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;
117 mptcp_sock_graft(msk->first, sk->sk_socket);
122 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
124 sk_drops_add(sk, skb);
128 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
129 struct sk_buff *from)
134 if (MPTCP_SKB_CB(from)->offset ||
135 !skb_try_coalesce(to, from, &fragstolen, &delta))
138 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
139 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
140 to->len, MPTCP_SKB_CB(from)->end_seq);
141 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
142 kfree_skb_partial(from, fragstolen);
143 atomic_add(delta, &sk->sk_rmem_alloc);
144 sk_mem_charge(sk, delta);
148 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
149 struct sk_buff *from)
151 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
154 return mptcp_try_coalesce((struct sock *)msk, to, from);
157 /* "inspired" by tcp_data_queue_ofo(), main differences:
159 * - don't cope with sacks
161 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
163 struct sock *sk = (struct sock *)msk;
164 struct rb_node **p, *parent;
165 u64 seq, end_seq, max_seq;
166 struct sk_buff *skb1;
168 seq = MPTCP_SKB_CB(skb)->map_seq;
169 end_seq = MPTCP_SKB_CB(skb)->end_seq;
170 max_seq = READ_ONCE(msk->rcv_wnd_sent);
172 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
173 RB_EMPTY_ROOT(&msk->out_of_order_queue));
174 if (after64(end_seq, max_seq)) {
177 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
178 (unsigned long long)end_seq - (unsigned long)max_seq,
179 (unsigned long long)msk->rcv_wnd_sent);
180 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
184 p = &msk->out_of_order_queue.rb_node;
185 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
186 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
187 rb_link_node(&skb->rbnode, NULL, p);
188 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
189 msk->ooo_last_skb = skb;
193 /* with 2 subflows, adding at end of ooo queue is quite likely
194 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
196 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
197 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
198 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
202 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
203 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
204 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
205 parent = &msk->ooo_last_skb->rbnode;
206 p = &parent->rb_right;
210 /* Find place to insert this segment. Handle overlaps on the way. */
214 skb1 = rb_to_skb(parent);
215 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
216 p = &parent->rb_left;
219 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
220 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
221 /* All the bits are present. Drop. */
223 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
226 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
230 * continue traversing
233 /* skb's seq == skb1's seq and skb covers skb1.
234 * Replace skb1 with skb.
236 rb_replace_node(&skb1->rbnode, &skb->rbnode,
237 &msk->out_of_order_queue);
238 mptcp_drop(sk, skb1);
239 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
242 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
246 p = &parent->rb_right;
250 /* Insert segment into RB tree. */
251 rb_link_node(&skb->rbnode, parent, p);
252 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
255 /* Remove other segments covered by skb. */
256 while ((skb1 = skb_rb_next(skb)) != NULL) {
257 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
259 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
260 mptcp_drop(sk, skb1);
261 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
263 /* If there is no skb after us, we are the last_skb ! */
265 msk->ooo_last_skb = skb;
269 skb_set_owner_r(skb, sk);
272 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
273 struct sk_buff *skb, unsigned int offset,
276 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
277 struct sock *sk = (struct sock *)msk;
278 struct sk_buff *tail;
280 __skb_unlink(skb, &ssk->sk_receive_queue);
285 /* try to fetch required memory from subflow */
286 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
287 if (ssk->sk_forward_alloc < skb->truesize)
289 __sk_mem_reclaim(ssk, skb->truesize);
290 if (!sk_rmem_schedule(sk, skb, skb->truesize))
294 /* the skb map_seq accounts for the skb offset:
295 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
298 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
299 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
300 MPTCP_SKB_CB(skb)->offset = offset;
302 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
304 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
305 tail = skb_peek_tail(&sk->sk_receive_queue);
306 if (tail && mptcp_try_coalesce(sk, tail, skb))
309 skb_set_owner_r(skb, sk);
310 __skb_queue_tail(&sk->sk_receive_queue, skb);
312 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
313 mptcp_data_queue_ofo(msk, skb);
317 /* old data, keep it simple and drop the whole pkt, sender
318 * will retransmit as needed, if needed.
320 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
326 static void mptcp_stop_timer(struct sock *sk)
328 struct inet_connection_sock *icsk = inet_csk(sk);
330 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
331 mptcp_sk(sk)->timer_ival = 0;
334 static void mptcp_close_wake_up(struct sock *sk)
336 if (sock_flag(sk, SOCK_DEAD))
339 sk->sk_state_change(sk);
340 if (sk->sk_shutdown == SHUTDOWN_MASK ||
341 sk->sk_state == TCP_CLOSE)
342 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
344 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
347 static bool mptcp_pending_data_fin_ack(struct sock *sk)
349 struct mptcp_sock *msk = mptcp_sk(sk);
351 return !__mptcp_check_fallback(msk) &&
352 ((1 << sk->sk_state) &
353 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
354 msk->write_seq == READ_ONCE(msk->snd_una);
357 static void mptcp_check_data_fin_ack(struct sock *sk)
359 struct mptcp_sock *msk = mptcp_sk(sk);
361 /* Look for an acknowledged DATA_FIN */
362 if (mptcp_pending_data_fin_ack(sk)) {
363 mptcp_stop_timer(sk);
365 WRITE_ONCE(msk->snd_data_fin_enable, 0);
367 switch (sk->sk_state) {
369 inet_sk_state_store(sk, TCP_FIN_WAIT2);
373 inet_sk_state_store(sk, TCP_CLOSE);
377 mptcp_close_wake_up(sk);
381 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
383 struct mptcp_sock *msk = mptcp_sk(sk);
385 if (READ_ONCE(msk->rcv_data_fin) &&
386 ((1 << sk->sk_state) &
387 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
388 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
390 if (msk->ack_seq == rcv_data_fin_seq) {
392 *seq = rcv_data_fin_seq;
401 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
403 long tout = ssk && inet_csk(ssk)->icsk_pending ?
404 inet_csk(ssk)->icsk_timeout - jiffies : 0;
407 tout = mptcp_sk(sk)->timer_ival;
408 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
411 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
413 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
415 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
416 if (subflow->request_join && !subflow->fully_established)
419 /* only send if our side has not closed yet */
420 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
423 static bool tcp_can_send_ack(const struct sock *ssk)
425 return !((1 << inet_sk_state_load(ssk)) &
426 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
429 static void mptcp_send_ack(struct mptcp_sock *msk)
431 struct mptcp_subflow_context *subflow;
433 mptcp_for_each_subflow(msk, subflow) {
434 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
437 if (tcp_can_send_ack(ssk))
443 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
448 ret = tcp_can_send_ack(ssk);
450 tcp_cleanup_rbuf(ssk, 1);
455 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
457 struct sock *ack_hint = READ_ONCE(msk->ack_hint);
458 struct mptcp_subflow_context *subflow;
460 /* if the hinted ssk is still active, try to use it */
461 if (likely(ack_hint)) {
462 mptcp_for_each_subflow(msk, subflow) {
463 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
465 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk))
470 /* otherwise pick the first active subflow */
471 mptcp_for_each_subflow(msk, subflow)
472 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
476 static bool mptcp_check_data_fin(struct sock *sk)
478 struct mptcp_sock *msk = mptcp_sk(sk);
479 u64 rcv_data_fin_seq;
482 if (__mptcp_check_fallback(msk) || !msk->first)
485 /* Need to ack a DATA_FIN received from a peer while this side
486 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
487 * msk->rcv_data_fin was set when parsing the incoming options
488 * at the subflow level and the msk lock was not held, so this
489 * is the first opportunity to act on the DATA_FIN and change
492 * If we are caught up to the sequence number of the incoming
493 * DATA_FIN, send the DATA_ACK now and do state transition. If
494 * not caught up, do nothing and let the recv code send DATA_ACK
498 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
499 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
500 WRITE_ONCE(msk->rcv_data_fin, 0);
502 sk->sk_shutdown |= RCV_SHUTDOWN;
503 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
504 set_bit(MPTCP_DATA_READY, &msk->flags);
506 switch (sk->sk_state) {
507 case TCP_ESTABLISHED:
508 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
511 inet_sk_state_store(sk, TCP_CLOSING);
514 inet_sk_state_store(sk, TCP_CLOSE);
517 /* Other states not expected */
523 mptcp_set_timeout(sk, NULL);
525 mptcp_close_wake_up(sk);
530 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
534 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
535 struct sock *sk = (struct sock *)msk;
536 unsigned int moved = 0;
537 bool more_data_avail;
542 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
544 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
545 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
547 if (unlikely(ssk_rbuf > sk_rbuf)) {
548 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
553 pr_debug("msk=%p ssk=%p", msk, ssk);
556 u32 map_remaining, offset;
557 u32 seq = tp->copied_seq;
561 /* try to move as much data as available */
562 map_remaining = subflow->map_data_len -
563 mptcp_subflow_get_map_offset(subflow);
565 skb = skb_peek(&ssk->sk_receive_queue);
567 /* if no data is found, a racing workqueue/recvmsg
568 * already processed the new data, stop here or we
569 * can enter an infinite loop
576 if (__mptcp_check_fallback(msk)) {
577 /* if we are running under the workqueue, TCP could have
578 * collapsed skbs between dummy map creation and now
579 * be sure to adjust the size
581 map_remaining = skb->len;
582 subflow->map_data_len = skb->len;
585 offset = seq - TCP_SKB_CB(skb)->seq;
586 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
592 if (offset < skb->len) {
593 size_t len = skb->len - offset;
598 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
602 if (WARN_ON_ONCE(map_remaining < len))
606 sk_eat_skb(ssk, skb);
610 WRITE_ONCE(tp->copied_seq, seq);
611 more_data_avail = mptcp_subflow_data_available(ssk);
613 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
617 } while (more_data_avail);
618 WRITE_ONCE(msk->ack_hint, ssk);
624 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
626 struct sock *sk = (struct sock *)msk;
627 struct sk_buff *skb, *tail;
632 p = rb_first(&msk->out_of_order_queue);
633 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
636 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
640 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
642 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
645 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
649 end_seq = MPTCP_SKB_CB(skb)->end_seq;
650 tail = skb_peek_tail(&sk->sk_receive_queue);
651 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
652 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
654 /* skip overlapping data, if any */
655 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
656 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
658 MPTCP_SKB_CB(skb)->offset += delta;
659 __skb_queue_tail(&sk->sk_receive_queue, skb);
661 msk->ack_seq = end_seq;
667 /* In most cases we will be able to lock the mptcp socket. If its already
668 * owned, we need to defer to the work queue to avoid ABBA deadlock.
670 static void move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
672 struct sock *sk = (struct sock *)msk;
673 unsigned int moved = 0;
675 if (inet_sk_state_load(sk) == TCP_CLOSE)
680 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
681 __mptcp_ofo_queue(msk);
683 /* If the moves have caught up with the DATA_FIN sequence number
684 * it's time to ack the DATA_FIN and change socket state, but
685 * this is not a good place to change state. Let the workqueue
688 if (mptcp_pending_data_fin(sk, NULL))
689 mptcp_schedule_work(sk);
690 mptcp_data_unlock(sk);
693 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
695 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
696 struct mptcp_sock *msk = mptcp_sk(sk);
697 int sk_rbuf, ssk_rbuf;
700 /* The peer can send data while we are shutting down this
701 * subflow at msk destruction time, but we must avoid enqueuing
702 * more data to the msk receive queue
704 if (unlikely(subflow->disposable))
707 /* move_skbs_to_msk below can legitly clear the data_avail flag,
708 * but we will need later to properly woke the reader, cache its
711 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
713 set_bit(MPTCP_DATA_READY, &msk->flags);
715 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
716 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
717 if (unlikely(ssk_rbuf > sk_rbuf))
720 /* over limit? can't append more skbs to msk */
721 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
724 move_skbs_to_msk(msk, ssk);
728 sk->sk_data_ready(sk);
731 void __mptcp_flush_join_list(struct mptcp_sock *msk)
733 struct mptcp_subflow_context *subflow;
735 if (likely(list_empty(&msk->join_list)))
738 spin_lock_bh(&msk->join_list_lock);
739 list_for_each_entry(subflow, &msk->join_list, node)
740 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
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 (WARN_ON_ONCE(amount < 0))
883 if (amount <= sk->sk_forward_alloc)
886 /* under memory pressure try to reserve at most a single page
887 * otherwise try to reserve the full estimate and fallback
888 * to a single page before entering the error path
890 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
891 !sk_wmem_schedule(sk, amount)) {
892 if (amount <= PAGE_SIZE)
896 if (!sk_wmem_schedule(sk, amount))
901 msk->wmem_reserved = amount;
902 sk->sk_forward_alloc -= amount;
906 /* we will wait for memory on next allocation */
907 msk->wmem_reserved = -1;
910 static void __mptcp_update_wmem(struct sock *sk)
912 struct mptcp_sock *msk = mptcp_sk(sk);
914 if (!msk->wmem_reserved)
917 if (msk->wmem_reserved < 0)
918 msk->wmem_reserved = 0;
919 if (msk->wmem_reserved > 0) {
920 sk->sk_forward_alloc += msk->wmem_reserved;
921 msk->wmem_reserved = 0;
925 static bool mptcp_wmem_alloc(struct sock *sk, int size)
927 struct mptcp_sock *msk = mptcp_sk(sk);
929 /* check for pre-existing error condition */
930 if (msk->wmem_reserved < 0)
933 if (msk->wmem_reserved >= size)
937 if (!sk_wmem_schedule(sk, size)) {
938 mptcp_data_unlock(sk);
942 sk->sk_forward_alloc -= size;
943 msk->wmem_reserved += size;
944 mptcp_data_unlock(sk);
947 msk->wmem_reserved -= size;
951 static void mptcp_wmem_uncharge(struct sock *sk, int size)
953 struct mptcp_sock *msk = mptcp_sk(sk);
955 if (msk->wmem_reserved < 0)
956 msk->wmem_reserved = 0;
957 msk->wmem_reserved += size;
960 static void mptcp_mem_reclaim_partial(struct sock *sk)
962 struct mptcp_sock *msk = mptcp_sk(sk);
964 /* if we are experiencing a transint allocation error,
965 * the forward allocation memory has been already
968 if (msk->wmem_reserved < 0)
972 sk->sk_forward_alloc += msk->wmem_reserved;
973 sk_mem_reclaim_partial(sk);
974 msk->wmem_reserved = sk->sk_forward_alloc;
975 sk->sk_forward_alloc = 0;
976 mptcp_data_unlock(sk);
979 static void dfrag_uncharge(struct sock *sk, int len)
981 sk_mem_uncharge(sk, len);
982 sk_wmem_queued_add(sk, -len);
985 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
987 int len = dfrag->data_len + dfrag->overhead;
989 list_del(&dfrag->list);
990 dfrag_uncharge(sk, len);
991 put_page(dfrag->page);
994 static void __mptcp_clean_una(struct sock *sk)
996 struct mptcp_sock *msk = mptcp_sk(sk);
997 struct mptcp_data_frag *dtmp, *dfrag;
998 bool cleaned = false;
1001 /* on fallback we just need to ignore snd_una, as this is really
1004 if (__mptcp_check_fallback(msk))
1005 msk->snd_una = READ_ONCE(msk->snd_nxt);
1007 snd_una = msk->snd_una;
1008 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1009 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1012 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1014 dfrag_clear(sk, dfrag);
1018 dfrag = mptcp_rtx_head(sk);
1019 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1020 u64 delta = snd_una - dfrag->data_seq;
1022 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1025 dfrag->data_seq += delta;
1026 dfrag->offset += delta;
1027 dfrag->data_len -= delta;
1028 dfrag->already_sent -= delta;
1030 dfrag_uncharge(sk, delta);
1036 if (tcp_under_memory_pressure(sk)) {
1037 __mptcp_update_wmem(sk);
1038 sk_mem_reclaim_partial(sk);
1042 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1043 if (msk->timer_ival)
1044 mptcp_stop_timer(sk);
1046 mptcp_reset_timer(sk);
1050 static void mptcp_enter_memory_pressure(struct sock *sk)
1052 struct mptcp_subflow_context *subflow;
1053 struct mptcp_sock *msk = mptcp_sk(sk);
1056 sk_stream_moderate_sndbuf(sk);
1057 mptcp_for_each_subflow(msk, subflow) {
1058 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1061 tcp_enter_memory_pressure(ssk);
1062 sk_stream_moderate_sndbuf(ssk);
1067 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1070 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1072 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1073 pfrag, sk->sk_allocation)))
1076 mptcp_enter_memory_pressure(sk);
1080 static struct mptcp_data_frag *
1081 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1084 int offset = ALIGN(orig_offset, sizeof(long));
1085 struct mptcp_data_frag *dfrag;
1087 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1088 dfrag->data_len = 0;
1089 dfrag->data_seq = msk->write_seq;
1090 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1091 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1092 dfrag->already_sent = 0;
1093 dfrag->page = pfrag->page;
1098 struct mptcp_sendmsg_info {
1106 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1109 u64 window_end = mptcp_wnd_end(msk);
1111 if (__mptcp_check_fallback(msk))
1114 if (!before64(data_seq + avail_size, window_end)) {
1115 u64 allowed_size = window_end - data_seq;
1117 return min_t(unsigned int, allowed_size, avail_size);
1123 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1125 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1129 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1133 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1135 struct sk_buff *skb;
1137 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1139 if (likely(__mptcp_add_ext(skb, gfp))) {
1140 skb_reserve(skb, MAX_TCP_HEADER);
1141 skb->reserved_tailroom = skb->end - skb->tail;
1146 mptcp_enter_memory_pressure(sk);
1151 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1152 struct sk_buff_head *skbs, int *total_ts)
1154 struct mptcp_sock *msk = mptcp_sk(sk);
1155 struct sk_buff *skb;
1158 if (unlikely(tcp_under_memory_pressure(sk))) {
1159 mptcp_mem_reclaim_partial(sk);
1161 /* under pressure pre-allocate at most a single skb */
1162 if (msk->skb_tx_cache.qlen)
1164 space_needed = msk->size_goal_cache;
1166 space_needed = msk->tx_pending_data + size -
1167 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1170 while (space_needed > 0) {
1171 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1172 if (unlikely(!skb)) {
1173 /* under memory pressure, try to pass the caller a
1174 * single skb to allow forward progress
1176 while (skbs->qlen > 1) {
1177 skb = __skb_dequeue_tail(skbs);
1180 return skbs->qlen > 0;
1183 *total_ts += skb->truesize;
1184 __skb_queue_tail(skbs, skb);
1185 space_needed -= msk->size_goal_cache;
1190 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1192 struct mptcp_sock *msk = mptcp_sk(sk);
1193 struct sk_buff *skb;
1195 if (ssk->sk_tx_skb_cache) {
1196 skb = ssk->sk_tx_skb_cache;
1197 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1198 !__mptcp_add_ext(skb, gfp)))
1203 skb = skb_peek(&msk->skb_tx_cache);
1205 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1206 skb = __skb_dequeue(&msk->skb_tx_cache);
1207 if (WARN_ON_ONCE(!skb))
1210 mptcp_wmem_uncharge(sk, skb->truesize);
1211 ssk->sk_tx_skb_cache = skb;
1215 /* over memory limit, no point to try to allocate a new skb */
1219 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1223 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1224 ssk->sk_tx_skb_cache = skb;
1231 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1233 return !ssk->sk_tx_skb_cache &&
1234 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1235 tcp_under_memory_pressure(sk);
1238 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1240 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1241 mptcp_mem_reclaim_partial(sk);
1242 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1245 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1246 struct mptcp_data_frag *dfrag,
1247 struct mptcp_sendmsg_info *info)
1249 u64 data_seq = dfrag->data_seq + info->sent;
1250 struct mptcp_sock *msk = mptcp_sk(sk);
1251 bool zero_window_probe = false;
1252 struct mptcp_ext *mpext = NULL;
1253 struct sk_buff *skb, *tail;
1254 bool can_collapse = false;
1259 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
1260 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1262 /* compute send limit */
1263 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1264 avail_size = info->size_goal;
1265 msk->size_goal_cache = info->size_goal;
1266 skb = tcp_write_queue_tail(ssk);
1268 /* Limit the write to the size available in the
1269 * current skb, if any, so that we create at most a new skb.
1270 * Explicitly tells TCP internals to avoid collapsing on later
1271 * queue management operation, to avoid breaking the ext <->
1272 * SSN association set here
1274 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1275 can_collapse = (info->size_goal - skb->len > 0) &&
1276 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1277 if (!can_collapse) {
1278 TCP_SKB_CB(skb)->eor = 1;
1280 size_bias = skb->len;
1281 avail_size = info->size_goal - skb->len;
1285 /* Zero window and all data acked? Probe. */
1286 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1287 if (avail_size == 0) {
1288 u64 snd_una = READ_ONCE(msk->snd_una);
1290 if (skb || snd_una != msk->snd_nxt)
1292 zero_window_probe = true;
1293 data_seq = snd_una - 1;
1297 if (WARN_ON_ONCE(info->sent > info->limit ||
1298 info->limit > dfrag->data_len))
1301 ret = info->limit - info->sent;
1302 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1303 dfrag->page, dfrag->offset + info->sent, &ret);
1305 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1309 /* if the tail skb is still the cached one, collapsing really happened.
1312 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1313 mpext->data_len += ret;
1314 WARN_ON_ONCE(!can_collapse);
1315 WARN_ON_ONCE(zero_window_probe);
1319 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1320 if (WARN_ON_ONCE(!mpext)) {
1321 /* should never reach here, stream corrupted */
1325 memset(mpext, 0, sizeof(*mpext));
1326 mpext->data_seq = data_seq;
1327 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1328 mpext->data_len = ret;
1332 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1333 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1336 if (zero_window_probe) {
1337 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1340 tcp_push_pending_frames(ssk);
1343 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1347 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1348 sizeof(struct tcphdr) - \
1349 MAX_TCP_OPTION_SPACE - \
1350 sizeof(struct ipv6hdr) - \
1351 sizeof(struct frag_hdr))
1353 struct subflow_send_info {
1358 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1360 struct subflow_send_info send_info[2];
1361 struct mptcp_subflow_context *subflow;
1362 int i, nr_active = 0;
1367 sock_owned_by_me((struct sock *)msk);
1369 if (__mptcp_check_fallback(msk)) {
1372 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1375 /* re-use last subflow, if the burst allow that */
1376 if (msk->last_snd && msk->snd_burst > 0 &&
1377 sk_stream_memory_free(msk->last_snd) &&
1378 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1379 return msk->last_snd;
1381 /* pick the subflow with the lower wmem/wspace ratio */
1382 for (i = 0; i < 2; ++i) {
1383 send_info[i].ssk = NULL;
1384 send_info[i].ratio = -1;
1386 mptcp_for_each_subflow(msk, subflow) {
1387 ssk = mptcp_subflow_tcp_sock(subflow);
1388 if (!mptcp_subflow_active(subflow))
1391 nr_active += !subflow->backup;
1392 if (!sk_stream_memory_free(subflow->tcp_sock))
1395 pace = READ_ONCE(ssk->sk_pacing_rate);
1399 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1401 if (ratio < send_info[subflow->backup].ratio) {
1402 send_info[subflow->backup].ssk = ssk;
1403 send_info[subflow->backup].ratio = ratio;
1407 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1408 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1409 send_info[1].ssk, send_info[1].ratio);
1411 /* pick the best backup if no other subflow is active */
1413 send_info[0].ssk = send_info[1].ssk;
1415 if (send_info[0].ssk) {
1416 msk->last_snd = send_info[0].ssk;
1417 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1418 sk_stream_wspace(msk->last_snd));
1419 return msk->last_snd;
1425 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1426 struct mptcp_sendmsg_info *info)
1428 mptcp_set_timeout(sk, ssk);
1429 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1433 static void mptcp_push_pending(struct sock *sk, unsigned int flags)
1435 struct sock *prev_ssk = NULL, *ssk = NULL;
1436 struct mptcp_sock *msk = mptcp_sk(sk);
1437 struct mptcp_sendmsg_info info = {
1440 struct mptcp_data_frag *dfrag;
1441 int len, copied = 0;
1443 while ((dfrag = mptcp_send_head(sk))) {
1444 info.sent = dfrag->already_sent;
1445 info.limit = dfrag->data_len;
1446 len = dfrag->data_len - dfrag->already_sent;
1451 __mptcp_flush_join_list(msk);
1452 ssk = mptcp_subflow_get_send(msk);
1454 /* try to keep the subflow socket lock across
1455 * consecutive xmit on the same socket
1457 if (ssk != prev_ssk && prev_ssk)
1458 mptcp_push_release(sk, prev_ssk, &info);
1462 if (ssk != prev_ssk || !prev_ssk)
1465 /* keep it simple and always provide a new skb for the
1466 * subflow, even if we will not use it when collapsing
1467 * on the pending one
1469 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1470 mptcp_push_release(sk, ssk, &info);
1474 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1476 mptcp_push_release(sk, ssk, &info);
1481 dfrag->already_sent += ret;
1482 msk->snd_nxt += ret;
1483 msk->snd_burst -= ret;
1484 msk->tx_pending_data -= ret;
1488 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1491 /* at this point we held the socket lock for the last subflow we used */
1493 mptcp_push_release(sk, ssk, &info);
1497 /* start the timer, if it's not pending */
1498 if (!mptcp_timer_pending(sk))
1499 mptcp_reset_timer(sk);
1500 __mptcp_check_send_data_fin(sk);
1504 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1506 struct mptcp_sock *msk = mptcp_sk(sk);
1507 struct mptcp_sendmsg_info info;
1508 struct mptcp_data_frag *dfrag;
1509 int len, copied = 0;
1512 while ((dfrag = mptcp_send_head(sk))) {
1513 info.sent = dfrag->already_sent;
1514 info.limit = dfrag->data_len;
1515 len = dfrag->data_len - dfrag->already_sent;
1519 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1520 __mptcp_update_wmem(sk);
1521 sk_mem_reclaim_partial(sk);
1523 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1526 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1531 dfrag->already_sent += ret;
1532 msk->snd_nxt += ret;
1533 msk->snd_burst -= ret;
1534 msk->tx_pending_data -= ret;
1538 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1542 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1543 * not going to flush it via release_sock()
1545 __mptcp_update_wmem(sk);
1547 mptcp_set_timeout(sk, ssk);
1548 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1550 if (msk->snd_data_fin_enable &&
1551 msk->snd_nxt + 1 == msk->write_seq)
1552 mptcp_schedule_work(sk);
1556 static void mptcp_set_nospace(struct sock *sk)
1558 /* enable autotune */
1559 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1561 /* will be cleared on avail space */
1562 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1565 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1567 struct mptcp_sock *msk = mptcp_sk(sk);
1568 struct page_frag *pfrag;
1573 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1576 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1578 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1580 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1581 ret = sk_stream_wait_connect(sk, &timeo);
1586 pfrag = sk_page_frag(sk);
1588 while (msg_data_left(msg)) {
1589 int total_ts, frag_truesize = 0;
1590 struct mptcp_data_frag *dfrag;
1591 struct sk_buff_head skbs;
1592 bool dfrag_collapsed;
1593 size_t psize, offset;
1595 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1600 /* reuse tail pfrag, if possible, or carve a new one from the
1603 dfrag = mptcp_pending_tail(sk);
1604 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1605 if (!dfrag_collapsed) {
1606 if (!sk_stream_memory_free(sk))
1607 goto wait_for_memory;
1609 if (!mptcp_page_frag_refill(sk, pfrag))
1610 goto wait_for_memory;
1612 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1613 frag_truesize = dfrag->overhead;
1616 /* we do not bound vs wspace, to allow a single packet.
1617 * memory accounting will prevent execessive memory usage
1620 offset = dfrag->offset + dfrag->data_len;
1621 psize = pfrag->size - offset;
1622 psize = min_t(size_t, psize, msg_data_left(msg));
1623 total_ts = psize + frag_truesize;
1624 __skb_queue_head_init(&skbs);
1625 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1626 goto wait_for_memory;
1628 if (!mptcp_wmem_alloc(sk, total_ts)) {
1629 __skb_queue_purge(&skbs);
1630 goto wait_for_memory;
1633 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1634 if (copy_page_from_iter(dfrag->page, offset, psize,
1635 &msg->msg_iter) != psize) {
1636 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1641 /* data successfully copied into the write queue */
1643 dfrag->data_len += psize;
1644 frag_truesize += psize;
1645 pfrag->offset += frag_truesize;
1646 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1647 msk->tx_pending_data += psize;
1649 /* charge data on mptcp pending queue to the msk socket
1650 * Note: we charge such data both to sk and ssk
1652 sk_wmem_queued_add(sk, frag_truesize);
1653 if (!dfrag_collapsed) {
1654 get_page(dfrag->page);
1655 list_add_tail(&dfrag->list, &msk->rtx_queue);
1656 if (!msk->first_pending)
1657 WRITE_ONCE(msk->first_pending, dfrag);
1659 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1660 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1666 mptcp_set_nospace(sk);
1667 mptcp_push_pending(sk, msg->msg_flags);
1668 ret = sk_stream_wait_memory(sk, &timeo);
1674 mptcp_push_pending(sk, msg->msg_flags);
1678 return copied ? : ret;
1681 static void mptcp_wait_data(struct sock *sk, long *timeo)
1683 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1684 struct mptcp_sock *msk = mptcp_sk(sk);
1686 add_wait_queue(sk_sleep(sk), &wait);
1687 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1689 sk_wait_event(sk, timeo,
1690 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1692 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1693 remove_wait_queue(sk_sleep(sk), &wait);
1696 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1700 struct sk_buff *skb;
1703 while ((skb = skb_peek(&msk->receive_queue)) != NULL) {
1704 u32 offset = MPTCP_SKB_CB(skb)->offset;
1705 u32 data_len = skb->len - offset;
1706 u32 count = min_t(size_t, len - copied, data_len);
1709 err = skb_copy_datagram_msg(skb, offset, msg, count);
1710 if (unlikely(err < 0)) {
1718 if (count < data_len) {
1719 MPTCP_SKB_CB(skb)->offset += count;
1723 /* we will bulk release the skb memory later */
1724 skb->destructor = NULL;
1725 msk->rmem_released += skb->truesize;
1726 __skb_unlink(skb, &msk->receive_queue);
1736 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1738 * Only difference: Use highest rtt estimate of the subflows in use.
1740 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1742 struct mptcp_subflow_context *subflow;
1743 struct sock *sk = (struct sock *)msk;
1744 u32 time, advmss = 1;
1747 sock_owned_by_me(sk);
1752 msk->rcvq_space.copied += copied;
1754 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1755 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1757 rtt_us = msk->rcvq_space.rtt_us;
1758 if (rtt_us && time < (rtt_us >> 3))
1762 mptcp_for_each_subflow(msk, subflow) {
1763 const struct tcp_sock *tp;
1767 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1769 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1770 sf_advmss = READ_ONCE(tp->advmss);
1772 rtt_us = max(sf_rtt_us, rtt_us);
1773 advmss = max(sf_advmss, advmss);
1776 msk->rcvq_space.rtt_us = rtt_us;
1777 if (time < (rtt_us >> 3) || rtt_us == 0)
1780 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1783 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1784 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1788 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1790 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1792 do_div(grow, msk->rcvq_space.space);
1793 rcvwin += (grow << 1);
1795 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1796 while (tcp_win_from_space(sk, rcvmem) < advmss)
1799 do_div(rcvwin, advmss);
1800 rcvbuf = min_t(u64, rcvwin * rcvmem,
1801 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1803 if (rcvbuf > sk->sk_rcvbuf) {
1806 window_clamp = tcp_win_from_space(sk, rcvbuf);
1807 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1809 /* Make subflows follow along. If we do not do this, we
1810 * get drops at subflow level if skbs can't be moved to
1811 * the mptcp rx queue fast enough (announced rcv_win can
1812 * exceed ssk->sk_rcvbuf).
1814 mptcp_for_each_subflow(msk, subflow) {
1818 ssk = mptcp_subflow_tcp_sock(subflow);
1819 slow = lock_sock_fast(ssk);
1820 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1821 tcp_sk(ssk)->window_clamp = window_clamp;
1822 tcp_cleanup_rbuf(ssk, 1);
1823 unlock_sock_fast(ssk, slow);
1828 msk->rcvq_space.space = msk->rcvq_space.copied;
1830 msk->rcvq_space.copied = 0;
1831 msk->rcvq_space.time = mstamp;
1834 static void __mptcp_update_rmem(struct sock *sk)
1836 struct mptcp_sock *msk = mptcp_sk(sk);
1838 if (!msk->rmem_released)
1841 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1842 sk_mem_uncharge(sk, msk->rmem_released);
1843 msk->rmem_released = 0;
1846 static void __mptcp_splice_receive_queue(struct sock *sk)
1848 struct mptcp_sock *msk = mptcp_sk(sk);
1850 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1853 static bool __mptcp_move_skbs(struct mptcp_sock *msk, unsigned int rcv)
1855 struct sock *sk = (struct sock *)msk;
1856 unsigned int moved = 0;
1859 __mptcp_flush_join_list(msk);
1861 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1864 /* we can have data pending in the subflows only if the msk
1865 * receive buffer was full at subflow_data_ready() time,
1866 * that is an unlikely slow path.
1871 slowpath = lock_sock_fast(ssk);
1872 mptcp_data_lock(sk);
1873 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1874 mptcp_data_unlock(sk);
1876 WRITE_ONCE(msk->rmem_pending, min(rcv, moved));
1877 tcp_cleanup_rbuf(ssk, 1);
1878 WRITE_ONCE(msk->rmem_pending, 0);
1880 unlock_sock_fast(ssk, slowpath);
1883 /* acquire the data lock only if some input data is pending */
1885 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1886 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1887 mptcp_data_lock(sk);
1888 __mptcp_update_rmem(sk);
1889 ret |= __mptcp_ofo_queue(msk);
1890 __mptcp_splice_receive_queue(sk);
1891 mptcp_data_unlock(sk);
1894 mptcp_check_data_fin((struct sock *)msk);
1895 return !skb_queue_empty(&msk->receive_queue);
1898 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1899 int nonblock, int flags, int *addr_len)
1901 struct mptcp_sock *msk = mptcp_sk(sk);
1906 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1909 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1910 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1915 timeo = sock_rcvtimeo(sk, nonblock);
1917 len = min_t(size_t, len, INT_MAX);
1918 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1920 while (copied < len) {
1921 int bytes_read, old_space;
1923 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1924 if (unlikely(bytes_read < 0)) {
1926 copied = bytes_read;
1930 copied += bytes_read;
1932 if (skb_queue_empty(&msk->receive_queue) &&
1933 __mptcp_move_skbs(msk, len - copied))
1936 /* be sure to advertise window change */
1937 old_space = READ_ONCE(msk->old_wspace);
1938 if ((tcp_space(sk) - old_space) >= old_space)
1939 mptcp_cleanup_rbuf(msk);
1941 /* only the master socket status is relevant here. The exit
1942 * conditions mirror closely tcp_recvmsg()
1944 if (copied >= target)
1949 sk->sk_state == TCP_CLOSE ||
1950 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1952 signal_pending(current))
1956 copied = sock_error(sk);
1960 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1961 mptcp_check_for_eof(msk);
1963 if (sk->sk_shutdown & RCV_SHUTDOWN) {
1964 /* race breaker: the shutdown could be after the
1965 * previous receive queue check
1967 if (__mptcp_move_skbs(msk, len - copied))
1972 if (sk->sk_state == TCP_CLOSE) {
1982 if (signal_pending(current)) {
1983 copied = sock_intr_errno(timeo);
1988 pr_debug("block timeout %ld", timeo);
1989 mptcp_wait_data(sk, &timeo);
1992 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1993 skb_queue_empty(&msk->receive_queue)) {
1994 /* entire backlog drained, clear DATA_READY. */
1995 clear_bit(MPTCP_DATA_READY, &msk->flags);
1997 /* .. race-breaker: ssk might have gotten new data
1998 * after last __mptcp_move_skbs() returned false.
2000 if (unlikely(__mptcp_move_skbs(msk, 0)))
2001 set_bit(MPTCP_DATA_READY, &msk->flags);
2002 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2003 /* data to read but mptcp_wait_data() cleared DATA_READY */
2004 set_bit(MPTCP_DATA_READY, &msk->flags);
2007 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2008 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2009 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2010 mptcp_rcv_space_adjust(msk, copied);
2016 static void mptcp_retransmit_handler(struct sock *sk)
2018 struct mptcp_sock *msk = mptcp_sk(sk);
2020 set_bit(MPTCP_WORK_RTX, &msk->flags);
2021 mptcp_schedule_work(sk);
2024 static void mptcp_retransmit_timer(struct timer_list *t)
2026 struct inet_connection_sock *icsk = from_timer(icsk, t,
2027 icsk_retransmit_timer);
2028 struct sock *sk = &icsk->icsk_inet.sk;
2031 if (!sock_owned_by_user(sk)) {
2032 mptcp_retransmit_handler(sk);
2034 /* delegate our work to tcp_release_cb() */
2035 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
2043 static void mptcp_timeout_timer(struct timer_list *t)
2045 struct sock *sk = from_timer(sk, t, sk_timer);
2047 mptcp_schedule_work(sk);
2051 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2054 * A backup subflow is returned only if that is the only kind available.
2056 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2058 struct mptcp_subflow_context *subflow;
2059 struct sock *backup = NULL;
2061 sock_owned_by_me((const struct sock *)msk);
2063 if (__mptcp_check_fallback(msk))
2066 mptcp_for_each_subflow(msk, subflow) {
2067 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2069 if (!mptcp_subflow_active(subflow))
2072 /* still data outstanding at TCP level? Don't retransmit. */
2073 if (!tcp_write_queue_empty(ssk)) {
2074 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2079 if (subflow->backup) {
2091 /* subflow sockets can be either outgoing (connect) or incoming
2094 * Outgoing subflows use in-kernel sockets.
2095 * Incoming subflows do not have their own 'struct socket' allocated,
2096 * so we need to use tcp_close() after detaching them from the mptcp
2099 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2100 struct mptcp_subflow_context *subflow)
2102 list_del(&subflow->node);
2104 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2106 /* if we are invoked by the msk cleanup code, the subflow is
2112 subflow->disposable = 1;
2114 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2115 * the ssk has been already destroyed, we just need to release the
2116 * reference owned by msk;
2118 if (!inet_csk(ssk)->icsk_ulp_ops) {
2119 kfree_rcu(subflow, rcu);
2121 /* otherwise tcp will dispose of the ssk and subflow ctx */
2122 __tcp_close(ssk, 0);
2124 /* close acquired an extra ref */
2132 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2137 static void pm_work(struct mptcp_sock *msk)
2139 struct mptcp_pm_data *pm = &msk->pm;
2141 spin_lock_bh(&msk->pm.lock);
2143 pr_debug("msk=%p status=%x", msk, pm->status);
2144 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
2145 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
2146 mptcp_pm_nl_add_addr_received(msk);
2148 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_SEND_ACK)) {
2149 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_SEND_ACK);
2150 mptcp_pm_nl_add_addr_send_ack(msk);
2152 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
2153 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
2154 mptcp_pm_nl_rm_addr_received(msk);
2156 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
2157 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
2158 mptcp_pm_nl_fully_established(msk);
2160 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
2161 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
2162 mptcp_pm_nl_subflow_established(msk);
2165 spin_unlock_bh(&msk->pm.lock);
2168 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2170 struct mptcp_subflow_context *subflow, *tmp;
2172 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2173 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2175 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2178 __mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2182 static bool mptcp_check_close_timeout(const struct sock *sk)
2184 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2185 struct mptcp_subflow_context *subflow;
2187 if (delta >= TCP_TIMEWAIT_LEN)
2190 /* if all subflows are in closed status don't bother with additional
2193 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2194 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2201 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2203 struct mptcp_subflow_context *subflow, *tmp;
2204 struct sock *sk = &msk->sk.icsk_inet.sk;
2206 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2209 mptcp_token_destroy(msk);
2211 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2212 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2215 if (tcp_sk->sk_state != TCP_CLOSE) {
2216 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2217 tcp_set_state(tcp_sk, TCP_CLOSE);
2219 release_sock(tcp_sk);
2222 inet_sk_state_store(sk, TCP_CLOSE);
2223 sk->sk_shutdown = SHUTDOWN_MASK;
2224 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2225 set_bit(MPTCP_DATA_READY, &msk->flags);
2226 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2228 mptcp_close_wake_up(sk);
2231 static void mptcp_worker(struct work_struct *work)
2233 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2234 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
2235 struct mptcp_sendmsg_info info = {};
2236 struct mptcp_data_frag *dfrag;
2241 state = sk->sk_state;
2242 if (unlikely(state == TCP_CLOSE))
2245 mptcp_check_data_fin_ack(sk);
2246 __mptcp_flush_join_list(msk);
2248 mptcp_check_fastclose(msk);
2250 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2251 __mptcp_close_subflow(msk);
2256 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2257 mptcp_check_for_eof(msk);
2259 __mptcp_check_send_data_fin(sk);
2260 mptcp_check_data_fin(sk);
2262 /* if the msk data is completely acked, or the socket timedout,
2263 * there is no point in keeping around an orphaned sk
2265 if (sock_flag(sk, SOCK_DEAD) &&
2266 (mptcp_check_close_timeout(sk) ||
2267 (state != sk->sk_state &&
2268 ((1 << inet_sk_state_load(sk)) & (TCPF_CLOSE | TCPF_FIN_WAIT2))))) {
2269 inet_sk_state_store(sk, TCP_CLOSE);
2270 __mptcp_destroy_sock(sk);
2274 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2277 dfrag = mptcp_rtx_head(sk);
2281 ssk = mptcp_subflow_get_retrans(msk);
2287 /* limit retransmission to the bytes already sent on some subflows */
2289 info.limit = dfrag->already_sent;
2290 while (info.sent < dfrag->already_sent) {
2291 if (!mptcp_alloc_tx_skb(sk, ssk))
2294 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2298 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2303 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2306 mptcp_set_timeout(sk, ssk);
2310 if (!mptcp_timer_pending(sk))
2311 mptcp_reset_timer(sk);
2318 static int __mptcp_init_sock(struct sock *sk)
2320 struct mptcp_sock *msk = mptcp_sk(sk);
2322 spin_lock_init(&msk->join_list_lock);
2324 INIT_LIST_HEAD(&msk->conn_list);
2325 INIT_LIST_HEAD(&msk->join_list);
2326 INIT_LIST_HEAD(&msk->rtx_queue);
2327 INIT_WORK(&msk->work, mptcp_worker);
2328 __skb_queue_head_init(&msk->receive_queue);
2329 __skb_queue_head_init(&msk->skb_tx_cache);
2330 msk->out_of_order_queue = RB_ROOT;
2331 msk->first_pending = NULL;
2332 msk->wmem_reserved = 0;
2333 msk->rmem_released = 0;
2334 msk->tx_pending_data = 0;
2335 msk->size_goal_cache = TCP_BASE_MSS;
2337 msk->ack_hint = NULL;
2339 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2341 mptcp_pm_data_init(msk);
2343 /* re-use the csk retrans timer for MPTCP-level retrans */
2344 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2345 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2349 static int mptcp_init_sock(struct sock *sk)
2351 struct net *net = sock_net(sk);
2354 ret = __mptcp_init_sock(sk);
2358 if (!mptcp_is_enabled(net))
2359 return -ENOPROTOOPT;
2361 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2364 ret = __mptcp_socket_create(mptcp_sk(sk));
2368 sk_sockets_allocated_inc(sk);
2369 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2370 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2375 static void __mptcp_clear_xmit(struct sock *sk)
2377 struct mptcp_sock *msk = mptcp_sk(sk);
2378 struct mptcp_data_frag *dtmp, *dfrag;
2379 struct sk_buff *skb;
2381 WRITE_ONCE(msk->first_pending, NULL);
2382 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2383 dfrag_clear(sk, dfrag);
2384 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2385 sk->sk_forward_alloc += skb->truesize;
2390 static void mptcp_cancel_work(struct sock *sk)
2392 struct mptcp_sock *msk = mptcp_sk(sk);
2394 if (cancel_work_sync(&msk->work))
2398 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2402 switch (ssk->sk_state) {
2404 if (!(how & RCV_SHUTDOWN))
2408 tcp_disconnect(ssk, O_NONBLOCK);
2411 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2412 pr_debug("Fallback");
2413 ssk->sk_shutdown |= how;
2414 tcp_shutdown(ssk, how);
2416 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2417 mptcp_set_timeout(sk, ssk);
2426 static const unsigned char new_state[16] = {
2427 /* current state: new state: action: */
2428 [0 /* (Invalid) */] = TCP_CLOSE,
2429 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2430 [TCP_SYN_SENT] = TCP_CLOSE,
2431 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2432 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2433 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2434 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2435 [TCP_CLOSE] = TCP_CLOSE,
2436 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2437 [TCP_LAST_ACK] = TCP_LAST_ACK,
2438 [TCP_LISTEN] = TCP_CLOSE,
2439 [TCP_CLOSING] = TCP_CLOSING,
2440 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2443 static int mptcp_close_state(struct sock *sk)
2445 int next = (int)new_state[sk->sk_state];
2446 int ns = next & TCP_STATE_MASK;
2448 inet_sk_state_store(sk, ns);
2450 return next & TCP_ACTION_FIN;
2453 static void __mptcp_check_send_data_fin(struct sock *sk)
2455 struct mptcp_subflow_context *subflow;
2456 struct mptcp_sock *msk = mptcp_sk(sk);
2458 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2459 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2460 msk->snd_nxt, msk->write_seq);
2462 /* we still need to enqueue subflows or not really shutting down,
2465 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2466 mptcp_send_head(sk))
2469 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2471 /* fallback socket will not get data_fin/ack, can move to the next
2474 if (__mptcp_check_fallback(msk)) {
2475 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2476 inet_sk_state_store(sk, TCP_CLOSE);
2477 mptcp_close_wake_up(sk);
2478 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2479 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2483 __mptcp_flush_join_list(msk);
2484 mptcp_for_each_subflow(msk, subflow) {
2485 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2487 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2491 static void __mptcp_wr_shutdown(struct sock *sk)
2493 struct mptcp_sock *msk = mptcp_sk(sk);
2495 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2496 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2497 !!mptcp_send_head(sk));
2499 /* will be ignored by fallback sockets */
2500 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2501 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2503 __mptcp_check_send_data_fin(sk);
2506 static void __mptcp_destroy_sock(struct sock *sk)
2508 struct mptcp_subflow_context *subflow, *tmp;
2509 struct mptcp_sock *msk = mptcp_sk(sk);
2510 LIST_HEAD(conn_list);
2512 pr_debug("msk=%p", msk);
2514 /* dispose the ancillatory tcp socket, if any */
2516 iput(SOCK_INODE(msk->subflow));
2517 msk->subflow = NULL;
2520 /* be sure to always acquire the join list lock, to sync vs
2521 * mptcp_finish_join().
2523 spin_lock_bh(&msk->join_list_lock);
2524 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2525 spin_unlock_bh(&msk->join_list_lock);
2526 list_splice_init(&msk->conn_list, &conn_list);
2528 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2529 sk_stop_timer(sk, &sk->sk_timer);
2532 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2533 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2534 __mptcp_close_ssk(sk, ssk, subflow);
2537 sk->sk_prot->destroy(sk);
2539 WARN_ON_ONCE(msk->wmem_reserved);
2540 WARN_ON_ONCE(msk->rmem_released);
2541 sk_stream_kill_queues(sk);
2542 xfrm_sk_free_policy(sk);
2543 sk_refcnt_debug_release(sk);
2547 static void mptcp_close(struct sock *sk, long timeout)
2549 struct mptcp_subflow_context *subflow;
2550 bool do_cancel_work = false;
2553 sk->sk_shutdown = SHUTDOWN_MASK;
2555 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2556 inet_sk_state_store(sk, TCP_CLOSE);
2560 if (mptcp_close_state(sk))
2561 __mptcp_wr_shutdown(sk);
2563 sk_stream_wait_close(sk, timeout);
2566 /* orphan all the subflows */
2567 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2568 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2569 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2570 bool slow = lock_sock_fast(ssk);
2573 unlock_sock_fast(ssk, slow);
2578 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2579 if (sk->sk_state == TCP_CLOSE) {
2580 __mptcp_destroy_sock(sk);
2581 do_cancel_work = true;
2583 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2587 mptcp_cancel_work(sk);
2591 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2593 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2594 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2595 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2597 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2598 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2601 msk6->saddr = ssk6->saddr;
2602 msk6->flow_label = ssk6->flow_label;
2606 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2607 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2608 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2609 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2610 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2611 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2614 static int mptcp_disconnect(struct sock *sk, int flags)
2616 struct mptcp_subflow_context *subflow;
2617 struct mptcp_sock *msk = mptcp_sk(sk);
2619 __mptcp_flush_join_list(msk);
2620 mptcp_for_each_subflow(msk, subflow) {
2621 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2624 tcp_disconnect(ssk, flags);
2630 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2631 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2633 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2635 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2639 struct sock *mptcp_sk_clone(const struct sock *sk,
2640 const struct mptcp_options_received *mp_opt,
2641 struct request_sock *req)
2643 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2644 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2645 struct mptcp_sock *msk;
2651 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2652 if (nsk->sk_family == AF_INET6)
2653 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2656 __mptcp_init_sock(nsk);
2658 msk = mptcp_sk(nsk);
2659 msk->local_key = subflow_req->local_key;
2660 msk->token = subflow_req->token;
2661 msk->subflow = NULL;
2662 WRITE_ONCE(msk->fully_established, false);
2664 msk->write_seq = subflow_req->idsn + 1;
2665 msk->snd_nxt = msk->write_seq;
2666 msk->snd_una = msk->write_seq;
2667 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2669 if (mp_opt->mp_capable) {
2670 msk->can_ack = true;
2671 msk->remote_key = mp_opt->sndr_key;
2672 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2674 WRITE_ONCE(msk->ack_seq, ack_seq);
2675 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2678 sock_reset_flag(nsk, SOCK_RCU_FREE);
2679 /* will be fully established after successful MPC subflow creation */
2680 inet_sk_state_store(nsk, TCP_SYN_RECV);
2682 security_inet_csk_clone(nsk, req);
2683 bh_unlock_sock(nsk);
2685 /* keep a single reference */
2690 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2692 const struct tcp_sock *tp = tcp_sk(ssk);
2694 msk->rcvq_space.copied = 0;
2695 msk->rcvq_space.rtt_us = 0;
2697 msk->rcvq_space.time = tp->tcp_mstamp;
2699 /* initial rcv_space offering made to peer */
2700 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2701 TCP_INIT_CWND * tp->advmss);
2702 if (msk->rcvq_space.space == 0)
2703 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2705 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2708 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2711 struct mptcp_sock *msk = mptcp_sk(sk);
2712 struct socket *listener;
2715 listener = __mptcp_nmpc_socket(msk);
2716 if (WARN_ON_ONCE(!listener)) {
2721 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2722 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2726 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2727 if (sk_is_mptcp(newsk)) {
2728 struct mptcp_subflow_context *subflow;
2729 struct sock *new_mptcp_sock;
2731 subflow = mptcp_subflow_ctx(newsk);
2732 new_mptcp_sock = subflow->conn;
2734 /* is_mptcp should be false if subflow->conn is missing, see
2735 * subflow_syn_recv_sock()
2737 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2738 tcp_sk(newsk)->is_mptcp = 0;
2742 /* acquire the 2nd reference for the owning socket */
2743 sock_hold(new_mptcp_sock);
2744 newsk = new_mptcp_sock;
2745 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2747 MPTCP_INC_STATS(sock_net(sk),
2748 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2754 void mptcp_destroy_common(struct mptcp_sock *msk)
2756 struct sock *sk = (struct sock *)msk;
2758 __mptcp_clear_xmit(sk);
2760 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2761 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2763 skb_rbtree_purge(&msk->out_of_order_queue);
2764 mptcp_token_destroy(msk);
2765 mptcp_pm_free_anno_list(msk);
2768 static void mptcp_destroy(struct sock *sk)
2770 struct mptcp_sock *msk = mptcp_sk(sk);
2772 mptcp_destroy_common(msk);
2773 sk_sockets_allocated_dec(sk);
2776 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2777 sockptr_t optval, unsigned int optlen)
2779 struct sock *sk = (struct sock *)msk;
2780 struct socket *ssock;
2787 ssock = __mptcp_nmpc_socket(msk);
2793 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2795 if (optname == SO_REUSEPORT)
2796 sk->sk_reuseport = ssock->sk->sk_reuseport;
2797 else if (optname == SO_REUSEADDR)
2798 sk->sk_reuse = ssock->sk->sk_reuse;
2804 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2807 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2808 sockptr_t optval, unsigned int optlen)
2810 struct sock *sk = (struct sock *)msk;
2811 int ret = -EOPNOTSUPP;
2812 struct socket *ssock;
2817 ssock = __mptcp_nmpc_socket(msk);
2823 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2825 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2834 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2835 sockptr_t optval, unsigned int optlen)
2837 struct mptcp_sock *msk = mptcp_sk(sk);
2840 pr_debug("msk=%p", msk);
2842 if (level == SOL_SOCKET)
2843 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2845 /* @@ the meaning of setsockopt() when the socket is connected and
2846 * there are multiple subflows is not yet defined. It is up to the
2847 * MPTCP-level socket to configure the subflows until the subflow
2848 * is in TCP fallback, when TCP socket options are passed through
2849 * to the one remaining subflow.
2852 ssk = __mptcp_tcp_fallback(msk);
2855 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2857 if (level == SOL_IPV6)
2858 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2863 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2864 char __user *optval, int __user *option)
2866 struct mptcp_sock *msk = mptcp_sk(sk);
2869 pr_debug("msk=%p", msk);
2871 /* @@ the meaning of setsockopt() when the socket is connected and
2872 * there are multiple subflows is not yet defined. It is up to the
2873 * MPTCP-level socket to configure the subflows until the subflow
2874 * is in TCP fallback, when socket options are passed through
2875 * to the one remaining subflow.
2878 ssk = __mptcp_tcp_fallback(msk);
2881 return tcp_getsockopt(ssk, level, optname, optval, option);
2886 void __mptcp_data_acked(struct sock *sk)
2888 if (!sock_owned_by_user(sk))
2889 __mptcp_clean_una(sk);
2891 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2893 if (mptcp_pending_data_fin_ack(sk))
2894 mptcp_schedule_work(sk);
2897 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2899 if (!mptcp_send_head(sk))
2902 if (!sock_owned_by_user(sk))
2903 __mptcp_subflow_push_pending(sk, ssk);
2905 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2908 #define MPTCP_DEFERRED_ALL (TCPF_WRITE_TIMER_DEFERRED)
2910 /* processes deferred events and flush wmem */
2911 static void mptcp_release_cb(struct sock *sk)
2913 unsigned long flags, nflags;
2915 /* push_pending may touch wmem_reserved, do it before the later
2918 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2919 __mptcp_clean_una(sk);
2920 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags)) {
2921 /* mptcp_push_pending() acquires the subflow socket lock
2923 * 1) can't be invoked in atomic scope
2924 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2925 * datapath acquires the msk socket spinlock while helding
2926 * the subflow socket lock
2929 spin_unlock_bh(&sk->sk_lock.slock);
2930 mptcp_push_pending(sk, 0);
2931 spin_lock_bh(&sk->sk_lock.slock);
2934 /* clear any wmem reservation and errors */
2935 __mptcp_update_wmem(sk);
2936 __mptcp_update_rmem(sk);
2939 flags = sk->sk_tsq_flags;
2940 if (!(flags & MPTCP_DEFERRED_ALL))
2942 nflags = flags & ~MPTCP_DEFERRED_ALL;
2943 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2945 sock_release_ownership(sk);
2947 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2948 mptcp_retransmit_handler(sk);
2953 static int mptcp_hash(struct sock *sk)
2955 /* should never be called,
2956 * we hash the TCP subflows not the master socket
2962 static void mptcp_unhash(struct sock *sk)
2964 /* called from sk_common_release(), but nothing to do here */
2967 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2969 struct mptcp_sock *msk = mptcp_sk(sk);
2970 struct socket *ssock;
2972 ssock = __mptcp_nmpc_socket(msk);
2973 pr_debug("msk=%p, subflow=%p", msk, ssock);
2974 if (WARN_ON_ONCE(!ssock))
2977 return inet_csk_get_port(ssock->sk, snum);
2980 void mptcp_finish_connect(struct sock *ssk)
2982 struct mptcp_subflow_context *subflow;
2983 struct mptcp_sock *msk;
2987 subflow = mptcp_subflow_ctx(ssk);
2991 pr_debug("msk=%p, token=%u", sk, subflow->token);
2993 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2995 subflow->map_seq = ack_seq;
2996 subflow->map_subflow_seq = 1;
2998 /* the socket is not connected yet, no msk/subflow ops can access/race
2999 * accessing the field below
3001 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3002 WRITE_ONCE(msk->local_key, subflow->local_key);
3003 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3004 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3005 WRITE_ONCE(msk->ack_seq, ack_seq);
3006 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3007 WRITE_ONCE(msk->can_ack, 1);
3008 WRITE_ONCE(msk->snd_una, msk->write_seq);
3010 mptcp_pm_new_connection(msk, 0);
3012 mptcp_rcv_space_init(msk, ssk);
3015 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3017 write_lock_bh(&sk->sk_callback_lock);
3018 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3019 sk_set_socket(sk, parent);
3020 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3021 write_unlock_bh(&sk->sk_callback_lock);
3024 bool mptcp_finish_join(struct sock *ssk)
3026 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3027 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3028 struct sock *parent = (void *)msk;
3029 struct socket *parent_sock;
3032 pr_debug("msk=%p, subflow=%p", msk, subflow);
3034 /* mptcp socket already closing? */
3035 if (!mptcp_is_fully_established(parent))
3038 if (!msk->pm.server_side)
3041 if (!mptcp_pm_allow_new_subflow(msk))
3044 /* active connections are already on conn_list, and we can't acquire
3046 * use the join list lock as synchronization point and double-check
3047 * msk status to avoid racing with __mptcp_destroy_sock()
3049 spin_lock_bh(&msk->join_list_lock);
3050 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3051 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3052 list_add_tail(&subflow->node, &msk->join_list);
3055 spin_unlock_bh(&msk->join_list_lock);
3059 /* attach to msk socket only after we are sure he will deal with us
3062 parent_sock = READ_ONCE(parent->sk_socket);
3063 if (parent_sock && !ssk->sk_socket)
3064 mptcp_sock_graft(ssk, parent_sock);
3065 subflow->map_seq = READ_ONCE(msk->ack_seq);
3069 static void mptcp_shutdown(struct sock *sk, int how)
3071 pr_debug("sk=%p, how=%d", sk, how);
3073 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3074 __mptcp_wr_shutdown(sk);
3077 static struct proto mptcp_prot = {
3079 .owner = THIS_MODULE,
3080 .init = mptcp_init_sock,
3081 .disconnect = mptcp_disconnect,
3082 .close = mptcp_close,
3083 .accept = mptcp_accept,
3084 .setsockopt = mptcp_setsockopt,
3085 .getsockopt = mptcp_getsockopt,
3086 .shutdown = mptcp_shutdown,
3087 .destroy = mptcp_destroy,
3088 .sendmsg = mptcp_sendmsg,
3089 .recvmsg = mptcp_recvmsg,
3090 .release_cb = mptcp_release_cb,
3092 .unhash = mptcp_unhash,
3093 .get_port = mptcp_get_port,
3094 .sockets_allocated = &mptcp_sockets_allocated,
3095 .memory_allocated = &tcp_memory_allocated,
3096 .memory_pressure = &tcp_memory_pressure,
3097 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3098 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3099 .sysctl_mem = sysctl_tcp_mem,
3100 .obj_size = sizeof(struct mptcp_sock),
3101 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3102 .no_autobind = true,
3105 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3107 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3108 struct socket *ssock;
3111 lock_sock(sock->sk);
3112 ssock = __mptcp_nmpc_socket(msk);
3118 err = ssock->ops->bind(ssock, uaddr, addr_len);
3120 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3123 release_sock(sock->sk);
3127 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3128 struct mptcp_subflow_context *subflow)
3130 subflow->request_mptcp = 0;
3131 __mptcp_do_fallback(msk);
3134 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3135 int addr_len, int flags)
3137 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3138 struct mptcp_subflow_context *subflow;
3139 struct socket *ssock;
3142 lock_sock(sock->sk);
3143 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3144 /* pending connection or invalid state, let existing subflow
3147 ssock = msk->subflow;
3151 ssock = __mptcp_nmpc_socket(msk);
3157 mptcp_token_destroy(msk);
3158 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3159 subflow = mptcp_subflow_ctx(ssock->sk);
3160 #ifdef CONFIG_TCP_MD5SIG
3161 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3164 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3165 mptcp_subflow_early_fallback(msk, subflow);
3167 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
3168 mptcp_subflow_early_fallback(msk, subflow);
3171 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3172 sock->state = ssock->state;
3174 /* on successful connect, the msk state will be moved to established by
3175 * subflow_finish_connect()
3177 if (!err || err == -EINPROGRESS)
3178 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3180 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3183 release_sock(sock->sk);
3187 static int mptcp_listen(struct socket *sock, int backlog)
3189 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3190 struct socket *ssock;
3193 pr_debug("msk=%p", msk);
3195 lock_sock(sock->sk);
3196 ssock = __mptcp_nmpc_socket(msk);
3202 mptcp_token_destroy(msk);
3203 inet_sk_state_store(sock->sk, TCP_LISTEN);
3204 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3206 err = ssock->ops->listen(ssock, backlog);
3207 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3209 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3212 release_sock(sock->sk);
3216 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3217 int flags, bool kern)
3219 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3220 struct socket *ssock;
3223 pr_debug("msk=%p", msk);
3225 lock_sock(sock->sk);
3226 if (sock->sk->sk_state != TCP_LISTEN)
3229 ssock = __mptcp_nmpc_socket(msk);
3233 clear_bit(MPTCP_DATA_READY, &msk->flags);
3234 sock_hold(ssock->sk);
3235 release_sock(sock->sk);
3237 err = ssock->ops->accept(sock, newsock, flags, kern);
3238 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3239 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3240 struct mptcp_subflow_context *subflow;
3241 struct sock *newsk = newsock->sk;
3244 slowpath = lock_sock_fast(newsk);
3246 /* PM/worker can now acquire the first subflow socket
3247 * lock without racing with listener queue cleanup,
3248 * we can notify it, if needed.
3250 subflow = mptcp_subflow_ctx(msk->first);
3251 list_add(&subflow->node, &msk->conn_list);
3252 sock_hold(msk->first);
3253 if (mptcp_is_fully_established(newsk))
3254 mptcp_pm_fully_established(msk);
3256 mptcp_copy_inaddrs(newsk, msk->first);
3257 mptcp_rcv_space_init(msk, msk->first);
3258 mptcp_propagate_sndbuf(newsk, msk->first);
3260 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3261 * This is needed so NOSPACE flag can be set from tcp stack.
3263 __mptcp_flush_join_list(msk);
3264 mptcp_for_each_subflow(msk, subflow) {
3265 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3267 if (!ssk->sk_socket)
3268 mptcp_sock_graft(ssk, newsock);
3270 unlock_sock_fast(newsk, slowpath);
3273 if (inet_csk_listen_poll(ssock->sk))
3274 set_bit(MPTCP_DATA_READY, &msk->flags);
3275 sock_put(ssock->sk);
3279 release_sock(sock->sk);
3283 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3285 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3289 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3291 struct sock *sk = (struct sock *)msk;
3293 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3296 if (sk_stream_is_writeable(sk))
3297 return EPOLLOUT | EPOLLWRNORM;
3299 mptcp_set_nospace(sk);
3300 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3301 if (sk_stream_is_writeable(sk))
3302 return EPOLLOUT | EPOLLWRNORM;
3307 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3308 struct poll_table_struct *wait)
3310 struct sock *sk = sock->sk;
3311 struct mptcp_sock *msk;
3316 sock_poll_wait(file, sock, wait);
3318 state = inet_sk_state_load(sk);
3319 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3320 if (state == TCP_LISTEN)
3321 return mptcp_check_readable(msk);
3323 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3324 mask |= mptcp_check_readable(msk);
3325 mask |= mptcp_check_writeable(msk);
3327 if (sk->sk_shutdown & RCV_SHUTDOWN)
3328 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3333 static const struct proto_ops mptcp_stream_ops = {
3335 .owner = THIS_MODULE,
3336 .release = inet_release,
3338 .connect = mptcp_stream_connect,
3339 .socketpair = sock_no_socketpair,
3340 .accept = mptcp_stream_accept,
3341 .getname = inet_getname,
3343 .ioctl = inet_ioctl,
3344 .gettstamp = sock_gettstamp,
3345 .listen = mptcp_listen,
3346 .shutdown = inet_shutdown,
3347 .setsockopt = sock_common_setsockopt,
3348 .getsockopt = sock_common_getsockopt,
3349 .sendmsg = inet_sendmsg,
3350 .recvmsg = inet_recvmsg,
3351 .mmap = sock_no_mmap,
3352 .sendpage = inet_sendpage,
3355 static struct inet_protosw mptcp_protosw = {
3356 .type = SOCK_STREAM,
3357 .protocol = IPPROTO_MPTCP,
3358 .prot = &mptcp_prot,
3359 .ops = &mptcp_stream_ops,
3360 .flags = INET_PROTOSW_ICSK,
3363 void __init mptcp_proto_init(void)
3365 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3367 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3368 panic("Failed to allocate MPTCP pcpu counter\n");
3370 mptcp_subflow_init();
3374 if (proto_register(&mptcp_prot, 1) != 0)
3375 panic("Failed to register MPTCP proto.\n");
3377 inet_register_protosw(&mptcp_protosw);
3379 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3382 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3383 static const struct proto_ops mptcp_v6_stream_ops = {
3385 .owner = THIS_MODULE,
3386 .release = inet6_release,
3388 .connect = mptcp_stream_connect,
3389 .socketpair = sock_no_socketpair,
3390 .accept = mptcp_stream_accept,
3391 .getname = inet6_getname,
3393 .ioctl = inet6_ioctl,
3394 .gettstamp = sock_gettstamp,
3395 .listen = mptcp_listen,
3396 .shutdown = inet_shutdown,
3397 .setsockopt = sock_common_setsockopt,
3398 .getsockopt = sock_common_getsockopt,
3399 .sendmsg = inet6_sendmsg,
3400 .recvmsg = inet6_recvmsg,
3401 .mmap = sock_no_mmap,
3402 .sendpage = inet_sendpage,
3403 #ifdef CONFIG_COMPAT
3404 .compat_ioctl = inet6_compat_ioctl,
3408 static struct proto mptcp_v6_prot;
3410 static void mptcp_v6_destroy(struct sock *sk)
3413 inet6_destroy_sock(sk);
3416 static struct inet_protosw mptcp_v6_protosw = {
3417 .type = SOCK_STREAM,
3418 .protocol = IPPROTO_MPTCP,
3419 .prot = &mptcp_v6_prot,
3420 .ops = &mptcp_v6_stream_ops,
3421 .flags = INET_PROTOSW_ICSK,
3424 int __init mptcp_proto_v6_init(void)
3428 mptcp_v6_prot = mptcp_prot;
3429 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3430 mptcp_v6_prot.slab = NULL;
3431 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3432 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3434 err = proto_register(&mptcp_v6_prot, 1);
3438 err = inet6_register_protosw(&mptcp_v6_protosw);
3440 proto_unregister(&mptcp_v6_prot);