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 #define CREATE_TRACE_POINTS
29 #include <trace/events/mptcp.h>
31 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
33 struct mptcp_sock msk;
44 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
46 static struct percpu_counter mptcp_sockets_allocated;
48 static void __mptcp_destroy_sock(struct sock *sk);
49 static void __mptcp_check_send_data_fin(struct sock *sk);
51 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
52 static struct net_device mptcp_napi_dev;
54 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
55 * completed yet or has failed, return the subflow socket.
56 * Otherwise return NULL.
58 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
60 if (!msk->subflow || READ_ONCE(msk->can_ack))
66 /* Returns end sequence number of the receiver's advertised window */
67 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
69 return READ_ONCE(msk->wnd_end);
72 static bool mptcp_is_tcpsk(struct sock *sk)
74 struct socket *sock = sk->sk_socket;
76 if (unlikely(sk->sk_prot == &tcp_prot)) {
77 /* we are being invoked after mptcp_accept() has
78 * accepted a non-mp-capable flow: sk is a tcp_sk,
81 * Hand the socket over to tcp so all further socket ops
84 sock->ops = &inet_stream_ops;
86 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
87 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
88 sock->ops = &inet6_stream_ops;
96 static int __mptcp_socket_create(struct mptcp_sock *msk)
98 struct mptcp_subflow_context *subflow;
99 struct sock *sk = (struct sock *)msk;
100 struct socket *ssock;
103 err = mptcp_subflow_create_socket(sk, &ssock);
107 msk->first = ssock->sk;
108 msk->subflow = ssock;
109 subflow = mptcp_subflow_ctx(ssock->sk);
110 list_add(&subflow->node, &msk->conn_list);
111 sock_hold(ssock->sk);
112 subflow->request_mptcp = 1;
113 mptcp_sock_graft(msk->first, sk->sk_socket);
118 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
120 sk_drops_add(sk, skb);
124 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
125 struct sk_buff *from)
130 if (MPTCP_SKB_CB(from)->offset ||
131 !skb_try_coalesce(to, from, &fragstolen, &delta))
134 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
135 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
136 to->len, MPTCP_SKB_CB(from)->end_seq);
137 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
138 kfree_skb_partial(from, fragstolen);
139 atomic_add(delta, &sk->sk_rmem_alloc);
140 sk_mem_charge(sk, delta);
144 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
145 struct sk_buff *from)
147 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
150 return mptcp_try_coalesce((struct sock *)msk, to, from);
153 /* "inspired" by tcp_data_queue_ofo(), main differences:
155 * - don't cope with sacks
157 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
159 struct sock *sk = (struct sock *)msk;
160 struct rb_node **p, *parent;
161 u64 seq, end_seq, max_seq;
162 struct sk_buff *skb1;
164 seq = MPTCP_SKB_CB(skb)->map_seq;
165 end_seq = MPTCP_SKB_CB(skb)->end_seq;
166 max_seq = READ_ONCE(msk->rcv_wnd_sent);
168 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
169 RB_EMPTY_ROOT(&msk->out_of_order_queue));
170 if (after64(end_seq, max_seq)) {
173 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
174 (unsigned long long)end_seq - (unsigned long)max_seq,
175 (unsigned long long)msk->rcv_wnd_sent);
176 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
180 p = &msk->out_of_order_queue.rb_node;
181 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
182 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
183 rb_link_node(&skb->rbnode, NULL, p);
184 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
185 msk->ooo_last_skb = skb;
189 /* with 2 subflows, adding at end of ooo queue is quite likely
190 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
192 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
193 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
194 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
198 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
199 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
200 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
201 parent = &msk->ooo_last_skb->rbnode;
202 p = &parent->rb_right;
206 /* Find place to insert this segment. Handle overlaps on the way. */
210 skb1 = rb_to_skb(parent);
211 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
212 p = &parent->rb_left;
215 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
216 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
217 /* All the bits are present. Drop. */
219 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
222 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
226 * continue traversing
229 /* skb's seq == skb1's seq and skb covers skb1.
230 * Replace skb1 with skb.
232 rb_replace_node(&skb1->rbnode, &skb->rbnode,
233 &msk->out_of_order_queue);
234 mptcp_drop(sk, skb1);
235 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
238 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
239 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
242 p = &parent->rb_right;
246 /* Insert segment into RB tree. */
247 rb_link_node(&skb->rbnode, parent, p);
248 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
251 /* Remove other segments covered by skb. */
252 while ((skb1 = skb_rb_next(skb)) != NULL) {
253 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
255 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
256 mptcp_drop(sk, skb1);
257 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
259 /* If there is no skb after us, we are the last_skb ! */
261 msk->ooo_last_skb = skb;
265 skb_set_owner_r(skb, sk);
268 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
269 struct sk_buff *skb, unsigned int offset,
272 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
273 struct sock *sk = (struct sock *)msk;
274 struct sk_buff *tail;
276 __skb_unlink(skb, &ssk->sk_receive_queue);
281 /* try to fetch required memory from subflow */
282 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
283 int amount = sk_mem_pages(skb->truesize) << SK_MEM_QUANTUM_SHIFT;
285 if (ssk->sk_forward_alloc < amount)
288 ssk->sk_forward_alloc -= amount;
289 sk->sk_forward_alloc += amount;
292 /* the skb map_seq accounts for the skb offset:
293 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
296 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
297 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
298 MPTCP_SKB_CB(skb)->offset = offset;
300 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
302 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
303 tail = skb_peek_tail(&sk->sk_receive_queue);
304 if (tail && mptcp_try_coalesce(sk, tail, skb))
307 skb_set_owner_r(skb, sk);
308 __skb_queue_tail(&sk->sk_receive_queue, skb);
310 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
311 mptcp_data_queue_ofo(msk, skb);
315 /* old data, keep it simple and drop the whole pkt, sender
316 * will retransmit as needed, if needed.
318 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
324 static void mptcp_stop_timer(struct sock *sk)
326 struct inet_connection_sock *icsk = inet_csk(sk);
328 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
329 mptcp_sk(sk)->timer_ival = 0;
332 static void mptcp_close_wake_up(struct sock *sk)
334 if (sock_flag(sk, SOCK_DEAD))
337 sk->sk_state_change(sk);
338 if (sk->sk_shutdown == SHUTDOWN_MASK ||
339 sk->sk_state == TCP_CLOSE)
340 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
342 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
345 static bool mptcp_pending_data_fin_ack(struct sock *sk)
347 struct mptcp_sock *msk = mptcp_sk(sk);
349 return !__mptcp_check_fallback(msk) &&
350 ((1 << sk->sk_state) &
351 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
352 msk->write_seq == READ_ONCE(msk->snd_una);
355 static void mptcp_check_data_fin_ack(struct sock *sk)
357 struct mptcp_sock *msk = mptcp_sk(sk);
359 /* Look for an acknowledged DATA_FIN */
360 if (mptcp_pending_data_fin_ack(sk)) {
361 WRITE_ONCE(msk->snd_data_fin_enable, 0);
363 switch (sk->sk_state) {
365 inet_sk_state_store(sk, TCP_FIN_WAIT2);
369 inet_sk_state_store(sk, TCP_CLOSE);
373 mptcp_close_wake_up(sk);
377 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
379 struct mptcp_sock *msk = mptcp_sk(sk);
381 if (READ_ONCE(msk->rcv_data_fin) &&
382 ((1 << sk->sk_state) &
383 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
384 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
386 if (msk->ack_seq == rcv_data_fin_seq) {
388 *seq = rcv_data_fin_seq;
397 static void mptcp_set_datafin_timeout(const struct sock *sk)
399 struct inet_connection_sock *icsk = inet_csk(sk);
401 mptcp_sk(sk)->timer_ival = min(TCP_RTO_MAX,
402 TCP_RTO_MIN << icsk->icsk_retransmits);
405 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
407 long tout = ssk && inet_csk(ssk)->icsk_pending ?
408 inet_csk(ssk)->icsk_timeout - jiffies : 0;
411 tout = mptcp_sk(sk)->timer_ival;
412 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
415 static bool tcp_can_send_ack(const struct sock *ssk)
417 return !((1 << inet_sk_state_load(ssk)) &
418 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
421 static void mptcp_send_ack(struct mptcp_sock *msk)
423 struct mptcp_subflow_context *subflow;
425 mptcp_for_each_subflow(msk, subflow) {
426 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
429 if (tcp_can_send_ack(ssk))
435 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
440 ret = tcp_can_send_ack(ssk);
442 tcp_cleanup_rbuf(ssk, 1);
447 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
449 struct sock *ack_hint = READ_ONCE(msk->ack_hint);
450 int old_space = READ_ONCE(msk->old_wspace);
451 struct mptcp_subflow_context *subflow;
452 struct sock *sk = (struct sock *)msk;
455 /* this is a simple superset of what tcp_cleanup_rbuf() implements
456 * so that we don't have to acquire the ssk socket lock most of the time
457 * to do actually nothing
459 cleanup = __mptcp_space(sk) - old_space >= max(0, old_space);
463 /* if the hinted ssk is still active, try to use it */
464 if (likely(ack_hint)) {
465 mptcp_for_each_subflow(msk, subflow) {
466 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
468 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk))
473 /* otherwise pick the first active subflow */
474 mptcp_for_each_subflow(msk, subflow)
475 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
479 static bool mptcp_check_data_fin(struct sock *sk)
481 struct mptcp_sock *msk = mptcp_sk(sk);
482 u64 rcv_data_fin_seq;
485 if (__mptcp_check_fallback(msk))
488 /* Need to ack a DATA_FIN received from a peer while this side
489 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
490 * msk->rcv_data_fin was set when parsing the incoming options
491 * at the subflow level and the msk lock was not held, so this
492 * is the first opportunity to act on the DATA_FIN and change
495 * If we are caught up to the sequence number of the incoming
496 * DATA_FIN, send the DATA_ACK now and do state transition. If
497 * not caught up, do nothing and let the recv code send DATA_ACK
501 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
502 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
503 WRITE_ONCE(msk->rcv_data_fin, 0);
505 sk->sk_shutdown |= RCV_SHUTDOWN;
506 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
507 set_bit(MPTCP_DATA_READY, &msk->flags);
509 switch (sk->sk_state) {
510 case TCP_ESTABLISHED:
511 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
514 inet_sk_state_store(sk, TCP_CLOSING);
517 inet_sk_state_store(sk, TCP_CLOSE);
520 /* Other states not expected */
526 mptcp_set_timeout(sk, NULL);
528 mptcp_close_wake_up(sk);
533 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
537 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
538 struct sock *sk = (struct sock *)msk;
539 unsigned int moved = 0;
540 bool more_data_avail;
545 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
547 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
548 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
550 if (unlikely(ssk_rbuf > sk_rbuf)) {
551 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
556 pr_debug("msk=%p ssk=%p", msk, ssk);
559 u32 map_remaining, offset;
560 u32 seq = tp->copied_seq;
564 /* try to move as much data as available */
565 map_remaining = subflow->map_data_len -
566 mptcp_subflow_get_map_offset(subflow);
568 skb = skb_peek(&ssk->sk_receive_queue);
570 /* if no data is found, a racing workqueue/recvmsg
571 * already processed the new data, stop here or we
572 * can enter an infinite loop
579 if (__mptcp_check_fallback(msk)) {
580 /* if we are running under the workqueue, TCP could have
581 * collapsed skbs between dummy map creation and now
582 * be sure to adjust the size
584 map_remaining = skb->len;
585 subflow->map_data_len = skb->len;
588 offset = seq - TCP_SKB_CB(skb)->seq;
589 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
595 if (offset < skb->len) {
596 size_t len = skb->len - offset;
601 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
605 if (WARN_ON_ONCE(map_remaining < len))
609 sk_eat_skb(ssk, skb);
613 WRITE_ONCE(tp->copied_seq, seq);
614 more_data_avail = mptcp_subflow_data_available(ssk);
616 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
620 } while (more_data_avail);
621 WRITE_ONCE(msk->ack_hint, ssk);
627 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
629 struct sock *sk = (struct sock *)msk;
630 struct sk_buff *skb, *tail;
635 p = rb_first(&msk->out_of_order_queue);
636 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
639 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
643 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
645 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
648 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
652 end_seq = MPTCP_SKB_CB(skb)->end_seq;
653 tail = skb_peek_tail(&sk->sk_receive_queue);
654 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
655 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
657 /* skip overlapping data, if any */
658 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
659 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
661 MPTCP_SKB_CB(skb)->offset += delta;
662 __skb_queue_tail(&sk->sk_receive_queue, skb);
664 msk->ack_seq = end_seq;
670 /* In most cases we will be able to lock the mptcp socket. If its already
671 * owned, we need to defer to the work queue to avoid ABBA deadlock.
673 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
675 struct sock *sk = (struct sock *)msk;
676 unsigned int moved = 0;
678 if (inet_sk_state_load(sk) == TCP_CLOSE)
681 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
682 __mptcp_ofo_queue(msk);
683 if (unlikely(ssk->sk_err)) {
684 if (!sock_owned_by_user(sk))
685 __mptcp_error_report(sk);
687 set_bit(MPTCP_ERROR_REPORT, &msk->flags);
690 /* If the moves have caught up with the DATA_FIN sequence number
691 * it's time to ack the DATA_FIN and change socket state, but
692 * this is not a good place to change state. Let the workqueue
695 if (mptcp_pending_data_fin(sk, NULL))
696 mptcp_schedule_work(sk);
700 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
702 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
703 struct mptcp_sock *msk = mptcp_sk(sk);
704 int sk_rbuf, ssk_rbuf;
706 /* The peer can send data while we are shutting down this
707 * subflow at msk destruction time, but we must avoid enqueuing
708 * more data to the msk receive queue
710 if (unlikely(subflow->disposable))
713 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
714 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
715 if (unlikely(ssk_rbuf > sk_rbuf))
718 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
719 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
722 /* Wake-up the reader only for in-sequence data */
724 if (move_skbs_to_msk(msk, ssk)) {
725 set_bit(MPTCP_DATA_READY, &msk->flags);
726 sk->sk_data_ready(sk);
728 mptcp_data_unlock(sk);
731 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
733 struct mptcp_subflow_context *subflow;
736 if (likely(list_empty(&msk->join_list)))
739 spin_lock_bh(&msk->join_list_lock);
740 list_for_each_entry(subflow, &msk->join_list, node) {
741 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
743 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
744 if (READ_ONCE(msk->setsockopt_seq) != sseq)
747 list_splice_tail_init(&msk->join_list, &msk->conn_list);
748 spin_unlock_bh(&msk->join_list_lock);
753 void __mptcp_flush_join_list(struct mptcp_sock *msk)
755 if (likely(!mptcp_do_flush_join_list(msk)))
758 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
759 mptcp_schedule_work((struct sock *)msk);
762 static void mptcp_flush_join_list(struct mptcp_sock *msk)
764 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
768 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
771 mptcp_sockopt_sync_all(msk);
774 static bool mptcp_timer_pending(struct sock *sk)
776 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
779 static void mptcp_reset_timer(struct sock *sk)
781 struct inet_connection_sock *icsk = inet_csk(sk);
784 /* prevent rescheduling on close */
785 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
788 /* should never be called with mptcp level timer cleared */
789 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
790 if (WARN_ON_ONCE(!tout))
792 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
795 bool mptcp_schedule_work(struct sock *sk)
797 if (inet_sk_state_load(sk) != TCP_CLOSE &&
798 schedule_work(&mptcp_sk(sk)->work)) {
799 /* each subflow already holds a reference to the sk, and the
800 * workqueue is invoked by a subflow, so sk can't go away here.
808 void mptcp_subflow_eof(struct sock *sk)
810 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
811 mptcp_schedule_work(sk);
814 static void mptcp_check_for_eof(struct mptcp_sock *msk)
816 struct mptcp_subflow_context *subflow;
817 struct sock *sk = (struct sock *)msk;
820 mptcp_for_each_subflow(msk, subflow)
821 receivers += !subflow->rx_eof;
825 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
826 /* hopefully temporary hack: propagate shutdown status
827 * to msk, when all subflows agree on it
829 sk->sk_shutdown |= RCV_SHUTDOWN;
831 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
832 set_bit(MPTCP_DATA_READY, &msk->flags);
833 sk->sk_data_ready(sk);
836 switch (sk->sk_state) {
837 case TCP_ESTABLISHED:
838 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
841 inet_sk_state_store(sk, TCP_CLOSING);
844 inet_sk_state_store(sk, TCP_CLOSE);
849 mptcp_close_wake_up(sk);
852 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
854 struct mptcp_subflow_context *subflow;
855 struct sock *sk = (struct sock *)msk;
857 sock_owned_by_me(sk);
859 mptcp_for_each_subflow(msk, subflow) {
860 if (READ_ONCE(subflow->data_avail))
861 return mptcp_subflow_tcp_sock(subflow);
867 static bool mptcp_skb_can_collapse_to(u64 write_seq,
868 const struct sk_buff *skb,
869 const struct mptcp_ext *mpext)
871 if (!tcp_skb_can_collapse_to(skb))
874 /* can collapse only if MPTCP level sequence is in order and this
875 * mapping has not been xmitted yet
877 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
881 /* we can append data to the given data frag if:
882 * - there is space available in the backing page_frag
883 * - the data frag tail matches the current page_frag free offset
884 * - the data frag end sequence number matches the current write seq
886 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
887 const struct page_frag *pfrag,
888 const struct mptcp_data_frag *df)
890 return df && pfrag->page == df->page &&
891 pfrag->size - pfrag->offset > 0 &&
892 pfrag->offset == (df->offset + df->data_len) &&
893 df->data_seq + df->data_len == msk->write_seq;
896 static int mptcp_wmem_with_overhead(struct sock *sk, int size)
898 struct mptcp_sock *msk = mptcp_sk(sk);
901 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
902 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache;
903 if (skbs < msk->skb_tx_cache.qlen)
906 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER);
909 static void __mptcp_wmem_reserve(struct sock *sk, int size)
911 int amount = mptcp_wmem_with_overhead(sk, size);
912 struct mptcp_sock *msk = mptcp_sk(sk);
914 WARN_ON_ONCE(msk->wmem_reserved);
915 if (WARN_ON_ONCE(amount < 0))
918 if (amount <= sk->sk_forward_alloc)
921 /* under memory pressure try to reserve at most a single page
922 * otherwise try to reserve the full estimate and fallback
923 * to a single page before entering the error path
925 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
926 !sk_wmem_schedule(sk, amount)) {
927 if (amount <= PAGE_SIZE)
931 if (!sk_wmem_schedule(sk, amount))
936 msk->wmem_reserved = amount;
937 sk->sk_forward_alloc -= amount;
941 /* we will wait for memory on next allocation */
942 msk->wmem_reserved = -1;
945 static void __mptcp_update_wmem(struct sock *sk)
947 struct mptcp_sock *msk = mptcp_sk(sk);
949 #ifdef CONFIG_LOCKDEP
950 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
953 if (!msk->wmem_reserved)
956 if (msk->wmem_reserved < 0)
957 msk->wmem_reserved = 0;
958 if (msk->wmem_reserved > 0) {
959 sk->sk_forward_alloc += msk->wmem_reserved;
960 msk->wmem_reserved = 0;
964 static bool mptcp_wmem_alloc(struct sock *sk, int size)
966 struct mptcp_sock *msk = mptcp_sk(sk);
968 /* check for pre-existing error condition */
969 if (msk->wmem_reserved < 0)
972 if (msk->wmem_reserved >= size)
976 if (!sk_wmem_schedule(sk, size)) {
977 mptcp_data_unlock(sk);
981 sk->sk_forward_alloc -= size;
982 msk->wmem_reserved += size;
983 mptcp_data_unlock(sk);
986 msk->wmem_reserved -= size;
990 static void mptcp_wmem_uncharge(struct sock *sk, int size)
992 struct mptcp_sock *msk = mptcp_sk(sk);
994 if (msk->wmem_reserved < 0)
995 msk->wmem_reserved = 0;
996 msk->wmem_reserved += size;
999 static void mptcp_mem_reclaim_partial(struct sock *sk)
1001 struct mptcp_sock *msk = mptcp_sk(sk);
1003 /* if we are experiencing a transint allocation error,
1004 * the forward allocation memory has been already
1007 if (msk->wmem_reserved < 0)
1010 mptcp_data_lock(sk);
1011 sk->sk_forward_alloc += msk->wmem_reserved;
1012 sk_mem_reclaim_partial(sk);
1013 msk->wmem_reserved = sk->sk_forward_alloc;
1014 sk->sk_forward_alloc = 0;
1015 mptcp_data_unlock(sk);
1018 static void dfrag_uncharge(struct sock *sk, int len)
1020 sk_mem_uncharge(sk, len);
1021 sk_wmem_queued_add(sk, -len);
1024 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1026 int len = dfrag->data_len + dfrag->overhead;
1028 list_del(&dfrag->list);
1029 dfrag_uncharge(sk, len);
1030 put_page(dfrag->page);
1033 static void __mptcp_clean_una(struct sock *sk)
1035 struct mptcp_sock *msk = mptcp_sk(sk);
1036 struct mptcp_data_frag *dtmp, *dfrag;
1037 bool cleaned = false;
1040 /* on fallback we just need to ignore snd_una, as this is really
1043 if (__mptcp_check_fallback(msk))
1044 msk->snd_una = READ_ONCE(msk->snd_nxt);
1046 snd_una = msk->snd_una;
1047 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1048 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1051 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1053 dfrag_clear(sk, dfrag);
1057 dfrag = mptcp_rtx_head(sk);
1058 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1059 u64 delta = snd_una - dfrag->data_seq;
1061 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1064 dfrag->data_seq += delta;
1065 dfrag->offset += delta;
1066 dfrag->data_len -= delta;
1067 dfrag->already_sent -= delta;
1069 dfrag_uncharge(sk, delta);
1075 if (tcp_under_memory_pressure(sk)) {
1076 __mptcp_update_wmem(sk);
1077 sk_mem_reclaim_partial(sk);
1081 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1082 if (msk->timer_ival && !mptcp_data_fin_enabled(msk))
1083 mptcp_stop_timer(sk);
1085 mptcp_reset_timer(sk);
1089 static void __mptcp_clean_una_wakeup(struct sock *sk)
1091 #ifdef CONFIG_LOCKDEP
1092 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
1094 __mptcp_clean_una(sk);
1095 mptcp_write_space(sk);
1098 static void mptcp_clean_una_wakeup(struct sock *sk)
1100 mptcp_data_lock(sk);
1101 __mptcp_clean_una_wakeup(sk);
1102 mptcp_data_unlock(sk);
1105 static void mptcp_enter_memory_pressure(struct sock *sk)
1107 struct mptcp_subflow_context *subflow;
1108 struct mptcp_sock *msk = mptcp_sk(sk);
1111 sk_stream_moderate_sndbuf(sk);
1112 mptcp_for_each_subflow(msk, subflow) {
1113 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1116 tcp_enter_memory_pressure(ssk);
1117 sk_stream_moderate_sndbuf(ssk);
1122 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1125 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1127 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1128 pfrag, sk->sk_allocation)))
1131 mptcp_enter_memory_pressure(sk);
1135 static struct mptcp_data_frag *
1136 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1139 int offset = ALIGN(orig_offset, sizeof(long));
1140 struct mptcp_data_frag *dfrag;
1142 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1143 dfrag->data_len = 0;
1144 dfrag->data_seq = msk->write_seq;
1145 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1146 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1147 dfrag->already_sent = 0;
1148 dfrag->page = pfrag->page;
1153 struct mptcp_sendmsg_info {
1161 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1164 u64 window_end = mptcp_wnd_end(msk);
1166 if (__mptcp_check_fallback(msk))
1169 if (!before64(data_seq + avail_size, window_end)) {
1170 u64 allowed_size = window_end - data_seq;
1172 return min_t(unsigned int, allowed_size, avail_size);
1178 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1180 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1184 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1188 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1190 struct sk_buff *skb;
1192 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1194 if (likely(__mptcp_add_ext(skb, gfp))) {
1195 skb_reserve(skb, MAX_TCP_HEADER);
1196 skb->reserved_tailroom = skb->end - skb->tail;
1201 mptcp_enter_memory_pressure(sk);
1206 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1207 struct sk_buff_head *skbs, int *total_ts)
1209 struct mptcp_sock *msk = mptcp_sk(sk);
1210 struct sk_buff *skb;
1213 if (unlikely(tcp_under_memory_pressure(sk))) {
1214 mptcp_mem_reclaim_partial(sk);
1216 /* under pressure pre-allocate at most a single skb */
1217 if (msk->skb_tx_cache.qlen)
1219 space_needed = msk->size_goal_cache;
1221 space_needed = msk->tx_pending_data + size -
1222 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1225 while (space_needed > 0) {
1226 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1227 if (unlikely(!skb)) {
1228 /* under memory pressure, try to pass the caller a
1229 * single skb to allow forward progress
1231 while (skbs->qlen > 1) {
1232 skb = __skb_dequeue_tail(skbs);
1233 *total_ts -= skb->truesize;
1236 return skbs->qlen > 0;
1239 *total_ts += skb->truesize;
1240 __skb_queue_tail(skbs, skb);
1241 space_needed -= msk->size_goal_cache;
1246 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1248 struct mptcp_sock *msk = mptcp_sk(sk);
1249 struct sk_buff *skb;
1251 if (ssk->sk_tx_skb_cache) {
1252 skb = ssk->sk_tx_skb_cache;
1253 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1254 !__mptcp_add_ext(skb, gfp)))
1259 skb = skb_peek(&msk->skb_tx_cache);
1261 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1262 skb = __skb_dequeue(&msk->skb_tx_cache);
1263 if (WARN_ON_ONCE(!skb))
1266 mptcp_wmem_uncharge(sk, skb->truesize);
1267 ssk->sk_tx_skb_cache = skb;
1271 /* over memory limit, no point to try to allocate a new skb */
1275 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1279 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1280 ssk->sk_tx_skb_cache = skb;
1287 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1289 return !ssk->sk_tx_skb_cache &&
1290 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1291 tcp_under_memory_pressure(sk);
1294 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1296 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1297 mptcp_mem_reclaim_partial(sk);
1298 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1301 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1302 struct mptcp_data_frag *dfrag,
1303 struct mptcp_sendmsg_info *info)
1305 u64 data_seq = dfrag->data_seq + info->sent;
1306 struct mptcp_sock *msk = mptcp_sk(sk);
1307 bool zero_window_probe = false;
1308 struct mptcp_ext *mpext = NULL;
1309 struct sk_buff *skb, *tail;
1310 bool can_collapse = false;
1315 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1316 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1318 /* compute send limit */
1319 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1320 avail_size = info->size_goal;
1321 msk->size_goal_cache = info->size_goal;
1322 skb = tcp_write_queue_tail(ssk);
1324 /* Limit the write to the size available in the
1325 * current skb, if any, so that we create at most a new skb.
1326 * Explicitly tells TCP internals to avoid collapsing on later
1327 * queue management operation, to avoid breaking the ext <->
1328 * SSN association set here
1330 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1331 can_collapse = (info->size_goal - skb->len > 0) &&
1332 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1333 if (!can_collapse) {
1334 TCP_SKB_CB(skb)->eor = 1;
1336 size_bias = skb->len;
1337 avail_size = info->size_goal - skb->len;
1341 /* Zero window and all data acked? Probe. */
1342 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1343 if (avail_size == 0) {
1344 u64 snd_una = READ_ONCE(msk->snd_una);
1346 if (skb || snd_una != msk->snd_nxt)
1348 zero_window_probe = true;
1349 data_seq = snd_una - 1;
1353 if (WARN_ON_ONCE(info->sent > info->limit ||
1354 info->limit > dfrag->data_len))
1357 ret = info->limit - info->sent;
1358 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1359 dfrag->page, dfrag->offset + info->sent, &ret);
1361 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1365 /* if the tail skb is still the cached one, collapsing really happened.
1368 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1369 mpext->data_len += ret;
1370 WARN_ON_ONCE(!can_collapse);
1371 WARN_ON_ONCE(zero_window_probe);
1375 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1376 if (WARN_ON_ONCE(!mpext)) {
1377 /* should never reach here, stream corrupted */
1381 memset(mpext, 0, sizeof(*mpext));
1382 mpext->data_seq = data_seq;
1383 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1384 mpext->data_len = ret;
1388 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1389 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1392 if (zero_window_probe) {
1393 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1396 tcp_push_pending_frames(ssk);
1399 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1403 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1404 sizeof(struct tcphdr) - \
1405 MAX_TCP_OPTION_SPACE - \
1406 sizeof(struct ipv6hdr) - \
1407 sizeof(struct frag_hdr))
1409 struct subflow_send_info {
1414 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1416 struct subflow_send_info send_info[2];
1417 struct mptcp_subflow_context *subflow;
1418 int i, nr_active = 0;
1423 sock_owned_by_me((struct sock *)msk);
1425 if (__mptcp_check_fallback(msk)) {
1428 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1431 /* re-use last subflow, if the burst allow that */
1432 if (msk->last_snd && msk->snd_burst > 0 &&
1433 sk_stream_memory_free(msk->last_snd) &&
1434 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1435 return msk->last_snd;
1437 /* pick the subflow with the lower wmem/wspace ratio */
1438 for (i = 0; i < 2; ++i) {
1439 send_info[i].ssk = NULL;
1440 send_info[i].ratio = -1;
1442 mptcp_for_each_subflow(msk, subflow) {
1443 trace_mptcp_subflow_get_send(subflow);
1444 ssk = mptcp_subflow_tcp_sock(subflow);
1445 if (!mptcp_subflow_active(subflow))
1448 nr_active += !subflow->backup;
1449 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1452 pace = READ_ONCE(ssk->sk_pacing_rate);
1456 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1458 if (ratio < send_info[subflow->backup].ratio) {
1459 send_info[subflow->backup].ssk = ssk;
1460 send_info[subflow->backup].ratio = ratio;
1464 /* pick the best backup if no other subflow is active */
1466 send_info[0].ssk = send_info[1].ssk;
1468 if (send_info[0].ssk) {
1469 msk->last_snd = send_info[0].ssk;
1470 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1471 tcp_sk(msk->last_snd)->snd_wnd);
1472 return msk->last_snd;
1478 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1479 struct mptcp_sendmsg_info *info)
1481 mptcp_set_timeout(sk, ssk);
1482 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1486 static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1488 struct sock *prev_ssk = NULL, *ssk = NULL;
1489 struct mptcp_sock *msk = mptcp_sk(sk);
1490 struct mptcp_sendmsg_info info = {
1493 struct mptcp_data_frag *dfrag;
1494 int len, copied = 0;
1496 while ((dfrag = mptcp_send_head(sk))) {
1497 info.sent = dfrag->already_sent;
1498 info.limit = dfrag->data_len;
1499 len = dfrag->data_len - dfrag->already_sent;
1504 mptcp_flush_join_list(msk);
1505 ssk = mptcp_subflow_get_send(msk);
1507 /* try to keep the subflow socket lock across
1508 * consecutive xmit on the same socket
1510 if (ssk != prev_ssk && prev_ssk)
1511 mptcp_push_release(sk, prev_ssk, &info);
1515 if (ssk != prev_ssk || !prev_ssk)
1518 /* keep it simple and always provide a new skb for the
1519 * subflow, even if we will not use it when collapsing
1520 * on the pending one
1522 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1523 mptcp_push_release(sk, ssk, &info);
1527 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1529 mptcp_push_release(sk, ssk, &info);
1534 dfrag->already_sent += ret;
1535 msk->snd_nxt += ret;
1536 msk->snd_burst -= ret;
1537 msk->tx_pending_data -= ret;
1541 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1544 /* at this point we held the socket lock for the last subflow we used */
1546 mptcp_push_release(sk, ssk, &info);
1550 /* start the timer, if it's not pending */
1551 if (!mptcp_timer_pending(sk))
1552 mptcp_reset_timer(sk);
1553 __mptcp_check_send_data_fin(sk);
1557 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1559 struct mptcp_sock *msk = mptcp_sk(sk);
1560 struct mptcp_sendmsg_info info;
1561 struct mptcp_data_frag *dfrag;
1562 struct sock *xmit_ssk;
1563 int len, copied = 0;
1567 while ((dfrag = mptcp_send_head(sk))) {
1568 info.sent = dfrag->already_sent;
1569 info.limit = dfrag->data_len;
1570 len = dfrag->data_len - dfrag->already_sent;
1574 /* the caller already invoked the packet scheduler,
1575 * check for a different subflow usage only after
1576 * spooling the first chunk of data
1578 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1581 if (xmit_ssk != ssk) {
1582 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1586 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1587 __mptcp_update_wmem(sk);
1588 sk_mem_reclaim_partial(sk);
1590 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1593 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1598 dfrag->already_sent += ret;
1599 msk->snd_nxt += ret;
1600 msk->snd_burst -= ret;
1601 msk->tx_pending_data -= ret;
1606 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1610 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1611 * not going to flush it via release_sock()
1613 __mptcp_update_wmem(sk);
1615 mptcp_set_timeout(sk, ssk);
1616 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1618 if (!mptcp_timer_pending(sk))
1619 mptcp_reset_timer(sk);
1621 if (msk->snd_data_fin_enable &&
1622 msk->snd_nxt + 1 == msk->write_seq)
1623 mptcp_schedule_work(sk);
1627 static void mptcp_set_nospace(struct sock *sk)
1629 /* enable autotune */
1630 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1632 /* will be cleared on avail space */
1633 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1636 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1638 struct mptcp_sock *msk = mptcp_sk(sk);
1639 struct page_frag *pfrag;
1644 /* we don't support FASTOPEN yet */
1645 if (msg->msg_flags & MSG_FASTOPEN)
1648 /* silently ignore everything else */
1649 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1651 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1653 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1655 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1656 ret = sk_stream_wait_connect(sk, &timeo);
1661 pfrag = sk_page_frag(sk);
1663 while (msg_data_left(msg)) {
1664 int total_ts, frag_truesize = 0;
1665 struct mptcp_data_frag *dfrag;
1666 struct sk_buff_head skbs;
1667 bool dfrag_collapsed;
1668 size_t psize, offset;
1670 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1675 /* reuse tail pfrag, if possible, or carve a new one from the
1678 dfrag = mptcp_pending_tail(sk);
1679 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1680 if (!dfrag_collapsed) {
1681 if (!sk_stream_memory_free(sk))
1682 goto wait_for_memory;
1684 if (!mptcp_page_frag_refill(sk, pfrag))
1685 goto wait_for_memory;
1687 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1688 frag_truesize = dfrag->overhead;
1691 /* we do not bound vs wspace, to allow a single packet.
1692 * memory accounting will prevent execessive memory usage
1695 offset = dfrag->offset + dfrag->data_len;
1696 psize = pfrag->size - offset;
1697 psize = min_t(size_t, psize, msg_data_left(msg));
1698 total_ts = psize + frag_truesize;
1699 __skb_queue_head_init(&skbs);
1700 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1701 goto wait_for_memory;
1703 if (!mptcp_wmem_alloc(sk, total_ts)) {
1704 __skb_queue_purge(&skbs);
1705 goto wait_for_memory;
1708 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1709 if (copy_page_from_iter(dfrag->page, offset, psize,
1710 &msg->msg_iter) != psize) {
1711 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1716 /* data successfully copied into the write queue */
1718 dfrag->data_len += psize;
1719 frag_truesize += psize;
1720 pfrag->offset += frag_truesize;
1721 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1722 msk->tx_pending_data += psize;
1724 /* charge data on mptcp pending queue to the msk socket
1725 * Note: we charge such data both to sk and ssk
1727 sk_wmem_queued_add(sk, frag_truesize);
1728 if (!dfrag_collapsed) {
1729 get_page(dfrag->page);
1730 list_add_tail(&dfrag->list, &msk->rtx_queue);
1731 if (!msk->first_pending)
1732 WRITE_ONCE(msk->first_pending, dfrag);
1734 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1735 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1741 mptcp_set_nospace(sk);
1742 __mptcp_push_pending(sk, msg->msg_flags);
1743 ret = sk_stream_wait_memory(sk, &timeo);
1749 __mptcp_push_pending(sk, msg->msg_flags);
1753 return copied ? : ret;
1756 static void mptcp_wait_data(struct sock *sk, long *timeo)
1758 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1759 struct mptcp_sock *msk = mptcp_sk(sk);
1761 add_wait_queue(sk_sleep(sk), &wait);
1762 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1764 sk_wait_event(sk, timeo,
1765 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1767 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1768 remove_wait_queue(sk_sleep(sk), &wait);
1771 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1773 size_t len, int flags)
1775 struct sk_buff *skb, *tmp;
1778 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1779 u32 offset = MPTCP_SKB_CB(skb)->offset;
1780 u32 data_len = skb->len - offset;
1781 u32 count = min_t(size_t, len - copied, data_len);
1784 if (!(flags & MSG_TRUNC)) {
1785 err = skb_copy_datagram_msg(skb, offset, msg, count);
1786 if (unlikely(err < 0)) {
1795 if (count < data_len) {
1796 if (!(flags & MSG_PEEK))
1797 MPTCP_SKB_CB(skb)->offset += count;
1801 if (!(flags & MSG_PEEK)) {
1802 /* we will bulk release the skb memory later */
1803 skb->destructor = NULL;
1804 msk->rmem_released += skb->truesize;
1805 __skb_unlink(skb, &msk->receive_queue);
1816 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1818 * Only difference: Use highest rtt estimate of the subflows in use.
1820 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1822 struct mptcp_subflow_context *subflow;
1823 struct sock *sk = (struct sock *)msk;
1824 u32 time, advmss = 1;
1827 sock_owned_by_me(sk);
1832 msk->rcvq_space.copied += copied;
1834 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1835 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1837 rtt_us = msk->rcvq_space.rtt_us;
1838 if (rtt_us && time < (rtt_us >> 3))
1842 mptcp_for_each_subflow(msk, subflow) {
1843 const struct tcp_sock *tp;
1847 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1849 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1850 sf_advmss = READ_ONCE(tp->advmss);
1852 rtt_us = max(sf_rtt_us, rtt_us);
1853 advmss = max(sf_advmss, advmss);
1856 msk->rcvq_space.rtt_us = rtt_us;
1857 if (time < (rtt_us >> 3) || rtt_us == 0)
1860 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1863 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1864 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1868 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1870 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1872 do_div(grow, msk->rcvq_space.space);
1873 rcvwin += (grow << 1);
1875 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1876 while (tcp_win_from_space(sk, rcvmem) < advmss)
1879 do_div(rcvwin, advmss);
1880 rcvbuf = min_t(u64, rcvwin * rcvmem,
1881 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1883 if (rcvbuf > sk->sk_rcvbuf) {
1886 window_clamp = tcp_win_from_space(sk, rcvbuf);
1887 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1889 /* Make subflows follow along. If we do not do this, we
1890 * get drops at subflow level if skbs can't be moved to
1891 * the mptcp rx queue fast enough (announced rcv_win can
1892 * exceed ssk->sk_rcvbuf).
1894 mptcp_for_each_subflow(msk, subflow) {
1898 ssk = mptcp_subflow_tcp_sock(subflow);
1899 slow = lock_sock_fast(ssk);
1900 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1901 tcp_sk(ssk)->window_clamp = window_clamp;
1902 tcp_cleanup_rbuf(ssk, 1);
1903 unlock_sock_fast(ssk, slow);
1908 msk->rcvq_space.space = msk->rcvq_space.copied;
1910 msk->rcvq_space.copied = 0;
1911 msk->rcvq_space.time = mstamp;
1914 static void __mptcp_update_rmem(struct sock *sk)
1916 struct mptcp_sock *msk = mptcp_sk(sk);
1918 if (!msk->rmem_released)
1921 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1922 sk_mem_uncharge(sk, msk->rmem_released);
1923 msk->rmem_released = 0;
1926 static void __mptcp_splice_receive_queue(struct sock *sk)
1928 struct mptcp_sock *msk = mptcp_sk(sk);
1930 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1933 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1935 struct sock *sk = (struct sock *)msk;
1936 unsigned int moved = 0;
1939 mptcp_flush_join_list(msk);
1941 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1944 /* we can have data pending in the subflows only if the msk
1945 * receive buffer was full at subflow_data_ready() time,
1946 * that is an unlikely slow path.
1951 slowpath = lock_sock_fast(ssk);
1952 mptcp_data_lock(sk);
1953 __mptcp_update_rmem(sk);
1954 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1955 mptcp_data_unlock(sk);
1956 tcp_cleanup_rbuf(ssk, moved);
1958 if (unlikely(ssk->sk_err))
1959 __mptcp_error_report(sk);
1960 unlock_sock_fast(ssk, slowpath);
1963 /* acquire the data lock only if some input data is pending */
1965 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1966 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1967 mptcp_data_lock(sk);
1968 __mptcp_update_rmem(sk);
1969 ret |= __mptcp_ofo_queue(msk);
1970 __mptcp_splice_receive_queue(sk);
1971 mptcp_data_unlock(sk);
1972 mptcp_cleanup_rbuf(msk);
1975 mptcp_check_data_fin((struct sock *)msk);
1976 return !skb_queue_empty(&msk->receive_queue);
1979 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1980 int nonblock, int flags, int *addr_len)
1982 struct mptcp_sock *msk = mptcp_sk(sk);
1987 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
1988 if (unlikely(flags & MSG_ERRQUEUE))
1989 return inet_recv_error(sk, msg, len, addr_len);
1991 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1992 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1997 timeo = sock_rcvtimeo(sk, nonblock);
1999 len = min_t(size_t, len, INT_MAX);
2000 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2002 while (copied < len) {
2005 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags);
2006 if (unlikely(bytes_read < 0)) {
2008 copied = bytes_read;
2012 copied += bytes_read;
2014 /* be sure to advertise window change */
2015 mptcp_cleanup_rbuf(msk);
2017 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2020 /* only the master socket status is relevant here. The exit
2021 * conditions mirror closely tcp_recvmsg()
2023 if (copied >= target)
2028 sk->sk_state == TCP_CLOSE ||
2029 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2031 signal_pending(current))
2035 copied = sock_error(sk);
2039 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2040 mptcp_check_for_eof(msk);
2042 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2043 /* race breaker: the shutdown could be after the
2044 * previous receive queue check
2046 if (__mptcp_move_skbs(msk))
2051 if (sk->sk_state == TCP_CLOSE) {
2061 if (signal_pending(current)) {
2062 copied = sock_intr_errno(timeo);
2067 pr_debug("block timeout %ld", timeo);
2068 mptcp_wait_data(sk, &timeo);
2071 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2072 skb_queue_empty(&msk->receive_queue)) {
2073 /* entire backlog drained, clear DATA_READY. */
2074 clear_bit(MPTCP_DATA_READY, &msk->flags);
2076 /* .. race-breaker: ssk might have gotten new data
2077 * after last __mptcp_move_skbs() returned false.
2079 if (unlikely(__mptcp_move_skbs(msk)))
2080 set_bit(MPTCP_DATA_READY, &msk->flags);
2081 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2082 /* data to read but mptcp_wait_data() cleared DATA_READY */
2083 set_bit(MPTCP_DATA_READY, &msk->flags);
2086 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2087 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2088 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2089 if (!(flags & MSG_PEEK))
2090 mptcp_rcv_space_adjust(msk, copied);
2096 static void mptcp_retransmit_timer(struct timer_list *t)
2098 struct inet_connection_sock *icsk = from_timer(icsk, t,
2099 icsk_retransmit_timer);
2100 struct sock *sk = &icsk->icsk_inet.sk;
2101 struct mptcp_sock *msk = mptcp_sk(sk);
2104 if (!sock_owned_by_user(sk)) {
2105 /* we need a process context to retransmit */
2106 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2107 mptcp_schedule_work(sk);
2109 /* delegate our work to tcp_release_cb() */
2110 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2116 static void mptcp_timeout_timer(struct timer_list *t)
2118 struct sock *sk = from_timer(sk, t, sk_timer);
2120 mptcp_schedule_work(sk);
2124 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2127 * A backup subflow is returned only if that is the only kind available.
2129 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2131 struct mptcp_subflow_context *subflow;
2132 struct sock *backup = NULL;
2134 sock_owned_by_me((const struct sock *)msk);
2136 if (__mptcp_check_fallback(msk))
2139 mptcp_for_each_subflow(msk, subflow) {
2140 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2142 if (!mptcp_subflow_active(subflow))
2145 /* still data outstanding at TCP level? Don't retransmit. */
2146 if (!tcp_write_queue_empty(ssk)) {
2147 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2152 if (subflow->backup) {
2164 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2167 iput(SOCK_INODE(msk->subflow));
2168 msk->subflow = NULL;
2172 /* subflow sockets can be either outgoing (connect) or incoming
2175 * Outgoing subflows use in-kernel sockets.
2176 * Incoming subflows do not have their own 'struct socket' allocated,
2177 * so we need to use tcp_close() after detaching them from the mptcp
2180 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2181 struct mptcp_subflow_context *subflow)
2183 struct mptcp_sock *msk = mptcp_sk(sk);
2185 list_del(&subflow->node);
2187 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2189 /* if we are invoked by the msk cleanup code, the subflow is
2195 subflow->disposable = 1;
2197 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2198 * the ssk has been already destroyed, we just need to release the
2199 * reference owned by msk;
2201 if (!inet_csk(ssk)->icsk_ulp_ops) {
2202 kfree_rcu(subflow, rcu);
2204 /* otherwise tcp will dispose of the ssk and subflow ctx */
2205 __tcp_close(ssk, 0);
2207 /* close acquired an extra ref */
2214 if (ssk == msk->last_snd)
2215 msk->last_snd = NULL;
2217 if (ssk == msk->ack_hint)
2218 msk->ack_hint = NULL;
2220 if (ssk == msk->first)
2223 if (msk->subflow && ssk == msk->subflow->sk)
2224 mptcp_dispose_initial_subflow(msk);
2227 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2228 struct mptcp_subflow_context *subflow)
2230 if (sk->sk_state == TCP_ESTABLISHED)
2231 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2232 __mptcp_close_ssk(sk, ssk, subflow);
2235 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2240 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2242 struct mptcp_subflow_context *subflow, *tmp;
2246 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2247 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2249 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2252 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2253 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2256 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2260 static bool mptcp_check_close_timeout(const struct sock *sk)
2262 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2263 struct mptcp_subflow_context *subflow;
2265 if (delta >= TCP_TIMEWAIT_LEN)
2268 /* if all subflows are in closed status don't bother with additional
2271 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2272 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2279 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2281 struct mptcp_subflow_context *subflow, *tmp;
2282 struct sock *sk = &msk->sk.icsk_inet.sk;
2284 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2287 mptcp_token_destroy(msk);
2289 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2290 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2293 if (tcp_sk->sk_state != TCP_CLOSE) {
2294 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2295 tcp_set_state(tcp_sk, TCP_CLOSE);
2297 release_sock(tcp_sk);
2300 inet_sk_state_store(sk, TCP_CLOSE);
2301 sk->sk_shutdown = SHUTDOWN_MASK;
2302 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2303 set_bit(MPTCP_DATA_READY, &msk->flags);
2304 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2306 mptcp_close_wake_up(sk);
2309 static void __mptcp_retrans(struct sock *sk)
2311 struct mptcp_sock *msk = mptcp_sk(sk);
2312 struct mptcp_sendmsg_info info = {};
2313 struct mptcp_data_frag *dfrag;
2318 mptcp_clean_una_wakeup(sk);
2319 dfrag = mptcp_rtx_head(sk);
2321 if (mptcp_data_fin_enabled(msk)) {
2322 struct inet_connection_sock *icsk = inet_csk(sk);
2324 icsk->icsk_retransmits++;
2325 mptcp_set_datafin_timeout(sk);
2326 mptcp_send_ack(msk);
2334 ssk = mptcp_subflow_get_retrans(msk);
2340 /* limit retransmission to the bytes already sent on some subflows */
2342 info.limit = dfrag->already_sent;
2343 while (info.sent < dfrag->already_sent) {
2344 if (!mptcp_alloc_tx_skb(sk, ssk))
2347 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2351 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2356 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2359 mptcp_set_timeout(sk, ssk);
2363 if (!mptcp_timer_pending(sk))
2364 mptcp_reset_timer(sk);
2367 static void mptcp_worker(struct work_struct *work)
2369 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2370 struct sock *sk = &msk->sk.icsk_inet.sk;
2374 state = sk->sk_state;
2375 if (unlikely(state == TCP_CLOSE))
2378 mptcp_check_data_fin_ack(sk);
2379 mptcp_flush_join_list(msk);
2381 mptcp_check_fastclose(msk);
2384 mptcp_pm_nl_work(msk);
2386 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2387 mptcp_check_for_eof(msk);
2389 __mptcp_check_send_data_fin(sk);
2390 mptcp_check_data_fin(sk);
2392 /* There is no point in keeping around an orphaned sk timedout or
2393 * closed, but we need the msk around to reply to incoming DATA_FIN,
2394 * even if it is orphaned and in FIN_WAIT2 state
2396 if (sock_flag(sk, SOCK_DEAD) &&
2397 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2398 inet_sk_state_store(sk, TCP_CLOSE);
2399 __mptcp_destroy_sock(sk);
2403 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2404 __mptcp_close_subflow(msk);
2406 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2407 __mptcp_retrans(sk);
2414 static int __mptcp_init_sock(struct sock *sk)
2416 struct mptcp_sock *msk = mptcp_sk(sk);
2418 spin_lock_init(&msk->join_list_lock);
2420 INIT_LIST_HEAD(&msk->conn_list);
2421 INIT_LIST_HEAD(&msk->join_list);
2422 INIT_LIST_HEAD(&msk->rtx_queue);
2423 INIT_WORK(&msk->work, mptcp_worker);
2424 __skb_queue_head_init(&msk->receive_queue);
2425 __skb_queue_head_init(&msk->skb_tx_cache);
2426 msk->out_of_order_queue = RB_ROOT;
2427 msk->first_pending = NULL;
2428 msk->wmem_reserved = 0;
2429 msk->rmem_released = 0;
2430 msk->tx_pending_data = 0;
2431 msk->size_goal_cache = TCP_BASE_MSS;
2433 msk->ack_hint = NULL;
2435 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2437 mptcp_pm_data_init(msk);
2439 /* re-use the csk retrans timer for MPTCP-level retrans */
2440 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2441 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2446 static int mptcp_init_sock(struct sock *sk)
2448 struct inet_connection_sock *icsk = inet_csk(sk);
2449 struct net *net = sock_net(sk);
2452 ret = __mptcp_init_sock(sk);
2456 if (!mptcp_is_enabled(net))
2457 return -ENOPROTOOPT;
2459 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2462 ret = __mptcp_socket_create(mptcp_sk(sk));
2466 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2467 * propagate the correct value
2469 tcp_assign_congestion_control(sk);
2470 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2472 /* no need to keep a reference to the ops, the name will suffice */
2473 tcp_cleanup_congestion_control(sk);
2474 icsk->icsk_ca_ops = NULL;
2476 sk_sockets_allocated_inc(sk);
2477 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2478 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2483 static void __mptcp_clear_xmit(struct sock *sk)
2485 struct mptcp_sock *msk = mptcp_sk(sk);
2486 struct mptcp_data_frag *dtmp, *dfrag;
2487 struct sk_buff *skb;
2489 WRITE_ONCE(msk->first_pending, NULL);
2490 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2491 dfrag_clear(sk, dfrag);
2492 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2493 sk->sk_forward_alloc += skb->truesize;
2498 static void mptcp_cancel_work(struct sock *sk)
2500 struct mptcp_sock *msk = mptcp_sk(sk);
2502 if (cancel_work_sync(&msk->work))
2506 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2510 switch (ssk->sk_state) {
2512 if (!(how & RCV_SHUTDOWN))
2516 tcp_disconnect(ssk, O_NONBLOCK);
2519 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2520 pr_debug("Fallback");
2521 ssk->sk_shutdown |= how;
2522 tcp_shutdown(ssk, how);
2524 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2525 mptcp_set_timeout(sk, ssk);
2527 if (!mptcp_timer_pending(sk))
2528 mptcp_reset_timer(sk);
2536 static const unsigned char new_state[16] = {
2537 /* current state: new state: action: */
2538 [0 /* (Invalid) */] = TCP_CLOSE,
2539 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2540 [TCP_SYN_SENT] = TCP_CLOSE,
2541 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2542 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2543 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2544 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2545 [TCP_CLOSE] = TCP_CLOSE,
2546 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2547 [TCP_LAST_ACK] = TCP_LAST_ACK,
2548 [TCP_LISTEN] = TCP_CLOSE,
2549 [TCP_CLOSING] = TCP_CLOSING,
2550 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2553 static int mptcp_close_state(struct sock *sk)
2555 int next = (int)new_state[sk->sk_state];
2556 int ns = next & TCP_STATE_MASK;
2558 inet_sk_state_store(sk, ns);
2560 return next & TCP_ACTION_FIN;
2563 static void __mptcp_check_send_data_fin(struct sock *sk)
2565 struct mptcp_subflow_context *subflow;
2566 struct mptcp_sock *msk = mptcp_sk(sk);
2568 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2569 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2570 msk->snd_nxt, msk->write_seq);
2572 /* we still need to enqueue subflows or not really shutting down,
2575 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2576 mptcp_send_head(sk))
2579 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2581 /* fallback socket will not get data_fin/ack, can move to the next
2584 if (__mptcp_check_fallback(msk)) {
2585 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2586 inet_sk_state_store(sk, TCP_CLOSE);
2587 mptcp_close_wake_up(sk);
2588 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2589 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2593 mptcp_flush_join_list(msk);
2594 mptcp_for_each_subflow(msk, subflow) {
2595 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2597 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2601 static void __mptcp_wr_shutdown(struct sock *sk)
2603 struct mptcp_sock *msk = mptcp_sk(sk);
2605 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2606 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2607 !!mptcp_send_head(sk));
2609 /* will be ignored by fallback sockets */
2610 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2611 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2613 __mptcp_check_send_data_fin(sk);
2616 static void __mptcp_destroy_sock(struct sock *sk)
2618 struct mptcp_subflow_context *subflow, *tmp;
2619 struct mptcp_sock *msk = mptcp_sk(sk);
2620 LIST_HEAD(conn_list);
2622 pr_debug("msk=%p", msk);
2626 /* be sure to always acquire the join list lock, to sync vs
2627 * mptcp_finish_join().
2629 spin_lock_bh(&msk->join_list_lock);
2630 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2631 spin_unlock_bh(&msk->join_list_lock);
2632 list_splice_init(&msk->conn_list, &conn_list);
2634 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2635 sk_stop_timer(sk, &sk->sk_timer);
2638 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2639 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2640 __mptcp_close_ssk(sk, ssk, subflow);
2643 sk->sk_prot->destroy(sk);
2645 WARN_ON_ONCE(msk->wmem_reserved);
2646 WARN_ON_ONCE(msk->rmem_released);
2647 sk_stream_kill_queues(sk);
2648 xfrm_sk_free_policy(sk);
2650 sk_refcnt_debug_release(sk);
2651 mptcp_dispose_initial_subflow(msk);
2655 static void mptcp_close(struct sock *sk, long timeout)
2657 struct mptcp_subflow_context *subflow;
2658 bool do_cancel_work = false;
2661 sk->sk_shutdown = SHUTDOWN_MASK;
2663 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2664 inet_sk_state_store(sk, TCP_CLOSE);
2668 if (mptcp_close_state(sk))
2669 __mptcp_wr_shutdown(sk);
2671 sk_stream_wait_close(sk, timeout);
2674 /* orphan all the subflows */
2675 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2676 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2677 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2678 bool slow = lock_sock_fast(ssk);
2681 unlock_sock_fast(ssk, slow);
2686 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2687 if (sk->sk_state == TCP_CLOSE) {
2688 __mptcp_destroy_sock(sk);
2689 do_cancel_work = true;
2691 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2695 mptcp_cancel_work(sk);
2697 if (mptcp_sk(sk)->token)
2698 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2703 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2705 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2706 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2707 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2709 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2710 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2713 msk6->saddr = ssk6->saddr;
2714 msk6->flow_label = ssk6->flow_label;
2718 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2719 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2720 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2721 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2722 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2723 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2726 static int mptcp_disconnect(struct sock *sk, int flags)
2728 struct mptcp_subflow_context *subflow;
2729 struct mptcp_sock *msk = mptcp_sk(sk);
2731 mptcp_do_flush_join_list(msk);
2733 mptcp_for_each_subflow(msk, subflow) {
2734 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2737 tcp_disconnect(ssk, flags);
2743 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2744 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2746 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2748 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2752 struct sock *mptcp_sk_clone(const struct sock *sk,
2753 const struct mptcp_options_received *mp_opt,
2754 struct request_sock *req)
2756 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2757 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2758 struct mptcp_sock *msk;
2764 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2765 if (nsk->sk_family == AF_INET6)
2766 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2769 __mptcp_init_sock(nsk);
2771 msk = mptcp_sk(nsk);
2772 msk->local_key = subflow_req->local_key;
2773 msk->token = subflow_req->token;
2774 msk->subflow = NULL;
2775 WRITE_ONCE(msk->fully_established, false);
2777 msk->write_seq = subflow_req->idsn + 1;
2778 msk->snd_nxt = msk->write_seq;
2779 msk->snd_una = msk->write_seq;
2780 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2781 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2783 if (mp_opt->mp_capable) {
2784 msk->can_ack = true;
2785 msk->remote_key = mp_opt->sndr_key;
2786 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2788 WRITE_ONCE(msk->ack_seq, ack_seq);
2789 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2792 sock_reset_flag(nsk, SOCK_RCU_FREE);
2793 /* will be fully established after successful MPC subflow creation */
2794 inet_sk_state_store(nsk, TCP_SYN_RECV);
2796 security_inet_csk_clone(nsk, req);
2797 bh_unlock_sock(nsk);
2799 /* keep a single reference */
2804 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2806 const struct tcp_sock *tp = tcp_sk(ssk);
2808 msk->rcvq_space.copied = 0;
2809 msk->rcvq_space.rtt_us = 0;
2811 msk->rcvq_space.time = tp->tcp_mstamp;
2813 /* initial rcv_space offering made to peer */
2814 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2815 TCP_INIT_CWND * tp->advmss);
2816 if (msk->rcvq_space.space == 0)
2817 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2819 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2822 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2825 struct mptcp_sock *msk = mptcp_sk(sk);
2826 struct socket *listener;
2829 listener = __mptcp_nmpc_socket(msk);
2830 if (WARN_ON_ONCE(!listener)) {
2835 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2836 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2840 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2841 if (sk_is_mptcp(newsk)) {
2842 struct mptcp_subflow_context *subflow;
2843 struct sock *new_mptcp_sock;
2845 subflow = mptcp_subflow_ctx(newsk);
2846 new_mptcp_sock = subflow->conn;
2848 /* is_mptcp should be false if subflow->conn is missing, see
2849 * subflow_syn_recv_sock()
2851 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2852 tcp_sk(newsk)->is_mptcp = 0;
2856 /* acquire the 2nd reference for the owning socket */
2857 sock_hold(new_mptcp_sock);
2858 newsk = new_mptcp_sock;
2859 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2861 MPTCP_INC_STATS(sock_net(sk),
2862 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2868 void mptcp_destroy_common(struct mptcp_sock *msk)
2870 struct sock *sk = (struct sock *)msk;
2872 __mptcp_clear_xmit(sk);
2874 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2875 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2877 skb_rbtree_purge(&msk->out_of_order_queue);
2878 mptcp_token_destroy(msk);
2879 mptcp_pm_free_anno_list(msk);
2882 static void mptcp_destroy(struct sock *sk)
2884 struct mptcp_sock *msk = mptcp_sk(sk);
2886 mptcp_destroy_common(msk);
2887 sk_sockets_allocated_dec(sk);
2890 void __mptcp_data_acked(struct sock *sk)
2892 if (!sock_owned_by_user(sk))
2893 __mptcp_clean_una(sk);
2895 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2897 if (mptcp_pending_data_fin_ack(sk))
2898 mptcp_schedule_work(sk);
2901 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2903 if (!mptcp_send_head(sk))
2906 if (!sock_owned_by_user(sk)) {
2907 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2909 if (xmit_ssk == ssk)
2910 __mptcp_subflow_push_pending(sk, ssk);
2912 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
2914 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2918 /* processes deferred events and flush wmem */
2919 static void mptcp_release_cb(struct sock *sk)
2922 unsigned long flags = 0;
2924 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2925 flags |= BIT(MPTCP_PUSH_PENDING);
2926 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2927 flags |= BIT(MPTCP_RETRANSMIT);
2931 /* the following actions acquire the subflow socket lock
2933 * 1) can't be invoked in atomic scope
2934 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2935 * datapath acquires the msk socket spinlock while helding
2936 * the subflow socket lock
2939 spin_unlock_bh(&sk->sk_lock.slock);
2940 if (flags & BIT(MPTCP_PUSH_PENDING))
2941 __mptcp_push_pending(sk, 0);
2942 if (flags & BIT(MPTCP_RETRANSMIT))
2943 __mptcp_retrans(sk);
2946 spin_lock_bh(&sk->sk_lock.slock);
2949 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2950 __mptcp_clean_una_wakeup(sk);
2951 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2952 __mptcp_error_report(sk);
2954 /* push_pending may touch wmem_reserved, ensure we do the cleanup
2957 __mptcp_update_wmem(sk);
2958 __mptcp_update_rmem(sk);
2961 void mptcp_subflow_process_delegated(struct sock *ssk)
2963 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2964 struct sock *sk = subflow->conn;
2966 mptcp_data_lock(sk);
2967 if (!sock_owned_by_user(sk))
2968 __mptcp_subflow_push_pending(sk, ssk);
2970 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2971 mptcp_data_unlock(sk);
2972 mptcp_subflow_delegated_done(subflow);
2975 static int mptcp_hash(struct sock *sk)
2977 /* should never be called,
2978 * we hash the TCP subflows not the master socket
2984 static void mptcp_unhash(struct sock *sk)
2986 /* called from sk_common_release(), but nothing to do here */
2989 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2991 struct mptcp_sock *msk = mptcp_sk(sk);
2992 struct socket *ssock;
2994 ssock = __mptcp_nmpc_socket(msk);
2995 pr_debug("msk=%p, subflow=%p", msk, ssock);
2996 if (WARN_ON_ONCE(!ssock))
2999 return inet_csk_get_port(ssock->sk, snum);
3002 void mptcp_finish_connect(struct sock *ssk)
3004 struct mptcp_subflow_context *subflow;
3005 struct mptcp_sock *msk;
3009 subflow = mptcp_subflow_ctx(ssk);
3013 pr_debug("msk=%p, token=%u", sk, subflow->token);
3015 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3017 subflow->map_seq = ack_seq;
3018 subflow->map_subflow_seq = 1;
3020 /* the socket is not connected yet, no msk/subflow ops can access/race
3021 * accessing the field below
3023 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3024 WRITE_ONCE(msk->local_key, subflow->local_key);
3025 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3026 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3027 WRITE_ONCE(msk->ack_seq, ack_seq);
3028 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3029 WRITE_ONCE(msk->can_ack, 1);
3030 WRITE_ONCE(msk->snd_una, msk->write_seq);
3032 mptcp_pm_new_connection(msk, ssk, 0);
3034 mptcp_rcv_space_init(msk, ssk);
3037 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3039 write_lock_bh(&sk->sk_callback_lock);
3040 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3041 sk_set_socket(sk, parent);
3042 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3043 write_unlock_bh(&sk->sk_callback_lock);
3046 bool mptcp_finish_join(struct sock *ssk)
3048 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3049 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3050 struct sock *parent = (void *)msk;
3051 struct socket *parent_sock;
3054 pr_debug("msk=%p, subflow=%p", msk, subflow);
3056 /* mptcp socket already closing? */
3057 if (!mptcp_is_fully_established(parent)) {
3058 subflow->reset_reason = MPTCP_RST_EMPTCP;
3062 if (!msk->pm.server_side)
3065 if (!mptcp_pm_allow_new_subflow(msk)) {
3066 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3070 /* active connections are already on conn_list, and we can't acquire
3072 * use the join list lock as synchronization point and double-check
3073 * msk status to avoid racing with __mptcp_destroy_sock()
3075 spin_lock_bh(&msk->join_list_lock);
3076 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3077 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3078 list_add_tail(&subflow->node, &msk->join_list);
3081 spin_unlock_bh(&msk->join_list_lock);
3083 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3087 /* attach to msk socket only after we are sure he will deal with us
3090 parent_sock = READ_ONCE(parent->sk_socket);
3091 if (parent_sock && !ssk->sk_socket)
3092 mptcp_sock_graft(ssk, parent_sock);
3093 subflow->map_seq = READ_ONCE(msk->ack_seq);
3095 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3099 static void mptcp_shutdown(struct sock *sk, int how)
3101 pr_debug("sk=%p, how=%d", sk, how);
3103 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3104 __mptcp_wr_shutdown(sk);
3107 static struct proto mptcp_prot = {
3109 .owner = THIS_MODULE,
3110 .init = mptcp_init_sock,
3111 .disconnect = mptcp_disconnect,
3112 .close = mptcp_close,
3113 .accept = mptcp_accept,
3114 .setsockopt = mptcp_setsockopt,
3115 .getsockopt = mptcp_getsockopt,
3116 .shutdown = mptcp_shutdown,
3117 .destroy = mptcp_destroy,
3118 .sendmsg = mptcp_sendmsg,
3119 .recvmsg = mptcp_recvmsg,
3120 .release_cb = mptcp_release_cb,
3122 .unhash = mptcp_unhash,
3123 .get_port = mptcp_get_port,
3124 .sockets_allocated = &mptcp_sockets_allocated,
3125 .memory_allocated = &tcp_memory_allocated,
3126 .memory_pressure = &tcp_memory_pressure,
3127 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3128 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3129 .sysctl_mem = sysctl_tcp_mem,
3130 .obj_size = sizeof(struct mptcp_sock),
3131 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3132 .no_autobind = true,
3135 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3137 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3138 struct socket *ssock;
3141 lock_sock(sock->sk);
3142 ssock = __mptcp_nmpc_socket(msk);
3148 err = ssock->ops->bind(ssock, uaddr, addr_len);
3150 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3153 release_sock(sock->sk);
3157 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3158 struct mptcp_subflow_context *subflow)
3160 subflow->request_mptcp = 0;
3161 __mptcp_do_fallback(msk);
3164 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3165 int addr_len, int flags)
3167 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3168 struct mptcp_subflow_context *subflow;
3169 struct socket *ssock;
3172 lock_sock(sock->sk);
3173 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3174 /* pending connection or invalid state, let existing subflow
3177 ssock = msk->subflow;
3181 ssock = __mptcp_nmpc_socket(msk);
3187 mptcp_token_destroy(msk);
3188 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3189 subflow = mptcp_subflow_ctx(ssock->sk);
3190 #ifdef CONFIG_TCP_MD5SIG
3191 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3194 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3195 mptcp_subflow_early_fallback(msk, subflow);
3197 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3198 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3199 mptcp_subflow_early_fallback(msk, subflow);
3201 if (likely(!__mptcp_check_fallback(msk)))
3202 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3205 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3206 sock->state = ssock->state;
3208 /* on successful connect, the msk state will be moved to established by
3209 * subflow_finish_connect()
3211 if (!err || err == -EINPROGRESS)
3212 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3214 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3217 release_sock(sock->sk);
3221 static int mptcp_listen(struct socket *sock, int backlog)
3223 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3224 struct socket *ssock;
3227 pr_debug("msk=%p", msk);
3229 lock_sock(sock->sk);
3230 ssock = __mptcp_nmpc_socket(msk);
3236 mptcp_token_destroy(msk);
3237 inet_sk_state_store(sock->sk, TCP_LISTEN);
3238 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3240 err = ssock->ops->listen(ssock, backlog);
3241 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3243 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3246 release_sock(sock->sk);
3250 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3251 int flags, bool kern)
3253 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3254 struct socket *ssock;
3257 pr_debug("msk=%p", msk);
3259 lock_sock(sock->sk);
3260 if (sock->sk->sk_state != TCP_LISTEN)
3263 ssock = __mptcp_nmpc_socket(msk);
3267 clear_bit(MPTCP_DATA_READY, &msk->flags);
3268 sock_hold(ssock->sk);
3269 release_sock(sock->sk);
3271 err = ssock->ops->accept(sock, newsock, flags, kern);
3272 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3273 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3274 struct mptcp_subflow_context *subflow;
3275 struct sock *newsk = newsock->sk;
3279 /* PM/worker can now acquire the first subflow socket
3280 * lock without racing with listener queue cleanup,
3281 * we can notify it, if needed.
3283 * Even if remote has reset the initial subflow by now
3284 * the refcnt is still at least one.
3286 subflow = mptcp_subflow_ctx(msk->first);
3287 list_add(&subflow->node, &msk->conn_list);
3288 sock_hold(msk->first);
3289 if (mptcp_is_fully_established(newsk))
3290 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3292 mptcp_copy_inaddrs(newsk, msk->first);
3293 mptcp_rcv_space_init(msk, msk->first);
3294 mptcp_propagate_sndbuf(newsk, msk->first);
3296 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3297 * This is needed so NOSPACE flag can be set from tcp stack.
3299 mptcp_flush_join_list(msk);
3300 mptcp_for_each_subflow(msk, subflow) {
3301 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3303 if (!ssk->sk_socket)
3304 mptcp_sock_graft(ssk, newsock);
3306 release_sock(newsk);
3309 if (inet_csk_listen_poll(ssock->sk))
3310 set_bit(MPTCP_DATA_READY, &msk->flags);
3311 sock_put(ssock->sk);
3315 release_sock(sock->sk);
3319 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3321 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3325 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3327 struct sock *sk = (struct sock *)msk;
3329 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3330 return EPOLLOUT | EPOLLWRNORM;
3332 if (sk_stream_is_writeable(sk))
3333 return EPOLLOUT | EPOLLWRNORM;
3335 mptcp_set_nospace(sk);
3336 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3337 if (sk_stream_is_writeable(sk))
3338 return EPOLLOUT | EPOLLWRNORM;
3343 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3344 struct poll_table_struct *wait)
3346 struct sock *sk = sock->sk;
3347 struct mptcp_sock *msk;
3352 sock_poll_wait(file, sock, wait);
3354 state = inet_sk_state_load(sk);
3355 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3356 if (state == TCP_LISTEN)
3357 return mptcp_check_readable(msk);
3359 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3360 mask |= mptcp_check_readable(msk);
3361 mask |= mptcp_check_writeable(msk);
3363 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3365 if (sk->sk_shutdown & RCV_SHUTDOWN)
3366 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3368 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3376 static const struct proto_ops mptcp_stream_ops = {
3378 .owner = THIS_MODULE,
3379 .release = inet_release,
3381 .connect = mptcp_stream_connect,
3382 .socketpair = sock_no_socketpair,
3383 .accept = mptcp_stream_accept,
3384 .getname = inet_getname,
3386 .ioctl = inet_ioctl,
3387 .gettstamp = sock_gettstamp,
3388 .listen = mptcp_listen,
3389 .shutdown = inet_shutdown,
3390 .setsockopt = sock_common_setsockopt,
3391 .getsockopt = sock_common_getsockopt,
3392 .sendmsg = inet_sendmsg,
3393 .recvmsg = inet_recvmsg,
3394 .mmap = sock_no_mmap,
3395 .sendpage = inet_sendpage,
3398 static struct inet_protosw mptcp_protosw = {
3399 .type = SOCK_STREAM,
3400 .protocol = IPPROTO_MPTCP,
3401 .prot = &mptcp_prot,
3402 .ops = &mptcp_stream_ops,
3403 .flags = INET_PROTOSW_ICSK,
3406 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3408 struct mptcp_delegated_action *delegated;
3409 struct mptcp_subflow_context *subflow;
3412 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3413 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3414 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3416 bh_lock_sock_nested(ssk);
3417 if (!sock_owned_by_user(ssk) &&
3418 mptcp_subflow_has_delegated_action(subflow))
3419 mptcp_subflow_process_delegated(ssk);
3420 /* ... elsewhere tcp_release_cb_override already processed
3421 * the action or will do at next release_sock().
3422 * In both case must dequeue the subflow here - on the same
3423 * CPU that scheduled it.
3425 bh_unlock_sock(ssk);
3428 if (++work_done == budget)
3432 /* always provide a 0 'work_done' argument, so that napi_complete_done
3433 * will not try accessing the NULL napi->dev ptr
3435 napi_complete_done(napi, 0);
3439 void __init mptcp_proto_init(void)
3441 struct mptcp_delegated_action *delegated;
3444 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3446 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3447 panic("Failed to allocate MPTCP pcpu counter\n");
3449 init_dummy_netdev(&mptcp_napi_dev);
3450 for_each_possible_cpu(cpu) {
3451 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3452 INIT_LIST_HEAD(&delegated->head);
3453 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3455 napi_enable(&delegated->napi);
3458 mptcp_subflow_init();
3462 if (proto_register(&mptcp_prot, 1) != 0)
3463 panic("Failed to register MPTCP proto.\n");
3465 inet_register_protosw(&mptcp_protosw);
3467 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3470 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3471 static const struct proto_ops mptcp_v6_stream_ops = {
3473 .owner = THIS_MODULE,
3474 .release = inet6_release,
3476 .connect = mptcp_stream_connect,
3477 .socketpair = sock_no_socketpair,
3478 .accept = mptcp_stream_accept,
3479 .getname = inet6_getname,
3481 .ioctl = inet6_ioctl,
3482 .gettstamp = sock_gettstamp,
3483 .listen = mptcp_listen,
3484 .shutdown = inet_shutdown,
3485 .setsockopt = sock_common_setsockopt,
3486 .getsockopt = sock_common_getsockopt,
3487 .sendmsg = inet6_sendmsg,
3488 .recvmsg = inet6_recvmsg,
3489 .mmap = sock_no_mmap,
3490 .sendpage = inet_sendpage,
3491 #ifdef CONFIG_COMPAT
3492 .compat_ioctl = inet6_compat_ioctl,
3496 static struct proto mptcp_v6_prot;
3498 static void mptcp_v6_destroy(struct sock *sk)
3501 inet6_destroy_sock(sk);
3504 static struct inet_protosw mptcp_v6_protosw = {
3505 .type = SOCK_STREAM,
3506 .protocol = IPPROTO_MPTCP,
3507 .prot = &mptcp_v6_prot,
3508 .ops = &mptcp_v6_stream_ops,
3509 .flags = INET_PROTOSW_ICSK,
3512 int __init mptcp_proto_v6_init(void)
3516 mptcp_v6_prot = mptcp_prot;
3517 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3518 mptcp_v6_prot.slab = NULL;
3519 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3520 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3522 err = proto_register(&mptcp_v6_prot, 1);
3526 err = inet6_register_protosw(&mptcp_v6_protosw);
3528 proto_unregister(&mptcp_v6_prot);