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 #define CREATE_TRACE_POINTS
29 #include <trace/events/mptcp.h>
31 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
33 struct mptcp_sock msk;
44 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
46 static struct percpu_counter mptcp_sockets_allocated;
48 static void __mptcp_destroy_sock(struct sock *sk);
49 static void __mptcp_check_send_data_fin(struct sock *sk);
51 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
52 static struct net_device mptcp_napi_dev;
54 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
55 * completed yet or has failed, return the subflow socket.
56 * Otherwise return NULL.
58 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
60 if (!msk->subflow || READ_ONCE(msk->can_ack))
66 /* Returns end sequence number of the receiver's advertised window */
67 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
69 return READ_ONCE(msk->wnd_end);
72 static bool mptcp_is_tcpsk(struct sock *sk)
74 struct socket *sock = sk->sk_socket;
76 if (unlikely(sk->sk_prot == &tcp_prot)) {
77 /* we are being invoked after mptcp_accept() has
78 * accepted a non-mp-capable flow: sk is a tcp_sk,
81 * Hand the socket over to tcp so all further socket ops
84 sock->ops = &inet_stream_ops;
86 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
87 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
88 sock->ops = &inet6_stream_ops;
96 static int __mptcp_socket_create(struct mptcp_sock *msk)
98 struct mptcp_subflow_context *subflow;
99 struct sock *sk = (struct sock *)msk;
100 struct socket *ssock;
103 err = mptcp_subflow_create_socket(sk, &ssock);
107 msk->first = ssock->sk;
108 msk->subflow = ssock;
109 subflow = mptcp_subflow_ctx(ssock->sk);
110 list_add(&subflow->node, &msk->conn_list);
111 sock_hold(ssock->sk);
112 subflow->request_mptcp = 1;
113 mptcp_sock_graft(msk->first, sk->sk_socket);
118 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
120 sk_drops_add(sk, skb);
124 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
125 struct sk_buff *from)
130 if (MPTCP_SKB_CB(from)->offset ||
131 !skb_try_coalesce(to, from, &fragstolen, &delta))
134 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
135 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
136 to->len, MPTCP_SKB_CB(from)->end_seq);
137 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
138 kfree_skb_partial(from, fragstolen);
139 atomic_add(delta, &sk->sk_rmem_alloc);
140 sk_mem_charge(sk, delta);
144 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
145 struct sk_buff *from)
147 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
150 return mptcp_try_coalesce((struct sock *)msk, to, from);
153 /* "inspired" by tcp_data_queue_ofo(), main differences:
155 * - don't cope with sacks
157 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
159 struct sock *sk = (struct sock *)msk;
160 struct rb_node **p, *parent;
161 u64 seq, end_seq, max_seq;
162 struct sk_buff *skb1;
164 seq = MPTCP_SKB_CB(skb)->map_seq;
165 end_seq = MPTCP_SKB_CB(skb)->end_seq;
166 max_seq = READ_ONCE(msk->rcv_wnd_sent);
168 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
169 RB_EMPTY_ROOT(&msk->out_of_order_queue));
170 if (after64(end_seq, max_seq)) {
173 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
174 (unsigned long long)end_seq - (unsigned long)max_seq,
175 (unsigned long long)msk->rcv_wnd_sent);
176 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
180 p = &msk->out_of_order_queue.rb_node;
181 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
182 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
183 rb_link_node(&skb->rbnode, NULL, p);
184 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
185 msk->ooo_last_skb = skb;
189 /* with 2 subflows, adding at end of ooo queue is quite likely
190 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
192 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
193 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
194 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
198 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
199 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
200 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
201 parent = &msk->ooo_last_skb->rbnode;
202 p = &parent->rb_right;
206 /* Find place to insert this segment. Handle overlaps on the way. */
210 skb1 = rb_to_skb(parent);
211 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
212 p = &parent->rb_left;
215 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
216 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
217 /* All the bits are present. Drop. */
219 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
222 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
226 * continue traversing
229 /* skb's seq == skb1's seq and skb covers skb1.
230 * Replace skb1 with skb.
232 rb_replace_node(&skb1->rbnode, &skb->rbnode,
233 &msk->out_of_order_queue);
234 mptcp_drop(sk, skb1);
235 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
238 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
239 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
242 p = &parent->rb_right;
246 /* Insert segment into RB tree. */
247 rb_link_node(&skb->rbnode, parent, p);
248 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
251 /* Remove other segments covered by skb. */
252 while ((skb1 = skb_rb_next(skb)) != NULL) {
253 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
255 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
256 mptcp_drop(sk, skb1);
257 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
259 /* If there is no skb after us, we are the last_skb ! */
261 msk->ooo_last_skb = skb;
265 skb_set_owner_r(skb, sk);
268 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
269 struct sk_buff *skb, unsigned int offset,
272 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
273 struct sock *sk = (struct sock *)msk;
274 struct sk_buff *tail;
276 __skb_unlink(skb, &ssk->sk_receive_queue);
281 /* try to fetch required memory from subflow */
282 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
283 if (ssk->sk_forward_alloc < skb->truesize)
285 __sk_mem_reclaim(ssk, skb->truesize);
286 if (!sk_rmem_schedule(sk, skb, skb->truesize))
290 /* the skb map_seq accounts for the skb offset:
291 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
294 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
295 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
296 MPTCP_SKB_CB(skb)->offset = offset;
298 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
300 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
301 tail = skb_peek_tail(&sk->sk_receive_queue);
302 if (tail && mptcp_try_coalesce(sk, tail, skb))
305 skb_set_owner_r(skb, sk);
306 __skb_queue_tail(&sk->sk_receive_queue, skb);
308 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
309 mptcp_data_queue_ofo(msk, skb);
313 /* old data, keep it simple and drop the whole pkt, sender
314 * will retransmit as needed, if needed.
316 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
322 static void mptcp_stop_timer(struct sock *sk)
324 struct inet_connection_sock *icsk = inet_csk(sk);
326 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
327 mptcp_sk(sk)->timer_ival = 0;
330 static void mptcp_close_wake_up(struct sock *sk)
332 if (sock_flag(sk, SOCK_DEAD))
335 sk->sk_state_change(sk);
336 if (sk->sk_shutdown == SHUTDOWN_MASK ||
337 sk->sk_state == TCP_CLOSE)
338 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
340 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
343 static bool mptcp_pending_data_fin_ack(struct sock *sk)
345 struct mptcp_sock *msk = mptcp_sk(sk);
347 return !__mptcp_check_fallback(msk) &&
348 ((1 << sk->sk_state) &
349 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
350 msk->write_seq == READ_ONCE(msk->snd_una);
353 static void mptcp_check_data_fin_ack(struct sock *sk)
355 struct mptcp_sock *msk = mptcp_sk(sk);
357 /* Look for an acknowledged DATA_FIN */
358 if (mptcp_pending_data_fin_ack(sk)) {
359 WRITE_ONCE(msk->snd_data_fin_enable, 0);
361 switch (sk->sk_state) {
363 inet_sk_state_store(sk, TCP_FIN_WAIT2);
367 inet_sk_state_store(sk, TCP_CLOSE);
371 mptcp_close_wake_up(sk);
375 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
377 struct mptcp_sock *msk = mptcp_sk(sk);
379 if (READ_ONCE(msk->rcv_data_fin) &&
380 ((1 << sk->sk_state) &
381 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
382 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
384 if (msk->ack_seq == rcv_data_fin_seq) {
386 *seq = rcv_data_fin_seq;
395 static void mptcp_set_datafin_timeout(const struct sock *sk)
397 struct inet_connection_sock *icsk = inet_csk(sk);
399 mptcp_sk(sk)->timer_ival = min(TCP_RTO_MAX,
400 TCP_RTO_MIN << icsk->icsk_retransmits);
403 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
405 long tout = ssk && inet_csk(ssk)->icsk_pending ?
406 inet_csk(ssk)->icsk_timeout - jiffies : 0;
409 tout = mptcp_sk(sk)->timer_ival;
410 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
413 static bool tcp_can_send_ack(const struct sock *ssk)
415 return !((1 << inet_sk_state_load(ssk)) &
416 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
419 static void mptcp_send_ack(struct mptcp_sock *msk)
421 struct mptcp_subflow_context *subflow;
423 mptcp_for_each_subflow(msk, subflow) {
424 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
427 if (tcp_can_send_ack(ssk))
433 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
438 ret = tcp_can_send_ack(ssk);
440 tcp_cleanup_rbuf(ssk, 1);
445 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
447 struct sock *ack_hint = READ_ONCE(msk->ack_hint);
448 int old_space = READ_ONCE(msk->old_wspace);
449 struct mptcp_subflow_context *subflow;
450 struct sock *sk = (struct sock *)msk;
453 /* this is a simple superset of what tcp_cleanup_rbuf() implements
454 * so that we don't have to acquire the ssk socket lock most of the time
455 * to do actually nothing
457 cleanup = __mptcp_space(sk) - old_space >= max(0, old_space);
461 /* if the hinted ssk is still active, try to use it */
462 if (likely(ack_hint)) {
463 mptcp_for_each_subflow(msk, subflow) {
464 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
466 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk))
471 /* otherwise pick the first active subflow */
472 mptcp_for_each_subflow(msk, subflow)
473 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
477 static bool mptcp_check_data_fin(struct sock *sk)
479 struct mptcp_sock *msk = mptcp_sk(sk);
480 u64 rcv_data_fin_seq;
483 if (__mptcp_check_fallback(msk))
486 /* Need to ack a DATA_FIN received from a peer while this side
487 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
488 * msk->rcv_data_fin was set when parsing the incoming options
489 * at the subflow level and the msk lock was not held, so this
490 * is the first opportunity to act on the DATA_FIN and change
493 * If we are caught up to the sequence number of the incoming
494 * DATA_FIN, send the DATA_ACK now and do state transition. If
495 * not caught up, do nothing and let the recv code send DATA_ACK
499 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
500 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
501 WRITE_ONCE(msk->rcv_data_fin, 0);
503 sk->sk_shutdown |= RCV_SHUTDOWN;
504 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
505 set_bit(MPTCP_DATA_READY, &msk->flags);
507 switch (sk->sk_state) {
508 case TCP_ESTABLISHED:
509 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
512 inet_sk_state_store(sk, TCP_CLOSING);
515 inet_sk_state_store(sk, TCP_CLOSE);
518 /* Other states not expected */
524 mptcp_set_timeout(sk, NULL);
526 mptcp_close_wake_up(sk);
531 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
535 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
536 struct sock *sk = (struct sock *)msk;
537 unsigned int moved = 0;
538 bool more_data_avail;
543 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
545 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
546 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
548 if (unlikely(ssk_rbuf > sk_rbuf)) {
549 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
554 pr_debug("msk=%p ssk=%p", msk, ssk);
557 u32 map_remaining, offset;
558 u32 seq = tp->copied_seq;
562 /* try to move as much data as available */
563 map_remaining = subflow->map_data_len -
564 mptcp_subflow_get_map_offset(subflow);
566 skb = skb_peek(&ssk->sk_receive_queue);
568 /* if no data is found, a racing workqueue/recvmsg
569 * already processed the new data, stop here or we
570 * can enter an infinite loop
577 if (__mptcp_check_fallback(msk)) {
578 /* if we are running under the workqueue, TCP could have
579 * collapsed skbs between dummy map creation and now
580 * be sure to adjust the size
582 map_remaining = skb->len;
583 subflow->map_data_len = skb->len;
586 offset = seq - TCP_SKB_CB(skb)->seq;
587 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
593 if (offset < skb->len) {
594 size_t len = skb->len - offset;
599 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
603 if (WARN_ON_ONCE(map_remaining < len))
607 sk_eat_skb(ssk, skb);
611 WRITE_ONCE(tp->copied_seq, seq);
612 more_data_avail = mptcp_subflow_data_available(ssk);
614 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
618 } while (more_data_avail);
619 WRITE_ONCE(msk->ack_hint, ssk);
625 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
627 struct sock *sk = (struct sock *)msk;
628 struct sk_buff *skb, *tail;
633 p = rb_first(&msk->out_of_order_queue);
634 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
637 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
641 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
643 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
646 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
650 end_seq = MPTCP_SKB_CB(skb)->end_seq;
651 tail = skb_peek_tail(&sk->sk_receive_queue);
652 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
653 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
655 /* skip overlapping data, if any */
656 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
657 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
659 MPTCP_SKB_CB(skb)->offset += delta;
660 __skb_queue_tail(&sk->sk_receive_queue, skb);
662 msk->ack_seq = end_seq;
668 /* In most cases we will be able to lock the mptcp socket. If its already
669 * owned, we need to defer to the work queue to avoid ABBA deadlock.
671 static void move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
673 struct sock *sk = (struct sock *)msk;
674 unsigned int moved = 0;
676 if (inet_sk_state_load(sk) == TCP_CLOSE)
681 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
682 __mptcp_ofo_queue(msk);
684 /* If the moves have caught up with the DATA_FIN sequence number
685 * it's time to ack the DATA_FIN and change socket state, but
686 * this is not a good place to change state. Let the workqueue
689 if (mptcp_pending_data_fin(sk, NULL))
690 mptcp_schedule_work(sk);
691 mptcp_data_unlock(sk);
694 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
696 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
697 struct mptcp_sock *msk = mptcp_sk(sk);
698 int sk_rbuf, ssk_rbuf;
701 /* The peer can send data while we are shutting down this
702 * subflow at msk destruction time, but we must avoid enqueuing
703 * more data to the msk receive queue
705 if (unlikely(subflow->disposable))
708 /* move_skbs_to_msk below can legitly clear the data_avail flag,
709 * but we will need later to properly woke the reader, cache its
712 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
714 set_bit(MPTCP_DATA_READY, &msk->flags);
716 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
717 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
718 if (unlikely(ssk_rbuf > sk_rbuf))
721 /* over limit? can't append more skbs to msk */
722 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
725 move_skbs_to_msk(msk, ssk);
729 sk->sk_data_ready(sk);
732 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
734 struct mptcp_subflow_context *subflow;
737 if (likely(list_empty(&msk->join_list)))
740 spin_lock_bh(&msk->join_list_lock);
741 list_for_each_entry(subflow, &msk->join_list, node) {
742 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
744 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
745 if (READ_ONCE(msk->setsockopt_seq) != sseq)
748 list_splice_tail_init(&msk->join_list, &msk->conn_list);
749 spin_unlock_bh(&msk->join_list_lock);
754 void __mptcp_flush_join_list(struct mptcp_sock *msk)
756 if (likely(!mptcp_do_flush_join_list(msk)))
759 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
760 mptcp_schedule_work((struct sock *)msk);
763 static void mptcp_flush_join_list(struct mptcp_sock *msk)
765 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
769 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
772 mptcp_sockopt_sync_all(msk);
775 static bool mptcp_timer_pending(struct sock *sk)
777 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
780 static void mptcp_reset_timer(struct sock *sk)
782 struct inet_connection_sock *icsk = inet_csk(sk);
785 /* prevent rescheduling on close */
786 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
789 /* should never be called with mptcp level timer cleared */
790 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
791 if (WARN_ON_ONCE(!tout))
793 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
796 bool mptcp_schedule_work(struct sock *sk)
798 if (inet_sk_state_load(sk) != TCP_CLOSE &&
799 schedule_work(&mptcp_sk(sk)->work)) {
800 /* each subflow already holds a reference to the sk, and the
801 * workqueue is invoked by a subflow, so sk can't go away here.
809 void mptcp_subflow_eof(struct sock *sk)
811 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
812 mptcp_schedule_work(sk);
815 static void mptcp_check_for_eof(struct mptcp_sock *msk)
817 struct mptcp_subflow_context *subflow;
818 struct sock *sk = (struct sock *)msk;
821 mptcp_for_each_subflow(msk, subflow)
822 receivers += !subflow->rx_eof;
826 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
827 /* hopefully temporary hack: propagate shutdown status
828 * to msk, when all subflows agree on it
830 sk->sk_shutdown |= RCV_SHUTDOWN;
832 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
833 set_bit(MPTCP_DATA_READY, &msk->flags);
834 sk->sk_data_ready(sk);
837 switch (sk->sk_state) {
838 case TCP_ESTABLISHED:
839 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
842 inet_sk_state_store(sk, TCP_CLOSING);
845 inet_sk_state_store(sk, TCP_CLOSE);
850 mptcp_close_wake_up(sk);
853 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
855 struct mptcp_subflow_context *subflow;
856 struct sock *sk = (struct sock *)msk;
858 sock_owned_by_me(sk);
860 mptcp_for_each_subflow(msk, subflow) {
861 if (subflow->data_avail)
862 return mptcp_subflow_tcp_sock(subflow);
868 static bool mptcp_skb_can_collapse_to(u64 write_seq,
869 const struct sk_buff *skb,
870 const struct mptcp_ext *mpext)
872 if (!tcp_skb_can_collapse_to(skb))
875 /* can collapse only if MPTCP level sequence is in order and this
876 * mapping has not been xmitted yet
878 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
882 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
883 const struct page_frag *pfrag,
884 const struct mptcp_data_frag *df)
886 return df && pfrag->page == df->page &&
887 pfrag->size - pfrag->offset > 0 &&
888 df->data_seq + df->data_len == msk->write_seq;
891 static int mptcp_wmem_with_overhead(struct sock *sk, int size)
893 struct mptcp_sock *msk = mptcp_sk(sk);
896 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
897 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache;
898 if (skbs < msk->skb_tx_cache.qlen)
901 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER);
904 static void __mptcp_wmem_reserve(struct sock *sk, int size)
906 int amount = mptcp_wmem_with_overhead(sk, size);
907 struct mptcp_sock *msk = mptcp_sk(sk);
909 WARN_ON_ONCE(msk->wmem_reserved);
910 if (WARN_ON_ONCE(amount < 0))
913 if (amount <= sk->sk_forward_alloc)
916 /* under memory pressure try to reserve at most a single page
917 * otherwise try to reserve the full estimate and fallback
918 * to a single page before entering the error path
920 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
921 !sk_wmem_schedule(sk, amount)) {
922 if (amount <= PAGE_SIZE)
926 if (!sk_wmem_schedule(sk, amount))
931 msk->wmem_reserved = amount;
932 sk->sk_forward_alloc -= amount;
936 /* we will wait for memory on next allocation */
937 msk->wmem_reserved = -1;
940 static void __mptcp_update_wmem(struct sock *sk)
942 struct mptcp_sock *msk = mptcp_sk(sk);
944 if (!msk->wmem_reserved)
947 if (msk->wmem_reserved < 0)
948 msk->wmem_reserved = 0;
949 if (msk->wmem_reserved > 0) {
950 sk->sk_forward_alloc += msk->wmem_reserved;
951 msk->wmem_reserved = 0;
955 static bool mptcp_wmem_alloc(struct sock *sk, int size)
957 struct mptcp_sock *msk = mptcp_sk(sk);
959 /* check for pre-existing error condition */
960 if (msk->wmem_reserved < 0)
963 if (msk->wmem_reserved >= size)
967 if (!sk_wmem_schedule(sk, size)) {
968 mptcp_data_unlock(sk);
972 sk->sk_forward_alloc -= size;
973 msk->wmem_reserved += size;
974 mptcp_data_unlock(sk);
977 msk->wmem_reserved -= size;
981 static void mptcp_wmem_uncharge(struct sock *sk, int size)
983 struct mptcp_sock *msk = mptcp_sk(sk);
985 if (msk->wmem_reserved < 0)
986 msk->wmem_reserved = 0;
987 msk->wmem_reserved += size;
990 static void mptcp_mem_reclaim_partial(struct sock *sk)
992 struct mptcp_sock *msk = mptcp_sk(sk);
994 /* if we are experiencing a transint allocation error,
995 * the forward allocation memory has been already
998 if (msk->wmem_reserved < 0)
1001 mptcp_data_lock(sk);
1002 sk->sk_forward_alloc += msk->wmem_reserved;
1003 sk_mem_reclaim_partial(sk);
1004 msk->wmem_reserved = sk->sk_forward_alloc;
1005 sk->sk_forward_alloc = 0;
1006 mptcp_data_unlock(sk);
1009 static void dfrag_uncharge(struct sock *sk, int len)
1011 sk_mem_uncharge(sk, len);
1012 sk_wmem_queued_add(sk, -len);
1015 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1017 int len = dfrag->data_len + dfrag->overhead;
1019 list_del(&dfrag->list);
1020 dfrag_uncharge(sk, len);
1021 put_page(dfrag->page);
1024 static void __mptcp_clean_una(struct sock *sk)
1026 struct mptcp_sock *msk = mptcp_sk(sk);
1027 struct mptcp_data_frag *dtmp, *dfrag;
1028 bool cleaned = false;
1031 /* on fallback we just need to ignore snd_una, as this is really
1034 if (__mptcp_check_fallback(msk))
1035 msk->snd_una = READ_ONCE(msk->snd_nxt);
1037 snd_una = msk->snd_una;
1038 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1039 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1042 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1044 dfrag_clear(sk, dfrag);
1048 dfrag = mptcp_rtx_head(sk);
1049 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1050 u64 delta = snd_una - dfrag->data_seq;
1052 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1055 dfrag->data_seq += delta;
1056 dfrag->offset += delta;
1057 dfrag->data_len -= delta;
1058 dfrag->already_sent -= delta;
1060 dfrag_uncharge(sk, delta);
1066 if (tcp_under_memory_pressure(sk)) {
1067 __mptcp_update_wmem(sk);
1068 sk_mem_reclaim_partial(sk);
1072 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1073 if (msk->timer_ival && !mptcp_data_fin_enabled(msk))
1074 mptcp_stop_timer(sk);
1076 mptcp_reset_timer(sk);
1080 static void __mptcp_clean_una_wakeup(struct sock *sk)
1082 __mptcp_clean_una(sk);
1083 mptcp_write_space(sk);
1086 static void mptcp_enter_memory_pressure(struct sock *sk)
1088 struct mptcp_subflow_context *subflow;
1089 struct mptcp_sock *msk = mptcp_sk(sk);
1092 sk_stream_moderate_sndbuf(sk);
1093 mptcp_for_each_subflow(msk, subflow) {
1094 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1097 tcp_enter_memory_pressure(ssk);
1098 sk_stream_moderate_sndbuf(ssk);
1103 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1106 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1108 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1109 pfrag, sk->sk_allocation)))
1112 mptcp_enter_memory_pressure(sk);
1116 static struct mptcp_data_frag *
1117 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1120 int offset = ALIGN(orig_offset, sizeof(long));
1121 struct mptcp_data_frag *dfrag;
1123 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1124 dfrag->data_len = 0;
1125 dfrag->data_seq = msk->write_seq;
1126 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1127 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1128 dfrag->already_sent = 0;
1129 dfrag->page = pfrag->page;
1134 struct mptcp_sendmsg_info {
1142 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1145 u64 window_end = mptcp_wnd_end(msk);
1147 if (__mptcp_check_fallback(msk))
1150 if (!before64(data_seq + avail_size, window_end)) {
1151 u64 allowed_size = window_end - data_seq;
1153 return min_t(unsigned int, allowed_size, avail_size);
1159 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1161 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1165 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1169 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1171 struct sk_buff *skb;
1173 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1175 if (likely(__mptcp_add_ext(skb, gfp))) {
1176 skb_reserve(skb, MAX_TCP_HEADER);
1177 skb->reserved_tailroom = skb->end - skb->tail;
1182 mptcp_enter_memory_pressure(sk);
1187 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1188 struct sk_buff_head *skbs, int *total_ts)
1190 struct mptcp_sock *msk = mptcp_sk(sk);
1191 struct sk_buff *skb;
1194 if (unlikely(tcp_under_memory_pressure(sk))) {
1195 mptcp_mem_reclaim_partial(sk);
1197 /* under pressure pre-allocate at most a single skb */
1198 if (msk->skb_tx_cache.qlen)
1200 space_needed = msk->size_goal_cache;
1202 space_needed = msk->tx_pending_data + size -
1203 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1206 while (space_needed > 0) {
1207 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1208 if (unlikely(!skb)) {
1209 /* under memory pressure, try to pass the caller a
1210 * single skb to allow forward progress
1212 while (skbs->qlen > 1) {
1213 skb = __skb_dequeue_tail(skbs);
1214 *total_ts -= skb->truesize;
1217 return skbs->qlen > 0;
1220 *total_ts += skb->truesize;
1221 __skb_queue_tail(skbs, skb);
1222 space_needed -= msk->size_goal_cache;
1227 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1229 struct mptcp_sock *msk = mptcp_sk(sk);
1230 struct sk_buff *skb;
1232 if (ssk->sk_tx_skb_cache) {
1233 skb = ssk->sk_tx_skb_cache;
1234 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1235 !__mptcp_add_ext(skb, gfp)))
1240 skb = skb_peek(&msk->skb_tx_cache);
1242 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1243 skb = __skb_dequeue(&msk->skb_tx_cache);
1244 if (WARN_ON_ONCE(!skb))
1247 mptcp_wmem_uncharge(sk, skb->truesize);
1248 ssk->sk_tx_skb_cache = skb;
1252 /* over memory limit, no point to try to allocate a new skb */
1256 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1260 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1261 ssk->sk_tx_skb_cache = skb;
1268 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1270 return !ssk->sk_tx_skb_cache &&
1271 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1272 tcp_under_memory_pressure(sk);
1275 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1277 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1278 mptcp_mem_reclaim_partial(sk);
1279 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1282 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1283 struct mptcp_data_frag *dfrag,
1284 struct mptcp_sendmsg_info *info)
1286 u64 data_seq = dfrag->data_seq + info->sent;
1287 struct mptcp_sock *msk = mptcp_sk(sk);
1288 bool zero_window_probe = false;
1289 struct mptcp_ext *mpext = NULL;
1290 struct sk_buff *skb, *tail;
1291 bool can_collapse = false;
1296 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1297 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1299 /* compute send limit */
1300 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1301 avail_size = info->size_goal;
1302 msk->size_goal_cache = info->size_goal;
1303 skb = tcp_write_queue_tail(ssk);
1305 /* Limit the write to the size available in the
1306 * current skb, if any, so that we create at most a new skb.
1307 * Explicitly tells TCP internals to avoid collapsing on later
1308 * queue management operation, to avoid breaking the ext <->
1309 * SSN association set here
1311 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1312 can_collapse = (info->size_goal - skb->len > 0) &&
1313 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1314 if (!can_collapse) {
1315 TCP_SKB_CB(skb)->eor = 1;
1317 size_bias = skb->len;
1318 avail_size = info->size_goal - skb->len;
1322 /* Zero window and all data acked? Probe. */
1323 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1324 if (avail_size == 0) {
1325 u64 snd_una = READ_ONCE(msk->snd_una);
1327 if (skb || snd_una != msk->snd_nxt)
1329 zero_window_probe = true;
1330 data_seq = snd_una - 1;
1334 if (WARN_ON_ONCE(info->sent > info->limit ||
1335 info->limit > dfrag->data_len))
1338 ret = info->limit - info->sent;
1339 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1340 dfrag->page, dfrag->offset + info->sent, &ret);
1342 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1346 /* if the tail skb is still the cached one, collapsing really happened.
1349 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1350 mpext->data_len += ret;
1351 WARN_ON_ONCE(!can_collapse);
1352 WARN_ON_ONCE(zero_window_probe);
1356 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1357 if (WARN_ON_ONCE(!mpext)) {
1358 /* should never reach here, stream corrupted */
1362 memset(mpext, 0, sizeof(*mpext));
1363 mpext->data_seq = data_seq;
1364 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1365 mpext->data_len = ret;
1369 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1370 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1373 if (zero_window_probe) {
1374 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1377 tcp_push_pending_frames(ssk);
1380 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1384 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1385 sizeof(struct tcphdr) - \
1386 MAX_TCP_OPTION_SPACE - \
1387 sizeof(struct ipv6hdr) - \
1388 sizeof(struct frag_hdr))
1390 struct subflow_send_info {
1395 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1397 struct subflow_send_info send_info[2];
1398 struct mptcp_subflow_context *subflow;
1399 int i, nr_active = 0;
1404 sock_owned_by_me((struct sock *)msk);
1406 if (__mptcp_check_fallback(msk)) {
1409 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1412 /* re-use last subflow, if the burst allow that */
1413 if (msk->last_snd && msk->snd_burst > 0 &&
1414 sk_stream_memory_free(msk->last_snd) &&
1415 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1416 return msk->last_snd;
1418 /* pick the subflow with the lower wmem/wspace ratio */
1419 for (i = 0; i < 2; ++i) {
1420 send_info[i].ssk = NULL;
1421 send_info[i].ratio = -1;
1423 mptcp_for_each_subflow(msk, subflow) {
1424 trace_mptcp_subflow_get_send(subflow);
1425 ssk = mptcp_subflow_tcp_sock(subflow);
1426 if (!mptcp_subflow_active(subflow))
1429 nr_active += !subflow->backup;
1430 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1433 pace = READ_ONCE(ssk->sk_pacing_rate);
1437 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1439 if (ratio < send_info[subflow->backup].ratio) {
1440 send_info[subflow->backup].ssk = ssk;
1441 send_info[subflow->backup].ratio = ratio;
1445 /* pick the best backup if no other subflow is active */
1447 send_info[0].ssk = send_info[1].ssk;
1449 if (send_info[0].ssk) {
1450 msk->last_snd = send_info[0].ssk;
1451 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1452 tcp_sk(msk->last_snd)->snd_wnd);
1453 return msk->last_snd;
1459 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1460 struct mptcp_sendmsg_info *info)
1462 mptcp_set_timeout(sk, ssk);
1463 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1467 static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1469 struct sock *prev_ssk = NULL, *ssk = NULL;
1470 struct mptcp_sock *msk = mptcp_sk(sk);
1471 struct mptcp_sendmsg_info info = {
1474 struct mptcp_data_frag *dfrag;
1475 int len, copied = 0;
1477 while ((dfrag = mptcp_send_head(sk))) {
1478 info.sent = dfrag->already_sent;
1479 info.limit = dfrag->data_len;
1480 len = dfrag->data_len - dfrag->already_sent;
1485 mptcp_flush_join_list(msk);
1486 ssk = mptcp_subflow_get_send(msk);
1488 /* try to keep the subflow socket lock across
1489 * consecutive xmit on the same socket
1491 if (ssk != prev_ssk && prev_ssk)
1492 mptcp_push_release(sk, prev_ssk, &info);
1496 if (ssk != prev_ssk || !prev_ssk)
1499 /* keep it simple and always provide a new skb for the
1500 * subflow, even if we will not use it when collapsing
1501 * on the pending one
1503 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1504 mptcp_push_release(sk, ssk, &info);
1508 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1510 mptcp_push_release(sk, ssk, &info);
1515 dfrag->already_sent += ret;
1516 msk->snd_nxt += ret;
1517 msk->snd_burst -= ret;
1518 msk->tx_pending_data -= ret;
1522 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1525 /* at this point we held the socket lock for the last subflow we used */
1527 mptcp_push_release(sk, ssk, &info);
1531 /* start the timer, if it's not pending */
1532 if (!mptcp_timer_pending(sk))
1533 mptcp_reset_timer(sk);
1534 __mptcp_check_send_data_fin(sk);
1538 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1540 struct mptcp_sock *msk = mptcp_sk(sk);
1541 struct mptcp_sendmsg_info info;
1542 struct mptcp_data_frag *dfrag;
1543 struct sock *xmit_ssk;
1544 int len, copied = 0;
1548 while ((dfrag = mptcp_send_head(sk))) {
1549 info.sent = dfrag->already_sent;
1550 info.limit = dfrag->data_len;
1551 len = dfrag->data_len - dfrag->already_sent;
1555 /* the caller already invoked the packet scheduler,
1556 * check for a different subflow usage only after
1557 * spooling the first chunk of data
1559 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1562 if (xmit_ssk != ssk) {
1563 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1567 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1568 __mptcp_update_wmem(sk);
1569 sk_mem_reclaim_partial(sk);
1571 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1574 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1579 dfrag->already_sent += ret;
1580 msk->snd_nxt += ret;
1581 msk->snd_burst -= ret;
1582 msk->tx_pending_data -= ret;
1587 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1591 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1592 * not going to flush it via release_sock()
1594 __mptcp_update_wmem(sk);
1596 mptcp_set_timeout(sk, ssk);
1597 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1599 if (!mptcp_timer_pending(sk))
1600 mptcp_reset_timer(sk);
1602 if (msk->snd_data_fin_enable &&
1603 msk->snd_nxt + 1 == msk->write_seq)
1604 mptcp_schedule_work(sk);
1608 static void mptcp_set_nospace(struct sock *sk)
1610 /* enable autotune */
1611 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1613 /* will be cleared on avail space */
1614 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1617 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1619 struct mptcp_sock *msk = mptcp_sk(sk);
1620 struct page_frag *pfrag;
1625 /* we don't support FASTOPEN yet */
1626 if (msg->msg_flags & MSG_FASTOPEN)
1629 /* silently ignore everything else */
1630 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1632 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1634 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1636 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1637 ret = sk_stream_wait_connect(sk, &timeo);
1642 pfrag = sk_page_frag(sk);
1644 while (msg_data_left(msg)) {
1645 int total_ts, frag_truesize = 0;
1646 struct mptcp_data_frag *dfrag;
1647 struct sk_buff_head skbs;
1648 bool dfrag_collapsed;
1649 size_t psize, offset;
1651 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1656 /* reuse tail pfrag, if possible, or carve a new one from the
1659 dfrag = mptcp_pending_tail(sk);
1660 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1661 if (!dfrag_collapsed) {
1662 if (!sk_stream_memory_free(sk))
1663 goto wait_for_memory;
1665 if (!mptcp_page_frag_refill(sk, pfrag))
1666 goto wait_for_memory;
1668 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1669 frag_truesize = dfrag->overhead;
1672 /* we do not bound vs wspace, to allow a single packet.
1673 * memory accounting will prevent execessive memory usage
1676 offset = dfrag->offset + dfrag->data_len;
1677 psize = pfrag->size - offset;
1678 psize = min_t(size_t, psize, msg_data_left(msg));
1679 total_ts = psize + frag_truesize;
1680 __skb_queue_head_init(&skbs);
1681 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1682 goto wait_for_memory;
1684 if (!mptcp_wmem_alloc(sk, total_ts)) {
1685 __skb_queue_purge(&skbs);
1686 goto wait_for_memory;
1689 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1690 if (copy_page_from_iter(dfrag->page, offset, psize,
1691 &msg->msg_iter) != psize) {
1692 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1697 /* data successfully copied into the write queue */
1699 dfrag->data_len += psize;
1700 frag_truesize += psize;
1701 pfrag->offset += frag_truesize;
1702 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1703 msk->tx_pending_data += psize;
1705 /* charge data on mptcp pending queue to the msk socket
1706 * Note: we charge such data both to sk and ssk
1708 sk_wmem_queued_add(sk, frag_truesize);
1709 if (!dfrag_collapsed) {
1710 get_page(dfrag->page);
1711 list_add_tail(&dfrag->list, &msk->rtx_queue);
1712 if (!msk->first_pending)
1713 WRITE_ONCE(msk->first_pending, dfrag);
1715 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1716 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1722 mptcp_set_nospace(sk);
1723 __mptcp_push_pending(sk, msg->msg_flags);
1724 ret = sk_stream_wait_memory(sk, &timeo);
1730 __mptcp_push_pending(sk, msg->msg_flags);
1734 return copied ? : ret;
1737 static void mptcp_wait_data(struct sock *sk, long *timeo)
1739 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1740 struct mptcp_sock *msk = mptcp_sk(sk);
1742 add_wait_queue(sk_sleep(sk), &wait);
1743 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1745 sk_wait_event(sk, timeo,
1746 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1748 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1749 remove_wait_queue(sk_sleep(sk), &wait);
1752 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1754 size_t len, int flags)
1756 struct sk_buff *skb, *tmp;
1759 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1760 u32 offset = MPTCP_SKB_CB(skb)->offset;
1761 u32 data_len = skb->len - offset;
1762 u32 count = min_t(size_t, len - copied, data_len);
1765 if (!(flags & MSG_TRUNC)) {
1766 err = skb_copy_datagram_msg(skb, offset, msg, count);
1767 if (unlikely(err < 0)) {
1776 if (count < data_len) {
1777 if (!(flags & MSG_PEEK))
1778 MPTCP_SKB_CB(skb)->offset += count;
1782 if (!(flags & MSG_PEEK)) {
1783 /* we will bulk release the skb memory later */
1784 skb->destructor = NULL;
1785 msk->rmem_released += skb->truesize;
1786 __skb_unlink(skb, &msk->receive_queue);
1797 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1799 * Only difference: Use highest rtt estimate of the subflows in use.
1801 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1803 struct mptcp_subflow_context *subflow;
1804 struct sock *sk = (struct sock *)msk;
1805 u32 time, advmss = 1;
1808 sock_owned_by_me(sk);
1813 msk->rcvq_space.copied += copied;
1815 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1816 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1818 rtt_us = msk->rcvq_space.rtt_us;
1819 if (rtt_us && time < (rtt_us >> 3))
1823 mptcp_for_each_subflow(msk, subflow) {
1824 const struct tcp_sock *tp;
1828 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1830 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1831 sf_advmss = READ_ONCE(tp->advmss);
1833 rtt_us = max(sf_rtt_us, rtt_us);
1834 advmss = max(sf_advmss, advmss);
1837 msk->rcvq_space.rtt_us = rtt_us;
1838 if (time < (rtt_us >> 3) || rtt_us == 0)
1841 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1844 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1845 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1849 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1851 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1853 do_div(grow, msk->rcvq_space.space);
1854 rcvwin += (grow << 1);
1856 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1857 while (tcp_win_from_space(sk, rcvmem) < advmss)
1860 do_div(rcvwin, advmss);
1861 rcvbuf = min_t(u64, rcvwin * rcvmem,
1862 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1864 if (rcvbuf > sk->sk_rcvbuf) {
1867 window_clamp = tcp_win_from_space(sk, rcvbuf);
1868 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1870 /* Make subflows follow along. If we do not do this, we
1871 * get drops at subflow level if skbs can't be moved to
1872 * the mptcp rx queue fast enough (announced rcv_win can
1873 * exceed ssk->sk_rcvbuf).
1875 mptcp_for_each_subflow(msk, subflow) {
1879 ssk = mptcp_subflow_tcp_sock(subflow);
1880 slow = lock_sock_fast(ssk);
1881 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1882 tcp_sk(ssk)->window_clamp = window_clamp;
1883 tcp_cleanup_rbuf(ssk, 1);
1884 unlock_sock_fast(ssk, slow);
1889 msk->rcvq_space.space = msk->rcvq_space.copied;
1891 msk->rcvq_space.copied = 0;
1892 msk->rcvq_space.time = mstamp;
1895 static void __mptcp_update_rmem(struct sock *sk)
1897 struct mptcp_sock *msk = mptcp_sk(sk);
1899 if (!msk->rmem_released)
1902 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1903 sk_mem_uncharge(sk, msk->rmem_released);
1904 msk->rmem_released = 0;
1907 static void __mptcp_splice_receive_queue(struct sock *sk)
1909 struct mptcp_sock *msk = mptcp_sk(sk);
1911 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1914 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1916 struct sock *sk = (struct sock *)msk;
1917 unsigned int moved = 0;
1920 mptcp_flush_join_list(msk);
1922 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1925 /* we can have data pending in the subflows only if the msk
1926 * receive buffer was full at subflow_data_ready() time,
1927 * that is an unlikely slow path.
1932 slowpath = lock_sock_fast(ssk);
1933 mptcp_data_lock(sk);
1934 __mptcp_update_rmem(sk);
1935 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1936 mptcp_data_unlock(sk);
1937 tcp_cleanup_rbuf(ssk, moved);
1938 unlock_sock_fast(ssk, slowpath);
1941 /* acquire the data lock only if some input data is pending */
1943 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1944 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1945 mptcp_data_lock(sk);
1946 __mptcp_update_rmem(sk);
1947 ret |= __mptcp_ofo_queue(msk);
1948 __mptcp_splice_receive_queue(sk);
1949 mptcp_data_unlock(sk);
1950 mptcp_cleanup_rbuf(msk);
1953 mptcp_check_data_fin((struct sock *)msk);
1954 return !skb_queue_empty(&msk->receive_queue);
1957 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1958 int nonblock, int flags, int *addr_len)
1960 struct mptcp_sock *msk = mptcp_sk(sk);
1965 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
1966 if (unlikely(flags & MSG_ERRQUEUE))
1967 return inet_recv_error(sk, msg, len, addr_len);
1969 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1970 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1975 timeo = sock_rcvtimeo(sk, nonblock);
1977 len = min_t(size_t, len, INT_MAX);
1978 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1980 while (copied < len) {
1983 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags);
1984 if (unlikely(bytes_read < 0)) {
1986 copied = bytes_read;
1990 copied += bytes_read;
1992 /* be sure to advertise window change */
1993 mptcp_cleanup_rbuf(msk);
1995 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
1998 /* only the master socket status is relevant here. The exit
1999 * conditions mirror closely tcp_recvmsg()
2001 if (copied >= target)
2006 sk->sk_state == TCP_CLOSE ||
2007 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2009 signal_pending(current))
2013 copied = sock_error(sk);
2017 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2018 mptcp_check_for_eof(msk);
2020 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2021 /* race breaker: the shutdown could be after the
2022 * previous receive queue check
2024 if (__mptcp_move_skbs(msk))
2029 if (sk->sk_state == TCP_CLOSE) {
2039 if (signal_pending(current)) {
2040 copied = sock_intr_errno(timeo);
2045 pr_debug("block timeout %ld", timeo);
2046 mptcp_wait_data(sk, &timeo);
2049 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2050 skb_queue_empty(&msk->receive_queue)) {
2051 /* entire backlog drained, clear DATA_READY. */
2052 clear_bit(MPTCP_DATA_READY, &msk->flags);
2054 /* .. race-breaker: ssk might have gotten new data
2055 * after last __mptcp_move_skbs() returned false.
2057 if (unlikely(__mptcp_move_skbs(msk)))
2058 set_bit(MPTCP_DATA_READY, &msk->flags);
2059 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2060 /* data to read but mptcp_wait_data() cleared DATA_READY */
2061 set_bit(MPTCP_DATA_READY, &msk->flags);
2064 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2065 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2066 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2067 if (!(flags & MSG_PEEK))
2068 mptcp_rcv_space_adjust(msk, copied);
2074 static void mptcp_retransmit_timer(struct timer_list *t)
2076 struct inet_connection_sock *icsk = from_timer(icsk, t,
2077 icsk_retransmit_timer);
2078 struct sock *sk = &icsk->icsk_inet.sk;
2079 struct mptcp_sock *msk = mptcp_sk(sk);
2082 if (!sock_owned_by_user(sk)) {
2083 /* we need a process context to retransmit */
2084 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2085 mptcp_schedule_work(sk);
2087 /* delegate our work to tcp_release_cb() */
2088 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2094 static void mptcp_timeout_timer(struct timer_list *t)
2096 struct sock *sk = from_timer(sk, t, sk_timer);
2098 mptcp_schedule_work(sk);
2102 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2105 * A backup subflow is returned only if that is the only kind available.
2107 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2109 struct mptcp_subflow_context *subflow;
2110 struct sock *backup = NULL;
2112 sock_owned_by_me((const struct sock *)msk);
2114 if (__mptcp_check_fallback(msk))
2117 mptcp_for_each_subflow(msk, subflow) {
2118 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2120 if (!mptcp_subflow_active(subflow))
2123 /* still data outstanding at TCP level? Don't retransmit. */
2124 if (!tcp_write_queue_empty(ssk)) {
2125 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2130 if (subflow->backup) {
2142 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2145 iput(SOCK_INODE(msk->subflow));
2146 msk->subflow = NULL;
2150 /* subflow sockets can be either outgoing (connect) or incoming
2153 * Outgoing subflows use in-kernel sockets.
2154 * Incoming subflows do not have their own 'struct socket' allocated,
2155 * so we need to use tcp_close() after detaching them from the mptcp
2158 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2159 struct mptcp_subflow_context *subflow)
2161 struct mptcp_sock *msk = mptcp_sk(sk);
2163 list_del(&subflow->node);
2165 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2167 /* if we are invoked by the msk cleanup code, the subflow is
2173 subflow->disposable = 1;
2175 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2176 * the ssk has been already destroyed, we just need to release the
2177 * reference owned by msk;
2179 if (!inet_csk(ssk)->icsk_ulp_ops) {
2180 kfree_rcu(subflow, rcu);
2182 /* otherwise tcp will dispose of the ssk and subflow ctx */
2183 __tcp_close(ssk, 0);
2185 /* close acquired an extra ref */
2192 if (ssk == msk->last_snd)
2193 msk->last_snd = NULL;
2195 if (ssk == msk->ack_hint)
2196 msk->ack_hint = NULL;
2198 if (ssk == msk->first)
2201 if (msk->subflow && ssk == msk->subflow->sk)
2202 mptcp_dispose_initial_subflow(msk);
2205 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2206 struct mptcp_subflow_context *subflow)
2208 if (sk->sk_state == TCP_ESTABLISHED)
2209 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2210 __mptcp_close_ssk(sk, ssk, subflow);
2213 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2218 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2220 struct mptcp_subflow_context *subflow, *tmp;
2224 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2225 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2227 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2230 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2231 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2234 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2238 static bool mptcp_check_close_timeout(const struct sock *sk)
2240 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2241 struct mptcp_subflow_context *subflow;
2243 if (delta >= TCP_TIMEWAIT_LEN)
2246 /* if all subflows are in closed status don't bother with additional
2249 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2250 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2257 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2259 struct mptcp_subflow_context *subflow, *tmp;
2260 struct sock *sk = &msk->sk.icsk_inet.sk;
2262 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2265 mptcp_token_destroy(msk);
2267 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2268 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2271 if (tcp_sk->sk_state != TCP_CLOSE) {
2272 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2273 tcp_set_state(tcp_sk, TCP_CLOSE);
2275 release_sock(tcp_sk);
2278 inet_sk_state_store(sk, TCP_CLOSE);
2279 sk->sk_shutdown = SHUTDOWN_MASK;
2280 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2281 set_bit(MPTCP_DATA_READY, &msk->flags);
2282 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2284 mptcp_close_wake_up(sk);
2287 static void __mptcp_retrans(struct sock *sk)
2289 struct mptcp_sock *msk = mptcp_sk(sk);
2290 struct mptcp_sendmsg_info info = {};
2291 struct mptcp_data_frag *dfrag;
2296 __mptcp_clean_una_wakeup(sk);
2297 dfrag = mptcp_rtx_head(sk);
2299 if (mptcp_data_fin_enabled(msk)) {
2300 struct inet_connection_sock *icsk = inet_csk(sk);
2302 icsk->icsk_retransmits++;
2303 mptcp_set_datafin_timeout(sk);
2304 mptcp_send_ack(msk);
2312 ssk = mptcp_subflow_get_retrans(msk);
2318 /* limit retransmission to the bytes already sent on some subflows */
2320 info.limit = dfrag->already_sent;
2321 while (info.sent < dfrag->already_sent) {
2322 if (!mptcp_alloc_tx_skb(sk, ssk))
2325 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2329 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2334 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2337 mptcp_set_timeout(sk, ssk);
2341 if (!mptcp_timer_pending(sk))
2342 mptcp_reset_timer(sk);
2345 static void mptcp_worker(struct work_struct *work)
2347 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2348 struct sock *sk = &msk->sk.icsk_inet.sk;
2352 state = sk->sk_state;
2353 if (unlikely(state == TCP_CLOSE))
2356 mptcp_check_data_fin_ack(sk);
2357 mptcp_flush_join_list(msk);
2359 mptcp_check_fastclose(msk);
2362 mptcp_pm_nl_work(msk);
2364 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2365 mptcp_check_for_eof(msk);
2367 __mptcp_check_send_data_fin(sk);
2368 mptcp_check_data_fin(sk);
2370 /* There is no point in keeping around an orphaned sk timedout or
2371 * closed, but we need the msk around to reply to incoming DATA_FIN,
2372 * even if it is orphaned and in FIN_WAIT2 state
2374 if (sock_flag(sk, SOCK_DEAD) &&
2375 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2376 inet_sk_state_store(sk, TCP_CLOSE);
2377 __mptcp_destroy_sock(sk);
2381 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2382 __mptcp_close_subflow(msk);
2384 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2385 __mptcp_retrans(sk);
2392 static int __mptcp_init_sock(struct sock *sk)
2394 struct mptcp_sock *msk = mptcp_sk(sk);
2396 spin_lock_init(&msk->join_list_lock);
2398 INIT_LIST_HEAD(&msk->conn_list);
2399 INIT_LIST_HEAD(&msk->join_list);
2400 INIT_LIST_HEAD(&msk->rtx_queue);
2401 INIT_WORK(&msk->work, mptcp_worker);
2402 __skb_queue_head_init(&msk->receive_queue);
2403 __skb_queue_head_init(&msk->skb_tx_cache);
2404 msk->out_of_order_queue = RB_ROOT;
2405 msk->first_pending = NULL;
2406 msk->wmem_reserved = 0;
2407 msk->rmem_released = 0;
2408 msk->tx_pending_data = 0;
2409 msk->size_goal_cache = TCP_BASE_MSS;
2411 msk->ack_hint = NULL;
2413 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2415 mptcp_pm_data_init(msk);
2417 /* re-use the csk retrans timer for MPTCP-level retrans */
2418 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2419 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2421 tcp_assign_congestion_control(sk);
2426 static int mptcp_init_sock(struct sock *sk)
2428 struct net *net = sock_net(sk);
2431 ret = __mptcp_init_sock(sk);
2435 if (!mptcp_is_enabled(net))
2436 return -ENOPROTOOPT;
2438 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2441 ret = __mptcp_socket_create(mptcp_sk(sk));
2445 sk_sockets_allocated_inc(sk);
2446 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2447 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2452 static void __mptcp_clear_xmit(struct sock *sk)
2454 struct mptcp_sock *msk = mptcp_sk(sk);
2455 struct mptcp_data_frag *dtmp, *dfrag;
2456 struct sk_buff *skb;
2458 WRITE_ONCE(msk->first_pending, NULL);
2459 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2460 dfrag_clear(sk, dfrag);
2461 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2462 sk->sk_forward_alloc += skb->truesize;
2467 static void mptcp_cancel_work(struct sock *sk)
2469 struct mptcp_sock *msk = mptcp_sk(sk);
2471 if (cancel_work_sync(&msk->work))
2475 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2479 switch (ssk->sk_state) {
2481 if (!(how & RCV_SHUTDOWN))
2485 tcp_disconnect(ssk, O_NONBLOCK);
2488 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2489 pr_debug("Fallback");
2490 ssk->sk_shutdown |= how;
2491 tcp_shutdown(ssk, how);
2493 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2494 mptcp_set_timeout(sk, ssk);
2496 if (!mptcp_timer_pending(sk))
2497 mptcp_reset_timer(sk);
2505 static const unsigned char new_state[16] = {
2506 /* current state: new state: action: */
2507 [0 /* (Invalid) */] = TCP_CLOSE,
2508 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2509 [TCP_SYN_SENT] = TCP_CLOSE,
2510 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2511 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2512 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2513 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2514 [TCP_CLOSE] = TCP_CLOSE,
2515 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2516 [TCP_LAST_ACK] = TCP_LAST_ACK,
2517 [TCP_LISTEN] = TCP_CLOSE,
2518 [TCP_CLOSING] = TCP_CLOSING,
2519 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2522 static int mptcp_close_state(struct sock *sk)
2524 int next = (int)new_state[sk->sk_state];
2525 int ns = next & TCP_STATE_MASK;
2527 inet_sk_state_store(sk, ns);
2529 return next & TCP_ACTION_FIN;
2532 static void __mptcp_check_send_data_fin(struct sock *sk)
2534 struct mptcp_subflow_context *subflow;
2535 struct mptcp_sock *msk = mptcp_sk(sk);
2537 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2538 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2539 msk->snd_nxt, msk->write_seq);
2541 /* we still need to enqueue subflows or not really shutting down,
2544 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2545 mptcp_send_head(sk))
2548 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2550 /* fallback socket will not get data_fin/ack, can move to the next
2553 if (__mptcp_check_fallback(msk)) {
2554 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2555 inet_sk_state_store(sk, TCP_CLOSE);
2556 mptcp_close_wake_up(sk);
2557 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2558 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2562 mptcp_flush_join_list(msk);
2563 mptcp_for_each_subflow(msk, subflow) {
2564 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2566 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2570 static void __mptcp_wr_shutdown(struct sock *sk)
2572 struct mptcp_sock *msk = mptcp_sk(sk);
2574 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2575 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2576 !!mptcp_send_head(sk));
2578 /* will be ignored by fallback sockets */
2579 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2580 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2582 __mptcp_check_send_data_fin(sk);
2585 static void __mptcp_destroy_sock(struct sock *sk)
2587 struct mptcp_subflow_context *subflow, *tmp;
2588 struct mptcp_sock *msk = mptcp_sk(sk);
2589 LIST_HEAD(conn_list);
2591 pr_debug("msk=%p", msk);
2595 /* be sure to always acquire the join list lock, to sync vs
2596 * mptcp_finish_join().
2598 spin_lock_bh(&msk->join_list_lock);
2599 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2600 spin_unlock_bh(&msk->join_list_lock);
2601 list_splice_init(&msk->conn_list, &conn_list);
2603 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2604 sk_stop_timer(sk, &sk->sk_timer);
2607 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2608 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2609 __mptcp_close_ssk(sk, ssk, subflow);
2612 sk->sk_prot->destroy(sk);
2614 WARN_ON_ONCE(msk->wmem_reserved);
2615 WARN_ON_ONCE(msk->rmem_released);
2616 sk_stream_kill_queues(sk);
2617 xfrm_sk_free_policy(sk);
2619 tcp_cleanup_congestion_control(sk);
2620 sk_refcnt_debug_release(sk);
2621 mptcp_dispose_initial_subflow(msk);
2625 static void mptcp_close(struct sock *sk, long timeout)
2627 struct mptcp_subflow_context *subflow;
2628 bool do_cancel_work = false;
2631 sk->sk_shutdown = SHUTDOWN_MASK;
2633 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2634 inet_sk_state_store(sk, TCP_CLOSE);
2638 if (mptcp_close_state(sk))
2639 __mptcp_wr_shutdown(sk);
2641 sk_stream_wait_close(sk, timeout);
2644 /* orphan all the subflows */
2645 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2646 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2647 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2648 bool slow = lock_sock_fast(ssk);
2651 unlock_sock_fast(ssk, slow);
2656 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2657 if (sk->sk_state == TCP_CLOSE) {
2658 __mptcp_destroy_sock(sk);
2659 do_cancel_work = true;
2661 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2665 mptcp_cancel_work(sk);
2667 if (mptcp_sk(sk)->token)
2668 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2673 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2675 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2676 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2677 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2679 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2680 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2683 msk6->saddr = ssk6->saddr;
2684 msk6->flow_label = ssk6->flow_label;
2688 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2689 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2690 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2691 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2692 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2693 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2696 static int mptcp_disconnect(struct sock *sk, int flags)
2698 struct mptcp_subflow_context *subflow;
2699 struct mptcp_sock *msk = mptcp_sk(sk);
2701 mptcp_do_flush_join_list(msk);
2703 mptcp_for_each_subflow(msk, subflow) {
2704 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2707 tcp_disconnect(ssk, flags);
2713 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2714 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2716 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2718 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2722 struct sock *mptcp_sk_clone(const struct sock *sk,
2723 const struct mptcp_options_received *mp_opt,
2724 struct request_sock *req)
2726 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2727 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2728 struct mptcp_sock *msk;
2734 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2735 if (nsk->sk_family == AF_INET6)
2736 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2739 __mptcp_init_sock(nsk);
2741 msk = mptcp_sk(nsk);
2742 msk->local_key = subflow_req->local_key;
2743 msk->token = subflow_req->token;
2744 msk->subflow = NULL;
2745 WRITE_ONCE(msk->fully_established, false);
2747 msk->write_seq = subflow_req->idsn + 1;
2748 msk->snd_nxt = msk->write_seq;
2749 msk->snd_una = msk->write_seq;
2750 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2751 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2753 if (mp_opt->mp_capable) {
2754 msk->can_ack = true;
2755 msk->remote_key = mp_opt->sndr_key;
2756 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2758 WRITE_ONCE(msk->ack_seq, ack_seq);
2759 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2762 sock_reset_flag(nsk, SOCK_RCU_FREE);
2763 /* will be fully established after successful MPC subflow creation */
2764 inet_sk_state_store(nsk, TCP_SYN_RECV);
2766 security_inet_csk_clone(nsk, req);
2767 bh_unlock_sock(nsk);
2769 /* keep a single reference */
2774 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2776 const struct tcp_sock *tp = tcp_sk(ssk);
2778 msk->rcvq_space.copied = 0;
2779 msk->rcvq_space.rtt_us = 0;
2781 msk->rcvq_space.time = tp->tcp_mstamp;
2783 /* initial rcv_space offering made to peer */
2784 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2785 TCP_INIT_CWND * tp->advmss);
2786 if (msk->rcvq_space.space == 0)
2787 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2789 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2792 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2795 struct mptcp_sock *msk = mptcp_sk(sk);
2796 struct socket *listener;
2799 listener = __mptcp_nmpc_socket(msk);
2800 if (WARN_ON_ONCE(!listener)) {
2805 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2806 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2810 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2811 if (sk_is_mptcp(newsk)) {
2812 struct mptcp_subflow_context *subflow;
2813 struct sock *new_mptcp_sock;
2815 subflow = mptcp_subflow_ctx(newsk);
2816 new_mptcp_sock = subflow->conn;
2818 /* is_mptcp should be false if subflow->conn is missing, see
2819 * subflow_syn_recv_sock()
2821 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2822 tcp_sk(newsk)->is_mptcp = 0;
2826 /* acquire the 2nd reference for the owning socket */
2827 sock_hold(new_mptcp_sock);
2828 newsk = new_mptcp_sock;
2829 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2831 MPTCP_INC_STATS(sock_net(sk),
2832 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2838 void mptcp_destroy_common(struct mptcp_sock *msk)
2840 struct sock *sk = (struct sock *)msk;
2842 __mptcp_clear_xmit(sk);
2844 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2845 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2847 skb_rbtree_purge(&msk->out_of_order_queue);
2848 mptcp_token_destroy(msk);
2849 mptcp_pm_free_anno_list(msk);
2852 static void mptcp_destroy(struct sock *sk)
2854 struct mptcp_sock *msk = mptcp_sk(sk);
2856 mptcp_destroy_common(msk);
2857 sk_sockets_allocated_dec(sk);
2860 void __mptcp_data_acked(struct sock *sk)
2862 if (!sock_owned_by_user(sk))
2863 __mptcp_clean_una(sk);
2865 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2867 if (mptcp_pending_data_fin_ack(sk))
2868 mptcp_schedule_work(sk);
2871 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2873 if (!mptcp_send_head(sk))
2876 if (!sock_owned_by_user(sk)) {
2877 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2879 if (xmit_ssk == ssk)
2880 __mptcp_subflow_push_pending(sk, ssk);
2882 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
2884 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2888 /* processes deferred events and flush wmem */
2889 static void mptcp_release_cb(struct sock *sk)
2892 unsigned long flags = 0;
2894 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2895 flags |= BIT(MPTCP_PUSH_PENDING);
2896 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2897 flags |= BIT(MPTCP_RETRANSMIT);
2901 /* the following actions acquire the subflow socket lock
2903 * 1) can't be invoked in atomic scope
2904 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2905 * datapath acquires the msk socket spinlock while helding
2906 * the subflow socket lock
2909 spin_unlock_bh(&sk->sk_lock.slock);
2910 if (flags & BIT(MPTCP_PUSH_PENDING))
2911 __mptcp_push_pending(sk, 0);
2912 if (flags & BIT(MPTCP_RETRANSMIT))
2913 __mptcp_retrans(sk);
2916 spin_lock_bh(&sk->sk_lock.slock);
2919 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2920 __mptcp_clean_una_wakeup(sk);
2921 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2922 __mptcp_error_report(sk);
2924 /* push_pending may touch wmem_reserved, ensure we do the cleanup
2927 __mptcp_update_wmem(sk);
2928 __mptcp_update_rmem(sk);
2931 void mptcp_subflow_process_delegated(struct sock *ssk)
2933 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2934 struct sock *sk = subflow->conn;
2936 mptcp_data_lock(sk);
2937 if (!sock_owned_by_user(sk))
2938 __mptcp_subflow_push_pending(sk, ssk);
2940 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2941 mptcp_data_unlock(sk);
2942 mptcp_subflow_delegated_done(subflow);
2945 static int mptcp_hash(struct sock *sk)
2947 /* should never be called,
2948 * we hash the TCP subflows not the master socket
2954 static void mptcp_unhash(struct sock *sk)
2956 /* called from sk_common_release(), but nothing to do here */
2959 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2961 struct mptcp_sock *msk = mptcp_sk(sk);
2962 struct socket *ssock;
2964 ssock = __mptcp_nmpc_socket(msk);
2965 pr_debug("msk=%p, subflow=%p", msk, ssock);
2966 if (WARN_ON_ONCE(!ssock))
2969 return inet_csk_get_port(ssock->sk, snum);
2972 void mptcp_finish_connect(struct sock *ssk)
2974 struct mptcp_subflow_context *subflow;
2975 struct mptcp_sock *msk;
2979 subflow = mptcp_subflow_ctx(ssk);
2983 pr_debug("msk=%p, token=%u", sk, subflow->token);
2985 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2987 subflow->map_seq = ack_seq;
2988 subflow->map_subflow_seq = 1;
2990 /* the socket is not connected yet, no msk/subflow ops can access/race
2991 * accessing the field below
2993 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2994 WRITE_ONCE(msk->local_key, subflow->local_key);
2995 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2996 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2997 WRITE_ONCE(msk->ack_seq, ack_seq);
2998 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2999 WRITE_ONCE(msk->can_ack, 1);
3000 WRITE_ONCE(msk->snd_una, msk->write_seq);
3002 mptcp_pm_new_connection(msk, ssk, 0);
3004 mptcp_rcv_space_init(msk, ssk);
3007 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3009 write_lock_bh(&sk->sk_callback_lock);
3010 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3011 sk_set_socket(sk, parent);
3012 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3013 write_unlock_bh(&sk->sk_callback_lock);
3016 bool mptcp_finish_join(struct sock *ssk)
3018 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3019 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3020 struct sock *parent = (void *)msk;
3021 struct socket *parent_sock;
3024 pr_debug("msk=%p, subflow=%p", msk, subflow);
3026 /* mptcp socket already closing? */
3027 if (!mptcp_is_fully_established(parent)) {
3028 subflow->reset_reason = MPTCP_RST_EMPTCP;
3032 if (!msk->pm.server_side)
3035 if (!mptcp_pm_allow_new_subflow(msk)) {
3036 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3040 /* active connections are already on conn_list, and we can't acquire
3042 * use the join list lock as synchronization point and double-check
3043 * msk status to avoid racing with __mptcp_destroy_sock()
3045 spin_lock_bh(&msk->join_list_lock);
3046 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3047 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3048 list_add_tail(&subflow->node, &msk->join_list);
3051 spin_unlock_bh(&msk->join_list_lock);
3053 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3057 /* attach to msk socket only after we are sure he will deal with us
3060 parent_sock = READ_ONCE(parent->sk_socket);
3061 if (parent_sock && !ssk->sk_socket)
3062 mptcp_sock_graft(ssk, parent_sock);
3063 subflow->map_seq = READ_ONCE(msk->ack_seq);
3065 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
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_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3169 mptcp_subflow_early_fallback(msk, subflow);
3171 if (likely(!__mptcp_check_fallback(msk)))
3172 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3175 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3176 sock->state = ssock->state;
3178 /* on successful connect, the msk state will be moved to established by
3179 * subflow_finish_connect()
3181 if (!err || err == -EINPROGRESS)
3182 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3184 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3187 release_sock(sock->sk);
3191 static int mptcp_listen(struct socket *sock, int backlog)
3193 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3194 struct socket *ssock;
3197 pr_debug("msk=%p", msk);
3199 lock_sock(sock->sk);
3200 ssock = __mptcp_nmpc_socket(msk);
3206 mptcp_token_destroy(msk);
3207 inet_sk_state_store(sock->sk, TCP_LISTEN);
3208 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3210 err = ssock->ops->listen(ssock, backlog);
3211 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3213 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3216 release_sock(sock->sk);
3220 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3221 int flags, bool kern)
3223 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3224 struct socket *ssock;
3227 pr_debug("msk=%p", msk);
3229 lock_sock(sock->sk);
3230 if (sock->sk->sk_state != TCP_LISTEN)
3233 ssock = __mptcp_nmpc_socket(msk);
3237 clear_bit(MPTCP_DATA_READY, &msk->flags);
3238 sock_hold(ssock->sk);
3239 release_sock(sock->sk);
3241 err = ssock->ops->accept(sock, newsock, flags, kern);
3242 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3243 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3244 struct mptcp_subflow_context *subflow;
3245 struct sock *newsk = newsock->sk;
3249 /* PM/worker can now acquire the first subflow socket
3250 * lock without racing with listener queue cleanup,
3251 * we can notify it, if needed.
3253 * Even if remote has reset the initial subflow by now
3254 * the refcnt is still at least one.
3256 subflow = mptcp_subflow_ctx(msk->first);
3257 list_add(&subflow->node, &msk->conn_list);
3258 sock_hold(msk->first);
3259 if (mptcp_is_fully_established(newsk))
3260 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3262 mptcp_copy_inaddrs(newsk, msk->first);
3263 mptcp_rcv_space_init(msk, msk->first);
3264 mptcp_propagate_sndbuf(newsk, msk->first);
3266 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3267 * This is needed so NOSPACE flag can be set from tcp stack.
3269 mptcp_flush_join_list(msk);
3270 mptcp_for_each_subflow(msk, subflow) {
3271 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3273 if (!ssk->sk_socket)
3274 mptcp_sock_graft(ssk, newsock);
3276 release_sock(newsk);
3279 if (inet_csk_listen_poll(ssock->sk))
3280 set_bit(MPTCP_DATA_READY, &msk->flags);
3281 sock_put(ssock->sk);
3285 release_sock(sock->sk);
3289 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3291 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3295 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3297 struct sock *sk = (struct sock *)msk;
3299 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3300 return EPOLLOUT | EPOLLWRNORM;
3302 if (sk_stream_is_writeable(sk))
3303 return EPOLLOUT | EPOLLWRNORM;
3305 mptcp_set_nospace(sk);
3306 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3307 if (sk_stream_is_writeable(sk))
3308 return EPOLLOUT | EPOLLWRNORM;
3313 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3314 struct poll_table_struct *wait)
3316 struct sock *sk = sock->sk;
3317 struct mptcp_sock *msk;
3322 sock_poll_wait(file, sock, wait);
3324 state = inet_sk_state_load(sk);
3325 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3326 if (state == TCP_LISTEN)
3327 return mptcp_check_readable(msk);
3329 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3330 mask |= mptcp_check_readable(msk);
3331 mask |= mptcp_check_writeable(msk);
3333 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3335 if (sk->sk_shutdown & RCV_SHUTDOWN)
3336 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3338 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3346 static const struct proto_ops mptcp_stream_ops = {
3348 .owner = THIS_MODULE,
3349 .release = inet_release,
3351 .connect = mptcp_stream_connect,
3352 .socketpair = sock_no_socketpair,
3353 .accept = mptcp_stream_accept,
3354 .getname = inet_getname,
3356 .ioctl = inet_ioctl,
3357 .gettstamp = sock_gettstamp,
3358 .listen = mptcp_listen,
3359 .shutdown = inet_shutdown,
3360 .setsockopt = sock_common_setsockopt,
3361 .getsockopt = sock_common_getsockopt,
3362 .sendmsg = inet_sendmsg,
3363 .recvmsg = inet_recvmsg,
3364 .mmap = sock_no_mmap,
3365 .sendpage = inet_sendpage,
3368 static struct inet_protosw mptcp_protosw = {
3369 .type = SOCK_STREAM,
3370 .protocol = IPPROTO_MPTCP,
3371 .prot = &mptcp_prot,
3372 .ops = &mptcp_stream_ops,
3373 .flags = INET_PROTOSW_ICSK,
3376 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3378 struct mptcp_delegated_action *delegated;
3379 struct mptcp_subflow_context *subflow;
3382 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3383 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3384 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3386 bh_lock_sock_nested(ssk);
3387 if (!sock_owned_by_user(ssk) &&
3388 mptcp_subflow_has_delegated_action(subflow))
3389 mptcp_subflow_process_delegated(ssk);
3390 /* ... elsewhere tcp_release_cb_override already processed
3391 * the action or will do at next release_sock().
3392 * In both case must dequeue the subflow here - on the same
3393 * CPU that scheduled it.
3395 bh_unlock_sock(ssk);
3398 if (++work_done == budget)
3402 /* always provide a 0 'work_done' argument, so that napi_complete_done
3403 * will not try accessing the NULL napi->dev ptr
3405 napi_complete_done(napi, 0);
3409 void __init mptcp_proto_init(void)
3411 struct mptcp_delegated_action *delegated;
3414 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3416 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3417 panic("Failed to allocate MPTCP pcpu counter\n");
3419 init_dummy_netdev(&mptcp_napi_dev);
3420 for_each_possible_cpu(cpu) {
3421 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3422 INIT_LIST_HEAD(&delegated->head);
3423 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3425 napi_enable(&delegated->napi);
3428 mptcp_subflow_init();
3432 if (proto_register(&mptcp_prot, 1) != 0)
3433 panic("Failed to register MPTCP proto.\n");
3435 inet_register_protosw(&mptcp_protosw);
3437 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3440 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3441 static const struct proto_ops mptcp_v6_stream_ops = {
3443 .owner = THIS_MODULE,
3444 .release = inet6_release,
3446 .connect = mptcp_stream_connect,
3447 .socketpair = sock_no_socketpair,
3448 .accept = mptcp_stream_accept,
3449 .getname = inet6_getname,
3451 .ioctl = inet6_ioctl,
3452 .gettstamp = sock_gettstamp,
3453 .listen = mptcp_listen,
3454 .shutdown = inet_shutdown,
3455 .setsockopt = sock_common_setsockopt,
3456 .getsockopt = sock_common_getsockopt,
3457 .sendmsg = inet6_sendmsg,
3458 .recvmsg = inet6_recvmsg,
3459 .mmap = sock_no_mmap,
3460 .sendpage = inet_sendpage,
3461 #ifdef CONFIG_COMPAT
3462 .compat_ioctl = inet6_compat_ioctl,
3466 static struct proto mptcp_v6_prot;
3468 static void mptcp_v6_destroy(struct sock *sk)
3471 inet6_destroy_sock(sk);
3474 static struct inet_protosw mptcp_v6_protosw = {
3475 .type = SOCK_STREAM,
3476 .protocol = IPPROTO_MPTCP,
3477 .prot = &mptcp_v6_prot,
3478 .ops = &mptcp_v6_stream_ops,
3479 .flags = INET_PROTOSW_ICSK,
3482 int __init mptcp_proto_v6_init(void)
3486 mptcp_v6_prot = mptcp_prot;
3487 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3488 mptcp_v6_prot.slab = NULL;
3489 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3490 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3492 err = proto_register(&mptcp_v6_prot, 1);
3496 err = inet6_register_protosw(&mptcp_v6_protosw);
3498 proto_unregister(&mptcp_v6_prot);