST)
542 stab = kzalloc(sizeof(*stab), GFP_USER);
544 return ERR_PTR(-ENOMEM);
546 bpf_map_init_from_attr(&stab->map, attr);
548 /* make sure page count doesn't overflow */
549 cost = (u64) stab->map.max_entries * sizeof(struct sock *);
550 if (cost >= U32_MAX - PAGE_SIZE)
553 stab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
555 /* if map size is larger than memlock limit, reject it early */
556 err = bpf_map_precharge_memlock(stab->map.pages);
561 stab->sock_map = bpf_map_area_alloc(stab->map.max_entries *
562 sizeof(struct sock *),
563 stab->map.numa_node);
573 static void smap_list_remove(struct smap_psock *psock, struct sock **entry)
575 struct smap_psock_map_entry *e, *tmp;
577 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
578 if (e->entry == entry) {
585 static void sock_map_free(struct bpf_map *map)
587 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
592 /* At this point no update, lookup or delete operations can happen.
593 * However, be aware we can still get a socket state event updates,
594 * and data ready callabacks that reference the psock from sk_user_data
595 * Also psock worker threads are still in-flight. So smap_release_sock
596 * will only free the psock after cancel_sync on the worker threads
597 * and a grace period expire to ensure psock is really safe to remove.
600 for (i = 0; i < stab->map.max_entries; i++) {
601 struct smap_psock *psock;
604 sock = xchg(&stab->sock_map[i], NULL);
608 write_lock_bh(&sock->sk_callback_lock);
609 psock = smap_psock_sk(sock);
610 /* This check handles a racing sock event that can get the
611 * sk_callback_lock before this case but after xchg happens
612 * causing the refcnt to hit zero and sock user data (psock)
613 * to be null and queued for garbage collection.
616 smap_list_remove(psock, &stab->sock_map[i]);
617 smap_release_sock(psock, sock);
619 write_unlock_bh(&sock->sk_callback_lock);
623 sock_map_remove_complete(stab);
626 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
628 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
629 u32 i = key ? *(u32 *)key : U32_MAX;
630 u32 *next = (u32 *)next_key;
632 if (i >= stab->map.max_entries) {
637 if (i == stab->map.max_entries - 1)
644 struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
646 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
648 if (key >= map->max_entries)
651 return READ_ONCE(stab->sock_map[key]);
654 static int sock_map_delete_elem(struct bpf_map *map, void *key)
656 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
657 struct smap_psock *psock;
661 if (k >= map->max_entries)
664 sock = xchg(&stab->sock_map[k], NULL);
668 write_lock_bh(&sock->sk_callback_lock);
669 psock = smap_psock_sk(sock);
673 if (psock->bpf_parse)
674 smap_stop_sock(psock, sock);
675 smap_list_remove(psock, &stab->sock_map[k]);
676 smap_release_sock(psock, sock);
678 write_unlock_bh(&sock->sk_callback_lock);
682 /* Locking notes: Concurrent updates, deletes, and lookups are allowed and are
683 * done inside rcu critical sections. This ensures on updates that the psock
684 * will not be released via smap_release_sock() until concurrent updates/deletes
685 * complete. All operations operate on sock_map using cmpxchg and xchg
686 * operations to ensure we do not get stale references. Any reads into the
687 * map must be done with READ_ONCE() because of this.
689 * A psock is destroyed via call_rcu and after any worker threads are cancelled
690 * and syncd so we are certain all references from the update/lookup/delete
691 * operations as well as references in the data path are no longer in use.
693 * Psocks may exist in multiple maps, but only a single set of parse/verdict
694 * programs may be inherited from the maps it belongs to. A reference count
695 * is kept with the total number of references to the psock from all maps. The
696 * psock will not be released until this reaches zero. The psock and sock
697 * user data data use the sk_callback_lock to protect critical data structures
698 * from concurrent access. This allows us to avoid two updates from modifying
699 * the user data in sock and the lock is required anyways for modifying
700 * callbacks, we simply increase its scope slightly.
703 * - psock must always be read inside RCU critical section
704 * - sk_user_data must only be modified inside sk_callback_lock and read
705 * inside RCU critical section.
706 * - psock->maps list must only be read & modified inside sk_callback_lock
707 * - sock_map must use READ_ONCE and (cmp)xchg operations
708 * - BPF verdict/parse programs must use READ_ONCE and xchg operations
710 static int sock_map_ctx_update_elem(struct bpf_sock_ops_kern *skops,
712 void *key, u64 flags)
714 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
715 struct smap_psock_map_entry *e = NULL;
716 struct bpf_prog *verdict, *parse;
717 struct sock *osock, *sock;
718 struct smap_psock *psock;
722 if (unlikely(flags > BPF_EXIST))
725 if (unlikely(i >= stab->map.max_entries))
728 sock = READ_ONCE(stab->sock_map[i]);
729 if (flags == BPF_EXIST && !sock)
731 else if (flags == BPF_NOEXIST && sock)
736 /* 1. If sock map has BPF programs those will be inherited by the
737 * sock being added. If the sock is already attached to BPF programs
738 * this results in an error.
740 verdict = READ_ONCE(stab->bpf_verdict);
741 parse = READ_ONCE(stab->bpf_parse);
743 if (parse && verdict) {
744 /* bpf prog refcnt may be zero if a concurrent attach operation
745 * removes the program after the above READ_ONCE() but before
746 * we increment the refcnt. If this is the case abort with an
749 verdict = bpf_prog_inc_not_zero(stab->bpf_verdict);
751 return PTR_ERR(verdict);
753 parse = bpf_prog_inc_not_zero(stab->bpf_parse);
755 bpf_prog_put(verdict);
756 return PTR_ERR(parse);
760 write_lock_bh(&sock->sk_callback_lock);
761 psock = smap_psock_sk(sock);
763 /* 2. Do not allow inheriting programs if psock exists and has
764 * already inherited programs. This would create confusion on
765 * which parser/verdict program is running. If no psock exists
766 * create one. Inside sk_callback_lock to ensure concurrent create
767 * doesn't update user data.
770 if (READ_ONCE(psock->bpf_parse) && parse) {
776 psock = smap_init_psock(sock, stab);
778 err = PTR_ERR(psock);
782 err = tcp_set_ulp_id(sock, TCP_ULP_BPF);
786 set_bit(SMAP_TX_RUNNING, &psock->state);
789 e = kzalloc(sizeof(*e), GFP_ATOMIC | __GFP_NOWARN);
794 e->entry = &stab->sock_map[i];
796 /* 3. At this point we have a reference to a valid psock that is
797 * running. Attach any BPF programs needed.
799 if (parse && verdict && !psock->strp_enabled) {
800 err = smap_init_sock(psock, sock);
803 smap_init_progs(psock, stab, verdict, parse);
804 smap_start_sock(psock, sock);
807 /* 4. Place psock in sockmap for use and stop any programs on
808 * the old sock assuming its not the same sock we are replacing
809 * it with. Because we can only have a single set of programs if
810 * old_sock has a strp we can stop it.
812 list_add_tail(&e->list, &psock->maps);
813 write_unlock_bh(&sock->sk_callback_lock);
815 osock = xchg(&stab->sock_map[i], sock);
817 struct smap_psock *opsock = smap_psock_sk(osock);
819 write_lock_bh(&osock->sk_callback_lock);
820 if (osock != sock && parse)
821 smap_stop_sock(opsock, osock);
822 smap_list_remove(opsock, &stab->sock_map[i]);
823 smap_release_sock(opsock, osock);
824 write_unlock_bh(&osock->sk_callback_lock);
828 smap_release_sock(psock, sock);
831 bpf_prog_put(verdict);
834 write_unlock_bh(&sock->sk_callback_lock);
839 int sock_map_prog(struct bpf_map *map, struct bpf_prog *prog, u32 type)
841 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
842 struct bpf_prog *orig;
844 if (unlikely(map->map_type != BPF_MAP_TYPE_SOCKMAP))
848 case BPF_SK_SKB_STREAM_PARSER:
849 orig = xchg(&stab->bpf_parse, prog);
851 case BPF_SK_SKB_STREAM_VERDICT:
852 orig = xchg(&stab->bpf_verdict, prog);
864 static void *sock_map_lookup(struct bpf_map *map, void *key)
869 static int sock_map_update_elem(struct bpf_map *map,
870 void *key, void *value, u64 flags)
872 struct bpf_sock_ops_kern skops;
873 u32 fd = *(u32 *)value;
874 struct socket *socket;
877 socket = sockfd_lookup(fd, &err);
881 skops.sk = socket->sk;
887 if (skops.sk->sk_type != SOCK_STREAM ||
888 skops.sk->sk_protocol != IPPROTO_TCP) {
893 err = sock_map_ctx_update_elem(&skops, map, key, flags);
898 static void sock_map_release(struct bpf_map *map, struct file *map_file)
900 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
901 struct bpf_prog *orig;
903 orig = xchg(&stab->bpf_parse, NULL);
906 orig = xchg(&stab->bpf_verdict, NULL);
911 const struct bpf_map_ops sock_map_ops = {
912 .map_alloc = sock_map_alloc,
913 .map_free = sock_map_free,
914 .map_lookup_elem = sock_map_lookup,
915 .map_get_next_key = sock_map_get_next_key,
916 .map_update_elem = sock_map_update_elem,
917 .map_delete_elem = sock_map_delete_elem,
918 .map_release = sock_map_release,
921 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, bpf_sock,
922 struct bpf_map *, map, void *, key, u64, flags)
924 WARN_ON_ONCE(!rcu_read_lock_held());
925 return sock_map_ctx_update_elem(bpf_sock, map, key, flags);
928 const struct bpf_func_proto bpf_sock_map_update_proto = {
929 .func = bpf_sock_map_update,
932 .ret_type = RET_INTEGER,
933 .arg1_type = ARG_PTR_TO_CTX,
934 .arg2_type = ARG_CONST_MAP_PTR,
935 .arg3_type = ARG_PTR_TO_MAP_KEY,
936 .arg4_type = ARG_ANYTHING,