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;
173 seq = MPTCP_SKB_CB(skb)->map_seq;
174 end_seq = MPTCP_SKB_CB(skb)->end_seq;
175 space = tcp_space(sk);
176 max_seq = space > 0 ? space + msk->ack_seq : msk->ack_seq;
178 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
179 RB_EMPTY_ROOT(&msk->out_of_order_queue));
180 if (after64(seq, max_seq)) {
183 pr_debug("oow by %ld", (unsigned long)seq - (unsigned long)max_seq);
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 void mptcp_send_ack(struct mptcp_sock *msk)
412 struct mptcp_subflow_context *subflow;
414 mptcp_for_each_subflow(msk, subflow) {
415 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
423 static bool mptcp_check_data_fin(struct sock *sk)
425 struct mptcp_sock *msk = mptcp_sk(sk);
426 u64 rcv_data_fin_seq;
429 if (__mptcp_check_fallback(msk) || !msk->first)
432 /* Need to ack a DATA_FIN received from a peer while this side
433 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
434 * msk->rcv_data_fin was set when parsing the incoming options
435 * at the subflow level and the msk lock was not held, so this
436 * is the first opportunity to act on the DATA_FIN and change
439 * If we are caught up to the sequence number of the incoming
440 * DATA_FIN, send the DATA_ACK now and do state transition. If
441 * not caught up, do nothing and let the recv code send DATA_ACK
445 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
446 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
447 WRITE_ONCE(msk->rcv_data_fin, 0);
449 sk->sk_shutdown |= RCV_SHUTDOWN;
450 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
451 set_bit(MPTCP_DATA_READY, &msk->flags);
453 switch (sk->sk_state) {
454 case TCP_ESTABLISHED:
455 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
458 inet_sk_state_store(sk, TCP_CLOSING);
461 inet_sk_state_store(sk, TCP_CLOSE);
464 /* Other states not expected */
470 mptcp_set_timeout(sk, NULL);
472 mptcp_close_wake_up(sk);
477 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
481 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
482 struct sock *sk = (struct sock *)msk;
483 unsigned int moved = 0;
484 bool more_data_avail;
490 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
492 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
493 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
495 if (unlikely(ssk_rbuf > sk_rbuf)) {
496 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
501 pr_debug("msk=%p ssk=%p", msk, ssk);
503 old_copied_seq = tp->copied_seq;
505 u32 map_remaining, offset;
506 u32 seq = tp->copied_seq;
510 /* try to move as much data as available */
511 map_remaining = subflow->map_data_len -
512 mptcp_subflow_get_map_offset(subflow);
514 skb = skb_peek(&ssk->sk_receive_queue);
516 /* if no data is found, a racing workqueue/recvmsg
517 * already processed the new data, stop here or we
518 * can enter an infinite loop
525 if (__mptcp_check_fallback(msk)) {
526 /* if we are running under the workqueue, TCP could have
527 * collapsed skbs between dummy map creation and now
528 * be sure to adjust the size
530 map_remaining = skb->len;
531 subflow->map_data_len = skb->len;
534 offset = seq - TCP_SKB_CB(skb)->seq;
535 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
541 if (offset < skb->len) {
542 size_t len = skb->len - offset;
547 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
551 if (WARN_ON_ONCE(map_remaining < len))
555 sk_eat_skb(ssk, skb);
559 WRITE_ONCE(tp->copied_seq, seq);
560 more_data_avail = mptcp_subflow_data_available(ssk);
562 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
566 } while (more_data_avail);
569 if (tp->copied_seq != old_copied_seq)
570 tcp_cleanup_rbuf(ssk, 1);
575 static bool mptcp_ofo_queue(struct mptcp_sock *msk)
577 struct sock *sk = (struct sock *)msk;
578 struct sk_buff *skb, *tail;
583 p = rb_first(&msk->out_of_order_queue);
584 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
587 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
591 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
593 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
596 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
600 end_seq = MPTCP_SKB_CB(skb)->end_seq;
601 tail = skb_peek_tail(&sk->sk_receive_queue);
602 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
603 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
605 /* skip overlapping data, if any */
606 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
607 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
609 MPTCP_SKB_CB(skb)->offset += delta;
610 __skb_queue_tail(&sk->sk_receive_queue, skb);
612 msk->ack_seq = end_seq;
618 /* In most cases we will be able to lock the mptcp socket. If its already
619 * owned, we need to defer to the work queue to avoid ABBA deadlock.
621 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
623 struct sock *sk = (struct sock *)msk;
624 unsigned int moved = 0;
626 if (READ_ONCE(sk->sk_lock.owned))
629 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
632 /* must re-check after taking the lock */
633 if (!READ_ONCE(sk->sk_lock.owned)) {
634 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
635 mptcp_ofo_queue(msk);
637 /* If the moves have caught up with the DATA_FIN sequence number
638 * it's time to ack the DATA_FIN and change socket state, but
639 * this is not a good place to change state. Let the workqueue
642 if (mptcp_pending_data_fin(sk, NULL))
643 mptcp_schedule_work(sk);
646 spin_unlock_bh(&sk->sk_lock.slock);
651 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
653 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
654 struct mptcp_sock *msk = mptcp_sk(sk);
655 int sk_rbuf, ssk_rbuf;
658 /* move_skbs_to_msk below can legitly clear the data_avail flag,
659 * but we will need later to properly woke the reader, cache its
662 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
664 set_bit(MPTCP_DATA_READY, &msk->flags);
666 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
667 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
668 if (unlikely(ssk_rbuf > sk_rbuf))
671 /* over limit? can't append more skbs to msk */
672 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
675 if (move_skbs_to_msk(msk, ssk))
678 /* mptcp socket is owned, release_cb should retry */
679 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
680 &sk->sk_tsq_flags)) {
683 /* need to try again, its possible release_cb() has already
684 * been called after the test_and_set_bit() above.
686 move_skbs_to_msk(msk, ssk);
690 sk->sk_data_ready(sk);
693 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
695 if (likely(list_empty(&msk->join_list)))
698 spin_lock_bh(&msk->join_list_lock);
699 list_splice_tail_init(&msk->join_list, &msk->conn_list);
700 spin_unlock_bh(&msk->join_list_lock);
703 static bool mptcp_timer_pending(struct sock *sk)
705 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
708 static void mptcp_reset_timer(struct sock *sk)
710 struct inet_connection_sock *icsk = inet_csk(sk);
713 /* prevent rescheduling on close */
714 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
717 /* should never be called with mptcp level timer cleared */
718 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
719 if (WARN_ON_ONCE(!tout))
721 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
724 bool mptcp_schedule_work(struct sock *sk)
726 if (inet_sk_state_load(sk) != TCP_CLOSE &&
727 schedule_work(&mptcp_sk(sk)->work)) {
728 /* each subflow already holds a reference to the sk, and the
729 * workqueue is invoked by a subflow, so sk can't go away here.
737 void mptcp_data_acked(struct sock *sk)
739 mptcp_reset_timer(sk);
741 if ((test_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags) ||
742 mptcp_send_head(sk) ||
743 (inet_sk_state_load(sk) != TCP_ESTABLISHED)))
744 mptcp_schedule_work(sk);
747 void mptcp_subflow_eof(struct sock *sk)
749 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
750 mptcp_schedule_work(sk);
753 static void mptcp_check_for_eof(struct mptcp_sock *msk)
755 struct mptcp_subflow_context *subflow;
756 struct sock *sk = (struct sock *)msk;
759 mptcp_for_each_subflow(msk, subflow)
760 receivers += !subflow->rx_eof;
764 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
765 /* hopefully temporary hack: propagate shutdown status
766 * to msk, when all subflows agree on it
768 sk->sk_shutdown |= RCV_SHUTDOWN;
770 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
771 set_bit(MPTCP_DATA_READY, &msk->flags);
772 sk->sk_data_ready(sk);
775 switch (sk->sk_state) {
776 case TCP_ESTABLISHED:
777 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
780 /* fallback sockets skip TCP_CLOSING - TCP will take care */
781 inet_sk_state_store(sk, TCP_CLOSE);
786 mptcp_close_wake_up(sk);
789 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
791 const struct sock *sk = (const struct sock *)msk;
793 if (!msk->cached_ext)
794 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
796 return !!msk->cached_ext;
799 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
801 struct mptcp_subflow_context *subflow;
802 struct sock *sk = (struct sock *)msk;
804 sock_owned_by_me(sk);
806 mptcp_for_each_subflow(msk, subflow) {
807 if (subflow->data_avail)
808 return mptcp_subflow_tcp_sock(subflow);
814 static bool mptcp_skb_can_collapse_to(u64 write_seq,
815 const struct sk_buff *skb,
816 const struct mptcp_ext *mpext)
818 if (!tcp_skb_can_collapse_to(skb))
821 /* can collapse only if MPTCP level sequence is in order and this
822 * mapping has not been xmitted yet
824 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
828 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
829 const struct page_frag *pfrag,
830 const struct mptcp_data_frag *df)
832 return df && pfrag->page == df->page &&
833 pfrag->size - pfrag->offset > 0 &&
834 df->data_seq + df->data_len == msk->write_seq;
837 static void dfrag_uncharge(struct sock *sk, int len)
839 sk_mem_uncharge(sk, len);
840 sk_wmem_queued_add(sk, -len);
843 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
845 int len = dfrag->data_len + dfrag->overhead;
847 list_del(&dfrag->list);
848 dfrag_uncharge(sk, len);
849 put_page(dfrag->page);
852 static void mptcp_clean_una(struct sock *sk)
854 struct mptcp_sock *msk = mptcp_sk(sk);
855 struct mptcp_data_frag *dtmp, *dfrag;
856 bool cleaned = false;
859 /* on fallback we just need to ignore snd_una, as this is really
862 if (__mptcp_check_fallback(msk))
863 atomic64_set(&msk->snd_una, msk->snd_nxt);
865 snd_una = atomic64_read(&msk->snd_una);
867 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
868 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
871 if (WARN_ON_ONCE(dfrag == msk->first_pending))
873 dfrag_clear(sk, dfrag);
877 dfrag = mptcp_rtx_head(sk);
878 if (dfrag && after64(snd_una, dfrag->data_seq)) {
879 u64 delta = snd_una - dfrag->data_seq;
881 if (WARN_ON_ONCE(delta > dfrag->already_sent))
884 dfrag->data_seq += delta;
885 dfrag->offset += delta;
886 dfrag->data_len -= delta;
887 dfrag->already_sent -= delta;
889 dfrag_uncharge(sk, delta);
895 sk_mem_reclaim_partial(sk);
898 static void mptcp_clean_una_wakeup(struct sock *sk)
900 struct mptcp_sock *msk = mptcp_sk(sk);
904 /* Only wake up writers if a subflow is ready */
905 if (sk_stream_is_writeable(sk)) {
906 clear_bit(MPTCP_NOSPACE, &msk->flags);
907 sk_stream_write_space(sk);
911 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
914 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
916 struct mptcp_subflow_context *subflow;
917 struct mptcp_sock *msk = mptcp_sk(sk);
920 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
921 pfrag, sk->sk_allocation)))
924 sk_stream_moderate_sndbuf(sk);
925 mptcp_for_each_subflow(msk, subflow) {
926 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
929 tcp_enter_memory_pressure(ssk);
930 sk_stream_moderate_sndbuf(ssk);
936 static struct mptcp_data_frag *
937 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
940 int offset = ALIGN(orig_offset, sizeof(long));
941 struct mptcp_data_frag *dfrag;
943 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
945 dfrag->data_seq = msk->write_seq;
946 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
947 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
948 dfrag->already_sent = 0;
949 dfrag->page = pfrag->page;
954 struct mptcp_sendmsg_info {
962 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
965 u64 window_end = mptcp_wnd_end(msk);
967 if (__mptcp_check_fallback(msk))
970 if (!before64(data_seq + avail_size, window_end)) {
971 u64 allowed_size = window_end - data_seq;
973 return min_t(unsigned int, allowed_size, avail_size);
979 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
980 struct mptcp_data_frag *dfrag,
981 struct mptcp_sendmsg_info *info)
983 u64 data_seq = dfrag->data_seq + info->sent;
984 struct mptcp_sock *msk = mptcp_sk(sk);
985 bool zero_window_probe = false;
986 struct mptcp_ext *mpext = NULL;
987 struct sk_buff *skb, *tail;
988 bool can_collapse = false;
992 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
993 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
995 /* compute send limit */
996 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
997 avail_size = info->size_goal;
998 skb = tcp_write_queue_tail(ssk);
1000 /* Limit the write to the size available in the
1001 * current skb, if any, so that we create at most a new skb.
1002 * Explicitly tells TCP internals to avoid collapsing on later
1003 * queue management operation, to avoid breaking the ext <->
1004 * SSN association set here
1006 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1007 can_collapse = (info->size_goal - skb->len > 0) &&
1008 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1010 TCP_SKB_CB(skb)->eor = 1;
1012 avail_size = info->size_goal - skb->len;
1015 /* Zero window and all data acked? Probe. */
1016 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1017 if (avail_size == 0) {
1018 if (skb || atomic64_read(&msk->snd_una) != msk->snd_nxt)
1020 zero_window_probe = true;
1021 data_seq = atomic64_read(&msk->snd_una) - 1;
1025 if (WARN_ON_ONCE(info->sent > info->limit ||
1026 info->limit > dfrag->data_len))
1029 ret = info->limit - info->sent;
1030 tail = tcp_build_frag(ssk, avail_size, info->flags, dfrag->page,
1031 dfrag->offset + info->sent, &ret);
1033 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1037 /* if the tail skb is still the cached one, collapsing really happened.
1040 WARN_ON_ONCE(!can_collapse);
1041 mpext->data_len += ret;
1042 WARN_ON_ONCE(zero_window_probe);
1046 mpext = __skb_ext_set(tail, SKB_EXT_MPTCP, msk->cached_ext);
1047 msk->cached_ext = NULL;
1049 memset(mpext, 0, sizeof(*mpext));
1050 mpext->data_seq = data_seq;
1051 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1052 mpext->data_len = ret;
1056 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1057 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1060 if (zero_window_probe) {
1061 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1064 tcp_push_pending_frames(ssk);
1067 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1071 static void mptcp_nospace(struct mptcp_sock *msk)
1073 struct mptcp_subflow_context *subflow;
1075 set_bit(MPTCP_NOSPACE, &msk->flags);
1076 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
1078 mptcp_for_each_subflow(msk, subflow) {
1079 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1080 bool ssk_writeable = sk_stream_is_writeable(ssk);
1081 struct socket *sock = READ_ONCE(ssk->sk_socket);
1083 if (ssk_writeable || !sock)
1086 /* enables ssk->write_space() callbacks */
1087 set_bit(SOCK_NOSPACE, &sock->flags);
1090 /* mptcp_data_acked() could run just before we set the NOSPACE bit,
1091 * so explicitly check for snd_una value
1093 mptcp_clean_una((struct sock *)msk);
1096 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1098 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1100 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
1101 if (subflow->request_join && !subflow->fully_established)
1104 /* only send if our side has not closed yet */
1105 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
1108 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1109 sizeof(struct tcphdr) - \
1110 MAX_TCP_OPTION_SPACE - \
1111 sizeof(struct ipv6hdr) - \
1112 sizeof(struct frag_hdr))
1114 struct subflow_send_info {
1119 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk,
1122 struct subflow_send_info send_info[2];
1123 struct mptcp_subflow_context *subflow;
1124 int i, nr_active = 0;
1129 sock_owned_by_me((struct sock *)msk);
1132 if (!mptcp_ext_cache_refill(msk))
1135 if (__mptcp_check_fallback(msk)) {
1138 *sndbuf = msk->first->sk_sndbuf;
1139 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1142 /* re-use last subflow, if the burst allow that */
1143 if (msk->last_snd && msk->snd_burst > 0 &&
1144 sk_stream_memory_free(msk->last_snd) &&
1145 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1146 mptcp_for_each_subflow(msk, subflow) {
1147 ssk = mptcp_subflow_tcp_sock(subflow);
1148 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1150 return msk->last_snd;
1153 /* pick the subflow with the lower wmem/wspace ratio */
1154 for (i = 0; i < 2; ++i) {
1155 send_info[i].ssk = NULL;
1156 send_info[i].ratio = -1;
1158 mptcp_for_each_subflow(msk, subflow) {
1159 ssk = mptcp_subflow_tcp_sock(subflow);
1160 if (!mptcp_subflow_active(subflow))
1163 nr_active += !subflow->backup;
1164 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1165 if (!sk_stream_memory_free(subflow->tcp_sock))
1168 pace = READ_ONCE(ssk->sk_pacing_rate);
1172 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1174 if (ratio < send_info[subflow->backup].ratio) {
1175 send_info[subflow->backup].ssk = ssk;
1176 send_info[subflow->backup].ratio = ratio;
1180 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1181 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1182 send_info[1].ssk, send_info[1].ratio);
1184 /* pick the best backup if no other subflow is active */
1186 send_info[0].ssk = send_info[1].ssk;
1188 if (send_info[0].ssk) {
1189 msk->last_snd = send_info[0].ssk;
1190 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1191 sk_stream_wspace(msk->last_snd));
1192 return msk->last_snd;
1197 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1198 struct mptcp_sendmsg_info *info)
1200 mptcp_set_timeout(sk, ssk);
1201 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1205 static void mptcp_push_pending(struct sock *sk, unsigned int flags)
1207 struct sock *prev_ssk = NULL, *ssk = NULL;
1208 struct mptcp_sock *msk = mptcp_sk(sk);
1209 struct mptcp_sendmsg_info info = {
1212 struct mptcp_data_frag *dfrag;
1213 int len, copied = 0;
1216 while ((dfrag = mptcp_send_head(sk))) {
1217 info.sent = dfrag->already_sent;
1218 info.limit = dfrag->data_len;
1219 len = dfrag->data_len - dfrag->already_sent;
1224 __mptcp_flush_join_list(msk);
1225 ssk = mptcp_subflow_get_send(msk, &sndbuf);
1227 /* do auto tuning */
1228 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1229 sndbuf > READ_ONCE(sk->sk_sndbuf))
1230 WRITE_ONCE(sk->sk_sndbuf, sndbuf);
1232 /* try to keep the subflow socket lock across
1233 * consecutive xmit on the same socket
1235 if (ssk != prev_ssk && prev_ssk)
1236 mptcp_push_release(sk, prev_ssk, &info);
1240 if (ssk != prev_ssk || !prev_ssk)
1243 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1245 mptcp_push_release(sk, ssk, &info);
1250 dfrag->already_sent += ret;
1251 msk->snd_nxt += ret;
1252 msk->snd_burst -= ret;
1256 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1259 /* at this point we held the socket lock for the last subflow we used */
1261 mptcp_push_release(sk, ssk, &info);
1265 /* start the timer, if it's not pending */
1266 if (!mptcp_timer_pending(sk))
1267 mptcp_reset_timer(sk);
1268 __mptcp_check_send_data_fin(sk);
1272 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1274 struct mptcp_sock *msk = mptcp_sk(sk);
1275 struct page_frag *pfrag;
1280 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1285 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1287 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1288 ret = sk_stream_wait_connect(sk, &timeo);
1293 pfrag = sk_page_frag(sk);
1294 mptcp_clean_una(sk);
1296 while (msg_data_left(msg)) {
1297 struct mptcp_data_frag *dfrag;
1298 int frag_truesize = 0;
1299 bool dfrag_collapsed;
1300 size_t psize, offset;
1302 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1307 /* reuse tail pfrag, if possible, or carve a new one from the
1310 dfrag = mptcp_pending_tail(sk);
1311 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1312 if (!dfrag_collapsed) {
1313 if (!sk_stream_memory_free(sk)) {
1314 mptcp_push_pending(sk, msg->msg_flags);
1315 if (!sk_stream_memory_free(sk))
1316 goto wait_for_memory;
1318 if (!mptcp_page_frag_refill(sk, pfrag))
1319 goto wait_for_memory;
1321 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1322 frag_truesize = dfrag->overhead;
1325 /* we do not bound vs wspace, to allow a single packet.
1326 * memory accounting will prevent execessive memory usage
1329 offset = dfrag->offset + dfrag->data_len;
1330 psize = pfrag->size - offset;
1331 psize = min_t(size_t, psize, msg_data_left(msg));
1332 if (!sk_wmem_schedule(sk, psize + frag_truesize))
1333 goto wait_for_memory;
1335 if (copy_page_from_iter(dfrag->page, offset, psize,
1336 &msg->msg_iter) != psize) {
1341 /* data successfully copied into the write queue */
1343 dfrag->data_len += psize;
1344 frag_truesize += psize;
1345 pfrag->offset += frag_truesize;
1346 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1348 /* charge data on mptcp pending queue to the msk socket
1349 * Note: we charge such data both to sk and ssk
1351 sk_wmem_queued_add(sk, frag_truesize);
1352 sk->sk_forward_alloc -= frag_truesize;
1353 if (!dfrag_collapsed) {
1354 get_page(dfrag->page);
1355 list_add_tail(&dfrag->list, &msk->rtx_queue);
1356 if (!msk->first_pending)
1357 WRITE_ONCE(msk->first_pending, dfrag);
1359 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1360 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1363 if (!mptcp_ext_cache_refill(msk))
1364 goto wait_for_memory;
1369 if (mptcp_timer_pending(sk))
1370 mptcp_reset_timer(sk);
1371 ret = sk_stream_wait_memory(sk, &timeo);
1377 mptcp_push_pending(sk, msg->msg_flags);
1381 return copied ? : ret;
1384 static void mptcp_wait_data(struct sock *sk, long *timeo)
1386 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1387 struct mptcp_sock *msk = mptcp_sk(sk);
1389 add_wait_queue(sk_sleep(sk), &wait);
1390 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1392 sk_wait_event(sk, timeo,
1393 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1395 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1396 remove_wait_queue(sk_sleep(sk), &wait);
1399 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1403 struct sock *sk = (struct sock *)msk;
1404 struct sk_buff *skb;
1407 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1408 u32 offset = MPTCP_SKB_CB(skb)->offset;
1409 u32 data_len = skb->len - offset;
1410 u32 count = min_t(size_t, len - copied, data_len);
1413 err = skb_copy_datagram_msg(skb, offset, msg, count);
1414 if (unlikely(err < 0)) {
1422 if (count < data_len) {
1423 MPTCP_SKB_CB(skb)->offset += count;
1427 __skb_unlink(skb, &sk->sk_receive_queue);
1437 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1439 * Only difference: Use highest rtt estimate of the subflows in use.
1441 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1443 struct mptcp_subflow_context *subflow;
1444 struct sock *sk = (struct sock *)msk;
1445 u32 time, advmss = 1;
1448 sock_owned_by_me(sk);
1453 msk->rcvq_space.copied += copied;
1455 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1456 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1458 rtt_us = msk->rcvq_space.rtt_us;
1459 if (rtt_us && time < (rtt_us >> 3))
1463 mptcp_for_each_subflow(msk, subflow) {
1464 const struct tcp_sock *tp;
1468 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1470 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1471 sf_advmss = READ_ONCE(tp->advmss);
1473 rtt_us = max(sf_rtt_us, rtt_us);
1474 advmss = max(sf_advmss, advmss);
1477 msk->rcvq_space.rtt_us = rtt_us;
1478 if (time < (rtt_us >> 3) || rtt_us == 0)
1481 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1484 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1485 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1489 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1491 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1493 do_div(grow, msk->rcvq_space.space);
1494 rcvwin += (grow << 1);
1496 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1497 while (tcp_win_from_space(sk, rcvmem) < advmss)
1500 do_div(rcvwin, advmss);
1501 rcvbuf = min_t(u64, rcvwin * rcvmem,
1502 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1504 if (rcvbuf > sk->sk_rcvbuf) {
1507 window_clamp = tcp_win_from_space(sk, rcvbuf);
1508 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1510 /* Make subflows follow along. If we do not do this, we
1511 * get drops at subflow level if skbs can't be moved to
1512 * the mptcp rx queue fast enough (announced rcv_win can
1513 * exceed ssk->sk_rcvbuf).
1515 mptcp_for_each_subflow(msk, subflow) {
1519 ssk = mptcp_subflow_tcp_sock(subflow);
1520 slow = lock_sock_fast(ssk);
1521 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1522 tcp_sk(ssk)->window_clamp = window_clamp;
1523 tcp_cleanup_rbuf(ssk, 1);
1524 unlock_sock_fast(ssk, slow);
1529 msk->rcvq_space.space = msk->rcvq_space.copied;
1531 msk->rcvq_space.copied = 0;
1532 msk->rcvq_space.time = mstamp;
1535 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1537 unsigned int moved = 0;
1540 /* avoid looping forever below on racing close */
1541 if (((struct sock *)msk)->sk_state == TCP_CLOSE)
1544 __mptcp_flush_join_list(msk);
1546 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1552 slowpath = lock_sock_fast(ssk);
1553 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1554 unlock_sock_fast(ssk, slowpath);
1557 if (mptcp_ofo_queue(msk) || moved > 0) {
1558 if (!mptcp_check_data_fin((struct sock *)msk))
1559 mptcp_send_ack(msk);
1565 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1566 int nonblock, int flags, int *addr_len)
1568 struct mptcp_sock *msk = mptcp_sk(sk);
1573 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1577 timeo = sock_rcvtimeo(sk, nonblock);
1579 len = min_t(size_t, len, INT_MAX);
1580 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1581 __mptcp_flush_join_list(msk);
1583 while (len > (size_t)copied) {
1586 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1587 if (unlikely(bytes_read < 0)) {
1589 copied = bytes_read;
1593 copied += bytes_read;
1595 if (skb_queue_empty(&sk->sk_receive_queue) &&
1596 __mptcp_move_skbs(msk))
1599 /* only the master socket status is relevant here. The exit
1600 * conditions mirror closely tcp_recvmsg()
1602 if (copied >= target)
1607 sk->sk_state == TCP_CLOSE ||
1608 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1610 signal_pending(current))
1614 copied = sock_error(sk);
1618 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1619 mptcp_check_for_eof(msk);
1621 if (sk->sk_shutdown & RCV_SHUTDOWN)
1624 if (sk->sk_state == TCP_CLOSE) {
1634 if (signal_pending(current)) {
1635 copied = sock_intr_errno(timeo);
1640 pr_debug("block timeout %ld", timeo);
1641 mptcp_wait_data(sk, &timeo);
1644 if (skb_queue_empty(&sk->sk_receive_queue)) {
1645 /* entire backlog drained, clear DATA_READY. */
1646 clear_bit(MPTCP_DATA_READY, &msk->flags);
1648 /* .. race-breaker: ssk might have gotten new data
1649 * after last __mptcp_move_skbs() returned false.
1651 if (unlikely(__mptcp_move_skbs(msk)))
1652 set_bit(MPTCP_DATA_READY, &msk->flags);
1653 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1654 /* data to read but mptcp_wait_data() cleared DATA_READY */
1655 set_bit(MPTCP_DATA_READY, &msk->flags);
1658 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
1659 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
1660 skb_queue_empty(&sk->sk_receive_queue), copied);
1661 mptcp_rcv_space_adjust(msk, copied);
1667 static void mptcp_retransmit_handler(struct sock *sk)
1669 struct mptcp_sock *msk = mptcp_sk(sk);
1671 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->snd_nxt)) {
1672 mptcp_stop_timer(sk);
1674 set_bit(MPTCP_WORK_RTX, &msk->flags);
1675 mptcp_schedule_work(sk);
1679 static void mptcp_retransmit_timer(struct timer_list *t)
1681 struct inet_connection_sock *icsk = from_timer(icsk, t,
1682 icsk_retransmit_timer);
1683 struct sock *sk = &icsk->icsk_inet.sk;
1686 if (!sock_owned_by_user(sk)) {
1687 mptcp_retransmit_handler(sk);
1689 /* delegate our work to tcp_release_cb() */
1690 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1698 static void mptcp_timeout_timer(struct timer_list *t)
1700 struct sock *sk = from_timer(sk, t, sk_timer);
1702 mptcp_schedule_work(sk);
1705 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1708 * A backup subflow is returned only if that is the only kind available.
1710 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1712 struct mptcp_subflow_context *subflow;
1713 struct sock *backup = NULL;
1715 sock_owned_by_me((const struct sock *)msk);
1717 if (__mptcp_check_fallback(msk))
1720 mptcp_for_each_subflow(msk, subflow) {
1721 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1723 if (!mptcp_subflow_active(subflow))
1726 /* still data outstanding at TCP level? Don't retransmit. */
1727 if (!tcp_write_queue_empty(ssk))
1730 if (subflow->backup) {
1742 /* subflow sockets can be either outgoing (connect) or incoming
1745 * Outgoing subflows use in-kernel sockets.
1746 * Incoming subflows do not have their own 'struct socket' allocated,
1747 * so we need to use tcp_close() after detaching them from the mptcp
1750 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1751 struct mptcp_subflow_context *subflow)
1753 bool dispose_socket = false;
1754 struct socket *sock;
1756 list_del(&subflow->node);
1760 /* if we are invoked by the msk cleanup code, the subflow is
1763 sock = ssk->sk_socket;
1765 dispose_socket = sock != sk->sk_socket;
1769 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
1770 * the ssk has been already destroyed, we just need to release the
1771 * reference owned by msk;
1773 if (!inet_csk(ssk)->icsk_ulp_ops) {
1774 kfree_rcu(subflow, rcu);
1776 /* otherwise ask tcp do dispose of ssk and subflow ctx */
1777 subflow->disposable = 1;
1778 __tcp_close(ssk, 0);
1780 /* close acquired an extra ref */
1785 iput(SOCK_INODE(sock));
1790 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1795 static void pm_work(struct mptcp_sock *msk)
1797 struct mptcp_pm_data *pm = &msk->pm;
1799 spin_lock_bh(&msk->pm.lock);
1801 pr_debug("msk=%p status=%x", msk, pm->status);
1802 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1803 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1804 mptcp_pm_nl_add_addr_received(msk);
1806 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
1807 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
1808 mptcp_pm_nl_rm_addr_received(msk);
1810 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1811 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1812 mptcp_pm_nl_fully_established(msk);
1814 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1815 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1816 mptcp_pm_nl_subflow_established(msk);
1819 spin_unlock_bh(&msk->pm.lock);
1822 static void __mptcp_close_subflow(struct mptcp_sock *msk)
1824 struct mptcp_subflow_context *subflow, *tmp;
1826 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
1827 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1829 if (inet_sk_state_load(ssk) != TCP_CLOSE)
1832 __mptcp_close_ssk((struct sock *)msk, ssk, subflow);
1836 static bool mptcp_check_close_timeout(const struct sock *sk)
1838 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
1839 struct mptcp_subflow_context *subflow;
1841 if (delta >= TCP_TIMEWAIT_LEN)
1844 /* if all subflows are in closed status don't bother with additional
1847 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
1848 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
1855 static void mptcp_worker(struct work_struct *work)
1857 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1858 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1859 struct mptcp_sendmsg_info info = {};
1860 struct mptcp_data_frag *dfrag;
1865 set_bit(MPTCP_WORKER_RUNNING, &msk->flags);
1866 state = sk->sk_state;
1867 if (unlikely(state == TCP_CLOSE))
1870 mptcp_clean_una_wakeup(sk);
1871 mptcp_check_data_fin_ack(sk);
1872 __mptcp_flush_join_list(msk);
1873 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
1874 __mptcp_close_subflow(msk);
1876 __mptcp_move_skbs(msk);
1877 if (mptcp_send_head(sk))
1878 mptcp_push_pending(sk, 0);
1883 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1884 mptcp_check_for_eof(msk);
1886 mptcp_check_data_fin(sk);
1888 /* if the msk data is completely acked, or the socket timedout,
1889 * there is no point in keeping around an orphaned sk
1891 if (sock_flag(sk, SOCK_DEAD) &&
1892 (mptcp_check_close_timeout(sk) ||
1893 (state != sk->sk_state &&
1894 ((1 << inet_sk_state_load(sk)) & (TCPF_CLOSE | TCPF_FIN_WAIT2))))) {
1895 inet_sk_state_store(sk, TCP_CLOSE);
1896 __mptcp_destroy_sock(sk);
1900 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1903 dfrag = mptcp_rtx_head(sk);
1907 if (!mptcp_ext_cache_refill(msk))
1910 ssk = mptcp_subflow_get_retrans(msk);
1916 /* limit retransmission to the bytes already sent on some subflows */
1918 info.limit = dfrag->already_sent;
1919 while (info.sent < dfrag->already_sent) {
1920 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1924 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1928 if (!mptcp_ext_cache_refill(msk))
1932 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1935 mptcp_set_timeout(sk, ssk);
1939 if (!mptcp_timer_pending(sk))
1940 mptcp_reset_timer(sk);
1943 clear_bit(MPTCP_WORKER_RUNNING, &msk->flags);
1948 static int __mptcp_init_sock(struct sock *sk)
1950 struct mptcp_sock *msk = mptcp_sk(sk);
1952 spin_lock_init(&msk->join_list_lock);
1954 INIT_LIST_HEAD(&msk->conn_list);
1955 INIT_LIST_HEAD(&msk->join_list);
1956 INIT_LIST_HEAD(&msk->rtx_queue);
1957 INIT_WORK(&msk->work, mptcp_worker);
1958 msk->out_of_order_queue = RB_ROOT;
1959 msk->first_pending = NULL;
1962 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1964 mptcp_pm_data_init(msk);
1966 /* re-use the csk retrans timer for MPTCP-level retrans */
1967 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1968 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
1972 static int mptcp_init_sock(struct sock *sk)
1974 struct net *net = sock_net(sk);
1977 ret = __mptcp_init_sock(sk);
1981 if (!mptcp_is_enabled(net))
1982 return -ENOPROTOOPT;
1984 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1987 ret = __mptcp_socket_create(mptcp_sk(sk));
1991 sk_sockets_allocated_inc(sk);
1992 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1993 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
1998 static void __mptcp_clear_xmit(struct sock *sk)
2000 struct mptcp_sock *msk = mptcp_sk(sk);
2001 struct mptcp_data_frag *dtmp, *dfrag;
2003 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2005 WRITE_ONCE(msk->first_pending, NULL);
2006 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2007 dfrag_clear(sk, dfrag);
2010 static void mptcp_cancel_work(struct sock *sk)
2012 struct mptcp_sock *msk = mptcp_sk(sk);
2014 /* if called by the work itself, do not try to cancel the work, or
2017 if (!test_bit(MPTCP_WORKER_RUNNING, &msk->flags) &&
2018 cancel_work_sync(&msk->work))
2022 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2026 switch (ssk->sk_state) {
2028 if (!(how & RCV_SHUTDOWN))
2032 tcp_disconnect(ssk, O_NONBLOCK);
2035 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2036 pr_debug("Fallback");
2037 ssk->sk_shutdown |= how;
2038 tcp_shutdown(ssk, how);
2040 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2041 mptcp_set_timeout(sk, ssk);
2050 static const unsigned char new_state[16] = {
2051 /* current state: new state: action: */
2052 [0 /* (Invalid) */] = TCP_CLOSE,
2053 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2054 [TCP_SYN_SENT] = TCP_CLOSE,
2055 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2056 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2057 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2058 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2059 [TCP_CLOSE] = TCP_CLOSE,
2060 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2061 [TCP_LAST_ACK] = TCP_LAST_ACK,
2062 [TCP_LISTEN] = TCP_CLOSE,
2063 [TCP_CLOSING] = TCP_CLOSING,
2064 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2067 static int mptcp_close_state(struct sock *sk)
2069 int next = (int)new_state[sk->sk_state];
2070 int ns = next & TCP_STATE_MASK;
2072 inet_sk_state_store(sk, ns);
2074 return next & TCP_ACTION_FIN;
2077 static void __mptcp_check_send_data_fin(struct sock *sk)
2079 struct mptcp_subflow_context *subflow;
2080 struct mptcp_sock *msk = mptcp_sk(sk);
2082 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2083 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2084 msk->snd_nxt, msk->write_seq);
2086 /* we still need to enqueue subflows or not really shutting down,
2089 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2090 mptcp_send_head(sk))
2093 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2095 /* fallback socket will not get data_fin/ack, can move to close now */
2096 if (__mptcp_check_fallback(msk) && sk->sk_state == TCP_LAST_ACK) {
2097 inet_sk_state_store(sk, TCP_CLOSE);
2098 mptcp_close_wake_up(sk);
2101 __mptcp_flush_join_list(msk);
2102 mptcp_for_each_subflow(msk, subflow) {
2103 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2105 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2109 static void __mptcp_wr_shutdown(struct sock *sk)
2111 struct mptcp_sock *msk = mptcp_sk(sk);
2113 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2114 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2115 !!mptcp_send_head(sk));
2117 /* will be ignored by fallback sockets */
2118 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2119 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2121 __mptcp_check_send_data_fin(sk);
2124 static void __mptcp_destroy_sock(struct sock *sk)
2126 struct mptcp_subflow_context *subflow, *tmp;
2127 struct mptcp_sock *msk = mptcp_sk(sk);
2128 LIST_HEAD(conn_list);
2130 pr_debug("msk=%p", msk);
2132 /* be sure to always acquire the join list lock, to sync vs
2133 * mptcp_finish_join().
2135 spin_lock_bh(&msk->join_list_lock);
2136 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2137 spin_unlock_bh(&msk->join_list_lock);
2138 list_splice_init(&msk->conn_list, &conn_list);
2140 __mptcp_clear_xmit(sk);
2141 sk_stop_timer(sk, &sk->sk_timer);
2144 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2145 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2146 __mptcp_close_ssk(sk, ssk, subflow);
2149 sk->sk_prot->destroy(sk);
2151 sk_stream_kill_queues(sk);
2152 xfrm_sk_free_policy(sk);
2153 sk_refcnt_debug_release(sk);
2157 static void mptcp_close(struct sock *sk, long timeout)
2159 struct mptcp_subflow_context *subflow;
2160 bool do_cancel_work = false;
2163 sk->sk_shutdown = SHUTDOWN_MASK;
2165 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2166 inet_sk_state_store(sk, TCP_CLOSE);
2170 if (mptcp_close_state(sk))
2171 __mptcp_wr_shutdown(sk);
2173 sk_stream_wait_close(sk, timeout);
2176 /* orphan all the subflows */
2177 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2178 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2179 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2180 bool slow, dispose_socket;
2181 struct socket *sock;
2183 slow = lock_sock_fast(ssk);
2184 sock = ssk->sk_socket;
2185 dispose_socket = sock && sock != sk->sk_socket;
2187 unlock_sock_fast(ssk, slow);
2189 /* for the outgoing subflows we additionally need to free
2190 * the associated socket
2193 iput(SOCK_INODE(sock));
2198 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2199 if (sk->sk_state == TCP_CLOSE) {
2200 __mptcp_destroy_sock(sk);
2201 do_cancel_work = true;
2203 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2207 mptcp_cancel_work(sk);
2211 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2213 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2214 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2215 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2217 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2218 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2221 msk6->saddr = ssk6->saddr;
2222 msk6->flow_label = ssk6->flow_label;
2226 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2227 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2228 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2229 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2230 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2231 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2234 static int mptcp_disconnect(struct sock *sk, int flags)
2236 /* Should never be called.
2237 * inet_stream_connect() calls ->disconnect, but that
2238 * refers to the subflow socket, not the mptcp one.
2244 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2245 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2247 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2249 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2253 struct sock *mptcp_sk_clone(const struct sock *sk,
2254 const struct mptcp_options_received *mp_opt,
2255 struct request_sock *req)
2257 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2258 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2259 struct mptcp_sock *msk;
2265 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2266 if (nsk->sk_family == AF_INET6)
2267 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2270 __mptcp_init_sock(nsk);
2272 msk = mptcp_sk(nsk);
2273 msk->local_key = subflow_req->local_key;
2274 msk->token = subflow_req->token;
2275 msk->subflow = NULL;
2276 WRITE_ONCE(msk->fully_established, false);
2278 msk->write_seq = subflow_req->idsn + 1;
2279 msk->snd_nxt = msk->write_seq;
2280 atomic64_set(&msk->snd_una, msk->write_seq);
2281 atomic64_set(&msk->wnd_end, msk->snd_nxt + req->rsk_rcv_wnd);
2283 if (mp_opt->mp_capable) {
2284 msk->can_ack = true;
2285 msk->remote_key = mp_opt->sndr_key;
2286 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2288 WRITE_ONCE(msk->ack_seq, ack_seq);
2291 sock_reset_flag(nsk, SOCK_RCU_FREE);
2292 /* will be fully established after successful MPC subflow creation */
2293 inet_sk_state_store(nsk, TCP_SYN_RECV);
2294 bh_unlock_sock(nsk);
2296 /* keep a single reference */
2301 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2303 const struct tcp_sock *tp = tcp_sk(ssk);
2305 msk->rcvq_space.copied = 0;
2306 msk->rcvq_space.rtt_us = 0;
2308 msk->rcvq_space.time = tp->tcp_mstamp;
2310 /* initial rcv_space offering made to peer */
2311 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2312 TCP_INIT_CWND * tp->advmss);
2313 if (msk->rcvq_space.space == 0)
2314 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2316 atomic64_set(&msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2319 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2322 struct mptcp_sock *msk = mptcp_sk(sk);
2323 struct socket *listener;
2326 listener = __mptcp_nmpc_socket(msk);
2327 if (WARN_ON_ONCE(!listener)) {
2332 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2333 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2337 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2338 if (sk_is_mptcp(newsk)) {
2339 struct mptcp_subflow_context *subflow;
2340 struct sock *new_mptcp_sock;
2341 struct sock *ssk = newsk;
2343 subflow = mptcp_subflow_ctx(newsk);
2344 new_mptcp_sock = subflow->conn;
2346 /* is_mptcp should be false if subflow->conn is missing, see
2347 * subflow_syn_recv_sock()
2349 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2350 tcp_sk(newsk)->is_mptcp = 0;
2354 /* acquire the 2nd reference for the owning socket */
2355 sock_hold(new_mptcp_sock);
2358 bh_lock_sock(new_mptcp_sock);
2359 msk = mptcp_sk(new_mptcp_sock);
2362 newsk = new_mptcp_sock;
2363 mptcp_copy_inaddrs(newsk, ssk);
2364 list_add(&subflow->node, &msk->conn_list);
2367 mptcp_rcv_space_init(msk, ssk);
2368 bh_unlock_sock(new_mptcp_sock);
2370 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2373 MPTCP_INC_STATS(sock_net(sk),
2374 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2380 void mptcp_destroy_common(struct mptcp_sock *msk)
2382 skb_rbtree_purge(&msk->out_of_order_queue);
2383 mptcp_token_destroy(msk);
2384 mptcp_pm_free_anno_list(msk);
2387 static void mptcp_destroy(struct sock *sk)
2389 struct mptcp_sock *msk = mptcp_sk(sk);
2391 if (msk->cached_ext)
2392 __skb_ext_put(msk->cached_ext);
2394 mptcp_destroy_common(msk);
2395 sk_sockets_allocated_dec(sk);
2398 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2399 sockptr_t optval, unsigned int optlen)
2401 struct sock *sk = (struct sock *)msk;
2402 struct socket *ssock;
2409 ssock = __mptcp_nmpc_socket(msk);
2415 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2417 if (optname == SO_REUSEPORT)
2418 sk->sk_reuseport = ssock->sk->sk_reuseport;
2419 else if (optname == SO_REUSEADDR)
2420 sk->sk_reuse = ssock->sk->sk_reuse;
2426 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2429 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2430 sockptr_t optval, unsigned int optlen)
2432 struct sock *sk = (struct sock *)msk;
2433 int ret = -EOPNOTSUPP;
2434 struct socket *ssock;
2439 ssock = __mptcp_nmpc_socket(msk);
2445 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2447 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2456 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2457 sockptr_t optval, unsigned int optlen)
2459 struct mptcp_sock *msk = mptcp_sk(sk);
2462 pr_debug("msk=%p", msk);
2464 if (level == SOL_SOCKET)
2465 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2467 /* @@ the meaning of setsockopt() when the socket is connected and
2468 * there are multiple subflows is not yet defined. It is up to the
2469 * MPTCP-level socket to configure the subflows until the subflow
2470 * is in TCP fallback, when TCP socket options are passed through
2471 * to the one remaining subflow.
2474 ssk = __mptcp_tcp_fallback(msk);
2477 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2479 if (level == SOL_IPV6)
2480 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2485 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2486 char __user *optval, int __user *option)
2488 struct mptcp_sock *msk = mptcp_sk(sk);
2491 pr_debug("msk=%p", msk);
2493 /* @@ the meaning of setsockopt() when the socket is connected and
2494 * there are multiple subflows is not yet defined. It is up to the
2495 * MPTCP-level socket to configure the subflows until the subflow
2496 * is in TCP fallback, when socket options are passed through
2497 * to the one remaining subflow.
2500 ssk = __mptcp_tcp_fallback(msk);
2503 return tcp_getsockopt(ssk, level, optname, optval, option);
2508 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
2509 TCPF_WRITE_TIMER_DEFERRED)
2511 /* this is very alike tcp_release_cb() but we must handle differently a
2512 * different set of events
2514 static void mptcp_release_cb(struct sock *sk)
2516 unsigned long flags, nflags;
2519 flags = sk->sk_tsq_flags;
2520 if (!(flags & MPTCP_DEFERRED_ALL))
2522 nflags = flags & ~MPTCP_DEFERRED_ALL;
2523 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2525 sock_release_ownership(sk);
2527 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
2528 struct mptcp_sock *msk = mptcp_sk(sk);
2531 ssk = mptcp_subflow_recv_lookup(msk);
2532 if (!ssk || sk->sk_state == TCP_CLOSE ||
2533 !schedule_work(&msk->work))
2537 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2538 mptcp_retransmit_handler(sk);
2543 static int mptcp_hash(struct sock *sk)
2545 /* should never be called,
2546 * we hash the TCP subflows not the master socket
2552 static void mptcp_unhash(struct sock *sk)
2554 /* called from sk_common_release(), but nothing to do here */
2557 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2559 struct mptcp_sock *msk = mptcp_sk(sk);
2560 struct socket *ssock;
2562 ssock = __mptcp_nmpc_socket(msk);
2563 pr_debug("msk=%p, subflow=%p", msk, ssock);
2564 if (WARN_ON_ONCE(!ssock))
2567 return inet_csk_get_port(ssock->sk, snum);
2570 void mptcp_finish_connect(struct sock *ssk)
2572 struct mptcp_subflow_context *subflow;
2573 struct mptcp_sock *msk;
2577 subflow = mptcp_subflow_ctx(ssk);
2581 pr_debug("msk=%p, token=%u", sk, subflow->token);
2583 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2585 subflow->map_seq = ack_seq;
2586 subflow->map_subflow_seq = 1;
2588 /* the socket is not connected yet, no msk/subflow ops can access/race
2589 * accessing the field below
2591 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2592 WRITE_ONCE(msk->local_key, subflow->local_key);
2593 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2594 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2595 WRITE_ONCE(msk->ack_seq, ack_seq);
2596 WRITE_ONCE(msk->can_ack, 1);
2597 atomic64_set(&msk->snd_una, msk->write_seq);
2599 mptcp_pm_new_connection(msk, 0);
2601 mptcp_rcv_space_init(msk, ssk);
2604 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2606 write_lock_bh(&sk->sk_callback_lock);
2607 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2608 sk_set_socket(sk, parent);
2609 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2610 write_unlock_bh(&sk->sk_callback_lock);
2613 bool mptcp_finish_join(struct sock *ssk)
2615 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2616 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2617 struct sock *parent = (void *)msk;
2618 struct socket *parent_sock;
2621 pr_debug("msk=%p, subflow=%p", msk, subflow);
2623 /* mptcp socket already closing? */
2624 if (!mptcp_is_fully_established(parent))
2627 if (!msk->pm.server_side)
2630 if (!mptcp_pm_allow_new_subflow(msk))
2633 /* active connections are already on conn_list, and we can't acquire
2635 * use the join list lock as synchronization point and double-check
2636 * msk status to avoid racing with __mptcp_destroy_sock()
2638 spin_lock_bh(&msk->join_list_lock);
2639 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2640 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
2641 list_add_tail(&subflow->node, &msk->join_list);
2644 spin_unlock_bh(&msk->join_list_lock);
2648 /* attach to msk socket only after we are sure he will deal with us
2651 parent_sock = READ_ONCE(parent->sk_socket);
2652 if (parent_sock && !ssk->sk_socket)
2653 mptcp_sock_graft(ssk, parent_sock);
2654 subflow->map_seq = READ_ONCE(msk->ack_seq);
2658 static struct proto mptcp_prot = {
2660 .owner = THIS_MODULE,
2661 .init = mptcp_init_sock,
2662 .disconnect = mptcp_disconnect,
2663 .close = mptcp_close,
2664 .accept = mptcp_accept,
2665 .setsockopt = mptcp_setsockopt,
2666 .getsockopt = mptcp_getsockopt,
2667 .shutdown = tcp_shutdown,
2668 .destroy = mptcp_destroy,
2669 .sendmsg = mptcp_sendmsg,
2670 .recvmsg = mptcp_recvmsg,
2671 .release_cb = mptcp_release_cb,
2673 .unhash = mptcp_unhash,
2674 .get_port = mptcp_get_port,
2675 .sockets_allocated = &mptcp_sockets_allocated,
2676 .memory_allocated = &tcp_memory_allocated,
2677 .memory_pressure = &tcp_memory_pressure,
2678 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2679 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
2680 .sysctl_mem = sysctl_tcp_mem,
2681 .obj_size = sizeof(struct mptcp_sock),
2682 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2683 .no_autobind = true,
2686 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2688 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2689 struct socket *ssock;
2692 lock_sock(sock->sk);
2693 ssock = __mptcp_nmpc_socket(msk);
2699 err = ssock->ops->bind(ssock, uaddr, addr_len);
2701 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2704 release_sock(sock->sk);
2708 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2709 struct mptcp_subflow_context *subflow)
2711 subflow->request_mptcp = 0;
2712 __mptcp_do_fallback(msk);
2715 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2716 int addr_len, int flags)
2718 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2719 struct mptcp_subflow_context *subflow;
2720 struct socket *ssock;
2723 lock_sock(sock->sk);
2724 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2725 /* pending connection or invalid state, let existing subflow
2728 ssock = msk->subflow;
2732 ssock = __mptcp_nmpc_socket(msk);
2738 mptcp_token_destroy(msk);
2739 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2740 subflow = mptcp_subflow_ctx(ssock->sk);
2741 #ifdef CONFIG_TCP_MD5SIG
2742 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2745 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2746 mptcp_subflow_early_fallback(msk, subflow);
2748 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2749 mptcp_subflow_early_fallback(msk, subflow);
2752 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2753 sock->state = ssock->state;
2755 /* on successful connect, the msk state will be moved to established by
2756 * subflow_finish_connect()
2758 if (!err || err == -EINPROGRESS)
2759 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2761 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2764 release_sock(sock->sk);
2768 static int mptcp_listen(struct socket *sock, int backlog)
2770 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2771 struct socket *ssock;
2774 pr_debug("msk=%p", msk);
2776 lock_sock(sock->sk);
2777 ssock = __mptcp_nmpc_socket(msk);
2783 mptcp_token_destroy(msk);
2784 inet_sk_state_store(sock->sk, TCP_LISTEN);
2785 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2787 err = ssock->ops->listen(ssock, backlog);
2788 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2790 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2793 release_sock(sock->sk);
2797 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2798 int flags, bool kern)
2800 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2801 struct socket *ssock;
2804 pr_debug("msk=%p", msk);
2806 lock_sock(sock->sk);
2807 if (sock->sk->sk_state != TCP_LISTEN)
2810 ssock = __mptcp_nmpc_socket(msk);
2814 clear_bit(MPTCP_DATA_READY, &msk->flags);
2815 sock_hold(ssock->sk);
2816 release_sock(sock->sk);
2818 err = ssock->ops->accept(sock, newsock, flags, kern);
2819 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2820 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2821 struct mptcp_subflow_context *subflow;
2823 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2824 * This is needed so NOSPACE flag can be set from tcp stack.
2826 __mptcp_flush_join_list(msk);
2827 mptcp_for_each_subflow(msk, subflow) {
2828 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2830 if (!ssk->sk_socket)
2831 mptcp_sock_graft(ssk, newsock);
2835 if (inet_csk_listen_poll(ssock->sk))
2836 set_bit(MPTCP_DATA_READY, &msk->flags);
2837 sock_put(ssock->sk);
2841 release_sock(sock->sk);
2845 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2847 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2851 static bool __mptcp_check_writeable(struct mptcp_sock *msk)
2853 struct sock *sk = (struct sock *)msk;
2854 bool mptcp_writable;
2856 mptcp_clean_una(sk);
2857 mptcp_writable = sk_stream_is_writeable(sk);
2858 if (!mptcp_writable)
2861 return mptcp_writable;
2864 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
2866 struct sock *sk = (struct sock *)msk;
2870 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
2873 if (sk_stream_is_writeable(sk))
2874 return EPOLLOUT | EPOLLWRNORM;
2876 slow = lock_sock_fast(sk);
2877 if (__mptcp_check_writeable(msk))
2878 ret = EPOLLOUT | EPOLLWRNORM;
2880 unlock_sock_fast(sk, slow);
2884 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2885 struct poll_table_struct *wait)
2887 struct sock *sk = sock->sk;
2888 struct mptcp_sock *msk;
2893 sock_poll_wait(file, sock, wait);
2895 state = inet_sk_state_load(sk);
2896 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
2897 if (state == TCP_LISTEN)
2898 return mptcp_check_readable(msk);
2900 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2901 mask |= mptcp_check_readable(msk);
2902 mask |= mptcp_check_writeable(msk);
2904 if (sk->sk_shutdown & RCV_SHUTDOWN)
2905 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2910 static int mptcp_shutdown(struct socket *sock, int how)
2912 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2913 struct sock *sk = sock->sk;
2916 pr_debug("sk=%p, how=%d", msk, how);
2921 if ((how & ~SHUTDOWN_MASK) || !how) {
2926 if (sock->state == SS_CONNECTING) {
2927 if ((1 << sk->sk_state) &
2928 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2929 sock->state = SS_DISCONNECTING;
2931 sock->state = SS_CONNECTED;
2934 sk->sk_shutdown |= how;
2935 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
2936 __mptcp_wr_shutdown(sk);
2938 /* Wake up anyone sleeping in poll. */
2939 sk->sk_state_change(sk);
2947 static const struct proto_ops mptcp_stream_ops = {
2949 .owner = THIS_MODULE,
2950 .release = inet_release,
2952 .connect = mptcp_stream_connect,
2953 .socketpair = sock_no_socketpair,
2954 .accept = mptcp_stream_accept,
2955 .getname = inet_getname,
2957 .ioctl = inet_ioctl,
2958 .gettstamp = sock_gettstamp,
2959 .listen = mptcp_listen,
2960 .shutdown = mptcp_shutdown,
2961 .setsockopt = sock_common_setsockopt,
2962 .getsockopt = sock_common_getsockopt,
2963 .sendmsg = inet_sendmsg,
2964 .recvmsg = inet_recvmsg,
2965 .mmap = sock_no_mmap,
2966 .sendpage = inet_sendpage,
2969 static struct inet_protosw mptcp_protosw = {
2970 .type = SOCK_STREAM,
2971 .protocol = IPPROTO_MPTCP,
2972 .prot = &mptcp_prot,
2973 .ops = &mptcp_stream_ops,
2974 .flags = INET_PROTOSW_ICSK,
2977 void __init mptcp_proto_init(void)
2979 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2981 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2982 panic("Failed to allocate MPTCP pcpu counter\n");
2984 mptcp_subflow_init();
2988 if (proto_register(&mptcp_prot, 1) != 0)
2989 panic("Failed to register MPTCP proto.\n");
2991 inet_register_protosw(&mptcp_protosw);
2993 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2996 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2997 static const struct proto_ops mptcp_v6_stream_ops = {
2999 .owner = THIS_MODULE,
3000 .release = inet6_release,
3002 .connect = mptcp_stream_connect,
3003 .socketpair = sock_no_socketpair,
3004 .accept = mptcp_stream_accept,
3005 .getname = inet6_getname,
3007 .ioctl = inet6_ioctl,
3008 .gettstamp = sock_gettstamp,
3009 .listen = mptcp_listen,
3010 .shutdown = mptcp_shutdown,
3011 .setsockopt = sock_common_setsockopt,
3012 .getsockopt = sock_common_getsockopt,
3013 .sendmsg = inet6_sendmsg,
3014 .recvmsg = inet6_recvmsg,
3015 .mmap = sock_no_mmap,
3016 .sendpage = inet_sendpage,
3017 #ifdef CONFIG_COMPAT
3018 .compat_ioctl = inet6_compat_ioctl,
3022 static struct proto mptcp_v6_prot;
3024 static void mptcp_v6_destroy(struct sock *sk)
3027 inet6_destroy_sock(sk);
3030 static struct inet_protosw mptcp_v6_protosw = {
3031 .type = SOCK_STREAM,
3032 .protocol = IPPROTO_MPTCP,
3033 .prot = &mptcp_v6_prot,
3034 .ops = &mptcp_v6_stream_ops,
3035 .flags = INET_PROTOSW_ICSK,
3038 int __init mptcp_proto_v6_init(void)
3042 mptcp_v6_prot = mptcp_prot;
3043 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3044 mptcp_v6_prot.slab = NULL;
3045 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3046 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3048 err = proto_register(&mptcp_v6_prot, 1);
3052 err = inet6_register_protosw(&mptcp_v6_protosw);
3054 proto_unregister(&mptcp_v6_prot);