1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
28 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
30 struct mptcp_sock msk;
41 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
43 static struct percpu_counter mptcp_sockets_allocated;
45 static void __mptcp_destroy_sock(struct sock *sk);
46 static void __mptcp_check_send_data_fin(struct sock *sk);
48 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
49 static struct net_device mptcp_napi_dev;
51 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
52 * completed yet or has failed, return the subflow socket.
53 * Otherwise return NULL.
55 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
57 if (!msk->subflow || READ_ONCE(msk->can_ack))
63 /* Returns end sequence number of the receiver's advertised window */
64 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
66 return READ_ONCE(msk->wnd_end);
69 static bool mptcp_is_tcpsk(struct sock *sk)
71 struct socket *sock = sk->sk_socket;
73 if (unlikely(sk->sk_prot == &tcp_prot)) {
74 /* we are being invoked after mptcp_accept() has
75 * accepted a non-mp-capable flow: sk is a tcp_sk,
78 * Hand the socket over to tcp so all further socket ops
81 sock->ops = &inet_stream_ops;
83 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
84 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
85 sock->ops = &inet6_stream_ops;
93 static int __mptcp_socket_create(struct mptcp_sock *msk)
95 struct mptcp_subflow_context *subflow;
96 struct sock *sk = (struct sock *)msk;
100 err = mptcp_subflow_create_socket(sk, &ssock);
104 msk->first = ssock->sk;
105 msk->subflow = ssock;
106 subflow = mptcp_subflow_ctx(ssock->sk);
107 list_add(&subflow->node, &msk->conn_list);
108 sock_hold(ssock->sk);
109 subflow->request_mptcp = 1;
110 mptcp_sock_graft(msk->first, sk->sk_socket);
115 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
117 sk_drops_add(sk, skb);
121 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
122 struct sk_buff *from)
127 if (MPTCP_SKB_CB(from)->offset ||
128 !skb_try_coalesce(to, from, &fragstolen, &delta))
131 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
132 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
133 to->len, MPTCP_SKB_CB(from)->end_seq);
134 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
135 kfree_skb_partial(from, fragstolen);
136 atomic_add(delta, &sk->sk_rmem_alloc);
137 sk_mem_charge(sk, delta);
141 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
142 struct sk_buff *from)
144 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
147 return mptcp_try_coalesce((struct sock *)msk, to, from);
150 /* "inspired" by tcp_data_queue_ofo(), main differences:
152 * - don't cope with sacks
154 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
156 struct sock *sk = (struct sock *)msk;
157 struct rb_node **p, *parent;
158 u64 seq, end_seq, max_seq;
159 struct sk_buff *skb1;
161 seq = MPTCP_SKB_CB(skb)->map_seq;
162 end_seq = MPTCP_SKB_CB(skb)->end_seq;
163 max_seq = READ_ONCE(msk->rcv_wnd_sent);
165 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
166 RB_EMPTY_ROOT(&msk->out_of_order_queue));
167 if (after64(end_seq, max_seq)) {
170 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
171 (unsigned long long)end_seq - (unsigned long)max_seq,
172 (unsigned long long)msk->rcv_wnd_sent);
173 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
177 p = &msk->out_of_order_queue.rb_node;
178 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
179 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
180 rb_link_node(&skb->rbnode, NULL, p);
181 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
182 msk->ooo_last_skb = skb;
186 /* with 2 subflows, adding at end of ooo queue is quite likely
187 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
189 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
190 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
191 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
195 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
196 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
197 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
198 parent = &msk->ooo_last_skb->rbnode;
199 p = &parent->rb_right;
203 /* Find place to insert this segment. Handle overlaps on the way. */
207 skb1 = rb_to_skb(parent);
208 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
209 p = &parent->rb_left;
212 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
213 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
214 /* All the bits are present. Drop. */
216 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
219 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
223 * continue traversing
226 /* skb's seq == skb1's seq and skb covers skb1.
227 * Replace skb1 with skb.
229 rb_replace_node(&skb1->rbnode, &skb->rbnode,
230 &msk->out_of_order_queue);
231 mptcp_drop(sk, skb1);
232 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
235 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
236 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
239 p = &parent->rb_right;
243 /* Insert segment into RB tree. */
244 rb_link_node(&skb->rbnode, parent, p);
245 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
248 /* Remove other segments covered by skb. */
249 while ((skb1 = skb_rb_next(skb)) != NULL) {
250 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
252 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
253 mptcp_drop(sk, skb1);
254 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
256 /* If there is no skb after us, we are the last_skb ! */
258 msk->ooo_last_skb = skb;
262 skb_set_owner_r(skb, sk);
265 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
266 struct sk_buff *skb, unsigned int offset,
269 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
270 struct sock *sk = (struct sock *)msk;
271 struct sk_buff *tail;
273 __skb_unlink(skb, &ssk->sk_receive_queue);
278 /* try to fetch required memory from subflow */
279 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
280 if (ssk->sk_forward_alloc < skb->truesize)
282 __sk_mem_reclaim(ssk, skb->truesize);
283 if (!sk_rmem_schedule(sk, skb, skb->truesize))
287 /* the skb map_seq accounts for the skb offset:
288 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
291 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
292 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
293 MPTCP_SKB_CB(skb)->offset = offset;
295 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
297 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
298 tail = skb_peek_tail(&sk->sk_receive_queue);
299 if (tail && mptcp_try_coalesce(sk, tail, skb))
302 skb_set_owner_r(skb, sk);
303 __skb_queue_tail(&sk->sk_receive_queue, skb);
305 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
306 mptcp_data_queue_ofo(msk, skb);
310 /* old data, keep it simple and drop the whole pkt, sender
311 * will retransmit as needed, if needed.
313 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
319 static void mptcp_stop_timer(struct sock *sk)
321 struct inet_connection_sock *icsk = inet_csk(sk);
323 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
324 mptcp_sk(sk)->timer_ival = 0;
327 static void mptcp_close_wake_up(struct sock *sk)
329 if (sock_flag(sk, SOCK_DEAD))
332 sk->sk_state_change(sk);
333 if (sk->sk_shutdown == SHUTDOWN_MASK ||
334 sk->sk_state == TCP_CLOSE)
335 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
337 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
340 static bool mptcp_pending_data_fin_ack(struct sock *sk)
342 struct mptcp_sock *msk = mptcp_sk(sk);
344 return !__mptcp_check_fallback(msk) &&
345 ((1 << sk->sk_state) &
346 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
347 msk->write_seq == READ_ONCE(msk->snd_una);
350 static void mptcp_check_data_fin_ack(struct sock *sk)
352 struct mptcp_sock *msk = mptcp_sk(sk);
354 /* Look for an acknowledged DATA_FIN */
355 if (mptcp_pending_data_fin_ack(sk)) {
356 WRITE_ONCE(msk->snd_data_fin_enable, 0);
358 switch (sk->sk_state) {
360 inet_sk_state_store(sk, TCP_FIN_WAIT2);
364 inet_sk_state_store(sk, TCP_CLOSE);
368 mptcp_close_wake_up(sk);
372 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
374 struct mptcp_sock *msk = mptcp_sk(sk);
376 if (READ_ONCE(msk->rcv_data_fin) &&
377 ((1 << sk->sk_state) &
378 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
379 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
381 if (msk->ack_seq == rcv_data_fin_seq) {
383 *seq = rcv_data_fin_seq;
392 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
394 long tout = ssk && inet_csk(ssk)->icsk_pending ?
395 inet_csk(ssk)->icsk_timeout - jiffies : 0;
398 tout = mptcp_sk(sk)->timer_ival;
399 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
402 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
404 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
406 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
407 if (subflow->request_join && !subflow->fully_established)
410 /* only send if our side has not closed yet */
411 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
414 static bool tcp_can_send_ack(const struct sock *ssk)
416 return !((1 << inet_sk_state_load(ssk)) &
417 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
420 static void mptcp_send_ack(struct mptcp_sock *msk)
422 struct mptcp_subflow_context *subflow;
424 mptcp_for_each_subflow(msk, subflow) {
425 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
428 if (tcp_can_send_ack(ssk))
434 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
439 ret = tcp_can_send_ack(ssk);
441 tcp_cleanup_rbuf(ssk, 1);
446 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
448 struct sock *ack_hint = READ_ONCE(msk->ack_hint);
449 int old_space = READ_ONCE(msk->old_wspace);
450 struct mptcp_subflow_context *subflow;
451 struct sock *sk = (struct sock *)msk;
454 /* this is a simple superset of what tcp_cleanup_rbuf() implements
455 * so that we don't have to acquire the ssk socket lock most of the time
456 * to do actually nothing
458 cleanup = __mptcp_space(sk) - old_space >= max(0, old_space);
462 /* if the hinted ssk is still active, try to use it */
463 if (likely(ack_hint)) {
464 mptcp_for_each_subflow(msk, subflow) {
465 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
467 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk))
472 /* otherwise pick the first active subflow */
473 mptcp_for_each_subflow(msk, subflow)
474 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
478 static bool mptcp_check_data_fin(struct sock *sk)
480 struct mptcp_sock *msk = mptcp_sk(sk);
481 u64 rcv_data_fin_seq;
484 if (__mptcp_check_fallback(msk))
487 /* Need to ack a DATA_FIN received from a peer while this side
488 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
489 * msk->rcv_data_fin was set when parsing the incoming options
490 * at the subflow level and the msk lock was not held, so this
491 * is the first opportunity to act on the DATA_FIN and change
494 * If we are caught up to the sequence number of the incoming
495 * DATA_FIN, send the DATA_ACK now and do state transition. If
496 * not caught up, do nothing and let the recv code send DATA_ACK
500 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
501 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
502 WRITE_ONCE(msk->rcv_data_fin, 0);
504 sk->sk_shutdown |= RCV_SHUTDOWN;
505 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
506 set_bit(MPTCP_DATA_READY, &msk->flags);
508 switch (sk->sk_state) {
509 case TCP_ESTABLISHED:
510 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
513 inet_sk_state_store(sk, TCP_CLOSING);
516 inet_sk_state_store(sk, TCP_CLOSE);
519 /* Other states not expected */
525 mptcp_set_timeout(sk, NULL);
527 mptcp_close_wake_up(sk);
532 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
536 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
537 struct sock *sk = (struct sock *)msk;
538 unsigned int moved = 0;
539 bool more_data_avail;
544 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
546 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
547 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
549 if (unlikely(ssk_rbuf > sk_rbuf)) {
550 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
555 pr_debug("msk=%p ssk=%p", msk, ssk);
558 u32 map_remaining, offset;
559 u32 seq = tp->copied_seq;
563 /* try to move as much data as available */
564 map_remaining = subflow->map_data_len -
565 mptcp_subflow_get_map_offset(subflow);
567 skb = skb_peek(&ssk->sk_receive_queue);
569 /* if no data is found, a racing workqueue/recvmsg
570 * already processed the new data, stop here or we
571 * can enter an infinite loop
578 if (__mptcp_check_fallback(msk)) {
579 /* if we are running under the workqueue, TCP could have
580 * collapsed skbs between dummy map creation and now
581 * be sure to adjust the size
583 map_remaining = skb->len;
584 subflow->map_data_len = skb->len;
587 offset = seq - TCP_SKB_CB(skb)->seq;
588 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
594 if (offset < skb->len) {
595 size_t len = skb->len - offset;
600 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
604 if (WARN_ON_ONCE(map_remaining < len))
608 sk_eat_skb(ssk, skb);
612 WRITE_ONCE(tp->copied_seq, seq);
613 more_data_avail = mptcp_subflow_data_available(ssk);
615 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
619 } while (more_data_avail);
620 WRITE_ONCE(msk->ack_hint, ssk);
626 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
628 struct sock *sk = (struct sock *)msk;
629 struct sk_buff *skb, *tail;
634 p = rb_first(&msk->out_of_order_queue);
635 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
638 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
642 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
644 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
647 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
651 end_seq = MPTCP_SKB_CB(skb)->end_seq;
652 tail = skb_peek_tail(&sk->sk_receive_queue);
653 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
654 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
656 /* skip overlapping data, if any */
657 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
658 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
660 MPTCP_SKB_CB(skb)->offset += delta;
661 __skb_queue_tail(&sk->sk_receive_queue, skb);
663 msk->ack_seq = end_seq;
669 /* In most cases we will be able to lock the mptcp socket. If its already
670 * owned, we need to defer to the work queue to avoid ABBA deadlock.
672 static void move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
674 struct sock *sk = (struct sock *)msk;
675 unsigned int moved = 0;
677 if (inet_sk_state_load(sk) == TCP_CLOSE)
682 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
683 __mptcp_ofo_queue(msk);
685 /* If the moves have caught up with the DATA_FIN sequence number
686 * it's time to ack the DATA_FIN and change socket state, but
687 * this is not a good place to change state. Let the workqueue
690 if (mptcp_pending_data_fin(sk, NULL))
691 mptcp_schedule_work(sk);
692 mptcp_data_unlock(sk);
695 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
697 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
698 struct mptcp_sock *msk = mptcp_sk(sk);
699 int sk_rbuf, ssk_rbuf;
702 /* The peer can send data while we are shutting down this
703 * subflow at msk destruction time, but we must avoid enqueuing
704 * more data to the msk receive queue
706 if (unlikely(subflow->disposable))
709 /* move_skbs_to_msk below can legitly clear the data_avail flag,
710 * but we will need later to properly woke the reader, cache its
713 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
715 set_bit(MPTCP_DATA_READY, &msk->flags);
717 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
718 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
719 if (unlikely(ssk_rbuf > sk_rbuf))
722 /* over limit? can't append more skbs to msk */
723 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
726 move_skbs_to_msk(msk, ssk);
730 sk->sk_data_ready(sk);
733 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
735 struct mptcp_subflow_context *subflow;
738 if (likely(list_empty(&msk->join_list)))
741 spin_lock_bh(&msk->join_list_lock);
742 list_for_each_entry(subflow, &msk->join_list, node) {
743 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
745 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
746 if (READ_ONCE(msk->setsockopt_seq) != sseq)
749 list_splice_tail_init(&msk->join_list, &msk->conn_list);
750 spin_unlock_bh(&msk->join_list_lock);
755 void __mptcp_flush_join_list(struct mptcp_sock *msk)
757 if (likely(!mptcp_do_flush_join_list(msk)))
760 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
761 mptcp_schedule_work((struct sock *)msk);
764 static void mptcp_flush_join_list(struct mptcp_sock *msk)
766 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
770 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
773 mptcp_sockopt_sync_all(msk);
776 static bool mptcp_timer_pending(struct sock *sk)
778 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
781 static void mptcp_reset_timer(struct sock *sk)
783 struct inet_connection_sock *icsk = inet_csk(sk);
786 /* prevent rescheduling on close */
787 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
790 /* should never be called with mptcp level timer cleared */
791 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
792 if (WARN_ON_ONCE(!tout))
794 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
797 bool mptcp_schedule_work(struct sock *sk)
799 if (inet_sk_state_load(sk) != TCP_CLOSE &&
800 schedule_work(&mptcp_sk(sk)->work)) {
801 /* each subflow already holds a reference to the sk, and the
802 * workqueue is invoked by a subflow, so sk can't go away here.
810 void mptcp_subflow_eof(struct sock *sk)
812 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
813 mptcp_schedule_work(sk);
816 static void mptcp_check_for_eof(struct mptcp_sock *msk)
818 struct mptcp_subflow_context *subflow;
819 struct sock *sk = (struct sock *)msk;
822 mptcp_for_each_subflow(msk, subflow)
823 receivers += !subflow->rx_eof;
827 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
828 /* hopefully temporary hack: propagate shutdown status
829 * to msk, when all subflows agree on it
831 sk->sk_shutdown |= RCV_SHUTDOWN;
833 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
834 set_bit(MPTCP_DATA_READY, &msk->flags);
835 sk->sk_data_ready(sk);
838 switch (sk->sk_state) {
839 case TCP_ESTABLISHED:
840 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
843 inet_sk_state_store(sk, TCP_CLOSING);
846 inet_sk_state_store(sk, TCP_CLOSE);
851 mptcp_close_wake_up(sk);
854 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
856 struct mptcp_subflow_context *subflow;
857 struct sock *sk = (struct sock *)msk;
859 sock_owned_by_me(sk);
861 mptcp_for_each_subflow(msk, subflow) {
862 if (subflow->data_avail)
863 return mptcp_subflow_tcp_sock(subflow);
869 static bool mptcp_skb_can_collapse_to(u64 write_seq,
870 const struct sk_buff *skb,
871 const struct mptcp_ext *mpext)
873 if (!tcp_skb_can_collapse_to(skb))
876 /* can collapse only if MPTCP level sequence is in order and this
877 * mapping has not been xmitted yet
879 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
883 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
884 const struct page_frag *pfrag,
885 const struct mptcp_data_frag *df)
887 return df && pfrag->page == df->page &&
888 pfrag->size - pfrag->offset > 0 &&
889 df->data_seq + df->data_len == msk->write_seq;
892 static int mptcp_wmem_with_overhead(struct sock *sk, int size)
894 struct mptcp_sock *msk = mptcp_sk(sk);
897 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
898 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache;
899 if (skbs < msk->skb_tx_cache.qlen)
902 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER);
905 static void __mptcp_wmem_reserve(struct sock *sk, int size)
907 int amount = mptcp_wmem_with_overhead(sk, size);
908 struct mptcp_sock *msk = mptcp_sk(sk);
910 WARN_ON_ONCE(msk->wmem_reserved);
911 if (WARN_ON_ONCE(amount < 0))
914 if (amount <= sk->sk_forward_alloc)
917 /* under memory pressure try to reserve at most a single page
918 * otherwise try to reserve the full estimate and fallback
919 * to a single page before entering the error path
921 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
922 !sk_wmem_schedule(sk, amount)) {
923 if (amount <= PAGE_SIZE)
927 if (!sk_wmem_schedule(sk, amount))
932 msk->wmem_reserved = amount;
933 sk->sk_forward_alloc -= amount;
937 /* we will wait for memory on next allocation */
938 msk->wmem_reserved = -1;
941 static void __mptcp_update_wmem(struct sock *sk)
943 struct mptcp_sock *msk = mptcp_sk(sk);
945 if (!msk->wmem_reserved)
948 if (msk->wmem_reserved < 0)
949 msk->wmem_reserved = 0;
950 if (msk->wmem_reserved > 0) {
951 sk->sk_forward_alloc += msk->wmem_reserved;
952 msk->wmem_reserved = 0;
956 static bool mptcp_wmem_alloc(struct sock *sk, int size)
958 struct mptcp_sock *msk = mptcp_sk(sk);
960 /* check for pre-existing error condition */
961 if (msk->wmem_reserved < 0)
964 if (msk->wmem_reserved >= size)
968 if (!sk_wmem_schedule(sk, size)) {
969 mptcp_data_unlock(sk);
973 sk->sk_forward_alloc -= size;
974 msk->wmem_reserved += size;
975 mptcp_data_unlock(sk);
978 msk->wmem_reserved -= size;
982 static void mptcp_wmem_uncharge(struct sock *sk, int size)
984 struct mptcp_sock *msk = mptcp_sk(sk);
986 if (msk->wmem_reserved < 0)
987 msk->wmem_reserved = 0;
988 msk->wmem_reserved += size;
991 static void mptcp_mem_reclaim_partial(struct sock *sk)
993 struct mptcp_sock *msk = mptcp_sk(sk);
995 /* if we are experiencing a transint allocation error,
996 * the forward allocation memory has been already
999 if (msk->wmem_reserved < 0)
1002 mptcp_data_lock(sk);
1003 sk->sk_forward_alloc += msk->wmem_reserved;
1004 sk_mem_reclaim_partial(sk);
1005 msk->wmem_reserved = sk->sk_forward_alloc;
1006 sk->sk_forward_alloc = 0;
1007 mptcp_data_unlock(sk);
1010 static void dfrag_uncharge(struct sock *sk, int len)
1012 sk_mem_uncharge(sk, len);
1013 sk_wmem_queued_add(sk, -len);
1016 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1018 int len = dfrag->data_len + dfrag->overhead;
1020 list_del(&dfrag->list);
1021 dfrag_uncharge(sk, len);
1022 put_page(dfrag->page);
1025 static void __mptcp_clean_una(struct sock *sk)
1027 struct mptcp_sock *msk = mptcp_sk(sk);
1028 struct mptcp_data_frag *dtmp, *dfrag;
1029 bool cleaned = false;
1032 /* on fallback we just need to ignore snd_una, as this is really
1035 if (__mptcp_check_fallback(msk))
1036 msk->snd_una = READ_ONCE(msk->snd_nxt);
1038 snd_una = msk->snd_una;
1039 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1040 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1043 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1045 dfrag_clear(sk, dfrag);
1049 dfrag = mptcp_rtx_head(sk);
1050 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1051 u64 delta = snd_una - dfrag->data_seq;
1053 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1056 dfrag->data_seq += delta;
1057 dfrag->offset += delta;
1058 dfrag->data_len -= delta;
1059 dfrag->already_sent -= delta;
1061 dfrag_uncharge(sk, delta);
1067 if (tcp_under_memory_pressure(sk)) {
1068 __mptcp_update_wmem(sk);
1069 sk_mem_reclaim_partial(sk);
1073 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1074 if (msk->timer_ival)
1075 mptcp_stop_timer(sk);
1077 mptcp_reset_timer(sk);
1081 static void __mptcp_clean_una_wakeup(struct sock *sk)
1083 __mptcp_clean_una(sk);
1084 mptcp_write_space(sk);
1087 static void mptcp_enter_memory_pressure(struct sock *sk)
1089 struct mptcp_subflow_context *subflow;
1090 struct mptcp_sock *msk = mptcp_sk(sk);
1093 sk_stream_moderate_sndbuf(sk);
1094 mptcp_for_each_subflow(msk, subflow) {
1095 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1098 tcp_enter_memory_pressure(ssk);
1099 sk_stream_moderate_sndbuf(ssk);
1104 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1107 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1109 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1110 pfrag, sk->sk_allocation)))
1113 mptcp_enter_memory_pressure(sk);
1117 static struct mptcp_data_frag *
1118 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1121 int offset = ALIGN(orig_offset, sizeof(long));
1122 struct mptcp_data_frag *dfrag;
1124 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1125 dfrag->data_len = 0;
1126 dfrag->data_seq = msk->write_seq;
1127 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1128 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1129 dfrag->already_sent = 0;
1130 dfrag->page = pfrag->page;
1135 struct mptcp_sendmsg_info {
1143 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1146 u64 window_end = mptcp_wnd_end(msk);
1148 if (__mptcp_check_fallback(msk))
1151 if (!before64(data_seq + avail_size, window_end)) {
1152 u64 allowed_size = window_end - data_seq;
1154 return min_t(unsigned int, allowed_size, avail_size);
1160 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1162 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1166 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1170 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1172 struct sk_buff *skb;
1174 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1176 if (likely(__mptcp_add_ext(skb, gfp))) {
1177 skb_reserve(skb, MAX_TCP_HEADER);
1178 skb->reserved_tailroom = skb->end - skb->tail;
1183 mptcp_enter_memory_pressure(sk);
1188 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1189 struct sk_buff_head *skbs, int *total_ts)
1191 struct mptcp_sock *msk = mptcp_sk(sk);
1192 struct sk_buff *skb;
1195 if (unlikely(tcp_under_memory_pressure(sk))) {
1196 mptcp_mem_reclaim_partial(sk);
1198 /* under pressure pre-allocate at most a single skb */
1199 if (msk->skb_tx_cache.qlen)
1201 space_needed = msk->size_goal_cache;
1203 space_needed = msk->tx_pending_data + size -
1204 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1207 while (space_needed > 0) {
1208 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1209 if (unlikely(!skb)) {
1210 /* under memory pressure, try to pass the caller a
1211 * single skb to allow forward progress
1213 while (skbs->qlen > 1) {
1214 skb = __skb_dequeue_tail(skbs);
1215 *total_ts -= skb->truesize;
1218 return skbs->qlen > 0;
1221 *total_ts += skb->truesize;
1222 __skb_queue_tail(skbs, skb);
1223 space_needed -= msk->size_goal_cache;
1228 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1230 struct mptcp_sock *msk = mptcp_sk(sk);
1231 struct sk_buff *skb;
1233 if (ssk->sk_tx_skb_cache) {
1234 skb = ssk->sk_tx_skb_cache;
1235 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1236 !__mptcp_add_ext(skb, gfp)))
1241 skb = skb_peek(&msk->skb_tx_cache);
1243 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1244 skb = __skb_dequeue(&msk->skb_tx_cache);
1245 if (WARN_ON_ONCE(!skb))
1248 mptcp_wmem_uncharge(sk, skb->truesize);
1249 ssk->sk_tx_skb_cache = skb;
1253 /* over memory limit, no point to try to allocate a new skb */
1257 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1261 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1262 ssk->sk_tx_skb_cache = skb;
1269 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1271 return !ssk->sk_tx_skb_cache &&
1272 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1273 tcp_under_memory_pressure(sk);
1276 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1278 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1279 mptcp_mem_reclaim_partial(sk);
1280 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1283 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1284 struct mptcp_data_frag *dfrag,
1285 struct mptcp_sendmsg_info *info)
1287 u64 data_seq = dfrag->data_seq + info->sent;
1288 struct mptcp_sock *msk = mptcp_sk(sk);
1289 bool zero_window_probe = false;
1290 struct mptcp_ext *mpext = NULL;
1291 struct sk_buff *skb, *tail;
1292 bool can_collapse = false;
1297 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
1298 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1300 /* compute send limit */
1301 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1302 avail_size = info->size_goal;
1303 msk->size_goal_cache = info->size_goal;
1304 skb = tcp_write_queue_tail(ssk);
1306 /* Limit the write to the size available in the
1307 * current skb, if any, so that we create at most a new skb.
1308 * Explicitly tells TCP internals to avoid collapsing on later
1309 * queue management operation, to avoid breaking the ext <->
1310 * SSN association set here
1312 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1313 can_collapse = (info->size_goal - skb->len > 0) &&
1314 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1315 if (!can_collapse) {
1316 TCP_SKB_CB(skb)->eor = 1;
1318 size_bias = skb->len;
1319 avail_size = info->size_goal - skb->len;
1323 /* Zero window and all data acked? Probe. */
1324 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1325 if (avail_size == 0) {
1326 u64 snd_una = READ_ONCE(msk->snd_una);
1328 if (skb || snd_una != msk->snd_nxt)
1330 zero_window_probe = true;
1331 data_seq = snd_una - 1;
1335 if (WARN_ON_ONCE(info->sent > info->limit ||
1336 info->limit > dfrag->data_len))
1339 ret = info->limit - info->sent;
1340 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1341 dfrag->page, dfrag->offset + info->sent, &ret);
1343 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1347 /* if the tail skb is still the cached one, collapsing really happened.
1350 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1351 mpext->data_len += ret;
1352 WARN_ON_ONCE(!can_collapse);
1353 WARN_ON_ONCE(zero_window_probe);
1357 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1358 if (WARN_ON_ONCE(!mpext)) {
1359 /* should never reach here, stream corrupted */
1363 memset(mpext, 0, sizeof(*mpext));
1364 mpext->data_seq = data_seq;
1365 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1366 mpext->data_len = ret;
1370 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1371 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1374 if (zero_window_probe) {
1375 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1378 tcp_push_pending_frames(ssk);
1381 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1385 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1386 sizeof(struct tcphdr) - \
1387 MAX_TCP_OPTION_SPACE - \
1388 sizeof(struct ipv6hdr) - \
1389 sizeof(struct frag_hdr))
1391 struct subflow_send_info {
1396 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1398 struct subflow_send_info send_info[2];
1399 struct mptcp_subflow_context *subflow;
1400 int i, nr_active = 0;
1405 sock_owned_by_me((struct sock *)msk);
1407 if (__mptcp_check_fallback(msk)) {
1410 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1413 /* re-use last subflow, if the burst allow that */
1414 if (msk->last_snd && msk->snd_burst > 0 &&
1415 sk_stream_memory_free(msk->last_snd) &&
1416 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1417 return msk->last_snd;
1419 /* pick the subflow with the lower wmem/wspace ratio */
1420 for (i = 0; i < 2; ++i) {
1421 send_info[i].ssk = NULL;
1422 send_info[i].ratio = -1;
1424 mptcp_for_each_subflow(msk, subflow) {
1425 ssk = mptcp_subflow_tcp_sock(subflow);
1426 if (!mptcp_subflow_active(subflow))
1429 nr_active += !subflow->backup;
1430 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1433 pace = READ_ONCE(ssk->sk_pacing_rate);
1437 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1439 if (ratio < send_info[subflow->backup].ratio) {
1440 send_info[subflow->backup].ssk = ssk;
1441 send_info[subflow->backup].ratio = ratio;
1445 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1446 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1447 send_info[1].ssk, send_info[1].ratio);
1449 /* pick the best backup if no other subflow is active */
1451 send_info[0].ssk = send_info[1].ssk;
1453 if (send_info[0].ssk) {
1454 msk->last_snd = send_info[0].ssk;
1455 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1456 tcp_sk(msk->last_snd)->snd_wnd);
1457 return msk->last_snd;
1463 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1464 struct mptcp_sendmsg_info *info)
1466 mptcp_set_timeout(sk, ssk);
1467 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1471 static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1473 struct sock *prev_ssk = NULL, *ssk = NULL;
1474 struct mptcp_sock *msk = mptcp_sk(sk);
1475 struct mptcp_sendmsg_info info = {
1478 struct mptcp_data_frag *dfrag;
1479 int len, copied = 0;
1481 while ((dfrag = mptcp_send_head(sk))) {
1482 info.sent = dfrag->already_sent;
1483 info.limit = dfrag->data_len;
1484 len = dfrag->data_len - dfrag->already_sent;
1489 mptcp_flush_join_list(msk);
1490 ssk = mptcp_subflow_get_send(msk);
1492 /* try to keep the subflow socket lock across
1493 * consecutive xmit on the same socket
1495 if (ssk != prev_ssk && prev_ssk)
1496 mptcp_push_release(sk, prev_ssk, &info);
1500 if (ssk != prev_ssk || !prev_ssk)
1503 /* keep it simple and always provide a new skb for the
1504 * subflow, even if we will not use it when collapsing
1505 * on the pending one
1507 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1508 mptcp_push_release(sk, ssk, &info);
1512 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1514 mptcp_push_release(sk, ssk, &info);
1519 dfrag->already_sent += ret;
1520 msk->snd_nxt += ret;
1521 msk->snd_burst -= ret;
1522 msk->tx_pending_data -= ret;
1526 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1529 /* at this point we held the socket lock for the last subflow we used */
1531 mptcp_push_release(sk, ssk, &info);
1535 /* start the timer, if it's not pending */
1536 if (!mptcp_timer_pending(sk))
1537 mptcp_reset_timer(sk);
1538 __mptcp_check_send_data_fin(sk);
1542 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1544 struct mptcp_sock *msk = mptcp_sk(sk);
1545 struct mptcp_sendmsg_info info;
1546 struct mptcp_data_frag *dfrag;
1547 struct sock *xmit_ssk;
1548 int len, copied = 0;
1552 while ((dfrag = mptcp_send_head(sk))) {
1553 info.sent = dfrag->already_sent;
1554 info.limit = dfrag->data_len;
1555 len = dfrag->data_len - dfrag->already_sent;
1559 /* the caller already invoked the packet scheduler,
1560 * check for a different subflow usage only after
1561 * spooling the first chunk of data
1563 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1566 if (xmit_ssk != ssk) {
1567 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1571 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1572 __mptcp_update_wmem(sk);
1573 sk_mem_reclaim_partial(sk);
1575 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1578 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1583 dfrag->already_sent += ret;
1584 msk->snd_nxt += ret;
1585 msk->snd_burst -= ret;
1586 msk->tx_pending_data -= ret;
1591 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1595 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1596 * not going to flush it via release_sock()
1598 __mptcp_update_wmem(sk);
1600 mptcp_set_timeout(sk, ssk);
1601 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1603 if (!mptcp_timer_pending(sk))
1604 mptcp_reset_timer(sk);
1606 if (msk->snd_data_fin_enable &&
1607 msk->snd_nxt + 1 == msk->write_seq)
1608 mptcp_schedule_work(sk);
1612 static void mptcp_set_nospace(struct sock *sk)
1614 /* enable autotune */
1615 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1617 /* will be cleared on avail space */
1618 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1621 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1623 struct mptcp_sock *msk = mptcp_sk(sk);
1624 struct page_frag *pfrag;
1629 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1632 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1634 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1636 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1637 ret = sk_stream_wait_connect(sk, &timeo);
1642 pfrag = sk_page_frag(sk);
1644 while (msg_data_left(msg)) {
1645 int total_ts, frag_truesize = 0;
1646 struct mptcp_data_frag *dfrag;
1647 struct sk_buff_head skbs;
1648 bool dfrag_collapsed;
1649 size_t psize, offset;
1651 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1656 /* reuse tail pfrag, if possible, or carve a new one from the
1659 dfrag = mptcp_pending_tail(sk);
1660 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1661 if (!dfrag_collapsed) {
1662 if (!sk_stream_memory_free(sk))
1663 goto wait_for_memory;
1665 if (!mptcp_page_frag_refill(sk, pfrag))
1666 goto wait_for_memory;
1668 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1669 frag_truesize = dfrag->overhead;
1672 /* we do not bound vs wspace, to allow a single packet.
1673 * memory accounting will prevent execessive memory usage
1676 offset = dfrag->offset + dfrag->data_len;
1677 psize = pfrag->size - offset;
1678 psize = min_t(size_t, psize, msg_data_left(msg));
1679 total_ts = psize + frag_truesize;
1680 __skb_queue_head_init(&skbs);
1681 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1682 goto wait_for_memory;
1684 if (!mptcp_wmem_alloc(sk, total_ts)) {
1685 __skb_queue_purge(&skbs);
1686 goto wait_for_memory;
1689 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1690 if (copy_page_from_iter(dfrag->page, offset, psize,
1691 &msg->msg_iter) != psize) {
1692 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1697 /* data successfully copied into the write queue */
1699 dfrag->data_len += psize;
1700 frag_truesize += psize;
1701 pfrag->offset += frag_truesize;
1702 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1703 msk->tx_pending_data += psize;
1705 /* charge data on mptcp pending queue to the msk socket
1706 * Note: we charge such data both to sk and ssk
1708 sk_wmem_queued_add(sk, frag_truesize);
1709 if (!dfrag_collapsed) {
1710 get_page(dfrag->page);
1711 list_add_tail(&dfrag->list, &msk->rtx_queue);
1712 if (!msk->first_pending)
1713 WRITE_ONCE(msk->first_pending, dfrag);
1715 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1716 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1722 mptcp_set_nospace(sk);
1723 __mptcp_push_pending(sk, msg->msg_flags);
1724 ret = sk_stream_wait_memory(sk, &timeo);
1730 __mptcp_push_pending(sk, msg->msg_flags);
1734 return copied ? : ret;
1737 static void mptcp_wait_data(struct sock *sk, long *timeo)
1739 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1740 struct mptcp_sock *msk = mptcp_sk(sk);
1742 add_wait_queue(sk_sleep(sk), &wait);
1743 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1745 sk_wait_event(sk, timeo,
1746 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1748 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1749 remove_wait_queue(sk_sleep(sk), &wait);
1752 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1756 struct sk_buff *skb;
1759 while ((skb = skb_peek(&msk->receive_queue)) != NULL) {
1760 u32 offset = MPTCP_SKB_CB(skb)->offset;
1761 u32 data_len = skb->len - offset;
1762 u32 count = min_t(size_t, len - copied, data_len);
1765 err = skb_copy_datagram_msg(skb, offset, msg, count);
1766 if (unlikely(err < 0)) {
1774 if (count < data_len) {
1775 MPTCP_SKB_CB(skb)->offset += count;
1779 /* we will bulk release the skb memory later */
1780 skb->destructor = NULL;
1781 msk->rmem_released += skb->truesize;
1782 __skb_unlink(skb, &msk->receive_queue);
1792 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1794 * Only difference: Use highest rtt estimate of the subflows in use.
1796 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1798 struct mptcp_subflow_context *subflow;
1799 struct sock *sk = (struct sock *)msk;
1800 u32 time, advmss = 1;
1803 sock_owned_by_me(sk);
1808 msk->rcvq_space.copied += copied;
1810 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1811 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1813 rtt_us = msk->rcvq_space.rtt_us;
1814 if (rtt_us && time < (rtt_us >> 3))
1818 mptcp_for_each_subflow(msk, subflow) {
1819 const struct tcp_sock *tp;
1823 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1825 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1826 sf_advmss = READ_ONCE(tp->advmss);
1828 rtt_us = max(sf_rtt_us, rtt_us);
1829 advmss = max(sf_advmss, advmss);
1832 msk->rcvq_space.rtt_us = rtt_us;
1833 if (time < (rtt_us >> 3) || rtt_us == 0)
1836 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1839 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1840 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1844 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1846 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1848 do_div(grow, msk->rcvq_space.space);
1849 rcvwin += (grow << 1);
1851 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1852 while (tcp_win_from_space(sk, rcvmem) < advmss)
1855 do_div(rcvwin, advmss);
1856 rcvbuf = min_t(u64, rcvwin * rcvmem,
1857 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1859 if (rcvbuf > sk->sk_rcvbuf) {
1862 window_clamp = tcp_win_from_space(sk, rcvbuf);
1863 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1865 /* Make subflows follow along. If we do not do this, we
1866 * get drops at subflow level if skbs can't be moved to
1867 * the mptcp rx queue fast enough (announced rcv_win can
1868 * exceed ssk->sk_rcvbuf).
1870 mptcp_for_each_subflow(msk, subflow) {
1874 ssk = mptcp_subflow_tcp_sock(subflow);
1875 slow = lock_sock_fast(ssk);
1876 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1877 tcp_sk(ssk)->window_clamp = window_clamp;
1878 tcp_cleanup_rbuf(ssk, 1);
1879 unlock_sock_fast(ssk, slow);
1884 msk->rcvq_space.space = msk->rcvq_space.copied;
1886 msk->rcvq_space.copied = 0;
1887 msk->rcvq_space.time = mstamp;
1890 static void __mptcp_update_rmem(struct sock *sk)
1892 struct mptcp_sock *msk = mptcp_sk(sk);
1894 if (!msk->rmem_released)
1897 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1898 sk_mem_uncharge(sk, msk->rmem_released);
1899 msk->rmem_released = 0;
1902 static void __mptcp_splice_receive_queue(struct sock *sk)
1904 struct mptcp_sock *msk = mptcp_sk(sk);
1906 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1909 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1911 struct sock *sk = (struct sock *)msk;
1912 unsigned int moved = 0;
1915 mptcp_flush_join_list(msk);
1917 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1920 /* we can have data pending in the subflows only if the msk
1921 * receive buffer was full at subflow_data_ready() time,
1922 * that is an unlikely slow path.
1927 slowpath = lock_sock_fast(ssk);
1928 mptcp_data_lock(sk);
1929 __mptcp_update_rmem(sk);
1930 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1931 mptcp_data_unlock(sk);
1932 tcp_cleanup_rbuf(ssk, moved);
1933 unlock_sock_fast(ssk, slowpath);
1936 /* acquire the data lock only if some input data is pending */
1938 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1939 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1940 mptcp_data_lock(sk);
1941 __mptcp_update_rmem(sk);
1942 ret |= __mptcp_ofo_queue(msk);
1943 __mptcp_splice_receive_queue(sk);
1944 mptcp_data_unlock(sk);
1945 mptcp_cleanup_rbuf(msk);
1948 mptcp_check_data_fin((struct sock *)msk);
1949 return !skb_queue_empty(&msk->receive_queue);
1952 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1953 int nonblock, int flags, int *addr_len)
1955 struct mptcp_sock *msk = mptcp_sk(sk);
1960 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1963 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1964 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1969 timeo = sock_rcvtimeo(sk, nonblock);
1971 len = min_t(size_t, len, INT_MAX);
1972 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1974 while (copied < len) {
1977 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1978 if (unlikely(bytes_read < 0)) {
1980 copied = bytes_read;
1984 copied += bytes_read;
1986 /* be sure to advertise window change */
1987 mptcp_cleanup_rbuf(msk);
1989 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
1992 /* only the master socket status is relevant here. The exit
1993 * conditions mirror closely tcp_recvmsg()
1995 if (copied >= target)
2000 sk->sk_state == TCP_CLOSE ||
2001 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2003 signal_pending(current))
2007 copied = sock_error(sk);
2011 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2012 mptcp_check_for_eof(msk);
2014 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2015 /* race breaker: the shutdown could be after the
2016 * previous receive queue check
2018 if (__mptcp_move_skbs(msk))
2023 if (sk->sk_state == TCP_CLOSE) {
2033 if (signal_pending(current)) {
2034 copied = sock_intr_errno(timeo);
2039 pr_debug("block timeout %ld", timeo);
2040 mptcp_wait_data(sk, &timeo);
2043 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2044 skb_queue_empty(&msk->receive_queue)) {
2045 /* entire backlog drained, clear DATA_READY. */
2046 clear_bit(MPTCP_DATA_READY, &msk->flags);
2048 /* .. race-breaker: ssk might have gotten new data
2049 * after last __mptcp_move_skbs() returned false.
2051 if (unlikely(__mptcp_move_skbs(msk)))
2052 set_bit(MPTCP_DATA_READY, &msk->flags);
2053 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2054 /* data to read but mptcp_wait_data() cleared DATA_READY */
2055 set_bit(MPTCP_DATA_READY, &msk->flags);
2058 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2059 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2060 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2061 mptcp_rcv_space_adjust(msk, copied);
2067 static void mptcp_retransmit_timer(struct timer_list *t)
2069 struct inet_connection_sock *icsk = from_timer(icsk, t,
2070 icsk_retransmit_timer);
2071 struct sock *sk = &icsk->icsk_inet.sk;
2072 struct mptcp_sock *msk = mptcp_sk(sk);
2075 if (!sock_owned_by_user(sk)) {
2076 /* we need a process context to retransmit */
2077 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2078 mptcp_schedule_work(sk);
2080 /* delegate our work to tcp_release_cb() */
2081 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2087 static void mptcp_timeout_timer(struct timer_list *t)
2089 struct sock *sk = from_timer(sk, t, sk_timer);
2091 mptcp_schedule_work(sk);
2095 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2098 * A backup subflow is returned only if that is the only kind available.
2100 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2102 struct mptcp_subflow_context *subflow;
2103 struct sock *backup = NULL;
2105 sock_owned_by_me((const struct sock *)msk);
2107 if (__mptcp_check_fallback(msk))
2110 mptcp_for_each_subflow(msk, subflow) {
2111 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2113 if (!mptcp_subflow_active(subflow))
2116 /* still data outstanding at TCP level? Don't retransmit. */
2117 if (!tcp_write_queue_empty(ssk)) {
2118 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2123 if (subflow->backup) {
2135 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2138 iput(SOCK_INODE(msk->subflow));
2139 msk->subflow = NULL;
2143 /* subflow sockets can be either outgoing (connect) or incoming
2146 * Outgoing subflows use in-kernel sockets.
2147 * Incoming subflows do not have their own 'struct socket' allocated,
2148 * so we need to use tcp_close() after detaching them from the mptcp
2151 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2152 struct mptcp_subflow_context *subflow)
2154 struct mptcp_sock *msk = mptcp_sk(sk);
2156 list_del(&subflow->node);
2158 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2160 /* if we are invoked by the msk cleanup code, the subflow is
2166 subflow->disposable = 1;
2168 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2169 * the ssk has been already destroyed, we just need to release the
2170 * reference owned by msk;
2172 if (!inet_csk(ssk)->icsk_ulp_ops) {
2173 kfree_rcu(subflow, rcu);
2175 /* otherwise tcp will dispose of the ssk and subflow ctx */
2176 __tcp_close(ssk, 0);
2178 /* close acquired an extra ref */
2185 if (ssk == msk->last_snd)
2186 msk->last_snd = NULL;
2188 if (ssk == msk->ack_hint)
2189 msk->ack_hint = NULL;
2191 if (ssk == msk->first)
2194 if (msk->subflow && ssk == msk->subflow->sk)
2195 mptcp_dispose_initial_subflow(msk);
2198 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2199 struct mptcp_subflow_context *subflow)
2201 if (sk->sk_state == TCP_ESTABLISHED)
2202 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2203 __mptcp_close_ssk(sk, ssk, subflow);
2206 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2211 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2213 struct mptcp_subflow_context *subflow, *tmp;
2217 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2218 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2220 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2223 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2224 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2227 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2231 static bool mptcp_check_close_timeout(const struct sock *sk)
2233 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2234 struct mptcp_subflow_context *subflow;
2236 if (delta >= TCP_TIMEWAIT_LEN)
2239 /* if all subflows are in closed status don't bother with additional
2242 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2243 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2250 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2252 struct mptcp_subflow_context *subflow, *tmp;
2253 struct sock *sk = &msk->sk.icsk_inet.sk;
2255 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2258 mptcp_token_destroy(msk);
2260 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2261 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2264 if (tcp_sk->sk_state != TCP_CLOSE) {
2265 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2266 tcp_set_state(tcp_sk, TCP_CLOSE);
2268 release_sock(tcp_sk);
2271 inet_sk_state_store(sk, TCP_CLOSE);
2272 sk->sk_shutdown = SHUTDOWN_MASK;
2273 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2274 set_bit(MPTCP_DATA_READY, &msk->flags);
2275 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2277 mptcp_close_wake_up(sk);
2280 static void __mptcp_retrans(struct sock *sk)
2282 struct mptcp_sock *msk = mptcp_sk(sk);
2283 struct mptcp_sendmsg_info info = {};
2284 struct mptcp_data_frag *dfrag;
2289 __mptcp_clean_una_wakeup(sk);
2290 dfrag = mptcp_rtx_head(sk);
2294 ssk = mptcp_subflow_get_retrans(msk);
2300 /* limit retransmission to the bytes already sent on some subflows */
2302 info.limit = dfrag->already_sent;
2303 while (info.sent < dfrag->already_sent) {
2304 if (!mptcp_alloc_tx_skb(sk, ssk))
2307 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2311 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2316 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2319 mptcp_set_timeout(sk, ssk);
2323 if (!mptcp_timer_pending(sk))
2324 mptcp_reset_timer(sk);
2327 static void mptcp_worker(struct work_struct *work)
2329 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2330 struct sock *sk = &msk->sk.icsk_inet.sk;
2334 state = sk->sk_state;
2335 if (unlikely(state == TCP_CLOSE))
2338 mptcp_check_data_fin_ack(sk);
2339 mptcp_flush_join_list(msk);
2341 mptcp_check_fastclose(msk);
2344 mptcp_pm_nl_work(msk);
2346 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2347 mptcp_check_for_eof(msk);
2349 __mptcp_check_send_data_fin(sk);
2350 mptcp_check_data_fin(sk);
2352 /* There is no point in keeping around an orphaned sk timedout or
2353 * closed, but we need the msk around to reply to incoming DATA_FIN,
2354 * even if it is orphaned and in FIN_WAIT2 state
2356 if (sock_flag(sk, SOCK_DEAD) &&
2357 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2358 inet_sk_state_store(sk, TCP_CLOSE);
2359 __mptcp_destroy_sock(sk);
2363 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2364 __mptcp_close_subflow(msk);
2366 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2367 __mptcp_retrans(sk);
2374 static int __mptcp_init_sock(struct sock *sk)
2376 struct mptcp_sock *msk = mptcp_sk(sk);
2378 spin_lock_init(&msk->join_list_lock);
2380 INIT_LIST_HEAD(&msk->conn_list);
2381 INIT_LIST_HEAD(&msk->join_list);
2382 INIT_LIST_HEAD(&msk->rtx_queue);
2383 INIT_WORK(&msk->work, mptcp_worker);
2384 __skb_queue_head_init(&msk->receive_queue);
2385 __skb_queue_head_init(&msk->skb_tx_cache);
2386 msk->out_of_order_queue = RB_ROOT;
2387 msk->first_pending = NULL;
2388 msk->wmem_reserved = 0;
2389 msk->rmem_released = 0;
2390 msk->tx_pending_data = 0;
2391 msk->size_goal_cache = TCP_BASE_MSS;
2393 msk->ack_hint = NULL;
2395 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2397 mptcp_pm_data_init(msk);
2399 /* re-use the csk retrans timer for MPTCP-level retrans */
2400 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2401 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2403 tcp_assign_congestion_control(sk);
2408 static int mptcp_init_sock(struct sock *sk)
2410 struct net *net = sock_net(sk);
2413 ret = __mptcp_init_sock(sk);
2417 if (!mptcp_is_enabled(net))
2418 return -ENOPROTOOPT;
2420 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2423 ret = __mptcp_socket_create(mptcp_sk(sk));
2427 sk_sockets_allocated_inc(sk);
2428 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2429 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2434 static void __mptcp_clear_xmit(struct sock *sk)
2436 struct mptcp_sock *msk = mptcp_sk(sk);
2437 struct mptcp_data_frag *dtmp, *dfrag;
2438 struct sk_buff *skb;
2440 WRITE_ONCE(msk->first_pending, NULL);
2441 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2442 dfrag_clear(sk, dfrag);
2443 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2444 sk->sk_forward_alloc += skb->truesize;
2449 static void mptcp_cancel_work(struct sock *sk)
2451 struct mptcp_sock *msk = mptcp_sk(sk);
2453 if (cancel_work_sync(&msk->work))
2457 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2461 switch (ssk->sk_state) {
2463 if (!(how & RCV_SHUTDOWN))
2467 tcp_disconnect(ssk, O_NONBLOCK);
2470 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2471 pr_debug("Fallback");
2472 ssk->sk_shutdown |= how;
2473 tcp_shutdown(ssk, how);
2475 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2476 mptcp_set_timeout(sk, ssk);
2485 static const unsigned char new_state[16] = {
2486 /* current state: new state: action: */
2487 [0 /* (Invalid) */] = TCP_CLOSE,
2488 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2489 [TCP_SYN_SENT] = TCP_CLOSE,
2490 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2491 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2492 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2493 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2494 [TCP_CLOSE] = TCP_CLOSE,
2495 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2496 [TCP_LAST_ACK] = TCP_LAST_ACK,
2497 [TCP_LISTEN] = TCP_CLOSE,
2498 [TCP_CLOSING] = TCP_CLOSING,
2499 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2502 static int mptcp_close_state(struct sock *sk)
2504 int next = (int)new_state[sk->sk_state];
2505 int ns = next & TCP_STATE_MASK;
2507 inet_sk_state_store(sk, ns);
2509 return next & TCP_ACTION_FIN;
2512 static void __mptcp_check_send_data_fin(struct sock *sk)
2514 struct mptcp_subflow_context *subflow;
2515 struct mptcp_sock *msk = mptcp_sk(sk);
2517 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2518 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2519 msk->snd_nxt, msk->write_seq);
2521 /* we still need to enqueue subflows or not really shutting down,
2524 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2525 mptcp_send_head(sk))
2528 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2530 /* fallback socket will not get data_fin/ack, can move to the next
2533 if (__mptcp_check_fallback(msk)) {
2534 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2535 inet_sk_state_store(sk, TCP_CLOSE);
2536 mptcp_close_wake_up(sk);
2537 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2538 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2542 mptcp_flush_join_list(msk);
2543 mptcp_for_each_subflow(msk, subflow) {
2544 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2546 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2550 static void __mptcp_wr_shutdown(struct sock *sk)
2552 struct mptcp_sock *msk = mptcp_sk(sk);
2554 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2555 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2556 !!mptcp_send_head(sk));
2558 /* will be ignored by fallback sockets */
2559 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2560 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2562 __mptcp_check_send_data_fin(sk);
2565 static void __mptcp_destroy_sock(struct sock *sk)
2567 struct mptcp_subflow_context *subflow, *tmp;
2568 struct mptcp_sock *msk = mptcp_sk(sk);
2569 LIST_HEAD(conn_list);
2571 pr_debug("msk=%p", msk);
2575 /* be sure to always acquire the join list lock, to sync vs
2576 * mptcp_finish_join().
2578 spin_lock_bh(&msk->join_list_lock);
2579 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2580 spin_unlock_bh(&msk->join_list_lock);
2581 list_splice_init(&msk->conn_list, &conn_list);
2583 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2584 sk_stop_timer(sk, &sk->sk_timer);
2587 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2588 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2589 __mptcp_close_ssk(sk, ssk, subflow);
2592 sk->sk_prot->destroy(sk);
2594 WARN_ON_ONCE(msk->wmem_reserved);
2595 WARN_ON_ONCE(msk->rmem_released);
2596 sk_stream_kill_queues(sk);
2597 xfrm_sk_free_policy(sk);
2599 tcp_cleanup_congestion_control(sk);
2600 sk_refcnt_debug_release(sk);
2601 mptcp_dispose_initial_subflow(msk);
2605 static void mptcp_close(struct sock *sk, long timeout)
2607 struct mptcp_subflow_context *subflow;
2608 bool do_cancel_work = false;
2611 sk->sk_shutdown = SHUTDOWN_MASK;
2613 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2614 inet_sk_state_store(sk, TCP_CLOSE);
2618 if (mptcp_close_state(sk))
2619 __mptcp_wr_shutdown(sk);
2621 sk_stream_wait_close(sk, timeout);
2624 /* orphan all the subflows */
2625 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2626 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2627 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2628 bool slow = lock_sock_fast(ssk);
2631 unlock_sock_fast(ssk, slow);
2636 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2637 if (sk->sk_state == TCP_CLOSE) {
2638 __mptcp_destroy_sock(sk);
2639 do_cancel_work = true;
2641 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2645 mptcp_cancel_work(sk);
2647 if (mptcp_sk(sk)->token)
2648 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2653 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2655 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2656 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2657 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2659 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2660 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2663 msk6->saddr = ssk6->saddr;
2664 msk6->flow_label = ssk6->flow_label;
2668 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2669 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2670 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2671 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2672 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2673 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2676 static int mptcp_disconnect(struct sock *sk, int flags)
2678 struct mptcp_subflow_context *subflow;
2679 struct mptcp_sock *msk = mptcp_sk(sk);
2681 mptcp_do_flush_join_list(msk);
2683 mptcp_for_each_subflow(msk, subflow) {
2684 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2687 tcp_disconnect(ssk, flags);
2693 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2694 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2696 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2698 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2702 struct sock *mptcp_sk_clone(const struct sock *sk,
2703 const struct mptcp_options_received *mp_opt,
2704 struct request_sock *req)
2706 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2707 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2708 struct mptcp_sock *msk;
2714 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2715 if (nsk->sk_family == AF_INET6)
2716 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2719 __mptcp_init_sock(nsk);
2721 msk = mptcp_sk(nsk);
2722 msk->local_key = subflow_req->local_key;
2723 msk->token = subflow_req->token;
2724 msk->subflow = NULL;
2725 WRITE_ONCE(msk->fully_established, false);
2727 msk->write_seq = subflow_req->idsn + 1;
2728 msk->snd_nxt = msk->write_seq;
2729 msk->snd_una = msk->write_seq;
2730 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2731 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2733 if (mp_opt->mp_capable) {
2734 msk->can_ack = true;
2735 msk->remote_key = mp_opt->sndr_key;
2736 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2738 WRITE_ONCE(msk->ack_seq, ack_seq);
2739 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2742 sock_reset_flag(nsk, SOCK_RCU_FREE);
2743 /* will be fully established after successful MPC subflow creation */
2744 inet_sk_state_store(nsk, TCP_SYN_RECV);
2746 security_inet_csk_clone(nsk, req);
2747 bh_unlock_sock(nsk);
2749 /* keep a single reference */
2754 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2756 const struct tcp_sock *tp = tcp_sk(ssk);
2758 msk->rcvq_space.copied = 0;
2759 msk->rcvq_space.rtt_us = 0;
2761 msk->rcvq_space.time = tp->tcp_mstamp;
2763 /* initial rcv_space offering made to peer */
2764 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2765 TCP_INIT_CWND * tp->advmss);
2766 if (msk->rcvq_space.space == 0)
2767 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2769 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2772 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2775 struct mptcp_sock *msk = mptcp_sk(sk);
2776 struct socket *listener;
2779 listener = __mptcp_nmpc_socket(msk);
2780 if (WARN_ON_ONCE(!listener)) {
2785 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2786 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2790 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2791 if (sk_is_mptcp(newsk)) {
2792 struct mptcp_subflow_context *subflow;
2793 struct sock *new_mptcp_sock;
2795 subflow = mptcp_subflow_ctx(newsk);
2796 new_mptcp_sock = subflow->conn;
2798 /* is_mptcp should be false if subflow->conn is missing, see
2799 * subflow_syn_recv_sock()
2801 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2802 tcp_sk(newsk)->is_mptcp = 0;
2806 /* acquire the 2nd reference for the owning socket */
2807 sock_hold(new_mptcp_sock);
2808 newsk = new_mptcp_sock;
2809 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2811 MPTCP_INC_STATS(sock_net(sk),
2812 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2818 void mptcp_destroy_common(struct mptcp_sock *msk)
2820 struct sock *sk = (struct sock *)msk;
2822 __mptcp_clear_xmit(sk);
2824 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2825 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2827 skb_rbtree_purge(&msk->out_of_order_queue);
2828 mptcp_token_destroy(msk);
2829 mptcp_pm_free_anno_list(msk);
2832 static void mptcp_destroy(struct sock *sk)
2834 struct mptcp_sock *msk = mptcp_sk(sk);
2836 mptcp_destroy_common(msk);
2837 sk_sockets_allocated_dec(sk);
2840 void __mptcp_data_acked(struct sock *sk)
2842 if (!sock_owned_by_user(sk))
2843 __mptcp_clean_una(sk);
2845 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2847 if (mptcp_pending_data_fin_ack(sk))
2848 mptcp_schedule_work(sk);
2851 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2853 if (!mptcp_send_head(sk))
2856 if (!sock_owned_by_user(sk)) {
2857 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2859 if (xmit_ssk == ssk)
2860 __mptcp_subflow_push_pending(sk, ssk);
2862 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
2864 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2868 /* processes deferred events and flush wmem */
2869 static void mptcp_release_cb(struct sock *sk)
2872 unsigned long flags = 0;
2874 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2875 flags |= BIT(MPTCP_PUSH_PENDING);
2876 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2877 flags |= BIT(MPTCP_RETRANSMIT);
2881 /* the following actions acquire the subflow socket lock
2883 * 1) can't be invoked in atomic scope
2884 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2885 * datapath acquires the msk socket spinlock while helding
2886 * the subflow socket lock
2889 spin_unlock_bh(&sk->sk_lock.slock);
2890 if (flags & BIT(MPTCP_PUSH_PENDING))
2891 __mptcp_push_pending(sk, 0);
2892 if (flags & BIT(MPTCP_RETRANSMIT))
2893 __mptcp_retrans(sk);
2896 spin_lock_bh(&sk->sk_lock.slock);
2899 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2900 __mptcp_clean_una_wakeup(sk);
2901 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2902 __mptcp_error_report(sk);
2904 /* push_pending may touch wmem_reserved, ensure we do the cleanup
2907 __mptcp_update_wmem(sk);
2908 __mptcp_update_rmem(sk);
2911 void mptcp_subflow_process_delegated(struct sock *ssk)
2913 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2914 struct sock *sk = subflow->conn;
2916 mptcp_data_lock(sk);
2917 if (!sock_owned_by_user(sk))
2918 __mptcp_subflow_push_pending(sk, ssk);
2920 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2921 mptcp_data_unlock(sk);
2922 mptcp_subflow_delegated_done(subflow);
2925 static int mptcp_hash(struct sock *sk)
2927 /* should never be called,
2928 * we hash the TCP subflows not the master socket
2934 static void mptcp_unhash(struct sock *sk)
2936 /* called from sk_common_release(), but nothing to do here */
2939 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2941 struct mptcp_sock *msk = mptcp_sk(sk);
2942 struct socket *ssock;
2944 ssock = __mptcp_nmpc_socket(msk);
2945 pr_debug("msk=%p, subflow=%p", msk, ssock);
2946 if (WARN_ON_ONCE(!ssock))
2949 return inet_csk_get_port(ssock->sk, snum);
2952 void mptcp_finish_connect(struct sock *ssk)
2954 struct mptcp_subflow_context *subflow;
2955 struct mptcp_sock *msk;
2959 subflow = mptcp_subflow_ctx(ssk);
2963 pr_debug("msk=%p, token=%u", sk, subflow->token);
2965 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2967 subflow->map_seq = ack_seq;
2968 subflow->map_subflow_seq = 1;
2970 /* the socket is not connected yet, no msk/subflow ops can access/race
2971 * accessing the field below
2973 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2974 WRITE_ONCE(msk->local_key, subflow->local_key);
2975 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2976 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2977 WRITE_ONCE(msk->ack_seq, ack_seq);
2978 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2979 WRITE_ONCE(msk->can_ack, 1);
2980 WRITE_ONCE(msk->snd_una, msk->write_seq);
2982 mptcp_pm_new_connection(msk, ssk, 0);
2984 mptcp_rcv_space_init(msk, ssk);
2987 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2989 write_lock_bh(&sk->sk_callback_lock);
2990 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2991 sk_set_socket(sk, parent);
2992 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2993 write_unlock_bh(&sk->sk_callback_lock);
2996 bool mptcp_finish_join(struct sock *ssk)
2998 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2999 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3000 struct sock *parent = (void *)msk;
3001 struct socket *parent_sock;
3004 pr_debug("msk=%p, subflow=%p", msk, subflow);
3006 /* mptcp socket already closing? */
3007 if (!mptcp_is_fully_established(parent)) {
3008 subflow->reset_reason = MPTCP_RST_EMPTCP;
3012 if (!msk->pm.server_side)
3015 if (!mptcp_pm_allow_new_subflow(msk)) {
3016 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3020 /* active connections are already on conn_list, and we can't acquire
3022 * use the join list lock as synchronization point and double-check
3023 * msk status to avoid racing with __mptcp_destroy_sock()
3025 spin_lock_bh(&msk->join_list_lock);
3026 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3027 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3028 list_add_tail(&subflow->node, &msk->join_list);
3031 spin_unlock_bh(&msk->join_list_lock);
3033 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3037 /* attach to msk socket only after we are sure he will deal with us
3040 parent_sock = READ_ONCE(parent->sk_socket);
3041 if (parent_sock && !ssk->sk_socket)
3042 mptcp_sock_graft(ssk, parent_sock);
3043 subflow->map_seq = READ_ONCE(msk->ack_seq);
3045 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3049 static void mptcp_shutdown(struct sock *sk, int how)
3051 pr_debug("sk=%p, how=%d", sk, how);
3053 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3054 __mptcp_wr_shutdown(sk);
3057 static struct proto mptcp_prot = {
3059 .owner = THIS_MODULE,
3060 .init = mptcp_init_sock,
3061 .disconnect = mptcp_disconnect,
3062 .close = mptcp_close,
3063 .accept = mptcp_accept,
3064 .setsockopt = mptcp_setsockopt,
3065 .getsockopt = mptcp_getsockopt,
3066 .shutdown = mptcp_shutdown,
3067 .destroy = mptcp_destroy,
3068 .sendmsg = mptcp_sendmsg,
3069 .recvmsg = mptcp_recvmsg,
3070 .release_cb = mptcp_release_cb,
3072 .unhash = mptcp_unhash,
3073 .get_port = mptcp_get_port,
3074 .sockets_allocated = &mptcp_sockets_allocated,
3075 .memory_allocated = &tcp_memory_allocated,
3076 .memory_pressure = &tcp_memory_pressure,
3077 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3078 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3079 .sysctl_mem = sysctl_tcp_mem,
3080 .obj_size = sizeof(struct mptcp_sock),
3081 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3082 .no_autobind = true,
3085 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3087 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3088 struct socket *ssock;
3091 lock_sock(sock->sk);
3092 ssock = __mptcp_nmpc_socket(msk);
3098 err = ssock->ops->bind(ssock, uaddr, addr_len);
3100 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3103 release_sock(sock->sk);
3107 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3108 struct mptcp_subflow_context *subflow)
3110 subflow->request_mptcp = 0;
3111 __mptcp_do_fallback(msk);
3114 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3115 int addr_len, int flags)
3117 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3118 struct mptcp_subflow_context *subflow;
3119 struct socket *ssock;
3122 lock_sock(sock->sk);
3123 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3124 /* pending connection or invalid state, let existing subflow
3127 ssock = msk->subflow;
3131 ssock = __mptcp_nmpc_socket(msk);
3137 mptcp_token_destroy(msk);
3138 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3139 subflow = mptcp_subflow_ctx(ssock->sk);
3140 #ifdef CONFIG_TCP_MD5SIG
3141 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3144 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3145 mptcp_subflow_early_fallback(msk, subflow);
3147 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3148 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3149 mptcp_subflow_early_fallback(msk, subflow);
3151 if (likely(!__mptcp_check_fallback(msk)))
3152 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3155 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3156 sock->state = ssock->state;
3158 /* on successful connect, the msk state will be moved to established by
3159 * subflow_finish_connect()
3161 if (!err || err == -EINPROGRESS)
3162 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3164 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3167 release_sock(sock->sk);
3171 static int mptcp_listen(struct socket *sock, int backlog)
3173 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3174 struct socket *ssock;
3177 pr_debug("msk=%p", msk);
3179 lock_sock(sock->sk);
3180 ssock = __mptcp_nmpc_socket(msk);
3186 mptcp_token_destroy(msk);
3187 inet_sk_state_store(sock->sk, TCP_LISTEN);
3188 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3190 err = ssock->ops->listen(ssock, backlog);
3191 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3193 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3196 release_sock(sock->sk);
3200 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3201 int flags, bool kern)
3203 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3204 struct socket *ssock;
3207 pr_debug("msk=%p", msk);
3209 lock_sock(sock->sk);
3210 if (sock->sk->sk_state != TCP_LISTEN)
3213 ssock = __mptcp_nmpc_socket(msk);
3217 clear_bit(MPTCP_DATA_READY, &msk->flags);
3218 sock_hold(ssock->sk);
3219 release_sock(sock->sk);
3221 err = ssock->ops->accept(sock, newsock, flags, kern);
3222 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3223 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3224 struct mptcp_subflow_context *subflow;
3225 struct sock *newsk = newsock->sk;
3229 /* PM/worker can now acquire the first subflow socket
3230 * lock without racing with listener queue cleanup,
3231 * we can notify it, if needed.
3233 * Even if remote has reset the initial subflow by now
3234 * the refcnt is still at least one.
3236 subflow = mptcp_subflow_ctx(msk->first);
3237 list_add(&subflow->node, &msk->conn_list);
3238 sock_hold(msk->first);
3239 if (mptcp_is_fully_established(newsk))
3240 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3242 mptcp_copy_inaddrs(newsk, msk->first);
3243 mptcp_rcv_space_init(msk, msk->first);
3244 mptcp_propagate_sndbuf(newsk, msk->first);
3246 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3247 * This is needed so NOSPACE flag can be set from tcp stack.
3249 mptcp_flush_join_list(msk);
3250 mptcp_for_each_subflow(msk, subflow) {
3251 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3253 if (!ssk->sk_socket)
3254 mptcp_sock_graft(ssk, newsock);
3256 release_sock(newsk);
3259 if (inet_csk_listen_poll(ssock->sk))
3260 set_bit(MPTCP_DATA_READY, &msk->flags);
3261 sock_put(ssock->sk);
3265 release_sock(sock->sk);
3269 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3271 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3275 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3277 struct sock *sk = (struct sock *)msk;
3279 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3280 return EPOLLOUT | EPOLLWRNORM;
3282 if (sk_stream_is_writeable(sk))
3283 return EPOLLOUT | EPOLLWRNORM;
3285 mptcp_set_nospace(sk);
3286 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3287 if (sk_stream_is_writeable(sk))
3288 return EPOLLOUT | EPOLLWRNORM;
3293 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3294 struct poll_table_struct *wait)
3296 struct sock *sk = sock->sk;
3297 struct mptcp_sock *msk;
3302 sock_poll_wait(file, sock, wait);
3304 state = inet_sk_state_load(sk);
3305 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3306 if (state == TCP_LISTEN)
3307 return mptcp_check_readable(msk);
3309 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3310 mask |= mptcp_check_readable(msk);
3311 mask |= mptcp_check_writeable(msk);
3313 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3315 if (sk->sk_shutdown & RCV_SHUTDOWN)
3316 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3318 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3326 static const struct proto_ops mptcp_stream_ops = {
3328 .owner = THIS_MODULE,
3329 .release = inet_release,
3331 .connect = mptcp_stream_connect,
3332 .socketpair = sock_no_socketpair,
3333 .accept = mptcp_stream_accept,
3334 .getname = inet_getname,
3336 .ioctl = inet_ioctl,
3337 .gettstamp = sock_gettstamp,
3338 .listen = mptcp_listen,
3339 .shutdown = inet_shutdown,
3340 .setsockopt = sock_common_setsockopt,
3341 .getsockopt = sock_common_getsockopt,
3342 .sendmsg = inet_sendmsg,
3343 .recvmsg = inet_recvmsg,
3344 .mmap = sock_no_mmap,
3345 .sendpage = inet_sendpage,
3348 static struct inet_protosw mptcp_protosw = {
3349 .type = SOCK_STREAM,
3350 .protocol = IPPROTO_MPTCP,
3351 .prot = &mptcp_prot,
3352 .ops = &mptcp_stream_ops,
3353 .flags = INET_PROTOSW_ICSK,
3356 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3358 struct mptcp_delegated_action *delegated;
3359 struct mptcp_subflow_context *subflow;
3362 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3363 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3364 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3366 bh_lock_sock_nested(ssk);
3367 if (!sock_owned_by_user(ssk) &&
3368 mptcp_subflow_has_delegated_action(subflow))
3369 mptcp_subflow_process_delegated(ssk);
3370 /* ... elsewhere tcp_release_cb_override already processed
3371 * the action or will do at next release_sock().
3372 * In both case must dequeue the subflow here - on the same
3373 * CPU that scheduled it.
3375 bh_unlock_sock(ssk);
3378 if (++work_done == budget)
3382 /* always provide a 0 'work_done' argument, so that napi_complete_done
3383 * will not try accessing the NULL napi->dev ptr
3385 napi_complete_done(napi, 0);
3389 void __init mptcp_proto_init(void)
3391 struct mptcp_delegated_action *delegated;
3394 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3396 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3397 panic("Failed to allocate MPTCP pcpu counter\n");
3399 init_dummy_netdev(&mptcp_napi_dev);
3400 for_each_possible_cpu(cpu) {
3401 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3402 INIT_LIST_HEAD(&delegated->head);
3403 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3405 napi_enable(&delegated->napi);
3408 mptcp_subflow_init();
3412 if (proto_register(&mptcp_prot, 1) != 0)
3413 panic("Failed to register MPTCP proto.\n");
3415 inet_register_protosw(&mptcp_protosw);
3417 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3420 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3421 static const struct proto_ops mptcp_v6_stream_ops = {
3423 .owner = THIS_MODULE,
3424 .release = inet6_release,
3426 .connect = mptcp_stream_connect,
3427 .socketpair = sock_no_socketpair,
3428 .accept = mptcp_stream_accept,
3429 .getname = inet6_getname,
3431 .ioctl = inet6_ioctl,
3432 .gettstamp = sock_gettstamp,
3433 .listen = mptcp_listen,
3434 .shutdown = inet_shutdown,
3435 .setsockopt = sock_common_setsockopt,
3436 .getsockopt = sock_common_getsockopt,
3437 .sendmsg = inet6_sendmsg,
3438 .recvmsg = inet6_recvmsg,
3439 .mmap = sock_no_mmap,
3440 .sendpage = inet_sendpage,
3441 #ifdef CONFIG_COMPAT
3442 .compat_ioctl = inet6_compat_ioctl,
3446 static struct proto mptcp_v6_prot;
3448 static void mptcp_v6_destroy(struct sock *sk)
3451 inet6_destroy_sock(sk);
3454 static struct inet_protosw mptcp_v6_protosw = {
3455 .type = SOCK_STREAM,
3456 .protocol = IPPROTO_MPTCP,
3457 .prot = &mptcp_v6_prot,
3458 .ops = &mptcp_v6_stream_ops,
3459 .flags = INET_PROTOSW_ICSK,
3462 int __init mptcp_proto_v6_init(void)
3466 mptcp_v6_prot = mptcp_prot;
3467 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3468 mptcp_v6_prot.slab = NULL;
3469 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3470 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3472 err = proto_register(&mptcp_v6_prot, 1);
3476 err = inet6_register_protosw(&mptcp_v6_protosw);
3478 proto_unregister(&mptcp_v6_prot);