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>
28 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
30 struct mptcp_sock msk;
41 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
43 static struct percpu_counter mptcp_sockets_allocated;
45 static void __mptcp_destroy_sock(struct sock *sk);
46 static void __mptcp_check_send_data_fin(struct sock *sk);
48 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
49 * completed yet or has failed, return the subflow socket.
50 * Otherwise return NULL.
52 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
54 if (!msk->subflow || READ_ONCE(msk->can_ack))
60 /* Returns end sequence number of the receiver's advertised window */
61 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
63 return atomic64_read(&msk->wnd_end);
66 static bool mptcp_is_tcpsk(struct sock *sk)
68 struct socket *sock = sk->sk_socket;
70 if (unlikely(sk->sk_prot == &tcp_prot)) {
71 /* we are being invoked after mptcp_accept() has
72 * accepted a non-mp-capable flow: sk is a tcp_sk,
75 * Hand the socket over to tcp so all further socket ops
78 sock->ops = &inet_stream_ops;
80 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
81 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
82 sock->ops = &inet6_stream_ops;
90 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
92 sock_owned_by_me((const struct sock *)msk);
94 if (likely(!__mptcp_check_fallback(msk)))
100 static int __mptcp_socket_create(struct mptcp_sock *msk)
102 struct mptcp_subflow_context *subflow;
103 struct sock *sk = (struct sock *)msk;
104 struct socket *ssock;
107 err = mptcp_subflow_create_socket(sk, &ssock);
111 msk->first = ssock->sk;
112 msk->subflow = ssock;
113 subflow = mptcp_subflow_ctx(ssock->sk);
114 list_add(&subflow->node, &msk->conn_list);
115 sock_hold(ssock->sk);
116 subflow->request_mptcp = 1;
118 /* accept() will wait on first subflow sk_wq, and we always wakes up
121 RCU_INIT_POINTER(msk->first->sk_wq, &sk->sk_socket->wq);
126 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
128 sk_drops_add(sk, skb);
132 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
133 struct sk_buff *from)
138 if (MPTCP_SKB_CB(from)->offset ||
139 !skb_try_coalesce(to, from, &fragstolen, &delta))
142 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
143 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
144 to->len, MPTCP_SKB_CB(from)->end_seq);
145 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
146 kfree_skb_partial(from, fragstolen);
147 atomic_add(delta, &sk->sk_rmem_alloc);
148 sk_mem_charge(sk, delta);
152 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
153 struct sk_buff *from)
155 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
158 return mptcp_try_coalesce((struct sock *)msk, to, from);
161 /* "inspired" by tcp_data_queue_ofo(), main differences:
163 * - don't cope with sacks
165 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
167 struct sock *sk = (struct sock *)msk;
168 struct rb_node **p, *parent;
169 u64 seq, end_seq, max_seq;
170 struct sk_buff *skb1;
173 seq = MPTCP_SKB_CB(skb)->map_seq;
174 end_seq = MPTCP_SKB_CB(skb)->end_seq;
175 space = tcp_space(sk);
176 max_seq = space > 0 ? space + msk->ack_seq : msk->ack_seq;
178 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
179 RB_EMPTY_ROOT(&msk->out_of_order_queue));
180 if (after64(seq, max_seq)) {
183 pr_debug("oow by %ld", (unsigned long)seq - (unsigned long)max_seq);
184 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
188 p = &msk->out_of_order_queue.rb_node;
189 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
190 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
191 rb_link_node(&skb->rbnode, NULL, p);
192 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
193 msk->ooo_last_skb = skb;
197 /* with 2 subflows, adding at end of ooo queue is quite likely
198 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
200 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
201 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
202 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
206 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
207 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
208 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
209 parent = &msk->ooo_last_skb->rbnode;
210 p = &parent->rb_right;
214 /* Find place to insert this segment. Handle overlaps on the way. */
218 skb1 = rb_to_skb(parent);
219 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
220 p = &parent->rb_left;
223 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
224 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
225 /* All the bits are present. Drop. */
227 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
230 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
234 * continue traversing
237 /* skb's seq == skb1's seq and skb covers skb1.
238 * Replace skb1 with skb.
240 rb_replace_node(&skb1->rbnode, &skb->rbnode,
241 &msk->out_of_order_queue);
242 mptcp_drop(sk, skb1);
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
246 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
250 p = &parent->rb_right;
254 /* Insert segment into RB tree. */
255 rb_link_node(&skb->rbnode, parent, p);
256 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
259 /* Remove other segments covered by skb. */
260 while ((skb1 = skb_rb_next(skb)) != NULL) {
261 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
263 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
264 mptcp_drop(sk, skb1);
265 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
267 /* If there is no skb after us, we are the last_skb ! */
269 msk->ooo_last_skb = skb;
273 skb_set_owner_r(skb, sk);
276 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
277 struct sk_buff *skb, unsigned int offset,
280 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
281 struct sock *sk = (struct sock *)msk;
282 struct sk_buff *tail;
284 __skb_unlink(skb, &ssk->sk_receive_queue);
289 /* try to fetch required memory from subflow */
290 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
291 if (ssk->sk_forward_alloc < skb->truesize)
293 __sk_mem_reclaim(ssk, skb->truesize);
294 if (!sk_rmem_schedule(sk, skb, skb->truesize))
298 /* the skb map_seq accounts for the skb offset:
299 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
302 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
303 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
304 MPTCP_SKB_CB(skb)->offset = offset;
306 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
308 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
309 tail = skb_peek_tail(&sk->sk_receive_queue);
310 if (tail && mptcp_try_coalesce(sk, tail, skb))
313 skb_set_owner_r(skb, sk);
314 __skb_queue_tail(&sk->sk_receive_queue, skb);
316 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
317 mptcp_data_queue_ofo(msk, skb);
321 /* old data, keep it simple and drop the whole pkt, sender
322 * will retransmit as needed, if needed.
324 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
330 static void mptcp_stop_timer(struct sock *sk)
332 struct inet_connection_sock *icsk = inet_csk(sk);
334 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
335 mptcp_sk(sk)->timer_ival = 0;
338 static void mptcp_close_wake_up(struct sock *sk)
340 if (sock_flag(sk, SOCK_DEAD))
343 sk->sk_state_change(sk);
344 if (sk->sk_shutdown == SHUTDOWN_MASK ||
345 sk->sk_state == TCP_CLOSE)
346 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
348 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
351 static void mptcp_check_data_fin_ack(struct sock *sk)
353 struct mptcp_sock *msk = mptcp_sk(sk);
355 if (__mptcp_check_fallback(msk))
358 /* Look for an acknowledged DATA_FIN */
359 if (((1 << sk->sk_state) &
360 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
361 msk->write_seq == atomic64_read(&msk->snd_una)) {
362 mptcp_stop_timer(sk);
364 WRITE_ONCE(msk->snd_data_fin_enable, 0);
366 switch (sk->sk_state) {
368 inet_sk_state_store(sk, TCP_FIN_WAIT2);
372 inet_sk_state_store(sk, TCP_CLOSE);
376 mptcp_close_wake_up(sk);
380 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
382 struct mptcp_sock *msk = mptcp_sk(sk);
384 if (READ_ONCE(msk->rcv_data_fin) &&
385 ((1 << sk->sk_state) &
386 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
387 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
389 if (msk->ack_seq == rcv_data_fin_seq) {
391 *seq = rcv_data_fin_seq;
400 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
402 long tout = ssk && inet_csk(ssk)->icsk_pending ?
403 inet_csk(ssk)->icsk_timeout - jiffies : 0;
406 tout = mptcp_sk(sk)->timer_ival;
407 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
410 static void mptcp_send_ack(struct mptcp_sock *msk)
412 struct mptcp_subflow_context *subflow;
414 mptcp_for_each_subflow(msk, subflow) {
415 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
423 static bool mptcp_check_data_fin(struct sock *sk)
425 struct mptcp_sock *msk = mptcp_sk(sk);
426 u64 rcv_data_fin_seq;
429 if (__mptcp_check_fallback(msk) || !msk->first)
432 /* Need to ack a DATA_FIN received from a peer while this side
433 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
434 * msk->rcv_data_fin was set when parsing the incoming options
435 * at the subflow level and the msk lock was not held, so this
436 * is the first opportunity to act on the DATA_FIN and change
439 * If we are caught up to the sequence number of the incoming
440 * DATA_FIN, send the DATA_ACK now and do state transition. If
441 * not caught up, do nothing and let the recv code send DATA_ACK
445 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
446 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
447 WRITE_ONCE(msk->rcv_data_fin, 0);
449 sk->sk_shutdown |= RCV_SHUTDOWN;
450 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
451 set_bit(MPTCP_DATA_READY, &msk->flags);
453 switch (sk->sk_state) {
454 case TCP_ESTABLISHED:
455 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
458 inet_sk_state_store(sk, TCP_CLOSING);
461 inet_sk_state_store(sk, TCP_CLOSE);
464 /* Other states not expected */
470 mptcp_set_timeout(sk, NULL);
472 mptcp_close_wake_up(sk);
477 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
481 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
482 struct sock *sk = (struct sock *)msk;
483 unsigned int moved = 0;
484 bool more_data_avail;
490 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
492 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
493 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
495 if (unlikely(ssk_rbuf > sk_rbuf)) {
496 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
501 pr_debug("msk=%p ssk=%p", msk, ssk);
503 old_copied_seq = tp->copied_seq;
505 u32 map_remaining, offset;
506 u32 seq = tp->copied_seq;
510 /* try to move as much data as available */
511 map_remaining = subflow->map_data_len -
512 mptcp_subflow_get_map_offset(subflow);
514 skb = skb_peek(&ssk->sk_receive_queue);
516 /* if no data is found, a racing workqueue/recvmsg
517 * already processed the new data, stop here or we
518 * can enter an infinite loop
525 if (__mptcp_check_fallback(msk)) {
526 /* if we are running under the workqueue, TCP could have
527 * collapsed skbs between dummy map creation and now
528 * be sure to adjust the size
530 map_remaining = skb->len;
531 subflow->map_data_len = skb->len;
534 offset = seq - TCP_SKB_CB(skb)->seq;
535 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
541 if (offset < skb->len) {
542 size_t len = skb->len - offset;
547 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
551 if (WARN_ON_ONCE(map_remaining < len))
555 sk_eat_skb(ssk, skb);
559 WRITE_ONCE(tp->copied_seq, seq);
560 more_data_avail = mptcp_subflow_data_available(ssk);
562 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
566 } while (more_data_avail);
569 if (tp->copied_seq != old_copied_seq)
570 tcp_cleanup_rbuf(ssk, 1);
575 static bool mptcp_ofo_queue(struct mptcp_sock *msk)
577 struct sock *sk = (struct sock *)msk;
578 struct sk_buff *skb, *tail;
583 p = rb_first(&msk->out_of_order_queue);
584 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
587 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
591 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
593 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
596 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
600 end_seq = MPTCP_SKB_CB(skb)->end_seq;
601 tail = skb_peek_tail(&sk->sk_receive_queue);
602 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
603 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
605 /* skip overlapping data, if any */
606 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
607 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
609 MPTCP_SKB_CB(skb)->offset += delta;
610 __skb_queue_tail(&sk->sk_receive_queue, skb);
612 msk->ack_seq = end_seq;
618 /* In most cases we will be able to lock the mptcp socket. If its already
619 * owned, we need to defer to the work queue to avoid ABBA deadlock.
621 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
623 struct sock *sk = (struct sock *)msk;
624 unsigned int moved = 0;
626 if (READ_ONCE(sk->sk_lock.owned))
629 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
632 /* must re-check after taking the lock */
633 if (!READ_ONCE(sk->sk_lock.owned)) {
634 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
635 mptcp_ofo_queue(msk);
637 /* If the moves have caught up with the DATA_FIN sequence number
638 * it's time to ack the DATA_FIN and change socket state, but
639 * this is not a good place to change state. Let the workqueue
642 if (mptcp_pending_data_fin(sk, NULL))
643 mptcp_schedule_work(sk);
646 spin_unlock_bh(&sk->sk_lock.slock);
651 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
653 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
654 struct mptcp_sock *msk = mptcp_sk(sk);
655 int sk_rbuf, ssk_rbuf;
658 /* move_skbs_to_msk below can legitly clear the data_avail flag,
659 * but we will need later to properly woke the reader, cache its
662 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
664 set_bit(MPTCP_DATA_READY, &msk->flags);
666 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
667 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
668 if (unlikely(ssk_rbuf > sk_rbuf))
671 /* over limit? can't append more skbs to msk */
672 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
675 if (move_skbs_to_msk(msk, ssk))
678 /* mptcp socket is owned, release_cb should retry */
679 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
680 &sk->sk_tsq_flags)) {
683 /* need to try again, its possible release_cb() has already
684 * been called after the test_and_set_bit() above.
686 move_skbs_to_msk(msk, ssk);
690 sk->sk_data_ready(sk);
693 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
695 if (likely(list_empty(&msk->join_list)))
698 spin_lock_bh(&msk->join_list_lock);
699 list_splice_tail_init(&msk->join_list, &msk->conn_list);
700 spin_unlock_bh(&msk->join_list_lock);
703 static bool mptcp_timer_pending(struct sock *sk)
705 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
708 static void mptcp_reset_timer(struct sock *sk)
710 struct inet_connection_sock *icsk = inet_csk(sk);
713 /* prevent rescheduling on close */
714 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
717 /* should never be called with mptcp level timer cleared */
718 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
719 if (WARN_ON_ONCE(!tout))
721 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
724 bool mptcp_schedule_work(struct sock *sk)
726 if (inet_sk_state_load(sk) != TCP_CLOSE &&
727 schedule_work(&mptcp_sk(sk)->work)) {
728 /* each subflow already holds a reference to the sk, and the
729 * workqueue is invoked by a subflow, so sk can't go away here.
737 void mptcp_data_acked(struct sock *sk)
739 mptcp_reset_timer(sk);
741 if ((test_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags) ||
742 mptcp_send_head(sk) ||
743 (inet_sk_state_load(sk) != TCP_ESTABLISHED)))
744 mptcp_schedule_work(sk);
747 void mptcp_subflow_eof(struct sock *sk)
749 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
750 mptcp_schedule_work(sk);
753 static void mptcp_check_for_eof(struct mptcp_sock *msk)
755 struct mptcp_subflow_context *subflow;
756 struct sock *sk = (struct sock *)msk;
759 mptcp_for_each_subflow(msk, subflow)
760 receivers += !subflow->rx_eof;
764 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
765 /* hopefully temporary hack: propagate shutdown status
766 * to msk, when all subflows agree on it
768 sk->sk_shutdown |= RCV_SHUTDOWN;
770 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
771 set_bit(MPTCP_DATA_READY, &msk->flags);
772 sk->sk_data_ready(sk);
775 switch (sk->sk_state) {
776 case TCP_ESTABLISHED:
777 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
780 inet_sk_state_store(sk, TCP_CLOSING);
783 inet_sk_state_store(sk, TCP_CLOSE);
788 mptcp_close_wake_up(sk);
791 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
793 const struct sock *sk = (const struct sock *)msk;
795 if (!msk->cached_ext)
796 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
798 return !!msk->cached_ext;
801 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
803 struct mptcp_subflow_context *subflow;
804 struct sock *sk = (struct sock *)msk;
806 sock_owned_by_me(sk);
808 mptcp_for_each_subflow(msk, subflow) {
809 if (subflow->data_avail)
810 return mptcp_subflow_tcp_sock(subflow);
816 static bool mptcp_skb_can_collapse_to(u64 write_seq,
817 const struct sk_buff *skb,
818 const struct mptcp_ext *mpext)
820 if (!tcp_skb_can_collapse_to(skb))
823 /* can collapse only if MPTCP level sequence is in order and this
824 * mapping has not been xmitted yet
826 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
830 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
831 const struct page_frag *pfrag,
832 const struct mptcp_data_frag *df)
834 return df && pfrag->page == df->page &&
835 pfrag->size - pfrag->offset > 0 &&
836 df->data_seq + df->data_len == msk->write_seq;
839 static void dfrag_uncharge(struct sock *sk, int len)
841 sk_mem_uncharge(sk, len);
842 sk_wmem_queued_add(sk, -len);
845 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
847 int len = dfrag->data_len + dfrag->overhead;
849 list_del(&dfrag->list);
850 dfrag_uncharge(sk, len);
851 put_page(dfrag->page);
854 static void mptcp_clean_una(struct sock *sk)
856 struct mptcp_sock *msk = mptcp_sk(sk);
857 struct mptcp_data_frag *dtmp, *dfrag;
858 bool cleaned = false;
861 /* on fallback we just need to ignore snd_una, as this is really
864 if (__mptcp_check_fallback(msk))
865 atomic64_set(&msk->snd_una, msk->snd_nxt);
867 snd_una = atomic64_read(&msk->snd_una);
869 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
870 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
873 if (WARN_ON_ONCE(dfrag == msk->first_pending))
875 dfrag_clear(sk, dfrag);
879 dfrag = mptcp_rtx_head(sk);
880 if (dfrag && after64(snd_una, dfrag->data_seq)) {
881 u64 delta = snd_una - dfrag->data_seq;
883 if (WARN_ON_ONCE(delta > dfrag->already_sent))
886 dfrag->data_seq += delta;
887 dfrag->offset += delta;
888 dfrag->data_len -= delta;
889 dfrag->already_sent -= delta;
891 dfrag_uncharge(sk, delta);
897 sk_mem_reclaim_partial(sk);
900 static void mptcp_clean_una_wakeup(struct sock *sk)
902 struct mptcp_sock *msk = mptcp_sk(sk);
906 /* Only wake up writers if a subflow is ready */
907 if (sk_stream_is_writeable(sk)) {
908 clear_bit(MPTCP_NOSPACE, &msk->flags);
909 sk_stream_write_space(sk);
913 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
916 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
918 struct mptcp_subflow_context *subflow;
919 struct mptcp_sock *msk = mptcp_sk(sk);
922 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
923 pfrag, sk->sk_allocation)))
926 sk_stream_moderate_sndbuf(sk);
927 mptcp_for_each_subflow(msk, subflow) {
928 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
931 tcp_enter_memory_pressure(ssk);
932 sk_stream_moderate_sndbuf(ssk);
938 static struct mptcp_data_frag *
939 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
942 int offset = ALIGN(orig_offset, sizeof(long));
943 struct mptcp_data_frag *dfrag;
945 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
947 dfrag->data_seq = msk->write_seq;
948 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
949 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
950 dfrag->already_sent = 0;
951 dfrag->page = pfrag->page;
956 struct mptcp_sendmsg_info {
964 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
967 u64 window_end = mptcp_wnd_end(msk);
969 if (__mptcp_check_fallback(msk))
972 if (!before64(data_seq + avail_size, window_end)) {
973 u64 allowed_size = window_end - data_seq;
975 return min_t(unsigned int, allowed_size, avail_size);
981 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
982 struct mptcp_data_frag *dfrag,
983 struct mptcp_sendmsg_info *info)
985 u64 data_seq = dfrag->data_seq + info->sent;
986 struct mptcp_sock *msk = mptcp_sk(sk);
987 bool zero_window_probe = false;
988 struct mptcp_ext *mpext = NULL;
989 struct sk_buff *skb, *tail;
990 bool can_collapse = false;
994 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
995 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
997 /* compute send limit */
998 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
999 avail_size = info->size_goal;
1000 skb = tcp_write_queue_tail(ssk);
1002 /* Limit the write to the size available in the
1003 * current skb, if any, so that we create at most a new skb.
1004 * Explicitly tells TCP internals to avoid collapsing on later
1005 * queue management operation, to avoid breaking the ext <->
1006 * SSN association set here
1008 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1009 can_collapse = (info->size_goal - skb->len > 0) &&
1010 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1012 TCP_SKB_CB(skb)->eor = 1;
1014 avail_size = info->size_goal - skb->len;
1017 /* Zero window and all data acked? Probe. */
1018 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1019 if (avail_size == 0) {
1020 if (skb || atomic64_read(&msk->snd_una) != msk->snd_nxt)
1022 zero_window_probe = true;
1023 data_seq = atomic64_read(&msk->snd_una) - 1;
1027 if (WARN_ON_ONCE(info->sent > info->limit ||
1028 info->limit > dfrag->data_len))
1031 ret = info->limit - info->sent;
1032 tail = tcp_build_frag(ssk, avail_size, info->flags, dfrag->page,
1033 dfrag->offset + info->sent, &ret);
1035 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1039 /* if the tail skb is still the cached one, collapsing really happened.
1042 WARN_ON_ONCE(!can_collapse);
1043 mpext->data_len += ret;
1044 WARN_ON_ONCE(zero_window_probe);
1048 mpext = __skb_ext_set(tail, SKB_EXT_MPTCP, msk->cached_ext);
1049 msk->cached_ext = NULL;
1051 memset(mpext, 0, sizeof(*mpext));
1052 mpext->data_seq = data_seq;
1053 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1054 mpext->data_len = ret;
1058 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1059 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1062 if (zero_window_probe) {
1063 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1066 tcp_push_pending_frames(ssk);
1069 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1073 static void mptcp_nospace(struct mptcp_sock *msk)
1075 struct mptcp_subflow_context *subflow;
1077 set_bit(MPTCP_NOSPACE, &msk->flags);
1078 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
1080 mptcp_for_each_subflow(msk, subflow) {
1081 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1082 bool ssk_writeable = sk_stream_is_writeable(ssk);
1083 struct socket *sock = READ_ONCE(ssk->sk_socket);
1085 if (ssk_writeable || !sock)
1088 /* enables ssk->write_space() callbacks */
1089 set_bit(SOCK_NOSPACE, &sock->flags);
1092 /* mptcp_data_acked() could run just before we set the NOSPACE bit,
1093 * so explicitly check for snd_una value
1095 mptcp_clean_una((struct sock *)msk);
1098 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1100 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1102 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
1103 if (subflow->request_join && !subflow->fully_established)
1106 /* only send if our side has not closed yet */
1107 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
1110 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1111 sizeof(struct tcphdr) - \
1112 MAX_TCP_OPTION_SPACE - \
1113 sizeof(struct ipv6hdr) - \
1114 sizeof(struct frag_hdr))
1116 struct subflow_send_info {
1121 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk,
1124 struct subflow_send_info send_info[2];
1125 struct mptcp_subflow_context *subflow;
1126 int i, nr_active = 0;
1131 sock_owned_by_me((struct sock *)msk);
1134 if (!mptcp_ext_cache_refill(msk))
1137 if (__mptcp_check_fallback(msk)) {
1140 *sndbuf = msk->first->sk_sndbuf;
1141 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1144 /* re-use last subflow, if the burst allow that */
1145 if (msk->last_snd && msk->snd_burst > 0 &&
1146 sk_stream_memory_free(msk->last_snd) &&
1147 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1148 mptcp_for_each_subflow(msk, subflow) {
1149 ssk = mptcp_subflow_tcp_sock(subflow);
1150 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1152 return msk->last_snd;
1155 /* pick the subflow with the lower wmem/wspace ratio */
1156 for (i = 0; i < 2; ++i) {
1157 send_info[i].ssk = NULL;
1158 send_info[i].ratio = -1;
1160 mptcp_for_each_subflow(msk, subflow) {
1161 ssk = mptcp_subflow_tcp_sock(subflow);
1162 if (!mptcp_subflow_active(subflow))
1165 nr_active += !subflow->backup;
1166 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1167 if (!sk_stream_memory_free(subflow->tcp_sock))
1170 pace = READ_ONCE(ssk->sk_pacing_rate);
1174 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1176 if (ratio < send_info[subflow->backup].ratio) {
1177 send_info[subflow->backup].ssk = ssk;
1178 send_info[subflow->backup].ratio = ratio;
1182 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1183 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1184 send_info[1].ssk, send_info[1].ratio);
1186 /* pick the best backup if no other subflow is active */
1188 send_info[0].ssk = send_info[1].ssk;
1190 if (send_info[0].ssk) {
1191 msk->last_snd = send_info[0].ssk;
1192 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1193 sk_stream_wspace(msk->last_snd));
1194 return msk->last_snd;
1199 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1200 struct mptcp_sendmsg_info *info)
1202 mptcp_set_timeout(sk, ssk);
1203 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1207 static void mptcp_push_pending(struct sock *sk, unsigned int flags)
1209 struct sock *prev_ssk = NULL, *ssk = NULL;
1210 struct mptcp_sock *msk = mptcp_sk(sk);
1211 struct mptcp_sendmsg_info info = {
1214 struct mptcp_data_frag *dfrag;
1215 int len, copied = 0;
1218 while ((dfrag = mptcp_send_head(sk))) {
1219 info.sent = dfrag->already_sent;
1220 info.limit = dfrag->data_len;
1221 len = dfrag->data_len - dfrag->already_sent;
1226 __mptcp_flush_join_list(msk);
1227 ssk = mptcp_subflow_get_send(msk, &sndbuf);
1229 /* do auto tuning */
1230 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1231 sndbuf > READ_ONCE(sk->sk_sndbuf))
1232 WRITE_ONCE(sk->sk_sndbuf, sndbuf);
1234 /* try to keep the subflow socket lock across
1235 * consecutive xmit on the same socket
1237 if (ssk != prev_ssk && prev_ssk)
1238 mptcp_push_release(sk, prev_ssk, &info);
1242 if (ssk != prev_ssk || !prev_ssk)
1245 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1247 mptcp_push_release(sk, ssk, &info);
1252 dfrag->already_sent += ret;
1253 msk->snd_nxt += ret;
1254 msk->snd_burst -= ret;
1258 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1261 /* at this point we held the socket lock for the last subflow we used */
1263 mptcp_push_release(sk, ssk, &info);
1267 /* start the timer, if it's not pending */
1268 if (!mptcp_timer_pending(sk))
1269 mptcp_reset_timer(sk);
1270 __mptcp_check_send_data_fin(sk);
1274 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1276 struct mptcp_sock *msk = mptcp_sk(sk);
1277 struct page_frag *pfrag;
1282 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1287 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1289 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1290 ret = sk_stream_wait_connect(sk, &timeo);
1295 pfrag = sk_page_frag(sk);
1296 mptcp_clean_una(sk);
1298 while (msg_data_left(msg)) {
1299 struct mptcp_data_frag *dfrag;
1300 int frag_truesize = 0;
1301 bool dfrag_collapsed;
1302 size_t psize, offset;
1304 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1309 /* reuse tail pfrag, if possible, or carve a new one from the
1312 dfrag = mptcp_pending_tail(sk);
1313 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1314 if (!dfrag_collapsed) {
1315 if (!sk_stream_memory_free(sk)) {
1316 mptcp_push_pending(sk, msg->msg_flags);
1317 if (!sk_stream_memory_free(sk))
1318 goto wait_for_memory;
1320 if (!mptcp_page_frag_refill(sk, pfrag))
1321 goto wait_for_memory;
1323 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1324 frag_truesize = dfrag->overhead;
1327 /* we do not bound vs wspace, to allow a single packet.
1328 * memory accounting will prevent execessive memory usage
1331 offset = dfrag->offset + dfrag->data_len;
1332 psize = pfrag->size - offset;
1333 psize = min_t(size_t, psize, msg_data_left(msg));
1334 if (!sk_wmem_schedule(sk, psize + frag_truesize))
1335 goto wait_for_memory;
1337 if (copy_page_from_iter(dfrag->page, offset, psize,
1338 &msg->msg_iter) != psize) {
1343 /* data successfully copied into the write queue */
1345 dfrag->data_len += psize;
1346 frag_truesize += psize;
1347 pfrag->offset += frag_truesize;
1348 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1350 /* charge data on mptcp pending queue to the msk socket
1351 * Note: we charge such data both to sk and ssk
1353 sk_wmem_queued_add(sk, frag_truesize);
1354 sk->sk_forward_alloc -= frag_truesize;
1355 if (!dfrag_collapsed) {
1356 get_page(dfrag->page);
1357 list_add_tail(&dfrag->list, &msk->rtx_queue);
1358 if (!msk->first_pending)
1359 WRITE_ONCE(msk->first_pending, dfrag);
1361 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1362 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1365 if (!mptcp_ext_cache_refill(msk))
1366 goto wait_for_memory;
1371 if (mptcp_timer_pending(sk))
1372 mptcp_reset_timer(sk);
1373 ret = sk_stream_wait_memory(sk, &timeo);
1379 mptcp_push_pending(sk, msg->msg_flags);
1383 return copied ? : ret;
1386 static void mptcp_wait_data(struct sock *sk, long *timeo)
1388 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1389 struct mptcp_sock *msk = mptcp_sk(sk);
1391 add_wait_queue(sk_sleep(sk), &wait);
1392 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1394 sk_wait_event(sk, timeo,
1395 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1397 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1398 remove_wait_queue(sk_sleep(sk), &wait);
1401 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1405 struct sock *sk = (struct sock *)msk;
1406 struct sk_buff *skb;
1409 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1410 u32 offset = MPTCP_SKB_CB(skb)->offset;
1411 u32 data_len = skb->len - offset;
1412 u32 count = min_t(size_t, len - copied, data_len);
1415 err = skb_copy_datagram_msg(skb, offset, msg, count);
1416 if (unlikely(err < 0)) {
1424 if (count < data_len) {
1425 MPTCP_SKB_CB(skb)->offset += count;
1429 __skb_unlink(skb, &sk->sk_receive_queue);
1439 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1441 * Only difference: Use highest rtt estimate of the subflows in use.
1443 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1445 struct mptcp_subflow_context *subflow;
1446 struct sock *sk = (struct sock *)msk;
1447 u32 time, advmss = 1;
1450 sock_owned_by_me(sk);
1455 msk->rcvq_space.copied += copied;
1457 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1458 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1460 rtt_us = msk->rcvq_space.rtt_us;
1461 if (rtt_us && time < (rtt_us >> 3))
1465 mptcp_for_each_subflow(msk, subflow) {
1466 const struct tcp_sock *tp;
1470 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1472 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1473 sf_advmss = READ_ONCE(tp->advmss);
1475 rtt_us = max(sf_rtt_us, rtt_us);
1476 advmss = max(sf_advmss, advmss);
1479 msk->rcvq_space.rtt_us = rtt_us;
1480 if (time < (rtt_us >> 3) || rtt_us == 0)
1483 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1486 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1487 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1491 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1493 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1495 do_div(grow, msk->rcvq_space.space);
1496 rcvwin += (grow << 1);
1498 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1499 while (tcp_win_from_space(sk, rcvmem) < advmss)
1502 do_div(rcvwin, advmss);
1503 rcvbuf = min_t(u64, rcvwin * rcvmem,
1504 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1506 if (rcvbuf > sk->sk_rcvbuf) {
1509 window_clamp = tcp_win_from_space(sk, rcvbuf);
1510 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1512 /* Make subflows follow along. If we do not do this, we
1513 * get drops at subflow level if skbs can't be moved to
1514 * the mptcp rx queue fast enough (announced rcv_win can
1515 * exceed ssk->sk_rcvbuf).
1517 mptcp_for_each_subflow(msk, subflow) {
1521 ssk = mptcp_subflow_tcp_sock(subflow);
1522 slow = lock_sock_fast(ssk);
1523 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1524 tcp_sk(ssk)->window_clamp = window_clamp;
1525 tcp_cleanup_rbuf(ssk, 1);
1526 unlock_sock_fast(ssk, slow);
1531 msk->rcvq_space.space = msk->rcvq_space.copied;
1533 msk->rcvq_space.copied = 0;
1534 msk->rcvq_space.time = mstamp;
1537 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1539 unsigned int moved = 0;
1542 /* avoid looping forever below on racing close */
1543 if (((struct sock *)msk)->sk_state == TCP_CLOSE)
1546 __mptcp_flush_join_list(msk);
1548 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1554 slowpath = lock_sock_fast(ssk);
1555 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1556 unlock_sock_fast(ssk, slowpath);
1559 if (mptcp_ofo_queue(msk) || moved > 0) {
1560 if (!mptcp_check_data_fin((struct sock *)msk))
1561 mptcp_send_ack(msk);
1567 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1568 int nonblock, int flags, int *addr_len)
1570 struct mptcp_sock *msk = mptcp_sk(sk);
1575 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1579 timeo = sock_rcvtimeo(sk, nonblock);
1581 len = min_t(size_t, len, INT_MAX);
1582 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1583 __mptcp_flush_join_list(msk);
1585 while (len > (size_t)copied) {
1588 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1589 if (unlikely(bytes_read < 0)) {
1591 copied = bytes_read;
1595 copied += bytes_read;
1597 if (skb_queue_empty(&sk->sk_receive_queue) &&
1598 __mptcp_move_skbs(msk))
1601 /* only the master socket status is relevant here. The exit
1602 * conditions mirror closely tcp_recvmsg()
1604 if (copied >= target)
1609 sk->sk_state == TCP_CLOSE ||
1610 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1612 signal_pending(current))
1616 copied = sock_error(sk);
1620 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1621 mptcp_check_for_eof(msk);
1623 if (sk->sk_shutdown & RCV_SHUTDOWN)
1626 if (sk->sk_state == TCP_CLOSE) {
1636 if (signal_pending(current)) {
1637 copied = sock_intr_errno(timeo);
1642 pr_debug("block timeout %ld", timeo);
1643 mptcp_wait_data(sk, &timeo);
1646 if (skb_queue_empty(&sk->sk_receive_queue)) {
1647 /* entire backlog drained, clear DATA_READY. */
1648 clear_bit(MPTCP_DATA_READY, &msk->flags);
1650 /* .. race-breaker: ssk might have gotten new data
1651 * after last __mptcp_move_skbs() returned false.
1653 if (unlikely(__mptcp_move_skbs(msk)))
1654 set_bit(MPTCP_DATA_READY, &msk->flags);
1655 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1656 /* data to read but mptcp_wait_data() cleared DATA_READY */
1657 set_bit(MPTCP_DATA_READY, &msk->flags);
1660 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
1661 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
1662 skb_queue_empty(&sk->sk_receive_queue), copied);
1663 mptcp_rcv_space_adjust(msk, copied);
1669 static void mptcp_retransmit_handler(struct sock *sk)
1671 struct mptcp_sock *msk = mptcp_sk(sk);
1673 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->snd_nxt)) {
1674 mptcp_stop_timer(sk);
1676 set_bit(MPTCP_WORK_RTX, &msk->flags);
1677 mptcp_schedule_work(sk);
1681 static void mptcp_retransmit_timer(struct timer_list *t)
1683 struct inet_connection_sock *icsk = from_timer(icsk, t,
1684 icsk_retransmit_timer);
1685 struct sock *sk = &icsk->icsk_inet.sk;
1688 if (!sock_owned_by_user(sk)) {
1689 mptcp_retransmit_handler(sk);
1691 /* delegate our work to tcp_release_cb() */
1692 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1700 static void mptcp_timeout_timer(struct timer_list *t)
1702 struct sock *sk = from_timer(sk, t, sk_timer);
1704 mptcp_schedule_work(sk);
1707 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1710 * A backup subflow is returned only if that is the only kind available.
1712 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1714 struct mptcp_subflow_context *subflow;
1715 struct sock *backup = NULL;
1717 sock_owned_by_me((const struct sock *)msk);
1719 if (__mptcp_check_fallback(msk))
1722 mptcp_for_each_subflow(msk, subflow) {
1723 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1725 if (!mptcp_subflow_active(subflow))
1728 /* still data outstanding at TCP level? Don't retransmit. */
1729 if (!tcp_write_queue_empty(ssk)) {
1730 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
1735 if (subflow->backup) {
1747 /* subflow sockets can be either outgoing (connect) or incoming
1750 * Outgoing subflows use in-kernel sockets.
1751 * Incoming subflows do not have their own 'struct socket' allocated,
1752 * so we need to use tcp_close() after detaching them from the mptcp
1755 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1756 struct mptcp_subflow_context *subflow)
1758 bool dispose_socket = false;
1759 struct socket *sock;
1761 list_del(&subflow->node);
1765 /* if we are invoked by the msk cleanup code, the subflow is
1768 sock = ssk->sk_socket;
1770 dispose_socket = sock != sk->sk_socket;
1774 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
1775 * the ssk has been already destroyed, we just need to release the
1776 * reference owned by msk;
1778 if (!inet_csk(ssk)->icsk_ulp_ops) {
1779 kfree_rcu(subflow, rcu);
1781 /* otherwise ask tcp do dispose of ssk and subflow ctx */
1782 subflow->disposable = 1;
1783 __tcp_close(ssk, 0);
1785 /* close acquired an extra ref */
1790 iput(SOCK_INODE(sock));
1795 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1800 static void pm_work(struct mptcp_sock *msk)
1802 struct mptcp_pm_data *pm = &msk->pm;
1804 spin_lock_bh(&msk->pm.lock);
1806 pr_debug("msk=%p status=%x", msk, pm->status);
1807 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1808 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1809 mptcp_pm_nl_add_addr_received(msk);
1811 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
1812 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
1813 mptcp_pm_nl_rm_addr_received(msk);
1815 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1816 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1817 mptcp_pm_nl_fully_established(msk);
1819 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1820 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1821 mptcp_pm_nl_subflow_established(msk);
1824 spin_unlock_bh(&msk->pm.lock);
1827 static void __mptcp_close_subflow(struct mptcp_sock *msk)
1829 struct mptcp_subflow_context *subflow, *tmp;
1831 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
1832 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1834 if (inet_sk_state_load(ssk) != TCP_CLOSE)
1837 __mptcp_close_ssk((struct sock *)msk, ssk, subflow);
1841 static bool mptcp_check_close_timeout(const struct sock *sk)
1843 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
1844 struct mptcp_subflow_context *subflow;
1846 if (delta >= TCP_TIMEWAIT_LEN)
1849 /* if all subflows are in closed status don't bother with additional
1852 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
1853 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
1860 static void mptcp_worker(struct work_struct *work)
1862 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1863 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1864 struct mptcp_sendmsg_info info = {};
1865 struct mptcp_data_frag *dfrag;
1870 state = sk->sk_state;
1871 if (unlikely(state == TCP_CLOSE))
1874 mptcp_clean_una_wakeup(sk);
1875 mptcp_check_data_fin_ack(sk);
1876 __mptcp_flush_join_list(msk);
1877 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
1878 __mptcp_close_subflow(msk);
1880 __mptcp_move_skbs(msk);
1881 if (mptcp_send_head(sk))
1882 mptcp_push_pending(sk, 0);
1887 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1888 mptcp_check_for_eof(msk);
1890 mptcp_check_data_fin(sk);
1892 /* if the msk data is completely acked, or the socket timedout,
1893 * there is no point in keeping around an orphaned sk
1895 if (sock_flag(sk, SOCK_DEAD) &&
1896 (mptcp_check_close_timeout(sk) ||
1897 (state != sk->sk_state &&
1898 ((1 << inet_sk_state_load(sk)) & (TCPF_CLOSE | TCPF_FIN_WAIT2))))) {
1899 inet_sk_state_store(sk, TCP_CLOSE);
1900 __mptcp_destroy_sock(sk);
1904 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1907 dfrag = mptcp_rtx_head(sk);
1911 if (!mptcp_ext_cache_refill(msk))
1914 ssk = mptcp_subflow_get_retrans(msk);
1920 /* limit retransmission to the bytes already sent on some subflows */
1922 info.limit = dfrag->already_sent;
1923 while (info.sent < dfrag->already_sent) {
1924 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1928 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1932 if (!mptcp_ext_cache_refill(msk))
1936 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1939 mptcp_set_timeout(sk, ssk);
1943 if (!mptcp_timer_pending(sk))
1944 mptcp_reset_timer(sk);
1951 static int __mptcp_init_sock(struct sock *sk)
1953 struct mptcp_sock *msk = mptcp_sk(sk);
1955 spin_lock_init(&msk->join_list_lock);
1957 INIT_LIST_HEAD(&msk->conn_list);
1958 INIT_LIST_HEAD(&msk->join_list);
1959 INIT_LIST_HEAD(&msk->rtx_queue);
1960 INIT_WORK(&msk->work, mptcp_worker);
1961 msk->out_of_order_queue = RB_ROOT;
1962 msk->first_pending = NULL;
1965 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1967 mptcp_pm_data_init(msk);
1969 /* re-use the csk retrans timer for MPTCP-level retrans */
1970 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1971 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
1975 static int mptcp_init_sock(struct sock *sk)
1977 struct net *net = sock_net(sk);
1980 ret = __mptcp_init_sock(sk);
1984 if (!mptcp_is_enabled(net))
1985 return -ENOPROTOOPT;
1987 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1990 ret = __mptcp_socket_create(mptcp_sk(sk));
1994 sk_sockets_allocated_inc(sk);
1995 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1996 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2001 static void __mptcp_clear_xmit(struct sock *sk)
2003 struct mptcp_sock *msk = mptcp_sk(sk);
2004 struct mptcp_data_frag *dtmp, *dfrag;
2006 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2008 WRITE_ONCE(msk->first_pending, NULL);
2009 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2010 dfrag_clear(sk, dfrag);
2013 static void mptcp_cancel_work(struct sock *sk)
2015 struct mptcp_sock *msk = mptcp_sk(sk);
2017 if (cancel_work_sync(&msk->work))
2021 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2025 switch (ssk->sk_state) {
2027 if (!(how & RCV_SHUTDOWN))
2031 tcp_disconnect(ssk, O_NONBLOCK);
2034 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2035 pr_debug("Fallback");
2036 ssk->sk_shutdown |= how;
2037 tcp_shutdown(ssk, how);
2039 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2040 mptcp_set_timeout(sk, ssk);
2049 static const unsigned char new_state[16] = {
2050 /* current state: new state: action: */
2051 [0 /* (Invalid) */] = TCP_CLOSE,
2052 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2053 [TCP_SYN_SENT] = TCP_CLOSE,
2054 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2055 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2056 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2057 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2058 [TCP_CLOSE] = TCP_CLOSE,
2059 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2060 [TCP_LAST_ACK] = TCP_LAST_ACK,
2061 [TCP_LISTEN] = TCP_CLOSE,
2062 [TCP_CLOSING] = TCP_CLOSING,
2063 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2066 static int mptcp_close_state(struct sock *sk)
2068 int next = (int)new_state[sk->sk_state];
2069 int ns = next & TCP_STATE_MASK;
2071 inet_sk_state_store(sk, ns);
2073 return next & TCP_ACTION_FIN;
2076 static void __mptcp_check_send_data_fin(struct sock *sk)
2078 struct mptcp_subflow_context *subflow;
2079 struct mptcp_sock *msk = mptcp_sk(sk);
2081 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2082 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2083 msk->snd_nxt, msk->write_seq);
2085 /* we still need to enqueue subflows or not really shutting down,
2088 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2089 mptcp_send_head(sk))
2092 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2094 /* fallback socket will not get data_fin/ack, can move to the next
2097 if (__mptcp_check_fallback(msk)) {
2098 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2099 inet_sk_state_store(sk, TCP_CLOSE);
2100 mptcp_close_wake_up(sk);
2101 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2102 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2106 __mptcp_flush_join_list(msk);
2107 mptcp_for_each_subflow(msk, subflow) {
2108 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2110 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2114 static void __mptcp_wr_shutdown(struct sock *sk)
2116 struct mptcp_sock *msk = mptcp_sk(sk);
2118 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2119 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2120 !!mptcp_send_head(sk));
2122 /* will be ignored by fallback sockets */
2123 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2124 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2126 __mptcp_check_send_data_fin(sk);
2129 static void __mptcp_destroy_sock(struct sock *sk)
2131 struct mptcp_subflow_context *subflow, *tmp;
2132 struct mptcp_sock *msk = mptcp_sk(sk);
2133 LIST_HEAD(conn_list);
2135 pr_debug("msk=%p", msk);
2137 /* be sure to always acquire the join list lock, to sync vs
2138 * mptcp_finish_join().
2140 spin_lock_bh(&msk->join_list_lock);
2141 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2142 spin_unlock_bh(&msk->join_list_lock);
2143 list_splice_init(&msk->conn_list, &conn_list);
2145 __mptcp_clear_xmit(sk);
2146 sk_stop_timer(sk, &sk->sk_timer);
2149 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2150 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2151 __mptcp_close_ssk(sk, ssk, subflow);
2154 sk->sk_prot->destroy(sk);
2156 sk_stream_kill_queues(sk);
2157 xfrm_sk_free_policy(sk);
2158 sk_refcnt_debug_release(sk);
2162 static void mptcp_close(struct sock *sk, long timeout)
2164 struct mptcp_subflow_context *subflow;
2165 bool do_cancel_work = false;
2168 sk->sk_shutdown = SHUTDOWN_MASK;
2170 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2171 inet_sk_state_store(sk, TCP_CLOSE);
2175 if (mptcp_close_state(sk))
2176 __mptcp_wr_shutdown(sk);
2178 sk_stream_wait_close(sk, timeout);
2181 /* orphan all the subflows */
2182 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2183 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2184 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2185 bool slow, dispose_socket;
2186 struct socket *sock;
2188 slow = lock_sock_fast(ssk);
2189 sock = ssk->sk_socket;
2190 dispose_socket = sock && sock != sk->sk_socket;
2192 unlock_sock_fast(ssk, slow);
2194 /* for the outgoing subflows we additionally need to free
2195 * the associated socket
2198 iput(SOCK_INODE(sock));
2203 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2204 if (sk->sk_state == TCP_CLOSE) {
2205 __mptcp_destroy_sock(sk);
2206 do_cancel_work = true;
2208 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2212 mptcp_cancel_work(sk);
2216 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2218 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2219 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2220 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2222 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2223 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2226 msk6->saddr = ssk6->saddr;
2227 msk6->flow_label = ssk6->flow_label;
2231 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2232 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2233 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2234 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2235 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2236 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2239 static int mptcp_disconnect(struct sock *sk, int flags)
2241 /* Should never be called.
2242 * inet_stream_connect() calls ->disconnect, but that
2243 * refers to the subflow socket, not the mptcp one.
2249 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2250 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2252 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2254 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2258 struct sock *mptcp_sk_clone(const struct sock *sk,
2259 const struct mptcp_options_received *mp_opt,
2260 struct request_sock *req)
2262 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2263 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2264 struct mptcp_sock *msk;
2270 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2271 if (nsk->sk_family == AF_INET6)
2272 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2275 __mptcp_init_sock(nsk);
2277 msk = mptcp_sk(nsk);
2278 msk->local_key = subflow_req->local_key;
2279 msk->token = subflow_req->token;
2280 msk->subflow = NULL;
2281 WRITE_ONCE(msk->fully_established, false);
2283 msk->write_seq = subflow_req->idsn + 1;
2284 msk->snd_nxt = msk->write_seq;
2285 atomic64_set(&msk->snd_una, msk->write_seq);
2286 atomic64_set(&msk->wnd_end, msk->snd_nxt + req->rsk_rcv_wnd);
2288 if (mp_opt->mp_capable) {
2289 msk->can_ack = true;
2290 msk->remote_key = mp_opt->sndr_key;
2291 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2293 WRITE_ONCE(msk->ack_seq, ack_seq);
2296 sock_reset_flag(nsk, SOCK_RCU_FREE);
2297 /* will be fully established after successful MPC subflow creation */
2298 inet_sk_state_store(nsk, TCP_SYN_RECV);
2299 bh_unlock_sock(nsk);
2301 /* keep a single reference */
2306 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2308 const struct tcp_sock *tp = tcp_sk(ssk);
2310 msk->rcvq_space.copied = 0;
2311 msk->rcvq_space.rtt_us = 0;
2313 msk->rcvq_space.time = tp->tcp_mstamp;
2315 /* initial rcv_space offering made to peer */
2316 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2317 TCP_INIT_CWND * tp->advmss);
2318 if (msk->rcvq_space.space == 0)
2319 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2321 atomic64_set(&msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2324 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2327 struct mptcp_sock *msk = mptcp_sk(sk);
2328 struct socket *listener;
2331 listener = __mptcp_nmpc_socket(msk);
2332 if (WARN_ON_ONCE(!listener)) {
2337 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2338 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2342 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2343 if (sk_is_mptcp(newsk)) {
2344 struct mptcp_subflow_context *subflow;
2345 struct sock *new_mptcp_sock;
2347 subflow = mptcp_subflow_ctx(newsk);
2348 new_mptcp_sock = subflow->conn;
2350 /* is_mptcp should be false if subflow->conn is missing, see
2351 * subflow_syn_recv_sock()
2353 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2354 tcp_sk(newsk)->is_mptcp = 0;
2358 /* acquire the 2nd reference for the owning socket */
2359 sock_hold(new_mptcp_sock);
2360 newsk = new_mptcp_sock;
2361 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2363 MPTCP_INC_STATS(sock_net(sk),
2364 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2370 void mptcp_destroy_common(struct mptcp_sock *msk)
2372 skb_rbtree_purge(&msk->out_of_order_queue);
2373 mptcp_token_destroy(msk);
2374 mptcp_pm_free_anno_list(msk);
2377 static void mptcp_destroy(struct sock *sk)
2379 struct mptcp_sock *msk = mptcp_sk(sk);
2381 if (msk->cached_ext)
2382 __skb_ext_put(msk->cached_ext);
2384 mptcp_destroy_common(msk);
2385 sk_sockets_allocated_dec(sk);
2388 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2389 sockptr_t optval, unsigned int optlen)
2391 struct sock *sk = (struct sock *)msk;
2392 struct socket *ssock;
2399 ssock = __mptcp_nmpc_socket(msk);
2405 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2407 if (optname == SO_REUSEPORT)
2408 sk->sk_reuseport = ssock->sk->sk_reuseport;
2409 else if (optname == SO_REUSEADDR)
2410 sk->sk_reuse = ssock->sk->sk_reuse;
2416 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2419 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2420 sockptr_t optval, unsigned int optlen)
2422 struct sock *sk = (struct sock *)msk;
2423 int ret = -EOPNOTSUPP;
2424 struct socket *ssock;
2429 ssock = __mptcp_nmpc_socket(msk);
2435 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2437 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2446 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2447 sockptr_t optval, unsigned int optlen)
2449 struct mptcp_sock *msk = mptcp_sk(sk);
2452 pr_debug("msk=%p", msk);
2454 if (level == SOL_SOCKET)
2455 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2457 /* @@ the meaning of setsockopt() when the socket is connected and
2458 * there are multiple subflows is not yet defined. It is up to the
2459 * MPTCP-level socket to configure the subflows until the subflow
2460 * is in TCP fallback, when TCP socket options are passed through
2461 * to the one remaining subflow.
2464 ssk = __mptcp_tcp_fallback(msk);
2467 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2469 if (level == SOL_IPV6)
2470 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2475 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2476 char __user *optval, int __user *option)
2478 struct mptcp_sock *msk = mptcp_sk(sk);
2481 pr_debug("msk=%p", msk);
2483 /* @@ the meaning of setsockopt() when the socket is connected and
2484 * there are multiple subflows is not yet defined. It is up to the
2485 * MPTCP-level socket to configure the subflows until the subflow
2486 * is in TCP fallback, when socket options are passed through
2487 * to the one remaining subflow.
2490 ssk = __mptcp_tcp_fallback(msk);
2493 return tcp_getsockopt(ssk, level, optname, optval, option);
2498 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
2499 TCPF_WRITE_TIMER_DEFERRED)
2501 /* this is very alike tcp_release_cb() but we must handle differently a
2502 * different set of events
2504 static void mptcp_release_cb(struct sock *sk)
2506 unsigned long flags, nflags;
2509 flags = sk->sk_tsq_flags;
2510 if (!(flags & MPTCP_DEFERRED_ALL))
2512 nflags = flags & ~MPTCP_DEFERRED_ALL;
2513 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2515 sock_release_ownership(sk);
2517 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
2518 struct mptcp_sock *msk = mptcp_sk(sk);
2521 ssk = mptcp_subflow_recv_lookup(msk);
2522 if (!ssk || sk->sk_state == TCP_CLOSE ||
2523 !schedule_work(&msk->work))
2527 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2528 mptcp_retransmit_handler(sk);
2533 static int mptcp_hash(struct sock *sk)
2535 /* should never be called,
2536 * we hash the TCP subflows not the master socket
2542 static void mptcp_unhash(struct sock *sk)
2544 /* called from sk_common_release(), but nothing to do here */
2547 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2549 struct mptcp_sock *msk = mptcp_sk(sk);
2550 struct socket *ssock;
2552 ssock = __mptcp_nmpc_socket(msk);
2553 pr_debug("msk=%p, subflow=%p", msk, ssock);
2554 if (WARN_ON_ONCE(!ssock))
2557 return inet_csk_get_port(ssock->sk, snum);
2560 void mptcp_finish_connect(struct sock *ssk)
2562 struct mptcp_subflow_context *subflow;
2563 struct mptcp_sock *msk;
2567 subflow = mptcp_subflow_ctx(ssk);
2571 pr_debug("msk=%p, token=%u", sk, subflow->token);
2573 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2575 subflow->map_seq = ack_seq;
2576 subflow->map_subflow_seq = 1;
2578 /* the socket is not connected yet, no msk/subflow ops can access/race
2579 * accessing the field below
2581 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2582 WRITE_ONCE(msk->local_key, subflow->local_key);
2583 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2584 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2585 WRITE_ONCE(msk->ack_seq, ack_seq);
2586 WRITE_ONCE(msk->can_ack, 1);
2587 atomic64_set(&msk->snd_una, msk->write_seq);
2589 mptcp_pm_new_connection(msk, 0);
2591 mptcp_rcv_space_init(msk, ssk);
2594 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2596 write_lock_bh(&sk->sk_callback_lock);
2597 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2598 sk_set_socket(sk, parent);
2599 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2600 write_unlock_bh(&sk->sk_callback_lock);
2603 bool mptcp_finish_join(struct sock *ssk)
2605 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2606 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2607 struct sock *parent = (void *)msk;
2608 struct socket *parent_sock;
2611 pr_debug("msk=%p, subflow=%p", msk, subflow);
2613 /* mptcp socket already closing? */
2614 if (!mptcp_is_fully_established(parent))
2617 if (!msk->pm.server_side)
2620 if (!mptcp_pm_allow_new_subflow(msk))
2623 /* active connections are already on conn_list, and we can't acquire
2625 * use the join list lock as synchronization point and double-check
2626 * msk status to avoid racing with __mptcp_destroy_sock()
2628 spin_lock_bh(&msk->join_list_lock);
2629 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2630 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
2631 list_add_tail(&subflow->node, &msk->join_list);
2634 spin_unlock_bh(&msk->join_list_lock);
2638 /* attach to msk socket only after we are sure he will deal with us
2641 parent_sock = READ_ONCE(parent->sk_socket);
2642 if (parent_sock && !ssk->sk_socket)
2643 mptcp_sock_graft(ssk, parent_sock);
2644 subflow->map_seq = READ_ONCE(msk->ack_seq);
2648 static struct proto mptcp_prot = {
2650 .owner = THIS_MODULE,
2651 .init = mptcp_init_sock,
2652 .disconnect = mptcp_disconnect,
2653 .close = mptcp_close,
2654 .accept = mptcp_accept,
2655 .setsockopt = mptcp_setsockopt,
2656 .getsockopt = mptcp_getsockopt,
2657 .shutdown = tcp_shutdown,
2658 .destroy = mptcp_destroy,
2659 .sendmsg = mptcp_sendmsg,
2660 .recvmsg = mptcp_recvmsg,
2661 .release_cb = mptcp_release_cb,
2663 .unhash = mptcp_unhash,
2664 .get_port = mptcp_get_port,
2665 .sockets_allocated = &mptcp_sockets_allocated,
2666 .memory_allocated = &tcp_memory_allocated,
2667 .memory_pressure = &tcp_memory_pressure,
2668 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2669 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
2670 .sysctl_mem = sysctl_tcp_mem,
2671 .obj_size = sizeof(struct mptcp_sock),
2672 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2673 .no_autobind = true,
2676 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2678 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2679 struct socket *ssock;
2682 lock_sock(sock->sk);
2683 ssock = __mptcp_nmpc_socket(msk);
2689 err = ssock->ops->bind(ssock, uaddr, addr_len);
2691 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2694 release_sock(sock->sk);
2698 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2699 struct mptcp_subflow_context *subflow)
2701 subflow->request_mptcp = 0;
2702 __mptcp_do_fallback(msk);
2705 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2706 int addr_len, int flags)
2708 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2709 struct mptcp_subflow_context *subflow;
2710 struct socket *ssock;
2713 lock_sock(sock->sk);
2714 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2715 /* pending connection or invalid state, let existing subflow
2718 ssock = msk->subflow;
2722 ssock = __mptcp_nmpc_socket(msk);
2728 mptcp_token_destroy(msk);
2729 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2730 subflow = mptcp_subflow_ctx(ssock->sk);
2731 #ifdef CONFIG_TCP_MD5SIG
2732 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2735 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2736 mptcp_subflow_early_fallback(msk, subflow);
2738 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2739 mptcp_subflow_early_fallback(msk, subflow);
2742 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2743 sock->state = ssock->state;
2745 /* on successful connect, the msk state will be moved to established by
2746 * subflow_finish_connect()
2748 if (!err || err == -EINPROGRESS)
2749 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2751 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2754 release_sock(sock->sk);
2758 static int mptcp_listen(struct socket *sock, int backlog)
2760 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2761 struct socket *ssock;
2764 pr_debug("msk=%p", msk);
2766 lock_sock(sock->sk);
2767 ssock = __mptcp_nmpc_socket(msk);
2773 mptcp_token_destroy(msk);
2774 inet_sk_state_store(sock->sk, TCP_LISTEN);
2775 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2777 err = ssock->ops->listen(ssock, backlog);
2778 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2780 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2783 release_sock(sock->sk);
2787 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2788 int flags, bool kern)
2790 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2791 struct socket *ssock;
2794 pr_debug("msk=%p", msk);
2796 lock_sock(sock->sk);
2797 if (sock->sk->sk_state != TCP_LISTEN)
2800 ssock = __mptcp_nmpc_socket(msk);
2804 clear_bit(MPTCP_DATA_READY, &msk->flags);
2805 sock_hold(ssock->sk);
2806 release_sock(sock->sk);
2808 err = ssock->ops->accept(sock, newsock, flags, kern);
2809 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2810 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2811 struct mptcp_subflow_context *subflow;
2812 struct sock *newsk = newsock->sk;
2815 slowpath = lock_sock_fast(newsk);
2816 mptcp_copy_inaddrs(newsk, msk->first);
2817 mptcp_rcv_space_init(msk, msk->first);
2819 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2820 * This is needed so NOSPACE flag can be set from tcp stack.
2822 __mptcp_flush_join_list(msk);
2823 mptcp_for_each_subflow(msk, subflow) {
2824 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2826 if (!ssk->sk_socket)
2827 mptcp_sock_graft(ssk, newsock);
2829 unlock_sock_fast(newsk, slowpath);
2832 if (inet_csk_listen_poll(ssock->sk))
2833 set_bit(MPTCP_DATA_READY, &msk->flags);
2834 sock_put(ssock->sk);
2838 release_sock(sock->sk);
2842 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2844 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2848 static bool __mptcp_check_writeable(struct mptcp_sock *msk)
2850 struct sock *sk = (struct sock *)msk;
2851 bool mptcp_writable;
2853 mptcp_clean_una(sk);
2854 mptcp_writable = sk_stream_is_writeable(sk);
2855 if (!mptcp_writable)
2858 return mptcp_writable;
2861 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
2863 struct sock *sk = (struct sock *)msk;
2867 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
2870 if (sk_stream_is_writeable(sk))
2871 return EPOLLOUT | EPOLLWRNORM;
2873 slow = lock_sock_fast(sk);
2874 if (__mptcp_check_writeable(msk))
2875 ret = EPOLLOUT | EPOLLWRNORM;
2877 unlock_sock_fast(sk, slow);
2881 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2882 struct poll_table_struct *wait)
2884 struct sock *sk = sock->sk;
2885 struct mptcp_sock *msk;
2890 sock_poll_wait(file, sock, wait);
2892 state = inet_sk_state_load(sk);
2893 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
2894 if (state == TCP_LISTEN)
2895 return mptcp_check_readable(msk);
2897 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2898 mask |= mptcp_check_readable(msk);
2899 mask |= mptcp_check_writeable(msk);
2901 if (sk->sk_shutdown & RCV_SHUTDOWN)
2902 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2907 static int mptcp_shutdown(struct socket *sock, int how)
2909 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2910 struct sock *sk = sock->sk;
2913 pr_debug("sk=%p, how=%d", msk, how);
2918 if ((how & ~SHUTDOWN_MASK) || !how) {
2923 if (sock->state == SS_CONNECTING) {
2924 if ((1 << sk->sk_state) &
2925 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2926 sock->state = SS_DISCONNECTING;
2928 sock->state = SS_CONNECTED;
2931 sk->sk_shutdown |= how;
2932 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
2933 __mptcp_wr_shutdown(sk);
2935 /* Wake up anyone sleeping in poll. */
2936 sk->sk_state_change(sk);
2944 static const struct proto_ops mptcp_stream_ops = {
2946 .owner = THIS_MODULE,
2947 .release = inet_release,
2949 .connect = mptcp_stream_connect,
2950 .socketpair = sock_no_socketpair,
2951 .accept = mptcp_stream_accept,
2952 .getname = inet_getname,
2954 .ioctl = inet_ioctl,
2955 .gettstamp = sock_gettstamp,
2956 .listen = mptcp_listen,
2957 .shutdown = mptcp_shutdown,
2958 .setsockopt = sock_common_setsockopt,
2959 .getsockopt = sock_common_getsockopt,
2960 .sendmsg = inet_sendmsg,
2961 .recvmsg = inet_recvmsg,
2962 .mmap = sock_no_mmap,
2963 .sendpage = inet_sendpage,
2966 static struct inet_protosw mptcp_protosw = {
2967 .type = SOCK_STREAM,
2968 .protocol = IPPROTO_MPTCP,
2969 .prot = &mptcp_prot,
2970 .ops = &mptcp_stream_ops,
2971 .flags = INET_PROTOSW_ICSK,
2974 void __init mptcp_proto_init(void)
2976 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2978 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2979 panic("Failed to allocate MPTCP pcpu counter\n");
2981 mptcp_subflow_init();
2985 if (proto_register(&mptcp_prot, 1) != 0)
2986 panic("Failed to register MPTCP proto.\n");
2988 inet_register_protosw(&mptcp_protosw);
2990 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2993 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2994 static const struct proto_ops mptcp_v6_stream_ops = {
2996 .owner = THIS_MODULE,
2997 .release = inet6_release,
2999 .connect = mptcp_stream_connect,
3000 .socketpair = sock_no_socketpair,
3001 .accept = mptcp_stream_accept,
3002 .getname = inet6_getname,
3004 .ioctl = inet6_ioctl,
3005 .gettstamp = sock_gettstamp,
3006 .listen = mptcp_listen,
3007 .shutdown = mptcp_shutdown,
3008 .setsockopt = sock_common_setsockopt,
3009 .getsockopt = sock_common_getsockopt,
3010 .sendmsg = inet6_sendmsg,
3011 .recvmsg = inet6_recvmsg,
3012 .mmap = sock_no_mmap,
3013 .sendpage = inet_sendpage,
3014 #ifdef CONFIG_COMPAT
3015 .compat_ioctl = inet6_compat_ioctl,
3019 static struct proto mptcp_v6_prot;
3021 static void mptcp_v6_destroy(struct sock *sk)
3024 inet6_destroy_sock(sk);
3027 static struct inet_protosw mptcp_v6_protosw = {
3028 .type = SOCK_STREAM,
3029 .protocol = IPPROTO_MPTCP,
3030 .prot = &mptcp_v6_prot,
3031 .ops = &mptcp_v6_stream_ops,
3032 .flags = INET_PROTOSW_ICSK,
3035 int __init mptcp_proto_v6_init(void)
3039 mptcp_v6_prot = mptcp_prot;
3040 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3041 mptcp_v6_prot.slab = NULL;
3042 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3043 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3045 err = proto_register(&mptcp_v6_prot, 1);
3049 err = inet6_register_protosw(&mptcp_v6_protosw);
3051 proto_unregister(&mptcp_v6_prot);