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 struct socket *mptcp_is_tcpsk(struct sock *sk)
57 struct socket *sock = sk->sk_socket;
62 if (unlikely(sk->sk_prot == &tcp_prot)) {
63 /* we are being invoked after mptcp_accept() has
64 * accepted a non-mp-capable flow: sk is a tcp_sk,
67 * Hand the socket over to tcp so all further socket ops
70 sock->ops = &inet_stream_ops;
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 sock->ops = &inet6_stream_ops;
82 static struct socket *__mptcp_tcp_fallback(struct mptcp_sock *msk)
86 sock_owned_by_me((const struct sock *)msk);
88 sock = mptcp_is_tcpsk((struct sock *)msk);
92 if (likely(!__mptcp_check_fallback(msk)))
98 static bool __mptcp_can_create_subflow(const struct mptcp_sock *msk)
103 static struct socket *__mptcp_socket_create(struct mptcp_sock *msk, int state)
105 struct mptcp_subflow_context *subflow;
106 struct sock *sk = (struct sock *)msk;
107 struct socket *ssock;
110 ssock = __mptcp_tcp_fallback(msk);
114 ssock = __mptcp_nmpc_socket(msk);
118 if (!__mptcp_can_create_subflow(msk))
119 return ERR_PTR(-EINVAL);
121 err = mptcp_subflow_create_socket(sk, &ssock);
125 msk->first = ssock->sk;
126 msk->subflow = ssock;
127 subflow = mptcp_subflow_ctx(ssock->sk);
128 list_add(&subflow->node, &msk->conn_list);
129 subflow->request_mptcp = 1;
131 /* accept() will wait on first subflow sk_wq, and we always wakes up
134 RCU_INIT_POINTER(msk->first->sk_wq, &sk->sk_socket->wq);
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 if (__mptcp_check_fallback(msk)) {
233 /* if we are running under the workqueue, TCP could have
234 * collapsed skbs between dummy map creation and now
235 * be sure to adjust the size
237 map_remaining = skb->len;
238 subflow->map_data_len = skb->len;
241 offset = seq - TCP_SKB_CB(skb)->seq;
242 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
248 if (offset < skb->len) {
249 size_t len = skb->len - offset;
254 __mptcp_move_skb(msk, ssk, skb, offset, len);
258 if (WARN_ON_ONCE(map_remaining < len))
262 sk_eat_skb(ssk, skb);
266 WRITE_ONCE(tp->copied_seq, seq);
267 more_data_avail = mptcp_subflow_data_available(ssk);
269 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf)) {
273 } while (more_data_avail);
280 /* In most cases we will be able to lock the mptcp socket. If its already
281 * owned, we need to defer to the work queue to avoid ABBA deadlock.
283 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
285 struct sock *sk = (struct sock *)msk;
286 unsigned int moved = 0;
288 if (READ_ONCE(sk->sk_lock.owned))
291 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
294 /* must re-check after taking the lock */
295 if (!READ_ONCE(sk->sk_lock.owned))
296 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
298 spin_unlock_bh(&sk->sk_lock.slock);
303 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
305 struct mptcp_sock *msk = mptcp_sk(sk);
307 set_bit(MPTCP_DATA_READY, &msk->flags);
309 if (atomic_read(&sk->sk_rmem_alloc) < READ_ONCE(sk->sk_rcvbuf) &&
310 move_skbs_to_msk(msk, ssk))
313 /* don't schedule if mptcp sk is (still) over limit */
314 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf))
317 /* mptcp socket is owned, release_cb should retry */
318 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
319 &sk->sk_tsq_flags)) {
322 /* need to try again, its possible release_cb() has already
323 * been called after the test_and_set_bit() above.
325 move_skbs_to_msk(msk, ssk);
328 sk->sk_data_ready(sk);
331 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
333 if (likely(list_empty(&msk->join_list)))
336 spin_lock_bh(&msk->join_list_lock);
337 list_splice_tail_init(&msk->join_list, &msk->conn_list);
338 spin_unlock_bh(&msk->join_list_lock);
341 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
343 long tout = ssk && inet_csk(ssk)->icsk_pending ?
344 inet_csk(ssk)->icsk_timeout - jiffies : 0;
347 tout = mptcp_sk(sk)->timer_ival;
348 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
351 static bool mptcp_timer_pending(struct sock *sk)
353 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
356 static void mptcp_reset_timer(struct sock *sk)
358 struct inet_connection_sock *icsk = inet_csk(sk);
361 /* should never be called with mptcp level timer cleared */
362 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
363 if (WARN_ON_ONCE(!tout))
365 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
368 void mptcp_data_acked(struct sock *sk)
370 mptcp_reset_timer(sk);
372 if (!sk_stream_is_writeable(sk) &&
373 schedule_work(&mptcp_sk(sk)->work))
377 void mptcp_subflow_eof(struct sock *sk)
379 struct mptcp_sock *msk = mptcp_sk(sk);
381 if (!test_and_set_bit(MPTCP_WORK_EOF, &msk->flags) &&
382 schedule_work(&msk->work))
386 static void mptcp_check_for_eof(struct mptcp_sock *msk)
388 struct mptcp_subflow_context *subflow;
389 struct sock *sk = (struct sock *)msk;
392 mptcp_for_each_subflow(msk, subflow)
393 receivers += !subflow->rx_eof;
395 if (!receivers && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
396 /* hopefully temporary hack: propagate shutdown status
397 * to msk, when all subflows agree on it
399 sk->sk_shutdown |= RCV_SHUTDOWN;
401 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
402 set_bit(MPTCP_DATA_READY, &msk->flags);
403 sk->sk_data_ready(sk);
407 static void mptcp_stop_timer(struct sock *sk)
409 struct inet_connection_sock *icsk = inet_csk(sk);
411 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
412 mptcp_sk(sk)->timer_ival = 0;
415 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
417 const struct sock *sk = (const struct sock *)msk;
419 if (!msk->cached_ext)
420 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
422 return !!msk->cached_ext;
425 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
427 struct mptcp_subflow_context *subflow;
428 struct sock *sk = (struct sock *)msk;
430 sock_owned_by_me(sk);
432 mptcp_for_each_subflow(msk, subflow) {
433 if (subflow->data_avail)
434 return mptcp_subflow_tcp_sock(subflow);
440 static bool mptcp_skb_can_collapse_to(u64 write_seq,
441 const struct sk_buff *skb,
442 const struct mptcp_ext *mpext)
444 if (!tcp_skb_can_collapse_to(skb))
447 /* can collapse only if MPTCP level sequence is in order */
448 return mpext && mpext->data_seq + mpext->data_len == write_seq;
451 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
452 const struct page_frag *pfrag,
453 const struct mptcp_data_frag *df)
455 return df && pfrag->page == df->page &&
456 df->data_seq + df->data_len == msk->write_seq;
459 static void dfrag_uncharge(struct sock *sk, int len)
461 sk_mem_uncharge(sk, len);
462 sk_wmem_queued_add(sk, -len);
465 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
467 int len = dfrag->data_len + dfrag->overhead;
469 list_del(&dfrag->list);
470 dfrag_uncharge(sk, len);
471 put_page(dfrag->page);
474 static void mptcp_clean_una(struct sock *sk)
476 struct mptcp_sock *msk = mptcp_sk(sk);
477 struct mptcp_data_frag *dtmp, *dfrag;
478 bool cleaned = false;
481 /* on fallback we just need to ignore snd_una, as this is really
484 if (__mptcp_check_fallback(msk))
485 atomic64_set(&msk->snd_una, msk->write_seq);
486 snd_una = atomic64_read(&msk->snd_una);
488 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
489 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
492 dfrag_clear(sk, dfrag);
496 dfrag = mptcp_rtx_head(sk);
497 if (dfrag && after64(snd_una, dfrag->data_seq)) {
498 u64 delta = dfrag->data_seq + dfrag->data_len - snd_una;
500 dfrag->data_seq += delta;
501 dfrag->data_len -= delta;
503 dfrag_uncharge(sk, delta);
508 sk_mem_reclaim_partial(sk);
510 /* Only wake up writers if a subflow is ready */
511 if (test_bit(MPTCP_SEND_SPACE, &msk->flags))
512 sk_stream_write_space(sk);
516 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
519 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
521 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
522 pfrag, sk->sk_allocation)))
525 sk->sk_prot->enter_memory_pressure(sk);
526 sk_stream_moderate_sndbuf(sk);
530 static struct mptcp_data_frag *
531 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
534 int offset = ALIGN(orig_offset, sizeof(long));
535 struct mptcp_data_frag *dfrag;
537 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
539 dfrag->data_seq = msk->write_seq;
540 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
541 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
542 dfrag->page = pfrag->page;
547 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
548 struct msghdr *msg, struct mptcp_data_frag *dfrag,
549 long *timeo, int *pmss_now,
552 int mss_now, avail_size, size_goal, offset, ret, frag_truesize = 0;
553 bool dfrag_collapsed, can_collapse = false;
554 struct mptcp_sock *msk = mptcp_sk(sk);
555 struct mptcp_ext *mpext = NULL;
556 bool retransmission = !!dfrag;
557 struct sk_buff *skb, *tail;
558 struct page_frag *pfrag;
563 /* use the mptcp page cache so that we can easily move the data
564 * from one substream to another, but do per subflow memory accounting
565 * Note: pfrag is used only !retransmission, but the compiler if
566 * fooled into a warning if we don't init here
568 pfrag = sk_page_frag(sk);
569 if (!retransmission) {
570 write_seq = &msk->write_seq;
573 write_seq = &dfrag->data_seq;
577 /* compute copy limit */
578 mss_now = tcp_send_mss(ssk, &size_goal, msg->msg_flags);
580 *ps_goal = size_goal;
581 avail_size = size_goal;
582 skb = tcp_write_queue_tail(ssk);
584 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
586 /* Limit the write to the size available in the
587 * current skb, if any, so that we create at most a new skb.
588 * Explicitly tells TCP internals to avoid collapsing on later
589 * queue management operation, to avoid breaking the ext <->
590 * SSN association set here
592 can_collapse = (size_goal - skb->len > 0) &&
593 mptcp_skb_can_collapse_to(*write_seq, skb, mpext);
595 TCP_SKB_CB(skb)->eor = 1;
597 avail_size = size_goal - skb->len;
600 if (!retransmission) {
601 /* reuse tail pfrag, if possible, or carve a new one from the
604 dfrag = mptcp_rtx_tail(sk);
605 offset = pfrag->offset;
606 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
607 if (!dfrag_collapsed) {
608 dfrag = mptcp_carve_data_frag(msk, pfrag, offset);
609 offset = dfrag->offset;
610 frag_truesize = dfrag->overhead;
612 psize = min_t(size_t, pfrag->size - offset, avail_size);
615 pr_debug("left=%zu", msg_data_left(msg));
616 psize = copy_page_from_iter(pfrag->page, offset,
617 min_t(size_t, msg_data_left(msg),
620 pr_debug("left=%zu", msg_data_left(msg));
624 if (!sk_wmem_schedule(sk, psize + dfrag->overhead))
627 offset = dfrag->offset;
628 psize = min_t(size_t, dfrag->data_len, avail_size);
631 /* tell the TCP stack to delay the push so that we can safely
632 * access the skb after the sendpages call
634 ret = do_tcp_sendpages(ssk, page, offset, psize,
635 msg->msg_flags | MSG_SENDPAGE_NOTLAST | MSG_DONTWAIT);
639 frag_truesize += ret;
640 if (!retransmission) {
641 if (unlikely(ret < psize))
642 iov_iter_revert(&msg->msg_iter, psize - ret);
644 /* send successful, keep track of sent data for mptcp-level
647 dfrag->data_len += ret;
648 if (!dfrag_collapsed) {
649 get_page(dfrag->page);
650 list_add_tail(&dfrag->list, &msk->rtx_queue);
651 sk_wmem_queued_add(sk, frag_truesize);
653 sk_wmem_queued_add(sk, ret);
656 /* charge data on mptcp rtx queue to the master socket
657 * Note: we charge such data both to sk and ssk
659 sk->sk_forward_alloc -= frag_truesize;
662 /* if the tail skb extension is still the cached one, collapsing
663 * really happened. Note: we can't check for 'same skb' as the sk_buff
664 * hdr on tail can be transmitted, freed and re-allocated by the
665 * do_tcp_sendpages() call
667 tail = tcp_write_queue_tail(ssk);
668 if (mpext && tail && mpext == skb_ext_find(tail, SKB_EXT_MPTCP)) {
669 WARN_ON_ONCE(!can_collapse);
670 mpext->data_len += ret;
674 skb = tcp_write_queue_tail(ssk);
675 mpext = __skb_ext_set(skb, SKB_EXT_MPTCP, msk->cached_ext);
676 msk->cached_ext = NULL;
678 memset(mpext, 0, sizeof(*mpext));
679 mpext->data_seq = *write_seq;
680 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
681 mpext->data_len = ret;
685 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
686 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
691 pfrag->offset += frag_truesize;
693 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
698 static void mptcp_nospace(struct mptcp_sock *msk, struct socket *sock)
700 clear_bit(MPTCP_SEND_SPACE, &msk->flags);
701 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
703 /* enables sk->write_space() callbacks */
704 set_bit(SOCK_NOSPACE, &sock->flags);
707 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
709 struct mptcp_subflow_context *subflow;
710 struct sock *backup = NULL;
712 sock_owned_by_me((const struct sock *)msk);
714 if (!mptcp_ext_cache_refill(msk))
717 mptcp_for_each_subflow(msk, subflow) {
718 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
720 if (!sk_stream_memory_free(ssk)) {
721 struct socket *sock = ssk->sk_socket;
724 mptcp_nospace(msk, sock);
729 if (subflow->backup) {
742 static void ssk_check_wmem(struct mptcp_sock *msk, struct sock *ssk)
746 if (likely(sk_stream_is_writeable(ssk)))
749 sock = READ_ONCE(ssk->sk_socket);
751 mptcp_nospace(msk, sock);
754 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
756 int mss_now = 0, size_goal = 0, ret = 0;
757 struct mptcp_sock *msk = mptcp_sk(sk);
758 struct page_frag *pfrag;
764 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
769 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
771 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
772 ret = sk_stream_wait_connect(sk, &timeo);
777 pfrag = sk_page_frag(sk);
782 __mptcp_flush_join_list(msk);
783 ssk = mptcp_subflow_get_send(msk);
784 while (!sk_stream_memory_free(sk) ||
786 !mptcp_page_frag_refill(ssk, pfrag)) {
788 /* make sure retransmit timer is
789 * running before we wait for memory.
791 * The retransmit timer might be needed
792 * to make the peer send an up-to-date
795 mptcp_set_timeout(sk, ssk);
796 if (!mptcp_timer_pending(sk))
797 mptcp_reset_timer(sk);
800 ret = sk_stream_wait_memory(sk, &timeo);
806 ssk = mptcp_subflow_get_send(msk);
807 if (list_empty(&msk->conn_list)) {
813 pr_debug("conn_list->subflow=%p", ssk);
816 tx_ok = msg_data_left(msg);
818 ret = mptcp_sendmsg_frag(sk, ssk, msg, NULL, &timeo, &mss_now,
821 if (ret == -EAGAIN && timeo > 0) {
822 mptcp_set_timeout(sk, ssk);
831 tx_ok = msg_data_left(msg);
835 if (!sk_stream_memory_free(ssk) ||
836 !mptcp_page_frag_refill(ssk, pfrag) ||
837 !mptcp_ext_cache_refill(msk)) {
838 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
839 tcp_push(ssk, msg->msg_flags, mss_now,
840 tcp_sk(ssk)->nonagle, size_goal);
841 mptcp_set_timeout(sk, ssk);
846 /* memory is charged to mptcp level socket as well, i.e.
847 * if msg is very large, mptcp socket may run out of buffer
848 * space. mptcp_clean_una() will release data that has
849 * been acked at mptcp level in the mean time, so there is
850 * a good chance we can continue sending data right away.
852 * Normally, when the tcp subflow can accept more data, then
853 * so can the MPTCP socket. However, we need to cope with
854 * peers that might lag behind in their MPTCP-level
855 * acknowledgements, i.e. data might have been acked at
856 * tcp level only. So, we must also check the MPTCP socket
857 * limits before we send more data.
859 if (unlikely(!sk_stream_memory_free(sk))) {
860 tcp_push(ssk, msg->msg_flags, mss_now,
861 tcp_sk(ssk)->nonagle, size_goal);
863 if (!sk_stream_memory_free(sk)) {
864 /* can't send more for now, need to wait for
865 * MPTCP-level ACKs from peer.
867 * Wakeup will happen via mptcp_clean_una().
869 mptcp_set_timeout(sk, ssk);
871 goto wait_for_sndbuf;
876 mptcp_set_timeout(sk, ssk);
879 tcp_push(ssk, msg->msg_flags, mss_now, tcp_sk(ssk)->nonagle,
882 /* start the timer, if it's not pending */
883 if (!mptcp_timer_pending(sk))
884 mptcp_reset_timer(sk);
887 ssk_check_wmem(msk, ssk);
894 static void mptcp_wait_data(struct sock *sk, long *timeo)
896 DEFINE_WAIT_FUNC(wait, woken_wake_function);
897 struct mptcp_sock *msk = mptcp_sk(sk);
899 add_wait_queue(sk_sleep(sk), &wait);
900 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
902 sk_wait_event(sk, timeo,
903 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
905 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
906 remove_wait_queue(sk_sleep(sk), &wait);
909 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
913 struct sock *sk = (struct sock *)msk;
917 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
918 u32 offset = MPTCP_SKB_CB(skb)->offset;
919 u32 data_len = skb->len - offset;
920 u32 count = min_t(size_t, len - copied, data_len);
923 err = skb_copy_datagram_msg(skb, offset, msg, count);
924 if (unlikely(err < 0)) {
932 if (count < data_len) {
933 MPTCP_SKB_CB(skb)->offset += count;
937 __skb_unlink(skb, &sk->sk_receive_queue);
947 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
949 unsigned int moved = 0;
953 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
959 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
966 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
967 int nonblock, int flags, int *addr_len)
969 struct mptcp_sock *msk = mptcp_sk(sk);
974 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
978 timeo = sock_rcvtimeo(sk, nonblock);
980 len = min_t(size_t, len, INT_MAX);
981 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
982 __mptcp_flush_join_list(msk);
984 while (len > (size_t)copied) {
987 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
988 if (unlikely(bytes_read < 0)) {
994 copied += bytes_read;
996 if (skb_queue_empty(&sk->sk_receive_queue) &&
997 __mptcp_move_skbs(msk))
1000 /* only the master socket status is relevant here. The exit
1001 * conditions mirror closely tcp_recvmsg()
1003 if (copied >= target)
1008 sk->sk_state == TCP_CLOSE ||
1009 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1011 signal_pending(current))
1015 copied = sock_error(sk);
1019 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1020 mptcp_check_for_eof(msk);
1022 if (sk->sk_shutdown & RCV_SHUTDOWN)
1025 if (sk->sk_state == TCP_CLOSE) {
1035 if (signal_pending(current)) {
1036 copied = sock_intr_errno(timeo);
1041 pr_debug("block timeout %ld", timeo);
1042 mptcp_wait_data(sk, &timeo);
1045 if (skb_queue_empty(&sk->sk_receive_queue)) {
1046 /* entire backlog drained, clear DATA_READY. */
1047 clear_bit(MPTCP_DATA_READY, &msk->flags);
1049 /* .. race-breaker: ssk might have gotten new data
1050 * after last __mptcp_move_skbs() returned false.
1052 if (unlikely(__mptcp_move_skbs(msk)))
1053 set_bit(MPTCP_DATA_READY, &msk->flags);
1054 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1055 /* data to read but mptcp_wait_data() cleared DATA_READY */
1056 set_bit(MPTCP_DATA_READY, &msk->flags);
1063 static void mptcp_retransmit_handler(struct sock *sk)
1065 struct mptcp_sock *msk = mptcp_sk(sk);
1067 if (atomic64_read(&msk->snd_una) == msk->write_seq) {
1068 mptcp_stop_timer(sk);
1070 set_bit(MPTCP_WORK_RTX, &msk->flags);
1071 if (schedule_work(&msk->work))
1076 static void mptcp_retransmit_timer(struct timer_list *t)
1078 struct inet_connection_sock *icsk = from_timer(icsk, t,
1079 icsk_retransmit_timer);
1080 struct sock *sk = &icsk->icsk_inet.sk;
1083 if (!sock_owned_by_user(sk)) {
1084 mptcp_retransmit_handler(sk);
1086 /* delegate our work to tcp_release_cb() */
1087 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1095 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1098 * A backup subflow is returned only if that is the only kind available.
1100 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1102 struct mptcp_subflow_context *subflow;
1103 struct sock *backup = NULL;
1105 sock_owned_by_me((const struct sock *)msk);
1107 mptcp_for_each_subflow(msk, subflow) {
1108 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1110 /* still data outstanding at TCP level? Don't retransmit. */
1111 if (!tcp_write_queue_empty(ssk))
1114 if (subflow->backup) {
1126 /* subflow sockets can be either outgoing (connect) or incoming
1129 * Outgoing subflows use in-kernel sockets.
1130 * Incoming subflows do not have their own 'struct socket' allocated,
1131 * so we need to use tcp_close() after detaching them from the mptcp
1134 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1135 struct mptcp_subflow_context *subflow,
1138 struct socket *sock = READ_ONCE(ssk->sk_socket);
1140 list_del(&subflow->node);
1142 if (sock && sock != sk->sk_socket) {
1143 /* outgoing subflow */
1146 /* incoming subflow */
1147 tcp_close(ssk, timeout);
1151 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1156 static void mptcp_worker(struct work_struct *work)
1158 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1159 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1160 int orig_len, orig_offset, mss_now = 0, size_goal = 0;
1161 struct mptcp_data_frag *dfrag;
1168 mptcp_clean_una(sk);
1169 __mptcp_flush_join_list(msk);
1170 __mptcp_move_skbs(msk);
1172 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1173 mptcp_check_for_eof(msk);
1175 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1178 dfrag = mptcp_rtx_head(sk);
1182 if (!mptcp_ext_cache_refill(msk))
1185 ssk = mptcp_subflow_get_retrans(msk);
1191 msg.msg_flags = MSG_DONTWAIT;
1192 orig_len = dfrag->data_len;
1193 orig_offset = dfrag->offset;
1194 orig_write_seq = dfrag->data_seq;
1195 while (dfrag->data_len > 0) {
1196 int ret = mptcp_sendmsg_frag(sk, ssk, &msg, dfrag, &timeo,
1197 &mss_now, &size_goal);
1201 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1203 dfrag->data_len -= ret;
1204 dfrag->offset += ret;
1206 if (!mptcp_ext_cache_refill(msk))
1210 tcp_push(ssk, msg.msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1213 dfrag->data_seq = orig_write_seq;
1214 dfrag->offset = orig_offset;
1215 dfrag->data_len = orig_len;
1217 mptcp_set_timeout(sk, ssk);
1221 if (!mptcp_timer_pending(sk))
1222 mptcp_reset_timer(sk);
1229 static int __mptcp_init_sock(struct sock *sk)
1231 struct mptcp_sock *msk = mptcp_sk(sk);
1233 spin_lock_init(&msk->join_list_lock);
1235 INIT_LIST_HEAD(&msk->conn_list);
1236 INIT_LIST_HEAD(&msk->join_list);
1237 INIT_LIST_HEAD(&msk->rtx_queue);
1238 __set_bit(MPTCP_SEND_SPACE, &msk->flags);
1239 INIT_WORK(&msk->work, mptcp_worker);
1242 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1244 mptcp_pm_data_init(msk);
1246 /* re-use the csk retrans timer for MPTCP-level retrans */
1247 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1252 static int mptcp_init_sock(struct sock *sk)
1254 struct net *net = sock_net(sk);
1257 if (!mptcp_is_enabled(net))
1258 return -ENOPROTOOPT;
1260 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1263 ret = __mptcp_init_sock(sk);
1267 sk_sockets_allocated_inc(sk);
1268 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[2];
1273 static void __mptcp_clear_xmit(struct sock *sk)
1275 struct mptcp_sock *msk = mptcp_sk(sk);
1276 struct mptcp_data_frag *dtmp, *dfrag;
1278 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
1280 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
1281 dfrag_clear(sk, dfrag);
1284 static void mptcp_cancel_work(struct sock *sk)
1286 struct mptcp_sock *msk = mptcp_sk(sk);
1288 if (cancel_work_sync(&msk->work))
1292 static void mptcp_subflow_shutdown(struct sock *ssk, int how,
1293 bool data_fin_tx_enable, u64 data_fin_tx_seq)
1297 switch (ssk->sk_state) {
1299 if (!(how & RCV_SHUTDOWN))
1303 tcp_disconnect(ssk, O_NONBLOCK);
1306 if (data_fin_tx_enable) {
1307 struct mptcp_subflow_context *subflow;
1309 subflow = mptcp_subflow_ctx(ssk);
1310 subflow->data_fin_tx_seq = data_fin_tx_seq;
1311 subflow->data_fin_tx_enable = 1;
1314 ssk->sk_shutdown |= how;
1315 tcp_shutdown(ssk, how);
1322 /* Called with msk lock held, releases such lock before returning */
1323 static void mptcp_close(struct sock *sk, long timeout)
1325 struct mptcp_subflow_context *subflow, *tmp;
1326 struct mptcp_sock *msk = mptcp_sk(sk);
1327 LIST_HEAD(conn_list);
1328 u64 data_fin_tx_seq;
1332 inet_sk_state_store(sk, TCP_CLOSE);
1334 /* be sure to always acquire the join list lock, to sync vs
1335 * mptcp_finish_join().
1337 spin_lock_bh(&msk->join_list_lock);
1338 list_splice_tail_init(&msk->join_list, &msk->conn_list);
1339 spin_unlock_bh(&msk->join_list_lock);
1340 list_splice_init(&msk->conn_list, &conn_list);
1342 data_fin_tx_seq = msk->write_seq;
1344 __mptcp_clear_xmit(sk);
1348 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
1349 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1351 subflow->data_fin_tx_seq = data_fin_tx_seq;
1352 subflow->data_fin_tx_enable = 1;
1353 __mptcp_close_ssk(sk, ssk, subflow, timeout);
1356 mptcp_cancel_work(sk);
1357 mptcp_pm_close(msk);
1359 __skb_queue_purge(&sk->sk_receive_queue);
1361 sk_common_release(sk);
1364 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
1366 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1367 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
1368 struct ipv6_pinfo *msk6 = inet6_sk(msk);
1370 msk->sk_v6_daddr = ssk->sk_v6_daddr;
1371 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
1374 msk6->saddr = ssk6->saddr;
1375 msk6->flow_label = ssk6->flow_label;
1379 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
1380 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
1381 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
1382 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
1383 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
1384 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
1387 static int mptcp_disconnect(struct sock *sk, int flags)
1389 /* Should never be called.
1390 * inet_stream_connect() calls ->disconnect, but that
1391 * refers to the subflow socket, not the mptcp one.
1397 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1398 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
1400 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
1402 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
1406 struct sock *mptcp_sk_clone(const struct sock *sk,
1407 const struct mptcp_options_received *mp_opt,
1408 struct request_sock *req)
1410 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
1411 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
1412 struct mptcp_sock *msk;
1418 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1419 if (nsk->sk_family == AF_INET6)
1420 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
1423 __mptcp_init_sock(nsk);
1425 msk = mptcp_sk(nsk);
1426 msk->local_key = subflow_req->local_key;
1427 msk->token = subflow_req->token;
1428 msk->subflow = NULL;
1430 msk->write_seq = subflow_req->idsn + 1;
1431 atomic64_set(&msk->snd_una, msk->write_seq);
1432 if (mp_opt->mp_capable) {
1433 msk->can_ack = true;
1434 msk->remote_key = mp_opt->sndr_key;
1435 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
1437 msk->ack_seq = ack_seq;
1440 sock_reset_flag(nsk, SOCK_RCU_FREE);
1441 /* will be fully established after successful MPC subflow creation */
1442 inet_sk_state_store(nsk, TCP_SYN_RECV);
1443 bh_unlock_sock(nsk);
1445 /* keep a single reference */
1450 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
1453 struct mptcp_sock *msk = mptcp_sk(sk);
1454 struct socket *listener;
1457 listener = __mptcp_nmpc_socket(msk);
1458 if (WARN_ON_ONCE(!listener)) {
1463 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
1464 newsk = inet_csk_accept(listener->sk, flags, err, kern);
1468 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
1470 if (sk_is_mptcp(newsk)) {
1471 struct mptcp_subflow_context *subflow;
1472 struct sock *new_mptcp_sock;
1473 struct sock *ssk = newsk;
1475 subflow = mptcp_subflow_ctx(newsk);
1476 new_mptcp_sock = subflow->conn;
1478 /* is_mptcp should be false if subflow->conn is missing, see
1479 * subflow_syn_recv_sock()
1481 if (WARN_ON_ONCE(!new_mptcp_sock)) {
1482 tcp_sk(newsk)->is_mptcp = 0;
1486 /* acquire the 2nd reference for the owning socket */
1487 sock_hold(new_mptcp_sock);
1490 bh_lock_sock(new_mptcp_sock);
1491 msk = mptcp_sk(new_mptcp_sock);
1494 newsk = new_mptcp_sock;
1495 mptcp_copy_inaddrs(newsk, ssk);
1496 list_add(&subflow->node, &msk->conn_list);
1497 inet_sk_state_store(newsk, TCP_ESTABLISHED);
1499 bh_unlock_sock(new_mptcp_sock);
1501 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
1504 MPTCP_INC_STATS(sock_net(sk),
1505 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
1511 static void mptcp_destroy(struct sock *sk)
1513 struct mptcp_sock *msk = mptcp_sk(sk);
1515 mptcp_token_destroy(msk);
1516 if (msk->cached_ext)
1517 __skb_ext_put(msk->cached_ext);
1519 sk_sockets_allocated_dec(sk);
1522 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
1523 char __user *optval, unsigned int optlen)
1525 struct mptcp_sock *msk = mptcp_sk(sk);
1526 struct socket *ssock;
1528 pr_debug("msk=%p", msk);
1530 /* @@ the meaning of setsockopt() when the socket is connected and
1531 * there are multiple subflows is not yet defined. It is up to the
1532 * MPTCP-level socket to configure the subflows until the subflow
1533 * is in TCP fallback, when TCP socket options are passed through
1534 * to the one remaining subflow.
1537 ssock = __mptcp_tcp_fallback(msk);
1540 return tcp_setsockopt(ssock->sk, level, optname, optval,
1546 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
1547 char __user *optval, int __user *option)
1549 struct mptcp_sock *msk = mptcp_sk(sk);
1550 struct socket *ssock;
1552 pr_debug("msk=%p", msk);
1554 /* @@ the meaning of setsockopt() when the socket is connected and
1555 * there are multiple subflows is not yet defined. It is up to the
1556 * MPTCP-level socket to configure the subflows until the subflow
1557 * is in TCP fallback, when socket options are passed through
1558 * to the one remaining subflow.
1561 ssock = __mptcp_tcp_fallback(msk);
1564 return tcp_getsockopt(ssock->sk, level, optname, optval,
1570 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
1571 TCPF_WRITE_TIMER_DEFERRED)
1573 /* this is very alike tcp_release_cb() but we must handle differently a
1574 * different set of events
1576 static void mptcp_release_cb(struct sock *sk)
1578 unsigned long flags, nflags;
1581 flags = sk->sk_tsq_flags;
1582 if (!(flags & MPTCP_DEFERRED_ALL))
1584 nflags = flags & ~MPTCP_DEFERRED_ALL;
1585 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1587 sock_release_ownership(sk);
1589 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1590 struct mptcp_sock *msk = mptcp_sk(sk);
1593 ssk = mptcp_subflow_recv_lookup(msk);
1594 if (!ssk || !schedule_work(&msk->work))
1598 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1599 mptcp_retransmit_handler(sk);
1604 static int mptcp_hash(struct sock *sk)
1606 /* should never be called,
1607 * we hash the TCP subflows not the master socket
1613 static void mptcp_unhash(struct sock *sk)
1615 /* called from sk_common_release(), but nothing to do here */
1618 static int mptcp_get_port(struct sock *sk, unsigned short snum)
1620 struct mptcp_sock *msk = mptcp_sk(sk);
1621 struct socket *ssock;
1623 ssock = __mptcp_nmpc_socket(msk);
1624 pr_debug("msk=%p, subflow=%p", msk, ssock);
1625 if (WARN_ON_ONCE(!ssock))
1628 return inet_csk_get_port(ssock->sk, snum);
1631 void mptcp_finish_connect(struct sock *ssk)
1633 struct mptcp_subflow_context *subflow;
1634 struct mptcp_sock *msk;
1638 subflow = mptcp_subflow_ctx(ssk);
1642 pr_debug("msk=%p, token=%u", sk, subflow->token);
1644 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
1646 subflow->map_seq = ack_seq;
1647 subflow->map_subflow_seq = 1;
1648 subflow->rel_write_seq = 1;
1650 /* the socket is not connected yet, no msk/subflow ops can access/race
1651 * accessing the field below
1653 WRITE_ONCE(msk->remote_key, subflow->remote_key);
1654 WRITE_ONCE(msk->local_key, subflow->local_key);
1655 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
1656 WRITE_ONCE(msk->ack_seq, ack_seq);
1657 WRITE_ONCE(msk->can_ack, 1);
1658 atomic64_set(&msk->snd_una, msk->write_seq);
1660 mptcp_pm_new_connection(msk, 0);
1663 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
1665 write_lock_bh(&sk->sk_callback_lock);
1666 rcu_assign_pointer(sk->sk_wq, &parent->wq);
1667 sk_set_socket(sk, parent);
1668 sk->sk_uid = SOCK_INODE(parent)->i_uid;
1669 write_unlock_bh(&sk->sk_callback_lock);
1672 bool mptcp_finish_join(struct sock *sk)
1674 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
1675 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
1676 struct sock *parent = (void *)msk;
1677 struct socket *parent_sock;
1680 pr_debug("msk=%p, subflow=%p", msk, subflow);
1682 /* mptcp socket already closing? */
1683 if (inet_sk_state_load(parent) != TCP_ESTABLISHED)
1686 if (!msk->pm.server_side)
1689 if (!mptcp_pm_allow_new_subflow(msk))
1692 /* active connections are already on conn_list, and we can't acquire
1694 * use the join list lock as synchronization point and double-check
1695 * msk status to avoid racing with mptcp_close()
1697 spin_lock_bh(&msk->join_list_lock);
1698 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
1699 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node)))
1700 list_add_tail(&subflow->node, &msk->join_list);
1701 spin_unlock_bh(&msk->join_list_lock);
1705 /* attach to msk socket only after we are sure he will deal with us
1708 parent_sock = READ_ONCE(parent->sk_socket);
1709 if (parent_sock && !sk->sk_socket)
1710 mptcp_sock_graft(sk, parent_sock);
1711 subflow->map_seq = msk->ack_seq;
1715 static bool mptcp_memory_free(const struct sock *sk, int wake)
1717 struct mptcp_sock *msk = mptcp_sk(sk);
1719 return wake ? test_bit(MPTCP_SEND_SPACE, &msk->flags) : true;
1722 static struct proto mptcp_prot = {
1724 .owner = THIS_MODULE,
1725 .init = mptcp_init_sock,
1726 .disconnect = mptcp_disconnect,
1727 .close = mptcp_close,
1728 .accept = mptcp_accept,
1729 .setsockopt = mptcp_setsockopt,
1730 .getsockopt = mptcp_getsockopt,
1731 .shutdown = tcp_shutdown,
1732 .destroy = mptcp_destroy,
1733 .sendmsg = mptcp_sendmsg,
1734 .recvmsg = mptcp_recvmsg,
1735 .release_cb = mptcp_release_cb,
1737 .unhash = mptcp_unhash,
1738 .get_port = mptcp_get_port,
1739 .sockets_allocated = &mptcp_sockets_allocated,
1740 .memory_allocated = &tcp_memory_allocated,
1741 .memory_pressure = &tcp_memory_pressure,
1742 .stream_memory_free = mptcp_memory_free,
1743 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
1744 .sysctl_mem = sysctl_tcp_mem,
1745 .obj_size = sizeof(struct mptcp_sock),
1746 .slab_flags = SLAB_TYPESAFE_BY_RCU,
1747 .no_autobind = true,
1750 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
1752 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1753 struct socket *ssock;
1756 lock_sock(sock->sk);
1757 ssock = __mptcp_socket_create(msk, MPTCP_SAME_STATE);
1758 if (IS_ERR(ssock)) {
1759 err = PTR_ERR(ssock);
1763 err = ssock->ops->bind(ssock, uaddr, addr_len);
1765 mptcp_copy_inaddrs(sock->sk, ssock->sk);
1768 release_sock(sock->sk);
1772 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
1773 int addr_len, int flags)
1775 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1776 struct mptcp_subflow_context *subflow;
1777 struct socket *ssock;
1780 lock_sock(sock->sk);
1781 if (sock->state != SS_UNCONNECTED && msk->subflow) {
1782 /* pending connection or invalid state, let existing subflow
1785 ssock = msk->subflow;
1789 mptcp_token_destroy(msk);
1790 ssock = __mptcp_socket_create(msk, TCP_SYN_SENT);
1791 if (IS_ERR(ssock)) {
1792 err = PTR_ERR(ssock);
1796 subflow = mptcp_subflow_ctx(ssock->sk);
1797 #ifdef CONFIG_TCP_MD5SIG
1798 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
1801 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
1802 subflow->request_mptcp = 0;
1804 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
1805 subflow->request_mptcp = 0;
1808 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
1809 sock->state = ssock->state;
1811 /* on successful connect, the msk state will be moved to established by
1812 * subflow_finish_connect()
1814 if (!err || err == EINPROGRESS)
1815 mptcp_copy_inaddrs(sock->sk, ssock->sk);
1817 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
1820 release_sock(sock->sk);
1824 static int mptcp_v4_getname(struct socket *sock, struct sockaddr *uaddr,
1827 if (sock->sk->sk_prot == &tcp_prot) {
1828 /* we are being invoked from __sys_accept4, after
1829 * mptcp_accept() has just accepted a non-mp-capable
1830 * flow: sk is a tcp_sk, not an mptcp one.
1832 * Hand the socket over to tcp so all further socket ops
1835 sock->ops = &inet_stream_ops;
1838 return inet_getname(sock, uaddr, peer);
1841 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1842 static int mptcp_v6_getname(struct socket *sock, struct sockaddr *uaddr,
1845 if (sock->sk->sk_prot == &tcpv6_prot) {
1846 /* we are being invoked from __sys_accept4 after
1847 * mptcp_accept() has accepted a non-mp-capable
1848 * subflow: sk is a tcp_sk, not mptcp.
1850 * Hand the socket over to tcp so all further
1851 * socket ops bypass mptcp.
1853 sock->ops = &inet6_stream_ops;
1856 return inet6_getname(sock, uaddr, peer);
1860 static int mptcp_listen(struct socket *sock, int backlog)
1862 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1863 struct socket *ssock;
1866 pr_debug("msk=%p", msk);
1868 lock_sock(sock->sk);
1869 mptcp_token_destroy(msk);
1870 ssock = __mptcp_socket_create(msk, TCP_LISTEN);
1871 if (IS_ERR(ssock)) {
1872 err = PTR_ERR(ssock);
1876 sock_set_flag(sock->sk, SOCK_RCU_FREE);
1878 err = ssock->ops->listen(ssock, backlog);
1879 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
1881 mptcp_copy_inaddrs(sock->sk, ssock->sk);
1884 release_sock(sock->sk);
1888 static bool is_tcp_proto(const struct proto *p)
1890 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1891 return p == &tcp_prot || p == &tcpv6_prot;
1893 return p == &tcp_prot;
1897 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
1898 int flags, bool kern)
1900 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1901 struct socket *ssock;
1904 pr_debug("msk=%p", msk);
1906 lock_sock(sock->sk);
1907 if (sock->sk->sk_state != TCP_LISTEN)
1910 ssock = __mptcp_nmpc_socket(msk);
1914 sock_hold(ssock->sk);
1915 release_sock(sock->sk);
1917 err = ssock->ops->accept(sock, newsock, flags, kern);
1918 if (err == 0 && !is_tcp_proto(newsock->sk->sk_prot)) {
1919 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
1920 struct mptcp_subflow_context *subflow;
1922 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
1923 * This is needed so NOSPACE flag can be set from tcp stack.
1925 __mptcp_flush_join_list(msk);
1926 list_for_each_entry(subflow, &msk->conn_list, node) {
1927 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1929 if (!ssk->sk_socket)
1930 mptcp_sock_graft(ssk, newsock);
1934 sock_put(ssock->sk);
1938 release_sock(sock->sk);
1942 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
1943 struct poll_table_struct *wait)
1945 struct sock *sk = sock->sk;
1946 struct mptcp_sock *msk;
1950 sock_poll_wait(file, sock, wait);
1952 if (test_bit(MPTCP_DATA_READY, &msk->flags))
1953 mask = EPOLLIN | EPOLLRDNORM;
1954 if (sk_stream_is_writeable(sk) &&
1955 test_bit(MPTCP_SEND_SPACE, &msk->flags))
1956 mask |= EPOLLOUT | EPOLLWRNORM;
1957 if (sk->sk_shutdown & RCV_SHUTDOWN)
1958 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1963 static int mptcp_shutdown(struct socket *sock, int how)
1965 struct mptcp_sock *msk = mptcp_sk(sock->sk);
1966 struct mptcp_subflow_context *subflow;
1969 pr_debug("sk=%p, how=%d", msk, how);
1971 lock_sock(sock->sk);
1972 if (how == SHUT_WR || how == SHUT_RDWR)
1973 inet_sk_state_store(sock->sk, TCP_FIN_WAIT1);
1977 if ((how & ~SHUTDOWN_MASK) || !how) {
1982 if (sock->state == SS_CONNECTING) {
1983 if ((1 << sock->sk->sk_state) &
1984 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
1985 sock->state = SS_DISCONNECTING;
1987 sock->state = SS_CONNECTED;
1990 __mptcp_flush_join_list(msk);
1991 mptcp_for_each_subflow(msk, subflow) {
1992 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
1994 mptcp_subflow_shutdown(tcp_sk, how, 1, msk->write_seq);
1997 /* Wake up anyone sleeping in poll. */
1998 sock->sk->sk_state_change(sock->sk);
2001 release_sock(sock->sk);
2006 static const struct proto_ops mptcp_stream_ops = {
2008 .owner = THIS_MODULE,
2009 .release = inet_release,
2011 .connect = mptcp_stream_connect,
2012 .socketpair = sock_no_socketpair,
2013 .accept = mptcp_stream_accept,
2014 .getname = mptcp_v4_getname,
2016 .ioctl = inet_ioctl,
2017 .gettstamp = sock_gettstamp,
2018 .listen = mptcp_listen,
2019 .shutdown = mptcp_shutdown,
2020 .setsockopt = sock_common_setsockopt,
2021 .getsockopt = sock_common_getsockopt,
2022 .sendmsg = inet_sendmsg,
2023 .recvmsg = inet_recvmsg,
2024 .mmap = sock_no_mmap,
2025 .sendpage = inet_sendpage,
2026 #ifdef CONFIG_COMPAT
2027 .compat_setsockopt = compat_sock_common_setsockopt,
2028 .compat_getsockopt = compat_sock_common_getsockopt,
2032 static struct inet_protosw mptcp_protosw = {
2033 .type = SOCK_STREAM,
2034 .protocol = IPPROTO_MPTCP,
2035 .prot = &mptcp_prot,
2036 .ops = &mptcp_stream_ops,
2037 .flags = INET_PROTOSW_ICSK,
2040 void __init mptcp_proto_init(void)
2042 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2044 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2045 panic("Failed to allocate MPTCP pcpu counter\n");
2047 mptcp_subflow_init();
2051 if (proto_register(&mptcp_prot, 1) != 0)
2052 panic("Failed to register MPTCP proto.\n");
2054 inet_register_protosw(&mptcp_protosw);
2056 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2059 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2060 static const struct proto_ops mptcp_v6_stream_ops = {
2062 .owner = THIS_MODULE,
2063 .release = inet6_release,
2065 .connect = mptcp_stream_connect,
2066 .socketpair = sock_no_socketpair,
2067 .accept = mptcp_stream_accept,
2068 .getname = mptcp_v6_getname,
2070 .ioctl = inet6_ioctl,
2071 .gettstamp = sock_gettstamp,
2072 .listen = mptcp_listen,
2073 .shutdown = mptcp_shutdown,
2074 .setsockopt = sock_common_setsockopt,
2075 .getsockopt = sock_common_getsockopt,
2076 .sendmsg = inet6_sendmsg,
2077 .recvmsg = inet6_recvmsg,
2078 .mmap = sock_no_mmap,
2079 .sendpage = inet_sendpage,
2080 #ifdef CONFIG_COMPAT
2081 .compat_ioctl = inet6_compat_ioctl,
2082 .compat_setsockopt = compat_sock_common_setsockopt,
2083 .compat_getsockopt = compat_sock_common_getsockopt,
2087 static struct proto mptcp_v6_prot;
2089 static void mptcp_v6_destroy(struct sock *sk)
2092 inet6_destroy_sock(sk);
2095 static struct inet_protosw mptcp_v6_protosw = {
2096 .type = SOCK_STREAM,
2097 .protocol = IPPROTO_MPTCP,
2098 .prot = &mptcp_v6_prot,
2099 .ops = &mptcp_v6_stream_ops,
2100 .flags = INET_PROTOSW_ICSK,
2103 int __init mptcp_proto_v6_init(void)
2107 mptcp_v6_prot = mptcp_prot;
2108 strcpy(mptcp_v6_prot.name, "MPTCPv6");
2109 mptcp_v6_prot.slab = NULL;
2110 mptcp_v6_prot.destroy = mptcp_v6_destroy;
2111 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
2113 err = proto_register(&mptcp_v6_prot, 1);
2117 err = inet6_register_protosw(&mptcp_v6_protosw);
2119 proto_unregister(&mptcp_v6_prot);