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 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);
198 inet_sk_state_store(sk, TCP_CLOSE);
199 sk->sk_state_change(sk);
203 if (sk->sk_shutdown == SHUTDOWN_MASK ||
204 sk->sk_state == TCP_CLOSE)
205 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
207 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
211 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
213 struct mptcp_sock *msk = mptcp_sk(sk);
215 if (READ_ONCE(msk->rcv_data_fin) &&
216 ((1 << sk->sk_state) &
217 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
218 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
220 if (msk->ack_seq == rcv_data_fin_seq) {
222 *seq = rcv_data_fin_seq;
231 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
233 long tout = ssk && inet_csk(ssk)->icsk_pending ?
234 inet_csk(ssk)->icsk_timeout - jiffies : 0;
237 tout = mptcp_sk(sk)->timer_ival;
238 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
241 static void mptcp_check_data_fin(struct sock *sk)
243 struct mptcp_sock *msk = mptcp_sk(sk);
244 u64 rcv_data_fin_seq;
246 if (__mptcp_check_fallback(msk) || !msk->first)
249 /* Need to ack a DATA_FIN received from a peer while this side
250 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
251 * msk->rcv_data_fin was set when parsing the incoming options
252 * at the subflow level and the msk lock was not held, so this
253 * is the first opportunity to act on the DATA_FIN and change
256 * If we are caught up to the sequence number of the incoming
257 * DATA_FIN, send the DATA_ACK now and do state transition. If
258 * not caught up, do nothing and let the recv code send DATA_ACK
262 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
263 struct mptcp_subflow_context *subflow;
266 WRITE_ONCE(msk->rcv_data_fin, 0);
268 sk->sk_shutdown |= RCV_SHUTDOWN;
269 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
270 set_bit(MPTCP_DATA_READY, &msk->flags);
272 switch (sk->sk_state) {
273 case TCP_ESTABLISHED:
274 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
277 inet_sk_state_store(sk, TCP_CLOSING);
280 inet_sk_state_store(sk, TCP_CLOSE);
281 // @@ Close subflows now?
284 /* Other states not expected */
289 mptcp_set_timeout(sk, NULL);
290 mptcp_for_each_subflow(msk, subflow) {
291 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
298 sk->sk_state_change(sk);
300 if (sk->sk_shutdown == SHUTDOWN_MASK ||
301 sk->sk_state == TCP_CLOSE)
302 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
304 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
308 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
312 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
313 struct sock *sk = (struct sock *)msk;
314 unsigned int moved = 0;
315 bool more_data_avail;
319 if (!mptcp_subflow_dsn_valid(msk, ssk)) {
326 u32 map_remaining, offset;
327 u32 seq = tp->copied_seq;
331 /* try to move as much data as available */
332 map_remaining = subflow->map_data_len -
333 mptcp_subflow_get_map_offset(subflow);
335 skb = skb_peek(&ssk->sk_receive_queue);
339 if (__mptcp_check_fallback(msk)) {
340 /* if we are running under the workqueue, TCP could have
341 * collapsed skbs between dummy map creation and now
342 * be sure to adjust the size
344 map_remaining = skb->len;
345 subflow->map_data_len = skb->len;
348 offset = seq - TCP_SKB_CB(skb)->seq;
349 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
355 if (offset < skb->len) {
356 size_t len = skb->len - offset;
361 __mptcp_move_skb(msk, ssk, skb, offset, len);
365 if (WARN_ON_ONCE(map_remaining < len))
369 sk_eat_skb(ssk, skb);
373 WRITE_ONCE(tp->copied_seq, seq);
374 more_data_avail = mptcp_subflow_data_available(ssk);
376 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf)) {
380 } while (more_data_avail);
384 /* If the moves have caught up with the DATA_FIN sequence number
385 * it's time to ack the DATA_FIN and change socket state, but
386 * this is not a good place to change state. Let the workqueue
389 if (mptcp_pending_data_fin(sk, NULL) &&
390 schedule_work(&msk->work))
396 /* In most cases we will be able to lock the mptcp socket. If its already
397 * owned, we need to defer to the work queue to avoid ABBA deadlock.
399 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
401 struct sock *sk = (struct sock *)msk;
402 unsigned int moved = 0;
404 if (READ_ONCE(sk->sk_lock.owned))
407 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
410 /* must re-check after taking the lock */
411 if (!READ_ONCE(sk->sk_lock.owned))
412 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
414 spin_unlock_bh(&sk->sk_lock.slock);
419 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
421 struct mptcp_sock *msk = mptcp_sk(sk);
423 set_bit(MPTCP_DATA_READY, &msk->flags);
425 if (atomic_read(&sk->sk_rmem_alloc) < READ_ONCE(sk->sk_rcvbuf) &&
426 move_skbs_to_msk(msk, ssk))
429 /* don't schedule if mptcp sk is (still) over limit */
430 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf))
433 /* mptcp socket is owned, release_cb should retry */
434 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
435 &sk->sk_tsq_flags)) {
438 /* need to try again, its possible release_cb() has already
439 * been called after the test_and_set_bit() above.
441 move_skbs_to_msk(msk, ssk);
444 sk->sk_data_ready(sk);
447 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
449 if (likely(list_empty(&msk->join_list)))
452 spin_lock_bh(&msk->join_list_lock);
453 list_splice_tail_init(&msk->join_list, &msk->conn_list);
454 spin_unlock_bh(&msk->join_list_lock);
457 static bool mptcp_timer_pending(struct sock *sk)
459 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
462 static void mptcp_reset_timer(struct sock *sk)
464 struct inet_connection_sock *icsk = inet_csk(sk);
467 /* should never be called with mptcp level timer cleared */
468 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
469 if (WARN_ON_ONCE(!tout))
471 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
474 void mptcp_data_acked(struct sock *sk)
476 mptcp_reset_timer(sk);
478 if ((!sk_stream_is_writeable(sk) ||
479 (inet_sk_state_load(sk) != TCP_ESTABLISHED)) &&
480 schedule_work(&mptcp_sk(sk)->work))
484 void mptcp_subflow_eof(struct sock *sk)
486 struct mptcp_sock *msk = mptcp_sk(sk);
488 if (!test_and_set_bit(MPTCP_WORK_EOF, &msk->flags) &&
489 schedule_work(&msk->work))
493 static void mptcp_check_for_eof(struct mptcp_sock *msk)
495 struct mptcp_subflow_context *subflow;
496 struct sock *sk = (struct sock *)msk;
499 mptcp_for_each_subflow(msk, subflow)
500 receivers += !subflow->rx_eof;
502 if (!receivers && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
503 /* hopefully temporary hack: propagate shutdown status
504 * to msk, when all subflows agree on it
506 sk->sk_shutdown |= RCV_SHUTDOWN;
508 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
509 set_bit(MPTCP_DATA_READY, &msk->flags);
510 sk->sk_data_ready(sk);
514 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
516 const struct sock *sk = (const struct sock *)msk;
518 if (!msk->cached_ext)
519 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
521 return !!msk->cached_ext;
524 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
526 struct mptcp_subflow_context *subflow;
527 struct sock *sk = (struct sock *)msk;
529 sock_owned_by_me(sk);
531 mptcp_for_each_subflow(msk, subflow) {
532 if (subflow->data_avail)
533 return mptcp_subflow_tcp_sock(subflow);
539 static bool mptcp_skb_can_collapse_to(u64 write_seq,
540 const struct sk_buff *skb,
541 const struct mptcp_ext *mpext)
543 if (!tcp_skb_can_collapse_to(skb))
546 /* can collapse only if MPTCP level sequence is in order */
547 return mpext && mpext->data_seq + mpext->data_len == write_seq;
550 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
551 const struct page_frag *pfrag,
552 const struct mptcp_data_frag *df)
554 return df && pfrag->page == df->page &&
555 df->data_seq + df->data_len == msk->write_seq;
558 static void dfrag_uncharge(struct sock *sk, int len)
560 sk_mem_uncharge(sk, len);
561 sk_wmem_queued_add(sk, -len);
564 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
566 int len = dfrag->data_len + dfrag->overhead;
568 list_del(&dfrag->list);
569 dfrag_uncharge(sk, len);
570 put_page(dfrag->page);
573 static void mptcp_clean_una(struct sock *sk)
575 struct mptcp_sock *msk = mptcp_sk(sk);
576 struct mptcp_data_frag *dtmp, *dfrag;
577 bool cleaned = false;
580 /* on fallback we just need to ignore snd_una, as this is really
583 if (__mptcp_check_fallback(msk))
584 atomic64_set(&msk->snd_una, msk->write_seq);
585 snd_una = atomic64_read(&msk->snd_una);
587 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
588 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
591 dfrag_clear(sk, dfrag);
595 dfrag = mptcp_rtx_head(sk);
596 if (dfrag && after64(snd_una, dfrag->data_seq)) {
597 u64 delta = snd_una - dfrag->data_seq;
599 if (WARN_ON_ONCE(delta > dfrag->data_len))
602 dfrag->data_seq += delta;
603 dfrag->offset += delta;
604 dfrag->data_len -= delta;
606 dfrag_uncharge(sk, delta);
612 sk_mem_reclaim_partial(sk);
614 /* Only wake up writers if a subflow is ready */
615 if (test_bit(MPTCP_SEND_SPACE, &msk->flags))
616 sk_stream_write_space(sk);
620 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
623 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
625 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
626 pfrag, sk->sk_allocation)))
629 sk->sk_prot->enter_memory_pressure(sk);
630 sk_stream_moderate_sndbuf(sk);
634 static struct mptcp_data_frag *
635 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
638 int offset = ALIGN(orig_offset, sizeof(long));
639 struct mptcp_data_frag *dfrag;
641 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
643 dfrag->data_seq = msk->write_seq;
644 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
645 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
646 dfrag->page = pfrag->page;
651 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
652 struct msghdr *msg, struct mptcp_data_frag *dfrag,
653 long *timeo, int *pmss_now,
656 int mss_now, avail_size, size_goal, offset, ret, frag_truesize = 0;
657 bool dfrag_collapsed, can_collapse = false;
658 struct mptcp_sock *msk = mptcp_sk(sk);
659 struct mptcp_ext *mpext = NULL;
660 bool retransmission = !!dfrag;
661 struct sk_buff *skb, *tail;
662 struct page_frag *pfrag;
667 /* use the mptcp page cache so that we can easily move the data
668 * from one substream to another, but do per subflow memory accounting
669 * Note: pfrag is used only !retransmission, but the compiler if
670 * fooled into a warning if we don't init here
672 pfrag = sk_page_frag(sk);
673 if (!retransmission) {
674 write_seq = &msk->write_seq;
677 write_seq = &dfrag->data_seq;
681 /* compute copy limit */
682 mss_now = tcp_send_mss(ssk, &size_goal, msg->msg_flags);
684 *ps_goal = size_goal;
685 avail_size = size_goal;
686 skb = tcp_write_queue_tail(ssk);
688 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
690 /* Limit the write to the size available in the
691 * current skb, if any, so that we create at most a new skb.
692 * Explicitly tells TCP internals to avoid collapsing on later
693 * queue management operation, to avoid breaking the ext <->
694 * SSN association set here
696 can_collapse = (size_goal - skb->len > 0) &&
697 mptcp_skb_can_collapse_to(*write_seq, skb, mpext);
699 TCP_SKB_CB(skb)->eor = 1;
701 avail_size = size_goal - skb->len;
704 if (!retransmission) {
705 /* reuse tail pfrag, if possible, or carve a new one from the
708 dfrag = mptcp_rtx_tail(sk);
709 offset = pfrag->offset;
710 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
711 if (!dfrag_collapsed) {
712 dfrag = mptcp_carve_data_frag(msk, pfrag, offset);
713 offset = dfrag->offset;
714 frag_truesize = dfrag->overhead;
716 psize = min_t(size_t, pfrag->size - offset, avail_size);
719 pr_debug("left=%zu", msg_data_left(msg));
720 psize = copy_page_from_iter(pfrag->page, offset,
721 min_t(size_t, msg_data_left(msg),
724 pr_debug("left=%zu", msg_data_left(msg));
728 if (!sk_wmem_schedule(sk, psize + dfrag->overhead)) {
729 iov_iter_revert(&msg->msg_iter, psize);
733 offset = dfrag->offset;
734 psize = min_t(size_t, dfrag->data_len, avail_size);
737 /* tell the TCP stack to delay the push so that we can safely
738 * access the skb after the sendpages call
740 ret = do_tcp_sendpages(ssk, page, offset, psize,
741 msg->msg_flags | MSG_SENDPAGE_NOTLAST | MSG_DONTWAIT);
744 iov_iter_revert(&msg->msg_iter, psize);
748 frag_truesize += ret;
749 if (!retransmission) {
750 if (unlikely(ret < psize))
751 iov_iter_revert(&msg->msg_iter, psize - ret);
753 /* send successful, keep track of sent data for mptcp-level
756 dfrag->data_len += ret;
757 if (!dfrag_collapsed) {
758 get_page(dfrag->page);
759 list_add_tail(&dfrag->list, &msk->rtx_queue);
760 sk_wmem_queued_add(sk, frag_truesize);
762 sk_wmem_queued_add(sk, ret);
765 /* charge data on mptcp rtx queue to the master socket
766 * Note: we charge such data both to sk and ssk
768 sk->sk_forward_alloc -= frag_truesize;
771 /* if the tail skb extension is still the cached one, collapsing
772 * really happened. Note: we can't check for 'same skb' as the sk_buff
773 * hdr on tail can be transmitted, freed and re-allocated by the
774 * do_tcp_sendpages() call
776 tail = tcp_write_queue_tail(ssk);
777 if (mpext && tail && mpext == skb_ext_find(tail, SKB_EXT_MPTCP)) {
778 WARN_ON_ONCE(!can_collapse);
779 mpext->data_len += ret;
783 skb = tcp_write_queue_tail(ssk);
784 mpext = __skb_ext_set(skb, SKB_EXT_MPTCP, msk->cached_ext);
785 msk->cached_ext = NULL;
787 memset(mpext, 0, sizeof(*mpext));
788 mpext->data_seq = *write_seq;
789 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
790 mpext->data_len = ret;
794 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
795 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
800 pfrag->offset += frag_truesize;
801 WRITE_ONCE(*write_seq, *write_seq + ret);
802 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
807 static void mptcp_nospace(struct mptcp_sock *msk, struct socket *sock)
809 clear_bit(MPTCP_SEND_SPACE, &msk->flags);
810 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
812 /* enables sk->write_space() callbacks */
813 set_bit(SOCK_NOSPACE, &sock->flags);
816 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
818 struct mptcp_subflow_context *subflow;
819 struct sock *backup = NULL;
821 sock_owned_by_me((const struct sock *)msk);
823 if (!mptcp_ext_cache_refill(msk))
826 mptcp_for_each_subflow(msk, subflow) {
827 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
829 if (!sk_stream_memory_free(ssk)) {
830 struct socket *sock = ssk->sk_socket;
833 mptcp_nospace(msk, sock);
838 if (subflow->backup) {
851 static void ssk_check_wmem(struct mptcp_sock *msk, struct sock *ssk)
855 if (likely(sk_stream_is_writeable(ssk)))
858 sock = READ_ONCE(ssk->sk_socket);
860 mptcp_nospace(msk, sock);
863 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
865 int mss_now = 0, size_goal = 0, ret = 0;
866 struct mptcp_sock *msk = mptcp_sk(sk);
867 struct page_frag *pfrag;
873 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
878 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
880 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
881 ret = sk_stream_wait_connect(sk, &timeo);
886 pfrag = sk_page_frag(sk);
890 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
896 __mptcp_flush_join_list(msk);
897 ssk = mptcp_subflow_get_send(msk);
898 while (!sk_stream_memory_free(sk) ||
900 !mptcp_page_frag_refill(ssk, pfrag)) {
902 /* make sure retransmit timer is
903 * running before we wait for memory.
905 * The retransmit timer might be needed
906 * to make the peer send an up-to-date
909 mptcp_set_timeout(sk, ssk);
910 if (!mptcp_timer_pending(sk))
911 mptcp_reset_timer(sk);
914 ret = sk_stream_wait_memory(sk, &timeo);
920 ssk = mptcp_subflow_get_send(msk);
921 if (list_empty(&msk->conn_list)) {
927 pr_debug("conn_list->subflow=%p", ssk);
930 tx_ok = msg_data_left(msg);
932 ret = mptcp_sendmsg_frag(sk, ssk, msg, NULL, &timeo, &mss_now,
935 if (ret == -EAGAIN && timeo > 0) {
936 mptcp_set_timeout(sk, ssk);
945 tx_ok = msg_data_left(msg);
949 if (!sk_stream_memory_free(ssk) ||
950 !mptcp_page_frag_refill(ssk, pfrag) ||
951 !mptcp_ext_cache_refill(msk)) {
952 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
953 tcp_push(ssk, msg->msg_flags, mss_now,
954 tcp_sk(ssk)->nonagle, size_goal);
955 mptcp_set_timeout(sk, ssk);
960 /* memory is charged to mptcp level socket as well, i.e.
961 * if msg is very large, mptcp socket may run out of buffer
962 * space. mptcp_clean_una() will release data that has
963 * been acked at mptcp level in the mean time, so there is
964 * a good chance we can continue sending data right away.
966 * Normally, when the tcp subflow can accept more data, then
967 * so can the MPTCP socket. However, we need to cope with
968 * peers that might lag behind in their MPTCP-level
969 * acknowledgements, i.e. data might have been acked at
970 * tcp level only. So, we must also check the MPTCP socket
971 * limits before we send more data.
973 if (unlikely(!sk_stream_memory_free(sk))) {
974 tcp_push(ssk, msg->msg_flags, mss_now,
975 tcp_sk(ssk)->nonagle, size_goal);
977 if (!sk_stream_memory_free(sk)) {
978 /* can't send more for now, need to wait for
979 * MPTCP-level ACKs from peer.
981 * Wakeup will happen via mptcp_clean_una().
983 mptcp_set_timeout(sk, ssk);
985 goto wait_for_sndbuf;
990 mptcp_set_timeout(sk, ssk);
992 tcp_push(ssk, msg->msg_flags, mss_now, tcp_sk(ssk)->nonagle,
995 /* start the timer, if it's not pending */
996 if (!mptcp_timer_pending(sk))
997 mptcp_reset_timer(sk);
1000 ssk_check_wmem(msk, ssk);
1004 return copied ? : ret;
1007 static void mptcp_wait_data(struct sock *sk, long *timeo)
1009 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1010 struct mptcp_sock *msk = mptcp_sk(sk);
1012 add_wait_queue(sk_sleep(sk), &wait);
1013 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1015 sk_wait_event(sk, timeo,
1016 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1018 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1019 remove_wait_queue(sk_sleep(sk), &wait);
1022 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1026 struct sock *sk = (struct sock *)msk;
1027 struct sk_buff *skb;
1030 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1031 u32 offset = MPTCP_SKB_CB(skb)->offset;
1032 u32 data_len = skb->len - offset;
1033 u32 count = min_t(size_t, len - copied, data_len);
1036 err = skb_copy_datagram_msg(skb, offset, msg, count);
1037 if (unlikely(err < 0)) {
1045 if (count < data_len) {
1046 MPTCP_SKB_CB(skb)->offset += count;
1050 __skb_unlink(skb, &sk->sk_receive_queue);
1060 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1062 * Only difference: Use highest rtt estimate of the subflows in use.
1064 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1066 struct mptcp_subflow_context *subflow;
1067 struct sock *sk = (struct sock *)msk;
1068 u32 time, advmss = 1;
1071 sock_owned_by_me(sk);
1076 msk->rcvq_space.copied += copied;
1078 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1079 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1081 rtt_us = msk->rcvq_space.rtt_us;
1082 if (rtt_us && time < (rtt_us >> 3))
1086 mptcp_for_each_subflow(msk, subflow) {
1087 const struct tcp_sock *tp;
1091 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1093 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1094 sf_advmss = READ_ONCE(tp->advmss);
1096 rtt_us = max(sf_rtt_us, rtt_us);
1097 advmss = max(sf_advmss, advmss);
1100 msk->rcvq_space.rtt_us = rtt_us;
1101 if (time < (rtt_us >> 3) || rtt_us == 0)
1104 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1107 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1108 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1112 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1114 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1116 do_div(grow, msk->rcvq_space.space);
1117 rcvwin += (grow << 1);
1119 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1120 while (tcp_win_from_space(sk, rcvmem) < advmss)
1123 do_div(rcvwin, advmss);
1124 rcvbuf = min_t(u64, rcvwin * rcvmem,
1125 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1127 if (rcvbuf > sk->sk_rcvbuf) {
1130 window_clamp = tcp_win_from_space(sk, rcvbuf);
1131 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1133 /* Make subflows follow along. If we do not do this, we
1134 * get drops at subflow level if skbs can't be moved to
1135 * the mptcp rx queue fast enough (announced rcv_win can
1136 * exceed ssk->sk_rcvbuf).
1138 mptcp_for_each_subflow(msk, subflow) {
1141 ssk = mptcp_subflow_tcp_sock(subflow);
1142 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1143 tcp_sk(ssk)->window_clamp = window_clamp;
1148 msk->rcvq_space.space = msk->rcvq_space.copied;
1150 msk->rcvq_space.copied = 0;
1151 msk->rcvq_space.time = mstamp;
1154 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1156 unsigned int moved = 0;
1160 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1166 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1173 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1174 int nonblock, int flags, int *addr_len)
1176 struct mptcp_sock *msk = mptcp_sk(sk);
1181 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1185 timeo = sock_rcvtimeo(sk, nonblock);
1187 len = min_t(size_t, len, INT_MAX);
1188 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1189 __mptcp_flush_join_list(msk);
1191 while (len > (size_t)copied) {
1194 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1195 if (unlikely(bytes_read < 0)) {
1197 copied = bytes_read;
1201 copied += bytes_read;
1203 if (skb_queue_empty(&sk->sk_receive_queue) &&
1204 __mptcp_move_skbs(msk))
1207 /* only the master socket status is relevant here. The exit
1208 * conditions mirror closely tcp_recvmsg()
1210 if (copied >= target)
1215 sk->sk_state == TCP_CLOSE ||
1216 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1218 signal_pending(current))
1222 copied = sock_error(sk);
1226 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1227 mptcp_check_for_eof(msk);
1229 if (sk->sk_shutdown & RCV_SHUTDOWN)
1232 if (sk->sk_state == TCP_CLOSE) {
1242 if (signal_pending(current)) {
1243 copied = sock_intr_errno(timeo);
1248 pr_debug("block timeout %ld", timeo);
1249 mptcp_wait_data(sk, &timeo);
1252 if (skb_queue_empty(&sk->sk_receive_queue)) {
1253 /* entire backlog drained, clear DATA_READY. */
1254 clear_bit(MPTCP_DATA_READY, &msk->flags);
1256 /* .. race-breaker: ssk might have gotten new data
1257 * after last __mptcp_move_skbs() returned false.
1259 if (unlikely(__mptcp_move_skbs(msk)))
1260 set_bit(MPTCP_DATA_READY, &msk->flags);
1261 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1262 /* data to read but mptcp_wait_data() cleared DATA_READY */
1263 set_bit(MPTCP_DATA_READY, &msk->flags);
1266 mptcp_rcv_space_adjust(msk, copied);
1272 static void mptcp_retransmit_handler(struct sock *sk)
1274 struct mptcp_sock *msk = mptcp_sk(sk);
1276 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->write_seq)) {
1277 mptcp_stop_timer(sk);
1279 set_bit(MPTCP_WORK_RTX, &msk->flags);
1280 if (schedule_work(&msk->work))
1285 static void mptcp_retransmit_timer(struct timer_list *t)
1287 struct inet_connection_sock *icsk = from_timer(icsk, t,
1288 icsk_retransmit_timer);
1289 struct sock *sk = &icsk->icsk_inet.sk;
1292 if (!sock_owned_by_user(sk)) {
1293 mptcp_retransmit_handler(sk);
1295 /* delegate our work to tcp_release_cb() */
1296 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1304 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1307 * A backup subflow is returned only if that is the only kind available.
1309 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1311 struct mptcp_subflow_context *subflow;
1312 struct sock *backup = NULL;
1314 sock_owned_by_me((const struct sock *)msk);
1316 mptcp_for_each_subflow(msk, subflow) {
1317 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1319 /* still data outstanding at TCP level? Don't retransmit. */
1320 if (!tcp_write_queue_empty(ssk))
1323 if (subflow->backup) {
1335 /* subflow sockets can be either outgoing (connect) or incoming
1338 * Outgoing subflows use in-kernel sockets.
1339 * Incoming subflows do not have their own 'struct socket' allocated,
1340 * so we need to use tcp_close() after detaching them from the mptcp
1343 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1344 struct mptcp_subflow_context *subflow,
1347 struct socket *sock = READ_ONCE(ssk->sk_socket);
1349 list_del(&subflow->node);
1351 if (sock && sock != sk->sk_socket) {
1352 /* outgoing subflow */
1355 /* incoming subflow */
1356 tcp_close(ssk, timeout);
1360 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1365 static void pm_work(struct mptcp_sock *msk)
1367 struct mptcp_pm_data *pm = &msk->pm;
1369 spin_lock_bh(&msk->pm.lock);
1371 pr_debug("msk=%p status=%x", msk, pm->status);
1372 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1373 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1374 mptcp_pm_nl_add_addr_received(msk);
1376 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1377 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1378 mptcp_pm_nl_fully_established(msk);
1380 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1381 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1382 mptcp_pm_nl_subflow_established(msk);
1385 spin_unlock_bh(&msk->pm.lock);
1388 static void mptcp_worker(struct work_struct *work)
1390 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1391 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1392 int orig_len, orig_offset, mss_now = 0, size_goal = 0;
1393 struct mptcp_data_frag *dfrag;
1396 struct msghdr msg = {
1397 .msg_flags = MSG_DONTWAIT,
1402 mptcp_clean_una(sk);
1403 mptcp_check_data_fin_ack(sk);
1404 __mptcp_flush_join_list(msk);
1405 __mptcp_move_skbs(msk);
1410 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1411 mptcp_check_for_eof(msk);
1413 mptcp_check_data_fin(sk);
1415 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1418 dfrag = mptcp_rtx_head(sk);
1422 if (!mptcp_ext_cache_refill(msk))
1425 ssk = mptcp_subflow_get_retrans(msk);
1431 orig_len = dfrag->data_len;
1432 orig_offset = dfrag->offset;
1433 orig_write_seq = dfrag->data_seq;
1434 while (dfrag->data_len > 0) {
1435 int ret = mptcp_sendmsg_frag(sk, ssk, &msg, dfrag, &timeo,
1436 &mss_now, &size_goal);
1440 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1442 dfrag->data_len -= ret;
1443 dfrag->offset += ret;
1445 if (!mptcp_ext_cache_refill(msk))
1449 tcp_push(ssk, msg.msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1452 dfrag->data_seq = orig_write_seq;
1453 dfrag->offset = orig_offset;
1454 dfrag->data_len = orig_len;
1456 mptcp_set_timeout(sk, ssk);
1460 if (!mptcp_timer_pending(sk))
1461 mptcp_reset_timer(sk);
1468 static int __mptcp_init_sock(struct sock *sk)
1470 struct mptcp_sock *msk = mptcp_sk(sk);
1472 spin_lock_init(&msk->join_list_lock);
1474 INIT_LIST_HEAD(&msk->conn_list);
1475 INIT_LIST_HEAD(&msk->join_list);
1476 INIT_LIST_HEAD(&msk->rtx_queue);
1477 __set_bit(MPTCP_SEND_SPACE, &msk->flags);
1478 INIT_WORK(&msk->work, mptcp_worker);
1481 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1483 mptcp_pm_data_init(msk);
1485 /* re-use the csk retrans timer for MPTCP-level retrans */
1486 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1491 static int mptcp_init_sock(struct sock *sk)
1493 struct net *net = sock_net(sk);
1496 if (!mptcp_is_enabled(net))
1497 return -ENOPROTOOPT;
1499 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1502 ret = __mptcp_init_sock(sk);
1506 ret = __mptcp_socket_create(mptcp_sk(sk));
1510 sk_sockets_allocated_inc(sk);
1511 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1512 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[2];
1517 static void __mptcp_clear_xmit(struct sock *sk)
1519 struct mptcp_sock *msk = mptcp_sk(sk);
1520 struct mptcp_data_frag *dtmp, *dfrag;
1522 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
1524 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
1525 dfrag_clear(sk, dfrag);
1528 static void mptcp_cancel_work(struct sock *sk)
1530 struct mptcp_sock *msk = mptcp_sk(sk);
1532 if (cancel_work_sync(&msk->work))
1536 static void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
1540 switch (ssk->sk_state) {
1542 if (!(how & RCV_SHUTDOWN))
1546 tcp_disconnect(ssk, O_NONBLOCK);
1549 if (__mptcp_check_fallback(mptcp_sk(sk))) {
1550 pr_debug("Fallback");
1551 ssk->sk_shutdown |= how;
1552 tcp_shutdown(ssk, how);
1554 pr_debug("Sending DATA_FIN on subflow %p", ssk);
1555 mptcp_set_timeout(sk, ssk);
1564 static const unsigned char new_state[16] = {
1565 /* current state: new state: action: */
1566 [0 /* (Invalid) */] = TCP_CLOSE,
1567 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1568 [TCP_SYN_SENT] = TCP_CLOSE,
1569 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1570 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1571 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1572 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
1573 [TCP_CLOSE] = TCP_CLOSE,
1574 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1575 [TCP_LAST_ACK] = TCP_LAST_ACK,
1576 [TCP_LISTEN] = TCP_CLOSE,
1577 [TCP_CLOSING] = TCP_CLOSING,
1578 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1581 static int mptcp_close_state(struct sock *sk)
1583 int next = (int)new_state[sk->sk_state];
1584 int ns = next & TCP_STATE_MASK;
1586 inet_sk_state_store(sk, ns);
1588 return next & TCP_ACTION_FIN;
1591 static void mptcp_close(struct sock *sk, long timeout)
1593 struct mptcp_subflow_context *subflow, *tmp;
1594 struct mptcp_sock *msk = mptcp_sk(sk);
1595 LIST_HEAD(conn_list);
1598 sk->sk_shutdown = SHUTDOWN_MASK;
1600 if (sk->sk_state == TCP_LISTEN) {
1601 inet_sk_state_store(sk, TCP_CLOSE);
1603 } else if (sk->sk_state == TCP_CLOSE) {
1607 if (__mptcp_check_fallback(msk)) {
1609 } else if (mptcp_close_state(sk)) {
1610 pr_debug("Sending DATA_FIN sk=%p", sk);
1611 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
1612 WRITE_ONCE(msk->snd_data_fin_enable, 1);
1614 mptcp_for_each_subflow(msk, subflow) {
1615 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
1617 mptcp_subflow_shutdown(sk, tcp_sk, SHUTDOWN_MASK);
1621 sk_stream_wait_close(sk, timeout);
1624 inet_sk_state_store(sk, TCP_CLOSE);
1627 /* be sure to always acquire the join list lock, to sync vs
1628 * mptcp_finish_join().
1630 spin_lock_bh(&msk->join_list_lock);
1631 list_splice_tail_init(&msk->join_list, &msk->conn_list);
1632 spin_unlock_bh(&msk->join_list_lock);
1633 list_splice_init(&msk->conn_list, &conn_list);
1635 __mptcp_clear_xmit(sk);
1639 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
1640 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1641 __mptcp_close_ssk(sk, ssk, subflow, timeout);
1644 mptcp_cancel_work(sk);
1646 __skb_queue_purge(&sk->sk_receive_queue);
1648 sk_common_release(sk);
1651 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
1653 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1654 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
1655 struct ipv6_pinfo *msk6 = inet6_sk(msk);
1657 msk->sk_v6_daddr = ssk->sk_v6_daddr;
1658 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
1661 msk6->saddr = ssk6->saddr;
1662 msk6->flow_label = ssk6->flow_label;
1666 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
1667 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
1668 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
1669 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
1670 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
1671 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
1674 static int mptcp_disconnect(struct sock *sk, int flags)
1676 /* Should never be called.
1677 * inet_stream_connect() calls ->disconnect, but that
1678 * refers to the subflow socket, not the mptcp one.
1684 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1685 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
1687 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
1689 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
1693 struct sock *mptcp_sk_clone(const struct sock *sk,
1694 const struct mptcp_options_received *mp_opt,
1695 struct request_sock *req)
1697 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
1698 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
1699 struct mptcp_sock *msk;
1705 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1706 if (nsk->sk_family == AF_INET6)
1707 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
1710 __mptcp_init_sock(nsk);
1712 msk = mptcp_sk(nsk);
1713 msk->local_key = subflow_req->local_key;
1714 msk->token = subflow_req->token;
1715 msk->subflow = NULL;
1716 WRITE_ONCE(msk->fully_established, false);
1718 msk->write_seq = subflow_req->idsn + 1;
1719 atomic64_set(&msk->snd_una, msk->write_seq);
1720 if (mp_opt->mp_capable) {
1721 msk->can_ack = true;
1722 msk->remote_key = mp_opt->sndr_key;
1723 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
1725 msk->ack_seq = ack_seq;
1728 sock_reset_flag(nsk, SOCK_RCU_FREE);
1729 /* will be fully established after successful MPC subflow creation */
1730 inet_sk_state_store(nsk, TCP_SYN_RECV);
1731 bh_unlock_sock(nsk);
1733 /* keep a single reference */
1738 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
1740 const struct tcp_sock *tp = tcp_sk(ssk);
1742 msk->rcvq_space.copied = 0;
1743 msk->rcvq_space.rtt_us = 0;
1745 msk->rcvq_space.time = tp->tcp_mstamp;
1747 /* initial rcv_space offering made to peer */
1748 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
1749 TCP_INIT_CWND * tp->advmss);
1750 if (msk->rcvq_space.space == 0)
1751 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
1754 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
1757 struct mptcp_sock *msk = mptcp_sk(sk);
1758 struct socket *listener;
1761 listener = __mptcp_nmpc_socket(msk);
1762 if (WARN_ON_ONCE(!listener)) {
1767 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
1768 newsk = inet_csk_accept(listener->sk, flags, err, kern);
1772 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
1773 if (sk_is_mptcp(newsk)) {
1774 struct mptcp_subflow_context *subflow;
1775 struct sock *new_mptcp_sock;
1776 struct sock *ssk = newsk;
1778 subflow = mptcp_subflow_ctx(newsk);
1779 new_mptcp_sock = subflow->conn;
1781 /* is_mptcp should be false if subflow->conn is missing, see
1782 * subflow_syn_recv_sock()
1784 if (WARN_ON_ONCE(!new_mptcp_sock)) {
1785 tcp_sk(newsk)->is_mptcp = 0;
1789 /* acquire the 2nd reference for the owning socket */
1790 sock_hold(new_mptcp_sock);
1793 bh_lock_sock(new_mptcp_sock);
1794 msk = mptcp_sk(new_mptcp_sock);
1797 newsk = new_mptcp_sock;
1798 mptcp_copy_inaddrs(newsk, ssk);
1799 list_add(&subflow->node, &msk->conn_list);
1801 mptcp_rcv_space_init(msk, ssk);
1802 bh_unlock_sock(new_mptcp_sock);
1804 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
1807 MPTCP_INC_STATS(sock_net(sk),
1808 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
1814 static void mptcp_destroy(struct sock *sk)
1816 struct mptcp_sock *msk = mptcp_sk(sk);
1818 mptcp_token_destroy(msk);
1819 if (msk->cached_ext)
1820 __skb_ext_put(msk->cached_ext);
1822 sk_sockets_allocated_dec(sk);
1825 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
1826 sockptr_t optval, unsigned int optlen)
1828 struct sock *sk = (struct sock *)msk;
1829 struct socket *ssock;
1836 ssock = __mptcp_nmpc_socket(msk);
1842 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
1844 if (optname == SO_REUSEPORT)
1845 sk->sk_reuseport = ssock->sk->sk_reuseport;
1846 else if (optname == SO_REUSEADDR)
1847 sk->sk_reuse = ssock->sk->sk_reuse;
1853 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
1856 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
1857 sockptr_t optval, unsigned int optlen)
1859 struct sock *sk = (struct sock *)msk;
1860 int ret = -EOPNOTSUPP;
1861 struct socket *ssock;
1866 ssock = __mptcp_nmpc_socket(msk);
1872 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
1874 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
1883 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
1884 sockptr_t optval, unsigned int optlen)
1886 struct mptcp_sock *msk = mptcp_sk(sk);
1889 pr_debug("msk=%p", msk);
1891 if (level == SOL_SOCKET)
1892 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
1894 /* @@ the meaning of setsockopt() when the socket is connected and
1895 * there are multiple subflows is not yet defined. It is up to the
1896 * MPTCP-level socket to configure the subflows until the subflow
1897 * is in TCP fallback, when TCP socket options are passed through
1898 * to the one remaining subflow.
1901 ssk = __mptcp_tcp_fallback(msk);
1904 return tcp_setsockopt(ssk, level, optname, optval, optlen);
1906 if (level == SOL_IPV6)
1907 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
1912 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
1913 char __user *optval, int __user *option)
1915 struct mptcp_sock *msk = mptcp_sk(sk);
1918 pr_debug("msk=%p", msk);
1920 /* @@ the meaning of setsockopt() when the socket is connected and
1921 * there are multiple subflows is not yet defined. It is up to the
1922 * MPTCP-level socket to configure the subflows until the subflow
1923 * is in TCP fallback, when socket options are passed through
1924 * to the one remaining subflow.
1927 ssk = __mptcp_tcp_fallback(msk);
1930 return tcp_getsockopt(ssk, level, optname, optval, option);
1935 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
1936 TCPF_WRITE_TIMER_DEFERRED)
1938 /* this is very alike tcp_release_cb() but we must handle differently a
1939 * different set of events
1941 static void mptcp_release_cb(struct sock *sk)
1943 unsigned long flags, nflags;
1946 flags = sk->sk_tsq_flags;
1947 if (!(flags & MPTCP_DEFERRED_ALL))
1949 nflags = flags & ~MPTCP_DEFERRED_ALL;
1950 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1952 sock_release_ownership(sk);
1954 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1955 struct mptcp_sock *msk = mptcp_sk(sk);
1958 ssk = mptcp_subflow_recv_lookup(msk);
1959 if (!ssk || !schedule_work(&msk->work))
1963 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1964 mptcp_retransmit_handler(sk);
1969 static int mptcp_hash(struct sock *sk)
1971 /* should never be called,
1972 * we hash the TCP subflows not the master socket
1978 static void mptcp_unhash(struct sock *sk)
1980 /* called from sk_common_release(), but nothing to do here */
1983 static int mptcp_get_port(struct sock *sk, unsigned short snum)
1985 struct mptcp_sock *msk = mptcp_sk(sk);
1986 struct socket *ssock;
1988 ssock = __mptcp_nmpc_socket(msk);
1989 pr_debug("msk=%p, subflow=%p", msk, ssock);
1990 if (WARN_ON_ONCE(!ssock))
1993 return inet_csk_get_port(ssock->sk, snum);
1996 void mptcp_finish_connect(struct sock *ssk)
1998 struct mptcp_subflow_context *subflow;
1999 struct mptcp_sock *msk;
2003 subflow = mptcp_subflow_ctx(ssk);
2007 pr_debug("msk=%p, token=%u", sk, subflow->token);
2009 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2011 subflow->map_seq = ack_seq;
2012 subflow->map_subflow_seq = 1;
2014 /* the socket is not connected yet, no msk/subflow ops can access/race
2015 * accessing the field below
2017 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2018 WRITE_ONCE(msk->local_key, subflow->local_key);
2019 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2020 WRITE_ONCE(msk->ack_seq, ack_seq);
2021 WRITE_ONCE(msk->can_ack, 1);
2022 atomic64_set(&msk->snd_una, msk->write_seq);
2024 mptcp_pm_new_connection(msk, 0);
2026 mptcp_rcv_space_init(msk, ssk);
2029 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2031 write_lock_bh(&sk->sk_callback_lock);
2032 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2033 sk_set_socket(sk, parent);
2034 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2035 write_unlock_bh(&sk->sk_callback_lock);
2038 bool mptcp_finish_join(struct sock *sk)
2040 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
2041 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2042 struct sock *parent = (void *)msk;
2043 struct socket *parent_sock;
2046 pr_debug("msk=%p, subflow=%p", msk, subflow);
2048 /* mptcp socket already closing? */
2049 if (!mptcp_is_fully_established(parent))
2052 if (!msk->pm.server_side)
2055 if (!mptcp_pm_allow_new_subflow(msk))
2058 /* active connections are already on conn_list, and we can't acquire
2060 * use the join list lock as synchronization point and double-check
2061 * msk status to avoid racing with mptcp_close()
2063 spin_lock_bh(&msk->join_list_lock);
2064 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2065 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node)))
2066 list_add_tail(&subflow->node, &msk->join_list);
2067 spin_unlock_bh(&msk->join_list_lock);
2071 /* attach to msk socket only after we are sure he will deal with us
2074 parent_sock = READ_ONCE(parent->sk_socket);
2075 if (parent_sock && !sk->sk_socket)
2076 mptcp_sock_graft(sk, parent_sock);
2077 subflow->map_seq = msk->ack_seq;
2081 static bool mptcp_memory_free(const struct sock *sk, int wake)
2083 struct mptcp_sock *msk = mptcp_sk(sk);
2085 return wake ? test_bit(MPTCP_SEND_SPACE, &msk->flags) : true;
2088 static struct proto mptcp_prot = {
2090 .owner = THIS_MODULE,
2091 .init = mptcp_init_sock,
2092 .disconnect = mptcp_disconnect,
2093 .close = mptcp_close,
2094 .accept = mptcp_accept,
2095 .setsockopt = mptcp_setsockopt,
2096 .getsockopt = mptcp_getsockopt,
2097 .shutdown = tcp_shutdown,
2098 .destroy = mptcp_destroy,
2099 .sendmsg = mptcp_sendmsg,
2100 .recvmsg = mptcp_recvmsg,
2101 .release_cb = mptcp_release_cb,
2103 .unhash = mptcp_unhash,
2104 .get_port = mptcp_get_port,
2105 .sockets_allocated = &mptcp_sockets_allocated,
2106 .memory_allocated = &tcp_memory_allocated,
2107 .memory_pressure = &tcp_memory_pressure,
2108 .stream_memory_free = mptcp_memory_free,
2109 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2110 .sysctl_mem = sysctl_tcp_mem,
2111 .obj_size = sizeof(struct mptcp_sock),
2112 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2113 .no_autobind = true,
2116 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2118 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2119 struct socket *ssock;
2122 lock_sock(sock->sk);
2123 ssock = __mptcp_nmpc_socket(msk);
2129 err = ssock->ops->bind(ssock, uaddr, addr_len);
2131 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2134 release_sock(sock->sk);
2138 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2139 struct mptcp_subflow_context *subflow)
2141 subflow->request_mptcp = 0;
2142 __mptcp_do_fallback(msk);
2145 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2146 int addr_len, int flags)
2148 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2149 struct mptcp_subflow_context *subflow;
2150 struct socket *ssock;
2153 lock_sock(sock->sk);
2154 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2155 /* pending connection or invalid state, let existing subflow
2158 ssock = msk->subflow;
2162 ssock = __mptcp_nmpc_socket(msk);
2168 mptcp_token_destroy(msk);
2169 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2170 subflow = mptcp_subflow_ctx(ssock->sk);
2171 #ifdef CONFIG_TCP_MD5SIG
2172 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2175 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2176 mptcp_subflow_early_fallback(msk, subflow);
2178 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2179 mptcp_subflow_early_fallback(msk, subflow);
2182 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2183 sock->state = ssock->state;
2185 /* on successful connect, the msk state will be moved to established by
2186 * subflow_finish_connect()
2188 if (!err || err == -EINPROGRESS)
2189 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2191 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2194 release_sock(sock->sk);
2198 static int mptcp_listen(struct socket *sock, int backlog)
2200 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2201 struct socket *ssock;
2204 pr_debug("msk=%p", msk);
2206 lock_sock(sock->sk);
2207 ssock = __mptcp_nmpc_socket(msk);
2213 mptcp_token_destroy(msk);
2214 inet_sk_state_store(sock->sk, TCP_LISTEN);
2215 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2217 err = ssock->ops->listen(ssock, backlog);
2218 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2220 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2223 release_sock(sock->sk);
2227 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2228 int flags, bool kern)
2230 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2231 struct socket *ssock;
2234 pr_debug("msk=%p", msk);
2236 lock_sock(sock->sk);
2237 if (sock->sk->sk_state != TCP_LISTEN)
2240 ssock = __mptcp_nmpc_socket(msk);
2244 clear_bit(MPTCP_DATA_READY, &msk->flags);
2245 sock_hold(ssock->sk);
2246 release_sock(sock->sk);
2248 err = ssock->ops->accept(sock, newsock, flags, kern);
2249 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2250 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2251 struct mptcp_subflow_context *subflow;
2253 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2254 * This is needed so NOSPACE flag can be set from tcp stack.
2256 __mptcp_flush_join_list(msk);
2257 mptcp_for_each_subflow(msk, subflow) {
2258 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2260 if (!ssk->sk_socket)
2261 mptcp_sock_graft(ssk, newsock);
2265 if (inet_csk_listen_poll(ssock->sk))
2266 set_bit(MPTCP_DATA_READY, &msk->flags);
2267 sock_put(ssock->sk);
2271 release_sock(sock->sk);
2275 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2277 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2281 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2282 struct poll_table_struct *wait)
2284 struct sock *sk = sock->sk;
2285 struct mptcp_sock *msk;
2290 sock_poll_wait(file, sock, wait);
2292 state = inet_sk_state_load(sk);
2293 if (state == TCP_LISTEN)
2294 return mptcp_check_readable(msk);
2296 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2297 mask |= mptcp_check_readable(msk);
2298 if (sk_stream_is_writeable(sk) &&
2299 test_bit(MPTCP_SEND_SPACE, &msk->flags))
2300 mask |= EPOLLOUT | EPOLLWRNORM;
2302 if (sk->sk_shutdown & RCV_SHUTDOWN)
2303 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2308 static int mptcp_shutdown(struct socket *sock, int how)
2310 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2311 struct mptcp_subflow_context *subflow;
2314 pr_debug("sk=%p, how=%d", msk, how);
2316 lock_sock(sock->sk);
2319 if ((how & ~SHUTDOWN_MASK) || !how) {
2324 if (sock->state == SS_CONNECTING) {
2325 if ((1 << sock->sk->sk_state) &
2326 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2327 sock->state = SS_DISCONNECTING;
2329 sock->state = SS_CONNECTED;
2332 /* If we've already sent a FIN, or it's a closed state, skip this. */
2333 if (__mptcp_check_fallback(msk)) {
2334 if (how == SHUT_WR || how == SHUT_RDWR)
2335 inet_sk_state_store(sock->sk, TCP_FIN_WAIT1);
2337 mptcp_for_each_subflow(msk, subflow) {
2338 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2340 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2342 } else if ((how & SEND_SHUTDOWN) &&
2343 ((1 << sock->sk->sk_state) &
2344 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2345 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) &&
2346 mptcp_close_state(sock->sk)) {
2347 __mptcp_flush_join_list(msk);
2349 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2350 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2352 mptcp_for_each_subflow(msk, subflow) {
2353 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2355 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2359 /* Wake up anyone sleeping in poll. */
2360 sock->sk->sk_state_change(sock->sk);
2363 release_sock(sock->sk);
2368 static const struct proto_ops mptcp_stream_ops = {
2370 .owner = THIS_MODULE,
2371 .release = inet_release,
2373 .connect = mptcp_stream_connect,
2374 .socketpair = sock_no_socketpair,
2375 .accept = mptcp_stream_accept,
2376 .getname = inet_getname,
2378 .ioctl = inet_ioctl,
2379 .gettstamp = sock_gettstamp,
2380 .listen = mptcp_listen,
2381 .shutdown = mptcp_shutdown,
2382 .setsockopt = sock_common_setsockopt,
2383 .getsockopt = sock_common_getsockopt,
2384 .sendmsg = inet_sendmsg,
2385 .recvmsg = inet_recvmsg,
2386 .mmap = sock_no_mmap,
2387 .sendpage = inet_sendpage,
2390 static struct inet_protosw mptcp_protosw = {
2391 .type = SOCK_STREAM,
2392 .protocol = IPPROTO_MPTCP,
2393 .prot = &mptcp_prot,
2394 .ops = &mptcp_stream_ops,
2395 .flags = INET_PROTOSW_ICSK,
2398 void __init mptcp_proto_init(void)
2400 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2402 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2403 panic("Failed to allocate MPTCP pcpu counter\n");
2405 mptcp_subflow_init();
2409 if (proto_register(&mptcp_prot, 1) != 0)
2410 panic("Failed to register MPTCP proto.\n");
2412 inet_register_protosw(&mptcp_protosw);
2414 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2417 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2418 static const struct proto_ops mptcp_v6_stream_ops = {
2420 .owner = THIS_MODULE,
2421 .release = inet6_release,
2423 .connect = mptcp_stream_connect,
2424 .socketpair = sock_no_socketpair,
2425 .accept = mptcp_stream_accept,
2426 .getname = inet6_getname,
2428 .ioctl = inet6_ioctl,
2429 .gettstamp = sock_gettstamp,
2430 .listen = mptcp_listen,
2431 .shutdown = mptcp_shutdown,
2432 .setsockopt = sock_common_setsockopt,
2433 .getsockopt = sock_common_getsockopt,
2434 .sendmsg = inet6_sendmsg,
2435 .recvmsg = inet6_recvmsg,
2436 .mmap = sock_no_mmap,
2437 .sendpage = inet_sendpage,
2438 #ifdef CONFIG_COMPAT
2439 .compat_ioctl = inet6_compat_ioctl,
2443 static struct proto mptcp_v6_prot;
2445 static void mptcp_v6_destroy(struct sock *sk)
2448 inet6_destroy_sock(sk);
2451 static struct inet_protosw mptcp_v6_protosw = {
2452 .type = SOCK_STREAM,
2453 .protocol = IPPROTO_MPTCP,
2454 .prot = &mptcp_v6_prot,
2455 .ops = &mptcp_v6_stream_ops,
2456 .flags = INET_PROTOSW_ICSK,
2459 int __init mptcp_proto_v6_init(void)
2463 mptcp_v6_prot = mptcp_prot;
2464 strcpy(mptcp_v6_prot.name, "MPTCPv6");
2465 mptcp_v6_prot.slab = NULL;
2466 mptcp_v6_prot.destroy = mptcp_v6_destroy;
2467 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
2469 err = proto_register(&mptcp_v6_prot, 1);
2473 err = inet6_register_protosw(&mptcp_v6_protosw);
2475 proto_unregister(&mptcp_v6_prot);