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>
42 #include <net/strparser.h>
44 #include <linux/ptr_ring.h>
45 #include <net/inet_common.h>
47 #define SOCK_CREATE_FLAG_MASK \
48 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
52 struct sock **sock_map;
53 struct bpf_prog *bpf_tx_msg;
54 struct bpf_prog *bpf_parse;
55 struct bpf_prog *bpf_verdict;
58 enum smap_psock_state {
62 struct smap_psock_map_entry {
63 struct list_head list;
71 /* datapath variables */
72 struct sk_buff_head rxqueue;
75 /* datapath error path cache across tx work invocations */
78 struct sk_buff *save_skb;
80 /* datapath variables for tx_msg ULP */
81 struct sock *sk_redir;
86 struct sk_msg_buff *cork;
87 struct list_head ingress;
89 struct strparser strp;
90 struct bpf_prog *bpf_tx_msg;
91 struct bpf_prog *bpf_parse;
92 struct bpf_prog *bpf_verdict;
93 struct list_head maps;
95 /* Back reference used when sock callback trigger sockmap operations */
99 struct work_struct tx_work;
100 struct work_struct gc_work;
102 struct proto *sk_proto;
103 void (*save_close)(struct sock *sk, long timeout);
104 void (*save_data_ready)(struct sock *sk);
105 void (*save_write_space)(struct sock *sk);
108 static void smap_release_sock(struct smap_psock *psock, struct sock *sock);
109 static int bpf_tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
110 int nonblock, int flags, int *addr_len);
111 static int bpf_tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size);
112 static int bpf_tcp_sendpage(struct sock *sk, struct page *page,
113 int offset, size_t size, int flags);
115 static inline struct smap_psock *smap_psock_sk(const struct sock *sk)
117 return rcu_dereference_sk_user_data(sk);
120 static bool bpf_tcp_stream_read(const struct sock *sk)
122 struct smap_psock *psock;
126 psock = smap_psock_sk(sk);
127 if (unlikely(!psock))
129 empty = list_empty(&psock->ingress);
135 static struct proto tcp_bpf_proto;
136 static int bpf_tcp_init(struct sock *sk)
138 struct smap_psock *psock;
141 psock = smap_psock_sk(sk);
142 if (unlikely(!psock)) {
147 if (unlikely(psock->sk_proto)) {
152 psock->save_close = sk->sk_prot->close;
153 psock->sk_proto = sk->sk_prot;
155 if (psock->bpf_tx_msg) {
156 tcp_bpf_proto.sendmsg = bpf_tcp_sendmsg;
157 tcp_bpf_proto.sendpage = bpf_tcp_sendpage;
158 tcp_bpf_proto.recvmsg = bpf_tcp_recvmsg;
159 tcp_bpf_proto.stream_memory_read = bpf_tcp_stream_read;
162 sk->sk_prot = &tcp_bpf_proto;
167 static void smap_release_sock(struct smap_psock *psock, struct sock *sock);
168 static int free_start_sg(struct sock *sk, struct sk_msg_buff *md);
170 static void bpf_tcp_release(struct sock *sk)
172 struct smap_psock *psock;
175 psock = smap_psock_sk(sk);
176 if (unlikely(!psock))
180 free_start_sg(psock->sock, psock->cork);
185 if (psock->sk_proto) {
186 sk->sk_prot = psock->sk_proto;
187 psock->sk_proto = NULL;
193 static void bpf_tcp_close(struct sock *sk, long timeout)
195 void (*close_fun)(struct sock *sk, long timeout);
196 struct smap_psock_map_entry *e, *tmp;
197 struct sk_msg_buff *md, *mtmp;
198 struct smap_psock *psock;
202 psock = smap_psock_sk(sk);
203 if (unlikely(!psock)) {
205 return sk->sk_prot->close(sk, timeout);
208 /* The psock may be destroyed anytime after exiting the RCU critial
209 * section so by the time we use close_fun the psock may no longer
210 * be valid. However, bpf_tcp_close is called with the sock lock
211 * held so the close hook and sk are still valid.
213 close_fun = psock->save_close;
215 write_lock_bh(&sk->sk_callback_lock);
217 free_start_sg(psock->sock, psock->cork);
222 list_for_each_entry_safe(md, mtmp, &psock->ingress, list) {
224 free_start_sg(psock->sock, md);
228 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
229 osk = cmpxchg(e->entry, sk, NULL);
232 smap_release_sock(psock, sk);
235 write_unlock_bh(&sk->sk_callback_lock);
237 close_fun(sk, timeout);
247 static struct tcp_ulp_ops bpf_tcp_ulp_ops __read_mostly = {
250 .user_visible = false,
252 .init = bpf_tcp_init,
253 .release = bpf_tcp_release,
256 static int memcopy_from_iter(struct sock *sk,
257 struct sk_msg_buff *md,
258 struct iov_iter *from, int bytes)
260 struct scatterlist *sg = md->sg_data;
261 int i = md->sg_curr, rc = -ENOSPC;
267 if (md->sg_copybreak >= sg[i].length) {
268 md->sg_copybreak = 0;
270 if (++i == MAX_SKB_FRAGS)
277 copy = sg[i].length - md->sg_copybreak;
278 to = sg_virt(&sg[i]) + md->sg_copybreak;
279 md->sg_copybreak += copy;
281 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY)
282 rc = copy_from_iter_nocache(to, copy, from);
284 rc = copy_from_iter(to, copy, from);
295 md->sg_copybreak = 0;
296 if (++i == MAX_SKB_FRAGS)
298 } while (i != md->sg_end);
304 static int bpf_tcp_push(struct sock *sk, int apply_bytes,
305 struct sk_msg_buff *md,
306 int flags, bool uncharge)
308 bool apply = apply_bytes;
309 struct scatterlist *sg;
315 sg = md->sg_data + md->sg_start;
316 size = (apply && apply_bytes < sg->length) ?
317 apply_bytes : sg->length;
320 tcp_rate_check_app_limited(sk);
323 ret = do_tcp_sendpages(sk, p, offset, size, flags);
331 sk_mem_uncharge(sk, ret);
345 sk_mem_uncharge(sk, ret);
350 if (md->sg_start == MAX_SKB_FRAGS)
352 sg_init_table(sg, 1);
354 if (md->sg_start == md->sg_end)
358 if (apply && !apply_bytes)
364 static inline void bpf_compute_data_pointers_sg(struct sk_msg_buff *md)
366 struct scatterlist *sg = md->sg_data + md->sg_start;
368 if (md->sg_copy[md->sg_start]) {
369 md->data = md->data_end = 0;
371 md->data = sg_virt(sg);
372 md->data_end = md->data + sg->length;
376 static void return_mem_sg(struct sock *sk, int bytes, struct sk_msg_buff *md)
378 struct scatterlist *sg = md->sg_data;
379 int i = md->sg_start;
382 int uncharge = (bytes < sg[i].length) ? bytes : sg[i].length;
384 sk_mem_uncharge(sk, uncharge);
389 if (i == MAX_SKB_FRAGS)
391 } while (i != md->sg_end);
394 static void free_bytes_sg(struct sock *sk, int bytes, struct sk_msg_buff *md)
396 struct scatterlist *sg = md->sg_data;
397 int i = md->sg_start, free;
399 while (bytes && sg[i].length) {
402 sg[i].length -= bytes;
403 sg[i].offset += bytes;
404 sk_mem_uncharge(sk, bytes);
408 sk_mem_uncharge(sk, sg[i].length);
409 put_page(sg_page(&sg[i]));
410 bytes -= sg[i].length;
416 if (i == MAX_SKB_FRAGS)
421 static int free_sg(struct sock *sk, int start, struct sk_msg_buff *md)
423 struct scatterlist *sg = md->sg_data;
424 int i = start, free = 0;
426 while (sg[i].length) {
427 free += sg[i].length;
428 sk_mem_uncharge(sk, sg[i].length);
429 put_page(sg_page(&sg[i]));
435 if (i == MAX_SKB_FRAGS)
442 static int free_start_sg(struct sock *sk, struct sk_msg_buff *md)
444 int free = free_sg(sk, md->sg_start, md);
446 md->sg_start = md->sg_end;
450 static int free_curr_sg(struct sock *sk, struct sk_msg_buff *md)
452 return free_sg(sk, md->sg_curr, md);
455 static int bpf_map_msg_verdict(int _rc, struct sk_msg_buff *md)
457 return ((_rc == SK_PASS) ?
458 (md->map ? __SK_REDIRECT : __SK_PASS) :
462 static unsigned int smap_do_tx_msg(struct sock *sk,
463 struct smap_psock *psock,
464 struct sk_msg_buff *md)
466 struct bpf_prog *prog;
467 unsigned int rc, _rc;
472 /* If the policy was removed mid-send then default to 'accept' */
473 prog = READ_ONCE(psock->bpf_tx_msg);
474 if (unlikely(!prog)) {
479 bpf_compute_data_pointers_sg(md);
480 rc = (*prog->bpf_func)(md, prog->insnsi);
481 psock->apply_bytes = md->apply_bytes;
483 /* Moving return codes from UAPI namespace into internal namespace */
484 _rc = bpf_map_msg_verdict(rc, md);
486 /* The psock has a refcount on the sock but not on the map and because
487 * we need to drop rcu read lock here its possible the map could be
488 * removed between here and when we need it to execute the sock
489 * redirect. So do the map lookup now for future use.
491 if (_rc == __SK_REDIRECT) {
493 sock_put(psock->sk_redir);
494 psock->sk_redir = do_msg_redirect_map(md);
495 if (!psock->sk_redir) {
499 sock_hold(psock->sk_redir);
508 static int bpf_tcp_ingress(struct sock *sk, int apply_bytes,
509 struct smap_psock *psock,
510 struct sk_msg_buff *md, int flags)
512 bool apply = apply_bytes;
513 size_t size, copied = 0;
514 struct sk_msg_buff *r;
517 r = kzalloc(sizeof(struct sk_msg_buff), __GFP_NOWARN | GFP_KERNEL);
522 r->sg_start = md->sg_start;
526 r->sg_data[i] = md->sg_data[i];
528 size = (apply && apply_bytes < md->sg_data[i].length) ?
529 apply_bytes : md->sg_data[i].length;
531 if (!sk_wmem_schedule(sk, size)) {
537 sk_mem_charge(sk, size);
538 r->sg_data[i].length = size;
539 md->sg_data[i].length -= size;
540 md->sg_data[i].offset += size;
543 if (md->sg_data[i].length) {
544 get_page(sg_page(&r->sg_data[i]));
545 r->sg_end = (i + 1) == MAX_SKB_FRAGS ? 0 : i + 1;
548 if (i == MAX_SKB_FRAGS)
558 } while (i != md->sg_end);
563 list_add_tail(&r->list, &psock->ingress);
564 sk->sk_data_ready(sk);
566 free_start_sg(sk, r);
574 static int bpf_tcp_sendmsg_do_redirect(struct sock *sk, int send,
575 struct sk_msg_buff *md,
578 struct smap_psock *psock;
579 struct scatterlist *sg;
580 int i, err, free = 0;
581 bool ingress = !!(md->flags & BPF_F_INGRESS);
586 psock = smap_psock_sk(sk);
587 if (unlikely(!psock))
590 if (!refcount_inc_not_zero(&psock->refcnt))
596 err = bpf_tcp_ingress(sk, send, psock, md, flags);
599 err = bpf_tcp_push(sk, send, md, flags, false);
602 smap_release_sock(psock, sk);
610 while (sg[i].length) {
611 free += sg[i].length;
612 put_page(sg_page(&sg[i]));
615 if (i == MAX_SKB_FRAGS)
621 static inline void bpf_md_init(struct smap_psock *psock)
623 if (!psock->apply_bytes) {
624 psock->eval = __SK_NONE;
625 if (psock->sk_redir) {
626 sock_put(psock->sk_redir);
627 psock->sk_redir = NULL;
632 static void apply_bytes_dec(struct smap_psock *psock, int i)
634 if (psock->apply_bytes) {
635 if (psock->apply_bytes < i)
636 psock->apply_bytes = 0;
638 psock->apply_bytes -= i;
642 static int bpf_exec_tx_verdict(struct smap_psock *psock,
643 struct sk_msg_buff *m,
645 int *copied, int flags)
647 bool cork = false, enospc = (m->sg_start == m->sg_end);
653 if (psock->eval == __SK_NONE)
654 psock->eval = smap_do_tx_msg(sk, psock, m);
657 m->cork_bytes > psock->sg_size && !enospc) {
658 psock->cork_bytes = m->cork_bytes - psock->sg_size;
660 psock->cork = kcalloc(1,
661 sizeof(struct sk_msg_buff),
662 GFP_ATOMIC | __GFP_NOWARN);
669 memcpy(psock->cork, m, sizeof(*m));
673 send = psock->sg_size;
674 if (psock->apply_bytes && psock->apply_bytes < send)
675 send = psock->apply_bytes;
677 switch (psock->eval) {
679 err = bpf_tcp_push(sk, send, m, flags, true);
681 *copied -= free_start_sg(sk, m);
685 apply_bytes_dec(psock, send);
686 psock->sg_size -= send;
689 redir = psock->sk_redir;
690 apply_bytes_dec(psock, send);
697 return_mem_sg(sk, send, m);
700 err = bpf_tcp_sendmsg_do_redirect(redir, send, m, flags);
704 free_start_sg(sk, m);
711 psock->sg_size -= send;
715 free_bytes_sg(sk, send, m);
716 apply_bytes_dec(psock, send);
718 psock->sg_size -= send;
726 m->sg_data[m->sg_start].page_link &&
727 m->sg_data[m->sg_start].length)
735 static int bpf_tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
736 int nonblock, int flags, int *addr_len)
738 struct iov_iter *iter = &msg->msg_iter;
739 struct smap_psock *psock;
742 if (unlikely(flags & MSG_ERRQUEUE))
743 return inet_recv_error(sk, msg, len, addr_len);
746 psock = smap_psock_sk(sk);
747 if (unlikely(!psock))
750 if (unlikely(!refcount_inc_not_zero(&psock->refcnt)))
754 if (!skb_queue_empty(&sk->sk_receive_queue))
755 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
758 while (copied != len) {
759 struct scatterlist *sg;
760 struct sk_msg_buff *md;
763 md = list_first_entry_or_null(&psock->ingress,
764 struct sk_msg_buff, list);
772 sg = &md->sg_data[i];
776 if (copied + copy > len)
779 n = copy_page_to_iter(page, sg->offset, copy, iter);
783 smap_release_sock(psock, sk);
790 sk_mem_uncharge(sk, copy);
794 if (i == MAX_SKB_FRAGS)
801 } while (i != md->sg_end);
804 if (!sg->length && md->sg_start == md->sg_end) {
807 consume_skb(md->skb);
813 smap_release_sock(psock, sk);
817 return tcp_recvmsg(sk, msg, len, nonblock, flags, addr_len);
821 static int bpf_tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
823 int flags = msg->msg_flags | MSG_NO_SHARED_FRAGS;
824 struct sk_msg_buff md = {0};
825 unsigned int sg_copy = 0;
826 struct smap_psock *psock;
827 int copied = 0, err = 0;
828 struct scatterlist *sg;
831 /* Its possible a sock event or user removed the psock _but_ the ops
832 * have not been reprogrammed yet so we get here. In this case fallback
833 * to tcp_sendmsg. Note this only works because we _only_ ever allow
834 * a single ULP there is no hierarchy here.
837 psock = smap_psock_sk(sk);
838 if (unlikely(!psock)) {
840 return tcp_sendmsg(sk, msg, size);
843 /* Increment the psock refcnt to ensure its not released while sending a
844 * message. Required because sk lookup and bpf programs are used in
845 * separate rcu critical sections. Its OK if we lose the map entry
846 * but we can't lose the sock reference.
848 if (!refcount_inc_not_zero(&psock->refcnt)) {
850 return tcp_sendmsg(sk, msg, size);
854 sg_init_marker(sg, MAX_SKB_FRAGS);
858 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
860 while (msg_data_left(msg)) {
861 struct sk_msg_buff *m;
870 copy = msg_data_left(msg);
871 if (!sk_stream_memory_free(sk))
872 goto wait_for_sndbuf;
874 m = psock->cork_bytes ? psock->cork : &md;
875 m->sg_curr = m->sg_copybreak ? m->sg_curr : m->sg_end;
876 err = sk_alloc_sg(sk, copy, m->sg_data,
877 m->sg_start, &m->sg_end, &sg_copy,
881 goto wait_for_memory;
886 err = memcopy_from_iter(sk, m, &msg->msg_iter, copy);
892 psock->sg_size += copy;
896 /* When bytes are being corked skip running BPF program and
897 * applying verdict unless there is no more buffer space. In
898 * the ENOSPC case simply run BPF prorgram with currently
899 * accumulated data. We don't have much choice at this point
900 * we could try extending the page frags or chaining complex
901 * frags but even in these cases _eventually_ we will hit an
902 * OOM scenario. More complex recovery schemes may be
903 * implemented in the future, but BPF programs must handle
904 * the case where apply_cork requests are not honored. The
905 * canonical method to verify this is to check data length.
907 if (psock->cork_bytes) {
908 if (copy > psock->cork_bytes)
909 psock->cork_bytes = 0;
911 psock->cork_bytes -= copy;
913 if (psock->cork_bytes && !enospc)
916 /* All cork bytes accounted for re-run filter */
917 psock->eval = __SK_NONE;
918 psock->cork_bytes = 0;
921 err = bpf_exec_tx_verdict(psock, m, sk, &copied, flags);
922 if (unlikely(err < 0))
926 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
928 err = sk_stream_wait_memory(sk, &timeo);
934 err = sk_stream_error(sk, msg->msg_flags, err);
937 smap_release_sock(psock, sk);
938 return copied ? copied : err;
941 static int bpf_tcp_sendpage(struct sock *sk, struct page *page,
942 int offset, size_t size, int flags)
944 struct sk_msg_buff md = {0}, *m = NULL;
945 int err = 0, copied = 0;
946 struct smap_psock *psock;
947 struct scatterlist *sg;
951 psock = smap_psock_sk(sk);
952 if (unlikely(!psock))
955 if (!refcount_inc_not_zero(&psock->refcnt))
961 if (psock->cork_bytes) {
963 sg = &m->sg_data[m->sg_end];
967 sg_init_marker(sg, MAX_SKB_FRAGS);
970 /* Catch case where ring is full and sendpage is stalled. */
971 if (unlikely(m->sg_end == m->sg_start &&
972 m->sg_data[m->sg_end].length))
975 psock->sg_size += size;
976 sg_set_page(sg, page, size, offset);
978 m->sg_copy[m->sg_end] = true;
979 sk_mem_charge(sk, size);
983 if (m->sg_end == MAX_SKB_FRAGS)
986 if (m->sg_end == m->sg_start)
989 if (psock->cork_bytes) {
990 if (size > psock->cork_bytes)
991 psock->cork_bytes = 0;
993 psock->cork_bytes -= size;
995 if (psock->cork_bytes && !enospc)
998 /* All cork bytes accounted for re-run filter */
999 psock->eval = __SK_NONE;
1000 psock->cork_bytes = 0;
1003 err = bpf_exec_tx_verdict(psock, m, sk, &copied, flags);
1006 smap_release_sock(psock, sk);
1007 return copied ? copied : err;
1010 return tcp_sendpage(sk, page, offset, size, flags);
1013 static void bpf_tcp_msg_add(struct smap_psock *psock,
1015 struct bpf_prog *tx_msg)
1017 struct bpf_prog *orig_tx_msg;
1019 orig_tx_msg = xchg(&psock->bpf_tx_msg, tx_msg);
1021 bpf_prog_put(orig_tx_msg);
1024 static int bpf_tcp_ulp_register(void)
1026 tcp_bpf_proto = tcp_prot;
1027 tcp_bpf_proto.close = bpf_tcp_close;
1028 /* Once BPF TX ULP is registered it is never unregistered. It
1029 * will be in the ULP list for the lifetime of the system. Doing
1030 * duplicate registers is not a problem.
1032 return tcp_register_ulp(&bpf_tcp_ulp_ops);
1035 static int smap_verdict_func(struct smap_psock *psock, struct sk_buff *skb)
1037 struct bpf_prog *prog = READ_ONCE(psock->bpf_verdict);
1040 if (unlikely(!prog))
1044 /* We need to ensure that BPF metadata for maps is also cleared
1045 * when we orphan the skb so that we don't have the possibility
1046 * to reference a stale map.
1048 TCP_SKB_CB(skb)->bpf.map = NULL;
1049 skb->sk = psock->sock;
1050 bpf_compute_data_pointers(skb);
1052 rc = (*prog->bpf_func)(skb, prog->insnsi);
1056 /* Moving return codes from UAPI namespace into internal namespace */
1057 return rc == SK_PASS ?
1058 (TCP_SKB_CB(skb)->bpf.map ? __SK_REDIRECT : __SK_PASS) :
1062 static int smap_do_ingress(struct smap_psock *psock, struct sk_buff *skb)
1064 struct sock *sk = psock->sock;
1065 int copied = 0, num_sg;
1066 struct sk_msg_buff *r;
1068 r = kzalloc(sizeof(struct sk_msg_buff), __GFP_NOWARN | GFP_ATOMIC);
1072 if (!sk_rmem_schedule(sk, skb, skb->len)) {
1077 sg_init_table(r->sg_data, MAX_SKB_FRAGS);
1078 num_sg = skb_to_sgvec(skb, r->sg_data, 0, skb->len);
1079 if (unlikely(num_sg < 0)) {
1083 sk_mem_charge(sk, skb->len);
1086 r->sg_end = num_sg == MAX_SKB_FRAGS ? 0 : num_sg;
1088 list_add_tail(&r->list, &psock->ingress);
1089 sk->sk_data_ready(sk);
1093 static void smap_do_verdict(struct smap_psock *psock, struct sk_buff *skb)
1095 struct smap_psock *peer;
1100 rc = smap_verdict_func(psock, skb);
1103 sk = do_sk_redirect_map(skb);
1109 peer = smap_psock_sk(sk);
1110 in = (TCP_SKB_CB(skb)->bpf.flags) & BPF_F_INGRESS;
1112 if (unlikely(!peer || sock_flag(sk, SOCK_DEAD) ||
1113 !test_bit(SMAP_TX_RUNNING, &peer->state))) {
1118 if (!in && sock_writeable(sk)) {
1119 skb_set_owner_w(skb, sk);
1120 skb_queue_tail(&peer->rxqueue, skb);
1121 schedule_work(&peer->tx_work);
1124 atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) {
1125 skb_queue_tail(&peer->rxqueue, skb);
1126 schedule_work(&peer->tx_work);
1129 /* Fall through and free skb otherwise */
1136 static void smap_report_sk_error(struct smap_psock *psock, int err)
1138 struct sock *sk = psock->sock;
1141 sk->sk_error_report(sk);
1144 static void smap_read_sock_strparser(struct strparser *strp,
1145 struct sk_buff *skb)
1147 struct smap_psock *psock;
1150 psock = container_of(strp, struct smap_psock, strp);
1151 smap_do_verdict(psock, skb);
1155 /* Called with lock held on socket */
1156 static void smap_data_ready(struct sock *sk)
1158 struct smap_psock *psock;
1161 psock = smap_psock_sk(sk);
1162 if (likely(psock)) {
1163 write_lock_bh(&sk->sk_callback_lock);
1164 strp_data_ready(&psock->strp);
1165 write_unlock_bh(&sk->sk_callback_lock);
1170 static void smap_tx_work(struct work_struct *w)
1172 struct smap_psock *psock;
1173 struct sk_buff *skb;
1176 psock = container_of(w, struct smap_psock, tx_work);
1178 /* lock sock to avoid losing sk_socket at some point during loop */
1179 lock_sock(psock->sock);
1180 if (psock->save_skb) {
1181 skb = psock->save_skb;
1182 rem = psock->save_rem;
1183 off = psock->save_off;
1184 psock->save_skb = NULL;
1188 while ((skb = skb_dequeue(&psock->rxqueue))) {
1194 flags = (TCP_SKB_CB(skb)->bpf.flags) & BPF_F_INGRESS;
1196 if (likely(psock->sock->sk_socket)) {
1198 n = smap_do_ingress(psock, skb);
1200 n = skb_send_sock_locked(psock->sock,
1208 /* Retry when space is available */
1209 psock->save_skb = skb;
1210 psock->save_rem = rem;
1211 psock->save_off = off;
1214 /* Hard errors break pipe and stop xmit */
1215 smap_report_sk_error(psock, n ? -n : EPIPE);
1216 clear_bit(SMAP_TX_RUNNING, &psock->state);
1228 release_sock(psock->sock);
1231 static void smap_write_space(struct sock *sk)
1233 struct smap_psock *psock;
1236 psock = smap_psock_sk(sk);
1237 if (likely(psock && test_bit(SMAP_TX_RUNNING, &psock->state)))
1238 schedule_work(&psock->tx_work);
1242 static void smap_stop_sock(struct smap_psock *psock, struct sock *sk)
1244 if (!psock->strp_enabled)
1246 sk->sk_data_ready = psock->save_data_ready;
1247 sk->sk_write_space = psock->save_write_space;
1248 psock->save_data_ready = NULL;
1249 psock->save_write_space = NULL;
1250 strp_stop(&psock->strp);
1251 psock->strp_enabled = false;
1254 static void smap_destroy_psock(struct rcu_head *rcu)
1256 struct smap_psock *psock = container_of(rcu,
1257 struct smap_psock, rcu);
1259 /* Now that a grace period has passed there is no longer
1260 * any reference to this sock in the sockmap so we can
1261 * destroy the psock, strparser, and bpf programs. But,
1262 * because we use workqueue sync operations we can not
1263 * do it in rcu context
1265 schedule_work(&psock->gc_work);
1268 static void smap_release_sock(struct smap_psock *psock, struct sock *sock)
1270 if (refcount_dec_and_test(&psock->refcnt)) {
1271 tcp_cleanup_ulp(sock);
1272 smap_stop_sock(psock, sock);
1273 clear_bit(SMAP_TX_RUNNING, &psock->state);
1274 rcu_assign_sk_user_data(sock, NULL);
1275 call_rcu_sched(&psock->rcu, smap_destroy_psock);
1279 static int smap_parse_func_strparser(struct strparser *strp,
1280 struct sk_buff *skb)
1282 struct smap_psock *psock;
1283 struct bpf_prog *prog;
1287 psock = container_of(strp, struct smap_psock, strp);
1288 prog = READ_ONCE(psock->bpf_parse);
1290 if (unlikely(!prog)) {
1295 /* Attach socket for bpf program to use if needed we can do this
1296 * because strparser clones the skb before handing it to a upper
1297 * layer, meaning skb_orphan has been called. We NULL sk on the
1298 * way out to ensure we don't trigger a BUG_ON in skb/sk operations
1299 * later and because we are not charging the memory of this skb to
1302 skb->sk = psock->sock;
1303 bpf_compute_data_pointers(skb);
1304 rc = (*prog->bpf_func)(skb, prog->insnsi);
1310 static int smap_read_sock_done(struct strparser *strp, int err)
1315 static int smap_init_sock(struct smap_psock *psock,
1318 static const struct strp_callbacks cb = {
1319 .rcv_msg = smap_read_sock_strparser,
1320 .parse_msg = smap_parse_func_strparser,
1321 .read_sock_done = smap_read_sock_done,
1324 return strp_init(&psock->strp, sk, &cb);
1327 static void smap_init_progs(struct smap_psock *psock,
1328 struct bpf_stab *stab,
1329 struct bpf_prog *verdict,
1330 struct bpf_prog *parse)
1332 struct bpf_prog *orig_parse, *orig_verdict;
1334 orig_parse = xchg(&psock->bpf_parse, parse);
1335 orig_verdict = xchg(&psock->bpf_verdict, verdict);
1338 bpf_prog_put(orig_verdict);
1340 bpf_prog_put(orig_parse);
1343 static void smap_start_sock(struct smap_psock *psock, struct sock *sk)
1345 if (sk->sk_data_ready == smap_data_ready)
1347 psock->save_data_ready = sk->sk_data_ready;
1348 psock->save_write_space = sk->sk_write_space;
1349 sk->sk_data_ready = smap_data_ready;
1350 sk->sk_write_space = smap_write_space;
1351 psock->strp_enabled = true;
1354 static void sock_map_remove_complete(struct bpf_stab *stab)
1356 bpf_map_area_free(stab->sock_map);
1360 static void smap_gc_work(struct work_struct *w)
1362 struct smap_psock_map_entry *e, *tmp;
1363 struct sk_msg_buff *md, *mtmp;
1364 struct smap_psock *psock;
1366 psock = container_of(w, struct smap_psock, gc_work);
1368 /* no callback lock needed because we already detached sockmap ops */
1369 if (psock->strp_enabled)
1370 strp_done(&psock->strp);
1372 cancel_work_sync(&psock->tx_work);
1373 __skb_queue_purge(&psock->rxqueue);
1375 /* At this point all strparser and xmit work must be complete */
1376 if (psock->bpf_parse)
1377 bpf_prog_put(psock->bpf_parse);
1378 if (psock->bpf_verdict)
1379 bpf_prog_put(psock->bpf_verdict);
1380 if (psock->bpf_tx_msg)
1381 bpf_prog_put(psock->bpf_tx_msg);
1384 free_start_sg(psock->sock, psock->cork);
1388 list_for_each_entry_safe(md, mtmp, &psock->ingress, list) {
1389 list_del(&md->list);
1390 free_start_sg(psock->sock, md);
1394 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
1399 if (psock->sk_redir)
1400 sock_put(psock->sk_redir);
1402 sock_put(psock->sock);
1406 static struct smap_psock *smap_init_psock(struct sock *sock,
1407 struct bpf_stab *stab)
1409 struct smap_psock *psock;
1411 psock = kzalloc_node(sizeof(struct smap_psock),
1412 GFP_ATOMIC | __GFP_NOWARN,
1413 stab->map.numa_node);
1415 return ERR_PTR(-ENOMEM);
1417 psock->eval = __SK_NONE;
1419 skb_queue_head_init(&psock->rxqueue);
1420 INIT_WORK(&psock->tx_work, smap_tx_work);
1421 INIT_WORK(&psock->gc_work, smap_gc_work);
1422 INIT_LIST_HEAD(&psock->maps);
1423 INIT_LIST_HEAD(&psock->ingress);
1424 refcount_set(&psock->refcnt, 1);
1426 rcu_assign_sk_user_data(sock, psock);
1431 static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
1433 struct bpf_stab *stab;
1437 if (!capable(CAP_NET_ADMIN))
1438 return ERR_PTR(-EPERM);
1440 /* check sanity of attributes */
1441 if (attr->max_entries == 0 || attr->key_size != 4 ||
1442 attr->value_size != 4 || attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
1443 return ERR_PTR(-EINVAL);
1445 err = bpf_tcp_ulp_register();
1446 if (err && err != -EEXIST)
1447 return ERR_PTR(err);
1449 stab = kzalloc(sizeof(*stab), GFP_USER);
1451 return ERR_PTR(-ENOMEM);
1453 bpf_map_init_from_attr(&stab->map, attr);
1455 /* make sure page count doesn't overflow */
1456 cost = (u64) stab->map.max_entries * sizeof(struct sock *);
1458 if (cost >= U32_MAX - PAGE_SIZE)
1461 stab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
1463 /* if map size is larger than memlock limit, reject it early */
1464 err = bpf_map_precharge_memlock(stab->map.pages);
1469 stab->sock_map = bpf_map_area_alloc(stab->map.max_entries *
1470 sizeof(struct sock *),
1471 stab->map.numa_node);
1472 if (!stab->sock_map)
1478 return ERR_PTR(err);
1481 static void smap_list_remove(struct smap_psock *psock, struct sock **entry)
1483 struct smap_psock_map_entry *e, *tmp;
1485 list_for_each_entry_safe(e, tmp, &psock->maps, list) {
1486 if (e->entry == entry) {
1493 static void sock_map_free(struct bpf_map *map)
1495 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
1500 /* At this point no update, lookup or delete operations can happen.
1501 * However, be aware we can still get a socket state event updates,
1502 * and data ready callabacks that reference the psock from sk_user_data
1503 * Also psock worker threads are still in-flight. So smap_release_sock
1504 * will only free the psock after cancel_sync on the worker threads
1505 * and a grace period expire to ensure psock is really safe to remove.
1508 for (i = 0; i < stab->map.max_entries; i++) {
1509 struct smap_psock *psock;
1512 sock = xchg(&stab->sock_map[i], NULL);
1516 write_lock_bh(&sock->sk_callback_lock);
1517 psock = smap_psock_sk(sock);
1518 /* This check handles a racing sock event that can get the
1519 * sk_callback_lock before this case but after xchg happens
1520 * causing the refcnt to hit zero and sock user data (psock)
1521 * to be null and queued for garbage collection.
1523 if (likely(psock)) {
1524 smap_list_remove(psock, &stab->sock_map[i]);
1525 smap_release_sock(psock, sock);
1527 write_unlock_bh(&sock->sk_callback_lock);
1531 sock_map_remove_complete(stab);
1534 static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
1536 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
1537 u32 i = key ? *(u32 *)key : U32_MAX;
1538 u32 *next = (u32 *)next_key;
1540 if (i >= stab->map.max_entries) {
1545 if (i == stab->map.max_entries - 1)
1552 struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
1554 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
1556 if (key >= map->max_entries)
1559 return READ_ONCE(stab->sock_map[key]);
1562 static int sock_map_delete_elem(struct bpf_map *map, void *key)
1564 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
1565 struct smap_psock *psock;
1566 int k = *(u32 *)key;
1569 if (k >= map->max_entries)
1572 sock = xchg(&stab->sock_map[k], NULL);
1576 write_lock_bh(&sock->sk_callback_lock);
1577 psock = smap_psock_sk(sock);
1581 if (psock->bpf_parse)
1582 smap_stop_sock(psock, sock);
1583 smap_list_remove(psock, &stab->sock_map[k]);
1584 smap_release_sock(psock, sock);
1586 write_unlock_bh(&sock->sk_callback_lock);
1590 /* Locking notes: Concurrent updates, deletes, and lookups are allowed and are
1591 * done inside rcu critical sections. This ensures on updates that the psock
1592 * will not be released via smap_release_sock() until concurrent updates/deletes
1593 * complete. All operations operate on sock_map using cmpxchg and xchg
1594 * operations to ensure we do not get stale references. Any reads into the
1595 * map must be done with READ_ONCE() because of this.
1597 * A psock is destroyed via call_rcu and after any worker threads are cancelled
1598 * and syncd so we are certain all references from the update/lookup/delete
1599 * operations as well as references in the data path are no longer in use.
1601 * Psocks may exist in multiple maps, but only a single set of parse/verdict
1602 * programs may be inherited from the maps it belongs to. A reference count
1603 * is kept with the total number of references to the psock from all maps. The
1604 * psock will not be released until this reaches zero. The psock and sock
1605 * user data data use the sk_callback_lock to protect critical data structures
1606 * from concurrent access. This allows us to avoid two updates from modifying
1607 * the user data in sock and the lock is required anyways for modifying
1608 * callbacks, we simply increase its scope slightly.
1611 * - psock must always be read inside RCU critical section
1612 * - sk_user_data must only be modified inside sk_callback_lock and read
1613 * inside RCU critical section.
1614 * - psock->maps list must only be read & modified inside sk_callback_lock
1615 * - sock_map must use READ_ONCE and (cmp)xchg operations
1616 * - BPF verdict/parse programs must use READ_ONCE and xchg operations
1618 static int sock_map_ctx_update_elem(struct bpf_sock_ops_kern *skops,
1619 struct bpf_map *map,
1620 void *key, u64 flags)
1622 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
1623 struct smap_psock_map_entry *e = NULL;
1624 struct bpf_prog *verdict, *parse, *tx_msg;
1625 struct sock *osock, *sock;
1626 struct smap_psock *psock;
1627 u32 i = *(u32 *)key;
1631 if (unlikely(flags > BPF_EXIST))
1634 if (unlikely(i >= stab->map.max_entries))
1637 sock = READ_ONCE(stab->sock_map[i]);
1638 if (flags == BPF_EXIST && !sock)
1640 else if (flags == BPF_NOEXIST && sock)
1645 /* 1. If sock map has BPF programs those will be inherited by the
1646 * sock being added. If the sock is already attached to BPF programs
1647 * this results in an error.
1649 verdict = READ_ONCE(stab->bpf_verdict);
1650 parse = READ_ONCE(stab->bpf_parse);
1651 tx_msg = READ_ONCE(stab->bpf_tx_msg);
1653 if (parse && verdict) {
1654 /* bpf prog refcnt may be zero if a concurrent attach operation
1655 * removes the program after the above READ_ONCE() but before
1656 * we increment the refcnt. If this is the case abort with an
1659 verdict = bpf_prog_inc_not_zero(stab->bpf_verdict);
1660 if (IS_ERR(verdict))
1661 return PTR_ERR(verdict);
1663 parse = bpf_prog_inc_not_zero(stab->bpf_parse);
1664 if (IS_ERR(parse)) {
1665 bpf_prog_put(verdict);
1666 return PTR_ERR(parse);
1671 tx_msg = bpf_prog_inc_not_zero(stab->bpf_tx_msg);
1672 if (IS_ERR(tx_msg)) {
1674 bpf_prog_put(verdict);
1676 bpf_prog_put(parse);
1677 return PTR_ERR(tx_msg);
1681 write_lock_bh(&sock->sk_callback_lock);
1682 psock = smap_psock_sk(sock);
1684 /* 2. Do not allow inheriting programs if psock exists and has
1685 * already inherited programs. This would create confusion on
1686 * which parser/verdict program is running. If no psock exists
1687 * create one. Inside sk_callback_lock to ensure concurrent create
1688 * doesn't update user data.
1691 if (READ_ONCE(psock->bpf_parse) && parse) {
1695 if (READ_ONCE(psock->bpf_tx_msg) && tx_msg) {
1699 if (!refcount_inc_not_zero(&psock->refcnt)) {
1704 psock = smap_init_psock(sock, stab);
1705 if (IS_ERR(psock)) {
1706 err = PTR_ERR(psock);
1710 set_bit(SMAP_TX_RUNNING, &psock->state);
1714 e = kzalloc(sizeof(*e), GFP_ATOMIC | __GFP_NOWARN);
1719 e->entry = &stab->sock_map[i];
1721 /* 3. At this point we have a reference to a valid psock that is
1722 * running. Attach any BPF programs needed.
1725 bpf_tcp_msg_add(psock, sock, tx_msg);
1727 err = tcp_set_ulp_id(sock, TCP_ULP_BPF);
1732 if (parse && verdict && !psock->strp_enabled) {
1733 err = smap_init_sock(psock, sock);
1736 smap_init_progs(psock, stab, verdict, parse);
1737 smap_start_sock(psock, sock);
1740 /* 4. Place psock in sockmap for use and stop any programs on
1741 * the old sock assuming its not the same sock we are replacing
1742 * it with. Because we can only have a single set of programs if
1743 * old_sock has a strp we can stop it.
1745 list_add_tail(&e->list, &psock->maps);
1746 write_unlock_bh(&sock->sk_callback_lock);
1748 osock = xchg(&stab->sock_map[i], sock);
1750 struct smap_psock *opsock = smap_psock_sk(osock);
1752 write_lock_bh(&osock->sk_callback_lock);
1753 smap_list_remove(opsock, &stab->sock_map[i]);
1754 smap_release_sock(opsock, osock);
1755 write_unlock_bh(&osock->sk_callback_lock);
1759 smap_release_sock(psock, sock);
1762 bpf_prog_put(verdict);
1764 bpf_prog_put(parse);
1766 bpf_prog_put(tx_msg);
1767 write_unlock_bh(&sock->sk_callback_lock);
1772 int sock_map_prog(struct bpf_map *map, struct bpf_prog *prog, u32 type)
1774 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
1775 struct bpf_prog *orig;
1777 if (unlikely(map->map_type != BPF_MAP_TYPE_SOCKMAP))
1781 case BPF_SK_MSG_VERDICT:
1782 orig = xchg(&stab->bpf_tx_msg, prog);
1784 case BPF_SK_SKB_STREAM_PARSER:
1785 orig = xchg(&stab->bpf_parse, prog);
1787 case BPF_SK_SKB_STREAM_VERDICT:
1788 orig = xchg(&stab->bpf_verdict, prog);
1800 static void *sock_map_lookup(struct bpf_map *map, void *key)
1805 static int sock_map_update_elem(struct bpf_map *map,
1806 void *key, void *value, u64 flags)
1808 struct bpf_sock_ops_kern skops;
1809 u32 fd = *(u32 *)value;
1810 struct socket *socket;
1813 socket = sockfd_lookup(fd, &err);
1817 skops.sk = socket->sk;
1823 if (skops.sk->sk_type != SOCK_STREAM ||
1824 skops.sk->sk_protocol != IPPROTO_TCP) {
1829 err = sock_map_ctx_update_elem(&skops, map, key, flags);
1834 static void sock_map_release(struct bpf_map *map, struct file *map_file)
1836 struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
1837 struct bpf_prog *orig;
1839 orig = xchg(&stab->bpf_parse, NULL);
1842 orig = xchg(&stab->bpf_verdict, NULL);
1846 orig = xchg(&stab->bpf_tx_msg, NULL);
1851 const struct bpf_map_ops sock_map_ops = {
1852 .map_alloc = sock_map_alloc,
1853 .map_free = sock_map_free,
1854 .map_lookup_elem = sock_map_lookup,
1855 .map_get_next_key = sock_map_get_next_key,
1856 .map_update_elem = sock_map_update_elem,
1857 .map_delete_elem = sock_map_delete_elem,
1858 .map_release = sock_map_release,
1861 BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, bpf_sock,
1862 struct bpf_map *, map, void *, key, u64, flags)
1864 WARN_ON_ONCE(!rcu_read_lock_held());
1865 return sock_map_ctx_update_elem(bpf_sock, map, key, flags);
1868 const struct bpf_func_proto bpf_sock_map_update_proto = {
1869 .func = bpf_sock_map_update,
1872 .ret_type = RET_INTEGER,
1873 .arg1_type = ARG_PTR_TO_CTX,
1874 .arg2_type = ARG_CONST_MAP_PTR,
1875 .arg3_type = ARG_PTR_TO_MAP_KEY,
1876 .arg4_type = ARG_ANYTHING,