1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
28 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
30 struct mptcp_sock msk;
41 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
43 static struct percpu_counter mptcp_sockets_allocated;
45 static void __mptcp_destroy_sock(struct sock *sk);
46 static void __mptcp_check_send_data_fin(struct sock *sk);
48 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
49 * completed yet or has failed, return the subflow socket.
50 * Otherwise return NULL.
52 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
54 if (!msk->subflow || READ_ONCE(msk->can_ack))
60 /* Returns end sequence number of the receiver's advertised window */
61 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
63 return atomic64_read(&msk->wnd_end);
66 static bool mptcp_is_tcpsk(struct sock *sk)
68 struct socket *sock = sk->sk_socket;
70 if (unlikely(sk->sk_prot == &tcp_prot)) {
71 /* we are being invoked after mptcp_accept() has
72 * accepted a non-mp-capable flow: sk is a tcp_sk,
75 * Hand the socket over to tcp so all further socket ops
78 sock->ops = &inet_stream_ops;
80 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
81 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
82 sock->ops = &inet6_stream_ops;
90 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
92 sock_owned_by_me((const struct sock *)msk);
94 if (likely(!__mptcp_check_fallback(msk)))
100 static int __mptcp_socket_create(struct mptcp_sock *msk)
102 struct mptcp_subflow_context *subflow;
103 struct sock *sk = (struct sock *)msk;
104 struct socket *ssock;
107 err = mptcp_subflow_create_socket(sk, &ssock);
111 msk->first = ssock->sk;
112 msk->subflow = ssock;
113 subflow = mptcp_subflow_ctx(ssock->sk);
114 list_add(&subflow->node, &msk->conn_list);
115 sock_hold(ssock->sk);
116 subflow->request_mptcp = 1;
118 /* accept() will wait on first subflow sk_wq, and we always wakes up
121 RCU_INIT_POINTER(msk->first->sk_wq, &sk->sk_socket->wq);
126 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
128 sk_drops_add(sk, skb);
132 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
133 struct sk_buff *from)
138 if (MPTCP_SKB_CB(from)->offset ||
139 !skb_try_coalesce(to, from, &fragstolen, &delta))
142 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
143 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
144 to->len, MPTCP_SKB_CB(from)->end_seq);
145 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
146 kfree_skb_partial(from, fragstolen);
147 atomic_add(delta, &sk->sk_rmem_alloc);
148 sk_mem_charge(sk, delta);
152 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
153 struct sk_buff *from)
155 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
158 return mptcp_try_coalesce((struct sock *)msk, to, from);
161 /* "inspired" by tcp_data_queue_ofo(), main differences:
163 * - don't cope with sacks
165 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
167 struct sock *sk = (struct sock *)msk;
168 struct rb_node **p, *parent;
169 u64 seq, end_seq, max_seq;
170 struct sk_buff *skb1;
172 seq = MPTCP_SKB_CB(skb)->map_seq;
173 end_seq = MPTCP_SKB_CB(skb)->end_seq;
174 max_seq = READ_ONCE(msk->rcv_wnd_sent);
176 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
177 RB_EMPTY_ROOT(&msk->out_of_order_queue));
178 if (after64(end_seq, max_seq)) {
181 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
182 (unsigned long long)end_seq - (unsigned long)max_seq,
183 (unsigned long long)msk->rcv_wnd_sent);
184 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
188 p = &msk->out_of_order_queue.rb_node;
189 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
190 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
191 rb_link_node(&skb->rbnode, NULL, p);
192 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
193 msk->ooo_last_skb = skb;
197 /* with 2 subflows, adding at end of ooo queue is quite likely
198 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
200 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
201 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
202 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
206 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
207 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
208 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
209 parent = &msk->ooo_last_skb->rbnode;
210 p = &parent->rb_right;
214 /* Find place to insert this segment. Handle overlaps on the way. */
218 skb1 = rb_to_skb(parent);
219 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
220 p = &parent->rb_left;
223 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
224 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
225 /* All the bits are present. Drop. */
227 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
230 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
234 * continue traversing
237 /* skb's seq == skb1's seq and skb covers skb1.
238 * Replace skb1 with skb.
240 rb_replace_node(&skb1->rbnode, &skb->rbnode,
241 &msk->out_of_order_queue);
242 mptcp_drop(sk, skb1);
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
246 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
250 p = &parent->rb_right;
254 /* Insert segment into RB tree. */
255 rb_link_node(&skb->rbnode, parent, p);
256 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
259 /* Remove other segments covered by skb. */
260 while ((skb1 = skb_rb_next(skb)) != NULL) {
261 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
263 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
264 mptcp_drop(sk, skb1);
265 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
267 /* If there is no skb after us, we are the last_skb ! */
269 msk->ooo_last_skb = skb;
273 skb_set_owner_r(skb, sk);
276 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
277 struct sk_buff *skb, unsigned int offset,
280 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
281 struct sock *sk = (struct sock *)msk;
282 struct sk_buff *tail;
284 __skb_unlink(skb, &ssk->sk_receive_queue);
289 /* try to fetch required memory from subflow */
290 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
291 if (ssk->sk_forward_alloc < skb->truesize)
293 __sk_mem_reclaim(ssk, skb->truesize);
294 if (!sk_rmem_schedule(sk, skb, skb->truesize))
298 /* the skb map_seq accounts for the skb offset:
299 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
302 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
303 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
304 MPTCP_SKB_CB(skb)->offset = offset;
306 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
308 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
309 tail = skb_peek_tail(&sk->sk_receive_queue);
310 if (tail && mptcp_try_coalesce(sk, tail, skb))
313 skb_set_owner_r(skb, sk);
314 __skb_queue_tail(&sk->sk_receive_queue, skb);
316 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
317 mptcp_data_queue_ofo(msk, skb);
321 /* old data, keep it simple and drop the whole pkt, sender
322 * will retransmit as needed, if needed.
324 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
330 static void mptcp_stop_timer(struct sock *sk)
332 struct inet_connection_sock *icsk = inet_csk(sk);
334 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
335 mptcp_sk(sk)->timer_ival = 0;
338 static void mptcp_close_wake_up(struct sock *sk)
340 if (sock_flag(sk, SOCK_DEAD))
343 sk->sk_state_change(sk);
344 if (sk->sk_shutdown == SHUTDOWN_MASK ||
345 sk->sk_state == TCP_CLOSE)
346 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
348 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
351 static void mptcp_check_data_fin_ack(struct sock *sk)
353 struct mptcp_sock *msk = mptcp_sk(sk);
355 if (__mptcp_check_fallback(msk))
358 /* Look for an acknowledged DATA_FIN */
359 if (((1 << sk->sk_state) &
360 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
361 msk->write_seq == atomic64_read(&msk->snd_una)) {
362 mptcp_stop_timer(sk);
364 WRITE_ONCE(msk->snd_data_fin_enable, 0);
366 switch (sk->sk_state) {
368 inet_sk_state_store(sk, TCP_FIN_WAIT2);
372 inet_sk_state_store(sk, TCP_CLOSE);
376 mptcp_close_wake_up(sk);
380 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
382 struct mptcp_sock *msk = mptcp_sk(sk);
384 if (READ_ONCE(msk->rcv_data_fin) &&
385 ((1 << sk->sk_state) &
386 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
387 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
389 if (msk->ack_seq == rcv_data_fin_seq) {
391 *seq = rcv_data_fin_seq;
400 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
402 long tout = ssk && inet_csk(ssk)->icsk_pending ?
403 inet_csk(ssk)->icsk_timeout - jiffies : 0;
406 tout = mptcp_sk(sk)->timer_ival;
407 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
410 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
412 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
414 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
415 if (subflow->request_join && !subflow->fully_established)
418 /* only send if our side has not closed yet */
419 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
422 static bool tcp_can_send_ack(const struct sock *ssk)
424 return !((1 << inet_sk_state_load(ssk)) &
425 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE));
428 static void mptcp_send_ack(struct mptcp_sock *msk)
430 struct mptcp_subflow_context *subflow;
432 mptcp_for_each_subflow(msk, subflow) {
433 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
436 if (tcp_can_send_ack(ssk))
442 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
447 ret = tcp_can_send_ack(ssk);
449 tcp_cleanup_rbuf(ssk, 1);
454 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
456 struct mptcp_subflow_context *subflow;
458 /* if the hinted ssk is still active, try to use it */
459 if (likely(msk->ack_hint)) {
460 mptcp_for_each_subflow(msk, subflow) {
461 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
463 if (msk->ack_hint == ssk &&
464 mptcp_subflow_cleanup_rbuf(ssk))
469 /* otherwise pick the first active subflow */
470 mptcp_for_each_subflow(msk, subflow)
471 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
475 static bool mptcp_check_data_fin(struct sock *sk)
477 struct mptcp_sock *msk = mptcp_sk(sk);
478 u64 rcv_data_fin_seq;
481 if (__mptcp_check_fallback(msk) || !msk->first)
484 /* Need to ack a DATA_FIN received from a peer while this side
485 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
486 * msk->rcv_data_fin was set when parsing the incoming options
487 * at the subflow level and the msk lock was not held, so this
488 * is the first opportunity to act on the DATA_FIN and change
491 * If we are caught up to the sequence number of the incoming
492 * DATA_FIN, send the DATA_ACK now and do state transition. If
493 * not caught up, do nothing and let the recv code send DATA_ACK
497 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
498 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
499 WRITE_ONCE(msk->rcv_data_fin, 0);
501 sk->sk_shutdown |= RCV_SHUTDOWN;
502 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
503 set_bit(MPTCP_DATA_READY, &msk->flags);
505 switch (sk->sk_state) {
506 case TCP_ESTABLISHED:
507 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
510 inet_sk_state_store(sk, TCP_CLOSING);
513 inet_sk_state_store(sk, TCP_CLOSE);
516 /* Other states not expected */
522 mptcp_set_timeout(sk, NULL);
524 mptcp_close_wake_up(sk);
529 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
533 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
534 struct sock *sk = (struct sock *)msk;
535 unsigned int moved = 0;
536 bool more_data_avail;
541 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
543 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
544 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
546 if (unlikely(ssk_rbuf > sk_rbuf)) {
547 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
552 pr_debug("msk=%p ssk=%p", msk, ssk);
555 u32 map_remaining, offset;
556 u32 seq = tp->copied_seq;
560 /* try to move as much data as available */
561 map_remaining = subflow->map_data_len -
562 mptcp_subflow_get_map_offset(subflow);
564 skb = skb_peek(&ssk->sk_receive_queue);
566 /* if no data is found, a racing workqueue/recvmsg
567 * already processed the new data, stop here or we
568 * can enter an infinite loop
575 if (__mptcp_check_fallback(msk)) {
576 /* if we are running under the workqueue, TCP could have
577 * collapsed skbs between dummy map creation and now
578 * be sure to adjust the size
580 map_remaining = skb->len;
581 subflow->map_data_len = skb->len;
584 offset = seq - TCP_SKB_CB(skb)->seq;
585 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
591 if (offset < skb->len) {
592 size_t len = skb->len - offset;
597 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
601 if (WARN_ON_ONCE(map_remaining < len))
605 sk_eat_skb(ssk, skb);
609 WRITE_ONCE(tp->copied_seq, seq);
610 more_data_avail = mptcp_subflow_data_available(ssk);
612 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
616 } while (more_data_avail);
623 static bool mptcp_ofo_queue(struct mptcp_sock *msk)
625 struct sock *sk = (struct sock *)msk;
626 struct sk_buff *skb, *tail;
631 p = rb_first(&msk->out_of_order_queue);
632 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
635 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
639 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
641 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
644 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
648 end_seq = MPTCP_SKB_CB(skb)->end_seq;
649 tail = skb_peek_tail(&sk->sk_receive_queue);
650 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
651 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
653 /* skip overlapping data, if any */
654 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
655 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
657 MPTCP_SKB_CB(skb)->offset += delta;
658 __skb_queue_tail(&sk->sk_receive_queue, skb);
660 msk->ack_seq = end_seq;
666 /* In most cases we will be able to lock the mptcp socket. If its already
667 * owned, we need to defer to the work queue to avoid ABBA deadlock.
669 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
671 struct sock *sk = (struct sock *)msk;
672 unsigned int moved = 0;
674 if (READ_ONCE(sk->sk_lock.owned))
677 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
680 /* must re-check after taking the lock */
681 if (!READ_ONCE(sk->sk_lock.owned)) {
682 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
683 mptcp_ofo_queue(msk);
685 /* If the moves have caught up with the DATA_FIN sequence number
686 * it's time to ack the DATA_FIN and change socket state, but
687 * this is not a good place to change state. Let the workqueue
690 if (mptcp_pending_data_fin(sk, NULL))
691 mptcp_schedule_work(sk);
694 spin_unlock_bh(&sk->sk_lock.slock);
699 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
701 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
702 struct mptcp_sock *msk = mptcp_sk(sk);
703 int sk_rbuf, ssk_rbuf;
706 /* move_skbs_to_msk below can legitly clear the data_avail flag,
707 * but we will need later to properly woke the reader, cache its
710 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
712 set_bit(MPTCP_DATA_READY, &msk->flags);
714 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
715 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
716 if (unlikely(ssk_rbuf > sk_rbuf))
719 /* over limit? can't append more skbs to msk */
720 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
723 move_skbs_to_msk(msk, ssk);
727 sk->sk_data_ready(sk);
730 void __mptcp_flush_join_list(struct mptcp_sock *msk)
732 if (likely(list_empty(&msk->join_list)))
735 spin_lock_bh(&msk->join_list_lock);
736 list_splice_tail_init(&msk->join_list, &msk->conn_list);
737 spin_unlock_bh(&msk->join_list_lock);
740 static bool mptcp_timer_pending(struct sock *sk)
742 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
745 static void mptcp_reset_timer(struct sock *sk)
747 struct inet_connection_sock *icsk = inet_csk(sk);
750 /* prevent rescheduling on close */
751 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
754 /* should never be called with mptcp level timer cleared */
755 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
756 if (WARN_ON_ONCE(!tout))
758 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
761 bool mptcp_schedule_work(struct sock *sk)
763 if (inet_sk_state_load(sk) != TCP_CLOSE &&
764 schedule_work(&mptcp_sk(sk)->work)) {
765 /* each subflow already holds a reference to the sk, and the
766 * workqueue is invoked by a subflow, so sk can't go away here.
774 void mptcp_data_acked(struct sock *sk)
776 mptcp_reset_timer(sk);
778 if ((test_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags) ||
779 mptcp_send_head(sk) ||
780 (inet_sk_state_load(sk) != TCP_ESTABLISHED)))
781 mptcp_schedule_work(sk);
784 void mptcp_subflow_eof(struct sock *sk)
786 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
787 mptcp_schedule_work(sk);
790 static void mptcp_check_for_eof(struct mptcp_sock *msk)
792 struct mptcp_subflow_context *subflow;
793 struct sock *sk = (struct sock *)msk;
796 mptcp_for_each_subflow(msk, subflow)
797 receivers += !subflow->rx_eof;
801 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
802 /* hopefully temporary hack: propagate shutdown status
803 * to msk, when all subflows agree on it
805 sk->sk_shutdown |= RCV_SHUTDOWN;
807 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
808 set_bit(MPTCP_DATA_READY, &msk->flags);
809 sk->sk_data_ready(sk);
812 switch (sk->sk_state) {
813 case TCP_ESTABLISHED:
814 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
817 inet_sk_state_store(sk, TCP_CLOSING);
820 inet_sk_state_store(sk, TCP_CLOSE);
825 mptcp_close_wake_up(sk);
828 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
830 const struct sock *sk = (const struct sock *)msk;
832 if (!msk->cached_ext)
833 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
835 return !!msk->cached_ext;
838 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
840 struct mptcp_subflow_context *subflow;
841 struct sock *sk = (struct sock *)msk;
843 sock_owned_by_me(sk);
845 mptcp_for_each_subflow(msk, subflow) {
846 if (subflow->data_avail)
847 return mptcp_subflow_tcp_sock(subflow);
853 static bool mptcp_skb_can_collapse_to(u64 write_seq,
854 const struct sk_buff *skb,
855 const struct mptcp_ext *mpext)
857 if (!tcp_skb_can_collapse_to(skb))
860 /* can collapse only if MPTCP level sequence is in order and this
861 * mapping has not been xmitted yet
863 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
867 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
868 const struct page_frag *pfrag,
869 const struct mptcp_data_frag *df)
871 return df && pfrag->page == df->page &&
872 pfrag->size - pfrag->offset > 0 &&
873 df->data_seq + df->data_len == msk->write_seq;
876 static void dfrag_uncharge(struct sock *sk, int len)
878 sk_mem_uncharge(sk, len);
879 sk_wmem_queued_add(sk, -len);
882 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
884 int len = dfrag->data_len + dfrag->overhead;
886 list_del(&dfrag->list);
887 dfrag_uncharge(sk, len);
888 put_page(dfrag->page);
891 static void mptcp_clean_una(struct sock *sk)
893 struct mptcp_sock *msk = mptcp_sk(sk);
894 struct mptcp_data_frag *dtmp, *dfrag;
895 bool cleaned = false;
898 /* on fallback we just need to ignore snd_una, as this is really
901 if (__mptcp_check_fallback(msk))
902 atomic64_set(&msk->snd_una, msk->snd_nxt);
904 snd_una = atomic64_read(&msk->snd_una);
906 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
907 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
910 if (WARN_ON_ONCE(dfrag == msk->first_pending))
912 dfrag_clear(sk, dfrag);
916 dfrag = mptcp_rtx_head(sk);
917 if (dfrag && after64(snd_una, dfrag->data_seq)) {
918 u64 delta = snd_una - dfrag->data_seq;
920 if (WARN_ON_ONCE(delta > dfrag->already_sent))
923 dfrag->data_seq += delta;
924 dfrag->offset += delta;
925 dfrag->data_len -= delta;
926 dfrag->already_sent -= delta;
928 dfrag_uncharge(sk, delta);
934 sk_mem_reclaim_partial(sk);
937 static void mptcp_clean_una_wakeup(struct sock *sk)
939 struct mptcp_sock *msk = mptcp_sk(sk);
943 /* Only wake up writers if a subflow is ready */
944 if (sk_stream_is_writeable(sk)) {
945 clear_bit(MPTCP_NOSPACE, &msk->flags);
946 sk_stream_write_space(sk);
950 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
953 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
955 struct mptcp_subflow_context *subflow;
956 struct mptcp_sock *msk = mptcp_sk(sk);
959 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
960 pfrag, sk->sk_allocation)))
963 sk_stream_moderate_sndbuf(sk);
964 mptcp_for_each_subflow(msk, subflow) {
965 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
968 tcp_enter_memory_pressure(ssk);
969 sk_stream_moderate_sndbuf(ssk);
975 static struct mptcp_data_frag *
976 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
979 int offset = ALIGN(orig_offset, sizeof(long));
980 struct mptcp_data_frag *dfrag;
982 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
984 dfrag->data_seq = msk->write_seq;
985 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
986 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
987 dfrag->already_sent = 0;
988 dfrag->page = pfrag->page;
993 struct mptcp_sendmsg_info {
1001 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1004 u64 window_end = mptcp_wnd_end(msk);
1006 if (__mptcp_check_fallback(msk))
1009 if (!before64(data_seq + avail_size, window_end)) {
1010 u64 allowed_size = window_end - data_seq;
1012 return min_t(unsigned int, allowed_size, avail_size);
1018 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1019 struct mptcp_data_frag *dfrag,
1020 struct mptcp_sendmsg_info *info)
1022 u64 data_seq = dfrag->data_seq + info->sent;
1023 struct mptcp_sock *msk = mptcp_sk(sk);
1024 bool zero_window_probe = false;
1025 struct mptcp_ext *mpext = NULL;
1026 struct sk_buff *skb, *tail;
1027 bool can_collapse = false;
1031 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
1032 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1034 /* compute send limit */
1035 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1036 avail_size = info->size_goal;
1037 skb = tcp_write_queue_tail(ssk);
1039 /* Limit the write to the size available in the
1040 * current skb, if any, so that we create at most a new skb.
1041 * Explicitly tells TCP internals to avoid collapsing on later
1042 * queue management operation, to avoid breaking the ext <->
1043 * SSN association set here
1045 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1046 can_collapse = (info->size_goal - skb->len > 0) &&
1047 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1049 TCP_SKB_CB(skb)->eor = 1;
1051 avail_size = info->size_goal - skb->len;
1054 /* Zero window and all data acked? Probe. */
1055 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1056 if (avail_size == 0) {
1057 if (skb || atomic64_read(&msk->snd_una) != msk->snd_nxt)
1059 zero_window_probe = true;
1060 data_seq = atomic64_read(&msk->snd_una) - 1;
1064 if (WARN_ON_ONCE(info->sent > info->limit ||
1065 info->limit > dfrag->data_len))
1068 ret = info->limit - info->sent;
1069 tail = tcp_build_frag(ssk, avail_size, info->flags, dfrag->page,
1070 dfrag->offset + info->sent, &ret);
1072 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1076 /* if the tail skb is still the cached one, collapsing really happened.
1079 WARN_ON_ONCE(!can_collapse);
1080 mpext->data_len += ret;
1081 WARN_ON_ONCE(zero_window_probe);
1085 mpext = __skb_ext_set(tail, SKB_EXT_MPTCP, msk->cached_ext);
1086 msk->cached_ext = NULL;
1088 memset(mpext, 0, sizeof(*mpext));
1089 mpext->data_seq = data_seq;
1090 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1091 mpext->data_len = ret;
1095 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1096 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1099 if (zero_window_probe) {
1100 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1103 tcp_push_pending_frames(ssk);
1106 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1110 static void mptcp_nospace(struct mptcp_sock *msk)
1112 struct mptcp_subflow_context *subflow;
1114 set_bit(MPTCP_NOSPACE, &msk->flags);
1115 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
1117 mptcp_for_each_subflow(msk, subflow) {
1118 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1119 bool ssk_writeable = sk_stream_is_writeable(ssk);
1120 struct socket *sock = READ_ONCE(ssk->sk_socket);
1122 if (ssk_writeable || !sock)
1125 /* enables ssk->write_space() callbacks */
1126 set_bit(SOCK_NOSPACE, &sock->flags);
1129 /* mptcp_data_acked() could run just before we set the NOSPACE bit,
1130 * so explicitly check for snd_una value
1132 mptcp_clean_una((struct sock *)msk);
1135 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1136 sizeof(struct tcphdr) - \
1137 MAX_TCP_OPTION_SPACE - \
1138 sizeof(struct ipv6hdr) - \
1139 sizeof(struct frag_hdr))
1141 struct subflow_send_info {
1146 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk,
1149 struct subflow_send_info send_info[2];
1150 struct mptcp_subflow_context *subflow;
1151 int i, nr_active = 0;
1156 sock_owned_by_me((struct sock *)msk);
1159 if (!mptcp_ext_cache_refill(msk))
1162 if (__mptcp_check_fallback(msk)) {
1165 *sndbuf = msk->first->sk_sndbuf;
1166 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1169 /* re-use last subflow, if the burst allow that */
1170 if (msk->last_snd && msk->snd_burst > 0 &&
1171 sk_stream_memory_free(msk->last_snd) &&
1172 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1173 mptcp_for_each_subflow(msk, subflow) {
1174 ssk = mptcp_subflow_tcp_sock(subflow);
1175 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1177 return msk->last_snd;
1180 /* pick the subflow with the lower wmem/wspace ratio */
1181 for (i = 0; i < 2; ++i) {
1182 send_info[i].ssk = NULL;
1183 send_info[i].ratio = -1;
1185 mptcp_for_each_subflow(msk, subflow) {
1186 ssk = mptcp_subflow_tcp_sock(subflow);
1187 if (!mptcp_subflow_active(subflow))
1190 nr_active += !subflow->backup;
1191 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1192 if (!sk_stream_memory_free(subflow->tcp_sock))
1195 pace = READ_ONCE(ssk->sk_pacing_rate);
1199 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1201 if (ratio < send_info[subflow->backup].ratio) {
1202 send_info[subflow->backup].ssk = ssk;
1203 send_info[subflow->backup].ratio = ratio;
1207 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1208 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1209 send_info[1].ssk, send_info[1].ratio);
1211 /* pick the best backup if no other subflow is active */
1213 send_info[0].ssk = send_info[1].ssk;
1215 if (send_info[0].ssk) {
1216 msk->last_snd = send_info[0].ssk;
1217 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1218 sk_stream_wspace(msk->last_snd));
1219 return msk->last_snd;
1224 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1225 struct mptcp_sendmsg_info *info)
1227 mptcp_set_timeout(sk, ssk);
1228 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1232 static void mptcp_push_pending(struct sock *sk, unsigned int flags)
1234 struct sock *prev_ssk = NULL, *ssk = NULL;
1235 struct mptcp_sock *msk = mptcp_sk(sk);
1236 struct mptcp_sendmsg_info info = {
1239 struct mptcp_data_frag *dfrag;
1240 int len, copied = 0;
1243 while ((dfrag = mptcp_send_head(sk))) {
1244 info.sent = dfrag->already_sent;
1245 info.limit = dfrag->data_len;
1246 len = dfrag->data_len - dfrag->already_sent;
1251 __mptcp_flush_join_list(msk);
1252 ssk = mptcp_subflow_get_send(msk, &sndbuf);
1254 /* do auto tuning */
1255 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1256 sndbuf > READ_ONCE(sk->sk_sndbuf))
1257 WRITE_ONCE(sk->sk_sndbuf, sndbuf);
1259 /* try to keep the subflow socket lock across
1260 * consecutive xmit on the same socket
1262 if (ssk != prev_ssk && prev_ssk)
1263 mptcp_push_release(sk, prev_ssk, &info);
1267 if (ssk != prev_ssk || !prev_ssk)
1270 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1272 mptcp_push_release(sk, ssk, &info);
1277 dfrag->already_sent += ret;
1278 msk->snd_nxt += ret;
1279 msk->snd_burst -= ret;
1283 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1286 /* at this point we held the socket lock for the last subflow we used */
1288 mptcp_push_release(sk, ssk, &info);
1292 /* start the timer, if it's not pending */
1293 if (!mptcp_timer_pending(sk))
1294 mptcp_reset_timer(sk);
1295 __mptcp_check_send_data_fin(sk);
1299 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1301 struct mptcp_sock *msk = mptcp_sk(sk);
1302 struct page_frag *pfrag;
1307 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1312 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1314 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1315 ret = sk_stream_wait_connect(sk, &timeo);
1320 pfrag = sk_page_frag(sk);
1321 mptcp_clean_una(sk);
1323 while (msg_data_left(msg)) {
1324 struct mptcp_data_frag *dfrag;
1325 int frag_truesize = 0;
1326 bool dfrag_collapsed;
1327 size_t psize, offset;
1329 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1334 /* reuse tail pfrag, if possible, or carve a new one from the
1337 dfrag = mptcp_pending_tail(sk);
1338 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1339 if (!dfrag_collapsed) {
1340 if (!sk_stream_memory_free(sk)) {
1341 mptcp_push_pending(sk, msg->msg_flags);
1342 if (!sk_stream_memory_free(sk))
1343 goto wait_for_memory;
1345 if (!mptcp_page_frag_refill(sk, pfrag))
1346 goto wait_for_memory;
1348 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1349 frag_truesize = dfrag->overhead;
1352 /* we do not bound vs wspace, to allow a single packet.
1353 * memory accounting will prevent execessive memory usage
1356 offset = dfrag->offset + dfrag->data_len;
1357 psize = pfrag->size - offset;
1358 psize = min_t(size_t, psize, msg_data_left(msg));
1359 if (!sk_wmem_schedule(sk, psize + frag_truesize))
1360 goto wait_for_memory;
1362 if (copy_page_from_iter(dfrag->page, offset, psize,
1363 &msg->msg_iter) != psize) {
1368 /* data successfully copied into the write queue */
1370 dfrag->data_len += psize;
1371 frag_truesize += psize;
1372 pfrag->offset += frag_truesize;
1373 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1375 /* charge data on mptcp pending queue to the msk socket
1376 * Note: we charge such data both to sk and ssk
1378 sk_wmem_queued_add(sk, frag_truesize);
1379 sk->sk_forward_alloc -= frag_truesize;
1380 if (!dfrag_collapsed) {
1381 get_page(dfrag->page);
1382 list_add_tail(&dfrag->list, &msk->rtx_queue);
1383 if (!msk->first_pending)
1384 WRITE_ONCE(msk->first_pending, dfrag);
1386 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1387 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1390 if (!mptcp_ext_cache_refill(msk))
1391 goto wait_for_memory;
1396 if (mptcp_timer_pending(sk))
1397 mptcp_reset_timer(sk);
1398 ret = sk_stream_wait_memory(sk, &timeo);
1404 mptcp_push_pending(sk, msg->msg_flags);
1408 return copied ? : ret;
1411 static void mptcp_wait_data(struct sock *sk, long *timeo)
1413 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1414 struct mptcp_sock *msk = mptcp_sk(sk);
1416 add_wait_queue(sk_sleep(sk), &wait);
1417 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1419 sk_wait_event(sk, timeo,
1420 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1422 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1423 remove_wait_queue(sk_sleep(sk), &wait);
1426 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1430 struct sock *sk = (struct sock *)msk;
1431 struct sk_buff *skb;
1434 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1435 u32 offset = MPTCP_SKB_CB(skb)->offset;
1436 u32 data_len = skb->len - offset;
1437 u32 count = min_t(size_t, len - copied, data_len);
1440 err = skb_copy_datagram_msg(skb, offset, msg, count);
1441 if (unlikely(err < 0)) {
1449 if (count < data_len) {
1450 MPTCP_SKB_CB(skb)->offset += count;
1454 __skb_unlink(skb, &sk->sk_receive_queue);
1464 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1466 * Only difference: Use highest rtt estimate of the subflows in use.
1468 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1470 struct mptcp_subflow_context *subflow;
1471 struct sock *sk = (struct sock *)msk;
1472 u32 time, advmss = 1;
1475 sock_owned_by_me(sk);
1480 msk->rcvq_space.copied += copied;
1482 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1483 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1485 rtt_us = msk->rcvq_space.rtt_us;
1486 if (rtt_us && time < (rtt_us >> 3))
1490 mptcp_for_each_subflow(msk, subflow) {
1491 const struct tcp_sock *tp;
1495 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1497 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1498 sf_advmss = READ_ONCE(tp->advmss);
1500 rtt_us = max(sf_rtt_us, rtt_us);
1501 advmss = max(sf_advmss, advmss);
1504 msk->rcvq_space.rtt_us = rtt_us;
1505 if (time < (rtt_us >> 3) || rtt_us == 0)
1508 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1511 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1512 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1516 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1518 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1520 do_div(grow, msk->rcvq_space.space);
1521 rcvwin += (grow << 1);
1523 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1524 while (tcp_win_from_space(sk, rcvmem) < advmss)
1527 do_div(rcvwin, advmss);
1528 rcvbuf = min_t(u64, rcvwin * rcvmem,
1529 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1531 if (rcvbuf > sk->sk_rcvbuf) {
1534 window_clamp = tcp_win_from_space(sk, rcvbuf);
1535 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1537 /* Make subflows follow along. If we do not do this, we
1538 * get drops at subflow level if skbs can't be moved to
1539 * the mptcp rx queue fast enough (announced rcv_win can
1540 * exceed ssk->sk_rcvbuf).
1542 mptcp_for_each_subflow(msk, subflow) {
1546 ssk = mptcp_subflow_tcp_sock(subflow);
1547 slow = lock_sock_fast(ssk);
1548 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1549 tcp_sk(ssk)->window_clamp = window_clamp;
1550 tcp_cleanup_rbuf(ssk, 1);
1551 unlock_sock_fast(ssk, slow);
1556 msk->rcvq_space.space = msk->rcvq_space.copied;
1558 msk->rcvq_space.copied = 0;
1559 msk->rcvq_space.time = mstamp;
1562 static bool __mptcp_move_skbs(struct mptcp_sock *msk, unsigned int rcv)
1564 unsigned int moved = 0;
1567 /* avoid looping forever below on racing close */
1568 if (((struct sock *)msk)->sk_state == TCP_CLOSE)
1571 __mptcp_flush_join_list(msk);
1573 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1579 slowpath = lock_sock_fast(ssk);
1580 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1582 WRITE_ONCE(msk->rmem_pending, min(rcv, moved));
1583 tcp_cleanup_rbuf(ssk, 1);
1584 WRITE_ONCE(msk->rmem_pending, 0);
1586 unlock_sock_fast(ssk, slowpath);
1589 if (mptcp_ofo_queue(msk) || moved > 0) {
1590 mptcp_check_data_fin((struct sock *)msk);
1596 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1597 int nonblock, int flags, int *addr_len)
1599 struct mptcp_sock *msk = mptcp_sk(sk);
1604 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1608 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1613 timeo = sock_rcvtimeo(sk, nonblock);
1615 len = min_t(size_t, len, INT_MAX);
1616 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1617 __mptcp_flush_join_list(msk);
1620 int bytes_read, old_space;
1622 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1623 if (unlikely(bytes_read < 0)) {
1625 copied = bytes_read;
1629 copied += bytes_read;
1631 if (skb_queue_empty(&sk->sk_receive_queue) &&
1632 __mptcp_move_skbs(msk, len - copied))
1635 /* be sure to advertise window change */
1636 old_space = READ_ONCE(msk->old_wspace);
1637 if ((tcp_space(sk) - old_space) >= old_space)
1638 mptcp_cleanup_rbuf(msk);
1640 /* only the master socket status is relevant here. The exit
1641 * conditions mirror closely tcp_recvmsg()
1643 if (copied >= target)
1648 sk->sk_state == TCP_CLOSE ||
1649 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1651 signal_pending(current))
1655 copied = sock_error(sk);
1659 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1660 mptcp_check_for_eof(msk);
1662 if (sk->sk_shutdown & RCV_SHUTDOWN)
1665 if (sk->sk_state == TCP_CLOSE) {
1675 if (signal_pending(current)) {
1676 copied = sock_intr_errno(timeo);
1681 pr_debug("block timeout %ld", timeo);
1682 mptcp_wait_data(sk, &timeo);
1685 if (skb_queue_empty(&sk->sk_receive_queue)) {
1686 /* entire backlog drained, clear DATA_READY. */
1687 clear_bit(MPTCP_DATA_READY, &msk->flags);
1689 /* .. race-breaker: ssk might have gotten new data
1690 * after last __mptcp_move_skbs() returned false.
1692 if (unlikely(__mptcp_move_skbs(msk, 0)))
1693 set_bit(MPTCP_DATA_READY, &msk->flags);
1694 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1695 /* data to read but mptcp_wait_data() cleared DATA_READY */
1696 set_bit(MPTCP_DATA_READY, &msk->flags);
1699 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
1700 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
1701 skb_queue_empty(&sk->sk_receive_queue), copied);
1702 mptcp_rcv_space_adjust(msk, copied);
1708 static void mptcp_retransmit_handler(struct sock *sk)
1710 struct mptcp_sock *msk = mptcp_sk(sk);
1712 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->snd_nxt)) {
1713 mptcp_stop_timer(sk);
1715 set_bit(MPTCP_WORK_RTX, &msk->flags);
1716 mptcp_schedule_work(sk);
1720 static void mptcp_retransmit_timer(struct timer_list *t)
1722 struct inet_connection_sock *icsk = from_timer(icsk, t,
1723 icsk_retransmit_timer);
1724 struct sock *sk = &icsk->icsk_inet.sk;
1727 if (!sock_owned_by_user(sk)) {
1728 mptcp_retransmit_handler(sk);
1730 /* delegate our work to tcp_release_cb() */
1731 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1739 static void mptcp_timeout_timer(struct timer_list *t)
1741 struct sock *sk = from_timer(sk, t, sk_timer);
1743 mptcp_schedule_work(sk);
1747 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1750 * A backup subflow is returned only if that is the only kind available.
1752 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1754 struct mptcp_subflow_context *subflow;
1755 struct sock *backup = NULL;
1757 sock_owned_by_me((const struct sock *)msk);
1759 if (__mptcp_check_fallback(msk))
1762 mptcp_for_each_subflow(msk, subflow) {
1763 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1765 if (!mptcp_subflow_active(subflow))
1768 /* still data outstanding at TCP level? Don't retransmit. */
1769 if (!tcp_write_queue_empty(ssk)) {
1770 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
1775 if (subflow->backup) {
1787 /* subflow sockets can be either outgoing (connect) or incoming
1790 * Outgoing subflows use in-kernel sockets.
1791 * Incoming subflows do not have their own 'struct socket' allocated,
1792 * so we need to use tcp_close() after detaching them from the mptcp
1795 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1796 struct mptcp_subflow_context *subflow)
1798 bool dispose_socket = false;
1799 struct socket *sock;
1801 list_del(&subflow->node);
1805 /* if we are invoked by the msk cleanup code, the subflow is
1808 sock = ssk->sk_socket;
1810 dispose_socket = sock != sk->sk_socket;
1814 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
1815 * the ssk has been already destroyed, we just need to release the
1816 * reference owned by msk;
1818 if (!inet_csk(ssk)->icsk_ulp_ops) {
1819 kfree_rcu(subflow, rcu);
1821 /* otherwise ask tcp do dispose of ssk and subflow ctx */
1822 subflow->disposable = 1;
1823 __tcp_close(ssk, 0);
1825 /* close acquired an extra ref */
1830 iput(SOCK_INODE(sock));
1835 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1840 static void pm_work(struct mptcp_sock *msk)
1842 struct mptcp_pm_data *pm = &msk->pm;
1844 spin_lock_bh(&msk->pm.lock);
1846 pr_debug("msk=%p status=%x", msk, pm->status);
1847 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1848 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1849 mptcp_pm_nl_add_addr_received(msk);
1851 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_SEND_ACK)) {
1852 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_SEND_ACK);
1853 mptcp_pm_nl_add_addr_send_ack(msk);
1855 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
1856 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
1857 mptcp_pm_nl_rm_addr_received(msk);
1859 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1860 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1861 mptcp_pm_nl_fully_established(msk);
1863 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1864 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1865 mptcp_pm_nl_subflow_established(msk);
1868 spin_unlock_bh(&msk->pm.lock);
1871 static void __mptcp_close_subflow(struct mptcp_sock *msk)
1873 struct mptcp_subflow_context *subflow, *tmp;
1875 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
1876 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1878 if (inet_sk_state_load(ssk) != TCP_CLOSE)
1881 __mptcp_close_ssk((struct sock *)msk, ssk, subflow);
1885 static bool mptcp_check_close_timeout(const struct sock *sk)
1887 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
1888 struct mptcp_subflow_context *subflow;
1890 if (delta >= TCP_TIMEWAIT_LEN)
1893 /* if all subflows are in closed status don't bother with additional
1896 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
1897 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
1904 static void mptcp_worker(struct work_struct *work)
1906 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1907 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1908 struct mptcp_sendmsg_info info = {};
1909 struct mptcp_data_frag *dfrag;
1914 state = sk->sk_state;
1915 if (unlikely(state == TCP_CLOSE))
1918 mptcp_clean_una_wakeup(sk);
1919 mptcp_check_data_fin_ack(sk);
1920 __mptcp_flush_join_list(msk);
1921 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
1922 __mptcp_close_subflow(msk);
1924 if (mptcp_send_head(sk))
1925 mptcp_push_pending(sk, 0);
1930 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1931 mptcp_check_for_eof(msk);
1933 mptcp_check_data_fin(sk);
1935 /* if the msk data is completely acked, or the socket timedout,
1936 * there is no point in keeping around an orphaned sk
1938 if (sock_flag(sk, SOCK_DEAD) &&
1939 (mptcp_check_close_timeout(sk) ||
1940 (state != sk->sk_state &&
1941 ((1 << inet_sk_state_load(sk)) & (TCPF_CLOSE | TCPF_FIN_WAIT2))))) {
1942 inet_sk_state_store(sk, TCP_CLOSE);
1943 __mptcp_destroy_sock(sk);
1947 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1950 dfrag = mptcp_rtx_head(sk);
1954 if (!mptcp_ext_cache_refill(msk))
1957 ssk = mptcp_subflow_get_retrans(msk);
1963 /* limit retransmission to the bytes already sent on some subflows */
1965 info.limit = dfrag->already_sent;
1966 while (info.sent < dfrag->already_sent) {
1967 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1971 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1975 if (!mptcp_ext_cache_refill(msk))
1979 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1982 mptcp_set_timeout(sk, ssk);
1986 if (!mptcp_timer_pending(sk))
1987 mptcp_reset_timer(sk);
1994 static int __mptcp_init_sock(struct sock *sk)
1996 struct mptcp_sock *msk = mptcp_sk(sk);
1998 spin_lock_init(&msk->join_list_lock);
2000 INIT_LIST_HEAD(&msk->conn_list);
2001 INIT_LIST_HEAD(&msk->join_list);
2002 INIT_LIST_HEAD(&msk->rtx_queue);
2003 INIT_WORK(&msk->work, mptcp_worker);
2004 msk->out_of_order_queue = RB_ROOT;
2005 msk->first_pending = NULL;
2007 msk->ack_hint = NULL;
2009 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2011 mptcp_pm_data_init(msk);
2013 /* re-use the csk retrans timer for MPTCP-level retrans */
2014 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2015 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2019 static int mptcp_init_sock(struct sock *sk)
2021 struct net *net = sock_net(sk);
2024 ret = __mptcp_init_sock(sk);
2028 if (!mptcp_is_enabled(net))
2029 return -ENOPROTOOPT;
2031 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2034 ret = __mptcp_socket_create(mptcp_sk(sk));
2038 sk_sockets_allocated_inc(sk);
2039 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2040 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2045 static void __mptcp_clear_xmit(struct sock *sk)
2047 struct mptcp_sock *msk = mptcp_sk(sk);
2048 struct mptcp_data_frag *dtmp, *dfrag;
2050 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2052 WRITE_ONCE(msk->first_pending, NULL);
2053 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2054 dfrag_clear(sk, dfrag);
2057 static void mptcp_cancel_work(struct sock *sk)
2059 struct mptcp_sock *msk = mptcp_sk(sk);
2061 if (cancel_work_sync(&msk->work))
2065 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2069 switch (ssk->sk_state) {
2071 if (!(how & RCV_SHUTDOWN))
2075 tcp_disconnect(ssk, O_NONBLOCK);
2078 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2079 pr_debug("Fallback");
2080 ssk->sk_shutdown |= how;
2081 tcp_shutdown(ssk, how);
2083 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2084 mptcp_set_timeout(sk, ssk);
2093 static const unsigned char new_state[16] = {
2094 /* current state: new state: action: */
2095 [0 /* (Invalid) */] = TCP_CLOSE,
2096 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2097 [TCP_SYN_SENT] = TCP_CLOSE,
2098 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2099 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2100 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2101 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2102 [TCP_CLOSE] = TCP_CLOSE,
2103 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2104 [TCP_LAST_ACK] = TCP_LAST_ACK,
2105 [TCP_LISTEN] = TCP_CLOSE,
2106 [TCP_CLOSING] = TCP_CLOSING,
2107 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2110 static int mptcp_close_state(struct sock *sk)
2112 int next = (int)new_state[sk->sk_state];
2113 int ns = next & TCP_STATE_MASK;
2115 inet_sk_state_store(sk, ns);
2117 return next & TCP_ACTION_FIN;
2120 static void __mptcp_check_send_data_fin(struct sock *sk)
2122 struct mptcp_subflow_context *subflow;
2123 struct mptcp_sock *msk = mptcp_sk(sk);
2125 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2126 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2127 msk->snd_nxt, msk->write_seq);
2129 /* we still need to enqueue subflows or not really shutting down,
2132 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2133 mptcp_send_head(sk))
2136 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2138 /* fallback socket will not get data_fin/ack, can move to the next
2141 if (__mptcp_check_fallback(msk)) {
2142 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2143 inet_sk_state_store(sk, TCP_CLOSE);
2144 mptcp_close_wake_up(sk);
2145 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2146 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2150 __mptcp_flush_join_list(msk);
2151 mptcp_for_each_subflow(msk, subflow) {
2152 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2154 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2158 static void __mptcp_wr_shutdown(struct sock *sk)
2160 struct mptcp_sock *msk = mptcp_sk(sk);
2162 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2163 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2164 !!mptcp_send_head(sk));
2166 /* will be ignored by fallback sockets */
2167 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2168 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2170 __mptcp_check_send_data_fin(sk);
2173 static void __mptcp_destroy_sock(struct sock *sk)
2175 struct mptcp_subflow_context *subflow, *tmp;
2176 struct mptcp_sock *msk = mptcp_sk(sk);
2177 LIST_HEAD(conn_list);
2179 pr_debug("msk=%p", msk);
2181 /* be sure to always acquire the join list lock, to sync vs
2182 * mptcp_finish_join().
2184 spin_lock_bh(&msk->join_list_lock);
2185 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2186 spin_unlock_bh(&msk->join_list_lock);
2187 list_splice_init(&msk->conn_list, &conn_list);
2189 __mptcp_clear_xmit(sk);
2190 sk_stop_timer(sk, &sk->sk_timer);
2193 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2194 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2195 __mptcp_close_ssk(sk, ssk, subflow);
2198 sk->sk_prot->destroy(sk);
2200 sk_stream_kill_queues(sk);
2201 xfrm_sk_free_policy(sk);
2202 sk_refcnt_debug_release(sk);
2206 static void mptcp_close(struct sock *sk, long timeout)
2208 struct mptcp_subflow_context *subflow;
2209 bool do_cancel_work = false;
2212 sk->sk_shutdown = SHUTDOWN_MASK;
2214 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2215 inet_sk_state_store(sk, TCP_CLOSE);
2219 if (mptcp_close_state(sk))
2220 __mptcp_wr_shutdown(sk);
2222 sk_stream_wait_close(sk, timeout);
2225 /* orphan all the subflows */
2226 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2227 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2228 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2229 bool slow, dispose_socket;
2230 struct socket *sock;
2232 slow = lock_sock_fast(ssk);
2233 sock = ssk->sk_socket;
2234 dispose_socket = sock && sock != sk->sk_socket;
2236 unlock_sock_fast(ssk, slow);
2238 /* for the outgoing subflows we additionally need to free
2239 * the associated socket
2242 iput(SOCK_INODE(sock));
2247 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2248 if (sk->sk_state == TCP_CLOSE) {
2249 __mptcp_destroy_sock(sk);
2250 do_cancel_work = true;
2252 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2256 mptcp_cancel_work(sk);
2260 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2262 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2263 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2264 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2266 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2267 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2270 msk6->saddr = ssk6->saddr;
2271 msk6->flow_label = ssk6->flow_label;
2275 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2276 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2277 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2278 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2279 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2280 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2283 static int mptcp_disconnect(struct sock *sk, int flags)
2285 /* Should never be called.
2286 * inet_stream_connect() calls ->disconnect, but that
2287 * refers to the subflow socket, not the mptcp one.
2293 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2294 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2296 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2298 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2302 struct sock *mptcp_sk_clone(const struct sock *sk,
2303 const struct mptcp_options_received *mp_opt,
2304 struct request_sock *req)
2306 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2307 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2308 struct mptcp_sock *msk;
2314 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2315 if (nsk->sk_family == AF_INET6)
2316 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2319 __mptcp_init_sock(nsk);
2321 msk = mptcp_sk(nsk);
2322 msk->local_key = subflow_req->local_key;
2323 msk->token = subflow_req->token;
2324 msk->subflow = NULL;
2325 WRITE_ONCE(msk->fully_established, false);
2327 msk->write_seq = subflow_req->idsn + 1;
2328 msk->snd_nxt = msk->write_seq;
2329 atomic64_set(&msk->snd_una, msk->write_seq);
2330 atomic64_set(&msk->wnd_end, msk->snd_nxt + req->rsk_rcv_wnd);
2332 if (mp_opt->mp_capable) {
2333 msk->can_ack = true;
2334 msk->remote_key = mp_opt->sndr_key;
2335 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2337 WRITE_ONCE(msk->ack_seq, ack_seq);
2338 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2341 sock_reset_flag(nsk, SOCK_RCU_FREE);
2342 /* will be fully established after successful MPC subflow creation */
2343 inet_sk_state_store(nsk, TCP_SYN_RECV);
2344 bh_unlock_sock(nsk);
2346 /* keep a single reference */
2351 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2353 const struct tcp_sock *tp = tcp_sk(ssk);
2355 msk->rcvq_space.copied = 0;
2356 msk->rcvq_space.rtt_us = 0;
2358 msk->rcvq_space.time = tp->tcp_mstamp;
2360 /* initial rcv_space offering made to peer */
2361 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2362 TCP_INIT_CWND * tp->advmss);
2363 if (msk->rcvq_space.space == 0)
2364 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2366 atomic64_set(&msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2369 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2372 struct mptcp_sock *msk = mptcp_sk(sk);
2373 struct socket *listener;
2376 listener = __mptcp_nmpc_socket(msk);
2377 if (WARN_ON_ONCE(!listener)) {
2382 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2383 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2387 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2388 if (sk_is_mptcp(newsk)) {
2389 struct mptcp_subflow_context *subflow;
2390 struct sock *new_mptcp_sock;
2392 subflow = mptcp_subflow_ctx(newsk);
2393 new_mptcp_sock = subflow->conn;
2395 /* is_mptcp should be false if subflow->conn is missing, see
2396 * subflow_syn_recv_sock()
2398 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2399 tcp_sk(newsk)->is_mptcp = 0;
2403 /* acquire the 2nd reference for the owning socket */
2404 sock_hold(new_mptcp_sock);
2405 newsk = new_mptcp_sock;
2406 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2408 MPTCP_INC_STATS(sock_net(sk),
2409 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2415 void mptcp_destroy_common(struct mptcp_sock *msk)
2417 skb_rbtree_purge(&msk->out_of_order_queue);
2418 mptcp_token_destroy(msk);
2419 mptcp_pm_free_anno_list(msk);
2422 static void mptcp_destroy(struct sock *sk)
2424 struct mptcp_sock *msk = mptcp_sk(sk);
2426 if (msk->cached_ext)
2427 __skb_ext_put(msk->cached_ext);
2429 mptcp_destroy_common(msk);
2430 sk_sockets_allocated_dec(sk);
2433 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2434 sockptr_t optval, unsigned int optlen)
2436 struct sock *sk = (struct sock *)msk;
2437 struct socket *ssock;
2444 ssock = __mptcp_nmpc_socket(msk);
2450 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2452 if (optname == SO_REUSEPORT)
2453 sk->sk_reuseport = ssock->sk->sk_reuseport;
2454 else if (optname == SO_REUSEADDR)
2455 sk->sk_reuse = ssock->sk->sk_reuse;
2461 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2464 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2465 sockptr_t optval, unsigned int optlen)
2467 struct sock *sk = (struct sock *)msk;
2468 int ret = -EOPNOTSUPP;
2469 struct socket *ssock;
2474 ssock = __mptcp_nmpc_socket(msk);
2480 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2482 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2491 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2492 sockptr_t optval, unsigned int optlen)
2494 struct mptcp_sock *msk = mptcp_sk(sk);
2497 pr_debug("msk=%p", msk);
2499 if (level == SOL_SOCKET)
2500 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2502 /* @@ the meaning of setsockopt() when the socket is connected and
2503 * there are multiple subflows is not yet defined. It is up to the
2504 * MPTCP-level socket to configure the subflows until the subflow
2505 * is in TCP fallback, when TCP socket options are passed through
2506 * to the one remaining subflow.
2509 ssk = __mptcp_tcp_fallback(msk);
2512 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2514 if (level == SOL_IPV6)
2515 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2520 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2521 char __user *optval, int __user *option)
2523 struct mptcp_sock *msk = mptcp_sk(sk);
2526 pr_debug("msk=%p", msk);
2528 /* @@ the meaning of setsockopt() when the socket is connected and
2529 * there are multiple subflows is not yet defined. It is up to the
2530 * MPTCP-level socket to configure the subflows until the subflow
2531 * is in TCP fallback, when socket options are passed through
2532 * to the one remaining subflow.
2535 ssk = __mptcp_tcp_fallback(msk);
2538 return tcp_getsockopt(ssk, level, optname, optval, option);
2543 #define MPTCP_DEFERRED_ALL (TCPF_WRITE_TIMER_DEFERRED)
2545 /* this is very alike tcp_release_cb() but we must handle differently a
2546 * different set of events
2548 static void mptcp_release_cb(struct sock *sk)
2550 unsigned long flags, nflags;
2553 flags = sk->sk_tsq_flags;
2554 if (!(flags & MPTCP_DEFERRED_ALL))
2556 nflags = flags & ~MPTCP_DEFERRED_ALL;
2557 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2559 sock_release_ownership(sk);
2561 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2562 mptcp_retransmit_handler(sk);
2567 static int mptcp_hash(struct sock *sk)
2569 /* should never be called,
2570 * we hash the TCP subflows not the master socket
2576 static void mptcp_unhash(struct sock *sk)
2578 /* called from sk_common_release(), but nothing to do here */
2581 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2583 struct mptcp_sock *msk = mptcp_sk(sk);
2584 struct socket *ssock;
2586 ssock = __mptcp_nmpc_socket(msk);
2587 pr_debug("msk=%p, subflow=%p", msk, ssock);
2588 if (WARN_ON_ONCE(!ssock))
2591 return inet_csk_get_port(ssock->sk, snum);
2594 void mptcp_finish_connect(struct sock *ssk)
2596 struct mptcp_subflow_context *subflow;
2597 struct mptcp_sock *msk;
2601 subflow = mptcp_subflow_ctx(ssk);
2605 pr_debug("msk=%p, token=%u", sk, subflow->token);
2607 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2609 subflow->map_seq = ack_seq;
2610 subflow->map_subflow_seq = 1;
2612 /* the socket is not connected yet, no msk/subflow ops can access/race
2613 * accessing the field below
2615 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2616 WRITE_ONCE(msk->local_key, subflow->local_key);
2617 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2618 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2619 WRITE_ONCE(msk->ack_seq, ack_seq);
2620 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2621 WRITE_ONCE(msk->can_ack, 1);
2622 atomic64_set(&msk->snd_una, msk->write_seq);
2624 mptcp_pm_new_connection(msk, 0);
2626 mptcp_rcv_space_init(msk, ssk);
2629 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2631 write_lock_bh(&sk->sk_callback_lock);
2632 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2633 sk_set_socket(sk, parent);
2634 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2635 write_unlock_bh(&sk->sk_callback_lock);
2638 bool mptcp_finish_join(struct sock *ssk)
2640 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2641 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2642 struct sock *parent = (void *)msk;
2643 struct socket *parent_sock;
2646 pr_debug("msk=%p, subflow=%p", msk, subflow);
2648 /* mptcp socket already closing? */
2649 if (!mptcp_is_fully_established(parent))
2652 if (!msk->pm.server_side)
2655 if (!mptcp_pm_allow_new_subflow(msk))
2658 /* active connections are already on conn_list, and we can't acquire
2660 * use the join list lock as synchronization point and double-check
2661 * msk status to avoid racing with __mptcp_destroy_sock()
2663 spin_lock_bh(&msk->join_list_lock);
2664 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2665 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
2666 list_add_tail(&subflow->node, &msk->join_list);
2669 spin_unlock_bh(&msk->join_list_lock);
2673 /* attach to msk socket only after we are sure he will deal with us
2676 parent_sock = READ_ONCE(parent->sk_socket);
2677 if (parent_sock && !ssk->sk_socket)
2678 mptcp_sock_graft(ssk, parent_sock);
2679 subflow->map_seq = READ_ONCE(msk->ack_seq);
2683 static struct proto mptcp_prot = {
2685 .owner = THIS_MODULE,
2686 .init = mptcp_init_sock,
2687 .disconnect = mptcp_disconnect,
2688 .close = mptcp_close,
2689 .accept = mptcp_accept,
2690 .setsockopt = mptcp_setsockopt,
2691 .getsockopt = mptcp_getsockopt,
2692 .shutdown = tcp_shutdown,
2693 .destroy = mptcp_destroy,
2694 .sendmsg = mptcp_sendmsg,
2695 .recvmsg = mptcp_recvmsg,
2696 .release_cb = mptcp_release_cb,
2698 .unhash = mptcp_unhash,
2699 .get_port = mptcp_get_port,
2700 .sockets_allocated = &mptcp_sockets_allocated,
2701 .memory_allocated = &tcp_memory_allocated,
2702 .memory_pressure = &tcp_memory_pressure,
2703 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2704 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
2705 .sysctl_mem = sysctl_tcp_mem,
2706 .obj_size = sizeof(struct mptcp_sock),
2707 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2708 .no_autobind = true,
2711 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2713 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2714 struct socket *ssock;
2717 lock_sock(sock->sk);
2718 ssock = __mptcp_nmpc_socket(msk);
2724 err = ssock->ops->bind(ssock, uaddr, addr_len);
2726 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2729 release_sock(sock->sk);
2733 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2734 struct mptcp_subflow_context *subflow)
2736 subflow->request_mptcp = 0;
2737 __mptcp_do_fallback(msk);
2740 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2741 int addr_len, int flags)
2743 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2744 struct mptcp_subflow_context *subflow;
2745 struct socket *ssock;
2748 lock_sock(sock->sk);
2749 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2750 /* pending connection or invalid state, let existing subflow
2753 ssock = msk->subflow;
2757 ssock = __mptcp_nmpc_socket(msk);
2763 mptcp_token_destroy(msk);
2764 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2765 subflow = mptcp_subflow_ctx(ssock->sk);
2766 #ifdef CONFIG_TCP_MD5SIG
2767 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2770 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2771 mptcp_subflow_early_fallback(msk, subflow);
2773 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2774 mptcp_subflow_early_fallback(msk, subflow);
2777 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2778 sock->state = ssock->state;
2780 /* on successful connect, the msk state will be moved to established by
2781 * subflow_finish_connect()
2783 if (!err || err == -EINPROGRESS)
2784 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2786 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2789 release_sock(sock->sk);
2793 static int mptcp_listen(struct socket *sock, int backlog)
2795 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2796 struct socket *ssock;
2799 pr_debug("msk=%p", msk);
2801 lock_sock(sock->sk);
2802 ssock = __mptcp_nmpc_socket(msk);
2808 mptcp_token_destroy(msk);
2809 inet_sk_state_store(sock->sk, TCP_LISTEN);
2810 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2812 err = ssock->ops->listen(ssock, backlog);
2813 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2815 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2818 release_sock(sock->sk);
2822 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2823 int flags, bool kern)
2825 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2826 struct socket *ssock;
2829 pr_debug("msk=%p", msk);
2831 lock_sock(sock->sk);
2832 if (sock->sk->sk_state != TCP_LISTEN)
2835 ssock = __mptcp_nmpc_socket(msk);
2839 clear_bit(MPTCP_DATA_READY, &msk->flags);
2840 sock_hold(ssock->sk);
2841 release_sock(sock->sk);
2843 err = ssock->ops->accept(sock, newsock, flags, kern);
2844 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2845 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2846 struct mptcp_subflow_context *subflow;
2847 struct sock *newsk = newsock->sk;
2850 slowpath = lock_sock_fast(newsk);
2851 mptcp_copy_inaddrs(newsk, msk->first);
2852 mptcp_rcv_space_init(msk, msk->first);
2854 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2855 * This is needed so NOSPACE flag can be set from tcp stack.
2857 __mptcp_flush_join_list(msk);
2858 mptcp_for_each_subflow(msk, subflow) {
2859 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2861 if (!ssk->sk_socket)
2862 mptcp_sock_graft(ssk, newsock);
2864 unlock_sock_fast(newsk, slowpath);
2867 if (inet_csk_listen_poll(ssock->sk))
2868 set_bit(MPTCP_DATA_READY, &msk->flags);
2869 sock_put(ssock->sk);
2873 release_sock(sock->sk);
2877 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2879 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2883 static bool __mptcp_check_writeable(struct mptcp_sock *msk)
2885 struct sock *sk = (struct sock *)msk;
2886 bool mptcp_writable;
2888 mptcp_clean_una(sk);
2889 mptcp_writable = sk_stream_is_writeable(sk);
2890 if (!mptcp_writable)
2893 return mptcp_writable;
2896 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
2898 struct sock *sk = (struct sock *)msk;
2902 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
2905 if (sk_stream_is_writeable(sk))
2906 return EPOLLOUT | EPOLLWRNORM;
2908 slow = lock_sock_fast(sk);
2909 if (__mptcp_check_writeable(msk))
2910 ret = EPOLLOUT | EPOLLWRNORM;
2912 unlock_sock_fast(sk, slow);
2916 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2917 struct poll_table_struct *wait)
2919 struct sock *sk = sock->sk;
2920 struct mptcp_sock *msk;
2925 sock_poll_wait(file, sock, wait);
2927 state = inet_sk_state_load(sk);
2928 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
2929 if (state == TCP_LISTEN)
2930 return mptcp_check_readable(msk);
2932 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2933 mask |= mptcp_check_readable(msk);
2934 mask |= mptcp_check_writeable(msk);
2936 if (sk->sk_shutdown & RCV_SHUTDOWN)
2937 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2942 static int mptcp_shutdown(struct socket *sock, int how)
2944 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2945 struct sock *sk = sock->sk;
2948 pr_debug("sk=%p, how=%d", msk, how);
2953 if ((how & ~SHUTDOWN_MASK) || !how) {
2958 if (sock->state == SS_CONNECTING) {
2959 if ((1 << sk->sk_state) &
2960 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2961 sock->state = SS_DISCONNECTING;
2963 sock->state = SS_CONNECTED;
2966 sk->sk_shutdown |= how;
2967 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
2968 __mptcp_wr_shutdown(sk);
2970 /* Wake up anyone sleeping in poll. */
2971 sk->sk_state_change(sk);
2979 static const struct proto_ops mptcp_stream_ops = {
2981 .owner = THIS_MODULE,
2982 .release = inet_release,
2984 .connect = mptcp_stream_connect,
2985 .socketpair = sock_no_socketpair,
2986 .accept = mptcp_stream_accept,
2987 .getname = inet_getname,
2989 .ioctl = inet_ioctl,
2990 .gettstamp = sock_gettstamp,
2991 .listen = mptcp_listen,
2992 .shutdown = mptcp_shutdown,
2993 .setsockopt = sock_common_setsockopt,
2994 .getsockopt = sock_common_getsockopt,
2995 .sendmsg = inet_sendmsg,
2996 .recvmsg = inet_recvmsg,
2997 .mmap = sock_no_mmap,
2998 .sendpage = inet_sendpage,
3001 static struct inet_protosw mptcp_protosw = {
3002 .type = SOCK_STREAM,
3003 .protocol = IPPROTO_MPTCP,
3004 .prot = &mptcp_prot,
3005 .ops = &mptcp_stream_ops,
3006 .flags = INET_PROTOSW_ICSK,
3009 void __init mptcp_proto_init(void)
3011 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3013 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3014 panic("Failed to allocate MPTCP pcpu counter\n");
3016 mptcp_subflow_init();
3020 if (proto_register(&mptcp_prot, 1) != 0)
3021 panic("Failed to register MPTCP proto.\n");
3023 inet_register_protosw(&mptcp_protosw);
3025 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3028 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3029 static const struct proto_ops mptcp_v6_stream_ops = {
3031 .owner = THIS_MODULE,
3032 .release = inet6_release,
3034 .connect = mptcp_stream_connect,
3035 .socketpair = sock_no_socketpair,
3036 .accept = mptcp_stream_accept,
3037 .getname = inet6_getname,
3039 .ioctl = inet6_ioctl,
3040 .gettstamp = sock_gettstamp,
3041 .listen = mptcp_listen,
3042 .shutdown = mptcp_shutdown,
3043 .setsockopt = sock_common_setsockopt,
3044 .getsockopt = sock_common_getsockopt,
3045 .sendmsg = inet6_sendmsg,
3046 .recvmsg = inet6_recvmsg,
3047 .mmap = sock_no_mmap,
3048 .sendpage = inet_sendpage,
3049 #ifdef CONFIG_COMPAT
3050 .compat_ioctl = inet6_compat_ioctl,
3054 static struct proto mptcp_v6_prot;
3056 static void mptcp_v6_destroy(struct sock *sk)
3059 inet6_destroy_sock(sk);
3062 static struct inet_protosw mptcp_v6_protosw = {
3063 .type = SOCK_STREAM,
3064 .protocol = IPPROTO_MPTCP,
3065 .prot = &mptcp_v6_prot,
3066 .ops = &mptcp_v6_stream_ops,
3067 .flags = INET_PROTOSW_ICSK,
3070 int __init mptcp_proto_v6_init(void)
3074 mptcp_v6_prot = mptcp_prot;
3075 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3076 mptcp_v6_prot.slab = NULL;
3077 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3078 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3080 err = proto_register(&mptcp_v6_prot, 1);
3084 err = inet6_register_protosw(&mptcp_v6_protosw);
3086 proto_unregister(&mptcp_v6_prot);