1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
28 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
30 struct mptcp_sock msk;
41 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
43 static struct percpu_counter mptcp_sockets_allocated;
45 static void __mptcp_destroy_sock(struct sock *sk);
46 static void __mptcp_check_send_data_fin(struct sock *sk);
48 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
49 static struct net_device mptcp_napi_dev;
51 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
52 * completed yet or has failed, return the subflow socket.
53 * Otherwise return NULL.
55 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
57 if (!msk->subflow || READ_ONCE(msk->can_ack))
63 /* Returns end sequence number of the receiver's advertised window */
64 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
66 return READ_ONCE(msk->wnd_end);
69 static bool mptcp_is_tcpsk(struct sock *sk)
71 struct socket *sock = sk->sk_socket;
73 if (unlikely(sk->sk_prot == &tcp_prot)) {
74 /* we are being invoked after mptcp_accept() has
75 * accepted a non-mp-capable flow: sk is a tcp_sk,
78 * Hand the socket over to tcp so all further socket ops
81 sock->ops = &inet_stream_ops;
83 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
84 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
85 sock->ops = &inet6_stream_ops;
93 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
95 sock_owned_by_me((const struct sock *)msk);
97 if (likely(!__mptcp_check_fallback(msk)))
103 static int __mptcp_socket_create(struct mptcp_sock *msk)
105 struct mptcp_subflow_context *subflow;
106 struct sock *sk = (struct sock *)msk;
107 struct socket *ssock;
110 err = mptcp_subflow_create_socket(sk, &ssock);
114 msk->first = ssock->sk;
115 msk->subflow = ssock;
116 subflow = mptcp_subflow_ctx(ssock->sk);
117 list_add(&subflow->node, &msk->conn_list);
118 sock_hold(ssock->sk);
119 subflow->request_mptcp = 1;
120 mptcp_sock_graft(msk->first, sk->sk_socket);
125 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
127 sk_drops_add(sk, skb);
131 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
132 struct sk_buff *from)
137 if (MPTCP_SKB_CB(from)->offset ||
138 !skb_try_coalesce(to, from, &fragstolen, &delta))
141 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
142 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
143 to->len, MPTCP_SKB_CB(from)->end_seq);
144 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
145 kfree_skb_partial(from, fragstolen);
146 atomic_add(delta, &sk->sk_rmem_alloc);
147 sk_mem_charge(sk, delta);
151 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
152 struct sk_buff *from)
154 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
157 return mptcp_try_coalesce((struct sock *)msk, to, from);
160 /* "inspired" by tcp_data_queue_ofo(), main differences:
162 * - don't cope with sacks
164 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
166 struct sock *sk = (struct sock *)msk;
167 struct rb_node **p, *parent;
168 u64 seq, end_seq, max_seq;
169 struct sk_buff *skb1;
171 seq = MPTCP_SKB_CB(skb)->map_seq;
172 end_seq = MPTCP_SKB_CB(skb)->end_seq;
173 max_seq = READ_ONCE(msk->rcv_wnd_sent);
175 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
176 RB_EMPTY_ROOT(&msk->out_of_order_queue));
177 if (after64(end_seq, max_seq)) {
180 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
181 (unsigned long long)end_seq - (unsigned long)max_seq,
182 (unsigned long long)msk->rcv_wnd_sent);
183 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
187 p = &msk->out_of_order_queue.rb_node;
188 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
189 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
190 rb_link_node(&skb->rbnode, NULL, p);
191 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
192 msk->ooo_last_skb = skb;
196 /* with 2 subflows, adding at end of ooo queue is quite likely
197 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
199 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
200 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
201 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
205 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
206 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
207 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
208 parent = &msk->ooo_last_skb->rbnode;
209 p = &parent->rb_right;
213 /* Find place to insert this segment. Handle overlaps on the way. */
217 skb1 = rb_to_skb(parent);
218 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
219 p = &parent->rb_left;
222 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
223 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
224 /* All the bits are present. Drop. */
226 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
229 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
233 * continue traversing
236 /* skb's seq == skb1's seq and skb covers skb1.
237 * Replace skb1 with skb.
239 rb_replace_node(&skb1->rbnode, &skb->rbnode,
240 &msk->out_of_order_queue);
241 mptcp_drop(sk, skb1);
242 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
245 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
249 p = &parent->rb_right;
253 /* Insert segment into RB tree. */
254 rb_link_node(&skb->rbnode, parent, p);
255 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
258 /* Remove other segments covered by skb. */
259 while ((skb1 = skb_rb_next(skb)) != NULL) {
260 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
262 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
263 mptcp_drop(sk, skb1);
264 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
266 /* If there is no skb after us, we are the last_skb ! */
268 msk->ooo_last_skb = skb;
272 skb_set_owner_r(skb, sk);
275 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
276 struct sk_buff *skb, unsigned int offset,
279 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
280 struct sock *sk = (struct sock *)msk;
281 struct sk_buff *tail;
283 __skb_unlink(skb, &ssk->sk_receive_queue);
288 /* try to fetch required memory from subflow */
289 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
290 if (ssk->sk_forward_alloc < skb->truesize)
292 __sk_mem_reclaim(ssk, skb->truesize);
293 if (!sk_rmem_schedule(sk, skb, skb->truesize))
297 /* the skb map_seq accounts for the skb offset:
298 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
301 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
302 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
303 MPTCP_SKB_CB(skb)->offset = offset;
305 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
307 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
308 tail = skb_peek_tail(&sk->sk_receive_queue);
309 if (tail && mptcp_try_coalesce(sk, tail, skb))
312 skb_set_owner_r(skb, sk);
313 __skb_queue_tail(&sk->sk_receive_queue, skb);
315 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
316 mptcp_data_queue_ofo(msk, skb);
320 /* old data, keep it simple and drop the whole pkt, sender
321 * will retransmit as needed, if needed.
323 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
329 static void mptcp_stop_timer(struct sock *sk)
331 struct inet_connection_sock *icsk = inet_csk(sk);
333 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
334 mptcp_sk(sk)->timer_ival = 0;
337 static void mptcp_close_wake_up(struct sock *sk)
339 if (sock_flag(sk, SOCK_DEAD))
342 sk->sk_state_change(sk);
343 if (sk->sk_shutdown == SHUTDOWN_MASK ||
344 sk->sk_state == TCP_CLOSE)
345 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
347 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
350 static bool mptcp_pending_data_fin_ack(struct sock *sk)
352 struct mptcp_sock *msk = mptcp_sk(sk);
354 return !__mptcp_check_fallback(msk) &&
355 ((1 << sk->sk_state) &
356 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
357 msk->write_seq == READ_ONCE(msk->snd_una);
360 static void mptcp_check_data_fin_ack(struct sock *sk)
362 struct mptcp_sock *msk = mptcp_sk(sk);
364 /* Look for an acknowledged DATA_FIN */
365 if (mptcp_pending_data_fin_ack(sk)) {
366 WRITE_ONCE(msk->snd_data_fin_enable, 0);
368 switch (sk->sk_state) {
370 inet_sk_state_store(sk, TCP_FIN_WAIT2);
374 inet_sk_state_store(sk, TCP_CLOSE);
378 mptcp_close_wake_up(sk);
382 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
384 struct mptcp_sock *msk = mptcp_sk(sk);
386 if (READ_ONCE(msk->rcv_data_fin) &&
387 ((1 << sk->sk_state) &
388 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
389 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
391 if (msk->ack_seq == rcv_data_fin_seq) {
393 *seq = rcv_data_fin_seq;
402 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
404 long tout = ssk && inet_csk(ssk)->icsk_pending ?
405 inet_csk(ssk)->icsk_timeout - jiffies : 0;
408 tout = mptcp_sk(sk)->timer_ival;
409 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
412 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
414 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
416 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
417 if (subflow->request_join && !subflow->fully_established)
420 /* only send if our side has not closed yet */
421 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
424 static bool tcp_can_send_ack(const struct sock *ssk)
426 return !((1 << inet_sk_state_load(ssk)) &
427 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
430 static void mptcp_send_ack(struct mptcp_sock *msk)
432 struct mptcp_subflow_context *subflow;
434 mptcp_for_each_subflow(msk, subflow) {
435 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
438 if (tcp_can_send_ack(ssk))
444 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
449 ret = tcp_can_send_ack(ssk);
451 tcp_cleanup_rbuf(ssk, 1);
456 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
458 struct sock *ack_hint = READ_ONCE(msk->ack_hint);
459 int old_space = READ_ONCE(msk->old_wspace);
460 struct mptcp_subflow_context *subflow;
461 struct sock *sk = (struct sock *)msk;
464 /* this is a simple superset of what tcp_cleanup_rbuf() implements
465 * so that we don't have to acquire the ssk socket lock most of the time
466 * to do actually nothing
468 cleanup = __mptcp_space(sk) - old_space >= max(0, old_space);
472 /* if the hinted ssk is still active, try to use it */
473 if (likely(ack_hint)) {
474 mptcp_for_each_subflow(msk, subflow) {
475 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
477 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk))
482 /* otherwise pick the first active subflow */
483 mptcp_for_each_subflow(msk, subflow)
484 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
488 static bool mptcp_check_data_fin(struct sock *sk)
490 struct mptcp_sock *msk = mptcp_sk(sk);
491 u64 rcv_data_fin_seq;
494 if (__mptcp_check_fallback(msk) || !msk->first)
497 /* Need to ack a DATA_FIN received from a peer while this side
498 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
499 * msk->rcv_data_fin was set when parsing the incoming options
500 * at the subflow level and the msk lock was not held, so this
501 * is the first opportunity to act on the DATA_FIN and change
504 * If we are caught up to the sequence number of the incoming
505 * DATA_FIN, send the DATA_ACK now and do state transition. If
506 * not caught up, do nothing and let the recv code send DATA_ACK
510 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
511 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
512 WRITE_ONCE(msk->rcv_data_fin, 0);
514 sk->sk_shutdown |= RCV_SHUTDOWN;
515 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
516 set_bit(MPTCP_DATA_READY, &msk->flags);
518 switch (sk->sk_state) {
519 case TCP_ESTABLISHED:
520 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
523 inet_sk_state_store(sk, TCP_CLOSING);
526 inet_sk_state_store(sk, TCP_CLOSE);
529 /* Other states not expected */
535 mptcp_set_timeout(sk, NULL);
537 mptcp_close_wake_up(sk);
542 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
546 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
547 struct sock *sk = (struct sock *)msk;
548 unsigned int moved = 0;
549 bool more_data_avail;
554 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
556 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
557 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
559 if (unlikely(ssk_rbuf > sk_rbuf)) {
560 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
565 pr_debug("msk=%p ssk=%p", msk, ssk);
568 u32 map_remaining, offset;
569 u32 seq = tp->copied_seq;
573 /* try to move as much data as available */
574 map_remaining = subflow->map_data_len -
575 mptcp_subflow_get_map_offset(subflow);
577 skb = skb_peek(&ssk->sk_receive_queue);
579 /* if no data is found, a racing workqueue/recvmsg
580 * already processed the new data, stop here or we
581 * can enter an infinite loop
588 if (__mptcp_check_fallback(msk)) {
589 /* if we are running under the workqueue, TCP could have
590 * collapsed skbs between dummy map creation and now
591 * be sure to adjust the size
593 map_remaining = skb->len;
594 subflow->map_data_len = skb->len;
597 offset = seq - TCP_SKB_CB(skb)->seq;
598 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
604 if (offset < skb->len) {
605 size_t len = skb->len - offset;
610 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
614 if (WARN_ON_ONCE(map_remaining < len))
618 sk_eat_skb(ssk, skb);
622 WRITE_ONCE(tp->copied_seq, seq);
623 more_data_avail = mptcp_subflow_data_available(ssk);
625 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
629 } while (more_data_avail);
630 WRITE_ONCE(msk->ack_hint, ssk);
636 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
638 struct sock *sk = (struct sock *)msk;
639 struct sk_buff *skb, *tail;
644 p = rb_first(&msk->out_of_order_queue);
645 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
648 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
652 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
654 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
657 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
661 end_seq = MPTCP_SKB_CB(skb)->end_seq;
662 tail = skb_peek_tail(&sk->sk_receive_queue);
663 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
664 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
666 /* skip overlapping data, if any */
667 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
668 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
670 MPTCP_SKB_CB(skb)->offset += delta;
671 __skb_queue_tail(&sk->sk_receive_queue, skb);
673 msk->ack_seq = end_seq;
679 /* In most cases we will be able to lock the mptcp socket. If its already
680 * owned, we need to defer to the work queue to avoid ABBA deadlock.
682 static void move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
684 struct sock *sk = (struct sock *)msk;
685 unsigned int moved = 0;
687 if (inet_sk_state_load(sk) == TCP_CLOSE)
692 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
693 __mptcp_ofo_queue(msk);
695 /* If the moves have caught up with the DATA_FIN sequence number
696 * it's time to ack the DATA_FIN and change socket state, but
697 * this is not a good place to change state. Let the workqueue
700 if (mptcp_pending_data_fin(sk, NULL))
701 mptcp_schedule_work(sk);
702 mptcp_data_unlock(sk);
705 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
707 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
708 struct mptcp_sock *msk = mptcp_sk(sk);
709 int sk_rbuf, ssk_rbuf;
712 /* The peer can send data while we are shutting down this
713 * subflow at msk destruction time, but we must avoid enqueuing
714 * more data to the msk receive queue
716 if (unlikely(subflow->disposable))
719 /* move_skbs_to_msk below can legitly clear the data_avail flag,
720 * but we will need later to properly woke the reader, cache its
723 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
725 set_bit(MPTCP_DATA_READY, &msk->flags);
727 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
728 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
729 if (unlikely(ssk_rbuf > sk_rbuf))
732 /* over limit? can't append more skbs to msk */
733 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
736 move_skbs_to_msk(msk, ssk);
740 sk->sk_data_ready(sk);
743 void __mptcp_flush_join_list(struct mptcp_sock *msk)
745 struct mptcp_subflow_context *subflow;
747 if (likely(list_empty(&msk->join_list)))
750 spin_lock_bh(&msk->join_list_lock);
751 list_for_each_entry(subflow, &msk->join_list, node)
752 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
753 list_splice_tail_init(&msk->join_list, &msk->conn_list);
754 spin_unlock_bh(&msk->join_list_lock);
757 static bool mptcp_timer_pending(struct sock *sk)
759 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
762 static void mptcp_reset_timer(struct sock *sk)
764 struct inet_connection_sock *icsk = inet_csk(sk);
767 /* prevent rescheduling on close */
768 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
771 /* should never be called with mptcp level timer cleared */
772 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
773 if (WARN_ON_ONCE(!tout))
775 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
778 bool mptcp_schedule_work(struct sock *sk)
780 if (inet_sk_state_load(sk) != TCP_CLOSE &&
781 schedule_work(&mptcp_sk(sk)->work)) {
782 /* each subflow already holds a reference to the sk, and the
783 * workqueue is invoked by a subflow, so sk can't go away here.
791 void mptcp_subflow_eof(struct sock *sk)
793 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
794 mptcp_schedule_work(sk);
797 static void mptcp_check_for_eof(struct mptcp_sock *msk)
799 struct mptcp_subflow_context *subflow;
800 struct sock *sk = (struct sock *)msk;
803 mptcp_for_each_subflow(msk, subflow)
804 receivers += !subflow->rx_eof;
808 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
809 /* hopefully temporary hack: propagate shutdown status
810 * to msk, when all subflows agree on it
812 sk->sk_shutdown |= RCV_SHUTDOWN;
814 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
815 set_bit(MPTCP_DATA_READY, &msk->flags);
816 sk->sk_data_ready(sk);
819 switch (sk->sk_state) {
820 case TCP_ESTABLISHED:
821 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
824 inet_sk_state_store(sk, TCP_CLOSING);
827 inet_sk_state_store(sk, TCP_CLOSE);
832 mptcp_close_wake_up(sk);
835 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
837 struct mptcp_subflow_context *subflow;
838 struct sock *sk = (struct sock *)msk;
840 sock_owned_by_me(sk);
842 mptcp_for_each_subflow(msk, subflow) {
843 if (subflow->data_avail)
844 return mptcp_subflow_tcp_sock(subflow);
850 static bool mptcp_skb_can_collapse_to(u64 write_seq,
851 const struct sk_buff *skb,
852 const struct mptcp_ext *mpext)
854 if (!tcp_skb_can_collapse_to(skb))
857 /* can collapse only if MPTCP level sequence is in order and this
858 * mapping has not been xmitted yet
860 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
864 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
865 const struct page_frag *pfrag,
866 const struct mptcp_data_frag *df)
868 return df && pfrag->page == df->page &&
869 pfrag->size - pfrag->offset > 0 &&
870 df->data_seq + df->data_len == msk->write_seq;
873 static int mptcp_wmem_with_overhead(struct sock *sk, int size)
875 struct mptcp_sock *msk = mptcp_sk(sk);
878 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
879 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache;
880 if (skbs < msk->skb_tx_cache.qlen)
883 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER);
886 static void __mptcp_wmem_reserve(struct sock *sk, int size)
888 int amount = mptcp_wmem_with_overhead(sk, size);
889 struct mptcp_sock *msk = mptcp_sk(sk);
891 WARN_ON_ONCE(msk->wmem_reserved);
892 if (WARN_ON_ONCE(amount < 0))
895 if (amount <= sk->sk_forward_alloc)
898 /* under memory pressure try to reserve at most a single page
899 * otherwise try to reserve the full estimate and fallback
900 * to a single page before entering the error path
902 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
903 !sk_wmem_schedule(sk, amount)) {
904 if (amount <= PAGE_SIZE)
908 if (!sk_wmem_schedule(sk, amount))
913 msk->wmem_reserved = amount;
914 sk->sk_forward_alloc -= amount;
918 /* we will wait for memory on next allocation */
919 msk->wmem_reserved = -1;
922 static void __mptcp_update_wmem(struct sock *sk)
924 struct mptcp_sock *msk = mptcp_sk(sk);
926 if (!msk->wmem_reserved)
929 if (msk->wmem_reserved < 0)
930 msk->wmem_reserved = 0;
931 if (msk->wmem_reserved > 0) {
932 sk->sk_forward_alloc += msk->wmem_reserved;
933 msk->wmem_reserved = 0;
937 static bool mptcp_wmem_alloc(struct sock *sk, int size)
939 struct mptcp_sock *msk = mptcp_sk(sk);
941 /* check for pre-existing error condition */
942 if (msk->wmem_reserved < 0)
945 if (msk->wmem_reserved >= size)
949 if (!sk_wmem_schedule(sk, size)) {
950 mptcp_data_unlock(sk);
954 sk->sk_forward_alloc -= size;
955 msk->wmem_reserved += size;
956 mptcp_data_unlock(sk);
959 msk->wmem_reserved -= size;
963 static void mptcp_wmem_uncharge(struct sock *sk, int size)
965 struct mptcp_sock *msk = mptcp_sk(sk);
967 if (msk->wmem_reserved < 0)
968 msk->wmem_reserved = 0;
969 msk->wmem_reserved += size;
972 static void mptcp_mem_reclaim_partial(struct sock *sk)
974 struct mptcp_sock *msk = mptcp_sk(sk);
976 /* if we are experiencing a transint allocation error,
977 * the forward allocation memory has been already
980 if (msk->wmem_reserved < 0)
984 sk->sk_forward_alloc += msk->wmem_reserved;
985 sk_mem_reclaim_partial(sk);
986 msk->wmem_reserved = sk->sk_forward_alloc;
987 sk->sk_forward_alloc = 0;
988 mptcp_data_unlock(sk);
991 static void dfrag_uncharge(struct sock *sk, int len)
993 sk_mem_uncharge(sk, len);
994 sk_wmem_queued_add(sk, -len);
997 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
999 int len = dfrag->data_len + dfrag->overhead;
1001 list_del(&dfrag->list);
1002 dfrag_uncharge(sk, len);
1003 put_page(dfrag->page);
1006 static void __mptcp_clean_una(struct sock *sk)
1008 struct mptcp_sock *msk = mptcp_sk(sk);
1009 struct mptcp_data_frag *dtmp, *dfrag;
1010 bool cleaned = false;
1013 /* on fallback we just need to ignore snd_una, as this is really
1016 if (__mptcp_check_fallback(msk))
1017 msk->snd_una = READ_ONCE(msk->snd_nxt);
1019 snd_una = msk->snd_una;
1020 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1021 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1024 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1026 dfrag_clear(sk, dfrag);
1030 dfrag = mptcp_rtx_head(sk);
1031 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1032 u64 delta = snd_una - dfrag->data_seq;
1034 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1037 dfrag->data_seq += delta;
1038 dfrag->offset += delta;
1039 dfrag->data_len -= delta;
1040 dfrag->already_sent -= delta;
1042 dfrag_uncharge(sk, delta);
1048 if (tcp_under_memory_pressure(sk)) {
1049 __mptcp_update_wmem(sk);
1050 sk_mem_reclaim_partial(sk);
1054 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1055 if (msk->timer_ival)
1056 mptcp_stop_timer(sk);
1058 mptcp_reset_timer(sk);
1062 static void __mptcp_clean_una_wakeup(struct sock *sk)
1064 __mptcp_clean_una(sk);
1065 mptcp_write_space(sk);
1068 static void mptcp_enter_memory_pressure(struct sock *sk)
1070 struct mptcp_subflow_context *subflow;
1071 struct mptcp_sock *msk = mptcp_sk(sk);
1074 sk_stream_moderate_sndbuf(sk);
1075 mptcp_for_each_subflow(msk, subflow) {
1076 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1079 tcp_enter_memory_pressure(ssk);
1080 sk_stream_moderate_sndbuf(ssk);
1085 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1088 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1090 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1091 pfrag, sk->sk_allocation)))
1094 mptcp_enter_memory_pressure(sk);
1098 static struct mptcp_data_frag *
1099 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1102 int offset = ALIGN(orig_offset, sizeof(long));
1103 struct mptcp_data_frag *dfrag;
1105 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1106 dfrag->data_len = 0;
1107 dfrag->data_seq = msk->write_seq;
1108 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1109 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1110 dfrag->already_sent = 0;
1111 dfrag->page = pfrag->page;
1116 struct mptcp_sendmsg_info {
1124 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1127 u64 window_end = mptcp_wnd_end(msk);
1129 if (__mptcp_check_fallback(msk))
1132 if (!before64(data_seq + avail_size, window_end)) {
1133 u64 allowed_size = window_end - data_seq;
1135 return min_t(unsigned int, allowed_size, avail_size);
1141 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1143 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1147 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1151 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1153 struct sk_buff *skb;
1155 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1157 if (likely(__mptcp_add_ext(skb, gfp))) {
1158 skb_reserve(skb, MAX_TCP_HEADER);
1159 skb->reserved_tailroom = skb->end - skb->tail;
1164 mptcp_enter_memory_pressure(sk);
1169 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1170 struct sk_buff_head *skbs, int *total_ts)
1172 struct mptcp_sock *msk = mptcp_sk(sk);
1173 struct sk_buff *skb;
1176 if (unlikely(tcp_under_memory_pressure(sk))) {
1177 mptcp_mem_reclaim_partial(sk);
1179 /* under pressure pre-allocate at most a single skb */
1180 if (msk->skb_tx_cache.qlen)
1182 space_needed = msk->size_goal_cache;
1184 space_needed = msk->tx_pending_data + size -
1185 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1188 while (space_needed > 0) {
1189 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1190 if (unlikely(!skb)) {
1191 /* under memory pressure, try to pass the caller a
1192 * single skb to allow forward progress
1194 while (skbs->qlen > 1) {
1195 skb = __skb_dequeue_tail(skbs);
1196 *total_ts -= skb->truesize;
1199 return skbs->qlen > 0;
1202 *total_ts += skb->truesize;
1203 __skb_queue_tail(skbs, skb);
1204 space_needed -= msk->size_goal_cache;
1209 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1211 struct mptcp_sock *msk = mptcp_sk(sk);
1212 struct sk_buff *skb;
1214 if (ssk->sk_tx_skb_cache) {
1215 skb = ssk->sk_tx_skb_cache;
1216 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1217 !__mptcp_add_ext(skb, gfp)))
1222 skb = skb_peek(&msk->skb_tx_cache);
1224 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1225 skb = __skb_dequeue(&msk->skb_tx_cache);
1226 if (WARN_ON_ONCE(!skb))
1229 mptcp_wmem_uncharge(sk, skb->truesize);
1230 ssk->sk_tx_skb_cache = skb;
1234 /* over memory limit, no point to try to allocate a new skb */
1238 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1242 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1243 ssk->sk_tx_skb_cache = skb;
1250 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1252 return !ssk->sk_tx_skb_cache &&
1253 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1254 tcp_under_memory_pressure(sk);
1257 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1259 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1260 mptcp_mem_reclaim_partial(sk);
1261 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1264 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1265 struct mptcp_data_frag *dfrag,
1266 struct mptcp_sendmsg_info *info)
1268 u64 data_seq = dfrag->data_seq + info->sent;
1269 struct mptcp_sock *msk = mptcp_sk(sk);
1270 bool zero_window_probe = false;
1271 struct mptcp_ext *mpext = NULL;
1272 struct sk_buff *skb, *tail;
1273 bool can_collapse = false;
1278 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
1279 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1281 /* compute send limit */
1282 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1283 avail_size = info->size_goal;
1284 msk->size_goal_cache = info->size_goal;
1285 skb = tcp_write_queue_tail(ssk);
1287 /* Limit the write to the size available in the
1288 * current skb, if any, so that we create at most a new skb.
1289 * Explicitly tells TCP internals to avoid collapsing on later
1290 * queue management operation, to avoid breaking the ext <->
1291 * SSN association set here
1293 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1294 can_collapse = (info->size_goal - skb->len > 0) &&
1295 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1296 if (!can_collapse) {
1297 TCP_SKB_CB(skb)->eor = 1;
1299 size_bias = skb->len;
1300 avail_size = info->size_goal - skb->len;
1304 /* Zero window and all data acked? Probe. */
1305 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1306 if (avail_size == 0) {
1307 u64 snd_una = READ_ONCE(msk->snd_una);
1309 if (skb || snd_una != msk->snd_nxt)
1311 zero_window_probe = true;
1312 data_seq = snd_una - 1;
1316 if (WARN_ON_ONCE(info->sent > info->limit ||
1317 info->limit > dfrag->data_len))
1320 ret = info->limit - info->sent;
1321 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1322 dfrag->page, dfrag->offset + info->sent, &ret);
1324 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1328 /* if the tail skb is still the cached one, collapsing really happened.
1331 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1332 mpext->data_len += ret;
1333 WARN_ON_ONCE(!can_collapse);
1334 WARN_ON_ONCE(zero_window_probe);
1338 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1339 if (WARN_ON_ONCE(!mpext)) {
1340 /* should never reach here, stream corrupted */
1344 memset(mpext, 0, sizeof(*mpext));
1345 mpext->data_seq = data_seq;
1346 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1347 mpext->data_len = ret;
1351 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1352 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1355 if (zero_window_probe) {
1356 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1359 tcp_push_pending_frames(ssk);
1362 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1366 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1367 sizeof(struct tcphdr) - \
1368 MAX_TCP_OPTION_SPACE - \
1369 sizeof(struct ipv6hdr) - \
1370 sizeof(struct frag_hdr))
1372 struct subflow_send_info {
1377 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1379 struct subflow_send_info send_info[2];
1380 struct mptcp_subflow_context *subflow;
1381 int i, nr_active = 0;
1386 sock_owned_by_me((struct sock *)msk);
1388 if (__mptcp_check_fallback(msk)) {
1391 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1394 /* re-use last subflow, if the burst allow that */
1395 if (msk->last_snd && msk->snd_burst > 0 &&
1396 sk_stream_memory_free(msk->last_snd) &&
1397 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1398 return msk->last_snd;
1400 /* pick the subflow with the lower wmem/wspace ratio */
1401 for (i = 0; i < 2; ++i) {
1402 send_info[i].ssk = NULL;
1403 send_info[i].ratio = -1;
1405 mptcp_for_each_subflow(msk, subflow) {
1406 ssk = mptcp_subflow_tcp_sock(subflow);
1407 if (!mptcp_subflow_active(subflow))
1410 nr_active += !subflow->backup;
1411 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1414 pace = READ_ONCE(ssk->sk_pacing_rate);
1418 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1420 if (ratio < send_info[subflow->backup].ratio) {
1421 send_info[subflow->backup].ssk = ssk;
1422 send_info[subflow->backup].ratio = ratio;
1426 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1427 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1428 send_info[1].ssk, send_info[1].ratio);
1430 /* pick the best backup if no other subflow is active */
1432 send_info[0].ssk = send_info[1].ssk;
1434 if (send_info[0].ssk) {
1435 msk->last_snd = send_info[0].ssk;
1436 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1437 tcp_sk(msk->last_snd)->snd_wnd);
1438 return msk->last_snd;
1444 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1445 struct mptcp_sendmsg_info *info)
1447 mptcp_set_timeout(sk, ssk);
1448 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1452 static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1454 struct sock *prev_ssk = NULL, *ssk = NULL;
1455 struct mptcp_sock *msk = mptcp_sk(sk);
1456 struct mptcp_sendmsg_info info = {
1459 struct mptcp_data_frag *dfrag;
1460 int len, copied = 0;
1462 while ((dfrag = mptcp_send_head(sk))) {
1463 info.sent = dfrag->already_sent;
1464 info.limit = dfrag->data_len;
1465 len = dfrag->data_len - dfrag->already_sent;
1470 __mptcp_flush_join_list(msk);
1471 ssk = mptcp_subflow_get_send(msk);
1473 /* try to keep the subflow socket lock across
1474 * consecutive xmit on the same socket
1476 if (ssk != prev_ssk && prev_ssk)
1477 mptcp_push_release(sk, prev_ssk, &info);
1481 if (ssk != prev_ssk || !prev_ssk)
1484 /* keep it simple and always provide a new skb for the
1485 * subflow, even if we will not use it when collapsing
1486 * on the pending one
1488 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1489 mptcp_push_release(sk, ssk, &info);
1493 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1495 mptcp_push_release(sk, ssk, &info);
1500 dfrag->already_sent += ret;
1501 msk->snd_nxt += ret;
1502 msk->snd_burst -= ret;
1503 msk->tx_pending_data -= ret;
1507 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1510 /* at this point we held the socket lock for the last subflow we used */
1512 mptcp_push_release(sk, ssk, &info);
1516 /* start the timer, if it's not pending */
1517 if (!mptcp_timer_pending(sk))
1518 mptcp_reset_timer(sk);
1519 __mptcp_check_send_data_fin(sk);
1523 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1525 struct mptcp_sock *msk = mptcp_sk(sk);
1526 struct mptcp_sendmsg_info info;
1527 struct mptcp_data_frag *dfrag;
1528 struct sock *xmit_ssk;
1529 int len, copied = 0;
1533 while ((dfrag = mptcp_send_head(sk))) {
1534 info.sent = dfrag->already_sent;
1535 info.limit = dfrag->data_len;
1536 len = dfrag->data_len - dfrag->already_sent;
1540 /* the caller already invoked the packet scheduler,
1541 * check for a different subflow usage only after
1542 * spooling the first chunk of data
1544 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1547 if (xmit_ssk != ssk) {
1548 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1552 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1553 __mptcp_update_wmem(sk);
1554 sk_mem_reclaim_partial(sk);
1556 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1559 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1564 dfrag->already_sent += ret;
1565 msk->snd_nxt += ret;
1566 msk->snd_burst -= ret;
1567 msk->tx_pending_data -= ret;
1572 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1576 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1577 * not going to flush it via release_sock()
1579 __mptcp_update_wmem(sk);
1581 mptcp_set_timeout(sk, ssk);
1582 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1584 if (!mptcp_timer_pending(sk))
1585 mptcp_reset_timer(sk);
1587 if (msk->snd_data_fin_enable &&
1588 msk->snd_nxt + 1 == msk->write_seq)
1589 mptcp_schedule_work(sk);
1593 static void mptcp_set_nospace(struct sock *sk)
1595 /* enable autotune */
1596 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1598 /* will be cleared on avail space */
1599 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1602 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1604 struct mptcp_sock *msk = mptcp_sk(sk);
1605 struct page_frag *pfrag;
1610 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1613 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1615 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1617 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1618 ret = sk_stream_wait_connect(sk, &timeo);
1623 pfrag = sk_page_frag(sk);
1625 while (msg_data_left(msg)) {
1626 int total_ts, frag_truesize = 0;
1627 struct mptcp_data_frag *dfrag;
1628 struct sk_buff_head skbs;
1629 bool dfrag_collapsed;
1630 size_t psize, offset;
1632 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1637 /* reuse tail pfrag, if possible, or carve a new one from the
1640 dfrag = mptcp_pending_tail(sk);
1641 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1642 if (!dfrag_collapsed) {
1643 if (!sk_stream_memory_free(sk))
1644 goto wait_for_memory;
1646 if (!mptcp_page_frag_refill(sk, pfrag))
1647 goto wait_for_memory;
1649 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1650 frag_truesize = dfrag->overhead;
1653 /* we do not bound vs wspace, to allow a single packet.
1654 * memory accounting will prevent execessive memory usage
1657 offset = dfrag->offset + dfrag->data_len;
1658 psize = pfrag->size - offset;
1659 psize = min_t(size_t, psize, msg_data_left(msg));
1660 total_ts = psize + frag_truesize;
1661 __skb_queue_head_init(&skbs);
1662 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1663 goto wait_for_memory;
1665 if (!mptcp_wmem_alloc(sk, total_ts)) {
1666 __skb_queue_purge(&skbs);
1667 goto wait_for_memory;
1670 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1671 if (copy_page_from_iter(dfrag->page, offset, psize,
1672 &msg->msg_iter) != psize) {
1673 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1678 /* data successfully copied into the write queue */
1680 dfrag->data_len += psize;
1681 frag_truesize += psize;
1682 pfrag->offset += frag_truesize;
1683 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1684 msk->tx_pending_data += psize;
1686 /* charge data on mptcp pending queue to the msk socket
1687 * Note: we charge such data both to sk and ssk
1689 sk_wmem_queued_add(sk, frag_truesize);
1690 if (!dfrag_collapsed) {
1691 get_page(dfrag->page);
1692 list_add_tail(&dfrag->list, &msk->rtx_queue);
1693 if (!msk->first_pending)
1694 WRITE_ONCE(msk->first_pending, dfrag);
1696 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1697 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1703 mptcp_set_nospace(sk);
1704 __mptcp_push_pending(sk, msg->msg_flags);
1705 ret = sk_stream_wait_memory(sk, &timeo);
1711 __mptcp_push_pending(sk, msg->msg_flags);
1715 return copied ? : ret;
1718 static void mptcp_wait_data(struct sock *sk, long *timeo)
1720 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1721 struct mptcp_sock *msk = mptcp_sk(sk);
1723 add_wait_queue(sk_sleep(sk), &wait);
1724 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1726 sk_wait_event(sk, timeo,
1727 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1729 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1730 remove_wait_queue(sk_sleep(sk), &wait);
1733 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1737 struct sk_buff *skb;
1740 while ((skb = skb_peek(&msk->receive_queue)) != NULL) {
1741 u32 offset = MPTCP_SKB_CB(skb)->offset;
1742 u32 data_len = skb->len - offset;
1743 u32 count = min_t(size_t, len - copied, data_len);
1746 err = skb_copy_datagram_msg(skb, offset, msg, count);
1747 if (unlikely(err < 0)) {
1755 if (count < data_len) {
1756 MPTCP_SKB_CB(skb)->offset += count;
1760 /* we will bulk release the skb memory later */
1761 skb->destructor = NULL;
1762 msk->rmem_released += skb->truesize;
1763 __skb_unlink(skb, &msk->receive_queue);
1773 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1775 * Only difference: Use highest rtt estimate of the subflows in use.
1777 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1779 struct mptcp_subflow_context *subflow;
1780 struct sock *sk = (struct sock *)msk;
1781 u32 time, advmss = 1;
1784 sock_owned_by_me(sk);
1789 msk->rcvq_space.copied += copied;
1791 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1792 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1794 rtt_us = msk->rcvq_space.rtt_us;
1795 if (rtt_us && time < (rtt_us >> 3))
1799 mptcp_for_each_subflow(msk, subflow) {
1800 const struct tcp_sock *tp;
1804 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1806 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1807 sf_advmss = READ_ONCE(tp->advmss);
1809 rtt_us = max(sf_rtt_us, rtt_us);
1810 advmss = max(sf_advmss, advmss);
1813 msk->rcvq_space.rtt_us = rtt_us;
1814 if (time < (rtt_us >> 3) || rtt_us == 0)
1817 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1820 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1821 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1825 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1827 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1829 do_div(grow, msk->rcvq_space.space);
1830 rcvwin += (grow << 1);
1832 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1833 while (tcp_win_from_space(sk, rcvmem) < advmss)
1836 do_div(rcvwin, advmss);
1837 rcvbuf = min_t(u64, rcvwin * rcvmem,
1838 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1840 if (rcvbuf > sk->sk_rcvbuf) {
1843 window_clamp = tcp_win_from_space(sk, rcvbuf);
1844 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1846 /* Make subflows follow along. If we do not do this, we
1847 * get drops at subflow level if skbs can't be moved to
1848 * the mptcp rx queue fast enough (announced rcv_win can
1849 * exceed ssk->sk_rcvbuf).
1851 mptcp_for_each_subflow(msk, subflow) {
1855 ssk = mptcp_subflow_tcp_sock(subflow);
1856 slow = lock_sock_fast(ssk);
1857 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1858 tcp_sk(ssk)->window_clamp = window_clamp;
1859 tcp_cleanup_rbuf(ssk, 1);
1860 unlock_sock_fast(ssk, slow);
1865 msk->rcvq_space.space = msk->rcvq_space.copied;
1867 msk->rcvq_space.copied = 0;
1868 msk->rcvq_space.time = mstamp;
1871 static void __mptcp_update_rmem(struct sock *sk)
1873 struct mptcp_sock *msk = mptcp_sk(sk);
1875 if (!msk->rmem_released)
1878 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1879 sk_mem_uncharge(sk, msk->rmem_released);
1880 msk->rmem_released = 0;
1883 static void __mptcp_splice_receive_queue(struct sock *sk)
1885 struct mptcp_sock *msk = mptcp_sk(sk);
1887 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1890 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1892 struct sock *sk = (struct sock *)msk;
1893 unsigned int moved = 0;
1896 __mptcp_flush_join_list(msk);
1898 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1901 /* we can have data pending in the subflows only if the msk
1902 * receive buffer was full at subflow_data_ready() time,
1903 * that is an unlikely slow path.
1908 slowpath = lock_sock_fast(ssk);
1909 mptcp_data_lock(sk);
1910 __mptcp_update_rmem(sk);
1911 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1912 mptcp_data_unlock(sk);
1913 tcp_cleanup_rbuf(ssk, moved);
1914 unlock_sock_fast(ssk, slowpath);
1917 /* acquire the data lock only if some input data is pending */
1919 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1920 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1921 mptcp_data_lock(sk);
1922 __mptcp_update_rmem(sk);
1923 ret |= __mptcp_ofo_queue(msk);
1924 __mptcp_splice_receive_queue(sk);
1925 mptcp_data_unlock(sk);
1926 mptcp_cleanup_rbuf(msk);
1929 mptcp_check_data_fin((struct sock *)msk);
1930 return !skb_queue_empty(&msk->receive_queue);
1933 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1934 int nonblock, int flags, int *addr_len)
1936 struct mptcp_sock *msk = mptcp_sk(sk);
1941 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1944 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1945 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1950 timeo = sock_rcvtimeo(sk, nonblock);
1952 len = min_t(size_t, len, INT_MAX);
1953 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1955 while (copied < len) {
1958 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1959 if (unlikely(bytes_read < 0)) {
1961 copied = bytes_read;
1965 copied += bytes_read;
1967 /* be sure to advertise window change */
1968 mptcp_cleanup_rbuf(msk);
1970 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
1973 /* only the master socket status is relevant here. The exit
1974 * conditions mirror closely tcp_recvmsg()
1976 if (copied >= target)
1981 sk->sk_state == TCP_CLOSE ||
1982 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1984 signal_pending(current))
1988 copied = sock_error(sk);
1992 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1993 mptcp_check_for_eof(msk);
1995 if (sk->sk_shutdown & RCV_SHUTDOWN) {
1996 /* race breaker: the shutdown could be after the
1997 * previous receive queue check
1999 if (__mptcp_move_skbs(msk))
2004 if (sk->sk_state == TCP_CLOSE) {
2014 if (signal_pending(current)) {
2015 copied = sock_intr_errno(timeo);
2020 pr_debug("block timeout %ld", timeo);
2021 mptcp_wait_data(sk, &timeo);
2024 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2025 skb_queue_empty(&msk->receive_queue)) {
2026 /* entire backlog drained, clear DATA_READY. */
2027 clear_bit(MPTCP_DATA_READY, &msk->flags);
2029 /* .. race-breaker: ssk might have gotten new data
2030 * after last __mptcp_move_skbs() returned false.
2032 if (unlikely(__mptcp_move_skbs(msk)))
2033 set_bit(MPTCP_DATA_READY, &msk->flags);
2034 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2035 /* data to read but mptcp_wait_data() cleared DATA_READY */
2036 set_bit(MPTCP_DATA_READY, &msk->flags);
2039 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2040 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2041 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2042 mptcp_rcv_space_adjust(msk, copied);
2048 static void mptcp_retransmit_handler(struct sock *sk)
2050 struct mptcp_sock *msk = mptcp_sk(sk);
2052 set_bit(MPTCP_WORK_RTX, &msk->flags);
2053 mptcp_schedule_work(sk);
2056 static void mptcp_retransmit_timer(struct timer_list *t)
2058 struct inet_connection_sock *icsk = from_timer(icsk, t,
2059 icsk_retransmit_timer);
2060 struct sock *sk = &icsk->icsk_inet.sk;
2063 if (!sock_owned_by_user(sk)) {
2064 mptcp_retransmit_handler(sk);
2066 /* delegate our work to tcp_release_cb() */
2067 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
2075 static void mptcp_timeout_timer(struct timer_list *t)
2077 struct sock *sk = from_timer(sk, t, sk_timer);
2079 mptcp_schedule_work(sk);
2083 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2086 * A backup subflow is returned only if that is the only kind available.
2088 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2090 struct mptcp_subflow_context *subflow;
2091 struct sock *backup = NULL;
2093 sock_owned_by_me((const struct sock *)msk);
2095 if (__mptcp_check_fallback(msk))
2098 mptcp_for_each_subflow(msk, subflow) {
2099 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2101 if (!mptcp_subflow_active(subflow))
2104 /* still data outstanding at TCP level? Don't retransmit. */
2105 if (!tcp_write_queue_empty(ssk)) {
2106 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2111 if (subflow->backup) {
2123 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2126 iput(SOCK_INODE(msk->subflow));
2127 msk->subflow = NULL;
2131 /* subflow sockets can be either outgoing (connect) or incoming
2134 * Outgoing subflows use in-kernel sockets.
2135 * Incoming subflows do not have their own 'struct socket' allocated,
2136 * so we need to use tcp_close() after detaching them from the mptcp
2139 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2140 struct mptcp_subflow_context *subflow)
2142 struct mptcp_sock *msk = mptcp_sk(sk);
2144 list_del(&subflow->node);
2146 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2148 /* if we are invoked by the msk cleanup code, the subflow is
2154 subflow->disposable = 1;
2156 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2157 * the ssk has been already destroyed, we just need to release the
2158 * reference owned by msk;
2160 if (!inet_csk(ssk)->icsk_ulp_ops) {
2161 kfree_rcu(subflow, rcu);
2163 /* otherwise tcp will dispose of the ssk and subflow ctx */
2164 __tcp_close(ssk, 0);
2166 /* close acquired an extra ref */
2173 if (ssk == msk->last_snd)
2174 msk->last_snd = NULL;
2176 if (ssk == msk->ack_hint)
2177 msk->ack_hint = NULL;
2179 if (ssk == msk->first)
2182 if (msk->subflow && ssk == msk->subflow->sk)
2183 mptcp_dispose_initial_subflow(msk);
2186 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2187 struct mptcp_subflow_context *subflow)
2189 if (sk->sk_state == TCP_ESTABLISHED)
2190 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2191 __mptcp_close_ssk(sk, ssk, subflow);
2194 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2199 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2201 struct mptcp_subflow_context *subflow, *tmp;
2205 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2206 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2208 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2211 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2212 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2215 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2219 static bool mptcp_check_close_timeout(const struct sock *sk)
2221 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2222 struct mptcp_subflow_context *subflow;
2224 if (delta >= TCP_TIMEWAIT_LEN)
2227 /* if all subflows are in closed status don't bother with additional
2230 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2231 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2238 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2240 struct mptcp_subflow_context *subflow, *tmp;
2241 struct sock *sk = &msk->sk.icsk_inet.sk;
2243 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2246 mptcp_token_destroy(msk);
2248 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2249 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2252 if (tcp_sk->sk_state != TCP_CLOSE) {
2253 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2254 tcp_set_state(tcp_sk, TCP_CLOSE);
2256 release_sock(tcp_sk);
2259 inet_sk_state_store(sk, TCP_CLOSE);
2260 sk->sk_shutdown = SHUTDOWN_MASK;
2261 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2262 set_bit(MPTCP_DATA_READY, &msk->flags);
2263 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2265 mptcp_close_wake_up(sk);
2268 static void __mptcp_retrans(struct sock *sk)
2270 struct mptcp_sock *msk = mptcp_sk(sk);
2271 struct mptcp_sendmsg_info info = {};
2272 struct mptcp_data_frag *dfrag;
2277 __mptcp_clean_una_wakeup(sk);
2278 dfrag = mptcp_rtx_head(sk);
2282 ssk = mptcp_subflow_get_retrans(msk);
2288 /* limit retransmission to the bytes already sent on some subflows */
2290 info.limit = dfrag->already_sent;
2291 while (info.sent < dfrag->already_sent) {
2292 if (!mptcp_alloc_tx_skb(sk, ssk))
2295 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2299 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2304 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2307 mptcp_set_timeout(sk, ssk);
2311 if (!mptcp_timer_pending(sk))
2312 mptcp_reset_timer(sk);
2315 static void mptcp_worker(struct work_struct *work)
2317 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2318 struct sock *sk = &msk->sk.icsk_inet.sk;
2322 state = sk->sk_state;
2323 if (unlikely(state == TCP_CLOSE))
2326 mptcp_check_data_fin_ack(sk);
2327 __mptcp_flush_join_list(msk);
2329 mptcp_check_fastclose(msk);
2332 mptcp_pm_nl_work(msk);
2334 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2335 mptcp_check_for_eof(msk);
2337 __mptcp_check_send_data_fin(sk);
2338 mptcp_check_data_fin(sk);
2340 /* There is no point in keeping around an orphaned sk timedout or
2341 * closed, but we need the msk around to reply to incoming DATA_FIN,
2342 * even if it is orphaned and in FIN_WAIT2 state
2344 if (sock_flag(sk, SOCK_DEAD) &&
2345 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2346 inet_sk_state_store(sk, TCP_CLOSE);
2347 __mptcp_destroy_sock(sk);
2351 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2352 __mptcp_close_subflow(msk);
2354 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2355 __mptcp_retrans(sk);
2362 static int __mptcp_init_sock(struct sock *sk)
2364 struct mptcp_sock *msk = mptcp_sk(sk);
2366 spin_lock_init(&msk->join_list_lock);
2368 INIT_LIST_HEAD(&msk->conn_list);
2369 INIT_LIST_HEAD(&msk->join_list);
2370 INIT_LIST_HEAD(&msk->rtx_queue);
2371 INIT_WORK(&msk->work, mptcp_worker);
2372 __skb_queue_head_init(&msk->receive_queue);
2373 __skb_queue_head_init(&msk->skb_tx_cache);
2374 msk->out_of_order_queue = RB_ROOT;
2375 msk->first_pending = NULL;
2376 msk->wmem_reserved = 0;
2377 msk->rmem_released = 0;
2378 msk->tx_pending_data = 0;
2379 msk->size_goal_cache = TCP_BASE_MSS;
2381 msk->ack_hint = NULL;
2383 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2385 mptcp_pm_data_init(msk);
2387 /* re-use the csk retrans timer for MPTCP-level retrans */
2388 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2389 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2393 static int mptcp_init_sock(struct sock *sk)
2395 struct net *net = sock_net(sk);
2398 ret = __mptcp_init_sock(sk);
2402 if (!mptcp_is_enabled(net))
2403 return -ENOPROTOOPT;
2405 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2408 ret = __mptcp_socket_create(mptcp_sk(sk));
2412 sk_sockets_allocated_inc(sk);
2413 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2414 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2419 static void __mptcp_clear_xmit(struct sock *sk)
2421 struct mptcp_sock *msk = mptcp_sk(sk);
2422 struct mptcp_data_frag *dtmp, *dfrag;
2423 struct sk_buff *skb;
2425 WRITE_ONCE(msk->first_pending, NULL);
2426 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2427 dfrag_clear(sk, dfrag);
2428 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2429 sk->sk_forward_alloc += skb->truesize;
2434 static void mptcp_cancel_work(struct sock *sk)
2436 struct mptcp_sock *msk = mptcp_sk(sk);
2438 if (cancel_work_sync(&msk->work))
2442 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2446 switch (ssk->sk_state) {
2448 if (!(how & RCV_SHUTDOWN))
2452 tcp_disconnect(ssk, O_NONBLOCK);
2455 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2456 pr_debug("Fallback");
2457 ssk->sk_shutdown |= how;
2458 tcp_shutdown(ssk, how);
2460 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2461 mptcp_set_timeout(sk, ssk);
2470 static const unsigned char new_state[16] = {
2471 /* current state: new state: action: */
2472 [0 /* (Invalid) */] = TCP_CLOSE,
2473 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2474 [TCP_SYN_SENT] = TCP_CLOSE,
2475 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2476 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2477 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2478 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2479 [TCP_CLOSE] = TCP_CLOSE,
2480 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2481 [TCP_LAST_ACK] = TCP_LAST_ACK,
2482 [TCP_LISTEN] = TCP_CLOSE,
2483 [TCP_CLOSING] = TCP_CLOSING,
2484 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2487 static int mptcp_close_state(struct sock *sk)
2489 int next = (int)new_state[sk->sk_state];
2490 int ns = next & TCP_STATE_MASK;
2492 inet_sk_state_store(sk, ns);
2494 return next & TCP_ACTION_FIN;
2497 static void __mptcp_check_send_data_fin(struct sock *sk)
2499 struct mptcp_subflow_context *subflow;
2500 struct mptcp_sock *msk = mptcp_sk(sk);
2502 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2503 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2504 msk->snd_nxt, msk->write_seq);
2506 /* we still need to enqueue subflows or not really shutting down,
2509 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2510 mptcp_send_head(sk))
2513 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2515 /* fallback socket will not get data_fin/ack, can move to the next
2518 if (__mptcp_check_fallback(msk)) {
2519 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2520 inet_sk_state_store(sk, TCP_CLOSE);
2521 mptcp_close_wake_up(sk);
2522 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2523 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2527 __mptcp_flush_join_list(msk);
2528 mptcp_for_each_subflow(msk, subflow) {
2529 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2531 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2535 static void __mptcp_wr_shutdown(struct sock *sk)
2537 struct mptcp_sock *msk = mptcp_sk(sk);
2539 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2540 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2541 !!mptcp_send_head(sk));
2543 /* will be ignored by fallback sockets */
2544 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2545 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2547 __mptcp_check_send_data_fin(sk);
2550 static void __mptcp_destroy_sock(struct sock *sk)
2552 struct mptcp_subflow_context *subflow, *tmp;
2553 struct mptcp_sock *msk = mptcp_sk(sk);
2554 LIST_HEAD(conn_list);
2556 pr_debug("msk=%p", msk);
2560 /* be sure to always acquire the join list lock, to sync vs
2561 * mptcp_finish_join().
2563 spin_lock_bh(&msk->join_list_lock);
2564 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2565 spin_unlock_bh(&msk->join_list_lock);
2566 list_splice_init(&msk->conn_list, &conn_list);
2568 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2569 sk_stop_timer(sk, &sk->sk_timer);
2572 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2573 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2574 __mptcp_close_ssk(sk, ssk, subflow);
2577 sk->sk_prot->destroy(sk);
2579 WARN_ON_ONCE(msk->wmem_reserved);
2580 WARN_ON_ONCE(msk->rmem_released);
2581 sk_stream_kill_queues(sk);
2582 xfrm_sk_free_policy(sk);
2583 sk_refcnt_debug_release(sk);
2584 mptcp_dispose_initial_subflow(msk);
2588 static void mptcp_close(struct sock *sk, long timeout)
2590 struct mptcp_subflow_context *subflow;
2591 bool do_cancel_work = false;
2594 sk->sk_shutdown = SHUTDOWN_MASK;
2596 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2597 inet_sk_state_store(sk, TCP_CLOSE);
2601 if (mptcp_close_state(sk))
2602 __mptcp_wr_shutdown(sk);
2604 sk_stream_wait_close(sk, timeout);
2607 /* orphan all the subflows */
2608 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2609 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2610 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2611 bool slow = lock_sock_fast(ssk);
2614 unlock_sock_fast(ssk, slow);
2619 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2620 if (sk->sk_state == TCP_CLOSE) {
2621 __mptcp_destroy_sock(sk);
2622 do_cancel_work = true;
2624 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2628 mptcp_cancel_work(sk);
2630 if (mptcp_sk(sk)->token)
2631 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2636 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2638 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2639 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2640 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2642 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2643 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2646 msk6->saddr = ssk6->saddr;
2647 msk6->flow_label = ssk6->flow_label;
2651 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2652 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2653 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2654 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2655 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2656 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2659 static int mptcp_disconnect(struct sock *sk, int flags)
2661 struct mptcp_subflow_context *subflow;
2662 struct mptcp_sock *msk = mptcp_sk(sk);
2664 __mptcp_flush_join_list(msk);
2665 mptcp_for_each_subflow(msk, subflow) {
2666 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2669 tcp_disconnect(ssk, flags);
2675 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2676 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2678 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2680 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2684 struct sock *mptcp_sk_clone(const struct sock *sk,
2685 const struct mptcp_options_received *mp_opt,
2686 struct request_sock *req)
2688 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2689 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2690 struct mptcp_sock *msk;
2696 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2697 if (nsk->sk_family == AF_INET6)
2698 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2701 __mptcp_init_sock(nsk);
2703 msk = mptcp_sk(nsk);
2704 msk->local_key = subflow_req->local_key;
2705 msk->token = subflow_req->token;
2706 msk->subflow = NULL;
2707 WRITE_ONCE(msk->fully_established, false);
2709 msk->write_seq = subflow_req->idsn + 1;
2710 msk->snd_nxt = msk->write_seq;
2711 msk->snd_una = msk->write_seq;
2712 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2714 if (mp_opt->mp_capable) {
2715 msk->can_ack = true;
2716 msk->remote_key = mp_opt->sndr_key;
2717 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2719 WRITE_ONCE(msk->ack_seq, ack_seq);
2720 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2723 sock_reset_flag(nsk, SOCK_RCU_FREE);
2724 /* will be fully established after successful MPC subflow creation */
2725 inet_sk_state_store(nsk, TCP_SYN_RECV);
2727 security_inet_csk_clone(nsk, req);
2728 bh_unlock_sock(nsk);
2730 /* keep a single reference */
2735 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2737 const struct tcp_sock *tp = tcp_sk(ssk);
2739 msk->rcvq_space.copied = 0;
2740 msk->rcvq_space.rtt_us = 0;
2742 msk->rcvq_space.time = tp->tcp_mstamp;
2744 /* initial rcv_space offering made to peer */
2745 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2746 TCP_INIT_CWND * tp->advmss);
2747 if (msk->rcvq_space.space == 0)
2748 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2750 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2753 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2756 struct mptcp_sock *msk = mptcp_sk(sk);
2757 struct socket *listener;
2760 listener = __mptcp_nmpc_socket(msk);
2761 if (WARN_ON_ONCE(!listener)) {
2766 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2767 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2771 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2772 if (sk_is_mptcp(newsk)) {
2773 struct mptcp_subflow_context *subflow;
2774 struct sock *new_mptcp_sock;
2776 subflow = mptcp_subflow_ctx(newsk);
2777 new_mptcp_sock = subflow->conn;
2779 /* is_mptcp should be false if subflow->conn is missing, see
2780 * subflow_syn_recv_sock()
2782 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2783 tcp_sk(newsk)->is_mptcp = 0;
2787 /* acquire the 2nd reference for the owning socket */
2788 sock_hold(new_mptcp_sock);
2789 newsk = new_mptcp_sock;
2790 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2792 MPTCP_INC_STATS(sock_net(sk),
2793 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2799 void mptcp_destroy_common(struct mptcp_sock *msk)
2801 struct sock *sk = (struct sock *)msk;
2803 __mptcp_clear_xmit(sk);
2805 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2806 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2808 skb_rbtree_purge(&msk->out_of_order_queue);
2809 mptcp_token_destroy(msk);
2810 mptcp_pm_free_anno_list(msk);
2813 static void mptcp_destroy(struct sock *sk)
2815 struct mptcp_sock *msk = mptcp_sk(sk);
2817 mptcp_destroy_common(msk);
2818 sk_sockets_allocated_dec(sk);
2821 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2822 sockptr_t optval, unsigned int optlen)
2824 struct sock *sk = (struct sock *)msk;
2825 struct socket *ssock;
2832 ssock = __mptcp_nmpc_socket(msk);
2838 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2840 if (optname == SO_REUSEPORT)
2841 sk->sk_reuseport = ssock->sk->sk_reuseport;
2842 else if (optname == SO_REUSEADDR)
2843 sk->sk_reuse = ssock->sk->sk_reuse;
2849 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2852 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2853 sockptr_t optval, unsigned int optlen)
2855 struct sock *sk = (struct sock *)msk;
2856 int ret = -EOPNOTSUPP;
2857 struct socket *ssock;
2862 ssock = __mptcp_nmpc_socket(msk);
2868 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2870 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2879 static bool mptcp_unsupported(int level, int optname)
2881 if (level == SOL_IP) {
2883 case IP_ADD_MEMBERSHIP:
2884 case IP_ADD_SOURCE_MEMBERSHIP:
2885 case IP_DROP_MEMBERSHIP:
2886 case IP_DROP_SOURCE_MEMBERSHIP:
2887 case IP_BLOCK_SOURCE:
2888 case IP_UNBLOCK_SOURCE:
2889 case MCAST_JOIN_GROUP:
2890 case MCAST_LEAVE_GROUP:
2891 case MCAST_JOIN_SOURCE_GROUP:
2892 case MCAST_LEAVE_SOURCE_GROUP:
2893 case MCAST_BLOCK_SOURCE:
2894 case MCAST_UNBLOCK_SOURCE:
2895 case MCAST_MSFILTER:
2900 if (level == SOL_IPV6) {
2903 case IPV6_ADD_MEMBERSHIP:
2904 case IPV6_DROP_MEMBERSHIP:
2905 case IPV6_JOIN_ANYCAST:
2906 case IPV6_LEAVE_ANYCAST:
2907 case MCAST_JOIN_GROUP:
2908 case MCAST_LEAVE_GROUP:
2909 case MCAST_JOIN_SOURCE_GROUP:
2910 case MCAST_LEAVE_SOURCE_GROUP:
2911 case MCAST_BLOCK_SOURCE:
2912 case MCAST_UNBLOCK_SOURCE:
2913 case MCAST_MSFILTER:
2921 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2922 sockptr_t optval, unsigned int optlen)
2924 struct mptcp_sock *msk = mptcp_sk(sk);
2927 pr_debug("msk=%p", msk);
2929 if (mptcp_unsupported(level, optname))
2930 return -ENOPROTOOPT;
2932 if (level == SOL_SOCKET)
2933 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2935 /* @@ the meaning of setsockopt() when the socket is connected and
2936 * there are multiple subflows is not yet defined. It is up to the
2937 * MPTCP-level socket to configure the subflows until the subflow
2938 * is in TCP fallback, when TCP socket options are passed through
2939 * to the one remaining subflow.
2942 ssk = __mptcp_tcp_fallback(msk);
2945 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2947 if (level == SOL_IPV6)
2948 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2953 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2954 char __user *optval, int __user *option)
2956 struct mptcp_sock *msk = mptcp_sk(sk);
2959 pr_debug("msk=%p", msk);
2961 /* @@ the meaning of setsockopt() when the socket is connected and
2962 * there are multiple subflows is not yet defined. It is up to the
2963 * MPTCP-level socket to configure the subflows until the subflow
2964 * is in TCP fallback, when socket options are passed through
2965 * to the one remaining subflow.
2968 ssk = __mptcp_tcp_fallback(msk);
2971 return tcp_getsockopt(ssk, level, optname, optval, option);
2976 void __mptcp_data_acked(struct sock *sk)
2978 if (!sock_owned_by_user(sk))
2979 __mptcp_clean_una(sk);
2981 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2983 if (mptcp_pending_data_fin_ack(sk))
2984 mptcp_schedule_work(sk);
2987 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2989 if (!mptcp_send_head(sk))
2992 if (!sock_owned_by_user(sk)) {
2993 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2995 if (xmit_ssk == ssk)
2996 __mptcp_subflow_push_pending(sk, ssk);
2998 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
3000 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
3004 #define MPTCP_DEFERRED_ALL (TCPF_WRITE_TIMER_DEFERRED)
3006 /* processes deferred events and flush wmem */
3007 static void mptcp_release_cb(struct sock *sk)
3009 unsigned long flags, nflags;
3013 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
3014 flags |= BIT(MPTCP_PUSH_PENDING);
3018 /* the following actions acquire the subflow socket lock
3020 * 1) can't be invoked in atomic scope
3021 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3022 * datapath acquires the msk socket spinlock while helding
3023 * the subflow socket lock
3026 spin_unlock_bh(&sk->sk_lock.slock);
3027 if (flags & BIT(MPTCP_PUSH_PENDING))
3028 __mptcp_push_pending(sk, 0);
3031 spin_lock_bh(&sk->sk_lock.slock);
3034 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
3035 __mptcp_clean_una_wakeup(sk);
3036 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
3037 __mptcp_error_report(sk);
3039 /* push_pending may touch wmem_reserved, ensure we do the cleanup
3042 __mptcp_update_wmem(sk);
3043 __mptcp_update_rmem(sk);
3046 flags = sk->sk_tsq_flags;
3047 if (!(flags & MPTCP_DEFERRED_ALL))
3049 nflags = flags & ~MPTCP_DEFERRED_ALL;
3050 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
3052 sock_release_ownership(sk);
3054 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
3055 mptcp_retransmit_handler(sk);
3060 void mptcp_subflow_process_delegated(struct sock *ssk)
3062 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3063 struct sock *sk = subflow->conn;
3065 mptcp_data_lock(sk);
3066 if (!sock_owned_by_user(sk))
3067 __mptcp_subflow_push_pending(sk, ssk);
3069 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
3070 mptcp_data_unlock(sk);
3071 mptcp_subflow_delegated_done(subflow);
3074 static int mptcp_hash(struct sock *sk)
3076 /* should never be called,
3077 * we hash the TCP subflows not the master socket
3083 static void mptcp_unhash(struct sock *sk)
3085 /* called from sk_common_release(), but nothing to do here */
3088 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3090 struct mptcp_sock *msk = mptcp_sk(sk);
3091 struct socket *ssock;
3093 ssock = __mptcp_nmpc_socket(msk);
3094 pr_debug("msk=%p, subflow=%p", msk, ssock);
3095 if (WARN_ON_ONCE(!ssock))
3098 return inet_csk_get_port(ssock->sk, snum);
3101 void mptcp_finish_connect(struct sock *ssk)
3103 struct mptcp_subflow_context *subflow;
3104 struct mptcp_sock *msk;
3108 subflow = mptcp_subflow_ctx(ssk);
3112 pr_debug("msk=%p, token=%u", sk, subflow->token);
3114 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3116 subflow->map_seq = ack_seq;
3117 subflow->map_subflow_seq = 1;
3119 /* the socket is not connected yet, no msk/subflow ops can access/race
3120 * accessing the field below
3122 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3123 WRITE_ONCE(msk->local_key, subflow->local_key);
3124 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3125 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3126 WRITE_ONCE(msk->ack_seq, ack_seq);
3127 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3128 WRITE_ONCE(msk->can_ack, 1);
3129 WRITE_ONCE(msk->snd_una, msk->write_seq);
3131 mptcp_pm_new_connection(msk, ssk, 0);
3133 mptcp_rcv_space_init(msk, ssk);
3136 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3138 write_lock_bh(&sk->sk_callback_lock);
3139 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3140 sk_set_socket(sk, parent);
3141 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3142 write_unlock_bh(&sk->sk_callback_lock);
3145 bool mptcp_finish_join(struct sock *ssk)
3147 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3148 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3149 struct sock *parent = (void *)msk;
3150 struct socket *parent_sock;
3153 pr_debug("msk=%p, subflow=%p", msk, subflow);
3155 /* mptcp socket already closing? */
3156 if (!mptcp_is_fully_established(parent))
3159 if (!msk->pm.server_side)
3162 if (!mptcp_pm_allow_new_subflow(msk))
3165 /* active connections are already on conn_list, and we can't acquire
3167 * use the join list lock as synchronization point and double-check
3168 * msk status to avoid racing with __mptcp_destroy_sock()
3170 spin_lock_bh(&msk->join_list_lock);
3171 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3172 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3173 list_add_tail(&subflow->node, &msk->join_list);
3176 spin_unlock_bh(&msk->join_list_lock);
3180 /* attach to msk socket only after we are sure he will deal with us
3183 parent_sock = READ_ONCE(parent->sk_socket);
3184 if (parent_sock && !ssk->sk_socket)
3185 mptcp_sock_graft(ssk, parent_sock);
3186 subflow->map_seq = READ_ONCE(msk->ack_seq);
3188 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3192 static void mptcp_shutdown(struct sock *sk, int how)
3194 pr_debug("sk=%p, how=%d", sk, how);
3196 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3197 __mptcp_wr_shutdown(sk);
3200 static struct proto mptcp_prot = {
3202 .owner = THIS_MODULE,
3203 .init = mptcp_init_sock,
3204 .disconnect = mptcp_disconnect,
3205 .close = mptcp_close,
3206 .accept = mptcp_accept,
3207 .setsockopt = mptcp_setsockopt,
3208 .getsockopt = mptcp_getsockopt,
3209 .shutdown = mptcp_shutdown,
3210 .destroy = mptcp_destroy,
3211 .sendmsg = mptcp_sendmsg,
3212 .recvmsg = mptcp_recvmsg,
3213 .release_cb = mptcp_release_cb,
3215 .unhash = mptcp_unhash,
3216 .get_port = mptcp_get_port,
3217 .sockets_allocated = &mptcp_sockets_allocated,
3218 .memory_allocated = &tcp_memory_allocated,
3219 .memory_pressure = &tcp_memory_pressure,
3220 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3221 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3222 .sysctl_mem = sysctl_tcp_mem,
3223 .obj_size = sizeof(struct mptcp_sock),
3224 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3225 .no_autobind = true,
3228 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3230 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3231 struct socket *ssock;
3234 lock_sock(sock->sk);
3235 ssock = __mptcp_nmpc_socket(msk);
3241 err = ssock->ops->bind(ssock, uaddr, addr_len);
3243 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3246 release_sock(sock->sk);
3250 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3251 struct mptcp_subflow_context *subflow)
3253 subflow->request_mptcp = 0;
3254 __mptcp_do_fallback(msk);
3257 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3258 int addr_len, int flags)
3260 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3261 struct mptcp_subflow_context *subflow;
3262 struct socket *ssock;
3265 lock_sock(sock->sk);
3266 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3267 /* pending connection or invalid state, let existing subflow
3270 ssock = msk->subflow;
3274 ssock = __mptcp_nmpc_socket(msk);
3280 mptcp_token_destroy(msk);
3281 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3282 subflow = mptcp_subflow_ctx(ssock->sk);
3283 #ifdef CONFIG_TCP_MD5SIG
3284 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3287 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3288 mptcp_subflow_early_fallback(msk, subflow);
3290 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
3291 mptcp_subflow_early_fallback(msk, subflow);
3294 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3295 sock->state = ssock->state;
3297 /* on successful connect, the msk state will be moved to established by
3298 * subflow_finish_connect()
3300 if (!err || err == -EINPROGRESS)
3301 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3303 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3306 release_sock(sock->sk);
3310 static int mptcp_listen(struct socket *sock, int backlog)
3312 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3313 struct socket *ssock;
3316 pr_debug("msk=%p", msk);
3318 lock_sock(sock->sk);
3319 ssock = __mptcp_nmpc_socket(msk);
3325 mptcp_token_destroy(msk);
3326 inet_sk_state_store(sock->sk, TCP_LISTEN);
3327 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3329 err = ssock->ops->listen(ssock, backlog);
3330 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3332 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3335 release_sock(sock->sk);
3339 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3340 int flags, bool kern)
3342 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3343 struct socket *ssock;
3346 pr_debug("msk=%p", msk);
3348 lock_sock(sock->sk);
3349 if (sock->sk->sk_state != TCP_LISTEN)
3352 ssock = __mptcp_nmpc_socket(msk);
3356 clear_bit(MPTCP_DATA_READY, &msk->flags);
3357 sock_hold(ssock->sk);
3358 release_sock(sock->sk);
3360 err = ssock->ops->accept(sock, newsock, flags, kern);
3361 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3362 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3363 struct mptcp_subflow_context *subflow;
3364 struct sock *newsk = newsock->sk;
3368 /* PM/worker can now acquire the first subflow socket
3369 * lock without racing with listener queue cleanup,
3370 * we can notify it, if needed.
3372 * Even if remote has reset the initial subflow by now
3373 * the refcnt is still at least one.
3375 subflow = mptcp_subflow_ctx(msk->first);
3376 list_add(&subflow->node, &msk->conn_list);
3377 sock_hold(msk->first);
3378 if (mptcp_is_fully_established(newsk))
3379 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3381 mptcp_copy_inaddrs(newsk, msk->first);
3382 mptcp_rcv_space_init(msk, msk->first);
3383 mptcp_propagate_sndbuf(newsk, msk->first);
3385 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3386 * This is needed so NOSPACE flag can be set from tcp stack.
3388 __mptcp_flush_join_list(msk);
3389 mptcp_for_each_subflow(msk, subflow) {
3390 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3392 if (!ssk->sk_socket)
3393 mptcp_sock_graft(ssk, newsock);
3395 release_sock(newsk);
3398 if (inet_csk_listen_poll(ssock->sk))
3399 set_bit(MPTCP_DATA_READY, &msk->flags);
3400 sock_put(ssock->sk);
3404 release_sock(sock->sk);
3408 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3410 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3414 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3416 struct sock *sk = (struct sock *)msk;
3418 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3419 return EPOLLOUT | EPOLLWRNORM;
3421 if (sk_stream_is_writeable(sk))
3422 return EPOLLOUT | EPOLLWRNORM;
3424 mptcp_set_nospace(sk);
3425 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3426 if (sk_stream_is_writeable(sk))
3427 return EPOLLOUT | EPOLLWRNORM;
3432 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3433 struct poll_table_struct *wait)
3435 struct sock *sk = sock->sk;
3436 struct mptcp_sock *msk;
3441 sock_poll_wait(file, sock, wait);
3443 state = inet_sk_state_load(sk);
3444 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3445 if (state == TCP_LISTEN)
3446 return mptcp_check_readable(msk);
3448 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3449 mask |= mptcp_check_readable(msk);
3450 mask |= mptcp_check_writeable(msk);
3452 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3454 if (sk->sk_shutdown & RCV_SHUTDOWN)
3455 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3457 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3465 static const struct proto_ops mptcp_stream_ops = {
3467 .owner = THIS_MODULE,
3468 .release = inet_release,
3470 .connect = mptcp_stream_connect,
3471 .socketpair = sock_no_socketpair,
3472 .accept = mptcp_stream_accept,
3473 .getname = inet_getname,
3475 .ioctl = inet_ioctl,
3476 .gettstamp = sock_gettstamp,
3477 .listen = mptcp_listen,
3478 .shutdown = inet_shutdown,
3479 .setsockopt = sock_common_setsockopt,
3480 .getsockopt = sock_common_getsockopt,
3481 .sendmsg = inet_sendmsg,
3482 .recvmsg = inet_recvmsg,
3483 .mmap = sock_no_mmap,
3484 .sendpage = inet_sendpage,
3487 static struct inet_protosw mptcp_protosw = {
3488 .type = SOCK_STREAM,
3489 .protocol = IPPROTO_MPTCP,
3490 .prot = &mptcp_prot,
3491 .ops = &mptcp_stream_ops,
3492 .flags = INET_PROTOSW_ICSK,
3495 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3497 struct mptcp_delegated_action *delegated;
3498 struct mptcp_subflow_context *subflow;
3501 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3502 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3503 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3505 bh_lock_sock_nested(ssk);
3506 if (!sock_owned_by_user(ssk) &&
3507 mptcp_subflow_has_delegated_action(subflow))
3508 mptcp_subflow_process_delegated(ssk);
3509 /* ... elsewhere tcp_release_cb_override already processed
3510 * the action or will do at next release_sock().
3511 * In both case must dequeue the subflow here - on the same
3512 * CPU that scheduled it.
3514 bh_unlock_sock(ssk);
3517 if (++work_done == budget)
3521 /* always provide a 0 'work_done' argument, so that napi_complete_done
3522 * will not try accessing the NULL napi->dev ptr
3524 napi_complete_done(napi, 0);
3528 void __init mptcp_proto_init(void)
3530 struct mptcp_delegated_action *delegated;
3533 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3535 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3536 panic("Failed to allocate MPTCP pcpu counter\n");
3538 init_dummy_netdev(&mptcp_napi_dev);
3539 for_each_possible_cpu(cpu) {
3540 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3541 INIT_LIST_HEAD(&delegated->head);
3542 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3544 napi_enable(&delegated->napi);
3547 mptcp_subflow_init();
3551 if (proto_register(&mptcp_prot, 1) != 0)
3552 panic("Failed to register MPTCP proto.\n");
3554 inet_register_protosw(&mptcp_protosw);
3556 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3559 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3560 static const struct proto_ops mptcp_v6_stream_ops = {
3562 .owner = THIS_MODULE,
3563 .release = inet6_release,
3565 .connect = mptcp_stream_connect,
3566 .socketpair = sock_no_socketpair,
3567 .accept = mptcp_stream_accept,
3568 .getname = inet6_getname,
3570 .ioctl = inet6_ioctl,
3571 .gettstamp = sock_gettstamp,
3572 .listen = mptcp_listen,
3573 .shutdown = inet_shutdown,
3574 .setsockopt = sock_common_setsockopt,
3575 .getsockopt = sock_common_getsockopt,
3576 .sendmsg = inet6_sendmsg,
3577 .recvmsg = inet6_recvmsg,
3578 .mmap = sock_no_mmap,
3579 .sendpage = inet_sendpage,
3580 #ifdef CONFIG_COMPAT
3581 .compat_ioctl = inet6_compat_ioctl,
3585 static struct proto mptcp_v6_prot;
3587 static void mptcp_v6_destroy(struct sock *sk)
3590 inet6_destroy_sock(sk);
3593 static struct inet_protosw mptcp_v6_protosw = {
3594 .type = SOCK_STREAM,
3595 .protocol = IPPROTO_MPTCP,
3596 .prot = &mptcp_v6_prot,
3597 .ops = &mptcp_v6_stream_ops,
3598 .flags = INET_PROTOSW_ICSK,
3601 int __init mptcp_proto_v6_init(void)
3605 mptcp_v6_prot = mptcp_prot;
3606 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3607 mptcp_v6_prot.slab = NULL;
3608 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3609 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3611 err = proto_register(&mptcp_v6_prot, 1);
3615 err = inet6_register_protosw(&mptcp_v6_protosw);
3617 proto_unregister(&mptcp_v6_prot);