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 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
49 * completed yet or has failed, return the subflow socket.
50 * Otherwise return NULL.
52 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
54 if (!msk->subflow || READ_ONCE(msk->can_ack))
60 /* Returns end sequence number of the receiver's advertised window */
61 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
63 return READ_ONCE(msk->wnd_end);
66 static bool mptcp_is_tcpsk(struct sock *sk)
68 struct socket *sock = sk->sk_socket;
70 if (unlikely(sk->sk_prot == &tcp_prot)) {
71 /* we are being invoked after mptcp_accept() has
72 * accepted a non-mp-capable flow: sk is a tcp_sk,
75 * Hand the socket over to tcp so all further socket ops
78 sock->ops = &inet_stream_ops;
80 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
81 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
82 sock->ops = &inet6_stream_ops;
90 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
92 sock_owned_by_me((const struct sock *)msk);
94 if (likely(!__mptcp_check_fallback(msk)))
100 static int __mptcp_socket_create(struct mptcp_sock *msk)
102 struct mptcp_subflow_context *subflow;
103 struct sock *sk = (struct sock *)msk;
104 struct socket *ssock;
107 err = mptcp_subflow_create_socket(sk, &ssock);
111 msk->first = ssock->sk;
112 msk->subflow = ssock;
113 subflow = mptcp_subflow_ctx(ssock->sk);
114 list_add(&subflow->node, &msk->conn_list);
115 sock_hold(ssock->sk);
116 subflow->request_mptcp = 1;
118 /* accept() will wait on first subflow sk_wq, and we always wakes up
121 RCU_INIT_POINTER(msk->first->sk_wq, &sk->sk_socket->wq);
126 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
128 sk_drops_add(sk, skb);
132 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
133 struct sk_buff *from)
138 if (MPTCP_SKB_CB(from)->offset ||
139 !skb_try_coalesce(to, from, &fragstolen, &delta))
142 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
143 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
144 to->len, MPTCP_SKB_CB(from)->end_seq);
145 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
146 kfree_skb_partial(from, fragstolen);
147 atomic_add(delta, &sk->sk_rmem_alloc);
148 sk_mem_charge(sk, delta);
152 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
153 struct sk_buff *from)
155 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
158 return mptcp_try_coalesce((struct sock *)msk, to, from);
161 /* "inspired" by tcp_data_queue_ofo(), main differences:
163 * - don't cope with sacks
165 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
167 struct sock *sk = (struct sock *)msk;
168 struct rb_node **p, *parent;
169 u64 seq, end_seq, max_seq;
170 struct sk_buff *skb1;
172 seq = MPTCP_SKB_CB(skb)->map_seq;
173 end_seq = MPTCP_SKB_CB(skb)->end_seq;
174 max_seq = READ_ONCE(msk->rcv_wnd_sent);
176 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
177 RB_EMPTY_ROOT(&msk->out_of_order_queue));
178 if (after64(end_seq, max_seq)) {
181 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
182 (unsigned long long)end_seq - (unsigned long)max_seq,
183 (unsigned long long)msk->rcv_wnd_sent);
184 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
188 p = &msk->out_of_order_queue.rb_node;
189 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
190 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
191 rb_link_node(&skb->rbnode, NULL, p);
192 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
193 msk->ooo_last_skb = skb;
197 /* with 2 subflows, adding at end of ooo queue is quite likely
198 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
200 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
201 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
202 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
206 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
207 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
208 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
209 parent = &msk->ooo_last_skb->rbnode;
210 p = &parent->rb_right;
214 /* Find place to insert this segment. Handle overlaps on the way. */
218 skb1 = rb_to_skb(parent);
219 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
220 p = &parent->rb_left;
223 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
224 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
225 /* All the bits are present. Drop. */
227 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
230 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
234 * continue traversing
237 /* skb's seq == skb1's seq and skb covers skb1.
238 * Replace skb1 with skb.
240 rb_replace_node(&skb1->rbnode, &skb->rbnode,
241 &msk->out_of_order_queue);
242 mptcp_drop(sk, skb1);
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
246 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
250 p = &parent->rb_right;
254 /* Insert segment into RB tree. */
255 rb_link_node(&skb->rbnode, parent, p);
256 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
259 /* Remove other segments covered by skb. */
260 while ((skb1 = skb_rb_next(skb)) != NULL) {
261 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
263 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
264 mptcp_drop(sk, skb1);
265 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
267 /* If there is no skb after us, we are the last_skb ! */
269 msk->ooo_last_skb = skb;
273 skb_set_owner_r(skb, sk);
276 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
277 struct sk_buff *skb, unsigned int offset,
280 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
281 struct sock *sk = (struct sock *)msk;
282 struct sk_buff *tail;
284 __skb_unlink(skb, &ssk->sk_receive_queue);
289 /* try to fetch required memory from subflow */
290 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
291 if (ssk->sk_forward_alloc < skb->truesize)
293 __sk_mem_reclaim(ssk, skb->truesize);
294 if (!sk_rmem_schedule(sk, skb, skb->truesize))
298 /* the skb map_seq accounts for the skb offset:
299 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
302 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
303 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
304 MPTCP_SKB_CB(skb)->offset = offset;
306 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
308 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
309 tail = skb_peek_tail(&sk->sk_receive_queue);
310 if (tail && mptcp_try_coalesce(sk, tail, skb))
313 skb_set_owner_r(skb, sk);
314 __skb_queue_tail(&sk->sk_receive_queue, skb);
316 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
317 mptcp_data_queue_ofo(msk, skb);
321 /* old data, keep it simple and drop the whole pkt, sender
322 * will retransmit as needed, if needed.
324 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
330 static void mptcp_stop_timer(struct sock *sk)
332 struct inet_connection_sock *icsk = inet_csk(sk);
334 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
335 mptcp_sk(sk)->timer_ival = 0;
338 static void mptcp_close_wake_up(struct sock *sk)
340 if (sock_flag(sk, SOCK_DEAD))
343 sk->sk_state_change(sk);
344 if (sk->sk_shutdown == SHUTDOWN_MASK ||
345 sk->sk_state == TCP_CLOSE)
346 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
348 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
351 static bool mptcp_pending_data_fin_ack(struct sock *sk)
353 struct mptcp_sock *msk = mptcp_sk(sk);
355 return !__mptcp_check_fallback(msk) &&
356 ((1 << sk->sk_state) &
357 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
358 msk->write_seq == READ_ONCE(msk->snd_una);
361 static void mptcp_check_data_fin_ack(struct sock *sk)
363 struct mptcp_sock *msk = mptcp_sk(sk);
365 /* Look for an acknowledged DATA_FIN */
366 if (mptcp_pending_data_fin_ack(sk)) {
367 mptcp_stop_timer(sk);
369 WRITE_ONCE(msk->snd_data_fin_enable, 0);
371 switch (sk->sk_state) {
373 inet_sk_state_store(sk, TCP_FIN_WAIT2);
377 inet_sk_state_store(sk, TCP_CLOSE);
381 mptcp_close_wake_up(sk);
385 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
387 struct mptcp_sock *msk = mptcp_sk(sk);
389 if (READ_ONCE(msk->rcv_data_fin) &&
390 ((1 << sk->sk_state) &
391 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
392 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
394 if (msk->ack_seq == rcv_data_fin_seq) {
396 *seq = rcv_data_fin_seq;
405 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
407 long tout = ssk && inet_csk(ssk)->icsk_pending ?
408 inet_csk(ssk)->icsk_timeout - jiffies : 0;
411 tout = mptcp_sk(sk)->timer_ival;
412 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
415 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
417 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
419 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
420 if (subflow->request_join && !subflow->fully_established)
423 /* only send if our side has not closed yet */
424 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
427 static bool tcp_can_send_ack(const struct sock *ssk)
429 return !((1 << inet_sk_state_load(ssk)) &
430 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE));
433 static void mptcp_send_ack(struct mptcp_sock *msk)
435 struct mptcp_subflow_context *subflow;
437 mptcp_for_each_subflow(msk, subflow) {
438 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
441 if (tcp_can_send_ack(ssk))
447 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
452 ret = tcp_can_send_ack(ssk);
454 tcp_cleanup_rbuf(ssk, 1);
459 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
461 struct sock *ack_hint = READ_ONCE(msk->ack_hint);
462 struct mptcp_subflow_context *subflow;
464 /* if the hinted ssk is still active, try to use it */
465 if (likely(ack_hint)) {
466 mptcp_for_each_subflow(msk, subflow) {
467 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
469 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk))
474 /* otherwise pick the first active subflow */
475 mptcp_for_each_subflow(msk, subflow)
476 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
480 static bool mptcp_check_data_fin(struct sock *sk)
482 struct mptcp_sock *msk = mptcp_sk(sk);
483 u64 rcv_data_fin_seq;
486 if (__mptcp_check_fallback(msk) || !msk->first)
489 /* Need to ack a DATA_FIN received from a peer while this side
490 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
491 * msk->rcv_data_fin was set when parsing the incoming options
492 * at the subflow level and the msk lock was not held, so this
493 * is the first opportunity to act on the DATA_FIN and change
496 * If we are caught up to the sequence number of the incoming
497 * DATA_FIN, send the DATA_ACK now and do state transition. If
498 * not caught up, do nothing and let the recv code send DATA_ACK
502 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
503 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
504 WRITE_ONCE(msk->rcv_data_fin, 0);
506 sk->sk_shutdown |= RCV_SHUTDOWN;
507 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
508 set_bit(MPTCP_DATA_READY, &msk->flags);
510 switch (sk->sk_state) {
511 case TCP_ESTABLISHED:
512 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
515 inet_sk_state_store(sk, TCP_CLOSING);
518 inet_sk_state_store(sk, TCP_CLOSE);
521 /* Other states not expected */
527 mptcp_set_timeout(sk, NULL);
529 mptcp_close_wake_up(sk);
534 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
538 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
539 struct sock *sk = (struct sock *)msk;
540 unsigned int moved = 0;
541 bool more_data_avail;
546 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
548 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
549 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
551 if (unlikely(ssk_rbuf > sk_rbuf)) {
552 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
557 pr_debug("msk=%p ssk=%p", msk, ssk);
560 u32 map_remaining, offset;
561 u32 seq = tp->copied_seq;
565 /* try to move as much data as available */
566 map_remaining = subflow->map_data_len -
567 mptcp_subflow_get_map_offset(subflow);
569 skb = skb_peek(&ssk->sk_receive_queue);
571 /* if no data is found, a racing workqueue/recvmsg
572 * already processed the new data, stop here or we
573 * can enter an infinite loop
580 if (__mptcp_check_fallback(msk)) {
581 /* if we are running under the workqueue, TCP could have
582 * collapsed skbs between dummy map creation and now
583 * be sure to adjust the size
585 map_remaining = skb->len;
586 subflow->map_data_len = skb->len;
589 offset = seq - TCP_SKB_CB(skb)->seq;
590 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
596 if (offset < skb->len) {
597 size_t len = skb->len - offset;
602 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
606 if (WARN_ON_ONCE(map_remaining < len))
610 sk_eat_skb(ssk, skb);
614 WRITE_ONCE(tp->copied_seq, seq);
615 more_data_avail = mptcp_subflow_data_available(ssk);
617 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
621 } while (more_data_avail);
622 WRITE_ONCE(msk->ack_hint, ssk);
628 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
630 struct sock *sk = (struct sock *)msk;
631 struct sk_buff *skb, *tail;
636 p = rb_first(&msk->out_of_order_queue);
637 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
640 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
644 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
646 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
649 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
653 end_seq = MPTCP_SKB_CB(skb)->end_seq;
654 tail = skb_peek_tail(&sk->sk_receive_queue);
655 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
656 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
658 /* skip overlapping data, if any */
659 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
660 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
662 MPTCP_SKB_CB(skb)->offset += delta;
663 __skb_queue_tail(&sk->sk_receive_queue, skb);
665 msk->ack_seq = end_seq;
671 /* In most cases we will be able to lock the mptcp socket. If its already
672 * owned, we need to defer to the work queue to avoid ABBA deadlock.
674 static void move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
676 struct sock *sk = (struct sock *)msk;
677 unsigned int moved = 0;
679 if (inet_sk_state_load(sk) == TCP_CLOSE)
684 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
685 __mptcp_ofo_queue(msk);
687 /* If the moves have caught up with the DATA_FIN sequence number
688 * it's time to ack the DATA_FIN and change socket state, but
689 * this is not a good place to change state. Let the workqueue
692 if (mptcp_pending_data_fin(sk, NULL))
693 mptcp_schedule_work(sk);
694 mptcp_data_unlock(sk);
697 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
699 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
700 struct mptcp_sock *msk = mptcp_sk(sk);
701 int sk_rbuf, ssk_rbuf;
704 /* The peer can send data while we are shutting down this
705 * subflow at msk destruction time, but we must avoid enqueuing
706 * more data to the msk receive queue
708 if (unlikely(subflow->disposable))
711 /* move_skbs_to_msk below can legitly clear the data_avail flag,
712 * but we will need later to properly woke the reader, cache its
715 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
717 set_bit(MPTCP_DATA_READY, &msk->flags);
719 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
720 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
721 if (unlikely(ssk_rbuf > sk_rbuf))
724 /* over limit? can't append more skbs to msk */
725 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
728 move_skbs_to_msk(msk, ssk);
732 sk->sk_data_ready(sk);
735 void __mptcp_flush_join_list(struct mptcp_sock *msk)
737 if (likely(list_empty(&msk->join_list)))
740 spin_lock_bh(&msk->join_list_lock);
741 list_splice_tail_init(&msk->join_list, &msk->conn_list);
742 spin_unlock_bh(&msk->join_list_lock);
745 static bool mptcp_timer_pending(struct sock *sk)
747 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
750 static void mptcp_reset_timer(struct sock *sk)
752 struct inet_connection_sock *icsk = inet_csk(sk);
755 /* prevent rescheduling on close */
756 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
759 /* should never be called with mptcp level timer cleared */
760 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
761 if (WARN_ON_ONCE(!tout))
763 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
766 bool mptcp_schedule_work(struct sock *sk)
768 if (inet_sk_state_load(sk) != TCP_CLOSE &&
769 schedule_work(&mptcp_sk(sk)->work)) {
770 /* each subflow already holds a reference to the sk, and the
771 * workqueue is invoked by a subflow, so sk can't go away here.
779 void mptcp_subflow_eof(struct sock *sk)
781 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
782 mptcp_schedule_work(sk);
785 static void mptcp_check_for_eof(struct mptcp_sock *msk)
787 struct mptcp_subflow_context *subflow;
788 struct sock *sk = (struct sock *)msk;
791 mptcp_for_each_subflow(msk, subflow)
792 receivers += !subflow->rx_eof;
796 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
797 /* hopefully temporary hack: propagate shutdown status
798 * to msk, when all subflows agree on it
800 sk->sk_shutdown |= RCV_SHUTDOWN;
802 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
803 set_bit(MPTCP_DATA_READY, &msk->flags);
804 sk->sk_data_ready(sk);
807 switch (sk->sk_state) {
808 case TCP_ESTABLISHED:
809 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
812 inet_sk_state_store(sk, TCP_CLOSING);
815 inet_sk_state_store(sk, TCP_CLOSE);
820 mptcp_close_wake_up(sk);
823 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
825 struct mptcp_subflow_context *subflow;
826 struct sock *sk = (struct sock *)msk;
828 sock_owned_by_me(sk);
830 mptcp_for_each_subflow(msk, subflow) {
831 if (subflow->data_avail)
832 return mptcp_subflow_tcp_sock(subflow);
838 static bool mptcp_skb_can_collapse_to(u64 write_seq,
839 const struct sk_buff *skb,
840 const struct mptcp_ext *mpext)
842 if (!tcp_skb_can_collapse_to(skb))
845 /* can collapse only if MPTCP level sequence is in order and this
846 * mapping has not been xmitted yet
848 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
852 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
853 const struct page_frag *pfrag,
854 const struct mptcp_data_frag *df)
856 return df && pfrag->page == df->page &&
857 pfrag->size - pfrag->offset > 0 &&
858 df->data_seq + df->data_len == msk->write_seq;
861 static int mptcp_wmem_with_overhead(struct sock *sk, int size)
863 struct mptcp_sock *msk = mptcp_sk(sk);
866 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
867 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache;
868 if (skbs < msk->skb_tx_cache.qlen)
871 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER);
874 static void __mptcp_wmem_reserve(struct sock *sk, int size)
876 int amount = mptcp_wmem_with_overhead(sk, size);
877 struct mptcp_sock *msk = mptcp_sk(sk);
879 WARN_ON_ONCE(msk->wmem_reserved);
880 if (amount <= sk->sk_forward_alloc)
883 /* under memory pressure try to reserve at most a single page
884 * otherwise try to reserve the full estimate and fallback
885 * to a single page before entering the error path
887 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
888 !sk_wmem_schedule(sk, amount)) {
889 if (amount <= PAGE_SIZE)
893 if (!sk_wmem_schedule(sk, amount))
898 msk->wmem_reserved = amount;
899 sk->sk_forward_alloc -= amount;
903 /* we will wait for memory on next allocation */
904 msk->wmem_reserved = -1;
907 static void __mptcp_update_wmem(struct sock *sk)
909 struct mptcp_sock *msk = mptcp_sk(sk);
911 if (!msk->wmem_reserved)
914 if (msk->wmem_reserved < 0)
915 msk->wmem_reserved = 0;
916 if (msk->wmem_reserved > 0) {
917 sk->sk_forward_alloc += msk->wmem_reserved;
918 msk->wmem_reserved = 0;
922 static bool mptcp_wmem_alloc(struct sock *sk, int size)
924 struct mptcp_sock *msk = mptcp_sk(sk);
926 /* check for pre-existing error condition */
927 if (msk->wmem_reserved < 0)
930 if (msk->wmem_reserved >= size)
934 if (!sk_wmem_schedule(sk, size)) {
935 mptcp_data_unlock(sk);
939 sk->sk_forward_alloc -= size;
940 msk->wmem_reserved += size;
941 mptcp_data_unlock(sk);
944 msk->wmem_reserved -= size;
948 static void mptcp_wmem_uncharge(struct sock *sk, int size)
950 struct mptcp_sock *msk = mptcp_sk(sk);
952 if (msk->wmem_reserved < 0)
953 msk->wmem_reserved = 0;
954 msk->wmem_reserved += size;
957 static void mptcp_mem_reclaim_partial(struct sock *sk)
959 struct mptcp_sock *msk = mptcp_sk(sk);
961 /* if we are experiencing a transint allocation error,
962 * the forward allocation memory has been already
965 if (msk->wmem_reserved < 0)
969 sk->sk_forward_alloc += msk->wmem_reserved;
970 sk_mem_reclaim_partial(sk);
971 msk->wmem_reserved = sk->sk_forward_alloc;
972 sk->sk_forward_alloc = 0;
973 mptcp_data_unlock(sk);
976 static void dfrag_uncharge(struct sock *sk, int len)
978 sk_mem_uncharge(sk, len);
979 sk_wmem_queued_add(sk, -len);
982 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
984 int len = dfrag->data_len + dfrag->overhead;
986 list_del(&dfrag->list);
987 dfrag_uncharge(sk, len);
988 put_page(dfrag->page);
991 static void __mptcp_clean_una(struct sock *sk)
993 struct mptcp_sock *msk = mptcp_sk(sk);
994 struct mptcp_data_frag *dtmp, *dfrag;
995 bool cleaned = false;
998 /* on fallback we just need to ignore snd_una, as this is really
1001 if (__mptcp_check_fallback(msk))
1002 msk->snd_una = READ_ONCE(msk->snd_nxt);
1004 snd_una = msk->snd_una;
1005 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1006 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1009 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1011 dfrag_clear(sk, dfrag);
1015 dfrag = mptcp_rtx_head(sk);
1016 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1017 u64 delta = snd_una - dfrag->data_seq;
1019 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1022 dfrag->data_seq += delta;
1023 dfrag->offset += delta;
1024 dfrag->data_len -= delta;
1025 dfrag->already_sent -= delta;
1027 dfrag_uncharge(sk, delta);
1033 if (tcp_under_memory_pressure(sk)) {
1034 __mptcp_update_wmem(sk);
1035 sk_mem_reclaim_partial(sk);
1038 if (sk_stream_is_writeable(sk)) {
1039 /* pairs with memory barrier in mptcp_poll */
1041 if (test_and_clear_bit(MPTCP_NOSPACE, &msk->flags))
1042 sk_stream_write_space(sk);
1046 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1047 if (msk->timer_ival)
1048 mptcp_stop_timer(sk);
1050 mptcp_reset_timer(sk);
1054 static void mptcp_enter_memory_pressure(struct sock *sk)
1056 struct mptcp_subflow_context *subflow;
1057 struct mptcp_sock *msk = mptcp_sk(sk);
1060 sk_stream_moderate_sndbuf(sk);
1061 mptcp_for_each_subflow(msk, subflow) {
1062 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1065 tcp_enter_memory_pressure(ssk);
1066 sk_stream_moderate_sndbuf(ssk);
1071 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1074 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1076 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1077 pfrag, sk->sk_allocation)))
1080 mptcp_enter_memory_pressure(sk);
1084 static struct mptcp_data_frag *
1085 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1088 int offset = ALIGN(orig_offset, sizeof(long));
1089 struct mptcp_data_frag *dfrag;
1091 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1092 dfrag->data_len = 0;
1093 dfrag->data_seq = msk->write_seq;
1094 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1095 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1096 dfrag->already_sent = 0;
1097 dfrag->page = pfrag->page;
1102 struct mptcp_sendmsg_info {
1110 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1113 u64 window_end = mptcp_wnd_end(msk);
1115 if (__mptcp_check_fallback(msk))
1118 if (!before64(data_seq + avail_size, window_end)) {
1119 u64 allowed_size = window_end - data_seq;
1121 return min_t(unsigned int, allowed_size, avail_size);
1127 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1129 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1133 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1137 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1139 struct sk_buff *skb;
1141 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1143 if (likely(__mptcp_add_ext(skb, gfp))) {
1144 skb_reserve(skb, MAX_TCP_HEADER);
1145 skb->reserved_tailroom = skb->end - skb->tail;
1150 mptcp_enter_memory_pressure(sk);
1155 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1156 struct sk_buff_head *skbs, int *total_ts)
1158 struct mptcp_sock *msk = mptcp_sk(sk);
1159 struct sk_buff *skb;
1162 if (unlikely(tcp_under_memory_pressure(sk))) {
1163 mptcp_mem_reclaim_partial(sk);
1165 /* under pressure pre-allocate at most a single skb */
1166 if (msk->skb_tx_cache.qlen)
1168 space_needed = msk->size_goal_cache;
1170 space_needed = msk->tx_pending_data + size -
1171 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1174 while (space_needed > 0) {
1175 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1176 if (unlikely(!skb)) {
1177 /* under memory pressure, try to pass the caller a
1178 * single skb to allow forward progress
1180 while (skbs->qlen > 1) {
1181 skb = __skb_dequeue_tail(skbs);
1184 return skbs->qlen > 0;
1187 *total_ts += skb->truesize;
1188 __skb_queue_tail(skbs, skb);
1189 space_needed -= msk->size_goal_cache;
1194 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1196 struct mptcp_sock *msk = mptcp_sk(sk);
1197 struct sk_buff *skb;
1199 if (ssk->sk_tx_skb_cache) {
1200 skb = ssk->sk_tx_skb_cache;
1201 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1202 !__mptcp_add_ext(skb, gfp)))
1207 skb = skb_peek(&msk->skb_tx_cache);
1209 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1210 skb = __skb_dequeue(&msk->skb_tx_cache);
1211 if (WARN_ON_ONCE(!skb))
1214 mptcp_wmem_uncharge(sk, skb->truesize);
1215 ssk->sk_tx_skb_cache = skb;
1219 /* over memory limit, no point to try to allocate a new skb */
1223 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1227 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1228 ssk->sk_tx_skb_cache = skb;
1235 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1237 return !ssk->sk_tx_skb_cache &&
1238 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1239 tcp_under_memory_pressure(sk);
1242 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1244 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1245 mptcp_mem_reclaim_partial(sk);
1246 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1249 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1250 struct mptcp_data_frag *dfrag,
1251 struct mptcp_sendmsg_info *info)
1253 u64 data_seq = dfrag->data_seq + info->sent;
1254 struct mptcp_sock *msk = mptcp_sk(sk);
1255 bool zero_window_probe = false;
1256 struct mptcp_ext *mpext = NULL;
1257 struct sk_buff *skb, *tail;
1258 bool can_collapse = false;
1263 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
1264 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1266 /* compute send limit */
1267 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1268 avail_size = info->size_goal;
1269 msk->size_goal_cache = info->size_goal;
1270 skb = tcp_write_queue_tail(ssk);
1272 /* Limit the write to the size available in the
1273 * current skb, if any, so that we create at most a new skb.
1274 * Explicitly tells TCP internals to avoid collapsing on later
1275 * queue management operation, to avoid breaking the ext <->
1276 * SSN association set here
1278 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1279 can_collapse = (info->size_goal - skb->len > 0) &&
1280 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1281 if (!can_collapse) {
1282 TCP_SKB_CB(skb)->eor = 1;
1284 size_bias = skb->len;
1285 avail_size = info->size_goal - skb->len;
1289 /* Zero window and all data acked? Probe. */
1290 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1291 if (avail_size == 0) {
1292 u64 snd_una = READ_ONCE(msk->snd_una);
1294 if (skb || snd_una != msk->snd_nxt)
1296 zero_window_probe = true;
1297 data_seq = snd_una - 1;
1301 if (WARN_ON_ONCE(info->sent > info->limit ||
1302 info->limit > dfrag->data_len))
1305 ret = info->limit - info->sent;
1306 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1307 dfrag->page, dfrag->offset + info->sent, &ret);
1309 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1313 /* if the tail skb is still the cached one, collapsing really happened.
1316 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1317 mpext->data_len += ret;
1318 WARN_ON_ONCE(!can_collapse);
1319 WARN_ON_ONCE(zero_window_probe);
1323 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1324 if (WARN_ON_ONCE(!mpext)) {
1325 /* should never reach here, stream corrupted */
1329 memset(mpext, 0, sizeof(*mpext));
1330 mpext->data_seq = data_seq;
1331 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1332 mpext->data_len = ret;
1336 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1337 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1340 if (zero_window_probe) {
1341 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1344 tcp_push_pending_frames(ssk);
1347 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1351 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1352 sizeof(struct tcphdr) - \
1353 MAX_TCP_OPTION_SPACE - \
1354 sizeof(struct ipv6hdr) - \
1355 sizeof(struct frag_hdr))
1357 struct subflow_send_info {
1362 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk,
1365 struct subflow_send_info send_info[2];
1366 struct mptcp_subflow_context *subflow;
1367 int i, nr_active = 0;
1372 sock_owned_by_me((struct sock *)msk);
1375 if (__mptcp_check_fallback(msk)) {
1378 *sndbuf = msk->first->sk_sndbuf;
1379 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1382 /* re-use last subflow, if the burst allow that */
1383 if (msk->last_snd && msk->snd_burst > 0 &&
1384 sk_stream_memory_free(msk->last_snd) &&
1385 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1386 mptcp_for_each_subflow(msk, subflow) {
1387 ssk = mptcp_subflow_tcp_sock(subflow);
1388 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1390 return msk->last_snd;
1393 /* pick the subflow with the lower wmem/wspace ratio */
1394 for (i = 0; i < 2; ++i) {
1395 send_info[i].ssk = NULL;
1396 send_info[i].ratio = -1;
1398 mptcp_for_each_subflow(msk, subflow) {
1399 ssk = mptcp_subflow_tcp_sock(subflow);
1400 if (!mptcp_subflow_active(subflow))
1403 nr_active += !subflow->backup;
1404 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1405 if (!sk_stream_memory_free(subflow->tcp_sock))
1408 pace = READ_ONCE(ssk->sk_pacing_rate);
1412 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1414 if (ratio < send_info[subflow->backup].ratio) {
1415 send_info[subflow->backup].ssk = ssk;
1416 send_info[subflow->backup].ratio = ratio;
1420 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1421 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1422 send_info[1].ssk, send_info[1].ratio);
1424 /* pick the best backup if no other subflow is active */
1426 send_info[0].ssk = send_info[1].ssk;
1428 if (send_info[0].ssk) {
1429 msk->last_snd = send_info[0].ssk;
1430 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1431 sk_stream_wspace(msk->last_snd));
1432 return msk->last_snd;
1437 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1438 struct mptcp_sendmsg_info *info)
1440 mptcp_set_timeout(sk, ssk);
1441 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1445 static void mptcp_push_pending(struct sock *sk, unsigned int flags)
1447 struct sock *prev_ssk = NULL, *ssk = NULL;
1448 struct mptcp_sock *msk = mptcp_sk(sk);
1449 struct mptcp_sendmsg_info info = {
1452 struct mptcp_data_frag *dfrag;
1453 int len, copied = 0;
1456 while ((dfrag = mptcp_send_head(sk))) {
1457 info.sent = dfrag->already_sent;
1458 info.limit = dfrag->data_len;
1459 len = dfrag->data_len - dfrag->already_sent;
1464 __mptcp_flush_join_list(msk);
1465 ssk = mptcp_subflow_get_send(msk, &sndbuf);
1467 /* do auto tuning */
1468 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1469 sndbuf > READ_ONCE(sk->sk_sndbuf))
1470 WRITE_ONCE(sk->sk_sndbuf, sndbuf);
1472 /* try to keep the subflow socket lock across
1473 * consecutive xmit on the same socket
1475 if (ssk != prev_ssk && prev_ssk)
1476 mptcp_push_release(sk, prev_ssk, &info);
1480 if (ssk != prev_ssk || !prev_ssk)
1483 /* keep it simple and always provide a new skb for the
1484 * subflow, even if we will not use it when collapsing
1485 * on the pending one
1487 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1488 mptcp_push_release(sk, ssk, &info);
1492 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1494 mptcp_push_release(sk, ssk, &info);
1499 dfrag->already_sent += ret;
1500 msk->snd_nxt += ret;
1501 msk->snd_burst -= ret;
1502 msk->tx_pending_data -= ret;
1506 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1509 /* at this point we held the socket lock for the last subflow we used */
1511 mptcp_push_release(sk, ssk, &info);
1515 /* start the timer, if it's not pending */
1516 if (!mptcp_timer_pending(sk))
1517 mptcp_reset_timer(sk);
1518 __mptcp_check_send_data_fin(sk);
1522 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1524 struct mptcp_sock *msk = mptcp_sk(sk);
1525 struct mptcp_sendmsg_info info;
1526 struct mptcp_data_frag *dfrag;
1527 int len, copied = 0;
1530 while ((dfrag = mptcp_send_head(sk))) {
1531 info.sent = dfrag->already_sent;
1532 info.limit = dfrag->data_len;
1533 len = dfrag->data_len - dfrag->already_sent;
1537 /* do auto tuning */
1538 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1539 ssk->sk_sndbuf > READ_ONCE(sk->sk_sndbuf))
1540 WRITE_ONCE(sk->sk_sndbuf, ssk->sk_sndbuf);
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;
1561 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1565 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1566 * not going to flush it via release_sock()
1568 __mptcp_update_wmem(sk);
1570 mptcp_set_timeout(sk, ssk);
1571 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1573 if (msk->snd_data_fin_enable &&
1574 msk->snd_nxt + 1 == msk->write_seq)
1575 mptcp_schedule_work(sk);
1579 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1581 struct mptcp_sock *msk = mptcp_sk(sk);
1582 struct page_frag *pfrag;
1587 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1590 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, len));
1592 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1594 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1595 ret = sk_stream_wait_connect(sk, &timeo);
1600 pfrag = sk_page_frag(sk);
1602 while (msg_data_left(msg)) {
1603 int total_ts, frag_truesize = 0;
1604 struct mptcp_data_frag *dfrag;
1605 struct sk_buff_head skbs;
1606 bool dfrag_collapsed;
1607 size_t psize, offset;
1609 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1614 /* reuse tail pfrag, if possible, or carve a new one from the
1617 dfrag = mptcp_pending_tail(sk);
1618 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1619 if (!dfrag_collapsed) {
1620 if (!sk_stream_memory_free(sk))
1621 goto wait_for_memory;
1623 if (!mptcp_page_frag_refill(sk, pfrag))
1624 goto wait_for_memory;
1626 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1627 frag_truesize = dfrag->overhead;
1630 /* we do not bound vs wspace, to allow a single packet.
1631 * memory accounting will prevent execessive memory usage
1634 offset = dfrag->offset + dfrag->data_len;
1635 psize = pfrag->size - offset;
1636 psize = min_t(size_t, psize, msg_data_left(msg));
1637 total_ts = psize + frag_truesize;
1638 __skb_queue_head_init(&skbs);
1639 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1640 goto wait_for_memory;
1642 if (!mptcp_wmem_alloc(sk, total_ts)) {
1643 __skb_queue_purge(&skbs);
1644 goto wait_for_memory;
1647 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1648 if (copy_page_from_iter(dfrag->page, offset, psize,
1649 &msg->msg_iter) != psize) {
1650 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1655 /* data successfully copied into the write queue */
1657 dfrag->data_len += psize;
1658 frag_truesize += psize;
1659 pfrag->offset += frag_truesize;
1660 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1662 /* charge data on mptcp pending queue to the msk socket
1663 * Note: we charge such data both to sk and ssk
1665 sk_wmem_queued_add(sk, frag_truesize);
1666 if (!dfrag_collapsed) {
1667 get_page(dfrag->page);
1668 list_add_tail(&dfrag->list, &msk->rtx_queue);
1669 if (!msk->first_pending)
1670 WRITE_ONCE(msk->first_pending, dfrag);
1672 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1673 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1679 set_bit(MPTCP_NOSPACE, &msk->flags);
1680 mptcp_push_pending(sk, msg->msg_flags);
1681 ret = sk_stream_wait_memory(sk, &timeo);
1687 msk->tx_pending_data += copied;
1688 mptcp_push_pending(sk, msg->msg_flags);
1693 return copied ? : ret;
1696 static void mptcp_wait_data(struct sock *sk, long *timeo)
1698 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1699 struct mptcp_sock *msk = mptcp_sk(sk);
1701 add_wait_queue(sk_sleep(sk), &wait);
1702 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1704 sk_wait_event(sk, timeo,
1705 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1707 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1708 remove_wait_queue(sk_sleep(sk), &wait);
1711 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1715 struct sk_buff *skb;
1718 while ((skb = skb_peek(&msk->receive_queue)) != NULL) {
1719 u32 offset = MPTCP_SKB_CB(skb)->offset;
1720 u32 data_len = skb->len - offset;
1721 u32 count = min_t(size_t, len - copied, data_len);
1724 err = skb_copy_datagram_msg(skb, offset, msg, count);
1725 if (unlikely(err < 0)) {
1733 if (count < data_len) {
1734 MPTCP_SKB_CB(skb)->offset += count;
1738 /* we will bulk release the skb memory later */
1739 skb->destructor = NULL;
1740 msk->rmem_released += skb->truesize;
1741 __skb_unlink(skb, &msk->receive_queue);
1751 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1753 * Only difference: Use highest rtt estimate of the subflows in use.
1755 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1757 struct mptcp_subflow_context *subflow;
1758 struct sock *sk = (struct sock *)msk;
1759 u32 time, advmss = 1;
1762 sock_owned_by_me(sk);
1767 msk->rcvq_space.copied += copied;
1769 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1770 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1772 rtt_us = msk->rcvq_space.rtt_us;
1773 if (rtt_us && time < (rtt_us >> 3))
1777 mptcp_for_each_subflow(msk, subflow) {
1778 const struct tcp_sock *tp;
1782 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1784 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1785 sf_advmss = READ_ONCE(tp->advmss);
1787 rtt_us = max(sf_rtt_us, rtt_us);
1788 advmss = max(sf_advmss, advmss);
1791 msk->rcvq_space.rtt_us = rtt_us;
1792 if (time < (rtt_us >> 3) || rtt_us == 0)
1795 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1798 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1799 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1803 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1805 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1807 do_div(grow, msk->rcvq_space.space);
1808 rcvwin += (grow << 1);
1810 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1811 while (tcp_win_from_space(sk, rcvmem) < advmss)
1814 do_div(rcvwin, advmss);
1815 rcvbuf = min_t(u64, rcvwin * rcvmem,
1816 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1818 if (rcvbuf > sk->sk_rcvbuf) {
1821 window_clamp = tcp_win_from_space(sk, rcvbuf);
1822 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1824 /* Make subflows follow along. If we do not do this, we
1825 * get drops at subflow level if skbs can't be moved to
1826 * the mptcp rx queue fast enough (announced rcv_win can
1827 * exceed ssk->sk_rcvbuf).
1829 mptcp_for_each_subflow(msk, subflow) {
1833 ssk = mptcp_subflow_tcp_sock(subflow);
1834 slow = lock_sock_fast(ssk);
1835 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1836 tcp_sk(ssk)->window_clamp = window_clamp;
1837 tcp_cleanup_rbuf(ssk, 1);
1838 unlock_sock_fast(ssk, slow);
1843 msk->rcvq_space.space = msk->rcvq_space.copied;
1845 msk->rcvq_space.copied = 0;
1846 msk->rcvq_space.time = mstamp;
1849 static void __mptcp_update_rmem(struct sock *sk)
1851 struct mptcp_sock *msk = mptcp_sk(sk);
1853 if (!msk->rmem_released)
1856 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1857 sk_mem_uncharge(sk, msk->rmem_released);
1858 msk->rmem_released = 0;
1861 static void __mptcp_splice_receive_queue(struct sock *sk)
1863 struct mptcp_sock *msk = mptcp_sk(sk);
1865 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1868 static bool __mptcp_move_skbs(struct mptcp_sock *msk, unsigned int rcv)
1870 struct sock *sk = (struct sock *)msk;
1871 unsigned int moved = 0;
1874 __mptcp_flush_join_list(msk);
1876 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1879 /* we can have data pending in the subflows only if the msk
1880 * receive buffer was full at subflow_data_ready() time,
1881 * that is an unlikely slow path.
1886 slowpath = lock_sock_fast(ssk);
1887 mptcp_data_lock(sk);
1888 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1889 mptcp_data_unlock(sk);
1891 WRITE_ONCE(msk->rmem_pending, min(rcv, moved));
1892 tcp_cleanup_rbuf(ssk, 1);
1893 WRITE_ONCE(msk->rmem_pending, 0);
1895 unlock_sock_fast(ssk, slowpath);
1898 /* acquire the data lock only if some input data is pending */
1900 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1901 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1902 mptcp_data_lock(sk);
1903 __mptcp_update_rmem(sk);
1904 ret |= __mptcp_ofo_queue(msk);
1905 __mptcp_splice_receive_queue(sk);
1906 mptcp_data_unlock(sk);
1909 mptcp_check_data_fin((struct sock *)msk);
1910 return !skb_queue_empty(&msk->receive_queue);
1913 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1914 int nonblock, int flags, int *addr_len)
1916 struct mptcp_sock *msk = mptcp_sk(sk);
1921 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1924 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1925 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1930 timeo = sock_rcvtimeo(sk, nonblock);
1932 len = min_t(size_t, len, INT_MAX);
1933 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1935 while (copied < len) {
1936 int bytes_read, old_space;
1938 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1939 if (unlikely(bytes_read < 0)) {
1941 copied = bytes_read;
1945 copied += bytes_read;
1947 if (skb_queue_empty(&msk->receive_queue) &&
1948 __mptcp_move_skbs(msk, len - copied))
1951 /* be sure to advertise window change */
1952 old_space = READ_ONCE(msk->old_wspace);
1953 if ((tcp_space(sk) - old_space) >= old_space)
1954 mptcp_cleanup_rbuf(msk);
1956 /* only the master socket status is relevant here. The exit
1957 * conditions mirror closely tcp_recvmsg()
1959 if (copied >= target)
1964 sk->sk_state == TCP_CLOSE ||
1965 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1967 signal_pending(current))
1971 copied = sock_error(sk);
1975 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1976 mptcp_check_for_eof(msk);
1978 if (sk->sk_shutdown & RCV_SHUTDOWN) {
1979 /* race breaker: the shutdown could be after the
1980 * previous receive queue check
1982 if (__mptcp_move_skbs(msk, len - copied))
1987 if (sk->sk_state == TCP_CLOSE) {
1997 if (signal_pending(current)) {
1998 copied = sock_intr_errno(timeo);
2003 pr_debug("block timeout %ld", timeo);
2004 mptcp_wait_data(sk, &timeo);
2007 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2008 skb_queue_empty(&msk->receive_queue)) {
2009 /* entire backlog drained, clear DATA_READY. */
2010 clear_bit(MPTCP_DATA_READY, &msk->flags);
2012 /* .. race-breaker: ssk might have gotten new data
2013 * after last __mptcp_move_skbs() returned false.
2015 if (unlikely(__mptcp_move_skbs(msk, 0)))
2016 set_bit(MPTCP_DATA_READY, &msk->flags);
2017 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2018 /* data to read but mptcp_wait_data() cleared DATA_READY */
2019 set_bit(MPTCP_DATA_READY, &msk->flags);
2022 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2023 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2024 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2025 mptcp_rcv_space_adjust(msk, copied);
2031 static void mptcp_retransmit_handler(struct sock *sk)
2033 struct mptcp_sock *msk = mptcp_sk(sk);
2035 set_bit(MPTCP_WORK_RTX, &msk->flags);
2036 mptcp_schedule_work(sk);
2039 static void mptcp_retransmit_timer(struct timer_list *t)
2041 struct inet_connection_sock *icsk = from_timer(icsk, t,
2042 icsk_retransmit_timer);
2043 struct sock *sk = &icsk->icsk_inet.sk;
2046 if (!sock_owned_by_user(sk)) {
2047 mptcp_retransmit_handler(sk);
2049 /* delegate our work to tcp_release_cb() */
2050 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
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 /* subflow sockets can be either outgoing (connect) or incoming
2109 * Outgoing subflows use in-kernel sockets.
2110 * Incoming subflows do not have their own 'struct socket' allocated,
2111 * so we need to use tcp_close() after detaching them from the mptcp
2114 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2115 struct mptcp_subflow_context *subflow)
2117 bool dispose_socket = false;
2118 struct socket *sock;
2120 list_del(&subflow->node);
2124 /* if we are invoked by the msk cleanup code, the subflow is
2127 sock = ssk->sk_socket;
2129 dispose_socket = sock != sk->sk_socket;
2133 subflow->disposable = 1;
2135 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2136 * the ssk has been already destroyed, we just need to release the
2137 * reference owned by msk;
2139 if (!inet_csk(ssk)->icsk_ulp_ops) {
2140 kfree_rcu(subflow, rcu);
2142 /* otherwise tcp will dispose of the ssk and subflow ctx */
2143 __tcp_close(ssk, 0);
2145 /* close acquired an extra ref */
2150 iput(SOCK_INODE(sock));
2155 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2160 static void pm_work(struct mptcp_sock *msk)
2162 struct mptcp_pm_data *pm = &msk->pm;
2164 spin_lock_bh(&msk->pm.lock);
2166 pr_debug("msk=%p status=%x", msk, pm->status);
2167 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
2168 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
2169 mptcp_pm_nl_add_addr_received(msk);
2171 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_SEND_ACK)) {
2172 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_SEND_ACK);
2173 mptcp_pm_nl_add_addr_send_ack(msk);
2175 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
2176 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
2177 mptcp_pm_nl_rm_addr_received(msk);
2179 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
2180 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
2181 mptcp_pm_nl_fully_established(msk);
2183 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
2184 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
2185 mptcp_pm_nl_subflow_established(msk);
2188 spin_unlock_bh(&msk->pm.lock);
2191 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2193 struct mptcp_subflow_context *subflow, *tmp;
2195 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2196 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2198 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2201 __mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2205 static bool mptcp_check_close_timeout(const struct sock *sk)
2207 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2208 struct mptcp_subflow_context *subflow;
2210 if (delta >= TCP_TIMEWAIT_LEN)
2213 /* if all subflows are in closed status don't bother with additional
2216 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2217 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2224 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2226 struct mptcp_subflow_context *subflow, *tmp;
2227 struct sock *sk = &msk->sk.icsk_inet.sk;
2229 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2232 mptcp_token_destroy(msk);
2234 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2235 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2238 if (tcp_sk->sk_state != TCP_CLOSE) {
2239 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2240 tcp_set_state(tcp_sk, TCP_CLOSE);
2242 release_sock(tcp_sk);
2245 inet_sk_state_store(sk, TCP_CLOSE);
2246 sk->sk_shutdown = SHUTDOWN_MASK;
2247 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2248 set_bit(MPTCP_DATA_READY, &msk->flags);
2249 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2251 mptcp_close_wake_up(sk);
2254 static void mptcp_worker(struct work_struct *work)
2256 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2257 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
2258 struct mptcp_sendmsg_info info = {};
2259 struct mptcp_data_frag *dfrag;
2264 state = sk->sk_state;
2265 if (unlikely(state == TCP_CLOSE))
2268 mptcp_check_data_fin_ack(sk);
2269 __mptcp_flush_join_list(msk);
2271 mptcp_check_fastclose(msk);
2273 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2274 __mptcp_close_subflow(msk);
2279 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2280 mptcp_check_for_eof(msk);
2282 __mptcp_check_send_data_fin(sk);
2283 mptcp_check_data_fin(sk);
2285 /* if the msk data is completely acked, or the socket timedout,
2286 * there is no point in keeping around an orphaned sk
2288 if (sock_flag(sk, SOCK_DEAD) &&
2289 (mptcp_check_close_timeout(sk) ||
2290 (state != sk->sk_state &&
2291 ((1 << inet_sk_state_load(sk)) & (TCPF_CLOSE | TCPF_FIN_WAIT2))))) {
2292 inet_sk_state_store(sk, TCP_CLOSE);
2293 __mptcp_destroy_sock(sk);
2297 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2300 dfrag = mptcp_rtx_head(sk);
2304 ssk = mptcp_subflow_get_retrans(msk);
2310 /* limit retransmission to the bytes already sent on some subflows */
2312 info.limit = dfrag->already_sent;
2313 while (info.sent < dfrag->already_sent) {
2314 if (!mptcp_alloc_tx_skb(sk, ssk))
2317 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2321 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2326 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2329 mptcp_set_timeout(sk, ssk);
2333 if (!mptcp_timer_pending(sk))
2334 mptcp_reset_timer(sk);
2341 static int __mptcp_init_sock(struct sock *sk)
2343 struct mptcp_sock *msk = mptcp_sk(sk);
2345 spin_lock_init(&msk->join_list_lock);
2347 INIT_LIST_HEAD(&msk->conn_list);
2348 INIT_LIST_HEAD(&msk->join_list);
2349 INIT_LIST_HEAD(&msk->rtx_queue);
2350 INIT_WORK(&msk->work, mptcp_worker);
2351 __skb_queue_head_init(&msk->receive_queue);
2352 __skb_queue_head_init(&msk->skb_tx_cache);
2353 msk->out_of_order_queue = RB_ROOT;
2354 msk->first_pending = NULL;
2355 msk->wmem_reserved = 0;
2356 msk->rmem_released = 0;
2357 msk->tx_pending_data = 0;
2358 msk->size_goal_cache = TCP_BASE_MSS;
2360 msk->ack_hint = NULL;
2362 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2364 mptcp_pm_data_init(msk);
2366 /* re-use the csk retrans timer for MPTCP-level retrans */
2367 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2368 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2372 static int mptcp_init_sock(struct sock *sk)
2374 struct net *net = sock_net(sk);
2377 ret = __mptcp_init_sock(sk);
2381 if (!mptcp_is_enabled(net))
2382 return -ENOPROTOOPT;
2384 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2387 ret = __mptcp_socket_create(mptcp_sk(sk));
2391 sk_sockets_allocated_inc(sk);
2392 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2393 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2398 static void __mptcp_clear_xmit(struct sock *sk)
2400 struct mptcp_sock *msk = mptcp_sk(sk);
2401 struct mptcp_data_frag *dtmp, *dfrag;
2402 struct sk_buff *skb;
2404 WRITE_ONCE(msk->first_pending, NULL);
2405 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2406 dfrag_clear(sk, dfrag);
2407 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2408 sk->sk_forward_alloc += skb->truesize;
2413 static void mptcp_cancel_work(struct sock *sk)
2415 struct mptcp_sock *msk = mptcp_sk(sk);
2417 if (cancel_work_sync(&msk->work))
2421 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2425 switch (ssk->sk_state) {
2427 if (!(how & RCV_SHUTDOWN))
2431 tcp_disconnect(ssk, O_NONBLOCK);
2434 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2435 pr_debug("Fallback");
2436 ssk->sk_shutdown |= how;
2437 tcp_shutdown(ssk, how);
2439 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2440 mptcp_set_timeout(sk, ssk);
2449 static const unsigned char new_state[16] = {
2450 /* current state: new state: action: */
2451 [0 /* (Invalid) */] = TCP_CLOSE,
2452 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2453 [TCP_SYN_SENT] = TCP_CLOSE,
2454 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2455 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2456 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2457 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2458 [TCP_CLOSE] = TCP_CLOSE,
2459 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2460 [TCP_LAST_ACK] = TCP_LAST_ACK,
2461 [TCP_LISTEN] = TCP_CLOSE,
2462 [TCP_CLOSING] = TCP_CLOSING,
2463 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2466 static int mptcp_close_state(struct sock *sk)
2468 int next = (int)new_state[sk->sk_state];
2469 int ns = next & TCP_STATE_MASK;
2471 inet_sk_state_store(sk, ns);
2473 return next & TCP_ACTION_FIN;
2476 static void __mptcp_check_send_data_fin(struct sock *sk)
2478 struct mptcp_subflow_context *subflow;
2479 struct mptcp_sock *msk = mptcp_sk(sk);
2481 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2482 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2483 msk->snd_nxt, msk->write_seq);
2485 /* we still need to enqueue subflows or not really shutting down,
2488 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2489 mptcp_send_head(sk))
2492 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2494 /* fallback socket will not get data_fin/ack, can move to the next
2497 if (__mptcp_check_fallback(msk)) {
2498 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2499 inet_sk_state_store(sk, TCP_CLOSE);
2500 mptcp_close_wake_up(sk);
2501 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2502 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2506 __mptcp_flush_join_list(msk);
2507 mptcp_for_each_subflow(msk, subflow) {
2508 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2510 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2514 static void __mptcp_wr_shutdown(struct sock *sk)
2516 struct mptcp_sock *msk = mptcp_sk(sk);
2518 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2519 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2520 !!mptcp_send_head(sk));
2522 /* will be ignored by fallback sockets */
2523 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2524 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2526 __mptcp_check_send_data_fin(sk);
2529 static void __mptcp_destroy_sock(struct sock *sk)
2531 struct mptcp_subflow_context *subflow, *tmp;
2532 struct mptcp_sock *msk = mptcp_sk(sk);
2533 LIST_HEAD(conn_list);
2535 pr_debug("msk=%p", msk);
2537 /* be sure to always acquire the join list lock, to sync vs
2538 * mptcp_finish_join().
2540 spin_lock_bh(&msk->join_list_lock);
2541 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2542 spin_unlock_bh(&msk->join_list_lock);
2543 list_splice_init(&msk->conn_list, &conn_list);
2545 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2546 sk_stop_timer(sk, &sk->sk_timer);
2549 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2550 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2551 __mptcp_close_ssk(sk, ssk, subflow);
2554 sk->sk_prot->destroy(sk);
2556 WARN_ON_ONCE(msk->wmem_reserved);
2557 WARN_ON_ONCE(msk->rmem_released);
2558 sk_stream_kill_queues(sk);
2559 xfrm_sk_free_policy(sk);
2560 sk_refcnt_debug_release(sk);
2564 static void mptcp_close(struct sock *sk, long timeout)
2566 struct mptcp_subflow_context *subflow;
2567 bool do_cancel_work = false;
2570 sk->sk_shutdown = SHUTDOWN_MASK;
2572 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2573 inet_sk_state_store(sk, TCP_CLOSE);
2577 if (mptcp_close_state(sk))
2578 __mptcp_wr_shutdown(sk);
2580 sk_stream_wait_close(sk, timeout);
2583 /* orphan all the subflows */
2584 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2585 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2586 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2587 bool slow, dispose_socket;
2588 struct socket *sock;
2590 slow = lock_sock_fast(ssk);
2591 sock = ssk->sk_socket;
2592 dispose_socket = sock && sock != sk->sk_socket;
2594 unlock_sock_fast(ssk, slow);
2596 /* for the outgoing subflows we additionally need to free
2597 * the associated socket
2600 iput(SOCK_INODE(sock));
2605 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2606 if (sk->sk_state == TCP_CLOSE) {
2607 __mptcp_destroy_sock(sk);
2608 do_cancel_work = true;
2610 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2614 mptcp_cancel_work(sk);
2618 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2620 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2621 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2622 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2624 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2625 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2628 msk6->saddr = ssk6->saddr;
2629 msk6->flow_label = ssk6->flow_label;
2633 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2634 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2635 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2636 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2637 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2638 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2641 static int mptcp_disconnect(struct sock *sk, int flags)
2643 /* Should never be called.
2644 * inet_stream_connect() calls ->disconnect, but that
2645 * refers to the subflow socket, not the mptcp one.
2651 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2652 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2654 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2656 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2660 struct sock *mptcp_sk_clone(const struct sock *sk,
2661 const struct mptcp_options_received *mp_opt,
2662 struct request_sock *req)
2664 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2665 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2666 struct mptcp_sock *msk;
2672 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2673 if (nsk->sk_family == AF_INET6)
2674 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2677 __mptcp_init_sock(nsk);
2679 msk = mptcp_sk(nsk);
2680 msk->local_key = subflow_req->local_key;
2681 msk->token = subflow_req->token;
2682 msk->subflow = NULL;
2683 WRITE_ONCE(msk->fully_established, false);
2685 msk->write_seq = subflow_req->idsn + 1;
2686 msk->snd_nxt = msk->write_seq;
2687 msk->snd_una = msk->write_seq;
2688 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2690 if (mp_opt->mp_capable) {
2691 msk->can_ack = true;
2692 msk->remote_key = mp_opt->sndr_key;
2693 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2695 WRITE_ONCE(msk->ack_seq, ack_seq);
2696 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2699 sock_reset_flag(nsk, SOCK_RCU_FREE);
2700 /* will be fully established after successful MPC subflow creation */
2701 inet_sk_state_store(nsk, TCP_SYN_RECV);
2702 bh_unlock_sock(nsk);
2704 /* keep a single reference */
2709 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2711 const struct tcp_sock *tp = tcp_sk(ssk);
2713 msk->rcvq_space.copied = 0;
2714 msk->rcvq_space.rtt_us = 0;
2716 msk->rcvq_space.time = tp->tcp_mstamp;
2718 /* initial rcv_space offering made to peer */
2719 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2720 TCP_INIT_CWND * tp->advmss);
2721 if (msk->rcvq_space.space == 0)
2722 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2724 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2727 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2730 struct mptcp_sock *msk = mptcp_sk(sk);
2731 struct socket *listener;
2734 listener = __mptcp_nmpc_socket(msk);
2735 if (WARN_ON_ONCE(!listener)) {
2740 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2741 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2745 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2746 if (sk_is_mptcp(newsk)) {
2747 struct mptcp_subflow_context *subflow;
2748 struct sock *new_mptcp_sock;
2750 subflow = mptcp_subflow_ctx(newsk);
2751 new_mptcp_sock = subflow->conn;
2753 /* is_mptcp should be false if subflow->conn is missing, see
2754 * subflow_syn_recv_sock()
2756 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2757 tcp_sk(newsk)->is_mptcp = 0;
2761 /* acquire the 2nd reference for the owning socket */
2762 sock_hold(new_mptcp_sock);
2763 newsk = new_mptcp_sock;
2764 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2766 MPTCP_INC_STATS(sock_net(sk),
2767 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2773 void mptcp_destroy_common(struct mptcp_sock *msk)
2775 struct sock *sk = (struct sock *)msk;
2777 __mptcp_clear_xmit(sk);
2779 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2780 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2782 skb_rbtree_purge(&msk->out_of_order_queue);
2783 mptcp_token_destroy(msk);
2784 mptcp_pm_free_anno_list(msk);
2787 static void mptcp_destroy(struct sock *sk)
2789 struct mptcp_sock *msk = mptcp_sk(sk);
2791 mptcp_destroy_common(msk);
2792 sk_sockets_allocated_dec(sk);
2795 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2796 sockptr_t optval, unsigned int optlen)
2798 struct sock *sk = (struct sock *)msk;
2799 struct socket *ssock;
2806 ssock = __mptcp_nmpc_socket(msk);
2812 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2814 if (optname == SO_REUSEPORT)
2815 sk->sk_reuseport = ssock->sk->sk_reuseport;
2816 else if (optname == SO_REUSEADDR)
2817 sk->sk_reuse = ssock->sk->sk_reuse;
2823 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2826 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2827 sockptr_t optval, unsigned int optlen)
2829 struct sock *sk = (struct sock *)msk;
2830 int ret = -EOPNOTSUPP;
2831 struct socket *ssock;
2836 ssock = __mptcp_nmpc_socket(msk);
2842 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2844 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2853 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2854 sockptr_t optval, unsigned int optlen)
2856 struct mptcp_sock *msk = mptcp_sk(sk);
2859 pr_debug("msk=%p", msk);
2861 if (level == SOL_SOCKET)
2862 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2864 /* @@ the meaning of setsockopt() when the socket is connected and
2865 * there are multiple subflows is not yet defined. It is up to the
2866 * MPTCP-level socket to configure the subflows until the subflow
2867 * is in TCP fallback, when TCP socket options are passed through
2868 * to the one remaining subflow.
2871 ssk = __mptcp_tcp_fallback(msk);
2874 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2876 if (level == SOL_IPV6)
2877 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2882 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2883 char __user *optval, int __user *option)
2885 struct mptcp_sock *msk = mptcp_sk(sk);
2888 pr_debug("msk=%p", msk);
2890 /* @@ the meaning of setsockopt() when the socket is connected and
2891 * there are multiple subflows is not yet defined. It is up to the
2892 * MPTCP-level socket to configure the subflows until the subflow
2893 * is in TCP fallback, when socket options are passed through
2894 * to the one remaining subflow.
2897 ssk = __mptcp_tcp_fallback(msk);
2900 return tcp_getsockopt(ssk, level, optname, optval, option);
2905 void __mptcp_data_acked(struct sock *sk)
2907 if (!sock_owned_by_user(sk))
2908 __mptcp_clean_una(sk);
2910 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2912 if (mptcp_pending_data_fin_ack(sk))
2913 mptcp_schedule_work(sk);
2916 void __mptcp_wnd_updated(struct sock *sk, struct sock *ssk)
2918 if (!mptcp_send_head(sk))
2921 if (!sock_owned_by_user(sk))
2922 __mptcp_subflow_push_pending(sk, ssk);
2924 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2927 #define MPTCP_DEFERRED_ALL (TCPF_WRITE_TIMER_DEFERRED)
2929 /* processes deferred events and flush wmem */
2930 static void mptcp_release_cb(struct sock *sk)
2932 unsigned long flags, nflags;
2934 /* push_pending may touch wmem_reserved, do it before the later
2937 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2938 __mptcp_clean_una(sk);
2939 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags)) {
2940 /* mptcp_push_pending() acquires the subflow socket lock
2942 * 1) can't be invoked in atomic scope
2943 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2944 * datapath acquires the msk socket spinlock while helding
2945 * the subflow socket lock
2948 spin_unlock_bh(&sk->sk_lock.slock);
2949 mptcp_push_pending(sk, 0);
2950 spin_lock_bh(&sk->sk_lock.slock);
2953 /* clear any wmem reservation and errors */
2954 __mptcp_update_wmem(sk);
2955 __mptcp_update_rmem(sk);
2958 flags = sk->sk_tsq_flags;
2959 if (!(flags & MPTCP_DEFERRED_ALL))
2961 nflags = flags & ~MPTCP_DEFERRED_ALL;
2962 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2964 sock_release_ownership(sk);
2966 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2967 mptcp_retransmit_handler(sk);
2972 static int mptcp_hash(struct sock *sk)
2974 /* should never be called,
2975 * we hash the TCP subflows not the master socket
2981 static void mptcp_unhash(struct sock *sk)
2983 /* called from sk_common_release(), but nothing to do here */
2986 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2988 struct mptcp_sock *msk = mptcp_sk(sk);
2989 struct socket *ssock;
2991 ssock = __mptcp_nmpc_socket(msk);
2992 pr_debug("msk=%p, subflow=%p", msk, ssock);
2993 if (WARN_ON_ONCE(!ssock))
2996 return inet_csk_get_port(ssock->sk, snum);
2999 void mptcp_finish_connect(struct sock *ssk)
3001 struct mptcp_subflow_context *subflow;
3002 struct mptcp_sock *msk;
3006 subflow = mptcp_subflow_ctx(ssk);
3010 pr_debug("msk=%p, token=%u", sk, subflow->token);
3012 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3014 subflow->map_seq = ack_seq;
3015 subflow->map_subflow_seq = 1;
3017 /* the socket is not connected yet, no msk/subflow ops can access/race
3018 * accessing the field below
3020 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3021 WRITE_ONCE(msk->local_key, subflow->local_key);
3022 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3023 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3024 WRITE_ONCE(msk->ack_seq, ack_seq);
3025 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3026 WRITE_ONCE(msk->can_ack, 1);
3027 WRITE_ONCE(msk->snd_una, msk->write_seq);
3029 mptcp_pm_new_connection(msk, 0);
3031 mptcp_rcv_space_init(msk, ssk);
3034 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3036 write_lock_bh(&sk->sk_callback_lock);
3037 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3038 sk_set_socket(sk, parent);
3039 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3040 write_unlock_bh(&sk->sk_callback_lock);
3043 bool mptcp_finish_join(struct sock *ssk)
3045 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3046 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3047 struct sock *parent = (void *)msk;
3048 struct socket *parent_sock;
3051 pr_debug("msk=%p, subflow=%p", msk, subflow);
3053 /* mptcp socket already closing? */
3054 if (!mptcp_is_fully_established(parent))
3057 if (!msk->pm.server_side)
3060 if (!mptcp_pm_allow_new_subflow(msk))
3063 /* active connections are already on conn_list, and we can't acquire
3065 * use the join list lock as synchronization point and double-check
3066 * msk status to avoid racing with __mptcp_destroy_sock()
3068 spin_lock_bh(&msk->join_list_lock);
3069 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3070 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3071 list_add_tail(&subflow->node, &msk->join_list);
3074 spin_unlock_bh(&msk->join_list_lock);
3078 /* attach to msk socket only after we are sure he will deal with us
3081 parent_sock = READ_ONCE(parent->sk_socket);
3082 if (parent_sock && !ssk->sk_socket)
3083 mptcp_sock_graft(ssk, parent_sock);
3084 subflow->map_seq = READ_ONCE(msk->ack_seq);
3088 static struct proto mptcp_prot = {
3090 .owner = THIS_MODULE,
3091 .init = mptcp_init_sock,
3092 .disconnect = mptcp_disconnect,
3093 .close = mptcp_close,
3094 .accept = mptcp_accept,
3095 .setsockopt = mptcp_setsockopt,
3096 .getsockopt = mptcp_getsockopt,
3097 .shutdown = tcp_shutdown,
3098 .destroy = mptcp_destroy,
3099 .sendmsg = mptcp_sendmsg,
3100 .recvmsg = mptcp_recvmsg,
3101 .release_cb = mptcp_release_cb,
3103 .unhash = mptcp_unhash,
3104 .get_port = mptcp_get_port,
3105 .sockets_allocated = &mptcp_sockets_allocated,
3106 .memory_allocated = &tcp_memory_allocated,
3107 .memory_pressure = &tcp_memory_pressure,
3108 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3109 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3110 .sysctl_mem = sysctl_tcp_mem,
3111 .obj_size = sizeof(struct mptcp_sock),
3112 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3113 .no_autobind = true,
3116 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3118 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3119 struct socket *ssock;
3122 lock_sock(sock->sk);
3123 ssock = __mptcp_nmpc_socket(msk);
3129 err = ssock->ops->bind(ssock, uaddr, addr_len);
3131 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3134 release_sock(sock->sk);
3138 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3139 struct mptcp_subflow_context *subflow)
3141 subflow->request_mptcp = 0;
3142 __mptcp_do_fallback(msk);
3145 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3146 int addr_len, int flags)
3148 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3149 struct mptcp_subflow_context *subflow;
3150 struct socket *ssock;
3153 lock_sock(sock->sk);
3154 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3155 /* pending connection or invalid state, let existing subflow
3158 ssock = msk->subflow;
3162 ssock = __mptcp_nmpc_socket(msk);
3168 mptcp_token_destroy(msk);
3169 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3170 subflow = mptcp_subflow_ctx(ssock->sk);
3171 #ifdef CONFIG_TCP_MD5SIG
3172 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3175 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3176 mptcp_subflow_early_fallback(msk, subflow);
3178 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
3179 mptcp_subflow_early_fallback(msk, subflow);
3182 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3183 sock->state = ssock->state;
3185 /* on successful connect, the msk state will be moved to established by
3186 * subflow_finish_connect()
3188 if (!err || err == -EINPROGRESS)
3189 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3191 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3194 release_sock(sock->sk);
3198 static int mptcp_listen(struct socket *sock, int backlog)
3200 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3201 struct socket *ssock;
3204 pr_debug("msk=%p", msk);
3206 lock_sock(sock->sk);
3207 ssock = __mptcp_nmpc_socket(msk);
3213 mptcp_token_destroy(msk);
3214 inet_sk_state_store(sock->sk, TCP_LISTEN);
3215 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3217 err = ssock->ops->listen(ssock, backlog);
3218 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3220 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3223 release_sock(sock->sk);
3227 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3228 int flags, bool kern)
3230 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3231 struct socket *ssock;
3234 pr_debug("msk=%p", msk);
3236 lock_sock(sock->sk);
3237 if (sock->sk->sk_state != TCP_LISTEN)
3240 ssock = __mptcp_nmpc_socket(msk);
3244 clear_bit(MPTCP_DATA_READY, &msk->flags);
3245 sock_hold(ssock->sk);
3246 release_sock(sock->sk);
3248 err = ssock->ops->accept(sock, newsock, flags, kern);
3249 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3250 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3251 struct mptcp_subflow_context *subflow;
3252 struct sock *newsk = newsock->sk;
3255 slowpath = lock_sock_fast(newsk);
3257 /* PM/worker can now acquire the first subflow socket
3258 * lock without racing with listener queue cleanup,
3259 * we can notify it, if needed.
3261 subflow = mptcp_subflow_ctx(msk->first);
3262 list_add(&subflow->node, &msk->conn_list);
3263 sock_hold(msk->first);
3264 if (mptcp_is_fully_established(newsk))
3265 mptcp_pm_fully_established(msk);
3267 mptcp_copy_inaddrs(newsk, msk->first);
3268 mptcp_rcv_space_init(msk, msk->first);
3270 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3271 * This is needed so NOSPACE flag can be set from tcp stack.
3273 __mptcp_flush_join_list(msk);
3274 mptcp_for_each_subflow(msk, subflow) {
3275 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3277 if (!ssk->sk_socket)
3278 mptcp_sock_graft(ssk, newsock);
3280 unlock_sock_fast(newsk, slowpath);
3283 if (inet_csk_listen_poll(ssock->sk))
3284 set_bit(MPTCP_DATA_READY, &msk->flags);
3285 sock_put(ssock->sk);
3289 release_sock(sock->sk);
3293 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3295 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3299 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3301 struct sock *sk = (struct sock *)msk;
3303 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3306 if (sk_stream_is_writeable(sk))
3307 return EPOLLOUT | EPOLLWRNORM;
3309 set_bit(MPTCP_NOSPACE, &msk->flags);
3310 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3311 if (sk_stream_is_writeable(sk))
3312 return EPOLLOUT | EPOLLWRNORM;
3317 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3318 struct poll_table_struct *wait)
3320 struct sock *sk = sock->sk;
3321 struct mptcp_sock *msk;
3326 sock_poll_wait(file, sock, wait);
3328 state = inet_sk_state_load(sk);
3329 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3330 if (state == TCP_LISTEN)
3331 return mptcp_check_readable(msk);
3333 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3334 mask |= mptcp_check_readable(msk);
3335 mask |= mptcp_check_writeable(msk);
3337 if (sk->sk_shutdown & RCV_SHUTDOWN)
3338 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3343 static int mptcp_shutdown(struct socket *sock, int how)
3345 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3346 struct sock *sk = sock->sk;
3349 pr_debug("sk=%p, how=%d", msk, how);
3354 if ((how & ~SHUTDOWN_MASK) || !how) {
3359 if (sock->state == SS_CONNECTING) {
3360 if ((1 << sk->sk_state) &
3361 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
3362 sock->state = SS_DISCONNECTING;
3364 sock->state = SS_CONNECTED;
3367 sk->sk_shutdown |= how;
3368 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3369 __mptcp_wr_shutdown(sk);
3371 /* Wake up anyone sleeping in poll. */
3372 sk->sk_state_change(sk);
3380 static const struct proto_ops mptcp_stream_ops = {
3382 .owner = THIS_MODULE,
3383 .release = inet_release,
3385 .connect = mptcp_stream_connect,
3386 .socketpair = sock_no_socketpair,
3387 .accept = mptcp_stream_accept,
3388 .getname = inet_getname,
3390 .ioctl = inet_ioctl,
3391 .gettstamp = sock_gettstamp,
3392 .listen = mptcp_listen,
3393 .shutdown = mptcp_shutdown,
3394 .setsockopt = sock_common_setsockopt,
3395 .getsockopt = sock_common_getsockopt,
3396 .sendmsg = inet_sendmsg,
3397 .recvmsg = inet_recvmsg,
3398 .mmap = sock_no_mmap,
3399 .sendpage = inet_sendpage,
3402 static struct inet_protosw mptcp_protosw = {
3403 .type = SOCK_STREAM,
3404 .protocol = IPPROTO_MPTCP,
3405 .prot = &mptcp_prot,
3406 .ops = &mptcp_stream_ops,
3407 .flags = INET_PROTOSW_ICSK,
3410 void __init mptcp_proto_init(void)
3412 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3414 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3415 panic("Failed to allocate MPTCP pcpu counter\n");
3417 mptcp_subflow_init();
3421 if (proto_register(&mptcp_prot, 1) != 0)
3422 panic("Failed to register MPTCP proto.\n");
3424 inet_register_protosw(&mptcp_protosw);
3426 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3429 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3430 static const struct proto_ops mptcp_v6_stream_ops = {
3432 .owner = THIS_MODULE,
3433 .release = inet6_release,
3435 .connect = mptcp_stream_connect,
3436 .socketpair = sock_no_socketpair,
3437 .accept = mptcp_stream_accept,
3438 .getname = inet6_getname,
3440 .ioctl = inet6_ioctl,
3441 .gettstamp = sock_gettstamp,
3442 .listen = mptcp_listen,
3443 .shutdown = mptcp_shutdown,
3444 .setsockopt = sock_common_setsockopt,
3445 .getsockopt = sock_common_getsockopt,
3446 .sendmsg = inet6_sendmsg,
3447 .recvmsg = inet6_recvmsg,
3448 .mmap = sock_no_mmap,
3449 .sendpage = inet_sendpage,
3450 #ifdef CONFIG_COMPAT
3451 .compat_ioctl = inet6_compat_ioctl,
3455 static struct proto mptcp_v6_prot;
3457 static void mptcp_v6_destroy(struct sock *sk)
3460 inet6_destroy_sock(sk);
3463 static struct inet_protosw mptcp_v6_protosw = {
3464 .type = SOCK_STREAM,
3465 .protocol = IPPROTO_MPTCP,
3466 .prot = &mptcp_v6_prot,
3467 .ops = &mptcp_v6_stream_ops,
3468 .flags = INET_PROTOSW_ICSK,
3471 int __init mptcp_proto_v6_init(void)
3475 mptcp_v6_prot = mptcp_prot;
3476 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3477 mptcp_v6_prot.slab = NULL;
3478 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3479 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3481 err = proto_register(&mptcp_v6_prot, 1);
3485 err = inet6_register_protosw(&mptcp_v6_protosw);
3487 proto_unregister(&mptcp_v6_prot);