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 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
20 #include <net/transp_v6.h>
22 #include <net/mptcp.h>
26 #define MPTCP_SAME_STATE TCP_MAX_STATES
28 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
30 struct mptcp_sock msk;
39 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
41 static struct percpu_counter mptcp_sockets_allocated;
43 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
44 * completed yet or has failed, return the subflow socket.
45 * Otherwise return NULL.
47 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
49 if (!msk->subflow || READ_ONCE(msk->can_ack))
55 static bool __mptcp_needs_tcp_fallback(const struct mptcp_sock *msk)
57 return msk->first && !sk_is_mptcp(msk->first);
60 static struct socket *mptcp_is_tcpsk(struct sock *sk)
62 struct socket *sock = sk->sk_socket;
67 if (unlikely(sk->sk_prot == &tcp_prot)) {
68 /* we are being invoked after mptcp_accept() has
69 * accepted a non-mp-capable flow: sk is a tcp_sk,
72 * Hand the socket over to tcp so all further socket ops
75 sock->ops = &inet_stream_ops;
77 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
78 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
79 sock->ops = &inet6_stream_ops;
87 static struct socket *__mptcp_tcp_fallback(struct mptcp_sock *msk)
91 sock_owned_by_me((const struct sock *)msk);
93 sock = mptcp_is_tcpsk((struct sock *)msk);
97 if (likely(!__mptcp_needs_tcp_fallback(msk)))
103 static bool __mptcp_can_create_subflow(const struct mptcp_sock *msk)
108 static struct socket *__mptcp_socket_create(struct mptcp_sock *msk, int state)
110 struct mptcp_subflow_context *subflow;
111 struct sock *sk = (struct sock *)msk;
112 struct socket *ssock;
115 ssock = __mptcp_tcp_fallback(msk);
119 ssock = __mptcp_nmpc_socket(msk);
123 if (!__mptcp_can_create_subflow(msk))
124 return ERR_PTR(-EINVAL);
126 err = mptcp_subflow_create_socket(sk, &ssock);
130 msk->first = ssock->sk;
131 msk->subflow = ssock;
132 subflow = mptcp_subflow_ctx(ssock->sk);
133 list_add(&subflow->node, &msk->conn_list);
134 subflow->request_mptcp = 1;
137 if (state != MPTCP_SAME_STATE)
138 inet_sk_state_store(sk, state);
142 static void __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
144 unsigned int offset, size_t copy_len)
146 struct sock *sk = (struct sock *)msk;
147 struct sk_buff *tail;
149 __skb_unlink(skb, &ssk->sk_receive_queue);
153 msk->ack_seq += copy_len;
155 tail = skb_peek_tail(&sk->sk_receive_queue);
156 if (offset == 0 && tail) {
160 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
161 kfree_skb_partial(skb, fragstolen);
162 atomic_add(delta, &sk->sk_rmem_alloc);
163 sk_mem_charge(sk, delta);
168 skb_set_owner_r(skb, sk);
169 __skb_queue_tail(&sk->sk_receive_queue, skb);
170 MPTCP_SKB_CB(skb)->offset = offset;
173 /* both sockets must be locked */
174 static bool mptcp_subflow_dsn_valid(const struct mptcp_sock *msk,
177 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
178 u64 dsn = mptcp_subflow_get_mapped_dsn(subflow);
180 /* revalidate data sequence number.
182 * mptcp_subflow_data_available() is usually called
183 * without msk lock. Its unlikely (but possible)
184 * that msk->ack_seq has been advanced since the last
185 * call found in-sequence data.
187 if (likely(dsn == msk->ack_seq))
190 subflow->data_avail = 0;
191 return mptcp_subflow_data_available(ssk);
194 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
198 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
199 struct sock *sk = (struct sock *)msk;
200 unsigned int moved = 0;
201 bool more_data_avail;
205 if (!mptcp_subflow_dsn_valid(msk, ssk)) {
210 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
211 int rcvbuf = max(ssk->sk_rcvbuf, sk->sk_rcvbuf);
213 if (rcvbuf > sk->sk_rcvbuf)
214 sk->sk_rcvbuf = rcvbuf;
219 u32 map_remaining, offset;
220 u32 seq = tp->copied_seq;
224 /* try to move as much data as available */
225 map_remaining = subflow->map_data_len -
226 mptcp_subflow_get_map_offset(subflow);
228 skb = skb_peek(&ssk->sk_receive_queue);
232 offset = seq - TCP_SKB_CB(skb)->seq;
233 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
239 if (offset < skb->len) {
240 size_t len = skb->len - offset;
245 __mptcp_move_skb(msk, ssk, skb, offset, len);
249 if (WARN_ON_ONCE(map_remaining < len))
253 sk_eat_skb(ssk, skb);
257 WRITE_ONCE(tp->copied_seq, seq);
258 more_data_avail = mptcp_subflow_data_available(ssk);
260 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf)) {
264 } while (more_data_avail);
271 /* In most cases we will be able to lock the mptcp socket. If its already
272 * owned, we need to defer to the work queue to avoid ABBA deadlock.
274 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
276 struct sock *sk = (struct sock *)msk;
277 unsigned int moved = 0;
279 if (READ_ONCE(sk->sk_lock.owned))
282 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
285 /* must re-check after taking the lock */
286 if (!READ_ONCE(sk->sk_lock.owned))
287 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
289 spin_unlock_bh(&sk->sk_lock.slock);
294 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
296 struct mptcp_sock *msk = mptcp_sk(sk);
298 set_bit(MPTCP_DATA_READY, &msk->flags);
300 if (atomic_read(&sk->sk_rmem_alloc) < READ_ONCE(sk->sk_rcvbuf) &&
301 move_skbs_to_msk(msk, ssk))
304 /* don't schedule if mptcp sk is (still) over limit */
305 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf))
308 /* mptcp socket is owned, release_cb should retry */
309 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
310 &sk->sk_tsq_flags)) {
313 /* need to try again, its possible release_cb() has already
314 * been called after the test_and_set_bit() above.
316 move_skbs_to_msk(msk, ssk);
319 sk->sk_data_ready(sk);
322 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
324 if (likely(list_empty(&msk->join_list)))
327 spin_lock_bh(&msk->join_list_lock);
328 list_splice_tail_init(&msk->join_list, &msk->conn_list);
329 spin_unlock_bh(&msk->join_list_lock);
332 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
334 long tout = ssk && inet_csk(ssk)->icsk_pending ?
335 inet_csk(ssk)->icsk_timeout - jiffies : 0;
338 tout = mptcp_sk(sk)->timer_ival;
339 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
342 static bool mptcp_timer_pending(struct sock *sk)
344 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
347 static void mptcp_reset_timer(struct sock *sk)
349 struct inet_connection_sock *icsk = inet_csk(sk);
352 /* should never be called with mptcp level timer cleared */
353 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
354 if (WARN_ON_ONCE(!tout))
356 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
359 void mptcp_data_acked(struct sock *sk)
361 mptcp_reset_timer(sk);
363 if (!sk_stream_is_writeable(sk) &&
364 schedule_work(&mptcp_sk(sk)->work))
368 void mptcp_subflow_eof(struct sock *sk)
370 struct mptcp_sock *msk = mptcp_sk(sk);
372 if (!test_and_set_bit(MPTCP_WORK_EOF, &msk->flags) &&
373 schedule_work(&msk->work))
377 static void mptcp_stop_timer(struct sock *sk)
379 struct inet_connection_sock *icsk = inet_csk(sk);
381 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
382 mptcp_sk(sk)->timer_ival = 0;
385 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
387 const struct sock *sk = (const struct sock *)msk;
389 if (!msk->cached_ext)
390 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
392 return !!msk->cached_ext;
395 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
397 struct mptcp_subflow_context *subflow;
398 struct sock *sk = (struct sock *)msk;
400 sock_owned_by_me(sk);
402 mptcp_for_each_subflow(msk, subflow) {
403 if (subflow->data_avail)
404 return mptcp_subflow_tcp_sock(subflow);
410 static bool mptcp_skb_can_collapse_to(u64 write_seq,
411 const struct sk_buff *skb,
412 const struct mptcp_ext *mpext)
414 if (!tcp_skb_can_collapse_to(skb))
417 /* can collapse only if MPTCP level sequence is in order */
418 return mpext && mpext->data_seq + mpext->data_len == write_seq;
421 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
422 const struct page_frag *pfrag,
423 const struct mptcp_data_frag *df)
425 return df && pfrag->page == df->page &&
426 df->data_seq + df->data_len == msk->write_seq;
429 static void dfrag_uncharge(struct sock *sk, int len)
431 sk_mem_uncharge(sk, len);
432 sk_wmem_queued_add(sk, -len);
435 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
437 int len = dfrag->data_len + dfrag->overhead;
439 list_del(&dfrag->list);
440 dfrag_uncharge(sk, len);
441 put_page(dfrag->page);
444 static void mptcp_clean_una(struct sock *sk)
446 struct mptcp_sock *msk = mptcp_sk(sk);
447 struct mptcp_data_frag *dtmp, *dfrag;
448 u64 snd_una = atomic64_read(&msk->snd_una);
449 bool cleaned = false;
451 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
452 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
455 dfrag_clear(sk, dfrag);
459 dfrag = mptcp_rtx_head(sk);
460 if (dfrag && after64(snd_una, dfrag->data_seq)) {
461 u64 delta = dfrag->data_seq + dfrag->data_len - snd_una;
463 dfrag->data_seq += delta;
464 dfrag->data_len -= delta;
466 dfrag_uncharge(sk, delta);
471 sk_mem_reclaim_partial(sk);
473 /* Only wake up writers if a subflow is ready */
474 if (test_bit(MPTCP_SEND_SPACE, &msk->flags))
475 sk_stream_write_space(sk);
479 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
482 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
484 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
485 pfrag, sk->sk_allocation)))
488 sk->sk_prot->enter_memory_pressure(sk);
489 sk_stream_moderate_sndbuf(sk);
493 static struct mptcp_data_frag *
494 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
497 int offset = ALIGN(orig_offset, sizeof(long));
498 struct mptcp_data_frag *dfrag;
500 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
502 dfrag->data_seq = msk->write_seq;
503 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
504 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
505 dfrag->page = pfrag->page;
510 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
511 struct msghdr *msg, struct mptcp_data_frag *dfrag,
512 long *timeo, int *pmss_now,
515 int mss_now, avail_size, size_goal, offset, ret, frag_truesize = 0;
516 bool dfrag_collapsed, can_collapse = false;
517 struct mptcp_sock *msk = mptcp_sk(sk);
518 struct mptcp_ext *mpext = NULL;
519 bool retransmission = !!dfrag;
520 struct sk_buff *skb, *tail;
521 struct page_frag *pfrag;
526 /* use the mptcp page cache so that we can easily move the data
527 * from one substream to another, but do per subflow memory accounting
528 * Note: pfrag is used only !retransmission, but the compiler if
529 * fooled into a warning if we don't init here
531 pfrag = sk_page_frag(sk);
532 if (!retransmission) {
533 write_seq = &msk->write_seq;
536 write_seq = &dfrag->data_seq;
540 /* compute copy limit */
541 mss_now = tcp_send_mss(ssk, &size_goal, msg->msg_flags);
543 *ps_goal = size_goal;
544 avail_size = size_goal;
545 skb = tcp_write_queue_tail(ssk);
547 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
549 /* Limit the write to the size available in the
550 * current skb, if any, so that we create at most a new skb.
551 * Explicitly tells TCP internals to avoid collapsing on later
552 * queue management operation, to avoid breaking the ext <->
553 * SSN association set here
555 can_collapse = (size_goal - skb->len > 0) &&
556 mptcp_skb_can_collapse_to(*write_seq, skb, mpext);
558 TCP_SKB_CB(skb)->eor = 1;
560 avail_size = size_goal - skb->len;
563 if (!retransmission) {
564 /* reuse tail pfrag, if possible, or carve a new one from the
567 dfrag = mptcp_rtx_tail(sk);
568 offset = pfrag->offset;
569 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
570 if (!dfrag_collapsed) {
571 dfrag = mptcp_carve_data_frag(msk, pfrag, offset);
572 offset = dfrag->offset;
573 frag_truesize = dfrag->overhead;
575 psize = min_t(size_t, pfrag->size - offset, avail_size);
578 pr_debug("left=%zu", msg_data_left(msg));
579 psize = copy_page_from_iter(pfrag->page, offset,
580 min_t(size_t, msg_data_left(msg),
583 pr_debug("left=%zu", msg_data_left(msg));
587 if (!sk_wmem_schedule(sk, psize + dfrag->overhead))
590 offset = dfrag->offset;
591 psize = min_t(size_t, dfrag->data_len, avail_size);
594 /* tell the TCP stack to delay the push so that we can safely
595 * access the skb after the sendpages call
597 ret = do_tcp_sendpages(ssk, page, offset, psize,
598 msg->msg_flags | MSG_SENDPAGE_NOTLAST | MSG_DONTWAIT);
602 frag_truesize += ret;
603 if (!retransmission) {
604 if (unlikely(ret < psize))
605 iov_iter_revert(&msg->msg_iter, psize - ret);
607 /* send successful, keep track of sent data for mptcp-level
610 dfrag->data_len += ret;
611 if (!dfrag_collapsed) {
612 get_page(dfrag->page);
613 list_add_tail(&dfrag->list, &msk->rtx_queue);
614 sk_wmem_queued_add(sk, frag_truesize);
616 sk_wmem_queued_add(sk, ret);
619 /* charge data on mptcp rtx queue to the master socket
620 * Note: we charge such data both to sk and ssk
622 sk->sk_forward_alloc -= frag_truesize;
625 /* if the tail skb extension is still the cached one, collapsing
626 * really happened. Note: we can't check for 'same skb' as the sk_buff
627 * hdr on tail can be transmitted, freed and re-allocated by the
628 * do_tcp_sendpages() call
630 tail = tcp_write_queue_tail(ssk);
631 if (mpext && tail && mpext == skb_ext_find(tail, SKB_EXT_MPTCP)) {
632 WARN_ON_ONCE(!can_collapse);
633 mpext->data_len += ret;
637 skb = tcp_write_queue_tail(ssk);
638 mpext = __skb_ext_set(skb, SKB_EXT_MPTCP, msk->cached_ext);
639 msk->cached_ext = NULL;
641 memset(mpext, 0, sizeof(*mpext));
642 mpext->data_seq = *write_seq;
643 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
644 mpext->data_len = ret;
648 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
649 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
654 pfrag->offset += frag_truesize;
656 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
661 static void mptcp_nospace(struct mptcp_sock *msk, struct socket *sock)
663 clear_bit(MPTCP_SEND_SPACE, &msk->flags);
664 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
666 /* enables sk->write_space() callbacks */
667 set_bit(SOCK_NOSPACE, &sock->flags);
670 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
672 struct mptcp_subflow_context *subflow;
673 struct sock *backup = NULL;
675 sock_owned_by_me((const struct sock *)msk);
677 if (!mptcp_ext_cache_refill(msk))
680 mptcp_for_each_subflow(msk, subflow) {
681 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
683 if (!sk_stream_memory_free(ssk)) {
684 struct socket *sock = ssk->sk_socket;
687 mptcp_nospace(msk, sock);
692 if (subflow->backup) {
705 static void ssk_check_wmem(struct mptcp_sock *msk, struct sock *ssk)
709 if (likely(sk_stream_is_writeable(ssk)))
712 sock = READ_ONCE(ssk->sk_socket);
714 mptcp_nospace(msk, sock);
717 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
719 int mss_now = 0, size_goal = 0, ret = 0;
720 struct mptcp_sock *msk = mptcp_sk(sk);
721 struct page_frag *pfrag;
722 struct socket *ssock;
728 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
733 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
735 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
736 ret = sk_stream_wait_connect(sk, &timeo);
742 ssock = __mptcp_tcp_fallback(msk);
743 if (unlikely(ssock)) {
745 pr_debug("fallback passthrough");
746 ret = sock_sendmsg(ssock, msg);
747 return ret >= 0 ? ret + copied : (copied ? copied : ret);
750 pfrag = sk_page_frag(sk);
755 __mptcp_flush_join_list(msk);
756 ssk = mptcp_subflow_get_send(msk);
757 while (!sk_stream_memory_free(sk) ||
759 !mptcp_page_frag_refill(ssk, pfrag)) {
761 /* make sure retransmit timer is
762 * running before we wait for memory.
764 * The retransmit timer might be needed
765 * to make the peer send an up-to-date
768 mptcp_set_timeout(sk, ssk);
769 if (!mptcp_timer_pending(sk))
770 mptcp_reset_timer(sk);
773 ret = sk_stream_wait_memory(sk, &timeo);
779 ssk = mptcp_subflow_get_send(msk);
780 if (list_empty(&msk->conn_list)) {
786 pr_debug("conn_list->subflow=%p", ssk);
789 tx_ok = msg_data_left(msg);
791 ret = mptcp_sendmsg_frag(sk, ssk, msg, NULL, &timeo, &mss_now,
794 if (ret == -EAGAIN && timeo > 0) {
795 mptcp_set_timeout(sk, ssk);
801 if (ret == 0 && unlikely(__mptcp_needs_tcp_fallback(msk))) {
802 /* Can happen for passive sockets:
803 * 3WHS negotiated MPTCP, but first packet after is
804 * plain TCP (e.g. due to middlebox filtering unknown
815 tx_ok = msg_data_left(msg);
819 if (!sk_stream_memory_free(ssk) ||
820 !mptcp_page_frag_refill(ssk, pfrag) ||
821 !mptcp_ext_cache_refill(msk)) {
822 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
823 tcp_push(ssk, msg->msg_flags, mss_now,
824 tcp_sk(ssk)->nonagle, size_goal);
825 mptcp_set_timeout(sk, ssk);
830 /* memory is charged to mptcp level socket as well, i.e.
831 * if msg is very large, mptcp socket may run out of buffer
832 * space. mptcp_clean_una() will release data that has
833 * been acked at mptcp level in the mean time, so there is
834 * a good chance we can continue sending data right away.
836 * Normally, when the tcp subflow can accept more data, then
837 * so can the MPTCP socket. However, we need to cope with
838 * peers that might lag behind in their MPTCP-level
839 * acknowledgements, i.e. data might have been acked at
840 * tcp level only. So, we must also check the MPTCP socket
841 * limits before we send more data.
843 if (unlikely(!sk_stream_memory_free(sk))) {
844 tcp_push(ssk, msg->msg_flags, mss_now,
845 tcp_sk(ssk)->nonagle, size_goal);
847 if (!sk_stream_memory_free(sk)) {
848 /* can't send more for now, need to wait for
849 * MPTCP-level ACKs from peer.
851 * Wakeup will happen via mptcp_clean_una().
853 mptcp_set_timeout(sk, ssk);
855 goto wait_for_sndbuf;
860 mptcp_set_timeout(sk, ssk);
863 tcp_push(ssk, msg->msg_flags, mss_now, tcp_sk(ssk)->nonagle,
866 /* start the timer, if it's not pending */
867 if (!mptcp_timer_pending(sk))
868 mptcp_reset_timer(sk);
871 ssk_check_wmem(msk, ssk);
878 static void mptcp_wait_data(struct sock *sk, long *timeo)
880 DEFINE_WAIT_FUNC(wait, woken_wake_function);
881 struct mptcp_sock *msk = mptcp_sk(sk);
883 add_wait_queue(sk_sleep(sk), &wait);
884 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
886 sk_wait_event(sk, timeo,
887 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
889 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
890 remove_wait_queue(sk_sleep(sk), &wait);
893 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
897 struct sock *sk = (struct sock *)msk;
901 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
902 u32 offset = MPTCP_SKB_CB(skb)->offset;
903 u32 data_len = skb->len - offset;
904 u32 count = min_t(size_t, len - copied, data_len);
907 err = skb_copy_datagram_msg(skb, offset, msg, count);
908 if (unlikely(err < 0)) {
916 if (count < data_len) {
917 MPTCP_SKB_CB(skb)->offset += count;
921 __skb_unlink(skb, &sk->sk_receive_queue);
931 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
933 unsigned int moved = 0;
937 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
943 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
950 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
951 int nonblock, int flags, int *addr_len)
953 struct mptcp_sock *msk = mptcp_sk(sk);
954 struct socket *ssock;
959 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
963 ssock = __mptcp_tcp_fallback(msk);
964 if (unlikely(ssock)) {
967 pr_debug("fallback-read subflow=%p",
968 mptcp_subflow_ctx(ssock->sk));
969 copied = sock_recvmsg(ssock, msg, flags);
973 timeo = sock_rcvtimeo(sk, nonblock);
975 len = min_t(size_t, len, INT_MAX);
976 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
977 __mptcp_flush_join_list(msk);
979 while (len > (size_t)copied) {
982 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
983 if (unlikely(bytes_read < 0)) {
989 copied += bytes_read;
991 if (skb_queue_empty(&sk->sk_receive_queue) &&
992 __mptcp_move_skbs(msk))
995 /* only the master socket status is relevant here. The exit
996 * conditions mirror closely tcp_recvmsg()
998 if (copied >= target)
1003 sk->sk_state == TCP_CLOSE ||
1004 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1006 signal_pending(current))
1010 copied = sock_error(sk);
1014 if (sk->sk_shutdown & RCV_SHUTDOWN)
1017 if (sk->sk_state == TCP_CLOSE) {
1027 if (signal_pending(current)) {
1028 copied = sock_intr_errno(timeo);
1033 pr_debug("block timeout %ld", timeo);
1034 mptcp_wait_data(sk, &timeo);
1035 ssock = __mptcp_tcp_fallback(msk);
1036 if (unlikely(ssock))
1040 if (skb_queue_empty(&sk->sk_receive_queue)) {
1041 /* entire backlog drained, clear DATA_READY. */
1042 clear_bit(MPTCP_DATA_READY, &msk->flags);
1044 /* .. race-breaker: ssk might have gotten new data
1045 * after last __mptcp_move_skbs() returned false.
1047 if (unlikely(__mptcp_move_skbs(msk)))
1048 set_bit(MPTCP_DATA_READY, &msk->flags);
1049 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1050 /* data to read but mptcp_wait_data() cleared DATA_READY */
1051 set_bit(MPTCP_DATA_READY, &msk->flags);
1058 static void mptcp_retransmit_handler(struct sock *sk)
1060 struct mptcp_sock *msk = mptcp_sk(sk);
1062 if (atomic64_read(&msk->snd_una) == msk->write_seq) {
1063 mptcp_stop_timer(sk);
1065 set_bit(MPTCP_WORK_RTX, &msk->flags);
1066 if (schedule_work(&msk->work))
1071 static void mptcp_retransmit_timer(struct timer_list *t)
1073 struct inet_connection_sock *icsk = from_timer(icsk, t,
1074 icsk_retransmit_timer);
1075 struct sock *sk = &icsk->icsk_inet.sk;
1078 if (!sock_owned_by_user(sk)) {
1079 mptcp_retransmit_handler(sk);
1081 /* delegate our work to tcp_release_cb() */
1082 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1090 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1093 * A backup subflow is returned only if that is the only kind available.
1095 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1097 struct mptcp_subflow_context *subflow;
1098 struct sock *backup = NULL;
1100 sock_owned_by_me((const struct sock *)msk);
1102 mptcp_for_each_subflow(msk, subflow) {
1103 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1105 /* still data outstanding at TCP level? Don't retransmit. */
1106 if (!tcp_write_queue_empty(ssk))
1109 if (subflow->backup) {
1121 /* subflow sockets can be either outgoing (connect) or incoming
1124 * Outgoing subflows use in-kernel sockets.
1125 * Incoming subflows do not have their own 'struct socket' allocated,
1126 * so we need to use tcp_close() after detaching them from the mptcp
1129 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1130 struct mptcp_subflow_context *subflow,
1133 struct socket *sock = READ_ONCE(ssk->sk_socket);
1135 list_del(&subflow->node);
1137 if (sock && sock != sk->sk_socket) {
1138 /* outgoing subflow */
1141 /* incoming subflow */
1142 tcp_close(ssk, timeout);
1146 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1151 static void mptcp_check_for_eof(struct mptcp_sock *msk)
1153 struct mptcp_subflow_context *subflow;
1154 struct sock *sk = (struct sock *)msk;
1157 mptcp_for_each_subflow(msk, subflow)
1158 receivers += !subflow->rx_eof;
1160 if (!receivers && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1161 /* hopefully temporary hack: propagate shutdown status
1162 * to msk, when all subflows agree on it
1164 sk->sk_shutdown |= RCV_SHUTDOWN;
1166 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
1167 set_bit(MPTCP_DATA_READY, &msk->flags);
1168 sk->sk_data_ready(sk);
1172 static void mptcp_worker(struct work_struct *work)
1174 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1175 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1176 int orig_len, orig_offset, mss_now = 0, size_goal = 0;
1177 struct mptcp_data_frag *dfrag;
1184 mptcp_clean_una(sk);
1185 __mptcp_flush_join_list(msk);
1186 __mptcp_move_skbs(msk);
1188 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1189 mptcp_check_for_eof(msk);
1191 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1194 dfrag = mptcp_rtx_head(sk);
1198 if (!mptcp_ext_cache_refill(msk))
1201 ssk = mptcp_subflow_get_retrans(msk);
1207 msg.msg_flags = MSG_DONTWAIT;
1208 orig_len = dfrag->data_len;
1209 orig_offset = dfrag->offset;
1210 orig_write_seq = dfrag->data_seq;
1211 while (dfrag->data_len > 0) {
1212 int ret = mptcp_sendmsg_frag(sk, ssk, &msg, dfrag, &timeo,
1213 &mss_now, &size_goal);
1217 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1219 dfrag->data_len -= ret;
1220 dfrag->offset += ret;
1222 if (!mptcp_ext_cache_refill(msk))
1226 tcp_push(ssk, msg.msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1229 dfrag->data_seq = orig_write_seq;
1230 dfrag->offset = orig_offset;
1231 dfrag->data_len = orig_len;
1233 mptcp_set_timeout(sk, ssk);
1237 if (!mptcp_timer_pending(sk))
1238 mptcp_reset_timer(sk);
1245 static int __mptcp_init_sock(struct sock *sk)
1247 struct mptcp_sock *msk = mptcp_sk(sk);
1249 spin_lock_init(&msk->join_list_lock);
1251 INIT_LIST_HEAD(&msk->conn_list);
1252 INIT_LIST_HEAD(&msk->join_list);
1253 INIT_LIST_HEAD(&msk->rtx_queue);
1254 __set_bit(MPTCP_SEND_SPACE, &msk->flags);
1255 INIT_WORK(&msk->work, mptcp_worker);
1258 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1260 mptcp_pm_data_init(msk);
1262 /* re-use the csk retrans timer for MPTCP-level retrans */
1263 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1268 static int mptcp_init_sock(struct sock *sk)
1270 struct net *net = sock_net(sk);
1273 if (!mptcp_is_enabled(net))
1274 return -ENOPROTOOPT;
1276 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1279 ret = __mptcp_init_sock(sk);
1283 sk_sockets_allocated_inc(sk);
1284 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[2];
1289 static void __mptcp_clear_xmit(struct sock *sk)
1291 struct mptcp_sock *msk = mptcp_sk(sk);
1292 struct mptcp_data_frag *dtmp, *dfrag;
1294 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
1296 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
1297 dfrag_clear(sk, dfrag);
1300 static void mptcp_cancel_work(struct sock *sk)
1302 struct mptcp_sock *msk = mptcp_sk(sk);
1304 if (cancel_work_sync(&msk->work))
1308 static void mptcp_subflow_shutdown(struct sock *ssk, int how,
1309 bool data_fin_tx_enable, u64 data_fin_tx_seq)
1313 switch (ssk->sk_state) {
1315 if (!(how & RCV_SHUTDOWN))
1319 tcp_disconnect(ssk, O_NONBLOCK);
1322 if (data_fin_tx_enable) {
1323 struct mptcp_subflow_context *subflow;
1325 subflow = mptcp_subflow_ctx(ssk);
1326 subflow->data_fin_tx_seq = data_fin_tx_seq;
1327 subflow->data_fin_tx_enable = 1;
1330 ssk->sk_shutdown |= how;
1331 tcp_shutdown(ssk, how);
1335 /* Wake up anyone sleeping in poll. */
1336 ssk->sk_state_change(ssk);
1340 /* Called with msk lock held, releases such lock before returning */
1341 static void mptcp_close(struct sock *sk, long timeout)
1343 struct mptcp_subflow_context *subflow, *tmp;
1344 struct mptcp_sock *msk = mptcp_sk(sk);
1345 LIST_HEAD(conn_list);
1346 u64 data_fin_tx_seq;
1350 inet_sk_state_store(sk, TCP_CLOSE);
1352 /* be sure to always acquire the join list lock, to sync vs
1353 * mptcp_finish_join().
1355 spin_lock_bh(&msk->join_list_lock);
1356 list_splice_tail_init(&msk->join_list, &msk->conn_list);
1357 spin_unlock_bh(&msk->join_list_lock);
1358 list_splice_init(&msk->conn_list, &conn_list);
1360 data_fin_tx_seq = msk->write_seq;
1362 __mptcp_clear_xmit(sk);
1366 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
1367 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1369 subflow->data_fin_tx_seq = data_fin_tx_seq;
1370 subflow->data_fin_tx_enable = 1;
1371 __mptcp_close_ssk(sk, ssk, subflow, timeout);
1374 mptcp_cancel_work(sk);
1375 mptcp_pm_close(msk);
1377 __skb_queue_purge(&sk->sk_receive_queue);
1379 sk_common_release(sk);
1382 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
1384 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1385 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
1386 struct ipv6_pinfo *msk6 = inet6_sk(msk);
1388 msk->sk_v6_daddr = ssk->sk_v6_daddr;
1389 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
1392 msk6->saddr = ssk6->saddr;
1393 msk6->flow_label = ssk6->flow_label;
1397 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
1398 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
1399 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
1400 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
1401 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
1402 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
1405 static int mptcp_disconnect(struct sock *sk, int flags)
1407 /* Should never be called.
1408 * inet_stream_connect() calls ->disconnect, but that
1409 * refers to the subflow socket, not the mptcp one.
1415 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1416 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
1418 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
1420 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
1424 struct sock *mptcp_sk_clone(const struct sock *sk,
1425 const struct mptcp_options_received *mp_opt,
1426 struct request_sock *req)
1428 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
1429 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
1430 struct mptcp_sock *msk;
1436 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1437 if (nsk->sk_family == AF_INET6)
1438 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
1441 __mptcp_init_sock(nsk);
1443 msk = mptcp_sk(nsk);
1444 msk->local_key = subflow_req->local_key;
1445 msk->token = subflow_req->token;
1446 msk->subflow = NULL;
1448 if (unlikely(mptcp_token_new_accept(subflow_req->token, nsk))) {
1449 nsk->sk_state = TCP_CLOSE;
1450 bh_unlock_sock(nsk);
1452 /* we can't call into mptcp_close() here - possible BH context
1453 * free the sock directly.
1454 * sk_clone_lock() sets nsk refcnt to two, hence call sk_free()
1457 sk_common_release(nsk);
1462 msk->write_seq = subflow_req->idsn + 1;
1463 atomic64_set(&msk->snd_una, msk->write_seq);
1464 if (mp_opt->mp_capable) {
1465 msk->can_ack = true;
1466 msk->remote_key = mp_opt->sndr_key;
1467 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
1469 msk->ack_seq = ack_seq;
1472 sock_reset_flag(nsk, SOCK_RCU_FREE);
1473 /* will be fully established after successful MPC subflow creation */
1474 inet_sk_state_store(nsk, TCP_SYN_RECV);
1475 bh_unlock_sock(nsk);
1477 /* keep a single reference */
1482 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
1485 struct mptcp_sock *msk = mptcp_sk(sk);
1486 struct socket *listener;
1489 listener = __mptcp_nmpc_socket(msk);
1490 if (WARN_ON_ONCE(!listener)) {
1495 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
1496 newsk = inet_csk_accept(listener->sk, flags, err, kern);
1500 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
1502 if (sk_is_mptcp(newsk)) {
1503 struct mptcp_subflow_context *subflow;
1504 struct sock *new_mptcp_sock;
1505 struct sock *ssk = newsk;
1507 subflow = mptcp_subflow_ctx(newsk);
1508 new_mptcp_sock = subflow->conn;
1510 /* is_mptcp should be false if subflow->conn is missing, see
1511 * subflow_syn_recv_sock()
1513 if (WARN_ON_ONCE(!new_mptcp_sock)) {
1514 tcp_sk(newsk)->is_mptcp = 0;
1518 /* acquire the 2nd reference for the owning socket */
1519 sock_hold(new_mptcp_sock);
1522 bh_lock_sock(new_mptcp_sock);
1523 msk = mptcp_sk(new_mptcp_sock);
1526 newsk = new_mptcp_sock;
1527 mptcp_copy_inaddrs(newsk, ssk);
1528 list_add(&subflow->node, &msk->conn_list);
1529 inet_sk_state_store(newsk, TCP_ESTABLISHED);
1531 bh_unlock_sock(new_mptcp_sock);
1533 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
1536 MPTCP_INC_STATS(sock_net(sk),
1537 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
1543 static void mptcp_destroy(struct sock *sk)
1545 struct mptcp_sock *msk = mptcp_sk(sk);
1547 mptcp_token_destroy(msk->token);
1548 if (msk->cached_ext)
1549 __skb_ext_put(msk->cached_ext);
1551 sk_sockets_allocated_dec(sk);
1554 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
1555 char __user *optval, unsigned int optlen)
1557 struct mptcp_sock *msk = mptcp_sk(sk);
1558 struct socket *ssock;
1560 pr_debug("msk=%p", msk);
1562 /* @@ the meaning of setsockopt() when the socket is connected and
1563 * there are multiple subflows is not yet defined. It is up to the
1564 * MPTCP-level socket to configure the subflows until the subflow
1565 * is in TCP fallback, when TCP socket options are passed through
1566 * to the one remaining subflow.
1569 ssock = __mptcp_tcp_fallback(msk);
1572 return tcp_setsockopt(ssock->sk, level, optname, optval,
1578 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
1579 char __user *optval, int __user *option)
1581 struct mptcp_sock *msk = mptcp_sk(sk);
1582 struct socket *ssock;
1584 pr_debug("msk=%p", msk);
1586 /* @@ the meaning of setsockopt() when the socket is connected and
1587 * there are multiple subflows is not yet defined. It is up to the
1588 * MPTCP-level socket to configure the subflows until the subflow
1589 * is in TCP fallback, when socket options are passed through
1590 * to the one remaining subflow.
1593 ssock = __mptcp_tcp_fallback(msk);
1596 return tcp_getsockopt(ssock->sk, level, optname, optval,
1602 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
1603 TCPF_WRITE_TIMER_DEFERRED)
1605 /* this is very alike tcp_release_cb() but we must handle differently a
1606 * different set of events
1608 static void mptcp_release_cb(struct sock *sk)
1610 unsigned long flags, nflags;
1613 flags = sk->sk_tsq_flags;
1614 if (!(flags & MPTCP_DEFERRED_ALL))
1616 nflags = flags & ~MPTCP_DEFERRED_ALL;
1617 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1619 sock_release_ownership(sk);
1621 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1622 struct mptcp_sock *msk = mptcp_sk(sk);
1625 ssk = mptcp_subflow_recv_lookup(msk);
1626 if (!ssk || !schedule_work(&msk->work))
1630 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1631 mptcp_retransmit_handler(sk);
1636 static int mptcp_get_port(struct sock *sk, unsigned short snum)
1638 struct mptcp_sock *msk = mptcp_sk(sk);
1639 struct socket *ssock;
1641 ssock = __mptcp_nmpc_socket(msk);
1642 pr_debug("msk=%p, subflow=%p", msk, ssock);
1643 if (WARN_ON_ONCE(!ssock))
1646 return inet_csk_get_port(ssock->sk, snum);
1649 void mptcp_finish_connect(struct sock *ssk)
1651 struct mptcp_subflow_context *subflow;
1652 struct mptcp_sock *msk;
1656 subflow = mptcp_subflow_ctx(ssk);
1660 if (!subflow->mp_capable) {
1661 MPTCP_INC_STATS(sock_net(sk),
1662 MPTCP_MIB_MPCAPABLEACTIVEFALLBACK);
1666 pr_debug("msk=%p, token=%u", sk, subflow->token);
1668 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
1670 subflow->map_seq = ack_seq;
1671 subflow->map_subflow_seq = 1;
1672 subflow->rel_write_seq = 1;
1674 /* the socket is not connected yet, no msk/subflow ops can access/race
1675 * accessing the field below
1677 WRITE_ONCE(msk->remote_key, subflow->remote_key);
1678 WRITE_ONCE(msk->local_key, subflow->local_key);
1679 WRITE_ONCE(msk->token, subflow->token);
1680 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
1681 WRITE_ONCE(msk->ack_seq, ack_seq);
1682 WRITE_ONCE(msk->can_ack, 1);
1683 atomic64_set(&msk->snd_una, msk->write_seq);
1685 mptcp_pm_new_connection(msk, 0);
1688 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
1690 write_lock_bh(&sk->sk_callback_lock);
1691 rcu_assign_pointer(sk->sk_wq, &parent->wq);
1692 sk_set_socket(sk, parent);
1693 sk->sk_uid = SOCK_INODE(parent)->i_uid;
1694 write_unlock_bh(&sk->sk_callback_lock);
1697 bool mptcp_finish_join(struct sock *sk)
1699 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
1700 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
1701 struct sock *parent = (void *)msk;
1702 struct socket *parent_sock;
1705 pr_debug("msk=%p, subflow=%p", msk, subflow);
1707 /* mptcp socket already closing? */
1708 if (inet_sk_state_load(parent) != TCP_ESTABLISHED)
1711 if (!msk->pm.server_side)
1714 if (!mptcp_pm_allow_new_subflow(msk))
1717 /* active connections are already on conn_list, and we can't acquire
1719 * use the join list lock as synchronization point and double-check
1720 * msk status to avoid racing with mptcp_close()
1722 spin_lock_bh(&msk->join_list_lock);
1723 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
1724 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node)))
1725 list_add_tail(&subflow->node, &msk->join_list);
1726 spin_unlock_bh(&msk->join_list_lock);
1730 /* attach to msk socket only after we are sure he will deal with us
1733 parent_sock = READ_ONCE(parent->sk_socket);
1734 if (parent_sock && !sk->sk_socket)
1735 mptcp_sock_graft(sk, parent_sock);
1736 subflow->map_seq = msk->ack_seq;
1740 static bool mptcp_memory_free(const struct sock *sk, int wake)
1742 struct mptcp_sock *msk = mptcp_sk(sk);
1744 return wake ? test_bit(MPTCP_SEND_SPACE, &msk->flags) : true;
1747 static struct proto mptcp_prot = {
1749 .owner = THIS_MODULE,
1750 .init = mptcp_init_sock,
1751 .disconnect = mptcp_disconnect,
1752 .close = mptcp_close,
1753 .accept = mptcp_accept,
1754 .setsockopt = mptcp_setsockopt,
1755 .getsockopt = mptcp_getsockopt,
1756 .shutdown = tcp_shutdown,
1757 .destroy = mptcp_destroy,
1758 .sendmsg = mptcp_sendmsg,
1759 .recvmsg = mptcp_recvmsg,
1760 .release_cb = mptcp_release_cb,
1762 .unhash = inet_unhash,
1763 .get_port = mptcp_get_port,
1764 .sockets_allocated = &mptcp_sockets_allocated,
1765 .memory_allocated = &tcp_memory_allocated,
1766 .memory_pressure = &tcp_memory_pressure,
1767 .stream_memory_free = mptcp_memory_free,
1768 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
1769 .sysctl_mem = sysctl_tcp_mem,
1770 .obj_size = sizeof(struct mptcp_sock),
1771 .no_autobind = true,
1774 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
1776 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1777 struct socket *ssock;
1780 lock_sock(sock->sk);
1781 ssock = __mptcp_socket_create(msk, MPTCP_SAME_STATE);
1782 if (IS_ERR(ssock)) {
1783 err = PTR_ERR(ssock);
1787 err = ssock->ops->bind(ssock, uaddr, addr_len);
1789 mptcp_copy_inaddrs(sock->sk, ssock->sk);
1792 release_sock(sock->sk);
1796 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
1797 int addr_len, int flags)
1799 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1800 struct socket *ssock;
1803 lock_sock(sock->sk);
1804 if (sock->state != SS_UNCONNECTED && msk->subflow) {
1805 /* pending connection or invalid state, let existing subflow
1808 ssock = msk->subflow;
1812 ssock = __mptcp_socket_create(msk, TCP_SYN_SENT);
1813 if (IS_ERR(ssock)) {
1814 err = PTR_ERR(ssock);
1818 #ifdef CONFIG_TCP_MD5SIG
1819 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
1822 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
1823 mptcp_subflow_ctx(ssock->sk)->request_mptcp = 0;
1827 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
1828 sock->state = ssock->state;
1830 /* on successful connect, the msk state will be moved to established by
1831 * subflow_finish_connect()
1833 if (!err || err == EINPROGRESS)
1834 mptcp_copy_inaddrs(sock->sk, ssock->sk);
1836 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
1839 release_sock(sock->sk);
1843 static int mptcp_v4_getname(struct socket *sock, struct sockaddr *uaddr,
1846 if (sock->sk->sk_prot == &tcp_prot) {
1847 /* we are being invoked from __sys_accept4, after
1848 * mptcp_accept() has just accepted a non-mp-capable
1849 * flow: sk is a tcp_sk, not an mptcp one.
1851 * Hand the socket over to tcp so all further socket ops
1854 sock->ops = &inet_stream_ops;
1857 return inet_getname(sock, uaddr, peer);
1860 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1861 static int mptcp_v6_getname(struct socket *sock, struct sockaddr *uaddr,
1864 if (sock->sk->sk_prot == &tcpv6_prot) {
1865 /* we are being invoked from __sys_accept4 after
1866 * mptcp_accept() has accepted a non-mp-capable
1867 * subflow: sk is a tcp_sk, not mptcp.
1869 * Hand the socket over to tcp so all further
1870 * socket ops bypass mptcp.
1872 sock->ops = &inet6_stream_ops;
1875 return inet6_getname(sock, uaddr, peer);
1879 static int mptcp_listen(struct socket *sock, int backlog)
1881 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1882 struct socket *ssock;
1885 pr_debug("msk=%p", msk);
1887 lock_sock(sock->sk);
1888 ssock = __mptcp_socket_create(msk, TCP_LISTEN);
1889 if (IS_ERR(ssock)) {
1890 err = PTR_ERR(ssock);
1894 sock_set_flag(sock->sk, SOCK_RCU_FREE);
1896 err = ssock->ops->listen(ssock, backlog);
1897 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
1899 mptcp_copy_inaddrs(sock->sk, ssock->sk);
1902 release_sock(sock->sk);
1906 static bool is_tcp_proto(const struct proto *p)
1908 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1909 return p == &tcp_prot || p == &tcpv6_prot;
1911 return p == &tcp_prot;
1915 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
1916 int flags, bool kern)
1918 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1919 struct socket *ssock;
1922 pr_debug("msk=%p", msk);
1924 lock_sock(sock->sk);
1925 if (sock->sk->sk_state != TCP_LISTEN)
1928 ssock = __mptcp_nmpc_socket(msk);
1932 sock_hold(ssock->sk);
1933 release_sock(sock->sk);
1935 err = ssock->ops->accept(sock, newsock, flags, kern);
1936 if (err == 0 && !is_tcp_proto(newsock->sk->sk_prot)) {
1937 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
1938 struct mptcp_subflow_context *subflow;
1940 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
1941 * This is needed so NOSPACE flag can be set from tcp stack.
1943 __mptcp_flush_join_list(msk);
1944 list_for_each_entry(subflow, &msk->conn_list, node) {
1945 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1947 if (!ssk->sk_socket)
1948 mptcp_sock_graft(ssk, newsock);
1952 sock_put(ssock->sk);
1956 release_sock(sock->sk);
1960 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
1961 struct poll_table_struct *wait)
1963 struct sock *sk = sock->sk;
1964 struct mptcp_sock *msk;
1965 struct socket *ssock;
1970 ssock = __mptcp_tcp_fallback(msk);
1972 ssock = __mptcp_nmpc_socket(msk);
1974 mask = ssock->ops->poll(file, ssock, wait);
1980 sock_poll_wait(file, sock, wait);
1983 if (test_bit(MPTCP_DATA_READY, &msk->flags))
1984 mask = EPOLLIN | EPOLLRDNORM;
1985 if (sk_stream_is_writeable(sk) &&
1986 test_bit(MPTCP_SEND_SPACE, &msk->flags))
1987 mask |= EPOLLOUT | EPOLLWRNORM;
1988 if (sk->sk_shutdown & RCV_SHUTDOWN)
1989 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1996 static int mptcp_shutdown(struct socket *sock, int how)
1998 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1999 struct mptcp_subflow_context *subflow;
2000 struct socket *ssock;
2003 pr_debug("sk=%p, how=%d", msk, how);
2005 lock_sock(sock->sk);
2006 ssock = __mptcp_tcp_fallback(msk);
2008 release_sock(sock->sk);
2009 return inet_shutdown(ssock, how);
2012 if (how == SHUT_WR || how == SHUT_RDWR)
2013 inet_sk_state_store(sock->sk, TCP_FIN_WAIT1);
2017 if ((how & ~SHUTDOWN_MASK) || !how) {
2022 if (sock->state == SS_CONNECTING) {
2023 if ((1 << sock->sk->sk_state) &
2024 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2025 sock->state = SS_DISCONNECTING;
2027 sock->state = SS_CONNECTED;
2030 __mptcp_flush_join_list(msk);
2031 mptcp_for_each_subflow(msk, subflow) {
2032 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2034 mptcp_subflow_shutdown(tcp_sk, how, 1, msk->write_seq);
2038 release_sock(sock->sk);
2043 static const struct proto_ops mptcp_stream_ops = {
2045 .owner = THIS_MODULE,
2046 .release = inet_release,
2048 .connect = mptcp_stream_connect,
2049 .socketpair = sock_no_socketpair,
2050 .accept = mptcp_stream_accept,
2051 .getname = mptcp_v4_getname,
2053 .ioctl = inet_ioctl,
2054 .gettstamp = sock_gettstamp,
2055 .listen = mptcp_listen,
2056 .shutdown = mptcp_shutdown,
2057 .setsockopt = sock_common_setsockopt,
2058 .getsockopt = sock_common_getsockopt,
2059 .sendmsg = inet_sendmsg,
2060 .recvmsg = inet_recvmsg,
2061 .mmap = sock_no_mmap,
2062 .sendpage = inet_sendpage,
2063 #ifdef CONFIG_COMPAT
2064 .compat_setsockopt = compat_sock_common_setsockopt,
2065 .compat_getsockopt = compat_sock_common_getsockopt,
2069 static struct inet_protosw mptcp_protosw = {
2070 .type = SOCK_STREAM,
2071 .protocol = IPPROTO_MPTCP,
2072 .prot = &mptcp_prot,
2073 .ops = &mptcp_stream_ops,
2074 .flags = INET_PROTOSW_ICSK,
2077 void mptcp_proto_init(void)
2079 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2081 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2082 panic("Failed to allocate MPTCP pcpu counter\n");
2084 mptcp_subflow_init();
2087 if (proto_register(&mptcp_prot, 1) != 0)
2088 panic("Failed to register MPTCP proto.\n");
2090 inet_register_protosw(&mptcp_protosw);
2092 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2095 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2096 static const struct proto_ops mptcp_v6_stream_ops = {
2098 .owner = THIS_MODULE,
2099 .release = inet6_release,
2101 .connect = mptcp_stream_connect,
2102 .socketpair = sock_no_socketpair,
2103 .accept = mptcp_stream_accept,
2104 .getname = mptcp_v6_getname,
2106 .ioctl = inet6_ioctl,
2107 .gettstamp = sock_gettstamp,
2108 .listen = mptcp_listen,
2109 .shutdown = mptcp_shutdown,
2110 .setsockopt = sock_common_setsockopt,
2111 .getsockopt = sock_common_getsockopt,
2112 .sendmsg = inet6_sendmsg,
2113 .recvmsg = inet6_recvmsg,
2114 .mmap = sock_no_mmap,
2115 .sendpage = inet_sendpage,
2116 #ifdef CONFIG_COMPAT
2117 .compat_ioctl = inet6_compat_ioctl,
2118 .compat_setsockopt = compat_sock_common_setsockopt,
2119 .compat_getsockopt = compat_sock_common_getsockopt,
2123 static struct proto mptcp_v6_prot;
2125 static void mptcp_v6_destroy(struct sock *sk)
2128 inet6_destroy_sock(sk);
2131 static struct inet_protosw mptcp_v6_protosw = {
2132 .type = SOCK_STREAM,
2133 .protocol = IPPROTO_MPTCP,
2134 .prot = &mptcp_v6_prot,
2135 .ops = &mptcp_v6_stream_ops,
2136 .flags = INET_PROTOSW_ICSK,
2139 int mptcp_proto_v6_init(void)
2143 mptcp_v6_prot = mptcp_prot;
2144 strcpy(mptcp_v6_prot.name, "MPTCPv6");
2145 mptcp_v6_prot.slab = NULL;
2146 mptcp_v6_prot.destroy = mptcp_v6_destroy;
2147 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
2149 err = proto_register(&mptcp_v6_prot, 1);
2153 err = inet6_register_protosw(&mptcp_v6_protosw);
2155 proto_unregister(&mptcp_v6_prot);