1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
27 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
29 struct mptcp_sock msk;
40 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
42 static struct percpu_counter mptcp_sockets_allocated;
44 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
45 * completed yet or has failed, return the subflow socket.
46 * Otherwise return NULL.
48 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
50 if (!msk->subflow || READ_ONCE(msk->can_ack))
56 static bool mptcp_is_tcpsk(struct sock *sk)
58 struct socket *sock = sk->sk_socket;
60 if (unlikely(sk->sk_prot == &tcp_prot)) {
61 /* we are being invoked after mptcp_accept() has
62 * accepted a non-mp-capable flow: sk is a tcp_sk,
65 * Hand the socket over to tcp so all further socket ops
68 sock->ops = &inet_stream_ops;
70 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
71 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
72 sock->ops = &inet6_stream_ops;
80 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
82 sock_owned_by_me((const struct sock *)msk);
84 if (likely(!__mptcp_check_fallback(msk)))
90 static int __mptcp_socket_create(struct mptcp_sock *msk)
92 struct mptcp_subflow_context *subflow;
93 struct sock *sk = (struct sock *)msk;
97 err = mptcp_subflow_create_socket(sk, &ssock);
101 msk->first = ssock->sk;
102 msk->subflow = ssock;
103 subflow = mptcp_subflow_ctx(ssock->sk);
104 list_add(&subflow->node, &msk->conn_list);
105 subflow->request_mptcp = 1;
107 /* accept() will wait on first subflow sk_wq, and we always wakes up
110 RCU_INIT_POINTER(msk->first->sk_wq, &sk->sk_socket->wq);
115 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
117 sk_drops_add(sk, skb);
121 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
122 struct sk_buff *from)
127 if (MPTCP_SKB_CB(from)->offset ||
128 !skb_try_coalesce(to, from, &fragstolen, &delta))
131 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
132 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
133 to->len, MPTCP_SKB_CB(from)->end_seq);
134 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
135 kfree_skb_partial(from, fragstolen);
136 atomic_add(delta, &sk->sk_rmem_alloc);
137 sk_mem_charge(sk, delta);
141 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
142 struct sk_buff *from)
144 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
147 return mptcp_try_coalesce((struct sock *)msk, to, from);
150 /* "inspired" by tcp_data_queue_ofo(), main differences:
152 * - don't cope with sacks
154 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
156 struct sock *sk = (struct sock *)msk;
157 struct rb_node **p, *parent;
158 u64 seq, end_seq, max_seq;
159 struct sk_buff *skb1;
162 seq = MPTCP_SKB_CB(skb)->map_seq;
163 end_seq = MPTCP_SKB_CB(skb)->end_seq;
164 space = tcp_space(sk);
165 max_seq = space > 0 ? space + msk->ack_seq : msk->ack_seq;
167 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
168 RB_EMPTY_ROOT(&msk->out_of_order_queue));
169 if (after64(seq, max_seq)) {
172 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
176 p = &msk->out_of_order_queue.rb_node;
177 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
178 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
179 rb_link_node(&skb->rbnode, NULL, p);
180 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
181 msk->ooo_last_skb = skb;
185 /* with 2 subflows, adding at end of ooo queue is quite likely
186 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
188 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
189 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
190 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
194 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
195 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
196 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
197 parent = &msk->ooo_last_skb->rbnode;
198 p = &parent->rb_right;
202 /* Find place to insert this segment. Handle overlaps on the way. */
206 skb1 = rb_to_skb(parent);
207 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
208 p = &parent->rb_left;
211 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
212 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
213 /* All the bits are present. Drop. */
215 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
218 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
222 * continue traversing
225 /* skb's seq == skb1's seq and skb covers skb1.
226 * Replace skb1 with skb.
228 rb_replace_node(&skb1->rbnode, &skb->rbnode,
229 &msk->out_of_order_queue);
230 mptcp_drop(sk, skb1);
231 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
234 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
235 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
238 p = &parent->rb_right;
242 /* Insert segment into RB tree. */
243 rb_link_node(&skb->rbnode, parent, p);
244 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
247 /* Remove other segments covered by skb. */
248 while ((skb1 = skb_rb_next(skb)) != NULL) {
249 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
251 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
252 mptcp_drop(sk, skb1);
253 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
255 /* If there is no skb after us, we are the last_skb ! */
257 msk->ooo_last_skb = skb;
261 skb_set_owner_r(skb, sk);
264 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
265 struct sk_buff *skb, unsigned int offset,
268 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
269 struct sock *sk = (struct sock *)msk;
270 struct sk_buff *tail;
272 __skb_unlink(skb, &ssk->sk_receive_queue);
277 /* the skb map_seq accounts for the skb offset:
278 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
281 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
282 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
283 MPTCP_SKB_CB(skb)->offset = offset;
285 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
287 msk->ack_seq += copy_len;
288 tail = skb_peek_tail(&sk->sk_receive_queue);
289 if (tail && mptcp_try_coalesce(sk, tail, skb))
292 skb_set_owner_r(skb, sk);
293 __skb_queue_tail(&sk->sk_receive_queue, skb);
295 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
296 mptcp_data_queue_ofo(msk, skb);
300 /* old data, keep it simple and drop the whole pkt, sender
301 * will retransmit as needed, if needed.
303 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
308 static void mptcp_stop_timer(struct sock *sk)
310 struct inet_connection_sock *icsk = inet_csk(sk);
312 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
313 mptcp_sk(sk)->timer_ival = 0;
316 static void mptcp_check_data_fin_ack(struct sock *sk)
318 struct mptcp_sock *msk = mptcp_sk(sk);
320 if (__mptcp_check_fallback(msk))
323 /* Look for an acknowledged DATA_FIN */
324 if (((1 << sk->sk_state) &
325 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
326 msk->write_seq == atomic64_read(&msk->snd_una)) {
327 mptcp_stop_timer(sk);
329 WRITE_ONCE(msk->snd_data_fin_enable, 0);
331 switch (sk->sk_state) {
333 inet_sk_state_store(sk, TCP_FIN_WAIT2);
334 sk->sk_state_change(sk);
338 inet_sk_state_store(sk, TCP_CLOSE);
339 sk->sk_state_change(sk);
343 if (sk->sk_shutdown == SHUTDOWN_MASK ||
344 sk->sk_state == TCP_CLOSE)
345 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
347 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
351 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
353 struct mptcp_sock *msk = mptcp_sk(sk);
355 if (READ_ONCE(msk->rcv_data_fin) &&
356 ((1 << sk->sk_state) &
357 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
358 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
360 if (msk->ack_seq == rcv_data_fin_seq) {
362 *seq = rcv_data_fin_seq;
371 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
373 long tout = ssk && inet_csk(ssk)->icsk_pending ?
374 inet_csk(ssk)->icsk_timeout - jiffies : 0;
377 tout = mptcp_sk(sk)->timer_ival;
378 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
381 static void mptcp_check_data_fin(struct sock *sk)
383 struct mptcp_sock *msk = mptcp_sk(sk);
384 u64 rcv_data_fin_seq;
386 if (__mptcp_check_fallback(msk) || !msk->first)
389 /* Need to ack a DATA_FIN received from a peer while this side
390 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
391 * msk->rcv_data_fin was set when parsing the incoming options
392 * at the subflow level and the msk lock was not held, so this
393 * is the first opportunity to act on the DATA_FIN and change
396 * If we are caught up to the sequence number of the incoming
397 * DATA_FIN, send the DATA_ACK now and do state transition. If
398 * not caught up, do nothing and let the recv code send DATA_ACK
402 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
403 struct mptcp_subflow_context *subflow;
406 WRITE_ONCE(msk->rcv_data_fin, 0);
408 sk->sk_shutdown |= RCV_SHUTDOWN;
409 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
410 set_bit(MPTCP_DATA_READY, &msk->flags);
412 switch (sk->sk_state) {
413 case TCP_ESTABLISHED:
414 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
417 inet_sk_state_store(sk, TCP_CLOSING);
420 inet_sk_state_store(sk, TCP_CLOSE);
421 // @@ Close subflows now?
424 /* Other states not expected */
429 mptcp_set_timeout(sk, NULL);
430 mptcp_for_each_subflow(msk, subflow) {
431 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
438 sk->sk_state_change(sk);
440 if (sk->sk_shutdown == SHUTDOWN_MASK ||
441 sk->sk_state == TCP_CLOSE)
442 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
444 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
448 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
452 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
453 struct sock *sk = (struct sock *)msk;
454 unsigned int moved = 0;
455 bool more_data_avail;
459 pr_debug("msk=%p ssk=%p", msk, ssk);
462 u32 map_remaining, offset;
463 u32 seq = tp->copied_seq;
467 /* try to move as much data as available */
468 map_remaining = subflow->map_data_len -
469 mptcp_subflow_get_map_offset(subflow);
471 skb = skb_peek(&ssk->sk_receive_queue);
475 if (__mptcp_check_fallback(msk)) {
476 /* if we are running under the workqueue, TCP could have
477 * collapsed skbs between dummy map creation and now
478 * be sure to adjust the size
480 map_remaining = skb->len;
481 subflow->map_data_len = skb->len;
484 offset = seq - TCP_SKB_CB(skb)->seq;
485 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
491 if (offset < skb->len) {
492 size_t len = skb->len - offset;
497 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
501 if (WARN_ON_ONCE(map_remaining < len))
505 sk_eat_skb(ssk, skb);
509 WRITE_ONCE(tp->copied_seq, seq);
510 more_data_avail = mptcp_subflow_data_available(ssk);
512 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf)) {
516 } while (more_data_avail);
520 tcp_cleanup_rbuf(ssk, moved);
525 static bool mptcp_ofo_queue(struct mptcp_sock *msk)
527 struct sock *sk = (struct sock *)msk;
528 struct sk_buff *skb, *tail;
533 p = rb_first(&msk->out_of_order_queue);
534 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
537 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
541 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
543 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
546 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
550 end_seq = MPTCP_SKB_CB(skb)->end_seq;
551 tail = skb_peek_tail(&sk->sk_receive_queue);
552 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
553 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
555 /* skip overlapping data, if any */
556 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
557 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
559 MPTCP_SKB_CB(skb)->offset += delta;
560 __skb_queue_tail(&sk->sk_receive_queue, skb);
562 msk->ack_seq = end_seq;
568 /* In most cases we will be able to lock the mptcp socket. If its already
569 * owned, we need to defer to the work queue to avoid ABBA deadlock.
571 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
573 struct sock *sk = (struct sock *)msk;
574 unsigned int moved = 0;
576 if (READ_ONCE(sk->sk_lock.owned))
579 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
582 /* must re-check after taking the lock */
583 if (!READ_ONCE(sk->sk_lock.owned)) {
584 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
585 mptcp_ofo_queue(msk);
587 /* If the moves have caught up with the DATA_FIN sequence number
588 * it's time to ack the DATA_FIN and change socket state, but
589 * this is not a good place to change state. Let the workqueue
592 if (mptcp_pending_data_fin(sk, NULL) &&
593 schedule_work(&msk->work))
597 spin_unlock_bh(&sk->sk_lock.slock);
602 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
604 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
605 struct mptcp_sock *msk = mptcp_sk(sk);
608 /* move_skbs_to_msk below can legitly clear the data_avail flag,
609 * but we will need later to properly woke the reader, cache its
612 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
614 set_bit(MPTCP_DATA_READY, &msk->flags);
616 if (atomic_read(&sk->sk_rmem_alloc) < READ_ONCE(sk->sk_rcvbuf) &&
617 move_skbs_to_msk(msk, ssk))
620 /* don't schedule if mptcp sk is (still) over limit */
621 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf))
624 /* mptcp socket is owned, release_cb should retry */
625 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
626 &sk->sk_tsq_flags)) {
629 /* need to try again, its possible release_cb() has already
630 * been called after the test_and_set_bit() above.
632 move_skbs_to_msk(msk, ssk);
636 sk->sk_data_ready(sk);
639 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
641 if (likely(list_empty(&msk->join_list)))
644 spin_lock_bh(&msk->join_list_lock);
645 list_splice_tail_init(&msk->join_list, &msk->conn_list);
646 spin_unlock_bh(&msk->join_list_lock);
649 static bool mptcp_timer_pending(struct sock *sk)
651 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
654 static void mptcp_reset_timer(struct sock *sk)
656 struct inet_connection_sock *icsk = inet_csk(sk);
659 /* should never be called with mptcp level timer cleared */
660 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
661 if (WARN_ON_ONCE(!tout))
663 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
666 void mptcp_data_acked(struct sock *sk)
668 mptcp_reset_timer(sk);
670 if ((!test_bit(MPTCP_SEND_SPACE, &mptcp_sk(sk)->flags) ||
671 (inet_sk_state_load(sk) != TCP_ESTABLISHED)) &&
672 schedule_work(&mptcp_sk(sk)->work))
676 void mptcp_subflow_eof(struct sock *sk)
678 struct mptcp_sock *msk = mptcp_sk(sk);
680 if (!test_and_set_bit(MPTCP_WORK_EOF, &msk->flags) &&
681 schedule_work(&msk->work))
685 static void mptcp_check_for_eof(struct mptcp_sock *msk)
687 struct mptcp_subflow_context *subflow;
688 struct sock *sk = (struct sock *)msk;
691 mptcp_for_each_subflow(msk, subflow)
692 receivers += !subflow->rx_eof;
694 if (!receivers && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
695 /* hopefully temporary hack: propagate shutdown status
696 * to msk, when all subflows agree on it
698 sk->sk_shutdown |= RCV_SHUTDOWN;
700 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
701 set_bit(MPTCP_DATA_READY, &msk->flags);
702 sk->sk_data_ready(sk);
706 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
708 const struct sock *sk = (const struct sock *)msk;
710 if (!msk->cached_ext)
711 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
713 return !!msk->cached_ext;
716 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
718 struct mptcp_subflow_context *subflow;
719 struct sock *sk = (struct sock *)msk;
721 sock_owned_by_me(sk);
723 mptcp_for_each_subflow(msk, subflow) {
724 if (subflow->data_avail)
725 return mptcp_subflow_tcp_sock(subflow);
731 static bool mptcp_skb_can_collapse_to(u64 write_seq,
732 const struct sk_buff *skb,
733 const struct mptcp_ext *mpext)
735 if (!tcp_skb_can_collapse_to(skb))
738 /* can collapse only if MPTCP level sequence is in order */
739 return mpext && mpext->data_seq + mpext->data_len == write_seq;
742 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
743 const struct page_frag *pfrag,
744 const struct mptcp_data_frag *df)
746 return df && pfrag->page == df->page &&
747 df->data_seq + df->data_len == msk->write_seq;
750 static void dfrag_uncharge(struct sock *sk, int len)
752 sk_mem_uncharge(sk, len);
753 sk_wmem_queued_add(sk, -len);
756 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
758 int len = dfrag->data_len + dfrag->overhead;
760 list_del(&dfrag->list);
761 dfrag_uncharge(sk, len);
762 put_page(dfrag->page);
765 static bool mptcp_is_writeable(struct mptcp_sock *msk)
767 struct mptcp_subflow_context *subflow;
769 if (!sk_stream_is_writeable((struct sock *)msk))
772 mptcp_for_each_subflow(msk, subflow) {
773 if (sk_stream_is_writeable(subflow->tcp_sock))
779 static void mptcp_clean_una(struct sock *sk)
781 struct mptcp_sock *msk = mptcp_sk(sk);
782 struct mptcp_data_frag *dtmp, *dfrag;
783 bool cleaned = false;
786 /* on fallback we just need to ignore snd_una, as this is really
789 if (__mptcp_check_fallback(msk))
790 atomic64_set(&msk->snd_una, msk->write_seq);
791 snd_una = atomic64_read(&msk->snd_una);
793 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
794 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
797 dfrag_clear(sk, dfrag);
801 dfrag = mptcp_rtx_head(sk);
802 if (dfrag && after64(snd_una, dfrag->data_seq)) {
803 u64 delta = snd_una - dfrag->data_seq;
805 if (WARN_ON_ONCE(delta > dfrag->data_len))
808 dfrag->data_seq += delta;
809 dfrag->offset += delta;
810 dfrag->data_len -= delta;
812 dfrag_uncharge(sk, delta);
818 sk_mem_reclaim_partial(sk);
820 /* Only wake up writers if a subflow is ready */
821 if (mptcp_is_writeable(msk)) {
822 set_bit(MPTCP_SEND_SPACE, &mptcp_sk(sk)->flags);
823 smp_mb__after_atomic();
825 /* set SEND_SPACE before sk_stream_write_space clears
828 sk_stream_write_space(sk);
833 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
836 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
838 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
839 pfrag, sk->sk_allocation)))
842 sk->sk_prot->enter_memory_pressure(sk);
843 sk_stream_moderate_sndbuf(sk);
847 static struct mptcp_data_frag *
848 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
851 int offset = ALIGN(orig_offset, sizeof(long));
852 struct mptcp_data_frag *dfrag;
854 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
856 dfrag->data_seq = msk->write_seq;
857 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
858 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
859 dfrag->page = pfrag->page;
864 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
865 struct msghdr *msg, struct mptcp_data_frag *dfrag,
866 long *timeo, int *pmss_now,
869 int mss_now, avail_size, size_goal, offset, ret, frag_truesize = 0;
870 bool dfrag_collapsed, can_collapse = false;
871 struct mptcp_sock *msk = mptcp_sk(sk);
872 struct mptcp_ext *mpext = NULL;
873 bool retransmission = !!dfrag;
874 struct sk_buff *skb, *tail;
875 struct page_frag *pfrag;
880 /* use the mptcp page cache so that we can easily move the data
881 * from one substream to another, but do per subflow memory accounting
882 * Note: pfrag is used only !retransmission, but the compiler if
883 * fooled into a warning if we don't init here
885 pfrag = sk_page_frag(sk);
886 if (!retransmission) {
887 write_seq = &msk->write_seq;
890 write_seq = &dfrag->data_seq;
894 /* compute copy limit */
895 mss_now = tcp_send_mss(ssk, &size_goal, msg->msg_flags);
897 *ps_goal = size_goal;
898 avail_size = size_goal;
899 skb = tcp_write_queue_tail(ssk);
901 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
903 /* Limit the write to the size available in the
904 * current skb, if any, so that we create at most a new skb.
905 * Explicitly tells TCP internals to avoid collapsing on later
906 * queue management operation, to avoid breaking the ext <->
907 * SSN association set here
909 can_collapse = (size_goal - skb->len > 0) &&
910 mptcp_skb_can_collapse_to(*write_seq, skb, mpext);
912 TCP_SKB_CB(skb)->eor = 1;
914 avail_size = size_goal - skb->len;
917 if (!retransmission) {
918 /* reuse tail pfrag, if possible, or carve a new one from the
921 dfrag = mptcp_rtx_tail(sk);
922 offset = pfrag->offset;
923 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
924 if (!dfrag_collapsed) {
925 dfrag = mptcp_carve_data_frag(msk, pfrag, offset);
926 offset = dfrag->offset;
927 frag_truesize = dfrag->overhead;
929 psize = min_t(size_t, pfrag->size - offset, avail_size);
932 pr_debug("left=%zu", msg_data_left(msg));
933 psize = copy_page_from_iter(pfrag->page, offset,
934 min_t(size_t, msg_data_left(msg),
937 pr_debug("left=%zu", msg_data_left(msg));
941 if (!sk_wmem_schedule(sk, psize + dfrag->overhead)) {
942 iov_iter_revert(&msg->msg_iter, psize);
946 offset = dfrag->offset;
947 psize = min_t(size_t, dfrag->data_len, avail_size);
950 /* tell the TCP stack to delay the push so that we can safely
951 * access the skb after the sendpages call
953 ret = do_tcp_sendpages(ssk, page, offset, psize,
954 msg->msg_flags | MSG_SENDPAGE_NOTLAST | MSG_DONTWAIT);
957 iov_iter_revert(&msg->msg_iter, psize);
961 frag_truesize += ret;
962 if (!retransmission) {
963 if (unlikely(ret < psize))
964 iov_iter_revert(&msg->msg_iter, psize - ret);
966 /* send successful, keep track of sent data for mptcp-level
969 dfrag->data_len += ret;
970 if (!dfrag_collapsed) {
971 get_page(dfrag->page);
972 list_add_tail(&dfrag->list, &msk->rtx_queue);
973 sk_wmem_queued_add(sk, frag_truesize);
975 sk_wmem_queued_add(sk, ret);
978 /* charge data on mptcp rtx queue to the master socket
979 * Note: we charge such data both to sk and ssk
981 sk->sk_forward_alloc -= frag_truesize;
984 /* if the tail skb extension is still the cached one, collapsing
985 * really happened. Note: we can't check for 'same skb' as the sk_buff
986 * hdr on tail can be transmitted, freed and re-allocated by the
987 * do_tcp_sendpages() call
989 tail = tcp_write_queue_tail(ssk);
990 if (mpext && tail && mpext == skb_ext_find(tail, SKB_EXT_MPTCP)) {
991 WARN_ON_ONCE(!can_collapse);
992 mpext->data_len += ret;
996 skb = tcp_write_queue_tail(ssk);
997 mpext = __skb_ext_set(skb, SKB_EXT_MPTCP, msk->cached_ext);
998 msk->cached_ext = NULL;
1000 memset(mpext, 0, sizeof(*mpext));
1001 mpext->data_seq = *write_seq;
1002 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1003 mpext->data_len = ret;
1007 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1008 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1012 if (!retransmission)
1013 pfrag->offset += frag_truesize;
1014 WRITE_ONCE(*write_seq, *write_seq + ret);
1015 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1020 static void mptcp_nospace(struct mptcp_sock *msk)
1022 struct mptcp_subflow_context *subflow;
1024 clear_bit(MPTCP_SEND_SPACE, &msk->flags);
1025 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
1027 mptcp_for_each_subflow(msk, subflow) {
1028 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1029 struct socket *sock = READ_ONCE(ssk->sk_socket);
1031 /* enables ssk->write_space() callbacks */
1033 set_bit(SOCK_NOSPACE, &sock->flags);
1037 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1039 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1041 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
1042 if (subflow->request_join && !subflow->fully_established)
1045 /* only send if our side has not closed yet */
1046 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
1049 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1050 sizeof(struct tcphdr) - \
1051 MAX_TCP_OPTION_SPACE - \
1052 sizeof(struct ipv6hdr) - \
1053 sizeof(struct frag_hdr))
1055 struct subflow_send_info {
1060 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk,
1063 struct subflow_send_info send_info[2];
1064 struct mptcp_subflow_context *subflow;
1065 int i, nr_active = 0;
1070 sock_owned_by_me((struct sock *)msk);
1073 if (!mptcp_ext_cache_refill(msk))
1076 if (__mptcp_check_fallback(msk)) {
1079 *sndbuf = msk->first->sk_sndbuf;
1080 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1083 /* re-use last subflow, if the burst allow that */
1084 if (msk->last_snd && msk->snd_burst > 0 &&
1085 sk_stream_memory_free(msk->last_snd) &&
1086 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1087 mptcp_for_each_subflow(msk, subflow) {
1088 ssk = mptcp_subflow_tcp_sock(subflow);
1089 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1091 return msk->last_snd;
1094 /* pick the subflow with the lower wmem/wspace ratio */
1095 for (i = 0; i < 2; ++i) {
1096 send_info[i].ssk = NULL;
1097 send_info[i].ratio = -1;
1099 mptcp_for_each_subflow(msk, subflow) {
1100 ssk = mptcp_subflow_tcp_sock(subflow);
1101 if (!mptcp_subflow_active(subflow))
1104 nr_active += !subflow->backup;
1105 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1106 if (!sk_stream_memory_free(subflow->tcp_sock))
1109 pace = READ_ONCE(ssk->sk_pacing_rate);
1113 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1115 if (ratio < send_info[subflow->backup].ratio) {
1116 send_info[subflow->backup].ssk = ssk;
1117 send_info[subflow->backup].ratio = ratio;
1121 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1122 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1123 send_info[1].ssk, send_info[1].ratio);
1125 /* pick the best backup if no other subflow is active */
1127 send_info[0].ssk = send_info[1].ssk;
1129 if (send_info[0].ssk) {
1130 msk->last_snd = send_info[0].ssk;
1131 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1132 sk_stream_wspace(msk->last_snd));
1133 return msk->last_snd;
1138 static void ssk_check_wmem(struct mptcp_sock *msk)
1140 if (unlikely(!mptcp_is_writeable(msk)))
1144 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1146 int mss_now = 0, size_goal = 0, ret = 0;
1147 struct mptcp_sock *msk = mptcp_sk(sk);
1148 struct page_frag *pfrag;
1155 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1160 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1162 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1163 ret = sk_stream_wait_connect(sk, &timeo);
1168 pfrag = sk_page_frag(sk);
1170 mptcp_clean_una(sk);
1172 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1177 __mptcp_flush_join_list(msk);
1178 ssk = mptcp_subflow_get_send(msk, &sndbuf);
1179 while (!sk_stream_memory_free(sk) ||
1181 !mptcp_page_frag_refill(ssk, pfrag)) {
1183 /* make sure retransmit timer is
1184 * running before we wait for memory.
1186 * The retransmit timer might be needed
1187 * to make the peer send an up-to-date
1190 mptcp_set_timeout(sk, ssk);
1191 if (!mptcp_timer_pending(sk))
1192 mptcp_reset_timer(sk);
1196 ret = sk_stream_wait_memory(sk, &timeo);
1200 mptcp_clean_una(sk);
1202 ssk = mptcp_subflow_get_send(msk, &sndbuf);
1203 if (list_empty(&msk->conn_list)) {
1209 /* do auto tuning */
1210 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1211 sndbuf > READ_ONCE(sk->sk_sndbuf))
1212 WRITE_ONCE(sk->sk_sndbuf, sndbuf);
1214 pr_debug("conn_list->subflow=%p", ssk);
1217 tx_ok = msg_data_left(msg);
1219 ret = mptcp_sendmsg_frag(sk, ssk, msg, NULL, &timeo, &mss_now,
1222 if (ret == -EAGAIN && timeo > 0) {
1223 mptcp_set_timeout(sk, ssk);
1230 /* burst can be negative, we will try move to the next subflow
1231 * at selection time, if possible.
1233 msk->snd_burst -= ret;
1236 tx_ok = msg_data_left(msg);
1240 if (!sk_stream_memory_free(ssk) ||
1241 !mptcp_page_frag_refill(ssk, pfrag) ||
1242 !mptcp_ext_cache_refill(msk)) {
1243 tcp_push(ssk, msg->msg_flags, mss_now,
1244 tcp_sk(ssk)->nonagle, size_goal);
1245 mptcp_set_timeout(sk, ssk);
1250 /* memory is charged to mptcp level socket as well, i.e.
1251 * if msg is very large, mptcp socket may run out of buffer
1252 * space. mptcp_clean_una() will release data that has
1253 * been acked at mptcp level in the mean time, so there is
1254 * a good chance we can continue sending data right away.
1256 * Normally, when the tcp subflow can accept more data, then
1257 * so can the MPTCP socket. However, we need to cope with
1258 * peers that might lag behind in their MPTCP-level
1259 * acknowledgements, i.e. data might have been acked at
1260 * tcp level only. So, we must also check the MPTCP socket
1261 * limits before we send more data.
1263 if (unlikely(!sk_stream_memory_free(sk))) {
1264 tcp_push(ssk, msg->msg_flags, mss_now,
1265 tcp_sk(ssk)->nonagle, size_goal);
1266 mptcp_clean_una(sk);
1267 if (!sk_stream_memory_free(sk)) {
1268 /* can't send more for now, need to wait for
1269 * MPTCP-level ACKs from peer.
1271 * Wakeup will happen via mptcp_clean_una().
1273 mptcp_set_timeout(sk, ssk);
1280 mptcp_set_timeout(sk, ssk);
1282 tcp_push(ssk, msg->msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1285 /* start the timer, if it's not pending */
1286 if (!mptcp_timer_pending(sk))
1287 mptcp_reset_timer(sk);
1292 ssk_check_wmem(msk);
1294 return copied ? : ret;
1297 static void mptcp_wait_data(struct sock *sk, long *timeo)
1299 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1300 struct mptcp_sock *msk = mptcp_sk(sk);
1302 add_wait_queue(sk_sleep(sk), &wait);
1303 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1305 sk_wait_event(sk, timeo,
1306 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1308 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1309 remove_wait_queue(sk_sleep(sk), &wait);
1312 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1316 struct sock *sk = (struct sock *)msk;
1317 struct sk_buff *skb;
1320 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1321 u32 offset = MPTCP_SKB_CB(skb)->offset;
1322 u32 data_len = skb->len - offset;
1323 u32 count = min_t(size_t, len - copied, data_len);
1326 err = skb_copy_datagram_msg(skb, offset, msg, count);
1327 if (unlikely(err < 0)) {
1335 if (count < data_len) {
1336 MPTCP_SKB_CB(skb)->offset += count;
1340 __skb_unlink(skb, &sk->sk_receive_queue);
1350 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1352 * Only difference: Use highest rtt estimate of the subflows in use.
1354 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1356 struct mptcp_subflow_context *subflow;
1357 struct sock *sk = (struct sock *)msk;
1358 u32 time, advmss = 1;
1361 sock_owned_by_me(sk);
1366 msk->rcvq_space.copied += copied;
1368 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1369 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1371 rtt_us = msk->rcvq_space.rtt_us;
1372 if (rtt_us && time < (rtt_us >> 3))
1376 mptcp_for_each_subflow(msk, subflow) {
1377 const struct tcp_sock *tp;
1381 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1383 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1384 sf_advmss = READ_ONCE(tp->advmss);
1386 rtt_us = max(sf_rtt_us, rtt_us);
1387 advmss = max(sf_advmss, advmss);
1390 msk->rcvq_space.rtt_us = rtt_us;
1391 if (time < (rtt_us >> 3) || rtt_us == 0)
1394 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1397 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1398 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1402 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1404 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1406 do_div(grow, msk->rcvq_space.space);
1407 rcvwin += (grow << 1);
1409 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1410 while (tcp_win_from_space(sk, rcvmem) < advmss)
1413 do_div(rcvwin, advmss);
1414 rcvbuf = min_t(u64, rcvwin * rcvmem,
1415 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1417 if (rcvbuf > sk->sk_rcvbuf) {
1420 window_clamp = tcp_win_from_space(sk, rcvbuf);
1421 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1423 /* Make subflows follow along. If we do not do this, we
1424 * get drops at subflow level if skbs can't be moved to
1425 * the mptcp rx queue fast enough (announced rcv_win can
1426 * exceed ssk->sk_rcvbuf).
1428 mptcp_for_each_subflow(msk, subflow) {
1432 ssk = mptcp_subflow_tcp_sock(subflow);
1433 slow = lock_sock_fast(ssk);
1434 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1435 tcp_sk(ssk)->window_clamp = window_clamp;
1436 tcp_cleanup_rbuf(ssk, 1);
1437 unlock_sock_fast(ssk, slow);
1442 msk->rcvq_space.space = msk->rcvq_space.copied;
1444 msk->rcvq_space.copied = 0;
1445 msk->rcvq_space.time = mstamp;
1448 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1450 unsigned int moved = 0;
1453 /* avoid looping forever below on racing close */
1454 if (((struct sock *)msk)->sk_state == TCP_CLOSE)
1457 __mptcp_flush_join_list(msk);
1459 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1465 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1469 if (mptcp_ofo_queue(msk) || moved > 0) {
1470 mptcp_check_data_fin((struct sock *)msk);
1476 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1477 int nonblock, int flags, int *addr_len)
1479 struct mptcp_sock *msk = mptcp_sk(sk);
1484 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1488 timeo = sock_rcvtimeo(sk, nonblock);
1490 len = min_t(size_t, len, INT_MAX);
1491 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1492 __mptcp_flush_join_list(msk);
1494 while (len > (size_t)copied) {
1497 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1498 if (unlikely(bytes_read < 0)) {
1500 copied = bytes_read;
1504 copied += bytes_read;
1506 if (skb_queue_empty(&sk->sk_receive_queue) &&
1507 __mptcp_move_skbs(msk))
1510 /* only the master socket status is relevant here. The exit
1511 * conditions mirror closely tcp_recvmsg()
1513 if (copied >= target)
1518 sk->sk_state == TCP_CLOSE ||
1519 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1521 signal_pending(current))
1525 copied = sock_error(sk);
1529 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1530 mptcp_check_for_eof(msk);
1532 if (sk->sk_shutdown & RCV_SHUTDOWN)
1535 if (sk->sk_state == TCP_CLOSE) {
1545 if (signal_pending(current)) {
1546 copied = sock_intr_errno(timeo);
1551 pr_debug("block timeout %ld", timeo);
1552 mptcp_wait_data(sk, &timeo);
1555 if (skb_queue_empty(&sk->sk_receive_queue)) {
1556 /* entire backlog drained, clear DATA_READY. */
1557 clear_bit(MPTCP_DATA_READY, &msk->flags);
1559 /* .. race-breaker: ssk might have gotten new data
1560 * after last __mptcp_move_skbs() returned false.
1562 if (unlikely(__mptcp_move_skbs(msk)))
1563 set_bit(MPTCP_DATA_READY, &msk->flags);
1564 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1565 /* data to read but mptcp_wait_data() cleared DATA_READY */
1566 set_bit(MPTCP_DATA_READY, &msk->flags);
1569 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
1570 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
1571 skb_queue_empty(&sk->sk_receive_queue), copied);
1572 mptcp_rcv_space_adjust(msk, copied);
1578 static void mptcp_retransmit_handler(struct sock *sk)
1580 struct mptcp_sock *msk = mptcp_sk(sk);
1582 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->write_seq)) {
1583 mptcp_stop_timer(sk);
1585 set_bit(MPTCP_WORK_RTX, &msk->flags);
1586 if (schedule_work(&msk->work))
1591 static void mptcp_retransmit_timer(struct timer_list *t)
1593 struct inet_connection_sock *icsk = from_timer(icsk, t,
1594 icsk_retransmit_timer);
1595 struct sock *sk = &icsk->icsk_inet.sk;
1598 if (!sock_owned_by_user(sk)) {
1599 mptcp_retransmit_handler(sk);
1601 /* delegate our work to tcp_release_cb() */
1602 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1610 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1613 * A backup subflow is returned only if that is the only kind available.
1615 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1617 struct mptcp_subflow_context *subflow;
1618 struct sock *backup = NULL;
1620 sock_owned_by_me((const struct sock *)msk);
1622 if (__mptcp_check_fallback(msk))
1625 mptcp_for_each_subflow(msk, subflow) {
1626 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1628 if (!mptcp_subflow_active(subflow))
1631 /* still data outstanding at TCP level? Don't retransmit. */
1632 if (!tcp_write_queue_empty(ssk))
1635 if (subflow->backup) {
1647 /* subflow sockets can be either outgoing (connect) or incoming
1650 * Outgoing subflows use in-kernel sockets.
1651 * Incoming subflows do not have their own 'struct socket' allocated,
1652 * so we need to use tcp_close() after detaching them from the mptcp
1655 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1656 struct mptcp_subflow_context *subflow,
1659 struct socket *sock = READ_ONCE(ssk->sk_socket);
1661 list_del(&subflow->node);
1663 if (sock && sock != sk->sk_socket) {
1664 /* outgoing subflow */
1667 /* incoming subflow */
1668 tcp_close(ssk, timeout);
1672 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1677 static void pm_work(struct mptcp_sock *msk)
1679 struct mptcp_pm_data *pm = &msk->pm;
1681 spin_lock_bh(&msk->pm.lock);
1683 pr_debug("msk=%p status=%x", msk, pm->status);
1684 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1685 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1686 mptcp_pm_nl_add_addr_received(msk);
1688 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
1689 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
1690 mptcp_pm_nl_rm_addr_received(msk);
1692 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1693 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1694 mptcp_pm_nl_fully_established(msk);
1696 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1697 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1698 mptcp_pm_nl_subflow_established(msk);
1701 spin_unlock_bh(&msk->pm.lock);
1704 static void mptcp_worker(struct work_struct *work)
1706 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1707 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1708 int orig_len, orig_offset, mss_now = 0, size_goal = 0;
1709 struct mptcp_data_frag *dfrag;
1712 struct msghdr msg = {
1713 .msg_flags = MSG_DONTWAIT,
1718 mptcp_clean_una(sk);
1719 mptcp_check_data_fin_ack(sk);
1720 __mptcp_flush_join_list(msk);
1721 __mptcp_move_skbs(msk);
1726 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1727 mptcp_check_for_eof(msk);
1729 mptcp_check_data_fin(sk);
1731 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1734 dfrag = mptcp_rtx_head(sk);
1738 if (!mptcp_ext_cache_refill(msk))
1741 ssk = mptcp_subflow_get_retrans(msk);
1747 orig_len = dfrag->data_len;
1748 orig_offset = dfrag->offset;
1749 orig_write_seq = dfrag->data_seq;
1750 while (dfrag->data_len > 0) {
1751 int ret = mptcp_sendmsg_frag(sk, ssk, &msg, dfrag, &timeo,
1752 &mss_now, &size_goal);
1756 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1758 dfrag->data_len -= ret;
1759 dfrag->offset += ret;
1761 if (!mptcp_ext_cache_refill(msk))
1765 tcp_push(ssk, msg.msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1768 dfrag->data_seq = orig_write_seq;
1769 dfrag->offset = orig_offset;
1770 dfrag->data_len = orig_len;
1772 mptcp_set_timeout(sk, ssk);
1776 if (!mptcp_timer_pending(sk))
1777 mptcp_reset_timer(sk);
1784 static int __mptcp_init_sock(struct sock *sk)
1786 struct mptcp_sock *msk = mptcp_sk(sk);
1788 spin_lock_init(&msk->join_list_lock);
1790 INIT_LIST_HEAD(&msk->conn_list);
1791 INIT_LIST_HEAD(&msk->join_list);
1792 INIT_LIST_HEAD(&msk->rtx_queue);
1793 __set_bit(MPTCP_SEND_SPACE, &msk->flags);
1794 INIT_WORK(&msk->work, mptcp_worker);
1795 msk->out_of_order_queue = RB_ROOT;
1798 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1800 mptcp_pm_data_init(msk);
1802 /* re-use the csk retrans timer for MPTCP-level retrans */
1803 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1808 static int mptcp_init_sock(struct sock *sk)
1810 struct net *net = sock_net(sk);
1813 ret = __mptcp_init_sock(sk);
1817 if (!mptcp_is_enabled(net))
1818 return -ENOPROTOOPT;
1820 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1823 ret = __mptcp_socket_create(mptcp_sk(sk));
1827 sk_sockets_allocated_inc(sk);
1828 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1829 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
1834 static void __mptcp_clear_xmit(struct sock *sk)
1836 struct mptcp_sock *msk = mptcp_sk(sk);
1837 struct mptcp_data_frag *dtmp, *dfrag;
1839 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
1841 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
1842 dfrag_clear(sk, dfrag);
1845 static void mptcp_cancel_work(struct sock *sk)
1847 struct mptcp_sock *msk = mptcp_sk(sk);
1849 if (cancel_work_sync(&msk->work))
1853 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
1857 switch (ssk->sk_state) {
1859 if (!(how & RCV_SHUTDOWN))
1863 tcp_disconnect(ssk, O_NONBLOCK);
1866 if (__mptcp_check_fallback(mptcp_sk(sk))) {
1867 pr_debug("Fallback");
1868 ssk->sk_shutdown |= how;
1869 tcp_shutdown(ssk, how);
1871 pr_debug("Sending DATA_FIN on subflow %p", ssk);
1872 mptcp_set_timeout(sk, ssk);
1881 static const unsigned char new_state[16] = {
1882 /* current state: new state: action: */
1883 [0 /* (Invalid) */] = TCP_CLOSE,
1884 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1885 [TCP_SYN_SENT] = TCP_CLOSE,
1886 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1887 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1888 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1889 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
1890 [TCP_CLOSE] = TCP_CLOSE,
1891 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1892 [TCP_LAST_ACK] = TCP_LAST_ACK,
1893 [TCP_LISTEN] = TCP_CLOSE,
1894 [TCP_CLOSING] = TCP_CLOSING,
1895 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1898 static int mptcp_close_state(struct sock *sk)
1900 int next = (int)new_state[sk->sk_state];
1901 int ns = next & TCP_STATE_MASK;
1903 inet_sk_state_store(sk, ns);
1905 return next & TCP_ACTION_FIN;
1908 static void mptcp_close(struct sock *sk, long timeout)
1910 struct mptcp_subflow_context *subflow, *tmp;
1911 struct mptcp_sock *msk = mptcp_sk(sk);
1912 LIST_HEAD(conn_list);
1915 sk->sk_shutdown = SHUTDOWN_MASK;
1917 if (sk->sk_state == TCP_LISTEN) {
1918 inet_sk_state_store(sk, TCP_CLOSE);
1920 } else if (sk->sk_state == TCP_CLOSE) {
1924 if (__mptcp_check_fallback(msk)) {
1926 } else if (mptcp_close_state(sk)) {
1927 pr_debug("Sending DATA_FIN sk=%p", sk);
1928 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
1929 WRITE_ONCE(msk->snd_data_fin_enable, 1);
1931 mptcp_for_each_subflow(msk, subflow) {
1932 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
1934 mptcp_subflow_shutdown(sk, tcp_sk, SHUTDOWN_MASK);
1938 sk_stream_wait_close(sk, timeout);
1941 inet_sk_state_store(sk, TCP_CLOSE);
1944 /* be sure to always acquire the join list lock, to sync vs
1945 * mptcp_finish_join().
1947 spin_lock_bh(&msk->join_list_lock);
1948 list_splice_tail_init(&msk->join_list, &msk->conn_list);
1949 spin_unlock_bh(&msk->join_list_lock);
1950 list_splice_init(&msk->conn_list, &conn_list);
1952 __mptcp_clear_xmit(sk);
1956 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
1957 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1958 __mptcp_close_ssk(sk, ssk, subflow, timeout);
1961 mptcp_cancel_work(sk);
1963 __skb_queue_purge(&sk->sk_receive_queue);
1965 sk_common_release(sk);
1968 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
1970 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1971 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
1972 struct ipv6_pinfo *msk6 = inet6_sk(msk);
1974 msk->sk_v6_daddr = ssk->sk_v6_daddr;
1975 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
1978 msk6->saddr = ssk6->saddr;
1979 msk6->flow_label = ssk6->flow_label;
1983 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
1984 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
1985 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
1986 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
1987 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
1988 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
1991 static int mptcp_disconnect(struct sock *sk, int flags)
1993 /* Should never be called.
1994 * inet_stream_connect() calls ->disconnect, but that
1995 * refers to the subflow socket, not the mptcp one.
2001 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2002 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2004 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2006 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2010 struct sock *mptcp_sk_clone(const struct sock *sk,
2011 const struct mptcp_options_received *mp_opt,
2012 struct request_sock *req)
2014 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2015 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2016 struct mptcp_sock *msk;
2022 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2023 if (nsk->sk_family == AF_INET6)
2024 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2027 __mptcp_init_sock(nsk);
2029 msk = mptcp_sk(nsk);
2030 msk->local_key = subflow_req->local_key;
2031 msk->token = subflow_req->token;
2032 msk->subflow = NULL;
2033 WRITE_ONCE(msk->fully_established, false);
2035 msk->write_seq = subflow_req->idsn + 1;
2036 atomic64_set(&msk->snd_una, msk->write_seq);
2037 if (mp_opt->mp_capable) {
2038 msk->can_ack = true;
2039 msk->remote_key = mp_opt->sndr_key;
2040 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2042 msk->ack_seq = ack_seq;
2045 sock_reset_flag(nsk, SOCK_RCU_FREE);
2046 /* will be fully established after successful MPC subflow creation */
2047 inet_sk_state_store(nsk, TCP_SYN_RECV);
2048 bh_unlock_sock(nsk);
2050 /* keep a single reference */
2055 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2057 const struct tcp_sock *tp = tcp_sk(ssk);
2059 msk->rcvq_space.copied = 0;
2060 msk->rcvq_space.rtt_us = 0;
2062 msk->rcvq_space.time = tp->tcp_mstamp;
2064 /* initial rcv_space offering made to peer */
2065 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2066 TCP_INIT_CWND * tp->advmss);
2067 if (msk->rcvq_space.space == 0)
2068 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2071 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2074 struct mptcp_sock *msk = mptcp_sk(sk);
2075 struct socket *listener;
2078 listener = __mptcp_nmpc_socket(msk);
2079 if (WARN_ON_ONCE(!listener)) {
2084 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2085 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2089 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2090 if (sk_is_mptcp(newsk)) {
2091 struct mptcp_subflow_context *subflow;
2092 struct sock *new_mptcp_sock;
2093 struct sock *ssk = newsk;
2095 subflow = mptcp_subflow_ctx(newsk);
2096 new_mptcp_sock = subflow->conn;
2098 /* is_mptcp should be false if subflow->conn is missing, see
2099 * subflow_syn_recv_sock()
2101 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2102 tcp_sk(newsk)->is_mptcp = 0;
2106 /* acquire the 2nd reference for the owning socket */
2107 sock_hold(new_mptcp_sock);
2110 bh_lock_sock(new_mptcp_sock);
2111 msk = mptcp_sk(new_mptcp_sock);
2114 newsk = new_mptcp_sock;
2115 mptcp_copy_inaddrs(newsk, ssk);
2116 list_add(&subflow->node, &msk->conn_list);
2118 mptcp_rcv_space_init(msk, ssk);
2119 bh_unlock_sock(new_mptcp_sock);
2121 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2124 MPTCP_INC_STATS(sock_net(sk),
2125 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2131 void mptcp_destroy_common(struct mptcp_sock *msk)
2133 skb_rbtree_purge(&msk->out_of_order_queue);
2134 mptcp_token_destroy(msk);
2135 mptcp_pm_free_anno_list(msk);
2138 static void mptcp_destroy(struct sock *sk)
2140 struct mptcp_sock *msk = mptcp_sk(sk);
2142 if (msk->cached_ext)
2143 __skb_ext_put(msk->cached_ext);
2145 mptcp_destroy_common(msk);
2146 sk_sockets_allocated_dec(sk);
2149 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2150 sockptr_t optval, unsigned int optlen)
2152 struct sock *sk = (struct sock *)msk;
2153 struct socket *ssock;
2160 ssock = __mptcp_nmpc_socket(msk);
2166 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2168 if (optname == SO_REUSEPORT)
2169 sk->sk_reuseport = ssock->sk->sk_reuseport;
2170 else if (optname == SO_REUSEADDR)
2171 sk->sk_reuse = ssock->sk->sk_reuse;
2177 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2180 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2181 sockptr_t optval, unsigned int optlen)
2183 struct sock *sk = (struct sock *)msk;
2184 int ret = -EOPNOTSUPP;
2185 struct socket *ssock;
2190 ssock = __mptcp_nmpc_socket(msk);
2196 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2198 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2207 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2208 sockptr_t optval, unsigned int optlen)
2210 struct mptcp_sock *msk = mptcp_sk(sk);
2213 pr_debug("msk=%p", msk);
2215 if (level == SOL_SOCKET)
2216 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2218 /* @@ the meaning of setsockopt() when the socket is connected and
2219 * there are multiple subflows is not yet defined. It is up to the
2220 * MPTCP-level socket to configure the subflows until the subflow
2221 * is in TCP fallback, when TCP socket options are passed through
2222 * to the one remaining subflow.
2225 ssk = __mptcp_tcp_fallback(msk);
2228 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2230 if (level == SOL_IPV6)
2231 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2236 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2237 char __user *optval, int __user *option)
2239 struct mptcp_sock *msk = mptcp_sk(sk);
2242 pr_debug("msk=%p", msk);
2244 /* @@ the meaning of setsockopt() when the socket is connected and
2245 * there are multiple subflows is not yet defined. It is up to the
2246 * MPTCP-level socket to configure the subflows until the subflow
2247 * is in TCP fallback, when socket options are passed through
2248 * to the one remaining subflow.
2251 ssk = __mptcp_tcp_fallback(msk);
2254 return tcp_getsockopt(ssk, level, optname, optval, option);
2259 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
2260 TCPF_WRITE_TIMER_DEFERRED)
2262 /* this is very alike tcp_release_cb() but we must handle differently a
2263 * different set of events
2265 static void mptcp_release_cb(struct sock *sk)
2267 unsigned long flags, nflags;
2270 flags = sk->sk_tsq_flags;
2271 if (!(flags & MPTCP_DEFERRED_ALL))
2273 nflags = flags & ~MPTCP_DEFERRED_ALL;
2274 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2276 sock_release_ownership(sk);
2278 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
2279 struct mptcp_sock *msk = mptcp_sk(sk);
2282 ssk = mptcp_subflow_recv_lookup(msk);
2283 if (!ssk || !schedule_work(&msk->work))
2287 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2288 mptcp_retransmit_handler(sk);
2293 static int mptcp_hash(struct sock *sk)
2295 /* should never be called,
2296 * we hash the TCP subflows not the master socket
2302 static void mptcp_unhash(struct sock *sk)
2304 /* called from sk_common_release(), but nothing to do here */
2307 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2309 struct mptcp_sock *msk = mptcp_sk(sk);
2310 struct socket *ssock;
2312 ssock = __mptcp_nmpc_socket(msk);
2313 pr_debug("msk=%p, subflow=%p", msk, ssock);
2314 if (WARN_ON_ONCE(!ssock))
2317 return inet_csk_get_port(ssock->sk, snum);
2320 void mptcp_finish_connect(struct sock *ssk)
2322 struct mptcp_subflow_context *subflow;
2323 struct mptcp_sock *msk;
2327 subflow = mptcp_subflow_ctx(ssk);
2331 pr_debug("msk=%p, token=%u", sk, subflow->token);
2333 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2335 subflow->map_seq = ack_seq;
2336 subflow->map_subflow_seq = 1;
2338 /* the socket is not connected yet, no msk/subflow ops can access/race
2339 * accessing the field below
2341 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2342 WRITE_ONCE(msk->local_key, subflow->local_key);
2343 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2344 WRITE_ONCE(msk->ack_seq, ack_seq);
2345 WRITE_ONCE(msk->can_ack, 1);
2346 atomic64_set(&msk->snd_una, msk->write_seq);
2348 mptcp_pm_new_connection(msk, 0);
2350 mptcp_rcv_space_init(msk, ssk);
2353 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2355 write_lock_bh(&sk->sk_callback_lock);
2356 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2357 sk_set_socket(sk, parent);
2358 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2359 write_unlock_bh(&sk->sk_callback_lock);
2362 bool mptcp_finish_join(struct sock *sk)
2364 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
2365 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2366 struct sock *parent = (void *)msk;
2367 struct socket *parent_sock;
2370 pr_debug("msk=%p, subflow=%p", msk, subflow);
2372 /* mptcp socket already closing? */
2373 if (!mptcp_is_fully_established(parent))
2376 if (!msk->pm.server_side)
2379 if (!mptcp_pm_allow_new_subflow(msk))
2382 /* active connections are already on conn_list, and we can't acquire
2384 * use the join list lock as synchronization point and double-check
2385 * msk status to avoid racing with mptcp_close()
2387 spin_lock_bh(&msk->join_list_lock);
2388 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2389 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node)))
2390 list_add_tail(&subflow->node, &msk->join_list);
2391 spin_unlock_bh(&msk->join_list_lock);
2395 /* attach to msk socket only after we are sure he will deal with us
2398 parent_sock = READ_ONCE(parent->sk_socket);
2399 if (parent_sock && !sk->sk_socket)
2400 mptcp_sock_graft(sk, parent_sock);
2401 subflow->map_seq = msk->ack_seq;
2405 static bool mptcp_memory_free(const struct sock *sk, int wake)
2407 struct mptcp_sock *msk = mptcp_sk(sk);
2409 return wake ? test_bit(MPTCP_SEND_SPACE, &msk->flags) : true;
2412 static struct proto mptcp_prot = {
2414 .owner = THIS_MODULE,
2415 .init = mptcp_init_sock,
2416 .disconnect = mptcp_disconnect,
2417 .close = mptcp_close,
2418 .accept = mptcp_accept,
2419 .setsockopt = mptcp_setsockopt,
2420 .getsockopt = mptcp_getsockopt,
2421 .shutdown = tcp_shutdown,
2422 .destroy = mptcp_destroy,
2423 .sendmsg = mptcp_sendmsg,
2424 .recvmsg = mptcp_recvmsg,
2425 .release_cb = mptcp_release_cb,
2427 .unhash = mptcp_unhash,
2428 .get_port = mptcp_get_port,
2429 .sockets_allocated = &mptcp_sockets_allocated,
2430 .memory_allocated = &tcp_memory_allocated,
2431 .memory_pressure = &tcp_memory_pressure,
2432 .stream_memory_free = mptcp_memory_free,
2433 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2434 .sysctl_mem = sysctl_tcp_mem,
2435 .obj_size = sizeof(struct mptcp_sock),
2436 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2437 .no_autobind = true,
2440 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2442 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2443 struct socket *ssock;
2446 lock_sock(sock->sk);
2447 ssock = __mptcp_nmpc_socket(msk);
2453 err = ssock->ops->bind(ssock, uaddr, addr_len);
2455 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2458 release_sock(sock->sk);
2462 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2463 struct mptcp_subflow_context *subflow)
2465 subflow->request_mptcp = 0;
2466 __mptcp_do_fallback(msk);
2469 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2470 int addr_len, int flags)
2472 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2473 struct mptcp_subflow_context *subflow;
2474 struct socket *ssock;
2477 lock_sock(sock->sk);
2478 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2479 /* pending connection or invalid state, let existing subflow
2482 ssock = msk->subflow;
2486 ssock = __mptcp_nmpc_socket(msk);
2492 mptcp_token_destroy(msk);
2493 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2494 subflow = mptcp_subflow_ctx(ssock->sk);
2495 #ifdef CONFIG_TCP_MD5SIG
2496 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2499 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2500 mptcp_subflow_early_fallback(msk, subflow);
2502 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2503 mptcp_subflow_early_fallback(msk, subflow);
2506 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2507 sock->state = ssock->state;
2509 /* on successful connect, the msk state will be moved to established by
2510 * subflow_finish_connect()
2512 if (!err || err == -EINPROGRESS)
2513 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2515 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2518 release_sock(sock->sk);
2522 static int mptcp_listen(struct socket *sock, int backlog)
2524 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2525 struct socket *ssock;
2528 pr_debug("msk=%p", msk);
2530 lock_sock(sock->sk);
2531 ssock = __mptcp_nmpc_socket(msk);
2537 mptcp_token_destroy(msk);
2538 inet_sk_state_store(sock->sk, TCP_LISTEN);
2539 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2541 err = ssock->ops->listen(ssock, backlog);
2542 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2544 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2547 release_sock(sock->sk);
2551 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2552 int flags, bool kern)
2554 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2555 struct socket *ssock;
2558 pr_debug("msk=%p", msk);
2560 lock_sock(sock->sk);
2561 if (sock->sk->sk_state != TCP_LISTEN)
2564 ssock = __mptcp_nmpc_socket(msk);
2568 clear_bit(MPTCP_DATA_READY, &msk->flags);
2569 sock_hold(ssock->sk);
2570 release_sock(sock->sk);
2572 err = ssock->ops->accept(sock, newsock, flags, kern);
2573 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2574 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2575 struct mptcp_subflow_context *subflow;
2577 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2578 * This is needed so NOSPACE flag can be set from tcp stack.
2580 __mptcp_flush_join_list(msk);
2581 mptcp_for_each_subflow(msk, subflow) {
2582 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2584 if (!ssk->sk_socket)
2585 mptcp_sock_graft(ssk, newsock);
2589 if (inet_csk_listen_poll(ssock->sk))
2590 set_bit(MPTCP_DATA_READY, &msk->flags);
2591 sock_put(ssock->sk);
2595 release_sock(sock->sk);
2599 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2601 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2605 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2606 struct poll_table_struct *wait)
2608 struct sock *sk = sock->sk;
2609 struct mptcp_sock *msk;
2614 sock_poll_wait(file, sock, wait);
2616 state = inet_sk_state_load(sk);
2617 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
2618 if (state == TCP_LISTEN)
2619 return mptcp_check_readable(msk);
2621 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2622 mask |= mptcp_check_readable(msk);
2623 if (test_bit(MPTCP_SEND_SPACE, &msk->flags))
2624 mask |= EPOLLOUT | EPOLLWRNORM;
2626 if (sk->sk_shutdown & RCV_SHUTDOWN)
2627 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2632 static int mptcp_shutdown(struct socket *sock, int how)
2634 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2635 struct mptcp_subflow_context *subflow;
2638 pr_debug("sk=%p, how=%d", msk, how);
2640 lock_sock(sock->sk);
2643 if ((how & ~SHUTDOWN_MASK) || !how) {
2648 if (sock->state == SS_CONNECTING) {
2649 if ((1 << sock->sk->sk_state) &
2650 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2651 sock->state = SS_DISCONNECTING;
2653 sock->state = SS_CONNECTED;
2656 /* If we've already sent a FIN, or it's a closed state, skip this. */
2657 if (__mptcp_check_fallback(msk)) {
2658 if (how == SHUT_WR || how == SHUT_RDWR)
2659 inet_sk_state_store(sock->sk, TCP_FIN_WAIT1);
2661 mptcp_for_each_subflow(msk, subflow) {
2662 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2664 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2666 } else if ((how & SEND_SHUTDOWN) &&
2667 ((1 << sock->sk->sk_state) &
2668 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2669 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) &&
2670 mptcp_close_state(sock->sk)) {
2671 __mptcp_flush_join_list(msk);
2673 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2674 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2676 mptcp_for_each_subflow(msk, subflow) {
2677 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2679 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2683 /* Wake up anyone sleeping in poll. */
2684 sock->sk->sk_state_change(sock->sk);
2687 release_sock(sock->sk);
2692 static const struct proto_ops mptcp_stream_ops = {
2694 .owner = THIS_MODULE,
2695 .release = inet_release,
2697 .connect = mptcp_stream_connect,
2698 .socketpair = sock_no_socketpair,
2699 .accept = mptcp_stream_accept,
2700 .getname = inet_getname,
2702 .ioctl = inet_ioctl,
2703 .gettstamp = sock_gettstamp,
2704 .listen = mptcp_listen,
2705 .shutdown = mptcp_shutdown,
2706 .setsockopt = sock_common_setsockopt,
2707 .getsockopt = sock_common_getsockopt,
2708 .sendmsg = inet_sendmsg,
2709 .recvmsg = inet_recvmsg,
2710 .mmap = sock_no_mmap,
2711 .sendpage = inet_sendpage,
2714 static struct inet_protosw mptcp_protosw = {
2715 .type = SOCK_STREAM,
2716 .protocol = IPPROTO_MPTCP,
2717 .prot = &mptcp_prot,
2718 .ops = &mptcp_stream_ops,
2719 .flags = INET_PROTOSW_ICSK,
2722 void __init mptcp_proto_init(void)
2724 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2726 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2727 panic("Failed to allocate MPTCP pcpu counter\n");
2729 mptcp_subflow_init();
2733 if (proto_register(&mptcp_prot, 1) != 0)
2734 panic("Failed to register MPTCP proto.\n");
2736 inet_register_protosw(&mptcp_protosw);
2738 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2741 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2742 static const struct proto_ops mptcp_v6_stream_ops = {
2744 .owner = THIS_MODULE,
2745 .release = inet6_release,
2747 .connect = mptcp_stream_connect,
2748 .socketpair = sock_no_socketpair,
2749 .accept = mptcp_stream_accept,
2750 .getname = inet6_getname,
2752 .ioctl = inet6_ioctl,
2753 .gettstamp = sock_gettstamp,
2754 .listen = mptcp_listen,
2755 .shutdown = mptcp_shutdown,
2756 .setsockopt = sock_common_setsockopt,
2757 .getsockopt = sock_common_getsockopt,
2758 .sendmsg = inet6_sendmsg,
2759 .recvmsg = inet6_recvmsg,
2760 .mmap = sock_no_mmap,
2761 .sendpage = inet_sendpage,
2762 #ifdef CONFIG_COMPAT
2763 .compat_ioctl = inet6_compat_ioctl,
2767 static struct proto mptcp_v6_prot;
2769 static void mptcp_v6_destroy(struct sock *sk)
2772 inet6_destroy_sock(sk);
2775 static struct inet_protosw mptcp_v6_protosw = {
2776 .type = SOCK_STREAM,
2777 .protocol = IPPROTO_MPTCP,
2778 .prot = &mptcp_v6_prot,
2779 .ops = &mptcp_v6_stream_ops,
2780 .flags = INET_PROTOSW_ICSK,
2783 int __init mptcp_proto_v6_init(void)
2787 mptcp_v6_prot = mptcp_prot;
2788 strcpy(mptcp_v6_prot.name, "MPTCPv6");
2789 mptcp_v6_prot.slab = NULL;
2790 mptcp_v6_prot.destroy = mptcp_v6_destroy;
2791 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
2793 err = proto_register(&mptcp_v6_prot, 1);
2797 err = inet6_register_protosw(&mptcp_v6_protosw);
2799 proto_unregister(&mptcp_v6_prot);