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);
1264 /* start the timer, if it's not pending */
1265 if (!mptcp_timer_pending(sk))
1266 mptcp_reset_timer(sk);
1268 __mptcp_check_send_data_fin(sk);
1271 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1273 struct mptcp_sock *msk = mptcp_sk(sk);
1274 struct page_frag *pfrag;
1279 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1284 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1286 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1287 ret = sk_stream_wait_connect(sk, &timeo);
1292 pfrag = sk_page_frag(sk);
1293 mptcp_clean_una(sk);
1295 while (msg_data_left(msg)) {
1296 struct mptcp_data_frag *dfrag;
1297 int frag_truesize = 0;
1298 bool dfrag_collapsed;
1299 size_t psize, offset;
1301 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1306 /* reuse tail pfrag, if possible, or carve a new one from the
1309 dfrag = mptcp_pending_tail(sk);
1310 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1311 if (!dfrag_collapsed) {
1312 if (!sk_stream_memory_free(sk)) {
1313 mptcp_push_pending(sk, msg->msg_flags);
1314 if (!sk_stream_memory_free(sk))
1315 goto wait_for_memory;
1317 if (!mptcp_page_frag_refill(sk, pfrag))
1318 goto wait_for_memory;
1320 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1321 frag_truesize = dfrag->overhead;
1324 /* we do not bound vs wspace, to allow a single packet.
1325 * memory accounting will prevent execessive memory usage
1328 offset = dfrag->offset + dfrag->data_len;
1329 psize = pfrag->size - offset;
1330 psize = min_t(size_t, psize, msg_data_left(msg));
1331 if (!sk_wmem_schedule(sk, psize + frag_truesize))
1332 goto wait_for_memory;
1334 if (copy_page_from_iter(dfrag->page, offset, psize,
1335 &msg->msg_iter) != psize) {
1340 /* data successfully copied into the write queue */
1342 dfrag->data_len += psize;
1343 frag_truesize += psize;
1344 pfrag->offset += frag_truesize;
1345 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1347 /* charge data on mptcp pending queue to the msk socket
1348 * Note: we charge such data both to sk and ssk
1350 sk_wmem_queued_add(sk, frag_truesize);
1351 sk->sk_forward_alloc -= frag_truesize;
1352 if (!dfrag_collapsed) {
1353 get_page(dfrag->page);
1354 list_add_tail(&dfrag->list, &msk->rtx_queue);
1355 if (!msk->first_pending)
1356 WRITE_ONCE(msk->first_pending, dfrag);
1358 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1359 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1362 if (!mptcp_ext_cache_refill(msk))
1363 goto wait_for_memory;
1368 if (mptcp_timer_pending(sk))
1369 mptcp_reset_timer(sk);
1370 ret = sk_stream_wait_memory(sk, &timeo);
1376 mptcp_push_pending(sk, msg->msg_flags);
1380 return copied ? : ret;
1383 static void mptcp_wait_data(struct sock *sk, long *timeo)
1385 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1386 struct mptcp_sock *msk = mptcp_sk(sk);
1388 add_wait_queue(sk_sleep(sk), &wait);
1389 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1391 sk_wait_event(sk, timeo,
1392 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1394 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1395 remove_wait_queue(sk_sleep(sk), &wait);
1398 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1402 struct sock *sk = (struct sock *)msk;
1403 struct sk_buff *skb;
1406 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1407 u32 offset = MPTCP_SKB_CB(skb)->offset;
1408 u32 data_len = skb->len - offset;
1409 u32 count = min_t(size_t, len - copied, data_len);
1412 err = skb_copy_datagram_msg(skb, offset, msg, count);
1413 if (unlikely(err < 0)) {
1421 if (count < data_len) {
1422 MPTCP_SKB_CB(skb)->offset += count;
1426 __skb_unlink(skb, &sk->sk_receive_queue);
1436 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1438 * Only difference: Use highest rtt estimate of the subflows in use.
1440 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1442 struct mptcp_subflow_context *subflow;
1443 struct sock *sk = (struct sock *)msk;
1444 u32 time, advmss = 1;
1447 sock_owned_by_me(sk);
1452 msk->rcvq_space.copied += copied;
1454 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1455 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1457 rtt_us = msk->rcvq_space.rtt_us;
1458 if (rtt_us && time < (rtt_us >> 3))
1462 mptcp_for_each_subflow(msk, subflow) {
1463 const struct tcp_sock *tp;
1467 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1469 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1470 sf_advmss = READ_ONCE(tp->advmss);
1472 rtt_us = max(sf_rtt_us, rtt_us);
1473 advmss = max(sf_advmss, advmss);
1476 msk->rcvq_space.rtt_us = rtt_us;
1477 if (time < (rtt_us >> 3) || rtt_us == 0)
1480 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1483 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1484 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1488 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1490 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1492 do_div(grow, msk->rcvq_space.space);
1493 rcvwin += (grow << 1);
1495 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1496 while (tcp_win_from_space(sk, rcvmem) < advmss)
1499 do_div(rcvwin, advmss);
1500 rcvbuf = min_t(u64, rcvwin * rcvmem,
1501 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1503 if (rcvbuf > sk->sk_rcvbuf) {
1506 window_clamp = tcp_win_from_space(sk, rcvbuf);
1507 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1509 /* Make subflows follow along. If we do not do this, we
1510 * get drops at subflow level if skbs can't be moved to
1511 * the mptcp rx queue fast enough (announced rcv_win can
1512 * exceed ssk->sk_rcvbuf).
1514 mptcp_for_each_subflow(msk, subflow) {
1518 ssk = mptcp_subflow_tcp_sock(subflow);
1519 slow = lock_sock_fast(ssk);
1520 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1521 tcp_sk(ssk)->window_clamp = window_clamp;
1522 tcp_cleanup_rbuf(ssk, 1);
1523 unlock_sock_fast(ssk, slow);
1528 msk->rcvq_space.space = msk->rcvq_space.copied;
1530 msk->rcvq_space.copied = 0;
1531 msk->rcvq_space.time = mstamp;
1534 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1536 unsigned int moved = 0;
1539 /* avoid looping forever below on racing close */
1540 if (((struct sock *)msk)->sk_state == TCP_CLOSE)
1543 __mptcp_flush_join_list(msk);
1545 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1551 slowpath = lock_sock_fast(ssk);
1552 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1553 unlock_sock_fast(ssk, slowpath);
1556 if (mptcp_ofo_queue(msk) || moved > 0) {
1557 if (!mptcp_check_data_fin((struct sock *)msk))
1558 mptcp_send_ack(msk);
1564 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1565 int nonblock, int flags, int *addr_len)
1567 struct mptcp_sock *msk = mptcp_sk(sk);
1572 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1576 timeo = sock_rcvtimeo(sk, nonblock);
1578 len = min_t(size_t, len, INT_MAX);
1579 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1580 __mptcp_flush_join_list(msk);
1582 while (len > (size_t)copied) {
1585 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1586 if (unlikely(bytes_read < 0)) {
1588 copied = bytes_read;
1592 copied += bytes_read;
1594 if (skb_queue_empty(&sk->sk_receive_queue) &&
1595 __mptcp_move_skbs(msk))
1598 /* only the master socket status is relevant here. The exit
1599 * conditions mirror closely tcp_recvmsg()
1601 if (copied >= target)
1606 sk->sk_state == TCP_CLOSE ||
1607 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1609 signal_pending(current))
1613 copied = sock_error(sk);
1617 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1618 mptcp_check_for_eof(msk);
1620 if (sk->sk_shutdown & RCV_SHUTDOWN)
1623 if (sk->sk_state == TCP_CLOSE) {
1633 if (signal_pending(current)) {
1634 copied = sock_intr_errno(timeo);
1639 pr_debug("block timeout %ld", timeo);
1640 mptcp_wait_data(sk, &timeo);
1643 if (skb_queue_empty(&sk->sk_receive_queue)) {
1644 /* entire backlog drained, clear DATA_READY. */
1645 clear_bit(MPTCP_DATA_READY, &msk->flags);
1647 /* .. race-breaker: ssk might have gotten new data
1648 * after last __mptcp_move_skbs() returned false.
1650 if (unlikely(__mptcp_move_skbs(msk)))
1651 set_bit(MPTCP_DATA_READY, &msk->flags);
1652 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1653 /* data to read but mptcp_wait_data() cleared DATA_READY */
1654 set_bit(MPTCP_DATA_READY, &msk->flags);
1657 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
1658 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
1659 skb_queue_empty(&sk->sk_receive_queue), copied);
1660 mptcp_rcv_space_adjust(msk, copied);
1666 static void mptcp_retransmit_handler(struct sock *sk)
1668 struct mptcp_sock *msk = mptcp_sk(sk);
1670 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->snd_nxt)) {
1671 mptcp_stop_timer(sk);
1673 set_bit(MPTCP_WORK_RTX, &msk->flags);
1674 mptcp_schedule_work(sk);
1678 static void mptcp_retransmit_timer(struct timer_list *t)
1680 struct inet_connection_sock *icsk = from_timer(icsk, t,
1681 icsk_retransmit_timer);
1682 struct sock *sk = &icsk->icsk_inet.sk;
1685 if (!sock_owned_by_user(sk)) {
1686 mptcp_retransmit_handler(sk);
1688 /* delegate our work to tcp_release_cb() */
1689 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1697 static void mptcp_timeout_timer(struct timer_list *t)
1699 struct sock *sk = from_timer(sk, t, sk_timer);
1701 mptcp_schedule_work(sk);
1704 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1707 * A backup subflow is returned only if that is the only kind available.
1709 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1711 struct mptcp_subflow_context *subflow;
1712 struct sock *backup = NULL;
1714 sock_owned_by_me((const struct sock *)msk);
1716 if (__mptcp_check_fallback(msk))
1719 mptcp_for_each_subflow(msk, subflow) {
1720 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1722 if (!mptcp_subflow_active(subflow))
1725 /* still data outstanding at TCP level? Don't retransmit. */
1726 if (!tcp_write_queue_empty(ssk))
1729 if (subflow->backup) {
1741 /* subflow sockets can be either outgoing (connect) or incoming
1744 * Outgoing subflows use in-kernel sockets.
1745 * Incoming subflows do not have their own 'struct socket' allocated,
1746 * so we need to use tcp_close() after detaching them from the mptcp
1749 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1750 struct mptcp_subflow_context *subflow)
1752 bool dispose_socket = false;
1753 struct socket *sock;
1755 list_del(&subflow->node);
1759 /* if we are invoked by the msk cleanup code, the subflow is
1762 sock = ssk->sk_socket;
1764 dispose_socket = sock != sk->sk_socket;
1768 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
1769 * the ssk has been already destroyed, we just need to release the
1770 * reference owned by msk;
1772 if (!inet_csk(ssk)->icsk_ulp_ops) {
1773 kfree_rcu(subflow, rcu);
1775 /* otherwise ask tcp do dispose of ssk and subflow ctx */
1776 subflow->disposable = 1;
1777 __tcp_close(ssk, 0);
1779 /* close acquired an extra ref */
1784 iput(SOCK_INODE(sock));
1789 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1794 static void pm_work(struct mptcp_sock *msk)
1796 struct mptcp_pm_data *pm = &msk->pm;
1798 spin_lock_bh(&msk->pm.lock);
1800 pr_debug("msk=%p status=%x", msk, pm->status);
1801 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1802 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1803 mptcp_pm_nl_add_addr_received(msk);
1805 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
1806 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
1807 mptcp_pm_nl_rm_addr_received(msk);
1809 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1810 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1811 mptcp_pm_nl_fully_established(msk);
1813 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1814 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1815 mptcp_pm_nl_subflow_established(msk);
1818 spin_unlock_bh(&msk->pm.lock);
1821 static void __mptcp_close_subflow(struct mptcp_sock *msk)
1823 struct mptcp_subflow_context *subflow, *tmp;
1825 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
1826 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1828 if (inet_sk_state_load(ssk) != TCP_CLOSE)
1831 __mptcp_close_ssk((struct sock *)msk, ssk, subflow);
1835 static bool mptcp_check_close_timeout(const struct sock *sk)
1837 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
1838 struct mptcp_subflow_context *subflow;
1840 if (delta >= TCP_TIMEWAIT_LEN)
1843 /* if all subflows are in closed status don't bother with additional
1846 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
1847 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
1854 static void mptcp_worker(struct work_struct *work)
1856 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1857 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1858 struct mptcp_sendmsg_info info = {};
1859 struct mptcp_data_frag *dfrag;
1864 set_bit(MPTCP_WORKER_RUNNING, &msk->flags);
1865 state = sk->sk_state;
1866 if (unlikely(state == TCP_CLOSE))
1869 mptcp_clean_una_wakeup(sk);
1870 mptcp_check_data_fin_ack(sk);
1871 __mptcp_flush_join_list(msk);
1872 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
1873 __mptcp_close_subflow(msk);
1875 __mptcp_move_skbs(msk);
1876 if (mptcp_send_head(sk))
1877 mptcp_push_pending(sk, 0);
1882 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1883 mptcp_check_for_eof(msk);
1885 mptcp_check_data_fin(sk);
1887 /* if the msk data is completely acked, or the socket timedout,
1888 * there is no point in keeping around an orphaned sk
1890 if (sock_flag(sk, SOCK_DEAD) &&
1891 (mptcp_check_close_timeout(sk) ||
1892 (state != sk->sk_state &&
1893 ((1 << inet_sk_state_load(sk)) & (TCPF_CLOSE | TCPF_FIN_WAIT2))))) {
1894 inet_sk_state_store(sk, TCP_CLOSE);
1895 __mptcp_destroy_sock(sk);
1899 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1902 dfrag = mptcp_rtx_head(sk);
1906 if (!mptcp_ext_cache_refill(msk))
1909 ssk = mptcp_subflow_get_retrans(msk);
1915 /* limit retransmission to the bytes already sent on some subflows */
1917 info.limit = dfrag->already_sent;
1918 while (info.sent < dfrag->already_sent) {
1919 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1923 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1927 if (!mptcp_ext_cache_refill(msk))
1931 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1934 mptcp_set_timeout(sk, ssk);
1938 if (!mptcp_timer_pending(sk))
1939 mptcp_reset_timer(sk);
1942 clear_bit(MPTCP_WORKER_RUNNING, &msk->flags);
1947 static int __mptcp_init_sock(struct sock *sk)
1949 struct mptcp_sock *msk = mptcp_sk(sk);
1951 spin_lock_init(&msk->join_list_lock);
1953 INIT_LIST_HEAD(&msk->conn_list);
1954 INIT_LIST_HEAD(&msk->join_list);
1955 INIT_LIST_HEAD(&msk->rtx_queue);
1956 INIT_WORK(&msk->work, mptcp_worker);
1957 msk->out_of_order_queue = RB_ROOT;
1958 msk->first_pending = NULL;
1961 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1963 mptcp_pm_data_init(msk);
1965 /* re-use the csk retrans timer for MPTCP-level retrans */
1966 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1967 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
1971 static int mptcp_init_sock(struct sock *sk)
1973 struct net *net = sock_net(sk);
1976 ret = __mptcp_init_sock(sk);
1980 if (!mptcp_is_enabled(net))
1981 return -ENOPROTOOPT;
1983 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1986 ret = __mptcp_socket_create(mptcp_sk(sk));
1990 sk_sockets_allocated_inc(sk);
1991 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1992 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
1997 static void __mptcp_clear_xmit(struct sock *sk)
1999 struct mptcp_sock *msk = mptcp_sk(sk);
2000 struct mptcp_data_frag *dtmp, *dfrag;
2002 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2004 WRITE_ONCE(msk->first_pending, NULL);
2005 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2006 dfrag_clear(sk, dfrag);
2009 static void mptcp_cancel_work(struct sock *sk)
2011 struct mptcp_sock *msk = mptcp_sk(sk);
2013 /* if called by the work itself, do not try to cancel the work, or
2016 if (!test_bit(MPTCP_WORKER_RUNNING, &msk->flags) &&
2017 cancel_work_sync(&msk->work))
2021 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2025 switch (ssk->sk_state) {
2027 if (!(how & RCV_SHUTDOWN))
2031 tcp_disconnect(ssk, O_NONBLOCK);
2034 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2035 pr_debug("Fallback");
2036 ssk->sk_shutdown |= how;
2037 tcp_shutdown(ssk, how);
2039 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2040 mptcp_set_timeout(sk, ssk);
2049 static const unsigned char new_state[16] = {
2050 /* current state: new state: action: */
2051 [0 /* (Invalid) */] = TCP_CLOSE,
2052 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2053 [TCP_SYN_SENT] = TCP_CLOSE,
2054 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2055 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2056 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2057 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2058 [TCP_CLOSE] = TCP_CLOSE,
2059 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2060 [TCP_LAST_ACK] = TCP_LAST_ACK,
2061 [TCP_LISTEN] = TCP_CLOSE,
2062 [TCP_CLOSING] = TCP_CLOSING,
2063 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2066 static int mptcp_close_state(struct sock *sk)
2068 int next = (int)new_state[sk->sk_state];
2069 int ns = next & TCP_STATE_MASK;
2071 inet_sk_state_store(sk, ns);
2073 return next & TCP_ACTION_FIN;
2076 static void __mptcp_check_send_data_fin(struct sock *sk)
2078 struct mptcp_subflow_context *subflow;
2079 struct mptcp_sock *msk = mptcp_sk(sk);
2081 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2082 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2083 msk->snd_nxt, msk->write_seq);
2085 /* we still need to enqueue subflows or not really shutting down,
2088 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2089 mptcp_send_head(sk))
2092 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2094 /* fallback socket will not get data_fin/ack, can move to close now */
2095 if (__mptcp_check_fallback(msk) && sk->sk_state == TCP_LAST_ACK) {
2096 inet_sk_state_store(sk, TCP_CLOSE);
2097 mptcp_close_wake_up(sk);
2100 __mptcp_flush_join_list(msk);
2101 mptcp_for_each_subflow(msk, subflow) {
2102 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2104 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2108 static void __mptcp_wr_shutdown(struct sock *sk)
2110 struct mptcp_sock *msk = mptcp_sk(sk);
2112 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2113 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2114 !!mptcp_send_head(sk));
2116 /* will be ignored by fallback sockets */
2117 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2118 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2120 __mptcp_check_send_data_fin(sk);
2123 static void __mptcp_destroy_sock(struct sock *sk)
2125 struct mptcp_subflow_context *subflow, *tmp;
2126 struct mptcp_sock *msk = mptcp_sk(sk);
2127 LIST_HEAD(conn_list);
2129 pr_debug("msk=%p", msk);
2131 /* be sure to always acquire the join list lock, to sync vs
2132 * mptcp_finish_join().
2134 spin_lock_bh(&msk->join_list_lock);
2135 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2136 spin_unlock_bh(&msk->join_list_lock);
2137 list_splice_init(&msk->conn_list, &conn_list);
2139 __mptcp_clear_xmit(sk);
2140 sk_stop_timer(sk, &sk->sk_timer);
2143 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2144 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2145 __mptcp_close_ssk(sk, ssk, subflow);
2148 sk->sk_prot->destroy(sk);
2150 sk_stream_kill_queues(sk);
2151 xfrm_sk_free_policy(sk);
2152 sk_refcnt_debug_release(sk);
2156 static void mptcp_close(struct sock *sk, long timeout)
2158 struct mptcp_subflow_context *subflow;
2159 bool do_cancel_work = false;
2162 sk->sk_shutdown = SHUTDOWN_MASK;
2164 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2165 inet_sk_state_store(sk, TCP_CLOSE);
2169 if (mptcp_close_state(sk))
2170 __mptcp_wr_shutdown(sk);
2172 sk_stream_wait_close(sk, timeout);
2175 /* orphan all the subflows */
2176 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2177 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2178 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2179 bool slow, dispose_socket;
2180 struct socket *sock;
2182 slow = lock_sock_fast(ssk);
2183 sock = ssk->sk_socket;
2184 dispose_socket = sock && sock != sk->sk_socket;
2186 unlock_sock_fast(ssk, slow);
2188 /* for the outgoing subflows we additionally need to free
2189 * the associated socket
2192 iput(SOCK_INODE(sock));
2197 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2198 if (sk->sk_state == TCP_CLOSE) {
2199 __mptcp_destroy_sock(sk);
2200 do_cancel_work = true;
2202 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2206 mptcp_cancel_work(sk);
2210 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2212 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2213 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2214 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2216 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2217 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2220 msk6->saddr = ssk6->saddr;
2221 msk6->flow_label = ssk6->flow_label;
2225 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2226 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2227 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2228 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2229 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2230 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2233 static int mptcp_disconnect(struct sock *sk, int flags)
2235 /* Should never be called.
2236 * inet_stream_connect() calls ->disconnect, but that
2237 * refers to the subflow socket, not the mptcp one.
2243 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2244 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2246 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2248 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2252 struct sock *mptcp_sk_clone(const struct sock *sk,
2253 const struct mptcp_options_received *mp_opt,
2254 struct request_sock *req)
2256 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2257 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2258 struct mptcp_sock *msk;
2264 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2265 if (nsk->sk_family == AF_INET6)
2266 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2269 __mptcp_init_sock(nsk);
2271 msk = mptcp_sk(nsk);
2272 msk->local_key = subflow_req->local_key;
2273 msk->token = subflow_req->token;
2274 msk->subflow = NULL;
2275 WRITE_ONCE(msk->fully_established, false);
2277 msk->write_seq = subflow_req->idsn + 1;
2278 msk->snd_nxt = msk->write_seq;
2279 atomic64_set(&msk->snd_una, msk->write_seq);
2280 atomic64_set(&msk->wnd_end, msk->snd_nxt + req->rsk_rcv_wnd);
2282 if (mp_opt->mp_capable) {
2283 msk->can_ack = true;
2284 msk->remote_key = mp_opt->sndr_key;
2285 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2287 WRITE_ONCE(msk->ack_seq, ack_seq);
2290 sock_reset_flag(nsk, SOCK_RCU_FREE);
2291 /* will be fully established after successful MPC subflow creation */
2292 inet_sk_state_store(nsk, TCP_SYN_RECV);
2293 bh_unlock_sock(nsk);
2295 /* keep a single reference */
2300 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2302 const struct tcp_sock *tp = tcp_sk(ssk);
2304 msk->rcvq_space.copied = 0;
2305 msk->rcvq_space.rtt_us = 0;
2307 msk->rcvq_space.time = tp->tcp_mstamp;
2309 /* initial rcv_space offering made to peer */
2310 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2311 TCP_INIT_CWND * tp->advmss);
2312 if (msk->rcvq_space.space == 0)
2313 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2315 atomic64_set(&msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2318 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2321 struct mptcp_sock *msk = mptcp_sk(sk);
2322 struct socket *listener;
2325 listener = __mptcp_nmpc_socket(msk);
2326 if (WARN_ON_ONCE(!listener)) {
2331 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2332 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2336 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2337 if (sk_is_mptcp(newsk)) {
2338 struct mptcp_subflow_context *subflow;
2339 struct sock *new_mptcp_sock;
2340 struct sock *ssk = newsk;
2342 subflow = mptcp_subflow_ctx(newsk);
2343 new_mptcp_sock = subflow->conn;
2345 /* is_mptcp should be false if subflow->conn is missing, see
2346 * subflow_syn_recv_sock()
2348 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2349 tcp_sk(newsk)->is_mptcp = 0;
2353 /* acquire the 2nd reference for the owning socket */
2354 sock_hold(new_mptcp_sock);
2357 bh_lock_sock(new_mptcp_sock);
2358 msk = mptcp_sk(new_mptcp_sock);
2361 newsk = new_mptcp_sock;
2362 mptcp_copy_inaddrs(newsk, ssk);
2363 list_add(&subflow->node, &msk->conn_list);
2366 mptcp_rcv_space_init(msk, ssk);
2367 bh_unlock_sock(new_mptcp_sock);
2369 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2372 MPTCP_INC_STATS(sock_net(sk),
2373 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2379 void mptcp_destroy_common(struct mptcp_sock *msk)
2381 skb_rbtree_purge(&msk->out_of_order_queue);
2382 mptcp_token_destroy(msk);
2383 mptcp_pm_free_anno_list(msk);
2386 static void mptcp_destroy(struct sock *sk)
2388 struct mptcp_sock *msk = mptcp_sk(sk);
2390 if (msk->cached_ext)
2391 __skb_ext_put(msk->cached_ext);
2393 mptcp_destroy_common(msk);
2394 sk_sockets_allocated_dec(sk);
2397 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2398 sockptr_t optval, unsigned int optlen)
2400 struct sock *sk = (struct sock *)msk;
2401 struct socket *ssock;
2408 ssock = __mptcp_nmpc_socket(msk);
2414 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2416 if (optname == SO_REUSEPORT)
2417 sk->sk_reuseport = ssock->sk->sk_reuseport;
2418 else if (optname == SO_REUSEADDR)
2419 sk->sk_reuse = ssock->sk->sk_reuse;
2425 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2428 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2429 sockptr_t optval, unsigned int optlen)
2431 struct sock *sk = (struct sock *)msk;
2432 int ret = -EOPNOTSUPP;
2433 struct socket *ssock;
2438 ssock = __mptcp_nmpc_socket(msk);
2444 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2446 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2455 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2456 sockptr_t optval, unsigned int optlen)
2458 struct mptcp_sock *msk = mptcp_sk(sk);
2461 pr_debug("msk=%p", msk);
2463 if (level == SOL_SOCKET)
2464 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2466 /* @@ the meaning of setsockopt() when the socket is connected and
2467 * there are multiple subflows is not yet defined. It is up to the
2468 * MPTCP-level socket to configure the subflows until the subflow
2469 * is in TCP fallback, when TCP socket options are passed through
2470 * to the one remaining subflow.
2473 ssk = __mptcp_tcp_fallback(msk);
2476 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2478 if (level == SOL_IPV6)
2479 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2484 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2485 char __user *optval, int __user *option)
2487 struct mptcp_sock *msk = mptcp_sk(sk);
2490 pr_debug("msk=%p", msk);
2492 /* @@ the meaning of setsockopt() when the socket is connected and
2493 * there are multiple subflows is not yet defined. It is up to the
2494 * MPTCP-level socket to configure the subflows until the subflow
2495 * is in TCP fallback, when socket options are passed through
2496 * to the one remaining subflow.
2499 ssk = __mptcp_tcp_fallback(msk);
2502 return tcp_getsockopt(ssk, level, optname, optval, option);
2507 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
2508 TCPF_WRITE_TIMER_DEFERRED)
2510 /* this is very alike tcp_release_cb() but we must handle differently a
2511 * different set of events
2513 static void mptcp_release_cb(struct sock *sk)
2515 unsigned long flags, nflags;
2518 flags = sk->sk_tsq_flags;
2519 if (!(flags & MPTCP_DEFERRED_ALL))
2521 nflags = flags & ~MPTCP_DEFERRED_ALL;
2522 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2524 sock_release_ownership(sk);
2526 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
2527 struct mptcp_sock *msk = mptcp_sk(sk);
2530 ssk = mptcp_subflow_recv_lookup(msk);
2531 if (!ssk || sk->sk_state == TCP_CLOSE ||
2532 !schedule_work(&msk->work))
2536 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2537 mptcp_retransmit_handler(sk);
2542 static int mptcp_hash(struct sock *sk)
2544 /* should never be called,
2545 * we hash the TCP subflows not the master socket
2551 static void mptcp_unhash(struct sock *sk)
2553 /* called from sk_common_release(), but nothing to do here */
2556 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2558 struct mptcp_sock *msk = mptcp_sk(sk);
2559 struct socket *ssock;
2561 ssock = __mptcp_nmpc_socket(msk);
2562 pr_debug("msk=%p, subflow=%p", msk, ssock);
2563 if (WARN_ON_ONCE(!ssock))
2566 return inet_csk_get_port(ssock->sk, snum);
2569 void mptcp_finish_connect(struct sock *ssk)
2571 struct mptcp_subflow_context *subflow;
2572 struct mptcp_sock *msk;
2576 subflow = mptcp_subflow_ctx(ssk);
2580 pr_debug("msk=%p, token=%u", sk, subflow->token);
2582 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2584 subflow->map_seq = ack_seq;
2585 subflow->map_subflow_seq = 1;
2587 /* the socket is not connected yet, no msk/subflow ops can access/race
2588 * accessing the field below
2590 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2591 WRITE_ONCE(msk->local_key, subflow->local_key);
2592 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2593 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2594 WRITE_ONCE(msk->ack_seq, ack_seq);
2595 WRITE_ONCE(msk->can_ack, 1);
2596 atomic64_set(&msk->snd_una, msk->write_seq);
2598 mptcp_pm_new_connection(msk, 0);
2600 mptcp_rcv_space_init(msk, ssk);
2603 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2605 write_lock_bh(&sk->sk_callback_lock);
2606 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2607 sk_set_socket(sk, parent);
2608 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2609 write_unlock_bh(&sk->sk_callback_lock);
2612 bool mptcp_finish_join(struct sock *ssk)
2614 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2615 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2616 struct sock *parent = (void *)msk;
2617 struct socket *parent_sock;
2620 pr_debug("msk=%p, subflow=%p", msk, subflow);
2622 /* mptcp socket already closing? */
2623 if (!mptcp_is_fully_established(parent))
2626 if (!msk->pm.server_side)
2629 if (!mptcp_pm_allow_new_subflow(msk))
2632 /* active connections are already on conn_list, and we can't acquire
2634 * use the join list lock as synchronization point and double-check
2635 * msk status to avoid racing with __mptcp_destroy_sock()
2637 spin_lock_bh(&msk->join_list_lock);
2638 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2639 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
2640 list_add_tail(&subflow->node, &msk->join_list);
2643 spin_unlock_bh(&msk->join_list_lock);
2647 /* attach to msk socket only after we are sure he will deal with us
2650 parent_sock = READ_ONCE(parent->sk_socket);
2651 if (parent_sock && !ssk->sk_socket)
2652 mptcp_sock_graft(ssk, parent_sock);
2653 subflow->map_seq = READ_ONCE(msk->ack_seq);
2657 static struct proto mptcp_prot = {
2659 .owner = THIS_MODULE,
2660 .init = mptcp_init_sock,
2661 .disconnect = mptcp_disconnect,
2662 .close = mptcp_close,
2663 .accept = mptcp_accept,
2664 .setsockopt = mptcp_setsockopt,
2665 .getsockopt = mptcp_getsockopt,
2666 .shutdown = tcp_shutdown,
2667 .destroy = mptcp_destroy,
2668 .sendmsg = mptcp_sendmsg,
2669 .recvmsg = mptcp_recvmsg,
2670 .release_cb = mptcp_release_cb,
2672 .unhash = mptcp_unhash,
2673 .get_port = mptcp_get_port,
2674 .sockets_allocated = &mptcp_sockets_allocated,
2675 .memory_allocated = &tcp_memory_allocated,
2676 .memory_pressure = &tcp_memory_pressure,
2677 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2678 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
2679 .sysctl_mem = sysctl_tcp_mem,
2680 .obj_size = sizeof(struct mptcp_sock),
2681 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2682 .no_autobind = true,
2685 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2687 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2688 struct socket *ssock;
2691 lock_sock(sock->sk);
2692 ssock = __mptcp_nmpc_socket(msk);
2698 err = ssock->ops->bind(ssock, uaddr, addr_len);
2700 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2703 release_sock(sock->sk);
2707 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2708 struct mptcp_subflow_context *subflow)
2710 subflow->request_mptcp = 0;
2711 __mptcp_do_fallback(msk);
2714 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2715 int addr_len, int flags)
2717 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2718 struct mptcp_subflow_context *subflow;
2719 struct socket *ssock;
2722 lock_sock(sock->sk);
2723 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2724 /* pending connection or invalid state, let existing subflow
2727 ssock = msk->subflow;
2731 ssock = __mptcp_nmpc_socket(msk);
2737 mptcp_token_destroy(msk);
2738 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2739 subflow = mptcp_subflow_ctx(ssock->sk);
2740 #ifdef CONFIG_TCP_MD5SIG
2741 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2744 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2745 mptcp_subflow_early_fallback(msk, subflow);
2747 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2748 mptcp_subflow_early_fallback(msk, subflow);
2751 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2752 sock->state = ssock->state;
2754 /* on successful connect, the msk state will be moved to established by
2755 * subflow_finish_connect()
2757 if (!err || err == -EINPROGRESS)
2758 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2760 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2763 release_sock(sock->sk);
2767 static int mptcp_listen(struct socket *sock, int backlog)
2769 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2770 struct socket *ssock;
2773 pr_debug("msk=%p", msk);
2775 lock_sock(sock->sk);
2776 ssock = __mptcp_nmpc_socket(msk);
2782 mptcp_token_destroy(msk);
2783 inet_sk_state_store(sock->sk, TCP_LISTEN);
2784 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2786 err = ssock->ops->listen(ssock, backlog);
2787 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2789 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2792 release_sock(sock->sk);
2796 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2797 int flags, bool kern)
2799 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2800 struct socket *ssock;
2803 pr_debug("msk=%p", msk);
2805 lock_sock(sock->sk);
2806 if (sock->sk->sk_state != TCP_LISTEN)
2809 ssock = __mptcp_nmpc_socket(msk);
2813 clear_bit(MPTCP_DATA_READY, &msk->flags);
2814 sock_hold(ssock->sk);
2815 release_sock(sock->sk);
2817 err = ssock->ops->accept(sock, newsock, flags, kern);
2818 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2819 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2820 struct mptcp_subflow_context *subflow;
2822 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2823 * This is needed so NOSPACE flag can be set from tcp stack.
2825 __mptcp_flush_join_list(msk);
2826 mptcp_for_each_subflow(msk, subflow) {
2827 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2829 if (!ssk->sk_socket)
2830 mptcp_sock_graft(ssk, newsock);
2834 if (inet_csk_listen_poll(ssock->sk))
2835 set_bit(MPTCP_DATA_READY, &msk->flags);
2836 sock_put(ssock->sk);
2840 release_sock(sock->sk);
2844 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2846 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2850 static bool __mptcp_check_writeable(struct mptcp_sock *msk)
2852 struct sock *sk = (struct sock *)msk;
2853 bool mptcp_writable;
2855 mptcp_clean_una(sk);
2856 mptcp_writable = sk_stream_is_writeable(sk);
2857 if (!mptcp_writable)
2860 return mptcp_writable;
2863 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
2865 struct sock *sk = (struct sock *)msk;
2869 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
2872 if (sk_stream_is_writeable(sk))
2873 return EPOLLOUT | EPOLLWRNORM;
2875 slow = lock_sock_fast(sk);
2876 if (__mptcp_check_writeable(msk))
2877 ret = EPOLLOUT | EPOLLWRNORM;
2879 unlock_sock_fast(sk, slow);
2883 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2884 struct poll_table_struct *wait)
2886 struct sock *sk = sock->sk;
2887 struct mptcp_sock *msk;
2892 sock_poll_wait(file, sock, wait);
2894 state = inet_sk_state_load(sk);
2895 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
2896 if (state == TCP_LISTEN)
2897 return mptcp_check_readable(msk);
2899 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2900 mask |= mptcp_check_readable(msk);
2901 mask |= mptcp_check_writeable(msk);
2903 if (sk->sk_shutdown & RCV_SHUTDOWN)
2904 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2909 static int mptcp_shutdown(struct socket *sock, int how)
2911 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2912 struct sock *sk = sock->sk;
2915 pr_debug("sk=%p, how=%d", msk, how);
2920 if ((how & ~SHUTDOWN_MASK) || !how) {
2925 if (sock->state == SS_CONNECTING) {
2926 if ((1 << sk->sk_state) &
2927 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2928 sock->state = SS_DISCONNECTING;
2930 sock->state = SS_CONNECTED;
2933 sk->sk_shutdown |= how;
2934 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
2935 __mptcp_wr_shutdown(sk);
2937 /* Wake up anyone sleeping in poll. */
2938 sk->sk_state_change(sk);
2946 static const struct proto_ops mptcp_stream_ops = {
2948 .owner = THIS_MODULE,
2949 .release = inet_release,
2951 .connect = mptcp_stream_connect,
2952 .socketpair = sock_no_socketpair,
2953 .accept = mptcp_stream_accept,
2954 .getname = inet_getname,
2956 .ioctl = inet_ioctl,
2957 .gettstamp = sock_gettstamp,
2958 .listen = mptcp_listen,
2959 .shutdown = mptcp_shutdown,
2960 .setsockopt = sock_common_setsockopt,
2961 .getsockopt = sock_common_getsockopt,
2962 .sendmsg = inet_sendmsg,
2963 .recvmsg = inet_recvmsg,
2964 .mmap = sock_no_mmap,
2965 .sendpage = inet_sendpage,
2968 static struct inet_protosw mptcp_protosw = {
2969 .type = SOCK_STREAM,
2970 .protocol = IPPROTO_MPTCP,
2971 .prot = &mptcp_prot,
2972 .ops = &mptcp_stream_ops,
2973 .flags = INET_PROTOSW_ICSK,
2976 void __init mptcp_proto_init(void)
2978 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2980 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2981 panic("Failed to allocate MPTCP pcpu counter\n");
2983 mptcp_subflow_init();
2987 if (proto_register(&mptcp_prot, 1) != 0)
2988 panic("Failed to register MPTCP proto.\n");
2990 inet_register_protosw(&mptcp_protosw);
2992 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2995 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2996 static const struct proto_ops mptcp_v6_stream_ops = {
2998 .owner = THIS_MODULE,
2999 .release = inet6_release,
3001 .connect = mptcp_stream_connect,
3002 .socketpair = sock_no_socketpair,
3003 .accept = mptcp_stream_accept,
3004 .getname = inet6_getname,
3006 .ioctl = inet6_ioctl,
3007 .gettstamp = sock_gettstamp,
3008 .listen = mptcp_listen,
3009 .shutdown = mptcp_shutdown,
3010 .setsockopt = sock_common_setsockopt,
3011 .getsockopt = sock_common_getsockopt,
3012 .sendmsg = inet6_sendmsg,
3013 .recvmsg = inet6_recvmsg,
3014 .mmap = sock_no_mmap,
3015 .sendpage = inet_sendpage,
3016 #ifdef CONFIG_COMPAT
3017 .compat_ioctl = inet6_compat_ioctl,
3021 static struct proto mptcp_v6_prot;
3023 static void mptcp_v6_destroy(struct sock *sk)
3026 inet6_destroy_sock(sk);
3029 static struct inet_protosw mptcp_v6_protosw = {
3030 .type = SOCK_STREAM,
3031 .protocol = IPPROTO_MPTCP,
3032 .prot = &mptcp_v6_prot,
3033 .ops = &mptcp_v6_stream_ops,
3034 .flags = INET_PROTOSW_ICSK,
3037 int __init mptcp_proto_v6_init(void)
3041 mptcp_v6_prot = mptcp_prot;
3042 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3043 mptcp_v6_prot.slab = NULL;
3044 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3045 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3047 err = proto_register(&mptcp_v6_prot, 1);
3051 err = inet6_register_protosw(&mptcp_v6_protosw);
3053 proto_unregister(&mptcp_v6_prot);