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 void (*save_data_ready)(struct sock *sk);
90 void (*save_write_space)(struct sock *sk);
91 void (*save_state_change)(struct sock *sk);
94 static inline struct smap_psock *smap_psock_sk(const struct sock *sk)
96 return rcu_dereference_sk_user_data(sk);
105 static int smap_verdict_func(struct smap_psock *psock, struct sk_buff *skb)
107 struct bpf_prog *prog = READ_ONCE(psock->bpf_verdict);
114 /* We need to ensure that BPF metadata for maps is also cleared
115 * when we orphan the skb so that we don't have the possibility
116 * to reference a stale map.
118 TCP_SKB_CB(skb)->bpf.map = NULL;
119 skb->sk = psock->sock;
120 bpf_compute_data_pointers(skb);
122 rc = (*prog->bpf_func)(skb, prog->insnsi);
126 /* Moving return codes from UAPI namespace into internal namespace */
127 return rc == SK_PASS ?
128 (TCP_SKB_CB(skb)->bpf.map ? __SK_REDIRECT : __SK_PASS) :
132 static void smap_do_verdict(struct smap_psock *psock, struct sk_buff *skb)
137 rc = smap_verdict_func(psock, skb);
140 sk = do_sk_redirect_map(skb);
142 struct smap_psock *peer = smap_psock_sk(sk);
145 test_bit(SMAP_TX_RUNNING, &peer->state) &&
146 !sock_flag(sk, SOCK_DEAD) &&
147 sock_writeable(sk))) {
148 skb_set_owner_w(skb, sk);
149 skb_queue_tail(&peer->rxqueue, skb);
150 schedule_work(&peer->tx_work);
154 /* Fall through and free skb otherwise */
161 static void smap_report_sk_error(struct smap_psock *psock, int err)
163 struct sock *sk = psock->sock;
166 sk->sk_error_report(sk);
169 static void smap_release_sock(struct smap_psock *psock, struct sock *sock);
171 /* Called with lock_sock(sk) held */
172 static void smap_state_change(struct sock *sk)
174 struct smap_psock_map_entry *e, *tmp;
175 struct smap_psock *psock;
176 struct socket_wq *wq;
181 /* Allowing transitions into an established syn_recv states allows
182 * for early binding sockets to a smap object before the connection
185 switch (sk->sk_state) {
188 case TCP_ESTABLISHED:
198 /* Only release if the map entry is in fact the sock in
199 * question. There is a case where the operator deletes
200 * the sock from the map, but the TCP sock is closed before
201 * the psock is detached. Use cmpxchg to verify correct
204 psock = smap_psock_sk(sk);
205 if (unlikely(!psock))
207 write_lock_bh(&sk->sk_callback_lock);
208 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
209 osk = cmpxchg(e->entry, sk, NULL);
212 smap_release_sock(psock, sk);
215 write_unlock_bh(&sk->sk_callback_lock);
218 psock = smap_psock_sk(sk);
219 if (unlikely(!psock))
221 smap_report_sk_error(psock, EPIPE);
225 wq = rcu_dereference(sk->sk_wq);
226 if (skwq_has_sleeper(wq))
227 wake_up_interruptible_all(&wq->wait);
231 static void smap_read_sock_strparser(struct strparser *strp,
234 struct smap_psock *psock;
237 psock = container_of(strp, struct smap_psock, strp);
238 smap_do_verdict(psock, skb);
242 /* Called with lock held on socket */
243 static void smap_data_ready(struct sock *sk)
245 struct smap_psock *psock;
248 psock = smap_psock_sk(sk);
250 write_lock_bh(&sk->sk_callback_lock);
251 strp_data_ready(&psock->strp);
252 write_unlock_bh(&sk->sk_callback_lock);
257 static void smap_tx_work(struct work_struct *w)
259 struct smap_psock *psock;
263 psock = container_of(w, struct smap_psock, tx_work);
265 /* lock sock to avoid losing sk_socket at some point during loop */
266 lock_sock(psock->sock);
267 if (psock->save_skb) {
268 skb = psock->save_skb;
269 rem = psock->save_rem;
270 off = psock->save_off;
271 psock->save_skb = NULL;
275 while ((skb = skb_dequeue(&psock->rxqueue))) {
280 if (likely(psock->sock->sk_socket))
281 n = skb_send_sock_locked(psock->sock,
287 /* Retry when space is available */
288 psock->save_skb = skb;
289 psock->save_rem = rem;
290 psock->save_off = off;
293 /* Hard errors break pipe and stop xmit */
294 smap_report_sk_error(psock, n ? -n : EPIPE);
295 clear_bit(SMAP_TX_RUNNING, &psock->state);
305 release_sock(psock->sock);
308 static void smap_write_space(struct sock *sk)
310 struct smap_psock *psock;
313 psock = smap_psock_sk(sk);
314 if (likely(psock && test_bit(SMAP_TX_RUNNING, &psock->state)))
315 schedule_work(&psock->tx_work);
319 static void smap_stop_sock(struct smap_psock *psock, struct sock *sk)
321 if (!psock->strp_enabled)
323 sk->sk_data_ready = psock->save_data_ready;
324 sk->sk_write_space = psock->save_write_space;
325 sk->sk_state_change = psock->save_state_change;
326 psock->save_data_ready = NULL;
327 psock->save_write_space = NULL;
328 psock->save_state_change = NULL;
329 strp_stop(&psock->strp);
330 psock->strp_enabled = false;
333 static void smap_destroy_psock(struct rcu_head *rcu)
335 struct smap_psock *psock = container_of(rcu,
336 struct smap_psock, rcu);
338 /* Now that a grace period has passed there is no longer
339 * any reference to this sock in the sockmap so we can
340 * destroy the psock, strparser, and bpf programs. But,
341 * because we use workqueue sync operations we can not
342 * do it in rcu context
344 schedule_work(&psock->gc_work);
347 static void smap_release_sock(struct smap_psock *psock, struct sock *sock)
353 smap_stop_sock(psock, sock);
354 clear_bit(SMAP_TX_RUNNING, &psock->state);
355 rcu_assign_sk_user_data(sock, NULL);
356 call_rcu_sched(&psock->rcu, smap_destroy_psock);
359 static int smap_parse_func_strparser(struct strparser *strp,
362 struct smap_psock *psock;
363 struct bpf_prog *prog;
367 psock = container_of(strp, struct smap_psock, strp);
368 prog = READ_ONCE(psock->bpf_parse);
370 if (unlikely(!prog)) {
375 /* Attach socket for bpf program to use if needed we can do this
376 * because strparser clones the skb before handing it to a upper
377 * layer, meaning skb_orphan has been called. We NULL sk on the
378 * way out to ensure we don't trigger a BUG_ON in skb/sk operations
379 * later and because we are not charging the memory of this skb to
382 skb->sk = psock->sock;
383 bpf_compute_data_pointers(skb);
384 rc = (*prog->bpf_func)(skb, prog->insnsi);
391 static int smap_read_sock_done(struct strparser *strp, int err)
396 static int smap_init_sock(struct smap_psock *psock,
399 static const struct strp_callbacks cb = {
400 .rcv_msg = smap_read_sock_strparser,
401 .parse_msg = smap_parse_func_strparser,
402 .read_sock_done = smap_read_sock_done,
405 return strp_init(&psock->strp, sk, &cb);
408 static void smap_init_progs(struct smap_psock *psock,
409 struct bpf_stab *stab,
410 struct bpf_prog *verdict,
411 struct bpf_prog *parse)
413 struct bpf_prog *orig_parse, *orig_verdict;
415 orig_parse = xchg(&psock->bpf_parse, parse);
416 orig_verdict = xchg(&psock->bpf_verdict, verdict);
419 bpf_prog_put(orig_verdict);
421 bpf_prog_put(orig_parse);
424 static void smap_start_sock(struct smap_psock *psock, struct sock *sk)
426 if (sk->sk_data_ready == smap_data_ready)
428 psock->save_data_ready = sk->sk_data_ready;
429 psock->save_write_space = sk->sk_write_space;
430 psock->save_state_change = sk->sk_state_change;
431 sk->sk_data_ready = smap_data_ready;
432 sk->sk_write_space = smap_write_space;
433 sk->sk_state_change = smap_state_change;
434 psock->strp_enabled = true;
437 static void sock_map_remove_complete(struct bpf_stab *stab)
439 bpf_map_area_free(stab->sock_map);
443 static void smap_gc_work(struct work_struct *w)
445 struct smap_psock_map_entry *e, *tmp;
446 struct smap_psock *psock;
448 psock = container_of(w, struct smap_psock, gc_work);
450 /* no callback lock needed because we already detached sockmap ops */
451 if (psock->strp_enabled)
452 strp_done(&psock->strp);
454 cancel_work_sync(&psock->tx_work);
455 __skb_queue_purge(&psock->rxqueue);
457 /* At this point all strparser and xmit work must be complete */
458 if (psock->bpf_parse)
459 bpf_prog_put(psock->bpf_parse);
460 if (psock->bpf_verdict)
461 bpf_prog_put(psock->bpf_verdict);
463 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
468 sock_put(psock->sock);
472 static struct smap_psock *smap_init_psock(struct sock *sock,
473 struct bpf_stab *stab)
475 struct smap_psock *psock;
477 psock = kzalloc_node(sizeof(struct smap_psock),
478 GFP_ATOMIC | __GFP_NOWARN,
479 stab->map.numa_node);
481 return ERR_PTR(-ENOMEM);
484 skb_queue_head_init(&psock->rxqueue);
485 INIT_WORK(&psock->tx_work, smap_tx_work);
486 INIT_WORK(&psock->gc_work, smap_gc_work);
487 INIT_LIST_HEAD(&psock->maps);
490 rcu_assign_sk_user_data(sock, psock);
495 static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
497 struct bpf_stab *stab;
501 if (!capable(CAP_NET_ADMIN))
502 return ERR_PTR(-EPERM);
504 /* check sanity of attributes */
505 if (attr->max_entries == 0 || attr->key_size != 4 ||
506 attr->value_size != 4 || attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
507 return ERR_PTR(-EINVAL);
509 if (attr->value_size > KMALLOC_MAX_SIZE)
510 return ERR_PTR(-E2BIG);
512 stab = kzalloc(sizeof(*stab), GFP_USER);
514 return ERR_PTR(-ENOMEM);
516 bpf_map_init_from_attr(&stab->map, attr);
518 /* make sure page count doesn't overflow */
519 cost = (u64) stab->map.max_entries * sizeof(struct sock *);
520 if (cost >= U32_MAX - PAGE_SIZE)
523 stab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
525 /* if map size is larger than memlock limit, reject it early */
526 err = bpf_map_precharge_memlock(stab->map.pages);
531 stab->sock_map = bpf_map_area_alloc(stab->map.max_entries *
532 sizeof(struct sock *),
533 stab->map.numa_node);
543 static void smap_list_remove(struct smap_psock *psock, struct sock **entry)
545 struct smap_psock_map_entry *e, *tmp;
547 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
548 if (e->entry == entry) {
555 static void sock_map_free(struct bpf_map *map)
557 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
562 /* At this point no update, lookup or delete operations can happen.
563 * However, be aware we can still get a socket state event updates,
564 * and data ready callabacks that reference the psock from sk_user_data
565 * Also psock worker threads are still in-flight. So smap_release_sock
566 * will only free the psock after cancel_sync on the worker threads
567 * and a grace period expire to ensure psock is really safe to remove.
570 for (i = 0; i < stab->map.max_entries; i++) {
571 struct smap_psock *psock;
574 sock = xchg(&stab->sock_map[i], NULL);
578 write_lock_bh(&sock->sk_callback_lock);
579 psock = smap_psock_sk(sock);
580 /* This check handles a racing sock event that can get the
581 * sk_callback_lock before this case but after xchg happens
582 * causing the refcnt to hit zero and sock user data (psock)
583 * to be null and queued for garbage collection.
586 smap_list_remove(psock, &stab->sock_map[i]);
587 smap_release_sock(psock, sock);
589 write_unlock_bh(&sock->sk_callback_lock);
593 if (stab->bpf_verdict)
594 bpf_prog_put(stab->bpf_verdict);
596 bpf_prog_put(stab->bpf_parse);
598 sock_map_remove_complete(stab);
601 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
603 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
604 u32 i = key ? *(u32 *)key : U32_MAX;
605 u32 *next = (u32 *)next_key;
607 if (i >= stab->map.max_entries) {
612 if (i == stab->map.max_entries - 1)
619 struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
621 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
623 if (key >= map->max_entries)
626 return READ_ONCE(stab->sock_map[key]);
629 static int sock_map_delete_elem(struct bpf_map *map, void *key)
631 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
632 struct smap_psock *psock;
636 if (k >= map->max_entries)
639 sock = xchg(&stab->sock_map[k], NULL);
643 write_lock_bh(&sock->sk_callback_lock);
644 psock = smap_psock_sk(sock);
648 if (psock->bpf_parse)
649 smap_stop_sock(psock, sock);
650 smap_list_remove(psock, &stab->sock_map[k]);
651 smap_release_sock(psock, sock);
653 write_unlock_bh(&sock->sk_callback_lock);
657 /* Locking notes: Concurrent updates, deletes, and lookups are allowed and are
658 * done inside rcu critical sections. This ensures on updates that the psock
659 * will not be released via smap_release_sock() until concurrent updates/deletes
660 * complete. All operations operate on sock_map using cmpxchg and xchg
661 * operations to ensure we do not get stale references. Any reads into the
662 * map must be done with READ_ONCE() because of this.
664 * A psock is destroyed via call_rcu and after any worker threads are cancelled
665 * and syncd so we are certain all references from the update/lookup/delete
666 * operations as well as references in the data path are no longer in use.
668 * Psocks may exist in multiple maps, but only a single set of parse/verdict
669 * programs may be inherited from the maps it belongs to. A reference count
670 * is kept with the total number of references to the psock from all maps. The
671 * psock will not be released until this reaches zero. The psock and sock
672 * user data data use the sk_callback_lock to protect critical data structures
673 * from concurrent access. This allows us to avoid two updates from modifying
674 * the user data in sock and the lock is required anyways for modifying
675 * callbacks, we simply increase its scope slightly.
678 * - psock must always be read inside RCU critical section
679 * - sk_user_data must only be modified inside sk_callback_lock and read
680 * inside RCU critical section.
681 * - psock->maps list must only be read & modified inside sk_callback_lock
682 * - sock_map must use READ_ONCE and (cmp)xchg operations
683 * - BPF verdict/parse programs must use READ_ONCE and xchg operations
685 static int sock_map_ctx_update_elem(struct bpf_sock_ops_kern *skops,
687 void *key, u64 flags)
689 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
690 struct smap_psock_map_entry *e = NULL;
691 struct bpf_prog *verdict, *parse;
692 struct sock *osock, *sock;
693 struct smap_psock *psock;
697 if (unlikely(flags > BPF_EXIST))
700 if (unlikely(i >= stab->map.max_entries))
703 sock = READ_ONCE(stab->sock_map[i]);
704 if (flags == BPF_EXIST && !sock)
706 else if (flags == BPF_NOEXIST && sock)
711 /* 1. If sock map has BPF programs those will be inherited by the
712 * sock being added. If the sock is already attached to BPF programs
713 * this results in an error.
715 verdict = READ_ONCE(stab->bpf_verdict);
716 parse = READ_ONCE(stab->bpf_parse);
718 if (parse && verdict) {
719 /* bpf prog refcnt may be zero if a concurrent attach operation
720 * removes the program after the above READ_ONCE() but before
721 * we increment the refcnt. If this is the case abort with an
724 verdict = bpf_prog_inc_not_zero(stab->bpf_verdict);
726 return PTR_ERR(verdict);
728 parse = bpf_prog_inc_not_zero(stab->bpf_parse);
730 bpf_prog_put(verdict);
731 return PTR_ERR(parse);
735 write_lock_bh(&sock->sk_callback_lock);
736 psock = smap_psock_sk(sock);
738 /* 2. Do not allow inheriting programs if psock exists and has
739 * already inherited programs. This would create confusion on
740 * which parser/verdict program is running. If no psock exists
741 * create one. Inside sk_callback_lock to ensure concurrent create
742 * doesn't update user data.
745 if (READ_ONCE(psock->bpf_parse) && parse) {
751 psock = smap_init_psock(sock, stab);
753 err = PTR_ERR(psock);
757 set_bit(SMAP_TX_RUNNING, &psock->state);
760 e = kzalloc(sizeof(*e), GFP_ATOMIC | __GFP_NOWARN);
765 e->entry = &stab->sock_map[i];
767 /* 3. At this point we have a reference to a valid psock that is
768 * running. Attach any BPF programs needed.
770 if (parse && verdict && !psock->strp_enabled) {
771 err = smap_init_sock(psock, sock);
774 smap_init_progs(psock, stab, verdict, parse);
775 smap_start_sock(psock, sock);
778 /* 4. Place psock in sockmap for use and stop any programs on
779 * the old sock assuming its not the same sock we are replacing
780 * it with. Because we can only have a single set of programs if
781 * old_sock has a strp we can stop it.
783 list_add_tail(&e->list, &psock->maps);
784 write_unlock_bh(&sock->sk_callback_lock);
786 osock = xchg(&stab->sock_map[i], sock);
788 struct smap_psock *opsock = smap_psock_sk(osock);
790 write_lock_bh(&osock->sk_callback_lock);
791 if (osock != sock && parse)
792 smap_stop_sock(opsock, osock);
793 smap_list_remove(opsock, &stab->sock_map[i]);
794 smap_release_sock(opsock, osock);
795 write_unlock_bh(&osock->sk_callback_lock);
799 smap_release_sock(psock, sock);
802 bpf_prog_put(verdict);
805 write_unlock_bh(&sock->sk_callback_lock);
810 int sock_map_prog(struct bpf_map *map, struct bpf_prog *prog, u32 type)
812 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
813 struct bpf_prog *orig;
815 if (unlikely(map->map_type != BPF_MAP_TYPE_SOCKMAP))
819 case BPF_SK_SKB_STREAM_PARSER:
820 orig = xchg(&stab->bpf_parse, prog);
822 case BPF_SK_SKB_STREAM_VERDICT:
823 orig = xchg(&stab->bpf_verdict, prog);
835 static void *sock_map_lookup(struct bpf_map *map, void *key)
840 static int sock_map_update_elem(struct bpf_map *map,
841 void *key, void *value, u64 flags)
843 struct bpf_sock_ops_kern skops;
844 u32 fd = *(u32 *)value;
845 struct socket *socket;
848 socket = sockfd_lookup(fd, &err);
852 skops.sk = socket->sk;
858 if (skops.sk->sk_type != SOCK_STREAM ||
859 skops.sk->sk_protocol != IPPROTO_TCP) {
864 err = sock_map_ctx_update_elem(&skops, map, key, flags);
869 const struct bpf_map_ops sock_map_ops = {
870 .map_alloc = sock_map_alloc,
871 .map_free = sock_map_free,
872 .map_lookup_elem = sock_map_lookup,
873 .map_get_next_key = sock_map_get_next_key,
874 .map_update_elem = sock_map_update_elem,
875 .map_delete_elem = sock_map_delete_elem,
878 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, bpf_sock,
879 struct bpf_map *, map, void *, key, u64, flags)
881 WARN_ON_ONCE(!rcu_read_lock_held());
882 return sock_map_ctx_update_elem(bpf_sock, map, key, flags);
885 const struct bpf_func_proto bpf_sock_map_update_proto = {
886 .func = bpf_sock_map_update,
889 .ret_type = RET_INTEGER,
890 .arg1_type = ARG_PTR_TO_CTX,
891 .arg2_type = ARG_CONST_MAP_PTR,
892 .arg3_type = ARG_PTR_TO_MAP_KEY,
893 .arg4_type = ARG_ANYTHING,