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>
27 #define MPTCP_SAME_STATE TCP_MAX_STATES
29 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
31 struct mptcp_sock msk;
40 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
42 static struct percpu_counter mptcp_sockets_allocated;
44 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
45 * completed yet or has failed, return the subflow socket.
46 * Otherwise return NULL.
48 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
50 if (!msk->subflow || READ_ONCE(msk->can_ack))
56 static bool mptcp_is_tcpsk(struct sock *sk)
58 struct socket *sock = sk->sk_socket;
60 if (unlikely(sk->sk_prot == &tcp_prot)) {
61 /* we are being invoked after mptcp_accept() has
62 * accepted a non-mp-capable flow: sk is a tcp_sk,
65 * Hand the socket over to tcp so all further socket ops
68 sock->ops = &inet_stream_ops;
70 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
71 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
72 sock->ops = &inet6_stream_ops;
80 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
82 sock_owned_by_me((const struct sock *)msk);
84 if (likely(!__mptcp_check_fallback(msk)))
90 static int __mptcp_socket_create(struct mptcp_sock *msk)
92 struct mptcp_subflow_context *subflow;
93 struct sock *sk = (struct sock *)msk;
97 err = mptcp_subflow_create_socket(sk, &ssock);
101 msk->first = ssock->sk;
102 msk->subflow = ssock;
103 subflow = mptcp_subflow_ctx(ssock->sk);
104 list_add(&subflow->node, &msk->conn_list);
105 subflow->request_mptcp = 1;
107 /* accept() will wait on first subflow sk_wq, and we always wakes up
110 RCU_INIT_POINTER(msk->first->sk_wq, &sk->sk_socket->wq);
115 static void __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
117 unsigned int offset, size_t copy_len)
119 struct sock *sk = (struct sock *)msk;
120 struct sk_buff *tail;
122 __skb_unlink(skb, &ssk->sk_receive_queue);
126 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
128 tail = skb_peek_tail(&sk->sk_receive_queue);
129 if (offset == 0 && tail) {
133 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
134 kfree_skb_partial(skb, fragstolen);
135 atomic_add(delta, &sk->sk_rmem_alloc);
136 sk_mem_charge(sk, delta);
141 skb_set_owner_r(skb, sk);
142 __skb_queue_tail(&sk->sk_receive_queue, skb);
143 MPTCP_SKB_CB(skb)->offset = offset;
146 static void mptcp_stop_timer(struct sock *sk)
148 struct inet_connection_sock *icsk = inet_csk(sk);
150 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
151 mptcp_sk(sk)->timer_ival = 0;
154 /* both sockets must be locked */
155 static bool mptcp_subflow_dsn_valid(const struct mptcp_sock *msk,
158 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
159 u64 dsn = mptcp_subflow_get_mapped_dsn(subflow);
161 /* revalidate data sequence number.
163 * mptcp_subflow_data_available() is usually called
164 * without msk lock. Its unlikely (but possible)
165 * that msk->ack_seq has been advanced since the last
166 * call found in-sequence data.
168 if (likely(dsn == msk->ack_seq))
171 subflow->data_avail = 0;
172 return mptcp_subflow_data_available(ssk);
175 static void mptcp_check_data_fin_ack(struct sock *sk)
177 struct mptcp_sock *msk = mptcp_sk(sk);
179 if (__mptcp_check_fallback(msk))
182 /* Look for an acknowledged DATA_FIN */
183 if (((1 << sk->sk_state) &
184 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
185 msk->write_seq == atomic64_read(&msk->snd_una)) {
186 mptcp_stop_timer(sk);
188 WRITE_ONCE(msk->snd_data_fin_enable, 0);
190 switch (sk->sk_state) {
192 inet_sk_state_store(sk, TCP_FIN_WAIT2);
193 sk->sk_state_change(sk);
197 inet_sk_state_store(sk, TCP_CLOSE);
198 sk->sk_state_change(sk);
202 if (sk->sk_shutdown == SHUTDOWN_MASK ||
203 sk->sk_state == TCP_CLOSE)
204 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
206 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
210 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
212 struct mptcp_sock *msk = mptcp_sk(sk);
214 if (READ_ONCE(msk->rcv_data_fin) &&
215 ((1 << sk->sk_state) &
216 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
217 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
219 if (msk->ack_seq == rcv_data_fin_seq) {
221 *seq = rcv_data_fin_seq;
230 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
232 long tout = ssk && inet_csk(ssk)->icsk_pending ?
233 inet_csk(ssk)->icsk_timeout - jiffies : 0;
236 tout = mptcp_sk(sk)->timer_ival;
237 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
240 static void mptcp_check_data_fin(struct sock *sk)
242 struct mptcp_sock *msk = mptcp_sk(sk);
243 u64 rcv_data_fin_seq;
245 if (__mptcp_check_fallback(msk) || !msk->first)
248 /* Need to ack a DATA_FIN received from a peer while this side
249 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
250 * msk->rcv_data_fin was set when parsing the incoming options
251 * at the subflow level and the msk lock was not held, so this
252 * is the first opportunity to act on the DATA_FIN and change
255 * If we are caught up to the sequence number of the incoming
256 * DATA_FIN, send the DATA_ACK now and do state transition. If
257 * not caught up, do nothing and let the recv code send DATA_ACK
261 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
262 struct mptcp_subflow_context *subflow;
264 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
265 WRITE_ONCE(msk->rcv_data_fin, 0);
267 sk->sk_shutdown |= RCV_SHUTDOWN;
268 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
269 set_bit(MPTCP_DATA_READY, &msk->flags);
271 switch (sk->sk_state) {
272 case TCP_ESTABLISHED:
273 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
276 inet_sk_state_store(sk, TCP_CLOSING);
279 inet_sk_state_store(sk, TCP_CLOSE);
280 // @@ Close subflows now?
283 /* Other states not expected */
288 mptcp_set_timeout(sk, NULL);
289 mptcp_for_each_subflow(msk, subflow) {
290 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
297 sk->sk_state_change(sk);
299 if (sk->sk_shutdown == SHUTDOWN_MASK ||
300 sk->sk_state == TCP_CLOSE)
301 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
303 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
307 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
311 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
312 struct sock *sk = (struct sock *)msk;
313 unsigned int moved = 0;
314 bool more_data_avail;
318 if (!mptcp_subflow_dsn_valid(msk, ssk)) {
325 u32 map_remaining, offset;
326 u32 seq = tp->copied_seq;
330 /* try to move as much data as available */
331 map_remaining = subflow->map_data_len -
332 mptcp_subflow_get_map_offset(subflow);
334 skb = skb_peek(&ssk->sk_receive_queue);
338 if (__mptcp_check_fallback(msk)) {
339 /* if we are running under the workqueue, TCP could have
340 * collapsed skbs between dummy map creation and now
341 * be sure to adjust the size
343 map_remaining = skb->len;
344 subflow->map_data_len = skb->len;
347 offset = seq - TCP_SKB_CB(skb)->seq;
348 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
354 if (offset < skb->len) {
355 size_t len = skb->len - offset;
360 __mptcp_move_skb(msk, ssk, skb, offset, len);
364 if (WARN_ON_ONCE(map_remaining < len))
368 sk_eat_skb(ssk, skb);
372 WRITE_ONCE(tp->copied_seq, seq);
373 more_data_avail = mptcp_subflow_data_available(ssk);
375 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf)) {
379 } while (more_data_avail);
383 /* If the moves have caught up with the DATA_FIN sequence number
384 * it's time to ack the DATA_FIN and change socket state, but
385 * this is not a good place to change state. Let the workqueue
388 if (mptcp_pending_data_fin(sk, NULL) &&
389 schedule_work(&msk->work))
395 /* In most cases we will be able to lock the mptcp socket. If its already
396 * owned, we need to defer to the work queue to avoid ABBA deadlock.
398 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
400 struct sock *sk = (struct sock *)msk;
401 unsigned int moved = 0;
403 if (READ_ONCE(sk->sk_lock.owned))
406 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
409 /* must re-check after taking the lock */
410 if (!READ_ONCE(sk->sk_lock.owned))
411 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
413 spin_unlock_bh(&sk->sk_lock.slock);
418 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
420 struct mptcp_sock *msk = mptcp_sk(sk);
422 set_bit(MPTCP_DATA_READY, &msk->flags);
424 if (atomic_read(&sk->sk_rmem_alloc) < READ_ONCE(sk->sk_rcvbuf) &&
425 move_skbs_to_msk(msk, ssk))
428 /* don't schedule if mptcp sk is (still) over limit */
429 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf))
432 /* mptcp socket is owned, release_cb should retry */
433 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
434 &sk->sk_tsq_flags)) {
437 /* need to try again, its possible release_cb() has already
438 * been called after the test_and_set_bit() above.
440 move_skbs_to_msk(msk, ssk);
443 sk->sk_data_ready(sk);
446 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
448 if (likely(list_empty(&msk->join_list)))
451 spin_lock_bh(&msk->join_list_lock);
452 list_splice_tail_init(&msk->join_list, &msk->conn_list);
453 spin_unlock_bh(&msk->join_list_lock);
456 static bool mptcp_timer_pending(struct sock *sk)
458 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
461 static void mptcp_reset_timer(struct sock *sk)
463 struct inet_connection_sock *icsk = inet_csk(sk);
466 /* should never be called with mptcp level timer cleared */
467 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
468 if (WARN_ON_ONCE(!tout))
470 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
473 void mptcp_data_acked(struct sock *sk)
475 mptcp_reset_timer(sk);
477 if ((!sk_stream_is_writeable(sk) ||
478 (inet_sk_state_load(sk) != TCP_ESTABLISHED)) &&
479 schedule_work(&mptcp_sk(sk)->work))
483 void mptcp_subflow_eof(struct sock *sk)
485 struct mptcp_sock *msk = mptcp_sk(sk);
487 if (!test_and_set_bit(MPTCP_WORK_EOF, &msk->flags) &&
488 schedule_work(&msk->work))
492 static void mptcp_check_for_eof(struct mptcp_sock *msk)
494 struct mptcp_subflow_context *subflow;
495 struct sock *sk = (struct sock *)msk;
498 mptcp_for_each_subflow(msk, subflow)
499 receivers += !subflow->rx_eof;
501 if (!receivers && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
502 /* hopefully temporary hack: propagate shutdown status
503 * to msk, when all subflows agree on it
505 sk->sk_shutdown |= RCV_SHUTDOWN;
507 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
508 set_bit(MPTCP_DATA_READY, &msk->flags);
509 sk->sk_data_ready(sk);
513 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
515 const struct sock *sk = (const struct sock *)msk;
517 if (!msk->cached_ext)
518 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
520 return !!msk->cached_ext;
523 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
525 struct mptcp_subflow_context *subflow;
526 struct sock *sk = (struct sock *)msk;
528 sock_owned_by_me(sk);
530 mptcp_for_each_subflow(msk, subflow) {
531 if (subflow->data_avail)
532 return mptcp_subflow_tcp_sock(subflow);
538 static bool mptcp_skb_can_collapse_to(u64 write_seq,
539 const struct sk_buff *skb,
540 const struct mptcp_ext *mpext)
542 if (!tcp_skb_can_collapse_to(skb))
545 /* can collapse only if MPTCP level sequence is in order */
546 return mpext && mpext->data_seq + mpext->data_len == write_seq;
549 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
550 const struct page_frag *pfrag,
551 const struct mptcp_data_frag *df)
553 return df && pfrag->page == df->page &&
554 df->data_seq + df->data_len == msk->write_seq;
557 static void dfrag_uncharge(struct sock *sk, int len)
559 sk_mem_uncharge(sk, len);
560 sk_wmem_queued_add(sk, -len);
563 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
565 int len = dfrag->data_len + dfrag->overhead;
567 list_del(&dfrag->list);
568 dfrag_uncharge(sk, len);
569 put_page(dfrag->page);
572 static void mptcp_clean_una(struct sock *sk)
574 struct mptcp_sock *msk = mptcp_sk(sk);
575 struct mptcp_data_frag *dtmp, *dfrag;
576 bool cleaned = false;
579 /* on fallback we just need to ignore snd_una, as this is really
582 if (__mptcp_check_fallback(msk))
583 atomic64_set(&msk->snd_una, msk->write_seq);
584 snd_una = atomic64_read(&msk->snd_una);
586 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
587 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
590 dfrag_clear(sk, dfrag);
594 dfrag = mptcp_rtx_head(sk);
595 if (dfrag && after64(snd_una, dfrag->data_seq)) {
596 u64 delta = snd_una - dfrag->data_seq;
598 if (WARN_ON_ONCE(delta > dfrag->data_len))
601 dfrag->data_seq += delta;
602 dfrag->offset += delta;
603 dfrag->data_len -= delta;
605 dfrag_uncharge(sk, delta);
611 sk_mem_reclaim_partial(sk);
613 /* Only wake up writers if a subflow is ready */
614 if (test_bit(MPTCP_SEND_SPACE, &msk->flags))
615 sk_stream_write_space(sk);
619 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
622 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
624 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
625 pfrag, sk->sk_allocation)))
628 sk->sk_prot->enter_memory_pressure(sk);
629 sk_stream_moderate_sndbuf(sk);
633 static struct mptcp_data_frag *
634 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
637 int offset = ALIGN(orig_offset, sizeof(long));
638 struct mptcp_data_frag *dfrag;
640 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
642 dfrag->data_seq = msk->write_seq;
643 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
644 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
645 dfrag->page = pfrag->page;
650 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
651 struct msghdr *msg, struct mptcp_data_frag *dfrag,
652 long *timeo, int *pmss_now,
655 int mss_now, avail_size, size_goal, offset, ret, frag_truesize = 0;
656 bool dfrag_collapsed, can_collapse = false;
657 struct mptcp_sock *msk = mptcp_sk(sk);
658 struct mptcp_ext *mpext = NULL;
659 bool retransmission = !!dfrag;
660 struct sk_buff *skb, *tail;
661 struct page_frag *pfrag;
666 /* use the mptcp page cache so that we can easily move the data
667 * from one substream to another, but do per subflow memory accounting
668 * Note: pfrag is used only !retransmission, but the compiler if
669 * fooled into a warning if we don't init here
671 pfrag = sk_page_frag(sk);
672 if (!retransmission) {
673 write_seq = &msk->write_seq;
676 write_seq = &dfrag->data_seq;
680 /* compute copy limit */
681 mss_now = tcp_send_mss(ssk, &size_goal, msg->msg_flags);
683 *ps_goal = size_goal;
684 avail_size = size_goal;
685 skb = tcp_write_queue_tail(ssk);
687 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
689 /* Limit the write to the size available in the
690 * current skb, if any, so that we create at most a new skb.
691 * Explicitly tells TCP internals to avoid collapsing on later
692 * queue management operation, to avoid breaking the ext <->
693 * SSN association set here
695 can_collapse = (size_goal - skb->len > 0) &&
696 mptcp_skb_can_collapse_to(*write_seq, skb, mpext);
698 TCP_SKB_CB(skb)->eor = 1;
700 avail_size = size_goal - skb->len;
703 if (!retransmission) {
704 /* reuse tail pfrag, if possible, or carve a new one from the
707 dfrag = mptcp_rtx_tail(sk);
708 offset = pfrag->offset;
709 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
710 if (!dfrag_collapsed) {
711 dfrag = mptcp_carve_data_frag(msk, pfrag, offset);
712 offset = dfrag->offset;
713 frag_truesize = dfrag->overhead;
715 psize = min_t(size_t, pfrag->size - offset, avail_size);
718 pr_debug("left=%zu", msg_data_left(msg));
719 psize = copy_page_from_iter(pfrag->page, offset,
720 min_t(size_t, msg_data_left(msg),
723 pr_debug("left=%zu", msg_data_left(msg));
727 if (!sk_wmem_schedule(sk, psize + dfrag->overhead)) {
728 iov_iter_revert(&msg->msg_iter, psize);
732 offset = dfrag->offset;
733 psize = min_t(size_t, dfrag->data_len, avail_size);
736 /* tell the TCP stack to delay the push so that we can safely
737 * access the skb after the sendpages call
739 ret = do_tcp_sendpages(ssk, page, offset, psize,
740 msg->msg_flags | MSG_SENDPAGE_NOTLAST | MSG_DONTWAIT);
743 iov_iter_revert(&msg->msg_iter, psize);
747 frag_truesize += ret;
748 if (!retransmission) {
749 if (unlikely(ret < psize))
750 iov_iter_revert(&msg->msg_iter, psize - ret);
752 /* send successful, keep track of sent data for mptcp-level
755 dfrag->data_len += ret;
756 if (!dfrag_collapsed) {
757 get_page(dfrag->page);
758 list_add_tail(&dfrag->list, &msk->rtx_queue);
759 sk_wmem_queued_add(sk, frag_truesize);
761 sk_wmem_queued_add(sk, ret);
764 /* charge data on mptcp rtx queue to the master socket
765 * Note: we charge such data both to sk and ssk
767 sk->sk_forward_alloc -= frag_truesize;
770 /* if the tail skb extension is still the cached one, collapsing
771 * really happened. Note: we can't check for 'same skb' as the sk_buff
772 * hdr on tail can be transmitted, freed and re-allocated by the
773 * do_tcp_sendpages() call
775 tail = tcp_write_queue_tail(ssk);
776 if (mpext && tail && mpext == skb_ext_find(tail, SKB_EXT_MPTCP)) {
777 WARN_ON_ONCE(!can_collapse);
778 mpext->data_len += ret;
782 skb = tcp_write_queue_tail(ssk);
783 mpext = __skb_ext_set(skb, SKB_EXT_MPTCP, msk->cached_ext);
784 msk->cached_ext = NULL;
786 memset(mpext, 0, sizeof(*mpext));
787 mpext->data_seq = *write_seq;
788 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
789 mpext->data_len = ret;
793 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
794 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
799 pfrag->offset += frag_truesize;
800 WRITE_ONCE(*write_seq, *write_seq + ret);
801 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
806 static void mptcp_nospace(struct mptcp_sock *msk, struct socket *sock)
808 clear_bit(MPTCP_SEND_SPACE, &msk->flags);
809 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
811 /* enables sk->write_space() callbacks */
812 set_bit(SOCK_NOSPACE, &sock->flags);
815 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
817 struct mptcp_subflow_context *subflow;
818 struct sock *backup = NULL;
820 sock_owned_by_me((const struct sock *)msk);
822 if (!mptcp_ext_cache_refill(msk))
825 mptcp_for_each_subflow(msk, subflow) {
826 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
828 if (!sk_stream_memory_free(ssk)) {
829 struct socket *sock = ssk->sk_socket;
832 mptcp_nospace(msk, sock);
837 if (subflow->backup) {
850 static void ssk_check_wmem(struct mptcp_sock *msk, struct sock *ssk)
854 if (likely(sk_stream_is_writeable(ssk)))
857 sock = READ_ONCE(ssk->sk_socket);
859 mptcp_nospace(msk, sock);
862 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
864 int mss_now = 0, size_goal = 0, ret = 0;
865 struct mptcp_sock *msk = mptcp_sk(sk);
866 struct page_frag *pfrag;
872 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
877 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
879 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
880 ret = sk_stream_wait_connect(sk, &timeo);
885 pfrag = sk_page_frag(sk);
889 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
894 __mptcp_flush_join_list(msk);
895 ssk = mptcp_subflow_get_send(msk);
896 while (!sk_stream_memory_free(sk) ||
898 !mptcp_page_frag_refill(ssk, pfrag)) {
900 /* make sure retransmit timer is
901 * running before we wait for memory.
903 * The retransmit timer might be needed
904 * to make the peer send an up-to-date
907 mptcp_set_timeout(sk, ssk);
908 if (!mptcp_timer_pending(sk))
909 mptcp_reset_timer(sk);
912 ret = sk_stream_wait_memory(sk, &timeo);
918 ssk = mptcp_subflow_get_send(msk);
919 if (list_empty(&msk->conn_list)) {
925 pr_debug("conn_list->subflow=%p", ssk);
928 tx_ok = msg_data_left(msg);
930 ret = mptcp_sendmsg_frag(sk, ssk, msg, NULL, &timeo, &mss_now,
933 if (ret == -EAGAIN && timeo > 0) {
934 mptcp_set_timeout(sk, ssk);
943 tx_ok = msg_data_left(msg);
947 if (!sk_stream_memory_free(ssk) ||
948 !mptcp_page_frag_refill(ssk, pfrag) ||
949 !mptcp_ext_cache_refill(msk)) {
950 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
951 tcp_push(ssk, msg->msg_flags, mss_now,
952 tcp_sk(ssk)->nonagle, size_goal);
953 mptcp_set_timeout(sk, ssk);
958 /* memory is charged to mptcp level socket as well, i.e.
959 * if msg is very large, mptcp socket may run out of buffer
960 * space. mptcp_clean_una() will release data that has
961 * been acked at mptcp level in the mean time, so there is
962 * a good chance we can continue sending data right away.
964 * Normally, when the tcp subflow can accept more data, then
965 * so can the MPTCP socket. However, we need to cope with
966 * peers that might lag behind in their MPTCP-level
967 * acknowledgements, i.e. data might have been acked at
968 * tcp level only. So, we must also check the MPTCP socket
969 * limits before we send more data.
971 if (unlikely(!sk_stream_memory_free(sk))) {
972 tcp_push(ssk, msg->msg_flags, mss_now,
973 tcp_sk(ssk)->nonagle, size_goal);
975 if (!sk_stream_memory_free(sk)) {
976 /* can't send more for now, need to wait for
977 * MPTCP-level ACKs from peer.
979 * Wakeup will happen via mptcp_clean_una().
981 mptcp_set_timeout(sk, ssk);
988 mptcp_set_timeout(sk, ssk);
990 tcp_push(ssk, msg->msg_flags, mss_now, tcp_sk(ssk)->nonagle,
993 /* start the timer, if it's not pending */
994 if (!mptcp_timer_pending(sk))
995 mptcp_reset_timer(sk);
998 ssk_check_wmem(msk, ssk);
1002 return copied ? : ret;
1005 static void mptcp_wait_data(struct sock *sk, long *timeo)
1007 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1008 struct mptcp_sock *msk = mptcp_sk(sk);
1010 add_wait_queue(sk_sleep(sk), &wait);
1011 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1013 sk_wait_event(sk, timeo,
1014 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1016 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1017 remove_wait_queue(sk_sleep(sk), &wait);
1020 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1024 struct sock *sk = (struct sock *)msk;
1025 struct sk_buff *skb;
1028 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1029 u32 offset = MPTCP_SKB_CB(skb)->offset;
1030 u32 data_len = skb->len - offset;
1031 u32 count = min_t(size_t, len - copied, data_len);
1034 err = skb_copy_datagram_msg(skb, offset, msg, count);
1035 if (unlikely(err < 0)) {
1043 if (count < data_len) {
1044 MPTCP_SKB_CB(skb)->offset += count;
1048 __skb_unlink(skb, &sk->sk_receive_queue);
1058 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1060 * Only difference: Use highest rtt estimate of the subflows in use.
1062 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1064 struct mptcp_subflow_context *subflow;
1065 struct sock *sk = (struct sock *)msk;
1066 u32 time, advmss = 1;
1069 sock_owned_by_me(sk);
1074 msk->rcvq_space.copied += copied;
1076 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1077 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1079 rtt_us = msk->rcvq_space.rtt_us;
1080 if (rtt_us && time < (rtt_us >> 3))
1084 mptcp_for_each_subflow(msk, subflow) {
1085 const struct tcp_sock *tp;
1089 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1091 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1092 sf_advmss = READ_ONCE(tp->advmss);
1094 rtt_us = max(sf_rtt_us, rtt_us);
1095 advmss = max(sf_advmss, advmss);
1098 msk->rcvq_space.rtt_us = rtt_us;
1099 if (time < (rtt_us >> 3) || rtt_us == 0)
1102 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1105 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1106 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1110 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1112 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1114 do_div(grow, msk->rcvq_space.space);
1115 rcvwin += (grow << 1);
1117 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1118 while (tcp_win_from_space(sk, rcvmem) < advmss)
1121 do_div(rcvwin, advmss);
1122 rcvbuf = min_t(u64, rcvwin * rcvmem,
1123 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1125 if (rcvbuf > sk->sk_rcvbuf) {
1128 window_clamp = tcp_win_from_space(sk, rcvbuf);
1129 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1131 /* Make subflows follow along. If we do not do this, we
1132 * get drops at subflow level if skbs can't be moved to
1133 * the mptcp rx queue fast enough (announced rcv_win can
1134 * exceed ssk->sk_rcvbuf).
1136 mptcp_for_each_subflow(msk, subflow) {
1139 ssk = mptcp_subflow_tcp_sock(subflow);
1140 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1141 tcp_sk(ssk)->window_clamp = window_clamp;
1146 msk->rcvq_space.space = msk->rcvq_space.copied;
1148 msk->rcvq_space.copied = 0;
1149 msk->rcvq_space.time = mstamp;
1152 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1154 unsigned int moved = 0;
1158 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1164 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1171 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1172 int nonblock, int flags, int *addr_len)
1174 struct mptcp_sock *msk = mptcp_sk(sk);
1179 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1183 timeo = sock_rcvtimeo(sk, nonblock);
1185 len = min_t(size_t, len, INT_MAX);
1186 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1187 __mptcp_flush_join_list(msk);
1189 while (len > (size_t)copied) {
1192 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1193 if (unlikely(bytes_read < 0)) {
1195 copied = bytes_read;
1199 copied += bytes_read;
1201 if (skb_queue_empty(&sk->sk_receive_queue) &&
1202 __mptcp_move_skbs(msk))
1205 /* only the master socket status is relevant here. The exit
1206 * conditions mirror closely tcp_recvmsg()
1208 if (copied >= target)
1213 sk->sk_state == TCP_CLOSE ||
1214 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1216 signal_pending(current))
1220 copied = sock_error(sk);
1224 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1225 mptcp_check_for_eof(msk);
1227 if (sk->sk_shutdown & RCV_SHUTDOWN)
1230 if (sk->sk_state == TCP_CLOSE) {
1240 if (signal_pending(current)) {
1241 copied = sock_intr_errno(timeo);
1246 pr_debug("block timeout %ld", timeo);
1247 mptcp_wait_data(sk, &timeo);
1250 if (skb_queue_empty(&sk->sk_receive_queue)) {
1251 /* entire backlog drained, clear DATA_READY. */
1252 clear_bit(MPTCP_DATA_READY, &msk->flags);
1254 /* .. race-breaker: ssk might have gotten new data
1255 * after last __mptcp_move_skbs() returned false.
1257 if (unlikely(__mptcp_move_skbs(msk)))
1258 set_bit(MPTCP_DATA_READY, &msk->flags);
1259 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1260 /* data to read but mptcp_wait_data() cleared DATA_READY */
1261 set_bit(MPTCP_DATA_READY, &msk->flags);
1264 mptcp_rcv_space_adjust(msk, copied);
1270 static void mptcp_retransmit_handler(struct sock *sk)
1272 struct mptcp_sock *msk = mptcp_sk(sk);
1274 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->write_seq)) {
1275 mptcp_stop_timer(sk);
1277 set_bit(MPTCP_WORK_RTX, &msk->flags);
1278 if (schedule_work(&msk->work))
1283 static void mptcp_retransmit_timer(struct timer_list *t)
1285 struct inet_connection_sock *icsk = from_timer(icsk, t,
1286 icsk_retransmit_timer);
1287 struct sock *sk = &icsk->icsk_inet.sk;
1290 if (!sock_owned_by_user(sk)) {
1291 mptcp_retransmit_handler(sk);
1293 /* delegate our work to tcp_release_cb() */
1294 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1302 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1305 * A backup subflow is returned only if that is the only kind available.
1307 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1309 struct mptcp_subflow_context *subflow;
1310 struct sock *backup = NULL;
1312 sock_owned_by_me((const struct sock *)msk);
1314 mptcp_for_each_subflow(msk, subflow) {
1315 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1317 /* still data outstanding at TCP level? Don't retransmit. */
1318 if (!tcp_write_queue_empty(ssk))
1321 if (subflow->backup) {
1333 /* subflow sockets can be either outgoing (connect) or incoming
1336 * Outgoing subflows use in-kernel sockets.
1337 * Incoming subflows do not have their own 'struct socket' allocated,
1338 * so we need to use tcp_close() after detaching them from the mptcp
1341 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1342 struct mptcp_subflow_context *subflow,
1345 struct socket *sock = READ_ONCE(ssk->sk_socket);
1347 list_del(&subflow->node);
1349 if (sock && sock != sk->sk_socket) {
1350 /* outgoing subflow */
1353 /* incoming subflow */
1354 tcp_close(ssk, timeout);
1358 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1363 static void pm_work(struct mptcp_sock *msk)
1365 struct mptcp_pm_data *pm = &msk->pm;
1367 spin_lock_bh(&msk->pm.lock);
1369 pr_debug("msk=%p status=%x", msk, pm->status);
1370 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1371 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1372 mptcp_pm_nl_add_addr_received(msk);
1374 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1375 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1376 mptcp_pm_nl_fully_established(msk);
1378 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1379 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1380 mptcp_pm_nl_subflow_established(msk);
1383 spin_unlock_bh(&msk->pm.lock);
1386 static void mptcp_worker(struct work_struct *work)
1388 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1389 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1390 int orig_len, orig_offset, mss_now = 0, size_goal = 0;
1391 struct mptcp_data_frag *dfrag;
1394 struct msghdr msg = {
1395 .msg_flags = MSG_DONTWAIT,
1400 mptcp_clean_una(sk);
1401 mptcp_check_data_fin_ack(sk);
1402 __mptcp_flush_join_list(msk);
1403 __mptcp_move_skbs(msk);
1408 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1409 mptcp_check_for_eof(msk);
1411 mptcp_check_data_fin(sk);
1413 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1416 dfrag = mptcp_rtx_head(sk);
1420 if (!mptcp_ext_cache_refill(msk))
1423 ssk = mptcp_subflow_get_retrans(msk);
1429 orig_len = dfrag->data_len;
1430 orig_offset = dfrag->offset;
1431 orig_write_seq = dfrag->data_seq;
1432 while (dfrag->data_len > 0) {
1433 int ret = mptcp_sendmsg_frag(sk, ssk, &msg, dfrag, &timeo,
1434 &mss_now, &size_goal);
1438 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1440 dfrag->data_len -= ret;
1441 dfrag->offset += ret;
1443 if (!mptcp_ext_cache_refill(msk))
1447 tcp_push(ssk, msg.msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1450 dfrag->data_seq = orig_write_seq;
1451 dfrag->offset = orig_offset;
1452 dfrag->data_len = orig_len;
1454 mptcp_set_timeout(sk, ssk);
1458 if (!mptcp_timer_pending(sk))
1459 mptcp_reset_timer(sk);
1466 static int __mptcp_init_sock(struct sock *sk)
1468 struct mptcp_sock *msk = mptcp_sk(sk);
1470 spin_lock_init(&msk->join_list_lock);
1472 INIT_LIST_HEAD(&msk->conn_list);
1473 INIT_LIST_HEAD(&msk->join_list);
1474 INIT_LIST_HEAD(&msk->rtx_queue);
1475 __set_bit(MPTCP_SEND_SPACE, &msk->flags);
1476 INIT_WORK(&msk->work, mptcp_worker);
1479 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1481 mptcp_pm_data_init(msk);
1483 /* re-use the csk retrans timer for MPTCP-level retrans */
1484 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1489 static int mptcp_init_sock(struct sock *sk)
1491 struct net *net = sock_net(sk);
1494 if (!mptcp_is_enabled(net))
1495 return -ENOPROTOOPT;
1497 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1500 ret = __mptcp_init_sock(sk);
1504 ret = __mptcp_socket_create(mptcp_sk(sk));
1508 sk_sockets_allocated_inc(sk);
1509 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1510 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[2];
1515 static void __mptcp_clear_xmit(struct sock *sk)
1517 struct mptcp_sock *msk = mptcp_sk(sk);
1518 struct mptcp_data_frag *dtmp, *dfrag;
1520 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
1522 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
1523 dfrag_clear(sk, dfrag);
1526 static void mptcp_cancel_work(struct sock *sk)
1528 struct mptcp_sock *msk = mptcp_sk(sk);
1530 if (cancel_work_sync(&msk->work))
1534 static void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
1538 switch (ssk->sk_state) {
1540 if (!(how & RCV_SHUTDOWN))
1544 tcp_disconnect(ssk, O_NONBLOCK);
1547 if (__mptcp_check_fallback(mptcp_sk(sk))) {
1548 pr_debug("Fallback");
1549 ssk->sk_shutdown |= how;
1550 tcp_shutdown(ssk, how);
1552 pr_debug("Sending DATA_FIN on subflow %p", ssk);
1553 mptcp_set_timeout(sk, ssk);
1562 static const unsigned char new_state[16] = {
1563 /* current state: new state: action: */
1564 [0 /* (Invalid) */] = TCP_CLOSE,
1565 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1566 [TCP_SYN_SENT] = TCP_CLOSE,
1567 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1568 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1569 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1570 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
1571 [TCP_CLOSE] = TCP_CLOSE,
1572 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1573 [TCP_LAST_ACK] = TCP_LAST_ACK,
1574 [TCP_LISTEN] = TCP_CLOSE,
1575 [TCP_CLOSING] = TCP_CLOSING,
1576 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1579 static int mptcp_close_state(struct sock *sk)
1581 int next = (int)new_state[sk->sk_state];
1582 int ns = next & TCP_STATE_MASK;
1584 inet_sk_state_store(sk, ns);
1586 return next & TCP_ACTION_FIN;
1589 static void mptcp_close(struct sock *sk, long timeout)
1591 struct mptcp_subflow_context *subflow, *tmp;
1592 struct mptcp_sock *msk = mptcp_sk(sk);
1593 LIST_HEAD(conn_list);
1596 sk->sk_shutdown = SHUTDOWN_MASK;
1598 if (sk->sk_state == TCP_LISTEN) {
1599 inet_sk_state_store(sk, TCP_CLOSE);
1601 } else if (sk->sk_state == TCP_CLOSE) {
1605 if (__mptcp_check_fallback(msk)) {
1607 } else if (mptcp_close_state(sk)) {
1608 pr_debug("Sending DATA_FIN sk=%p", sk);
1609 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
1610 WRITE_ONCE(msk->snd_data_fin_enable, 1);
1612 mptcp_for_each_subflow(msk, subflow) {
1613 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
1615 mptcp_subflow_shutdown(sk, tcp_sk, SHUTDOWN_MASK);
1619 sk_stream_wait_close(sk, timeout);
1622 inet_sk_state_store(sk, TCP_CLOSE);
1625 /* be sure to always acquire the join list lock, to sync vs
1626 * mptcp_finish_join().
1628 spin_lock_bh(&msk->join_list_lock);
1629 list_splice_tail_init(&msk->join_list, &msk->conn_list);
1630 spin_unlock_bh(&msk->join_list_lock);
1631 list_splice_init(&msk->conn_list, &conn_list);
1633 __mptcp_clear_xmit(sk);
1637 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
1638 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1639 __mptcp_close_ssk(sk, ssk, subflow, timeout);
1642 mptcp_cancel_work(sk);
1644 __skb_queue_purge(&sk->sk_receive_queue);
1646 sk_common_release(sk);
1649 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
1651 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1652 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
1653 struct ipv6_pinfo *msk6 = inet6_sk(msk);
1655 msk->sk_v6_daddr = ssk->sk_v6_daddr;
1656 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
1659 msk6->saddr = ssk6->saddr;
1660 msk6->flow_label = ssk6->flow_label;
1664 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
1665 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
1666 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
1667 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
1668 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
1669 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
1672 static int mptcp_disconnect(struct sock *sk, int flags)
1674 /* Should never be called.
1675 * inet_stream_connect() calls ->disconnect, but that
1676 * refers to the subflow socket, not the mptcp one.
1682 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1683 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
1685 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
1687 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
1691 struct sock *mptcp_sk_clone(const struct sock *sk,
1692 const struct mptcp_options_received *mp_opt,
1693 struct request_sock *req)
1695 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
1696 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
1697 struct mptcp_sock *msk;
1703 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1704 if (nsk->sk_family == AF_INET6)
1705 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
1708 __mptcp_init_sock(nsk);
1710 msk = mptcp_sk(nsk);
1711 msk->local_key = subflow_req->local_key;
1712 msk->token = subflow_req->token;
1713 msk->subflow = NULL;
1714 WRITE_ONCE(msk->fully_established, false);
1716 msk->write_seq = subflow_req->idsn + 1;
1717 atomic64_set(&msk->snd_una, msk->write_seq);
1718 if (mp_opt->mp_capable) {
1719 msk->can_ack = true;
1720 msk->remote_key = mp_opt->sndr_key;
1721 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
1723 WRITE_ONCE(msk->ack_seq, ack_seq);
1726 sock_reset_flag(nsk, SOCK_RCU_FREE);
1727 /* will be fully established after successful MPC subflow creation */
1728 inet_sk_state_store(nsk, TCP_SYN_RECV);
1729 bh_unlock_sock(nsk);
1731 /* keep a single reference */
1736 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
1738 const struct tcp_sock *tp = tcp_sk(ssk);
1740 msk->rcvq_space.copied = 0;
1741 msk->rcvq_space.rtt_us = 0;
1743 msk->rcvq_space.time = tp->tcp_mstamp;
1745 /* initial rcv_space offering made to peer */
1746 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
1747 TCP_INIT_CWND * tp->advmss);
1748 if (msk->rcvq_space.space == 0)
1749 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
1752 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
1755 struct mptcp_sock *msk = mptcp_sk(sk);
1756 struct socket *listener;
1759 listener = __mptcp_nmpc_socket(msk);
1760 if (WARN_ON_ONCE(!listener)) {
1765 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
1766 newsk = inet_csk_accept(listener->sk, flags, err, kern);
1770 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
1771 if (sk_is_mptcp(newsk)) {
1772 struct mptcp_subflow_context *subflow;
1773 struct sock *new_mptcp_sock;
1774 struct sock *ssk = newsk;
1776 subflow = mptcp_subflow_ctx(newsk);
1777 new_mptcp_sock = subflow->conn;
1779 /* is_mptcp should be false if subflow->conn is missing, see
1780 * subflow_syn_recv_sock()
1782 if (WARN_ON_ONCE(!new_mptcp_sock)) {
1783 tcp_sk(newsk)->is_mptcp = 0;
1787 /* acquire the 2nd reference for the owning socket */
1788 sock_hold(new_mptcp_sock);
1791 bh_lock_sock(new_mptcp_sock);
1792 msk = mptcp_sk(new_mptcp_sock);
1795 newsk = new_mptcp_sock;
1796 mptcp_copy_inaddrs(newsk, ssk);
1797 list_add(&subflow->node, &msk->conn_list);
1799 mptcp_rcv_space_init(msk, ssk);
1800 bh_unlock_sock(new_mptcp_sock);
1802 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
1805 MPTCP_INC_STATS(sock_net(sk),
1806 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
1812 static void mptcp_destroy(struct sock *sk)
1814 struct mptcp_sock *msk = mptcp_sk(sk);
1816 mptcp_token_destroy(msk);
1817 if (msk->cached_ext)
1818 __skb_ext_put(msk->cached_ext);
1820 sk_sockets_allocated_dec(sk);
1823 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
1824 sockptr_t optval, unsigned int optlen)
1826 struct sock *sk = (struct sock *)msk;
1827 struct socket *ssock;
1834 ssock = __mptcp_nmpc_socket(msk);
1840 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
1842 if (optname == SO_REUSEPORT)
1843 sk->sk_reuseport = ssock->sk->sk_reuseport;
1844 else if (optname == SO_REUSEADDR)
1845 sk->sk_reuse = ssock->sk->sk_reuse;
1851 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
1854 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
1855 sockptr_t optval, unsigned int optlen)
1857 struct sock *sk = (struct sock *)msk;
1858 int ret = -EOPNOTSUPP;
1859 struct socket *ssock;
1864 ssock = __mptcp_nmpc_socket(msk);
1870 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
1872 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
1881 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
1882 sockptr_t optval, unsigned int optlen)
1884 struct mptcp_sock *msk = mptcp_sk(sk);
1887 pr_debug("msk=%p", msk);
1889 if (level == SOL_SOCKET)
1890 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
1892 /* @@ the meaning of setsockopt() when the socket is connected and
1893 * there are multiple subflows is not yet defined. It is up to the
1894 * MPTCP-level socket to configure the subflows until the subflow
1895 * is in TCP fallback, when TCP socket options are passed through
1896 * to the one remaining subflow.
1899 ssk = __mptcp_tcp_fallback(msk);
1902 return tcp_setsockopt(ssk, level, optname, optval, optlen);
1904 if (level == SOL_IPV6)
1905 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
1910 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
1911 char __user *optval, int __user *option)
1913 struct mptcp_sock *msk = mptcp_sk(sk);
1916 pr_debug("msk=%p", msk);
1918 /* @@ the meaning of setsockopt() when the socket is connected and
1919 * there are multiple subflows is not yet defined. It is up to the
1920 * MPTCP-level socket to configure the subflows until the subflow
1921 * is in TCP fallback, when socket options are passed through
1922 * to the one remaining subflow.
1925 ssk = __mptcp_tcp_fallback(msk);
1928 return tcp_getsockopt(ssk, level, optname, optval, option);
1933 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
1934 TCPF_WRITE_TIMER_DEFERRED)
1936 /* this is very alike tcp_release_cb() but we must handle differently a
1937 * different set of events
1939 static void mptcp_release_cb(struct sock *sk)
1941 unsigned long flags, nflags;
1944 flags = sk->sk_tsq_flags;
1945 if (!(flags & MPTCP_DEFERRED_ALL))
1947 nflags = flags & ~MPTCP_DEFERRED_ALL;
1948 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1950 sock_release_ownership(sk);
1952 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1953 struct mptcp_sock *msk = mptcp_sk(sk);
1956 ssk = mptcp_subflow_recv_lookup(msk);
1957 if (!ssk || !schedule_work(&msk->work))
1961 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1962 mptcp_retransmit_handler(sk);
1967 static int mptcp_hash(struct sock *sk)
1969 /* should never be called,
1970 * we hash the TCP subflows not the master socket
1976 static void mptcp_unhash(struct sock *sk)
1978 /* called from sk_common_release(), but nothing to do here */
1981 static int mptcp_get_port(struct sock *sk, unsigned short snum)
1983 struct mptcp_sock *msk = mptcp_sk(sk);
1984 struct socket *ssock;
1986 ssock = __mptcp_nmpc_socket(msk);
1987 pr_debug("msk=%p, subflow=%p", msk, ssock);
1988 if (WARN_ON_ONCE(!ssock))
1991 return inet_csk_get_port(ssock->sk, snum);
1994 void mptcp_finish_connect(struct sock *ssk)
1996 struct mptcp_subflow_context *subflow;
1997 struct mptcp_sock *msk;
2001 subflow = mptcp_subflow_ctx(ssk);
2005 pr_debug("msk=%p, token=%u", sk, subflow->token);
2007 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2009 subflow->map_seq = ack_seq;
2010 subflow->map_subflow_seq = 1;
2012 /* the socket is not connected yet, no msk/subflow ops can access/race
2013 * accessing the field below
2015 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2016 WRITE_ONCE(msk->local_key, subflow->local_key);
2017 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2018 WRITE_ONCE(msk->ack_seq, ack_seq);
2019 WRITE_ONCE(msk->can_ack, 1);
2020 atomic64_set(&msk->snd_una, msk->write_seq);
2022 mptcp_pm_new_connection(msk, 0);
2024 mptcp_rcv_space_init(msk, ssk);
2027 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2029 write_lock_bh(&sk->sk_callback_lock);
2030 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2031 sk_set_socket(sk, parent);
2032 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2033 write_unlock_bh(&sk->sk_callback_lock);
2036 bool mptcp_finish_join(struct sock *sk)
2038 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
2039 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2040 struct sock *parent = (void *)msk;
2041 struct socket *parent_sock;
2044 pr_debug("msk=%p, subflow=%p", msk, subflow);
2046 /* mptcp socket already closing? */
2047 if (!mptcp_is_fully_established(parent))
2050 if (!msk->pm.server_side)
2053 if (!mptcp_pm_allow_new_subflow(msk))
2056 /* active connections are already on conn_list, and we can't acquire
2058 * use the join list lock as synchronization point and double-check
2059 * msk status to avoid racing with mptcp_close()
2061 spin_lock_bh(&msk->join_list_lock);
2062 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2063 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node)))
2064 list_add_tail(&subflow->node, &msk->join_list);
2065 spin_unlock_bh(&msk->join_list_lock);
2069 /* attach to msk socket only after we are sure he will deal with us
2072 parent_sock = READ_ONCE(parent->sk_socket);
2073 if (parent_sock && !sk->sk_socket)
2074 mptcp_sock_graft(sk, parent_sock);
2075 subflow->map_seq = READ_ONCE(msk->ack_seq);
2079 static bool mptcp_memory_free(const struct sock *sk, int wake)
2081 struct mptcp_sock *msk = mptcp_sk(sk);
2083 return wake ? test_bit(MPTCP_SEND_SPACE, &msk->flags) : true;
2086 static struct proto mptcp_prot = {
2088 .owner = THIS_MODULE,
2089 .init = mptcp_init_sock,
2090 .disconnect = mptcp_disconnect,
2091 .close = mptcp_close,
2092 .accept = mptcp_accept,
2093 .setsockopt = mptcp_setsockopt,
2094 .getsockopt = mptcp_getsockopt,
2095 .shutdown = tcp_shutdown,
2096 .destroy = mptcp_destroy,
2097 .sendmsg = mptcp_sendmsg,
2098 .recvmsg = mptcp_recvmsg,
2099 .release_cb = mptcp_release_cb,
2101 .unhash = mptcp_unhash,
2102 .get_port = mptcp_get_port,
2103 .sockets_allocated = &mptcp_sockets_allocated,
2104 .memory_allocated = &tcp_memory_allocated,
2105 .memory_pressure = &tcp_memory_pressure,
2106 .stream_memory_free = mptcp_memory_free,
2107 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2108 .sysctl_mem = sysctl_tcp_mem,
2109 .obj_size = sizeof(struct mptcp_sock),
2110 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2111 .no_autobind = true,
2114 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2116 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2117 struct socket *ssock;
2120 lock_sock(sock->sk);
2121 ssock = __mptcp_nmpc_socket(msk);
2127 err = ssock->ops->bind(ssock, uaddr, addr_len);
2129 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2132 release_sock(sock->sk);
2136 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2137 struct mptcp_subflow_context *subflow)
2139 subflow->request_mptcp = 0;
2140 __mptcp_do_fallback(msk);
2143 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2144 int addr_len, int flags)
2146 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2147 struct mptcp_subflow_context *subflow;
2148 struct socket *ssock;
2151 lock_sock(sock->sk);
2152 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2153 /* pending connection or invalid state, let existing subflow
2156 ssock = msk->subflow;
2160 ssock = __mptcp_nmpc_socket(msk);
2166 mptcp_token_destroy(msk);
2167 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2168 subflow = mptcp_subflow_ctx(ssock->sk);
2169 #ifdef CONFIG_TCP_MD5SIG
2170 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2173 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2174 mptcp_subflow_early_fallback(msk, subflow);
2176 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2177 mptcp_subflow_early_fallback(msk, subflow);
2180 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2181 sock->state = ssock->state;
2183 /* on successful connect, the msk state will be moved to established by
2184 * subflow_finish_connect()
2186 if (!err || err == -EINPROGRESS)
2187 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2189 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2192 release_sock(sock->sk);
2196 static int mptcp_listen(struct socket *sock, int backlog)
2198 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2199 struct socket *ssock;
2202 pr_debug("msk=%p", msk);
2204 lock_sock(sock->sk);
2205 ssock = __mptcp_nmpc_socket(msk);
2211 mptcp_token_destroy(msk);
2212 inet_sk_state_store(sock->sk, TCP_LISTEN);
2213 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2215 err = ssock->ops->listen(ssock, backlog);
2216 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2218 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2221 release_sock(sock->sk);
2225 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2226 int flags, bool kern)
2228 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2229 struct socket *ssock;
2232 pr_debug("msk=%p", msk);
2234 lock_sock(sock->sk);
2235 if (sock->sk->sk_state != TCP_LISTEN)
2238 ssock = __mptcp_nmpc_socket(msk);
2242 clear_bit(MPTCP_DATA_READY, &msk->flags);
2243 sock_hold(ssock->sk);
2244 release_sock(sock->sk);
2246 err = ssock->ops->accept(sock, newsock, flags, kern);
2247 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2248 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2249 struct mptcp_subflow_context *subflow;
2251 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2252 * This is needed so NOSPACE flag can be set from tcp stack.
2254 __mptcp_flush_join_list(msk);
2255 mptcp_for_each_subflow(msk, subflow) {
2256 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2258 if (!ssk->sk_socket)
2259 mptcp_sock_graft(ssk, newsock);
2263 if (inet_csk_listen_poll(ssock->sk))
2264 set_bit(MPTCP_DATA_READY, &msk->flags);
2265 sock_put(ssock->sk);
2269 release_sock(sock->sk);
2273 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2275 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2279 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2280 struct poll_table_struct *wait)
2282 struct sock *sk = sock->sk;
2283 struct mptcp_sock *msk;
2288 sock_poll_wait(file, sock, wait);
2290 state = inet_sk_state_load(sk);
2291 if (state == TCP_LISTEN)
2292 return mptcp_check_readable(msk);
2294 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2295 mask |= mptcp_check_readable(msk);
2296 if (sk_stream_is_writeable(sk) &&
2297 test_bit(MPTCP_SEND_SPACE, &msk->flags))
2298 mask |= EPOLLOUT | EPOLLWRNORM;
2300 if (sk->sk_shutdown & RCV_SHUTDOWN)
2301 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2306 static int mptcp_shutdown(struct socket *sock, int how)
2308 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2309 struct mptcp_subflow_context *subflow;
2312 pr_debug("sk=%p, how=%d", msk, how);
2314 lock_sock(sock->sk);
2317 if ((how & ~SHUTDOWN_MASK) || !how) {
2322 if (sock->state == SS_CONNECTING) {
2323 if ((1 << sock->sk->sk_state) &
2324 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2325 sock->state = SS_DISCONNECTING;
2327 sock->state = SS_CONNECTED;
2330 /* If we've already sent a FIN, or it's a closed state, skip this. */
2331 if (__mptcp_check_fallback(msk)) {
2332 if (how == SHUT_WR || how == SHUT_RDWR)
2333 inet_sk_state_store(sock->sk, TCP_FIN_WAIT1);
2335 mptcp_for_each_subflow(msk, subflow) {
2336 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2338 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2340 } else if ((how & SEND_SHUTDOWN) &&
2341 ((1 << sock->sk->sk_state) &
2342 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2343 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) &&
2344 mptcp_close_state(sock->sk)) {
2345 __mptcp_flush_join_list(msk);
2347 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2348 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2350 mptcp_for_each_subflow(msk, subflow) {
2351 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2353 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2357 /* Wake up anyone sleeping in poll. */
2358 sock->sk->sk_state_change(sock->sk);
2361 release_sock(sock->sk);
2366 static const struct proto_ops mptcp_stream_ops = {
2368 .owner = THIS_MODULE,
2369 .release = inet_release,
2371 .connect = mptcp_stream_connect,
2372 .socketpair = sock_no_socketpair,
2373 .accept = mptcp_stream_accept,
2374 .getname = inet_getname,
2376 .ioctl = inet_ioctl,
2377 .gettstamp = sock_gettstamp,
2378 .listen = mptcp_listen,
2379 .shutdown = mptcp_shutdown,
2380 .setsockopt = sock_common_setsockopt,
2381 .getsockopt = sock_common_getsockopt,
2382 .sendmsg = inet_sendmsg,
2383 .recvmsg = inet_recvmsg,
2384 .mmap = sock_no_mmap,
2385 .sendpage = inet_sendpage,
2388 static struct inet_protosw mptcp_protosw = {
2389 .type = SOCK_STREAM,
2390 .protocol = IPPROTO_MPTCP,
2391 .prot = &mptcp_prot,
2392 .ops = &mptcp_stream_ops,
2393 .flags = INET_PROTOSW_ICSK,
2396 void __init mptcp_proto_init(void)
2398 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2400 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2401 panic("Failed to allocate MPTCP pcpu counter\n");
2403 mptcp_subflow_init();
2407 if (proto_register(&mptcp_prot, 1) != 0)
2408 panic("Failed to register MPTCP proto.\n");
2410 inet_register_protosw(&mptcp_protosw);
2412 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2415 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2416 static const struct proto_ops mptcp_v6_stream_ops = {
2418 .owner = THIS_MODULE,
2419 .release = inet6_release,
2421 .connect = mptcp_stream_connect,
2422 .socketpair = sock_no_socketpair,
2423 .accept = mptcp_stream_accept,
2424 .getname = inet6_getname,
2426 .ioctl = inet6_ioctl,
2427 .gettstamp = sock_gettstamp,
2428 .listen = mptcp_listen,
2429 .shutdown = mptcp_shutdown,
2430 .setsockopt = sock_common_setsockopt,
2431 .getsockopt = sock_common_getsockopt,
2432 .sendmsg = inet6_sendmsg,
2433 .recvmsg = inet6_recvmsg,
2434 .mmap = sock_no_mmap,
2435 .sendpage = inet_sendpage,
2436 #ifdef CONFIG_COMPAT
2437 .compat_ioctl = inet6_compat_ioctl,
2441 static struct proto mptcp_v6_prot;
2443 static void mptcp_v6_destroy(struct sock *sk)
2446 inet6_destroy_sock(sk);
2449 static struct inet_protosw mptcp_v6_protosw = {
2450 .type = SOCK_STREAM,
2451 .protocol = IPPROTO_MPTCP,
2452 .prot = &mptcp_v6_prot,
2453 .ops = &mptcp_v6_stream_ops,
2454 .flags = INET_PROTOSW_ICSK,
2457 int __init mptcp_proto_v6_init(void)
2461 mptcp_v6_prot = mptcp_prot;
2462 strcpy(mptcp_v6_prot.name, "MPTCPv6");
2463 mptcp_v6_prot.slab = NULL;
2464 mptcp_v6_prot.destroy = mptcp_v6_destroy;
2465 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
2467 err = proto_register(&mptcp_v6_prot, 1);
2471 err = inet6_register_protosw(&mptcp_v6_protosw);
2473 proto_unregister(&mptcp_v6_prot);