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 void __mptcp_flush_join_list(struct mptcp_sock *msk)
735 struct mptcp_subflow_context *subflow;
737 if (likely(list_empty(&msk->join_list)))
740 spin_lock_bh(&msk->join_list_lock);
741 list_for_each_entry(subflow, &msk->join_list, node)
742 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
743 list_splice_tail_init(&msk->join_list, &msk->conn_list);
744 spin_unlock_bh(&msk->join_list_lock);
747 static bool mptcp_timer_pending(struct sock *sk)
749 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
752 static void mptcp_reset_timer(struct sock *sk)
754 struct inet_connection_sock *icsk = inet_csk(sk);
757 /* prevent rescheduling on close */
758 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
761 /* should never be called with mptcp level timer cleared */
762 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
763 if (WARN_ON_ONCE(!tout))
765 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
768 bool mptcp_schedule_work(struct sock *sk)
770 if (inet_sk_state_load(sk) != TCP_CLOSE &&
771 schedule_work(&mptcp_sk(sk)->work)) {
772 /* each subflow already holds a reference to the sk, and the
773 * workqueue is invoked by a subflow, so sk can't go away here.
781 void mptcp_subflow_eof(struct sock *sk)
783 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
784 mptcp_schedule_work(sk);
787 static void mptcp_check_for_eof(struct mptcp_sock *msk)
789 struct mptcp_subflow_context *subflow;
790 struct sock *sk = (struct sock *)msk;
793 mptcp_for_each_subflow(msk, subflow)
794 receivers += !subflow->rx_eof;
798 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
799 /* hopefully temporary hack: propagate shutdown status
800 * to msk, when all subflows agree on it
802 sk->sk_shutdown |= RCV_SHUTDOWN;
804 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
805 set_bit(MPTCP_DATA_READY, &msk->flags);
806 sk->sk_data_ready(sk);
809 switch (sk->sk_state) {
810 case TCP_ESTABLISHED:
811 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
814 inet_sk_state_store(sk, TCP_CLOSING);
817 inet_sk_state_store(sk, TCP_CLOSE);
822 mptcp_close_wake_up(sk);
825 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
827 struct mptcp_subflow_context *subflow;
828 struct sock *sk = (struct sock *)msk;
830 sock_owned_by_me(sk);
832 mptcp_for_each_subflow(msk, subflow) {
833 if (subflow->data_avail)
834 return mptcp_subflow_tcp_sock(subflow);
840 static bool mptcp_skb_can_collapse_to(u64 write_seq,
841 const struct sk_buff *skb,
842 const struct mptcp_ext *mpext)
844 if (!tcp_skb_can_collapse_to(skb))
847 /* can collapse only if MPTCP level sequence is in order and this
848 * mapping has not been xmitted yet
850 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
854 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
855 const struct page_frag *pfrag,
856 const struct mptcp_data_frag *df)
858 return df && pfrag->page == df->page &&
859 pfrag->size - pfrag->offset > 0 &&
860 df->data_seq + df->data_len == msk->write_seq;
863 static int mptcp_wmem_with_overhead(struct sock *sk, int size)
865 struct mptcp_sock *msk = mptcp_sk(sk);
868 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
869 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache;
870 if (skbs < msk->skb_tx_cache.qlen)
873 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER);
876 static void __mptcp_wmem_reserve(struct sock *sk, int size)
878 int amount = mptcp_wmem_with_overhead(sk, size);
879 struct mptcp_sock *msk = mptcp_sk(sk);
881 WARN_ON_ONCE(msk->wmem_reserved);
882 if (WARN_ON_ONCE(amount < 0))
885 if (amount <= sk->sk_forward_alloc)
888 /* under memory pressure try to reserve at most a single page
889 * otherwise try to reserve the full estimate and fallback
890 * to a single page before entering the error path
892 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
893 !sk_wmem_schedule(sk, amount)) {
894 if (amount <= PAGE_SIZE)
898 if (!sk_wmem_schedule(sk, amount))
903 msk->wmem_reserved = amount;
904 sk->sk_forward_alloc -= amount;
908 /* we will wait for memory on next allocation */
909 msk->wmem_reserved = -1;
912 static void __mptcp_update_wmem(struct sock *sk)
914 struct mptcp_sock *msk = mptcp_sk(sk);
916 if (!msk->wmem_reserved)
919 if (msk->wmem_reserved < 0)
920 msk->wmem_reserved = 0;
921 if (msk->wmem_reserved > 0) {
922 sk->sk_forward_alloc += msk->wmem_reserved;
923 msk->wmem_reserved = 0;
927 static bool mptcp_wmem_alloc(struct sock *sk, int size)
929 struct mptcp_sock *msk = mptcp_sk(sk);
931 /* check for pre-existing error condition */
932 if (msk->wmem_reserved < 0)
935 if (msk->wmem_reserved >= size)
939 if (!sk_wmem_schedule(sk, size)) {
940 mptcp_data_unlock(sk);
944 sk->sk_forward_alloc -= size;
945 msk->wmem_reserved += size;
946 mptcp_data_unlock(sk);
949 msk->wmem_reserved -= size;
953 static void mptcp_wmem_uncharge(struct sock *sk, int size)
955 struct mptcp_sock *msk = mptcp_sk(sk);
957 if (msk->wmem_reserved < 0)
958 msk->wmem_reserved = 0;
959 msk->wmem_reserved += size;
962 static void mptcp_mem_reclaim_partial(struct sock *sk)
964 struct mptcp_sock *msk = mptcp_sk(sk);
966 /* if we are experiencing a transint allocation error,
967 * the forward allocation memory has been already
970 if (msk->wmem_reserved < 0)
974 sk->sk_forward_alloc += msk->wmem_reserved;
975 sk_mem_reclaim_partial(sk);
976 msk->wmem_reserved = sk->sk_forward_alloc;
977 sk->sk_forward_alloc = 0;
978 mptcp_data_unlock(sk);
981 static void dfrag_uncharge(struct sock *sk, int len)
983 sk_mem_uncharge(sk, len);
984 sk_wmem_queued_add(sk, -len);
987 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
989 int len = dfrag->data_len + dfrag->overhead;
991 list_del(&dfrag->list);
992 dfrag_uncharge(sk, len);
993 put_page(dfrag->page);
996 static void __mptcp_clean_una(struct sock *sk)
998 struct mptcp_sock *msk = mptcp_sk(sk);
999 struct mptcp_data_frag *dtmp, *dfrag;
1000 bool cleaned = false;
1003 /* on fallback we just need to ignore snd_una, as this is really
1006 if (__mptcp_check_fallback(msk))
1007 msk->snd_una = READ_ONCE(msk->snd_nxt);
1009 snd_una = msk->snd_una;
1010 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1011 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1014 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1016 dfrag_clear(sk, dfrag);
1020 dfrag = mptcp_rtx_head(sk);
1021 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1022 u64 delta = snd_una - dfrag->data_seq;
1024 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1027 dfrag->data_seq += delta;
1028 dfrag->offset += delta;
1029 dfrag->data_len -= delta;
1030 dfrag->already_sent -= delta;
1032 dfrag_uncharge(sk, delta);
1038 if (tcp_under_memory_pressure(sk)) {
1039 __mptcp_update_wmem(sk);
1040 sk_mem_reclaim_partial(sk);
1044 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1045 if (msk->timer_ival)
1046 mptcp_stop_timer(sk);
1048 mptcp_reset_timer(sk);
1052 static void __mptcp_clean_una_wakeup(struct sock *sk)
1054 __mptcp_clean_una(sk);
1055 mptcp_write_space(sk);
1058 static void mptcp_enter_memory_pressure(struct sock *sk)
1060 struct mptcp_subflow_context *subflow;
1061 struct mptcp_sock *msk = mptcp_sk(sk);
1064 sk_stream_moderate_sndbuf(sk);
1065 mptcp_for_each_subflow(msk, subflow) {
1066 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1069 tcp_enter_memory_pressure(ssk);
1070 sk_stream_moderate_sndbuf(ssk);
1075 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1078 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1080 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1081 pfrag, sk->sk_allocation)))
1084 mptcp_enter_memory_pressure(sk);
1088 static struct mptcp_data_frag *
1089 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1092 int offset = ALIGN(orig_offset, sizeof(long));
1093 struct mptcp_data_frag *dfrag;
1095 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1096 dfrag->data_len = 0;
1097 dfrag->data_seq = msk->write_seq;
1098 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1099 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1100 dfrag->already_sent = 0;
1101 dfrag->page = pfrag->page;
1106 struct mptcp_sendmsg_info {
1114 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1117 u64 window_end = mptcp_wnd_end(msk);
1119 if (__mptcp_check_fallback(msk))
1122 if (!before64(data_seq + avail_size, window_end)) {
1123 u64 allowed_size = window_end - data_seq;
1125 return min_t(unsigned int, allowed_size, avail_size);
1131 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1133 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1137 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1141 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1143 struct sk_buff *skb;
1145 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1147 if (likely(__mptcp_add_ext(skb, gfp))) {
1148 skb_reserve(skb, MAX_TCP_HEADER);
1149 skb->reserved_tailroom = skb->end - skb->tail;
1154 mptcp_enter_memory_pressure(sk);
1159 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1160 struct sk_buff_head *skbs, int *total_ts)
1162 struct mptcp_sock *msk = mptcp_sk(sk);
1163 struct sk_buff *skb;
1166 if (unlikely(tcp_under_memory_pressure(sk))) {
1167 mptcp_mem_reclaim_partial(sk);
1169 /* under pressure pre-allocate at most a single skb */
1170 if (msk->skb_tx_cache.qlen)
1172 space_needed = msk->size_goal_cache;
1174 space_needed = msk->tx_pending_data + size -
1175 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1178 while (space_needed > 0) {
1179 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1180 if (unlikely(!skb)) {
1181 /* under memory pressure, try to pass the caller a
1182 * single skb to allow forward progress
1184 while (skbs->qlen > 1) {
1185 skb = __skb_dequeue_tail(skbs);
1186 *total_ts -= skb->truesize;
1189 return skbs->qlen > 0;
1192 *total_ts += skb->truesize;
1193 __skb_queue_tail(skbs, skb);
1194 space_needed -= msk->size_goal_cache;
1199 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1201 struct mptcp_sock *msk = mptcp_sk(sk);
1202 struct sk_buff *skb;
1204 if (ssk->sk_tx_skb_cache) {
1205 skb = ssk->sk_tx_skb_cache;
1206 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1207 !__mptcp_add_ext(skb, gfp)))
1212 skb = skb_peek(&msk->skb_tx_cache);
1214 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1215 skb = __skb_dequeue(&msk->skb_tx_cache);
1216 if (WARN_ON_ONCE(!skb))
1219 mptcp_wmem_uncharge(sk, skb->truesize);
1220 ssk->sk_tx_skb_cache = skb;
1224 /* over memory limit, no point to try to allocate a new skb */
1228 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1232 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1233 ssk->sk_tx_skb_cache = skb;
1240 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1242 return !ssk->sk_tx_skb_cache &&
1243 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1244 tcp_under_memory_pressure(sk);
1247 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1249 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1250 mptcp_mem_reclaim_partial(sk);
1251 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1254 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1255 struct mptcp_data_frag *dfrag,
1256 struct mptcp_sendmsg_info *info)
1258 u64 data_seq = dfrag->data_seq + info->sent;
1259 struct mptcp_sock *msk = mptcp_sk(sk);
1260 bool zero_window_probe = false;
1261 struct mptcp_ext *mpext = NULL;
1262 struct sk_buff *skb, *tail;
1263 bool can_collapse = false;
1268 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
1269 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1271 /* compute send limit */
1272 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1273 avail_size = info->size_goal;
1274 msk->size_goal_cache = info->size_goal;
1275 skb = tcp_write_queue_tail(ssk);
1277 /* Limit the write to the size available in the
1278 * current skb, if any, so that we create at most a new skb.
1279 * Explicitly tells TCP internals to avoid collapsing on later
1280 * queue management operation, to avoid breaking the ext <->
1281 * SSN association set here
1283 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1284 can_collapse = (info->size_goal - skb->len > 0) &&
1285 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1286 if (!can_collapse) {
1287 TCP_SKB_CB(skb)->eor = 1;
1289 size_bias = skb->len;
1290 avail_size = info->size_goal - skb->len;
1294 /* Zero window and all data acked? Probe. */
1295 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1296 if (avail_size == 0) {
1297 u64 snd_una = READ_ONCE(msk->snd_una);
1299 if (skb || snd_una != msk->snd_nxt)
1301 zero_window_probe = true;
1302 data_seq = snd_una - 1;
1306 if (WARN_ON_ONCE(info->sent > info->limit ||
1307 info->limit > dfrag->data_len))
1310 ret = info->limit - info->sent;
1311 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1312 dfrag->page, dfrag->offset + info->sent, &ret);
1314 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1318 /* if the tail skb is still the cached one, collapsing really happened.
1321 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1322 mpext->data_len += ret;
1323 WARN_ON_ONCE(!can_collapse);
1324 WARN_ON_ONCE(zero_window_probe);
1328 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1329 if (WARN_ON_ONCE(!mpext)) {
1330 /* should never reach here, stream corrupted */
1334 memset(mpext, 0, sizeof(*mpext));
1335 mpext->data_seq = data_seq;
1336 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1337 mpext->data_len = ret;
1341 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1342 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1345 if (zero_window_probe) {
1346 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1349 tcp_push_pending_frames(ssk);
1352 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1356 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1357 sizeof(struct tcphdr) - \
1358 MAX_TCP_OPTION_SPACE - \
1359 sizeof(struct ipv6hdr) - \
1360 sizeof(struct frag_hdr))
1362 struct subflow_send_info {
1367 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1369 struct subflow_send_info send_info[2];
1370 struct mptcp_subflow_context *subflow;
1371 int i, nr_active = 0;
1376 sock_owned_by_me((struct sock *)msk);
1378 if (__mptcp_check_fallback(msk)) {
1381 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1384 /* re-use last subflow, if the burst allow that */
1385 if (msk->last_snd && msk->snd_burst > 0 &&
1386 sk_stream_memory_free(msk->last_snd) &&
1387 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1388 return msk->last_snd;
1390 /* pick the subflow with the lower wmem/wspace ratio */
1391 for (i = 0; i < 2; ++i) {
1392 send_info[i].ssk = NULL;
1393 send_info[i].ratio = -1;
1395 mptcp_for_each_subflow(msk, subflow) {
1396 ssk = mptcp_subflow_tcp_sock(subflow);
1397 if (!mptcp_subflow_active(subflow))
1400 nr_active += !subflow->backup;
1401 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1404 pace = READ_ONCE(ssk->sk_pacing_rate);
1408 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1410 if (ratio < send_info[subflow->backup].ratio) {
1411 send_info[subflow->backup].ssk = ssk;
1412 send_info[subflow->backup].ratio = ratio;
1416 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1417 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1418 send_info[1].ssk, send_info[1].ratio);
1420 /* pick the best backup if no other subflow is active */
1422 send_info[0].ssk = send_info[1].ssk;
1424 if (send_info[0].ssk) {
1425 msk->last_snd = send_info[0].ssk;
1426 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1427 tcp_sk(msk->last_snd)->snd_wnd);
1428 return msk->last_snd;
1434 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1435 struct mptcp_sendmsg_info *info)
1437 mptcp_set_timeout(sk, ssk);
1438 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1442 static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1444 struct sock *prev_ssk = NULL, *ssk = NULL;
1445 struct mptcp_sock *msk = mptcp_sk(sk);
1446 struct mptcp_sendmsg_info info = {
1449 struct mptcp_data_frag *dfrag;
1450 int len, copied = 0;
1452 while ((dfrag = mptcp_send_head(sk))) {
1453 info.sent = dfrag->already_sent;
1454 info.limit = dfrag->data_len;
1455 len = dfrag->data_len - dfrag->already_sent;
1460 __mptcp_flush_join_list(msk);
1461 ssk = mptcp_subflow_get_send(msk);
1463 /* try to keep the subflow socket lock across
1464 * consecutive xmit on the same socket
1466 if (ssk != prev_ssk && prev_ssk)
1467 mptcp_push_release(sk, prev_ssk, &info);
1471 if (ssk != prev_ssk || !prev_ssk)
1474 /* keep it simple and always provide a new skb for the
1475 * subflow, even if we will not use it when collapsing
1476 * on the pending one
1478 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1479 mptcp_push_release(sk, ssk, &info);
1483 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1485 mptcp_push_release(sk, ssk, &info);
1490 dfrag->already_sent += ret;
1491 msk->snd_nxt += ret;
1492 msk->snd_burst -= ret;
1493 msk->tx_pending_data -= ret;
1497 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1500 /* at this point we held the socket lock for the last subflow we used */
1502 mptcp_push_release(sk, ssk, &info);
1506 /* start the timer, if it's not pending */
1507 if (!mptcp_timer_pending(sk))
1508 mptcp_reset_timer(sk);
1509 __mptcp_check_send_data_fin(sk);
1513 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1515 struct mptcp_sock *msk = mptcp_sk(sk);
1516 struct mptcp_sendmsg_info info;
1517 struct mptcp_data_frag *dfrag;
1518 struct sock *xmit_ssk;
1519 int len, copied = 0;
1523 while ((dfrag = mptcp_send_head(sk))) {
1524 info.sent = dfrag->already_sent;
1525 info.limit = dfrag->data_len;
1526 len = dfrag->data_len - dfrag->already_sent;
1530 /* the caller already invoked the packet scheduler,
1531 * check for a different subflow usage only after
1532 * spooling the first chunk of data
1534 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1537 if (xmit_ssk != ssk) {
1538 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1542 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1543 __mptcp_update_wmem(sk);
1544 sk_mem_reclaim_partial(sk);
1546 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1549 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1554 dfrag->already_sent += ret;
1555 msk->snd_nxt += ret;
1556 msk->snd_burst -= ret;
1557 msk->tx_pending_data -= ret;
1562 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1566 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1567 * not going to flush it via release_sock()
1569 __mptcp_update_wmem(sk);
1571 mptcp_set_timeout(sk, ssk);
1572 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1574 if (!mptcp_timer_pending(sk))
1575 mptcp_reset_timer(sk);
1577 if (msk->snd_data_fin_enable &&
1578 msk->snd_nxt + 1 == msk->write_seq)
1579 mptcp_schedule_work(sk);
1583 static void mptcp_set_nospace(struct sock *sk)
1585 /* enable autotune */
1586 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1588 /* will be cleared on avail space */
1589 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1592 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1594 struct mptcp_sock *msk = mptcp_sk(sk);
1595 struct page_frag *pfrag;
1600 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1603 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1605 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1607 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1608 ret = sk_stream_wait_connect(sk, &timeo);
1613 pfrag = sk_page_frag(sk);
1615 while (msg_data_left(msg)) {
1616 int total_ts, frag_truesize = 0;
1617 struct mptcp_data_frag *dfrag;
1618 struct sk_buff_head skbs;
1619 bool dfrag_collapsed;
1620 size_t psize, offset;
1622 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1627 /* reuse tail pfrag, if possible, or carve a new one from the
1630 dfrag = mptcp_pending_tail(sk);
1631 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1632 if (!dfrag_collapsed) {
1633 if (!sk_stream_memory_free(sk))
1634 goto wait_for_memory;
1636 if (!mptcp_page_frag_refill(sk, pfrag))
1637 goto wait_for_memory;
1639 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1640 frag_truesize = dfrag->overhead;
1643 /* we do not bound vs wspace, to allow a single packet.
1644 * memory accounting will prevent execessive memory usage
1647 offset = dfrag->offset + dfrag->data_len;
1648 psize = pfrag->size - offset;
1649 psize = min_t(size_t, psize, msg_data_left(msg));
1650 total_ts = psize + frag_truesize;
1651 __skb_queue_head_init(&skbs);
1652 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1653 goto wait_for_memory;
1655 if (!mptcp_wmem_alloc(sk, total_ts)) {
1656 __skb_queue_purge(&skbs);
1657 goto wait_for_memory;
1660 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1661 if (copy_page_from_iter(dfrag->page, offset, psize,
1662 &msg->msg_iter) != psize) {
1663 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1668 /* data successfully copied into the write queue */
1670 dfrag->data_len += psize;
1671 frag_truesize += psize;
1672 pfrag->offset += frag_truesize;
1673 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1674 msk->tx_pending_data += psize;
1676 /* charge data on mptcp pending queue to the msk socket
1677 * Note: we charge such data both to sk and ssk
1679 sk_wmem_queued_add(sk, frag_truesize);
1680 if (!dfrag_collapsed) {
1681 get_page(dfrag->page);
1682 list_add_tail(&dfrag->list, &msk->rtx_queue);
1683 if (!msk->first_pending)
1684 WRITE_ONCE(msk->first_pending, dfrag);
1686 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1687 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1693 mptcp_set_nospace(sk);
1694 __mptcp_push_pending(sk, msg->msg_flags);
1695 ret = sk_stream_wait_memory(sk, &timeo);
1701 __mptcp_push_pending(sk, msg->msg_flags);
1705 return copied ? : ret;
1708 static void mptcp_wait_data(struct sock *sk, long *timeo)
1710 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1711 struct mptcp_sock *msk = mptcp_sk(sk);
1713 add_wait_queue(sk_sleep(sk), &wait);
1714 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1716 sk_wait_event(sk, timeo,
1717 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1719 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1720 remove_wait_queue(sk_sleep(sk), &wait);
1723 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1727 struct sk_buff *skb;
1730 while ((skb = skb_peek(&msk->receive_queue)) != NULL) {
1731 u32 offset = MPTCP_SKB_CB(skb)->offset;
1732 u32 data_len = skb->len - offset;
1733 u32 count = min_t(size_t, len - copied, data_len);
1736 err = skb_copy_datagram_msg(skb, offset, msg, count);
1737 if (unlikely(err < 0)) {
1745 if (count < data_len) {
1746 MPTCP_SKB_CB(skb)->offset += count;
1750 /* we will bulk release the skb memory later */
1751 skb->destructor = NULL;
1752 msk->rmem_released += skb->truesize;
1753 __skb_unlink(skb, &msk->receive_queue);
1763 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1765 * Only difference: Use highest rtt estimate of the subflows in use.
1767 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1769 struct mptcp_subflow_context *subflow;
1770 struct sock *sk = (struct sock *)msk;
1771 u32 time, advmss = 1;
1774 sock_owned_by_me(sk);
1779 msk->rcvq_space.copied += copied;
1781 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1782 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1784 rtt_us = msk->rcvq_space.rtt_us;
1785 if (rtt_us && time < (rtt_us >> 3))
1789 mptcp_for_each_subflow(msk, subflow) {
1790 const struct tcp_sock *tp;
1794 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1796 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1797 sf_advmss = READ_ONCE(tp->advmss);
1799 rtt_us = max(sf_rtt_us, rtt_us);
1800 advmss = max(sf_advmss, advmss);
1803 msk->rcvq_space.rtt_us = rtt_us;
1804 if (time < (rtt_us >> 3) || rtt_us == 0)
1807 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1810 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1811 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1815 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1817 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1819 do_div(grow, msk->rcvq_space.space);
1820 rcvwin += (grow << 1);
1822 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1823 while (tcp_win_from_space(sk, rcvmem) < advmss)
1826 do_div(rcvwin, advmss);
1827 rcvbuf = min_t(u64, rcvwin * rcvmem,
1828 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1830 if (rcvbuf > sk->sk_rcvbuf) {
1833 window_clamp = tcp_win_from_space(sk, rcvbuf);
1834 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1836 /* Make subflows follow along. If we do not do this, we
1837 * get drops at subflow level if skbs can't be moved to
1838 * the mptcp rx queue fast enough (announced rcv_win can
1839 * exceed ssk->sk_rcvbuf).
1841 mptcp_for_each_subflow(msk, subflow) {
1845 ssk = mptcp_subflow_tcp_sock(subflow);
1846 slow = lock_sock_fast(ssk);
1847 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1848 tcp_sk(ssk)->window_clamp = window_clamp;
1849 tcp_cleanup_rbuf(ssk, 1);
1850 unlock_sock_fast(ssk, slow);
1855 msk->rcvq_space.space = msk->rcvq_space.copied;
1857 msk->rcvq_space.copied = 0;
1858 msk->rcvq_space.time = mstamp;
1861 static void __mptcp_update_rmem(struct sock *sk)
1863 struct mptcp_sock *msk = mptcp_sk(sk);
1865 if (!msk->rmem_released)
1868 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1869 sk_mem_uncharge(sk, msk->rmem_released);
1870 msk->rmem_released = 0;
1873 static void __mptcp_splice_receive_queue(struct sock *sk)
1875 struct mptcp_sock *msk = mptcp_sk(sk);
1877 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1880 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1882 struct sock *sk = (struct sock *)msk;
1883 unsigned int moved = 0;
1886 __mptcp_flush_join_list(msk);
1888 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1891 /* we can have data pending in the subflows only if the msk
1892 * receive buffer was full at subflow_data_ready() time,
1893 * that is an unlikely slow path.
1898 slowpath = lock_sock_fast(ssk);
1899 mptcp_data_lock(sk);
1900 __mptcp_update_rmem(sk);
1901 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1902 mptcp_data_unlock(sk);
1903 tcp_cleanup_rbuf(ssk, moved);
1904 unlock_sock_fast(ssk, slowpath);
1907 /* acquire the data lock only if some input data is pending */
1909 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1910 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1911 mptcp_data_lock(sk);
1912 __mptcp_update_rmem(sk);
1913 ret |= __mptcp_ofo_queue(msk);
1914 __mptcp_splice_receive_queue(sk);
1915 mptcp_data_unlock(sk);
1916 mptcp_cleanup_rbuf(msk);
1919 mptcp_check_data_fin((struct sock *)msk);
1920 return !skb_queue_empty(&msk->receive_queue);
1923 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1924 int nonblock, int flags, int *addr_len)
1926 struct mptcp_sock *msk = mptcp_sk(sk);
1931 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1934 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1935 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1940 timeo = sock_rcvtimeo(sk, nonblock);
1942 len = min_t(size_t, len, INT_MAX);
1943 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1945 while (copied < len) {
1948 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1949 if (unlikely(bytes_read < 0)) {
1951 copied = bytes_read;
1955 copied += bytes_read;
1957 /* be sure to advertise window change */
1958 mptcp_cleanup_rbuf(msk);
1960 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
1963 /* only the master socket status is relevant here. The exit
1964 * conditions mirror closely tcp_recvmsg()
1966 if (copied >= target)
1971 sk->sk_state == TCP_CLOSE ||
1972 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1974 signal_pending(current))
1978 copied = sock_error(sk);
1982 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1983 mptcp_check_for_eof(msk);
1985 if (sk->sk_shutdown & RCV_SHUTDOWN) {
1986 /* race breaker: the shutdown could be after the
1987 * previous receive queue check
1989 if (__mptcp_move_skbs(msk))
1994 if (sk->sk_state == TCP_CLOSE) {
2004 if (signal_pending(current)) {
2005 copied = sock_intr_errno(timeo);
2010 pr_debug("block timeout %ld", timeo);
2011 mptcp_wait_data(sk, &timeo);
2014 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2015 skb_queue_empty(&msk->receive_queue)) {
2016 /* entire backlog drained, clear DATA_READY. */
2017 clear_bit(MPTCP_DATA_READY, &msk->flags);
2019 /* .. race-breaker: ssk might have gotten new data
2020 * after last __mptcp_move_skbs() returned false.
2022 if (unlikely(__mptcp_move_skbs(msk)))
2023 set_bit(MPTCP_DATA_READY, &msk->flags);
2024 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2025 /* data to read but mptcp_wait_data() cleared DATA_READY */
2026 set_bit(MPTCP_DATA_READY, &msk->flags);
2029 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2030 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2031 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2032 mptcp_rcv_space_adjust(msk, copied);
2038 static void mptcp_retransmit_timer(struct timer_list *t)
2040 struct inet_connection_sock *icsk = from_timer(icsk, t,
2041 icsk_retransmit_timer);
2042 struct sock *sk = &icsk->icsk_inet.sk;
2043 struct mptcp_sock *msk = mptcp_sk(sk);
2046 if (!sock_owned_by_user(sk)) {
2047 /* we need a process context to retransmit */
2048 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2049 mptcp_schedule_work(sk);
2051 /* delegate our work to tcp_release_cb() */
2052 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2058 static void mptcp_timeout_timer(struct timer_list *t)
2060 struct sock *sk = from_timer(sk, t, sk_timer);
2062 mptcp_schedule_work(sk);
2066 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2069 * A backup subflow is returned only if that is the only kind available.
2071 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2073 struct mptcp_subflow_context *subflow;
2074 struct sock *backup = NULL;
2076 sock_owned_by_me((const struct sock *)msk);
2078 if (__mptcp_check_fallback(msk))
2081 mptcp_for_each_subflow(msk, subflow) {
2082 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2084 if (!mptcp_subflow_active(subflow))
2087 /* still data outstanding at TCP level? Don't retransmit. */
2088 if (!tcp_write_queue_empty(ssk)) {
2089 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2094 if (subflow->backup) {
2106 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2109 iput(SOCK_INODE(msk->subflow));
2110 msk->subflow = NULL;
2114 /* subflow sockets can be either outgoing (connect) or incoming
2117 * Outgoing subflows use in-kernel sockets.
2118 * Incoming subflows do not have their own 'struct socket' allocated,
2119 * so we need to use tcp_close() after detaching them from the mptcp
2122 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2123 struct mptcp_subflow_context *subflow)
2125 struct mptcp_sock *msk = mptcp_sk(sk);
2127 list_del(&subflow->node);
2129 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2131 /* if we are invoked by the msk cleanup code, the subflow is
2137 subflow->disposable = 1;
2139 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2140 * the ssk has been already destroyed, we just need to release the
2141 * reference owned by msk;
2143 if (!inet_csk(ssk)->icsk_ulp_ops) {
2144 kfree_rcu(subflow, rcu);
2146 /* otherwise tcp will dispose of the ssk and subflow ctx */
2147 __tcp_close(ssk, 0);
2149 /* close acquired an extra ref */
2156 if (ssk == msk->last_snd)
2157 msk->last_snd = NULL;
2159 if (ssk == msk->ack_hint)
2160 msk->ack_hint = NULL;
2162 if (ssk == msk->first)
2165 if (msk->subflow && ssk == msk->subflow->sk)
2166 mptcp_dispose_initial_subflow(msk);
2169 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2170 struct mptcp_subflow_context *subflow)
2172 if (sk->sk_state == TCP_ESTABLISHED)
2173 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2174 __mptcp_close_ssk(sk, ssk, subflow);
2177 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2182 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2184 struct mptcp_subflow_context *subflow, *tmp;
2188 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2189 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2191 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2194 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2195 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2198 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2202 static bool mptcp_check_close_timeout(const struct sock *sk)
2204 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2205 struct mptcp_subflow_context *subflow;
2207 if (delta >= TCP_TIMEWAIT_LEN)
2210 /* if all subflows are in closed status don't bother with additional
2213 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2214 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2221 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2223 struct mptcp_subflow_context *subflow, *tmp;
2224 struct sock *sk = &msk->sk.icsk_inet.sk;
2226 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2229 mptcp_token_destroy(msk);
2231 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2232 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2235 if (tcp_sk->sk_state != TCP_CLOSE) {
2236 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2237 tcp_set_state(tcp_sk, TCP_CLOSE);
2239 release_sock(tcp_sk);
2242 inet_sk_state_store(sk, TCP_CLOSE);
2243 sk->sk_shutdown = SHUTDOWN_MASK;
2244 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2245 set_bit(MPTCP_DATA_READY, &msk->flags);
2246 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2248 mptcp_close_wake_up(sk);
2251 static void __mptcp_retrans(struct sock *sk)
2253 struct mptcp_sock *msk = mptcp_sk(sk);
2254 struct mptcp_sendmsg_info info = {};
2255 struct mptcp_data_frag *dfrag;
2260 __mptcp_clean_una_wakeup(sk);
2261 dfrag = mptcp_rtx_head(sk);
2265 ssk = mptcp_subflow_get_retrans(msk);
2271 /* limit retransmission to the bytes already sent on some subflows */
2273 info.limit = dfrag->already_sent;
2274 while (info.sent < dfrag->already_sent) {
2275 if (!mptcp_alloc_tx_skb(sk, ssk))
2278 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2282 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2287 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2290 mptcp_set_timeout(sk, ssk);
2294 if (!mptcp_timer_pending(sk))
2295 mptcp_reset_timer(sk);
2298 static void mptcp_worker(struct work_struct *work)
2300 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2301 struct sock *sk = &msk->sk.icsk_inet.sk;
2305 state = sk->sk_state;
2306 if (unlikely(state == TCP_CLOSE))
2309 mptcp_check_data_fin_ack(sk);
2310 __mptcp_flush_join_list(msk);
2312 mptcp_check_fastclose(msk);
2315 mptcp_pm_nl_work(msk);
2317 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2318 mptcp_check_for_eof(msk);
2320 __mptcp_check_send_data_fin(sk);
2321 mptcp_check_data_fin(sk);
2323 /* There is no point in keeping around an orphaned sk timedout or
2324 * closed, but we need the msk around to reply to incoming DATA_FIN,
2325 * even if it is orphaned and in FIN_WAIT2 state
2327 if (sock_flag(sk, SOCK_DEAD) &&
2328 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2329 inet_sk_state_store(sk, TCP_CLOSE);
2330 __mptcp_destroy_sock(sk);
2334 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2335 __mptcp_close_subflow(msk);
2337 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2338 __mptcp_retrans(sk);
2345 static int __mptcp_init_sock(struct sock *sk)
2347 struct mptcp_sock *msk = mptcp_sk(sk);
2349 spin_lock_init(&msk->join_list_lock);
2351 INIT_LIST_HEAD(&msk->conn_list);
2352 INIT_LIST_HEAD(&msk->join_list);
2353 INIT_LIST_HEAD(&msk->rtx_queue);
2354 INIT_WORK(&msk->work, mptcp_worker);
2355 __skb_queue_head_init(&msk->receive_queue);
2356 __skb_queue_head_init(&msk->skb_tx_cache);
2357 msk->out_of_order_queue = RB_ROOT;
2358 msk->first_pending = NULL;
2359 msk->wmem_reserved = 0;
2360 msk->rmem_released = 0;
2361 msk->tx_pending_data = 0;
2362 msk->size_goal_cache = TCP_BASE_MSS;
2364 msk->ack_hint = NULL;
2366 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2368 mptcp_pm_data_init(msk);
2370 /* re-use the csk retrans timer for MPTCP-level retrans */
2371 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2372 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2376 static int mptcp_init_sock(struct sock *sk)
2378 struct net *net = sock_net(sk);
2381 ret = __mptcp_init_sock(sk);
2385 if (!mptcp_is_enabled(net))
2386 return -ENOPROTOOPT;
2388 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2391 ret = __mptcp_socket_create(mptcp_sk(sk));
2395 sk_sockets_allocated_inc(sk);
2396 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2397 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2402 static void __mptcp_clear_xmit(struct sock *sk)
2404 struct mptcp_sock *msk = mptcp_sk(sk);
2405 struct mptcp_data_frag *dtmp, *dfrag;
2406 struct sk_buff *skb;
2408 WRITE_ONCE(msk->first_pending, NULL);
2409 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2410 dfrag_clear(sk, dfrag);
2411 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2412 sk->sk_forward_alloc += skb->truesize;
2417 static void mptcp_cancel_work(struct sock *sk)
2419 struct mptcp_sock *msk = mptcp_sk(sk);
2421 if (cancel_work_sync(&msk->work))
2425 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2429 switch (ssk->sk_state) {
2431 if (!(how & RCV_SHUTDOWN))
2435 tcp_disconnect(ssk, O_NONBLOCK);
2438 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2439 pr_debug("Fallback");
2440 ssk->sk_shutdown |= how;
2441 tcp_shutdown(ssk, how);
2443 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2444 mptcp_set_timeout(sk, ssk);
2453 static const unsigned char new_state[16] = {
2454 /* current state: new state: action: */
2455 [0 /* (Invalid) */] = TCP_CLOSE,
2456 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2457 [TCP_SYN_SENT] = TCP_CLOSE,
2458 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2459 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2460 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2461 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2462 [TCP_CLOSE] = TCP_CLOSE,
2463 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2464 [TCP_LAST_ACK] = TCP_LAST_ACK,
2465 [TCP_LISTEN] = TCP_CLOSE,
2466 [TCP_CLOSING] = TCP_CLOSING,
2467 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2470 static int mptcp_close_state(struct sock *sk)
2472 int next = (int)new_state[sk->sk_state];
2473 int ns = next & TCP_STATE_MASK;
2475 inet_sk_state_store(sk, ns);
2477 return next & TCP_ACTION_FIN;
2480 static void __mptcp_check_send_data_fin(struct sock *sk)
2482 struct mptcp_subflow_context *subflow;
2483 struct mptcp_sock *msk = mptcp_sk(sk);
2485 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2486 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2487 msk->snd_nxt, msk->write_seq);
2489 /* we still need to enqueue subflows or not really shutting down,
2492 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2493 mptcp_send_head(sk))
2496 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2498 /* fallback socket will not get data_fin/ack, can move to the next
2501 if (__mptcp_check_fallback(msk)) {
2502 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2503 inet_sk_state_store(sk, TCP_CLOSE);
2504 mptcp_close_wake_up(sk);
2505 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2506 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2510 __mptcp_flush_join_list(msk);
2511 mptcp_for_each_subflow(msk, subflow) {
2512 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2514 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2518 static void __mptcp_wr_shutdown(struct sock *sk)
2520 struct mptcp_sock *msk = mptcp_sk(sk);
2522 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2523 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2524 !!mptcp_send_head(sk));
2526 /* will be ignored by fallback sockets */
2527 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2528 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2530 __mptcp_check_send_data_fin(sk);
2533 static void __mptcp_destroy_sock(struct sock *sk)
2535 struct mptcp_subflow_context *subflow, *tmp;
2536 struct mptcp_sock *msk = mptcp_sk(sk);
2537 LIST_HEAD(conn_list);
2539 pr_debug("msk=%p", msk);
2543 /* be sure to always acquire the join list lock, to sync vs
2544 * mptcp_finish_join().
2546 spin_lock_bh(&msk->join_list_lock);
2547 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2548 spin_unlock_bh(&msk->join_list_lock);
2549 list_splice_init(&msk->conn_list, &conn_list);
2551 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2552 sk_stop_timer(sk, &sk->sk_timer);
2555 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2556 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2557 __mptcp_close_ssk(sk, ssk, subflow);
2560 sk->sk_prot->destroy(sk);
2562 WARN_ON_ONCE(msk->wmem_reserved);
2563 WARN_ON_ONCE(msk->rmem_released);
2564 sk_stream_kill_queues(sk);
2565 xfrm_sk_free_policy(sk);
2566 sk_refcnt_debug_release(sk);
2567 mptcp_dispose_initial_subflow(msk);
2571 static void mptcp_close(struct sock *sk, long timeout)
2573 struct mptcp_subflow_context *subflow;
2574 bool do_cancel_work = false;
2577 sk->sk_shutdown = SHUTDOWN_MASK;
2579 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2580 inet_sk_state_store(sk, TCP_CLOSE);
2584 if (mptcp_close_state(sk))
2585 __mptcp_wr_shutdown(sk);
2587 sk_stream_wait_close(sk, timeout);
2590 /* orphan all the subflows */
2591 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2592 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2593 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2594 bool slow = lock_sock_fast(ssk);
2597 unlock_sock_fast(ssk, slow);
2602 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2603 if (sk->sk_state == TCP_CLOSE) {
2604 __mptcp_destroy_sock(sk);
2605 do_cancel_work = true;
2607 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2611 mptcp_cancel_work(sk);
2613 if (mptcp_sk(sk)->token)
2614 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2619 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2621 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2622 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2623 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2625 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2626 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2629 msk6->saddr = ssk6->saddr;
2630 msk6->flow_label = ssk6->flow_label;
2634 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2635 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2636 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2637 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2638 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2639 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2642 static int mptcp_disconnect(struct sock *sk, int flags)
2644 struct mptcp_subflow_context *subflow;
2645 struct mptcp_sock *msk = mptcp_sk(sk);
2647 __mptcp_flush_join_list(msk);
2648 mptcp_for_each_subflow(msk, subflow) {
2649 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2652 tcp_disconnect(ssk, flags);
2658 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2659 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2661 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2663 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2667 struct sock *mptcp_sk_clone(const struct sock *sk,
2668 const struct mptcp_options_received *mp_opt,
2669 struct request_sock *req)
2671 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2672 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2673 struct mptcp_sock *msk;
2679 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2680 if (nsk->sk_family == AF_INET6)
2681 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2684 __mptcp_init_sock(nsk);
2686 msk = mptcp_sk(nsk);
2687 msk->local_key = subflow_req->local_key;
2688 msk->token = subflow_req->token;
2689 msk->subflow = NULL;
2690 WRITE_ONCE(msk->fully_established, false);
2692 msk->write_seq = subflow_req->idsn + 1;
2693 msk->snd_nxt = msk->write_seq;
2694 msk->snd_una = msk->write_seq;
2695 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2697 if (mp_opt->mp_capable) {
2698 msk->can_ack = true;
2699 msk->remote_key = mp_opt->sndr_key;
2700 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2702 WRITE_ONCE(msk->ack_seq, ack_seq);
2703 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2706 sock_reset_flag(nsk, SOCK_RCU_FREE);
2707 /* will be fully established after successful MPC subflow creation */
2708 inet_sk_state_store(nsk, TCP_SYN_RECV);
2710 security_inet_csk_clone(nsk, req);
2711 bh_unlock_sock(nsk);
2713 /* keep a single reference */
2718 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2720 const struct tcp_sock *tp = tcp_sk(ssk);
2722 msk->rcvq_space.copied = 0;
2723 msk->rcvq_space.rtt_us = 0;
2725 msk->rcvq_space.time = tp->tcp_mstamp;
2727 /* initial rcv_space offering made to peer */
2728 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2729 TCP_INIT_CWND * tp->advmss);
2730 if (msk->rcvq_space.space == 0)
2731 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2733 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2736 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2739 struct mptcp_sock *msk = mptcp_sk(sk);
2740 struct socket *listener;
2743 listener = __mptcp_nmpc_socket(msk);
2744 if (WARN_ON_ONCE(!listener)) {
2749 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2750 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2754 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2755 if (sk_is_mptcp(newsk)) {
2756 struct mptcp_subflow_context *subflow;
2757 struct sock *new_mptcp_sock;
2759 subflow = mptcp_subflow_ctx(newsk);
2760 new_mptcp_sock = subflow->conn;
2762 /* is_mptcp should be false if subflow->conn is missing, see
2763 * subflow_syn_recv_sock()
2765 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2766 tcp_sk(newsk)->is_mptcp = 0;
2770 /* acquire the 2nd reference for the owning socket */
2771 sock_hold(new_mptcp_sock);
2772 newsk = new_mptcp_sock;
2773 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2775 MPTCP_INC_STATS(sock_net(sk),
2776 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2782 void mptcp_destroy_common(struct mptcp_sock *msk)
2784 struct sock *sk = (struct sock *)msk;
2786 __mptcp_clear_xmit(sk);
2788 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2789 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2791 skb_rbtree_purge(&msk->out_of_order_queue);
2792 mptcp_token_destroy(msk);
2793 mptcp_pm_free_anno_list(msk);
2796 static void mptcp_destroy(struct sock *sk)
2798 struct mptcp_sock *msk = mptcp_sk(sk);
2800 mptcp_destroy_common(msk);
2801 sk_sockets_allocated_dec(sk);
2804 void __mptcp_data_acked(struct sock *sk)
2806 if (!sock_owned_by_user(sk))
2807 __mptcp_clean_una(sk);
2809 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2811 if (mptcp_pending_data_fin_ack(sk))
2812 mptcp_schedule_work(sk);
2815 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2817 if (!mptcp_send_head(sk))
2820 if (!sock_owned_by_user(sk)) {
2821 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2823 if (xmit_ssk == ssk)
2824 __mptcp_subflow_push_pending(sk, ssk);
2826 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
2828 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2832 /* processes deferred events and flush wmem */
2833 static void mptcp_release_cb(struct sock *sk)
2836 unsigned long flags = 0;
2838 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2839 flags |= BIT(MPTCP_PUSH_PENDING);
2840 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2841 flags |= BIT(MPTCP_RETRANSMIT);
2845 /* the following actions acquire the subflow socket lock
2847 * 1) can't be invoked in atomic scope
2848 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2849 * datapath acquires the msk socket spinlock while helding
2850 * the subflow socket lock
2853 spin_unlock_bh(&sk->sk_lock.slock);
2854 if (flags & BIT(MPTCP_PUSH_PENDING))
2855 __mptcp_push_pending(sk, 0);
2856 if (flags & BIT(MPTCP_RETRANSMIT))
2857 __mptcp_retrans(sk);
2860 spin_lock_bh(&sk->sk_lock.slock);
2863 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2864 __mptcp_clean_una_wakeup(sk);
2865 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2866 __mptcp_error_report(sk);
2868 /* push_pending may touch wmem_reserved, ensure we do the cleanup
2871 __mptcp_update_wmem(sk);
2872 __mptcp_update_rmem(sk);
2875 void mptcp_subflow_process_delegated(struct sock *ssk)
2877 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2878 struct sock *sk = subflow->conn;
2880 mptcp_data_lock(sk);
2881 if (!sock_owned_by_user(sk))
2882 __mptcp_subflow_push_pending(sk, ssk);
2884 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2885 mptcp_data_unlock(sk);
2886 mptcp_subflow_delegated_done(subflow);
2889 static int mptcp_hash(struct sock *sk)
2891 /* should never be called,
2892 * we hash the TCP subflows not the master socket
2898 static void mptcp_unhash(struct sock *sk)
2900 /* called from sk_common_release(), but nothing to do here */
2903 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2905 struct mptcp_sock *msk = mptcp_sk(sk);
2906 struct socket *ssock;
2908 ssock = __mptcp_nmpc_socket(msk);
2909 pr_debug("msk=%p, subflow=%p", msk, ssock);
2910 if (WARN_ON_ONCE(!ssock))
2913 return inet_csk_get_port(ssock->sk, snum);
2916 void mptcp_finish_connect(struct sock *ssk)
2918 struct mptcp_subflow_context *subflow;
2919 struct mptcp_sock *msk;
2923 subflow = mptcp_subflow_ctx(ssk);
2927 pr_debug("msk=%p, token=%u", sk, subflow->token);
2929 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2931 subflow->map_seq = ack_seq;
2932 subflow->map_subflow_seq = 1;
2934 /* the socket is not connected yet, no msk/subflow ops can access/race
2935 * accessing the field below
2937 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2938 WRITE_ONCE(msk->local_key, subflow->local_key);
2939 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2940 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2941 WRITE_ONCE(msk->ack_seq, ack_seq);
2942 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2943 WRITE_ONCE(msk->can_ack, 1);
2944 WRITE_ONCE(msk->snd_una, msk->write_seq);
2946 mptcp_pm_new_connection(msk, ssk, 0);
2948 mptcp_rcv_space_init(msk, ssk);
2951 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2953 write_lock_bh(&sk->sk_callback_lock);
2954 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2955 sk_set_socket(sk, parent);
2956 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2957 write_unlock_bh(&sk->sk_callback_lock);
2960 bool mptcp_finish_join(struct sock *ssk)
2962 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2963 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2964 struct sock *parent = (void *)msk;
2965 struct socket *parent_sock;
2968 pr_debug("msk=%p, subflow=%p", msk, subflow);
2970 /* mptcp socket already closing? */
2971 if (!mptcp_is_fully_established(parent)) {
2972 subflow->reset_reason = MPTCP_RST_EMPTCP;
2976 if (!msk->pm.server_side)
2979 if (!mptcp_pm_allow_new_subflow(msk)) {
2980 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
2984 /* active connections are already on conn_list, and we can't acquire
2986 * use the join list lock as synchronization point and double-check
2987 * msk status to avoid racing with __mptcp_destroy_sock()
2989 spin_lock_bh(&msk->join_list_lock);
2990 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2991 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
2992 list_add_tail(&subflow->node, &msk->join_list);
2995 spin_unlock_bh(&msk->join_list_lock);
2997 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3001 /* attach to msk socket only after we are sure he will deal with us
3004 parent_sock = READ_ONCE(parent->sk_socket);
3005 if (parent_sock && !ssk->sk_socket)
3006 mptcp_sock_graft(ssk, parent_sock);
3007 subflow->map_seq = READ_ONCE(msk->ack_seq);
3009 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3013 static void mptcp_shutdown(struct sock *sk, int how)
3015 pr_debug("sk=%p, how=%d", sk, how);
3017 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3018 __mptcp_wr_shutdown(sk);
3021 static struct proto mptcp_prot = {
3023 .owner = THIS_MODULE,
3024 .init = mptcp_init_sock,
3025 .disconnect = mptcp_disconnect,
3026 .close = mptcp_close,
3027 .accept = mptcp_accept,
3028 .setsockopt = mptcp_setsockopt,
3029 .getsockopt = mptcp_getsockopt,
3030 .shutdown = mptcp_shutdown,
3031 .destroy = mptcp_destroy,
3032 .sendmsg = mptcp_sendmsg,
3033 .recvmsg = mptcp_recvmsg,
3034 .release_cb = mptcp_release_cb,
3036 .unhash = mptcp_unhash,
3037 .get_port = mptcp_get_port,
3038 .sockets_allocated = &mptcp_sockets_allocated,
3039 .memory_allocated = &tcp_memory_allocated,
3040 .memory_pressure = &tcp_memory_pressure,
3041 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3042 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3043 .sysctl_mem = sysctl_tcp_mem,
3044 .obj_size = sizeof(struct mptcp_sock),
3045 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3046 .no_autobind = true,
3049 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3051 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3052 struct socket *ssock;
3055 lock_sock(sock->sk);
3056 ssock = __mptcp_nmpc_socket(msk);
3062 err = ssock->ops->bind(ssock, uaddr, addr_len);
3064 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3067 release_sock(sock->sk);
3071 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3072 struct mptcp_subflow_context *subflow)
3074 subflow->request_mptcp = 0;
3075 __mptcp_do_fallback(msk);
3078 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3079 int addr_len, int flags)
3081 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3082 struct mptcp_subflow_context *subflow;
3083 struct socket *ssock;
3086 lock_sock(sock->sk);
3087 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3088 /* pending connection or invalid state, let existing subflow
3091 ssock = msk->subflow;
3095 ssock = __mptcp_nmpc_socket(msk);
3101 mptcp_token_destroy(msk);
3102 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3103 subflow = mptcp_subflow_ctx(ssock->sk);
3104 #ifdef CONFIG_TCP_MD5SIG
3105 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3108 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3109 mptcp_subflow_early_fallback(msk, subflow);
3111 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3112 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3113 mptcp_subflow_early_fallback(msk, subflow);
3115 if (likely(!__mptcp_check_fallback(msk)))
3116 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3119 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3120 sock->state = ssock->state;
3122 /* on successful connect, the msk state will be moved to established by
3123 * subflow_finish_connect()
3125 if (!err || err == -EINPROGRESS)
3126 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3128 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3131 release_sock(sock->sk);
3135 static int mptcp_listen(struct socket *sock, int backlog)
3137 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3138 struct socket *ssock;
3141 pr_debug("msk=%p", msk);
3143 lock_sock(sock->sk);
3144 ssock = __mptcp_nmpc_socket(msk);
3150 mptcp_token_destroy(msk);
3151 inet_sk_state_store(sock->sk, TCP_LISTEN);
3152 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3154 err = ssock->ops->listen(ssock, backlog);
3155 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3157 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3160 release_sock(sock->sk);
3164 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3165 int flags, bool kern)
3167 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3168 struct socket *ssock;
3171 pr_debug("msk=%p", msk);
3173 lock_sock(sock->sk);
3174 if (sock->sk->sk_state != TCP_LISTEN)
3177 ssock = __mptcp_nmpc_socket(msk);
3181 clear_bit(MPTCP_DATA_READY, &msk->flags);
3182 sock_hold(ssock->sk);
3183 release_sock(sock->sk);
3185 err = ssock->ops->accept(sock, newsock, flags, kern);
3186 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3187 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3188 struct mptcp_subflow_context *subflow;
3189 struct sock *newsk = newsock->sk;
3193 /* PM/worker can now acquire the first subflow socket
3194 * lock without racing with listener queue cleanup,
3195 * we can notify it, if needed.
3197 * Even if remote has reset the initial subflow by now
3198 * the refcnt is still at least one.
3200 subflow = mptcp_subflow_ctx(msk->first);
3201 list_add(&subflow->node, &msk->conn_list);
3202 sock_hold(msk->first);
3203 if (mptcp_is_fully_established(newsk))
3204 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3206 mptcp_copy_inaddrs(newsk, msk->first);
3207 mptcp_rcv_space_init(msk, msk->first);
3208 mptcp_propagate_sndbuf(newsk, msk->first);
3210 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3211 * This is needed so NOSPACE flag can be set from tcp stack.
3213 __mptcp_flush_join_list(msk);
3214 mptcp_for_each_subflow(msk, subflow) {
3215 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3217 if (!ssk->sk_socket)
3218 mptcp_sock_graft(ssk, newsock);
3220 release_sock(newsk);
3223 if (inet_csk_listen_poll(ssock->sk))
3224 set_bit(MPTCP_DATA_READY, &msk->flags);
3225 sock_put(ssock->sk);
3229 release_sock(sock->sk);
3233 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3235 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3239 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3241 struct sock *sk = (struct sock *)msk;
3243 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3244 return EPOLLOUT | EPOLLWRNORM;
3246 if (sk_stream_is_writeable(sk))
3247 return EPOLLOUT | EPOLLWRNORM;
3249 mptcp_set_nospace(sk);
3250 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3251 if (sk_stream_is_writeable(sk))
3252 return EPOLLOUT | EPOLLWRNORM;
3257 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3258 struct poll_table_struct *wait)
3260 struct sock *sk = sock->sk;
3261 struct mptcp_sock *msk;
3266 sock_poll_wait(file, sock, wait);
3268 state = inet_sk_state_load(sk);
3269 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3270 if (state == TCP_LISTEN)
3271 return mptcp_check_readable(msk);
3273 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3274 mask |= mptcp_check_readable(msk);
3275 mask |= mptcp_check_writeable(msk);
3277 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3279 if (sk->sk_shutdown & RCV_SHUTDOWN)
3280 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3282 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3290 static const struct proto_ops mptcp_stream_ops = {
3292 .owner = THIS_MODULE,
3293 .release = inet_release,
3295 .connect = mptcp_stream_connect,
3296 .socketpair = sock_no_socketpair,
3297 .accept = mptcp_stream_accept,
3298 .getname = inet_getname,
3300 .ioctl = inet_ioctl,
3301 .gettstamp = sock_gettstamp,
3302 .listen = mptcp_listen,
3303 .shutdown = inet_shutdown,
3304 .setsockopt = sock_common_setsockopt,
3305 .getsockopt = sock_common_getsockopt,
3306 .sendmsg = inet_sendmsg,
3307 .recvmsg = inet_recvmsg,
3308 .mmap = sock_no_mmap,
3309 .sendpage = inet_sendpage,
3312 static struct inet_protosw mptcp_protosw = {
3313 .type = SOCK_STREAM,
3314 .protocol = IPPROTO_MPTCP,
3315 .prot = &mptcp_prot,
3316 .ops = &mptcp_stream_ops,
3317 .flags = INET_PROTOSW_ICSK,
3320 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3322 struct mptcp_delegated_action *delegated;
3323 struct mptcp_subflow_context *subflow;
3326 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3327 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3328 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3330 bh_lock_sock_nested(ssk);
3331 if (!sock_owned_by_user(ssk) &&
3332 mptcp_subflow_has_delegated_action(subflow))
3333 mptcp_subflow_process_delegated(ssk);
3334 /* ... elsewhere tcp_release_cb_override already processed
3335 * the action or will do at next release_sock().
3336 * In both case must dequeue the subflow here - on the same
3337 * CPU that scheduled it.
3339 bh_unlock_sock(ssk);
3342 if (++work_done == budget)
3346 /* always provide a 0 'work_done' argument, so that napi_complete_done
3347 * will not try accessing the NULL napi->dev ptr
3349 napi_complete_done(napi, 0);
3353 void __init mptcp_proto_init(void)
3355 struct mptcp_delegated_action *delegated;
3358 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3360 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3361 panic("Failed to allocate MPTCP pcpu counter\n");
3363 init_dummy_netdev(&mptcp_napi_dev);
3364 for_each_possible_cpu(cpu) {
3365 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3366 INIT_LIST_HEAD(&delegated->head);
3367 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3369 napi_enable(&delegated->napi);
3372 mptcp_subflow_init();
3376 if (proto_register(&mptcp_prot, 1) != 0)
3377 panic("Failed to register MPTCP proto.\n");
3379 inet_register_protosw(&mptcp_protosw);
3381 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3384 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3385 static const struct proto_ops mptcp_v6_stream_ops = {
3387 .owner = THIS_MODULE,
3388 .release = inet6_release,
3390 .connect = mptcp_stream_connect,
3391 .socketpair = sock_no_socketpair,
3392 .accept = mptcp_stream_accept,
3393 .getname = inet6_getname,
3395 .ioctl = inet6_ioctl,
3396 .gettstamp = sock_gettstamp,
3397 .listen = mptcp_listen,
3398 .shutdown = inet_shutdown,
3399 .setsockopt = sock_common_setsockopt,
3400 .getsockopt = sock_common_getsockopt,
3401 .sendmsg = inet6_sendmsg,
3402 .recvmsg = inet6_recvmsg,
3403 .mmap = sock_no_mmap,
3404 .sendpage = inet_sendpage,
3405 #ifdef CONFIG_COMPAT
3406 .compat_ioctl = inet6_compat_ioctl,
3410 static struct proto mptcp_v6_prot;
3412 static void mptcp_v6_destroy(struct sock *sk)
3415 inet6_destroy_sock(sk);
3418 static struct inet_protosw mptcp_v6_protosw = {
3419 .type = SOCK_STREAM,
3420 .protocol = IPPROTO_MPTCP,
3421 .prot = &mptcp_v6_prot,
3422 .ops = &mptcp_v6_stream_ops,
3423 .flags = INET_PROTOSW_ICSK,
3426 int __init mptcp_proto_v6_init(void)
3430 mptcp_v6_prot = mptcp_prot;
3431 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3432 mptcp_v6_prot.slab = NULL;
3433 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3434 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3436 err = proto_register(&mptcp_v6_prot, 1);
3440 err = inet6_register_protosw(&mptcp_v6_protosw);
3442 proto_unregister(&mptcp_v6_prot);