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;
45 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
48 MPTCP_CMSG_TS = BIT(0),
51 static struct percpu_counter mptcp_sockets_allocated;
53 static void __mptcp_destroy_sock(struct sock *sk);
54 static void __mptcp_check_send_data_fin(struct sock *sk);
56 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
57 static struct net_device mptcp_napi_dev;
59 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
60 * completed yet or has failed, return the subflow socket.
61 * Otherwise return NULL.
63 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
65 if (!msk->subflow || READ_ONCE(msk->can_ack))
71 /* Returns end sequence number of the receiver's advertised window */
72 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
74 return READ_ONCE(msk->wnd_end);
77 static bool mptcp_is_tcpsk(struct sock *sk)
79 struct socket *sock = sk->sk_socket;
81 if (unlikely(sk->sk_prot == &tcp_prot)) {
82 /* we are being invoked after mptcp_accept() has
83 * accepted a non-mp-capable flow: sk is a tcp_sk,
86 * Hand the socket over to tcp so all further socket ops
89 sock->ops = &inet_stream_ops;
91 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
92 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
93 sock->ops = &inet6_stream_ops;
101 static int __mptcp_socket_create(struct mptcp_sock *msk)
103 struct mptcp_subflow_context *subflow;
104 struct sock *sk = (struct sock *)msk;
105 struct socket *ssock;
108 err = mptcp_subflow_create_socket(sk, &ssock);
112 msk->first = ssock->sk;
113 msk->subflow = ssock;
114 subflow = mptcp_subflow_ctx(ssock->sk);
115 list_add(&subflow->node, &msk->conn_list);
116 sock_hold(ssock->sk);
117 subflow->request_mptcp = 1;
118 mptcp_sock_graft(msk->first, sk->sk_socket);
123 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
125 sk_drops_add(sk, skb);
129 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
130 struct sk_buff *from)
135 if (MPTCP_SKB_CB(from)->offset ||
136 !skb_try_coalesce(to, from, &fragstolen, &delta))
139 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
140 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
141 to->len, MPTCP_SKB_CB(from)->end_seq);
142 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
143 kfree_skb_partial(from, fragstolen);
144 atomic_add(delta, &sk->sk_rmem_alloc);
145 sk_mem_charge(sk, delta);
149 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
150 struct sk_buff *from)
152 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
155 return mptcp_try_coalesce((struct sock *)msk, to, from);
158 /* "inspired" by tcp_data_queue_ofo(), main differences:
160 * - don't cope with sacks
162 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
164 struct sock *sk = (struct sock *)msk;
165 struct rb_node **p, *parent;
166 u64 seq, end_seq, max_seq;
167 struct sk_buff *skb1;
169 seq = MPTCP_SKB_CB(skb)->map_seq;
170 end_seq = MPTCP_SKB_CB(skb)->end_seq;
171 max_seq = READ_ONCE(msk->rcv_wnd_sent);
173 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
174 RB_EMPTY_ROOT(&msk->out_of_order_queue));
175 if (after64(end_seq, max_seq)) {
178 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
179 (unsigned long long)end_seq - (unsigned long)max_seq,
180 (unsigned long long)msk->rcv_wnd_sent);
181 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
185 p = &msk->out_of_order_queue.rb_node;
186 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
187 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
188 rb_link_node(&skb->rbnode, NULL, p);
189 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
190 msk->ooo_last_skb = skb;
194 /* with 2 subflows, adding at end of ooo queue is quite likely
195 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
197 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
198 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
199 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
203 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
204 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
205 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
206 parent = &msk->ooo_last_skb->rbnode;
207 p = &parent->rb_right;
211 /* Find place to insert this segment. Handle overlaps on the way. */
215 skb1 = rb_to_skb(parent);
216 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
217 p = &parent->rb_left;
220 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
221 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
222 /* All the bits are present. Drop. */
224 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
227 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
231 * continue traversing
234 /* skb's seq == skb1's seq and skb covers skb1.
235 * Replace skb1 with skb.
237 rb_replace_node(&skb1->rbnode, &skb->rbnode,
238 &msk->out_of_order_queue);
239 mptcp_drop(sk, skb1);
240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
243 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
247 p = &parent->rb_right;
251 /* Insert segment into RB tree. */
252 rb_link_node(&skb->rbnode, parent, p);
253 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
256 /* Remove other segments covered by skb. */
257 while ((skb1 = skb_rb_next(skb)) != NULL) {
258 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
260 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
261 mptcp_drop(sk, skb1);
262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
264 /* If there is no skb after us, we are the last_skb ! */
266 msk->ooo_last_skb = skb;
270 skb_set_owner_r(skb, sk);
273 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
274 struct sk_buff *skb, unsigned int offset,
277 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
278 struct sock *sk = (struct sock *)msk;
279 struct sk_buff *tail;
282 __skb_unlink(skb, &ssk->sk_receive_queue);
287 /* try to fetch required memory from subflow */
288 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
289 int amount = sk_mem_pages(skb->truesize) << SK_MEM_QUANTUM_SHIFT;
291 if (ssk->sk_forward_alloc < amount)
294 ssk->sk_forward_alloc -= amount;
295 sk->sk_forward_alloc += amount;
298 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
300 /* the skb map_seq accounts for the skb offset:
301 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
304 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
305 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
306 MPTCP_SKB_CB(skb)->offset = offset;
307 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
309 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
311 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
312 tail = skb_peek_tail(&sk->sk_receive_queue);
313 if (tail && mptcp_try_coalesce(sk, tail, skb))
316 skb_set_owner_r(skb, sk);
317 __skb_queue_tail(&sk->sk_receive_queue, skb);
319 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
320 mptcp_data_queue_ofo(msk, skb);
324 /* old data, keep it simple and drop the whole pkt, sender
325 * will retransmit as needed, if needed.
327 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
333 static void mptcp_stop_timer(struct sock *sk)
335 struct inet_connection_sock *icsk = inet_csk(sk);
337 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
338 mptcp_sk(sk)->timer_ival = 0;
341 static void mptcp_close_wake_up(struct sock *sk)
343 if (sock_flag(sk, SOCK_DEAD))
346 sk->sk_state_change(sk);
347 if (sk->sk_shutdown == SHUTDOWN_MASK ||
348 sk->sk_state == TCP_CLOSE)
349 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
351 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
354 static bool mptcp_pending_data_fin_ack(struct sock *sk)
356 struct mptcp_sock *msk = mptcp_sk(sk);
358 return !__mptcp_check_fallback(msk) &&
359 ((1 << sk->sk_state) &
360 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
361 msk->write_seq == READ_ONCE(msk->snd_una);
364 static void mptcp_check_data_fin_ack(struct sock *sk)
366 struct mptcp_sock *msk = mptcp_sk(sk);
368 /* Look for an acknowledged DATA_FIN */
369 if (mptcp_pending_data_fin_ack(sk)) {
370 WRITE_ONCE(msk->snd_data_fin_enable, 0);
372 switch (sk->sk_state) {
374 inet_sk_state_store(sk, TCP_FIN_WAIT2);
378 inet_sk_state_store(sk, TCP_CLOSE);
382 mptcp_close_wake_up(sk);
386 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
388 struct mptcp_sock *msk = mptcp_sk(sk);
390 if (READ_ONCE(msk->rcv_data_fin) &&
391 ((1 << sk->sk_state) &
392 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
393 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
395 if (msk->ack_seq == rcv_data_fin_seq) {
397 *seq = rcv_data_fin_seq;
406 static void mptcp_set_datafin_timeout(const struct sock *sk)
408 struct inet_connection_sock *icsk = inet_csk(sk);
410 mptcp_sk(sk)->timer_ival = min(TCP_RTO_MAX,
411 TCP_RTO_MIN << icsk->icsk_retransmits);
414 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
416 long tout = ssk && inet_csk(ssk)->icsk_pending ?
417 inet_csk(ssk)->icsk_timeout - jiffies : 0;
420 tout = mptcp_sk(sk)->timer_ival;
421 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
424 static bool tcp_can_send_ack(const struct sock *ssk)
426 return !((1 << inet_sk_state_load(ssk)) &
427 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
430 static void mptcp_send_ack(struct mptcp_sock *msk)
432 struct mptcp_subflow_context *subflow;
434 mptcp_for_each_subflow(msk, subflow) {
435 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
438 slow = lock_sock_fast(ssk);
439 if (tcp_can_send_ack(ssk))
441 unlock_sock_fast(ssk, slow);
445 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
449 slow = lock_sock_fast(ssk);
450 if (tcp_can_send_ack(ssk))
451 tcp_cleanup_rbuf(ssk, 1);
452 unlock_sock_fast(ssk, slow);
455 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
457 const struct inet_connection_sock *icsk = inet_csk(ssk);
458 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
459 const struct tcp_sock *tp = tcp_sk(ssk);
461 return (ack_pending & ICSK_ACK_SCHED) &&
462 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
463 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
464 (rx_empty && ack_pending &
465 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
468 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
470 int old_space = READ_ONCE(msk->old_wspace);
471 struct mptcp_subflow_context *subflow;
472 struct sock *sk = (struct sock *)msk;
473 int space = __mptcp_space(sk);
474 bool cleanup, rx_empty;
476 cleanup = (space > 0) && (space >= (old_space << 1));
477 rx_empty = !atomic_read(&sk->sk_rmem_alloc);
479 mptcp_for_each_subflow(msk, subflow) {
480 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
482 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
483 mptcp_subflow_cleanup_rbuf(ssk);
487 static bool mptcp_check_data_fin(struct sock *sk)
489 struct mptcp_sock *msk = mptcp_sk(sk);
490 u64 rcv_data_fin_seq;
493 if (__mptcp_check_fallback(msk))
496 /* Need to ack a DATA_FIN received from a peer while this side
497 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
498 * msk->rcv_data_fin was set when parsing the incoming options
499 * at the subflow level and the msk lock was not held, so this
500 * is the first opportunity to act on the DATA_FIN and change
503 * If we are caught up to the sequence number of the incoming
504 * DATA_FIN, send the DATA_ACK now and do state transition. If
505 * not caught up, do nothing and let the recv code send DATA_ACK
509 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
510 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
511 WRITE_ONCE(msk->rcv_data_fin, 0);
513 sk->sk_shutdown |= RCV_SHUTDOWN;
514 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
515 set_bit(MPTCP_DATA_READY, &msk->flags);
517 switch (sk->sk_state) {
518 case TCP_ESTABLISHED:
519 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
522 inet_sk_state_store(sk, TCP_CLOSING);
525 inet_sk_state_store(sk, TCP_CLOSE);
528 /* Other states not expected */
534 mptcp_set_timeout(sk, NULL);
536 mptcp_close_wake_up(sk);
541 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
545 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
546 struct sock *sk = (struct sock *)msk;
547 unsigned int moved = 0;
548 bool more_data_avail;
553 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
555 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
556 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
558 if (unlikely(ssk_rbuf > sk_rbuf)) {
559 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
564 pr_debug("msk=%p ssk=%p", msk, ssk);
567 u32 map_remaining, offset;
568 u32 seq = tp->copied_seq;
572 /* try to move as much data as available */
573 map_remaining = subflow->map_data_len -
574 mptcp_subflow_get_map_offset(subflow);
576 skb = skb_peek(&ssk->sk_receive_queue);
578 /* if no data is found, a racing workqueue/recvmsg
579 * already processed the new data, stop here or we
580 * can enter an infinite loop
587 if (__mptcp_check_fallback(msk)) {
588 /* if we are running under the workqueue, TCP could have
589 * collapsed skbs between dummy map creation and now
590 * be sure to adjust the size
592 map_remaining = skb->len;
593 subflow->map_data_len = skb->len;
596 offset = seq - TCP_SKB_CB(skb)->seq;
597 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
603 if (offset < skb->len) {
604 size_t len = skb->len - offset;
609 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
613 if (WARN_ON_ONCE(map_remaining < len))
617 sk_eat_skb(ssk, skb);
621 WRITE_ONCE(tp->copied_seq, seq);
622 more_data_avail = mptcp_subflow_data_available(ssk);
624 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
628 } while (more_data_avail);
634 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
636 struct sock *sk = (struct sock *)msk;
637 struct sk_buff *skb, *tail;
642 p = rb_first(&msk->out_of_order_queue);
643 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
646 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
650 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
652 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
655 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
659 end_seq = MPTCP_SKB_CB(skb)->end_seq;
660 tail = skb_peek_tail(&sk->sk_receive_queue);
661 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
662 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
664 /* skip overlapping data, if any */
665 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
666 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
668 MPTCP_SKB_CB(skb)->offset += delta;
669 __skb_queue_tail(&sk->sk_receive_queue, skb);
671 msk->ack_seq = end_seq;
677 /* In most cases we will be able to lock the mptcp socket. If its already
678 * owned, we need to defer to the work queue to avoid ABBA deadlock.
680 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
682 struct sock *sk = (struct sock *)msk;
683 unsigned int moved = 0;
685 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
686 __mptcp_ofo_queue(msk);
687 if (unlikely(ssk->sk_err)) {
688 if (!sock_owned_by_user(sk))
689 __mptcp_error_report(sk);
691 set_bit(MPTCP_ERROR_REPORT, &msk->flags);
694 /* If the moves have caught up with the DATA_FIN sequence number
695 * it's time to ack the DATA_FIN and change socket state, but
696 * this is not a good place to change state. Let the workqueue
699 if (mptcp_pending_data_fin(sk, NULL))
700 mptcp_schedule_work(sk);
704 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
706 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
707 struct mptcp_sock *msk = mptcp_sk(sk);
708 int sk_rbuf, ssk_rbuf;
710 /* The peer can send data while we are shutting down this
711 * subflow at msk destruction time, but we must avoid enqueuing
712 * more data to the msk receive queue
714 if (unlikely(subflow->disposable))
717 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
718 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
719 if (unlikely(ssk_rbuf > sk_rbuf))
722 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
723 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
726 /* Wake-up the reader only for in-sequence data */
728 if (move_skbs_to_msk(msk, ssk)) {
729 set_bit(MPTCP_DATA_READY, &msk->flags);
730 sk->sk_data_ready(sk);
732 mptcp_data_unlock(sk);
735 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
737 struct mptcp_subflow_context *subflow;
740 if (likely(list_empty(&msk->join_list)))
743 spin_lock_bh(&msk->join_list_lock);
744 list_for_each_entry(subflow, &msk->join_list, node) {
745 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
747 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
748 if (READ_ONCE(msk->setsockopt_seq) != sseq)
751 list_splice_tail_init(&msk->join_list, &msk->conn_list);
752 spin_unlock_bh(&msk->join_list_lock);
757 void __mptcp_flush_join_list(struct mptcp_sock *msk)
759 if (likely(!mptcp_do_flush_join_list(msk)))
762 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
763 mptcp_schedule_work((struct sock *)msk);
766 static void mptcp_flush_join_list(struct mptcp_sock *msk)
768 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
772 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
775 mptcp_sockopt_sync_all(msk);
778 static bool mptcp_timer_pending(struct sock *sk)
780 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
783 static void mptcp_reset_timer(struct sock *sk)
785 struct inet_connection_sock *icsk = inet_csk(sk);
788 /* prevent rescheduling on close */
789 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
792 /* should never be called with mptcp level timer cleared */
793 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
794 if (WARN_ON_ONCE(!tout))
796 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
799 bool mptcp_schedule_work(struct sock *sk)
801 if (inet_sk_state_load(sk) != TCP_CLOSE &&
802 schedule_work(&mptcp_sk(sk)->work)) {
803 /* each subflow already holds a reference to the sk, and the
804 * workqueue is invoked by a subflow, so sk can't go away here.
812 void mptcp_subflow_eof(struct sock *sk)
814 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
815 mptcp_schedule_work(sk);
818 static void mptcp_check_for_eof(struct mptcp_sock *msk)
820 struct mptcp_subflow_context *subflow;
821 struct sock *sk = (struct sock *)msk;
824 mptcp_for_each_subflow(msk, subflow)
825 receivers += !subflow->rx_eof;
829 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
830 /* hopefully temporary hack: propagate shutdown status
831 * to msk, when all subflows agree on it
833 sk->sk_shutdown |= RCV_SHUTDOWN;
835 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
836 set_bit(MPTCP_DATA_READY, &msk->flags);
837 sk->sk_data_ready(sk);
840 switch (sk->sk_state) {
841 case TCP_ESTABLISHED:
842 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
845 inet_sk_state_store(sk, TCP_CLOSING);
848 inet_sk_state_store(sk, TCP_CLOSE);
853 mptcp_close_wake_up(sk);
856 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
858 struct mptcp_subflow_context *subflow;
859 struct sock *sk = (struct sock *)msk;
861 sock_owned_by_me(sk);
863 mptcp_for_each_subflow(msk, subflow) {
864 if (READ_ONCE(subflow->data_avail))
865 return mptcp_subflow_tcp_sock(subflow);
871 static bool mptcp_skb_can_collapse_to(u64 write_seq,
872 const struct sk_buff *skb,
873 const struct mptcp_ext *mpext)
875 if (!tcp_skb_can_collapse_to(skb))
878 /* can collapse only if MPTCP level sequence is in order and this
879 * mapping has not been xmitted yet
881 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
885 /* we can append data to the given data frag if:
886 * - there is space available in the backing page_frag
887 * - the data frag tail matches the current page_frag free offset
888 * - the data frag end sequence number matches the current write seq
890 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
891 const struct page_frag *pfrag,
892 const struct mptcp_data_frag *df)
894 return df && pfrag->page == df->page &&
895 pfrag->size - pfrag->offset > 0 &&
896 pfrag->offset == (df->offset + df->data_len) &&
897 df->data_seq + df->data_len == msk->write_seq;
900 static int mptcp_wmem_with_overhead(int size)
902 return size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
905 static void __mptcp_wmem_reserve(struct sock *sk, int size)
907 int amount = mptcp_wmem_with_overhead(size);
908 struct mptcp_sock *msk = mptcp_sk(sk);
910 WARN_ON_ONCE(msk->wmem_reserved);
911 if (WARN_ON_ONCE(amount < 0))
914 if (amount <= sk->sk_forward_alloc)
917 /* under memory pressure try to reserve at most a single page
918 * otherwise try to reserve the full estimate and fallback
919 * to a single page before entering the error path
921 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
922 !sk_wmem_schedule(sk, amount)) {
923 if (amount <= PAGE_SIZE)
927 if (!sk_wmem_schedule(sk, amount))
932 msk->wmem_reserved = amount;
933 sk->sk_forward_alloc -= amount;
937 /* we will wait for memory on next allocation */
938 msk->wmem_reserved = -1;
941 static void __mptcp_update_wmem(struct sock *sk)
943 struct mptcp_sock *msk = mptcp_sk(sk);
945 #ifdef CONFIG_LOCKDEP
946 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
949 if (!msk->wmem_reserved)
952 if (msk->wmem_reserved < 0)
953 msk->wmem_reserved = 0;
954 if (msk->wmem_reserved > 0) {
955 sk->sk_forward_alloc += msk->wmem_reserved;
956 msk->wmem_reserved = 0;
960 static bool mptcp_wmem_alloc(struct sock *sk, int size)
962 struct mptcp_sock *msk = mptcp_sk(sk);
964 /* check for pre-existing error condition */
965 if (msk->wmem_reserved < 0)
968 if (msk->wmem_reserved >= size)
972 if (!sk_wmem_schedule(sk, size)) {
973 mptcp_data_unlock(sk);
977 sk->sk_forward_alloc -= size;
978 msk->wmem_reserved += size;
979 mptcp_data_unlock(sk);
982 msk->wmem_reserved -= size;
986 static void mptcp_wmem_uncharge(struct sock *sk, int size)
988 struct mptcp_sock *msk = mptcp_sk(sk);
990 if (msk->wmem_reserved < 0)
991 msk->wmem_reserved = 0;
992 msk->wmem_reserved += size;
995 static void mptcp_mem_reclaim_partial(struct sock *sk)
997 struct mptcp_sock *msk = mptcp_sk(sk);
999 /* if we are experiencing a transint allocation error,
1000 * the forward allocation memory has been already
1003 if (msk->wmem_reserved < 0)
1006 mptcp_data_lock(sk);
1007 sk->sk_forward_alloc += msk->wmem_reserved;
1008 sk_mem_reclaim_partial(sk);
1009 msk->wmem_reserved = sk->sk_forward_alloc;
1010 sk->sk_forward_alloc = 0;
1011 mptcp_data_unlock(sk);
1014 static void dfrag_uncharge(struct sock *sk, int len)
1016 sk_mem_uncharge(sk, len);
1017 sk_wmem_queued_add(sk, -len);
1020 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1022 int len = dfrag->data_len + dfrag->overhead;
1024 list_del(&dfrag->list);
1025 dfrag_uncharge(sk, len);
1026 put_page(dfrag->page);
1029 static void __mptcp_clean_una(struct sock *sk)
1031 struct mptcp_sock *msk = mptcp_sk(sk);
1032 struct mptcp_data_frag *dtmp, *dfrag;
1033 bool cleaned = false;
1036 /* on fallback we just need to ignore snd_una, as this is really
1039 if (__mptcp_check_fallback(msk))
1040 msk->snd_una = READ_ONCE(msk->snd_nxt);
1042 snd_una = msk->snd_una;
1043 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1044 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1047 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1049 dfrag_clear(sk, dfrag);
1053 dfrag = mptcp_rtx_head(sk);
1054 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1055 u64 delta = snd_una - dfrag->data_seq;
1057 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1060 dfrag->data_seq += delta;
1061 dfrag->offset += delta;
1062 dfrag->data_len -= delta;
1063 dfrag->already_sent -= delta;
1065 dfrag_uncharge(sk, delta);
1071 if (tcp_under_memory_pressure(sk)) {
1072 __mptcp_update_wmem(sk);
1073 sk_mem_reclaim_partial(sk);
1077 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1078 if (msk->timer_ival && !mptcp_data_fin_enabled(msk))
1079 mptcp_stop_timer(sk);
1081 mptcp_reset_timer(sk);
1085 static void __mptcp_clean_una_wakeup(struct sock *sk)
1087 #ifdef CONFIG_LOCKDEP
1088 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
1090 __mptcp_clean_una(sk);
1091 mptcp_write_space(sk);
1094 static void mptcp_clean_una_wakeup(struct sock *sk)
1096 mptcp_data_lock(sk);
1097 __mptcp_clean_una_wakeup(sk);
1098 mptcp_data_unlock(sk);
1101 static void mptcp_enter_memory_pressure(struct sock *sk)
1103 struct mptcp_subflow_context *subflow;
1104 struct mptcp_sock *msk = mptcp_sk(sk);
1107 sk_stream_moderate_sndbuf(sk);
1108 mptcp_for_each_subflow(msk, subflow) {
1109 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1112 tcp_enter_memory_pressure(ssk);
1113 sk_stream_moderate_sndbuf(ssk);
1118 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1121 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1123 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1124 pfrag, sk->sk_allocation)))
1127 mptcp_enter_memory_pressure(sk);
1131 static struct mptcp_data_frag *
1132 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1135 int offset = ALIGN(orig_offset, sizeof(long));
1136 struct mptcp_data_frag *dfrag;
1138 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1139 dfrag->data_len = 0;
1140 dfrag->data_seq = msk->write_seq;
1141 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1142 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1143 dfrag->already_sent = 0;
1144 dfrag->page = pfrag->page;
1149 struct mptcp_sendmsg_info {
1157 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1160 u64 window_end = mptcp_wnd_end(msk);
1162 if (__mptcp_check_fallback(msk))
1165 if (!before64(data_seq + avail_size, window_end)) {
1166 u64 allowed_size = window_end - data_seq;
1168 return min_t(unsigned int, allowed_size, avail_size);
1174 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1176 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1180 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1184 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1186 struct sk_buff *skb;
1188 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1190 if (likely(__mptcp_add_ext(skb, gfp))) {
1191 skb_reserve(skb, MAX_TCP_HEADER);
1192 skb->reserved_tailroom = skb->end - skb->tail;
1197 mptcp_enter_memory_pressure(sk);
1202 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1204 struct sk_buff *skb;
1206 if (ssk->sk_tx_skb_cache) {
1207 skb = ssk->sk_tx_skb_cache;
1208 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1209 !__mptcp_add_ext(skb, gfp)))
1214 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1218 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1219 ssk->sk_tx_skb_cache = skb;
1226 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1228 return !ssk->sk_tx_skb_cache &&
1229 tcp_under_memory_pressure(sk);
1232 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1234 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1235 mptcp_mem_reclaim_partial(sk);
1236 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1239 /* note: this always recompute the csum on the whole skb, even
1240 * if we just appended a single frag. More status info needed
1242 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1244 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1245 __wsum csum = ~csum_unfold(mpext->csum);
1246 int offset = skb->len - added;
1248 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1251 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1252 struct mptcp_data_frag *dfrag,
1253 struct mptcp_sendmsg_info *info)
1255 u64 data_seq = dfrag->data_seq + info->sent;
1256 struct mptcp_sock *msk = mptcp_sk(sk);
1257 bool zero_window_probe = false;
1258 struct mptcp_ext *mpext = NULL;
1259 struct sk_buff *skb, *tail;
1260 bool can_collapse = false;
1265 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1266 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1268 /* compute send limit */
1269 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1270 avail_size = info->size_goal;
1271 skb = tcp_write_queue_tail(ssk);
1273 /* Limit the write to the size available in the
1274 * current skb, if any, so that we create at most a new skb.
1275 * Explicitly tells TCP internals to avoid collapsing on later
1276 * queue management operation, to avoid breaking the ext <->
1277 * SSN association set here
1279 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1280 can_collapse = (info->size_goal - skb->len > 0) &&
1281 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1282 if (!can_collapse) {
1283 TCP_SKB_CB(skb)->eor = 1;
1285 size_bias = skb->len;
1286 avail_size = info->size_goal - skb->len;
1290 /* Zero window and all data acked? Probe. */
1291 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1292 if (avail_size == 0) {
1293 u64 snd_una = READ_ONCE(msk->snd_una);
1295 if (skb || snd_una != msk->snd_nxt)
1297 zero_window_probe = true;
1298 data_seq = snd_una - 1;
1302 if (WARN_ON_ONCE(info->sent > info->limit ||
1303 info->limit > dfrag->data_len))
1306 ret = info->limit - info->sent;
1307 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1308 dfrag->page, dfrag->offset + info->sent, &ret);
1310 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1314 /* if the tail skb is still the cached one, collapsing really happened.
1317 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1318 mpext->data_len += ret;
1319 WARN_ON_ONCE(!can_collapse);
1320 WARN_ON_ONCE(zero_window_probe);
1324 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1325 if (WARN_ON_ONCE(!mpext)) {
1326 /* should never reach here, stream corrupted */
1330 memset(mpext, 0, sizeof(*mpext));
1331 mpext->data_seq = data_seq;
1332 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1333 mpext->data_len = ret;
1337 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1338 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1341 if (zero_window_probe) {
1342 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1344 if (READ_ONCE(msk->csum_enabled))
1345 mptcp_update_data_checksum(tail, ret);
1346 tcp_push_pending_frames(ssk);
1350 if (READ_ONCE(msk->csum_enabled))
1351 mptcp_update_data_checksum(tail, ret);
1352 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1356 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1357 sizeof(struct tcphdr) - \
1358 MAX_TCP_OPTION_SPACE - \
1359 sizeof(struct ipv6hdr) - \
1360 sizeof(struct frag_hdr))
1362 struct subflow_send_info {
1367 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1369 struct subflow_send_info send_info[2];
1370 struct mptcp_subflow_context *subflow;
1371 int i, nr_active = 0;
1376 sock_owned_by_me((struct sock *)msk);
1378 if (__mptcp_check_fallback(msk)) {
1381 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1384 /* re-use last subflow, if the burst allow that */
1385 if (msk->last_snd && msk->snd_burst > 0 &&
1386 sk_stream_memory_free(msk->last_snd) &&
1387 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1388 return msk->last_snd;
1390 /* pick the subflow with the lower wmem/wspace ratio */
1391 for (i = 0; i < 2; ++i) {
1392 send_info[i].ssk = NULL;
1393 send_info[i].ratio = -1;
1395 mptcp_for_each_subflow(msk, subflow) {
1396 trace_mptcp_subflow_get_send(subflow);
1397 ssk = mptcp_subflow_tcp_sock(subflow);
1398 if (!mptcp_subflow_active(subflow))
1401 nr_active += !subflow->backup;
1402 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1405 pace = READ_ONCE(ssk->sk_pacing_rate);
1409 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1411 if (ratio < send_info[subflow->backup].ratio) {
1412 send_info[subflow->backup].ssk = ssk;
1413 send_info[subflow->backup].ratio = ratio;
1417 /* pick the best backup if no other subflow is active */
1419 send_info[0].ssk = send_info[1].ssk;
1421 if (send_info[0].ssk) {
1422 msk->last_snd = send_info[0].ssk;
1423 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1424 tcp_sk(msk->last_snd)->snd_wnd);
1425 return msk->last_snd;
1431 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1432 struct mptcp_sendmsg_info *info)
1434 mptcp_set_timeout(sk, ssk);
1435 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1439 static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1441 struct sock *prev_ssk = NULL, *ssk = NULL;
1442 struct mptcp_sock *msk = mptcp_sk(sk);
1443 struct mptcp_sendmsg_info info = {
1446 struct mptcp_data_frag *dfrag;
1447 int len, copied = 0;
1449 while ((dfrag = mptcp_send_head(sk))) {
1450 info.sent = dfrag->already_sent;
1451 info.limit = dfrag->data_len;
1452 len = dfrag->data_len - dfrag->already_sent;
1457 mptcp_flush_join_list(msk);
1458 ssk = mptcp_subflow_get_send(msk);
1460 /* try to keep the subflow socket lock across
1461 * consecutive xmit on the same socket
1463 if (ssk != prev_ssk && prev_ssk)
1464 mptcp_push_release(sk, prev_ssk, &info);
1468 if (ssk != prev_ssk || !prev_ssk)
1471 /* keep it simple and always provide a new skb for the
1472 * subflow, even if we will not use it when collapsing
1473 * on the pending one
1475 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1476 mptcp_push_release(sk, ssk, &info);
1480 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1482 mptcp_push_release(sk, ssk, &info);
1487 dfrag->already_sent += ret;
1488 msk->snd_nxt += ret;
1489 msk->snd_burst -= ret;
1490 msk->tx_pending_data -= ret;
1494 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1497 /* at this point we held the socket lock for the last subflow we used */
1499 mptcp_push_release(sk, ssk, &info);
1503 /* start the timer, if it's not pending */
1504 if (!mptcp_timer_pending(sk))
1505 mptcp_reset_timer(sk);
1506 __mptcp_check_send_data_fin(sk);
1510 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1512 struct mptcp_sock *msk = mptcp_sk(sk);
1513 struct mptcp_sendmsg_info info;
1514 struct mptcp_data_frag *dfrag;
1515 struct sock *xmit_ssk;
1516 int len, copied = 0;
1520 while ((dfrag = mptcp_send_head(sk))) {
1521 info.sent = dfrag->already_sent;
1522 info.limit = dfrag->data_len;
1523 len = dfrag->data_len - dfrag->already_sent;
1527 /* the caller already invoked the packet scheduler,
1528 * check for a different subflow usage only after
1529 * spooling the first chunk of data
1531 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1534 if (xmit_ssk != ssk) {
1535 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1539 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1540 __mptcp_update_wmem(sk);
1541 sk_mem_reclaim_partial(sk);
1543 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1546 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1551 dfrag->already_sent += ret;
1552 msk->snd_nxt += ret;
1553 msk->snd_burst -= ret;
1554 msk->tx_pending_data -= ret;
1559 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1563 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1564 * not going to flush it via release_sock()
1566 __mptcp_update_wmem(sk);
1568 mptcp_set_timeout(sk, ssk);
1569 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1571 if (!mptcp_timer_pending(sk))
1572 mptcp_reset_timer(sk);
1574 if (msk->snd_data_fin_enable &&
1575 msk->snd_nxt + 1 == msk->write_seq)
1576 mptcp_schedule_work(sk);
1580 static void mptcp_set_nospace(struct sock *sk)
1582 /* enable autotune */
1583 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1585 /* will be cleared on avail space */
1586 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1589 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1591 struct mptcp_sock *msk = mptcp_sk(sk);
1592 struct page_frag *pfrag;
1597 /* we don't support FASTOPEN yet */
1598 if (msg->msg_flags & MSG_FASTOPEN)
1601 /* silently ignore everything else */
1602 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1604 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1606 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1608 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1609 ret = sk_stream_wait_connect(sk, &timeo);
1614 pfrag = sk_page_frag(sk);
1616 while (msg_data_left(msg)) {
1617 int total_ts, frag_truesize = 0;
1618 struct mptcp_data_frag *dfrag;
1619 bool dfrag_collapsed;
1620 size_t psize, offset;
1622 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1627 /* reuse tail pfrag, if possible, or carve a new one from the
1630 dfrag = mptcp_pending_tail(sk);
1631 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1632 if (!dfrag_collapsed) {
1633 if (!sk_stream_memory_free(sk))
1634 goto wait_for_memory;
1636 if (!mptcp_page_frag_refill(sk, pfrag))
1637 goto wait_for_memory;
1639 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1640 frag_truesize = dfrag->overhead;
1643 /* we do not bound vs wspace, to allow a single packet.
1644 * memory accounting will prevent execessive memory usage
1647 offset = dfrag->offset + dfrag->data_len;
1648 psize = pfrag->size - offset;
1649 psize = min_t(size_t, psize, msg_data_left(msg));
1650 total_ts = psize + frag_truesize;
1652 if (!mptcp_wmem_alloc(sk, total_ts))
1653 goto wait_for_memory;
1655 if (copy_page_from_iter(dfrag->page, offset, psize,
1656 &msg->msg_iter) != psize) {
1657 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1662 /* data successfully copied into the write queue */
1664 dfrag->data_len += psize;
1665 frag_truesize += psize;
1666 pfrag->offset += frag_truesize;
1667 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1668 msk->tx_pending_data += psize;
1670 /* charge data on mptcp pending queue to the msk socket
1671 * Note: we charge such data both to sk and ssk
1673 sk_wmem_queued_add(sk, frag_truesize);
1674 if (!dfrag_collapsed) {
1675 get_page(dfrag->page);
1676 list_add_tail(&dfrag->list, &msk->rtx_queue);
1677 if (!msk->first_pending)
1678 WRITE_ONCE(msk->first_pending, dfrag);
1680 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1681 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1687 mptcp_set_nospace(sk);
1688 __mptcp_push_pending(sk, msg->msg_flags);
1689 ret = sk_stream_wait_memory(sk, &timeo);
1695 __mptcp_push_pending(sk, msg->msg_flags);
1699 return copied ? : ret;
1702 static void mptcp_wait_data(struct sock *sk, long *timeo)
1704 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1705 struct mptcp_sock *msk = mptcp_sk(sk);
1707 add_wait_queue(sk_sleep(sk), &wait);
1708 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1710 sk_wait_event(sk, timeo,
1711 test_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1713 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1714 remove_wait_queue(sk_sleep(sk), &wait);
1717 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1719 size_t len, int flags,
1720 struct scm_timestamping_internal *tss,
1723 struct sk_buff *skb, *tmp;
1726 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1727 u32 offset = MPTCP_SKB_CB(skb)->offset;
1728 u32 data_len = skb->len - offset;
1729 u32 count = min_t(size_t, len - copied, data_len);
1732 if (!(flags & MSG_TRUNC)) {
1733 err = skb_copy_datagram_msg(skb, offset, msg, count);
1734 if (unlikely(err < 0)) {
1741 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1742 tcp_update_recv_tstamps(skb, tss);
1743 *cmsg_flags |= MPTCP_CMSG_TS;
1748 if (count < data_len) {
1749 if (!(flags & MSG_PEEK))
1750 MPTCP_SKB_CB(skb)->offset += count;
1754 if (!(flags & MSG_PEEK)) {
1755 /* we will bulk release the skb memory later */
1756 skb->destructor = NULL;
1757 msk->rmem_released += skb->truesize;
1758 __skb_unlink(skb, &msk->receive_queue);
1769 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1771 * Only difference: Use highest rtt estimate of the subflows in use.
1773 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1775 struct mptcp_subflow_context *subflow;
1776 struct sock *sk = (struct sock *)msk;
1777 u32 time, advmss = 1;
1780 sock_owned_by_me(sk);
1785 msk->rcvq_space.copied += copied;
1787 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1788 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1790 rtt_us = msk->rcvq_space.rtt_us;
1791 if (rtt_us && time < (rtt_us >> 3))
1795 mptcp_for_each_subflow(msk, subflow) {
1796 const struct tcp_sock *tp;
1800 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1802 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1803 sf_advmss = READ_ONCE(tp->advmss);
1805 rtt_us = max(sf_rtt_us, rtt_us);
1806 advmss = max(sf_advmss, advmss);
1809 msk->rcvq_space.rtt_us = rtt_us;
1810 if (time < (rtt_us >> 3) || rtt_us == 0)
1813 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1816 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1817 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1821 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1823 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1825 do_div(grow, msk->rcvq_space.space);
1826 rcvwin += (grow << 1);
1828 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1829 while (tcp_win_from_space(sk, rcvmem) < advmss)
1832 do_div(rcvwin, advmss);
1833 rcvbuf = min_t(u64, rcvwin * rcvmem,
1834 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1836 if (rcvbuf > sk->sk_rcvbuf) {
1839 window_clamp = tcp_win_from_space(sk, rcvbuf);
1840 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1842 /* Make subflows follow along. If we do not do this, we
1843 * get drops at subflow level if skbs can't be moved to
1844 * the mptcp rx queue fast enough (announced rcv_win can
1845 * exceed ssk->sk_rcvbuf).
1847 mptcp_for_each_subflow(msk, subflow) {
1851 ssk = mptcp_subflow_tcp_sock(subflow);
1852 slow = lock_sock_fast(ssk);
1853 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1854 tcp_sk(ssk)->window_clamp = window_clamp;
1855 tcp_cleanup_rbuf(ssk, 1);
1856 unlock_sock_fast(ssk, slow);
1861 msk->rcvq_space.space = msk->rcvq_space.copied;
1863 msk->rcvq_space.copied = 0;
1864 msk->rcvq_space.time = mstamp;
1867 static void __mptcp_update_rmem(struct sock *sk)
1869 struct mptcp_sock *msk = mptcp_sk(sk);
1871 if (!msk->rmem_released)
1874 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1875 sk_mem_uncharge(sk, msk->rmem_released);
1876 msk->rmem_released = 0;
1879 static void __mptcp_splice_receive_queue(struct sock *sk)
1881 struct mptcp_sock *msk = mptcp_sk(sk);
1883 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1886 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1888 struct sock *sk = (struct sock *)msk;
1889 unsigned int moved = 0;
1892 mptcp_flush_join_list(msk);
1894 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1897 /* we can have data pending in the subflows only if the msk
1898 * receive buffer was full at subflow_data_ready() time,
1899 * that is an unlikely slow path.
1904 slowpath = lock_sock_fast(ssk);
1905 mptcp_data_lock(sk);
1906 __mptcp_update_rmem(sk);
1907 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1908 mptcp_data_unlock(sk);
1910 if (unlikely(ssk->sk_err))
1911 __mptcp_error_report(sk);
1912 unlock_sock_fast(ssk, slowpath);
1915 /* acquire the data lock only if some input data is pending */
1917 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1918 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1919 mptcp_data_lock(sk);
1920 __mptcp_update_rmem(sk);
1921 ret |= __mptcp_ofo_queue(msk);
1922 __mptcp_splice_receive_queue(sk);
1923 mptcp_data_unlock(sk);
1926 mptcp_check_data_fin((struct sock *)msk);
1927 return !skb_queue_empty(&msk->receive_queue);
1930 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1931 int nonblock, int flags, int *addr_len)
1933 struct mptcp_sock *msk = mptcp_sk(sk);
1934 struct scm_timestamping_internal tss;
1935 int copied = 0, cmsg_flags = 0;
1939 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
1940 if (unlikely(flags & MSG_ERRQUEUE))
1941 return inet_recv_error(sk, msg, len, addr_len);
1943 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1944 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1949 timeo = sock_rcvtimeo(sk, nonblock);
1951 len = min_t(size_t, len, INT_MAX);
1952 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1954 while (copied < len) {
1957 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
1958 if (unlikely(bytes_read < 0)) {
1960 copied = bytes_read;
1964 copied += bytes_read;
1966 /* be sure to advertise window change */
1967 mptcp_cleanup_rbuf(msk);
1969 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
1972 /* only the master socket status is relevant here. The exit
1973 * conditions mirror closely tcp_recvmsg()
1975 if (copied >= target)
1980 sk->sk_state == TCP_CLOSE ||
1981 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1983 signal_pending(current))
1987 copied = sock_error(sk);
1991 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1992 mptcp_check_for_eof(msk);
1994 if (sk->sk_shutdown & RCV_SHUTDOWN) {
1995 /* race breaker: the shutdown could be after the
1996 * previous receive queue check
1998 if (__mptcp_move_skbs(msk))
2003 if (sk->sk_state == TCP_CLOSE) {
2013 if (signal_pending(current)) {
2014 copied = sock_intr_errno(timeo);
2019 pr_debug("block timeout %ld", timeo);
2020 mptcp_wait_data(sk, &timeo);
2023 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2024 skb_queue_empty(&msk->receive_queue)) {
2025 /* entire backlog drained, clear DATA_READY. */
2026 clear_bit(MPTCP_DATA_READY, &msk->flags);
2028 /* .. race-breaker: ssk might have gotten new data
2029 * after last __mptcp_move_skbs() returned false.
2031 if (unlikely(__mptcp_move_skbs(msk)))
2032 set_bit(MPTCP_DATA_READY, &msk->flags);
2036 if (cmsg_flags && copied >= 0) {
2037 if (cmsg_flags & MPTCP_CMSG_TS)
2038 tcp_recv_timestamp(msg, sk, &tss);
2041 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2042 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2043 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2044 if (!(flags & MSG_PEEK))
2045 mptcp_rcv_space_adjust(msk, copied);
2051 static void mptcp_retransmit_timer(struct timer_list *t)
2053 struct inet_connection_sock *icsk = from_timer(icsk, t,
2054 icsk_retransmit_timer);
2055 struct sock *sk = &icsk->icsk_inet.sk;
2056 struct mptcp_sock *msk = mptcp_sk(sk);
2059 if (!sock_owned_by_user(sk)) {
2060 /* we need a process context to retransmit */
2061 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2062 mptcp_schedule_work(sk);
2064 /* delegate our work to tcp_release_cb() */
2065 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2071 static void mptcp_timeout_timer(struct timer_list *t)
2073 struct sock *sk = from_timer(sk, t, sk_timer);
2075 mptcp_schedule_work(sk);
2079 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2082 * A backup subflow is returned only if that is the only kind available.
2084 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2086 struct mptcp_subflow_context *subflow;
2087 struct sock *backup = NULL;
2089 sock_owned_by_me((const struct sock *)msk);
2091 if (__mptcp_check_fallback(msk))
2094 mptcp_for_each_subflow(msk, subflow) {
2095 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2097 if (!mptcp_subflow_active(subflow))
2100 /* still data outstanding at TCP level? Don't retransmit. */
2101 if (!tcp_write_queue_empty(ssk)) {
2102 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2107 if (subflow->backup) {
2119 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2122 iput(SOCK_INODE(msk->subflow));
2123 msk->subflow = NULL;
2127 /* subflow sockets can be either outgoing (connect) or incoming
2130 * Outgoing subflows use in-kernel sockets.
2131 * Incoming subflows do not have their own 'struct socket' allocated,
2132 * so we need to use tcp_close() after detaching them from the mptcp
2135 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2136 struct mptcp_subflow_context *subflow)
2138 struct mptcp_sock *msk = mptcp_sk(sk);
2140 list_del(&subflow->node);
2142 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2144 /* if we are invoked by the msk cleanup code, the subflow is
2150 subflow->disposable = 1;
2152 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2153 * the ssk has been already destroyed, we just need to release the
2154 * reference owned by msk;
2156 if (!inet_csk(ssk)->icsk_ulp_ops) {
2157 kfree_rcu(subflow, rcu);
2159 /* otherwise tcp will dispose of the ssk and subflow ctx */
2160 __tcp_close(ssk, 0);
2162 /* close acquired an extra ref */
2169 if (ssk == msk->last_snd)
2170 msk->last_snd = NULL;
2172 if (ssk == msk->first)
2175 if (msk->subflow && ssk == msk->subflow->sk)
2176 mptcp_dispose_initial_subflow(msk);
2179 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2180 struct mptcp_subflow_context *subflow)
2182 if (sk->sk_state == TCP_ESTABLISHED)
2183 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2184 __mptcp_close_ssk(sk, ssk, subflow);
2187 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2192 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2194 struct mptcp_subflow_context *subflow, *tmp;
2198 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2199 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2201 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2204 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2205 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2208 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2212 static bool mptcp_check_close_timeout(const struct sock *sk)
2214 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2215 struct mptcp_subflow_context *subflow;
2217 if (delta >= TCP_TIMEWAIT_LEN)
2220 /* if all subflows are in closed status don't bother with additional
2223 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2224 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2231 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2233 struct mptcp_subflow_context *subflow, *tmp;
2234 struct sock *sk = &msk->sk.icsk_inet.sk;
2236 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2239 mptcp_token_destroy(msk);
2241 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2242 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2245 slow = lock_sock_fast(tcp_sk);
2246 if (tcp_sk->sk_state != TCP_CLOSE) {
2247 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2248 tcp_set_state(tcp_sk, TCP_CLOSE);
2250 unlock_sock_fast(tcp_sk, slow);
2253 inet_sk_state_store(sk, TCP_CLOSE);
2254 sk->sk_shutdown = SHUTDOWN_MASK;
2255 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2256 set_bit(MPTCP_DATA_READY, &msk->flags);
2257 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2259 mptcp_close_wake_up(sk);
2262 static void __mptcp_retrans(struct sock *sk)
2264 struct mptcp_sock *msk = mptcp_sk(sk);
2265 struct mptcp_sendmsg_info info = {};
2266 struct mptcp_data_frag *dfrag;
2271 mptcp_clean_una_wakeup(sk);
2272 dfrag = mptcp_rtx_head(sk);
2274 if (mptcp_data_fin_enabled(msk)) {
2275 struct inet_connection_sock *icsk = inet_csk(sk);
2277 icsk->icsk_retransmits++;
2278 mptcp_set_datafin_timeout(sk);
2279 mptcp_send_ack(msk);
2287 ssk = mptcp_subflow_get_retrans(msk);
2293 /* limit retransmission to the bytes already sent on some subflows */
2295 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2296 while (info.sent < info.limit) {
2297 if (!mptcp_alloc_tx_skb(sk, ssk))
2300 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2304 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2309 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2310 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2314 mptcp_set_timeout(sk, ssk);
2318 if (!mptcp_timer_pending(sk))
2319 mptcp_reset_timer(sk);
2322 static void mptcp_worker(struct work_struct *work)
2324 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2325 struct sock *sk = &msk->sk.icsk_inet.sk;
2329 state = sk->sk_state;
2330 if (unlikely(state == TCP_CLOSE))
2333 mptcp_check_data_fin_ack(sk);
2334 mptcp_flush_join_list(msk);
2336 mptcp_check_fastclose(msk);
2339 mptcp_pm_nl_work(msk);
2341 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2342 mptcp_check_for_eof(msk);
2344 __mptcp_check_send_data_fin(sk);
2345 mptcp_check_data_fin(sk);
2347 /* There is no point in keeping around an orphaned sk timedout or
2348 * closed, but we need the msk around to reply to incoming DATA_FIN,
2349 * even if it is orphaned and in FIN_WAIT2 state
2351 if (sock_flag(sk, SOCK_DEAD) &&
2352 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2353 inet_sk_state_store(sk, TCP_CLOSE);
2354 __mptcp_destroy_sock(sk);
2358 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2359 __mptcp_close_subflow(msk);
2361 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2362 __mptcp_retrans(sk);
2369 static int __mptcp_init_sock(struct sock *sk)
2371 struct mptcp_sock *msk = mptcp_sk(sk);
2373 spin_lock_init(&msk->join_list_lock);
2375 INIT_LIST_HEAD(&msk->conn_list);
2376 INIT_LIST_HEAD(&msk->join_list);
2377 INIT_LIST_HEAD(&msk->rtx_queue);
2378 INIT_WORK(&msk->work, mptcp_worker);
2379 __skb_queue_head_init(&msk->receive_queue);
2380 msk->out_of_order_queue = RB_ROOT;
2381 msk->first_pending = NULL;
2382 msk->wmem_reserved = 0;
2383 msk->rmem_released = 0;
2384 msk->tx_pending_data = 0;
2387 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2388 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2390 mptcp_pm_data_init(msk);
2392 /* re-use the csk retrans timer for MPTCP-level retrans */
2393 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2394 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2399 static int mptcp_init_sock(struct sock *sk)
2401 struct inet_connection_sock *icsk = inet_csk(sk);
2402 struct net *net = sock_net(sk);
2405 ret = __mptcp_init_sock(sk);
2409 if (!mptcp_is_enabled(net))
2410 return -ENOPROTOOPT;
2412 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2415 ret = __mptcp_socket_create(mptcp_sk(sk));
2419 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2420 * propagate the correct value
2422 tcp_assign_congestion_control(sk);
2423 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2425 /* no need to keep a reference to the ops, the name will suffice */
2426 tcp_cleanup_congestion_control(sk);
2427 icsk->icsk_ca_ops = NULL;
2429 sk_sockets_allocated_inc(sk);
2430 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2431 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2436 static void __mptcp_clear_xmit(struct sock *sk)
2438 struct mptcp_sock *msk = mptcp_sk(sk);
2439 struct mptcp_data_frag *dtmp, *dfrag;
2441 WRITE_ONCE(msk->first_pending, NULL);
2442 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2443 dfrag_clear(sk, dfrag);
2446 static void mptcp_cancel_work(struct sock *sk)
2448 struct mptcp_sock *msk = mptcp_sk(sk);
2450 if (cancel_work_sync(&msk->work))
2454 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2458 switch (ssk->sk_state) {
2460 if (!(how & RCV_SHUTDOWN))
2464 tcp_disconnect(ssk, O_NONBLOCK);
2467 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2468 pr_debug("Fallback");
2469 ssk->sk_shutdown |= how;
2470 tcp_shutdown(ssk, how);
2472 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2473 mptcp_set_timeout(sk, ssk);
2475 if (!mptcp_timer_pending(sk))
2476 mptcp_reset_timer(sk);
2484 static const unsigned char new_state[16] = {
2485 /* current state: new state: action: */
2486 [0 /* (Invalid) */] = TCP_CLOSE,
2487 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2488 [TCP_SYN_SENT] = TCP_CLOSE,
2489 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2490 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2491 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2492 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2493 [TCP_CLOSE] = TCP_CLOSE,
2494 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2495 [TCP_LAST_ACK] = TCP_LAST_ACK,
2496 [TCP_LISTEN] = TCP_CLOSE,
2497 [TCP_CLOSING] = TCP_CLOSING,
2498 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2501 static int mptcp_close_state(struct sock *sk)
2503 int next = (int)new_state[sk->sk_state];
2504 int ns = next & TCP_STATE_MASK;
2506 inet_sk_state_store(sk, ns);
2508 return next & TCP_ACTION_FIN;
2511 static void __mptcp_check_send_data_fin(struct sock *sk)
2513 struct mptcp_subflow_context *subflow;
2514 struct mptcp_sock *msk = mptcp_sk(sk);
2516 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2517 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2518 msk->snd_nxt, msk->write_seq);
2520 /* we still need to enqueue subflows or not really shutting down,
2523 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2524 mptcp_send_head(sk))
2527 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2529 /* fallback socket will not get data_fin/ack, can move to the next
2532 if (__mptcp_check_fallback(msk)) {
2533 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2534 inet_sk_state_store(sk, TCP_CLOSE);
2535 mptcp_close_wake_up(sk);
2536 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2537 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2541 mptcp_flush_join_list(msk);
2542 mptcp_for_each_subflow(msk, subflow) {
2543 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2545 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2549 static void __mptcp_wr_shutdown(struct sock *sk)
2551 struct mptcp_sock *msk = mptcp_sk(sk);
2553 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2554 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2555 !!mptcp_send_head(sk));
2557 /* will be ignored by fallback sockets */
2558 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2559 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2561 __mptcp_check_send_data_fin(sk);
2564 static void __mptcp_destroy_sock(struct sock *sk)
2566 struct mptcp_subflow_context *subflow, *tmp;
2567 struct mptcp_sock *msk = mptcp_sk(sk);
2568 LIST_HEAD(conn_list);
2570 pr_debug("msk=%p", msk);
2574 /* be sure to always acquire the join list lock, to sync vs
2575 * mptcp_finish_join().
2577 spin_lock_bh(&msk->join_list_lock);
2578 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2579 spin_unlock_bh(&msk->join_list_lock);
2580 list_splice_init(&msk->conn_list, &conn_list);
2582 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2583 sk_stop_timer(sk, &sk->sk_timer);
2586 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2587 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2588 __mptcp_close_ssk(sk, ssk, subflow);
2591 sk->sk_prot->destroy(sk);
2593 WARN_ON_ONCE(msk->wmem_reserved);
2594 WARN_ON_ONCE(msk->rmem_released);
2595 sk_stream_kill_queues(sk);
2596 xfrm_sk_free_policy(sk);
2598 sk_refcnt_debug_release(sk);
2599 mptcp_dispose_initial_subflow(msk);
2603 static void mptcp_close(struct sock *sk, long timeout)
2605 struct mptcp_subflow_context *subflow;
2606 bool do_cancel_work = false;
2609 sk->sk_shutdown = SHUTDOWN_MASK;
2611 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2612 inet_sk_state_store(sk, TCP_CLOSE);
2616 if (mptcp_close_state(sk))
2617 __mptcp_wr_shutdown(sk);
2619 sk_stream_wait_close(sk, timeout);
2622 /* orphan all the subflows */
2623 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2624 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2625 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2626 bool slow = lock_sock_fast(ssk);
2629 unlock_sock_fast(ssk, slow);
2634 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2635 if (sk->sk_state == TCP_CLOSE) {
2636 __mptcp_destroy_sock(sk);
2637 do_cancel_work = true;
2639 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2643 mptcp_cancel_work(sk);
2645 if (mptcp_sk(sk)->token)
2646 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2651 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2653 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2654 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2655 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2657 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2658 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2661 msk6->saddr = ssk6->saddr;
2662 msk6->flow_label = ssk6->flow_label;
2666 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2667 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2668 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2669 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2670 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2671 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2674 static int mptcp_disconnect(struct sock *sk, int flags)
2676 struct mptcp_subflow_context *subflow;
2677 struct mptcp_sock *msk = mptcp_sk(sk);
2679 mptcp_do_flush_join_list(msk);
2681 mptcp_for_each_subflow(msk, subflow) {
2682 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2685 tcp_disconnect(ssk, flags);
2691 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2692 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2694 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2696 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2700 struct sock *mptcp_sk_clone(const struct sock *sk,
2701 const struct mptcp_options_received *mp_opt,
2702 struct request_sock *req)
2704 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2705 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2706 struct mptcp_sock *msk;
2712 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2713 if (nsk->sk_family == AF_INET6)
2714 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2717 __mptcp_init_sock(nsk);
2719 msk = mptcp_sk(nsk);
2720 msk->local_key = subflow_req->local_key;
2721 msk->token = subflow_req->token;
2722 msk->subflow = NULL;
2723 WRITE_ONCE(msk->fully_established, false);
2724 if (mp_opt->csum_reqd)
2725 WRITE_ONCE(msk->csum_enabled, true);
2727 msk->write_seq = subflow_req->idsn + 1;
2728 msk->snd_nxt = msk->write_seq;
2729 msk->snd_una = msk->write_seq;
2730 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2731 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2733 if (mp_opt->mp_capable) {
2734 msk->can_ack = true;
2735 msk->remote_key = mp_opt->sndr_key;
2736 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2738 WRITE_ONCE(msk->ack_seq, ack_seq);
2739 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2742 sock_reset_flag(nsk, SOCK_RCU_FREE);
2743 /* will be fully established after successful MPC subflow creation */
2744 inet_sk_state_store(nsk, TCP_SYN_RECV);
2746 security_inet_csk_clone(nsk, req);
2747 bh_unlock_sock(nsk);
2749 /* keep a single reference */
2754 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2756 const struct tcp_sock *tp = tcp_sk(ssk);
2758 msk->rcvq_space.copied = 0;
2759 msk->rcvq_space.rtt_us = 0;
2761 msk->rcvq_space.time = tp->tcp_mstamp;
2763 /* initial rcv_space offering made to peer */
2764 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2765 TCP_INIT_CWND * tp->advmss);
2766 if (msk->rcvq_space.space == 0)
2767 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2769 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2772 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2775 struct mptcp_sock *msk = mptcp_sk(sk);
2776 struct socket *listener;
2779 listener = __mptcp_nmpc_socket(msk);
2780 if (WARN_ON_ONCE(!listener)) {
2785 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2786 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2790 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2791 if (sk_is_mptcp(newsk)) {
2792 struct mptcp_subflow_context *subflow;
2793 struct sock *new_mptcp_sock;
2795 subflow = mptcp_subflow_ctx(newsk);
2796 new_mptcp_sock = subflow->conn;
2798 /* is_mptcp should be false if subflow->conn is missing, see
2799 * subflow_syn_recv_sock()
2801 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2802 tcp_sk(newsk)->is_mptcp = 0;
2806 /* acquire the 2nd reference for the owning socket */
2807 sock_hold(new_mptcp_sock);
2808 newsk = new_mptcp_sock;
2809 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2811 MPTCP_INC_STATS(sock_net(sk),
2812 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2818 void mptcp_destroy_common(struct mptcp_sock *msk)
2820 struct sock *sk = (struct sock *)msk;
2822 __mptcp_clear_xmit(sk);
2824 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2825 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2827 skb_rbtree_purge(&msk->out_of_order_queue);
2828 mptcp_token_destroy(msk);
2829 mptcp_pm_free_anno_list(msk);
2832 static void mptcp_destroy(struct sock *sk)
2834 struct mptcp_sock *msk = mptcp_sk(sk);
2836 mptcp_destroy_common(msk);
2837 sk_sockets_allocated_dec(sk);
2840 void __mptcp_data_acked(struct sock *sk)
2842 if (!sock_owned_by_user(sk))
2843 __mptcp_clean_una(sk);
2845 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2847 if (mptcp_pending_data_fin_ack(sk))
2848 mptcp_schedule_work(sk);
2851 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2853 if (!mptcp_send_head(sk))
2856 if (!sock_owned_by_user(sk)) {
2857 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2859 if (xmit_ssk == ssk)
2860 __mptcp_subflow_push_pending(sk, ssk);
2862 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
2864 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2868 /* processes deferred events and flush wmem */
2869 static void mptcp_release_cb(struct sock *sk)
2872 unsigned long flags = 0;
2874 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2875 flags |= BIT(MPTCP_PUSH_PENDING);
2876 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2877 flags |= BIT(MPTCP_RETRANSMIT);
2881 /* the following actions acquire the subflow socket lock
2883 * 1) can't be invoked in atomic scope
2884 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2885 * datapath acquires the msk socket spinlock while helding
2886 * the subflow socket lock
2889 spin_unlock_bh(&sk->sk_lock.slock);
2890 if (flags & BIT(MPTCP_PUSH_PENDING))
2891 __mptcp_push_pending(sk, 0);
2892 if (flags & BIT(MPTCP_RETRANSMIT))
2893 __mptcp_retrans(sk);
2896 spin_lock_bh(&sk->sk_lock.slock);
2899 /* be sure to set the current sk state before tacking actions
2900 * depending on sk_state
2902 if (test_and_clear_bit(MPTCP_CONNECTED, &mptcp_sk(sk)->flags))
2903 __mptcp_set_connected(sk);
2904 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2905 __mptcp_clean_una_wakeup(sk);
2906 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2907 __mptcp_error_report(sk);
2909 /* push_pending may touch wmem_reserved, ensure we do the cleanup
2912 __mptcp_update_wmem(sk);
2913 __mptcp_update_rmem(sk);
2916 void mptcp_subflow_process_delegated(struct sock *ssk)
2918 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2919 struct sock *sk = subflow->conn;
2921 mptcp_data_lock(sk);
2922 if (!sock_owned_by_user(sk))
2923 __mptcp_subflow_push_pending(sk, ssk);
2925 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2926 mptcp_data_unlock(sk);
2927 mptcp_subflow_delegated_done(subflow);
2930 static int mptcp_hash(struct sock *sk)
2932 /* should never be called,
2933 * we hash the TCP subflows not the master socket
2939 static void mptcp_unhash(struct sock *sk)
2941 /* called from sk_common_release(), but nothing to do here */
2944 static int mptcp_get_port(struct sock *sk, unsigned short snum)
2946 struct mptcp_sock *msk = mptcp_sk(sk);
2947 struct socket *ssock;
2949 ssock = __mptcp_nmpc_socket(msk);
2950 pr_debug("msk=%p, subflow=%p", msk, ssock);
2951 if (WARN_ON_ONCE(!ssock))
2954 return inet_csk_get_port(ssock->sk, snum);
2957 void mptcp_finish_connect(struct sock *ssk)
2959 struct mptcp_subflow_context *subflow;
2960 struct mptcp_sock *msk;
2964 subflow = mptcp_subflow_ctx(ssk);
2968 pr_debug("msk=%p, token=%u", sk, subflow->token);
2970 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2972 subflow->map_seq = ack_seq;
2973 subflow->map_subflow_seq = 1;
2975 /* the socket is not connected yet, no msk/subflow ops can access/race
2976 * accessing the field below
2978 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2979 WRITE_ONCE(msk->local_key, subflow->local_key);
2980 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2981 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2982 WRITE_ONCE(msk->ack_seq, ack_seq);
2983 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2984 WRITE_ONCE(msk->can_ack, 1);
2985 WRITE_ONCE(msk->snd_una, msk->write_seq);
2987 mptcp_pm_new_connection(msk, ssk, 0);
2989 mptcp_rcv_space_init(msk, ssk);
2992 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2994 write_lock_bh(&sk->sk_callback_lock);
2995 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2996 sk_set_socket(sk, parent);
2997 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2998 write_unlock_bh(&sk->sk_callback_lock);
3001 bool mptcp_finish_join(struct sock *ssk)
3003 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3004 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3005 struct sock *parent = (void *)msk;
3006 struct socket *parent_sock;
3009 pr_debug("msk=%p, subflow=%p", msk, subflow);
3011 /* mptcp socket already closing? */
3012 if (!mptcp_is_fully_established(parent)) {
3013 subflow->reset_reason = MPTCP_RST_EMPTCP;
3017 if (!msk->pm.server_side)
3020 if (!mptcp_pm_allow_new_subflow(msk)) {
3021 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3025 /* active connections are already on conn_list, and we can't acquire
3027 * use the join list lock as synchronization point and double-check
3028 * msk status to avoid racing with __mptcp_destroy_sock()
3030 spin_lock_bh(&msk->join_list_lock);
3031 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3032 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3033 list_add_tail(&subflow->node, &msk->join_list);
3036 spin_unlock_bh(&msk->join_list_lock);
3038 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3042 /* attach to msk socket only after we are sure he will deal with us
3045 parent_sock = READ_ONCE(parent->sk_socket);
3046 if (parent_sock && !ssk->sk_socket)
3047 mptcp_sock_graft(ssk, parent_sock);
3048 subflow->map_seq = READ_ONCE(msk->ack_seq);
3050 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3054 static void mptcp_shutdown(struct sock *sk, int how)
3056 pr_debug("sk=%p, how=%d", sk, how);
3058 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3059 __mptcp_wr_shutdown(sk);
3062 static struct proto mptcp_prot = {
3064 .owner = THIS_MODULE,
3065 .init = mptcp_init_sock,
3066 .disconnect = mptcp_disconnect,
3067 .close = mptcp_close,
3068 .accept = mptcp_accept,
3069 .setsockopt = mptcp_setsockopt,
3070 .getsockopt = mptcp_getsockopt,
3071 .shutdown = mptcp_shutdown,
3072 .destroy = mptcp_destroy,
3073 .sendmsg = mptcp_sendmsg,
3074 .recvmsg = mptcp_recvmsg,
3075 .release_cb = mptcp_release_cb,
3077 .unhash = mptcp_unhash,
3078 .get_port = mptcp_get_port,
3079 .sockets_allocated = &mptcp_sockets_allocated,
3080 .memory_allocated = &tcp_memory_allocated,
3081 .memory_pressure = &tcp_memory_pressure,
3082 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3083 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3084 .sysctl_mem = sysctl_tcp_mem,
3085 .obj_size = sizeof(struct mptcp_sock),
3086 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3087 .no_autobind = true,
3090 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3092 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3093 struct socket *ssock;
3096 lock_sock(sock->sk);
3097 ssock = __mptcp_nmpc_socket(msk);
3103 err = ssock->ops->bind(ssock, uaddr, addr_len);
3105 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3108 release_sock(sock->sk);
3112 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3113 struct mptcp_subflow_context *subflow)
3115 subflow->request_mptcp = 0;
3116 __mptcp_do_fallback(msk);
3119 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3120 int addr_len, int flags)
3122 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3123 struct mptcp_subflow_context *subflow;
3124 struct socket *ssock;
3127 lock_sock(sock->sk);
3128 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3129 /* pending connection or invalid state, let existing subflow
3132 ssock = msk->subflow;
3136 ssock = __mptcp_nmpc_socket(msk);
3142 mptcp_token_destroy(msk);
3143 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3144 subflow = mptcp_subflow_ctx(ssock->sk);
3145 #ifdef CONFIG_TCP_MD5SIG
3146 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3149 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3150 mptcp_subflow_early_fallback(msk, subflow);
3152 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3153 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3154 mptcp_subflow_early_fallback(msk, subflow);
3156 if (likely(!__mptcp_check_fallback(msk)))
3157 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3160 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3161 sock->state = ssock->state;
3163 /* on successful connect, the msk state will be moved to established by
3164 * subflow_finish_connect()
3166 if (!err || err == -EINPROGRESS)
3167 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3169 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3172 release_sock(sock->sk);
3176 static int mptcp_listen(struct socket *sock, int backlog)
3178 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3179 struct socket *ssock;
3182 pr_debug("msk=%p", msk);
3184 lock_sock(sock->sk);
3185 ssock = __mptcp_nmpc_socket(msk);
3191 mptcp_token_destroy(msk);
3192 inet_sk_state_store(sock->sk, TCP_LISTEN);
3193 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3195 err = ssock->ops->listen(ssock, backlog);
3196 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3198 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3201 release_sock(sock->sk);
3205 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3206 int flags, bool kern)
3208 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3209 struct socket *ssock;
3212 pr_debug("msk=%p", msk);
3214 lock_sock(sock->sk);
3215 if (sock->sk->sk_state != TCP_LISTEN)
3218 ssock = __mptcp_nmpc_socket(msk);
3222 clear_bit(MPTCP_DATA_READY, &msk->flags);
3223 sock_hold(ssock->sk);
3224 release_sock(sock->sk);
3226 err = ssock->ops->accept(sock, newsock, flags, kern);
3227 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3228 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3229 struct mptcp_subflow_context *subflow;
3230 struct sock *newsk = newsock->sk;
3234 /* PM/worker can now acquire the first subflow socket
3235 * lock without racing with listener queue cleanup,
3236 * we can notify it, if needed.
3238 * Even if remote has reset the initial subflow by now
3239 * the refcnt is still at least one.
3241 subflow = mptcp_subflow_ctx(msk->first);
3242 list_add(&subflow->node, &msk->conn_list);
3243 sock_hold(msk->first);
3244 if (mptcp_is_fully_established(newsk))
3245 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3247 mptcp_copy_inaddrs(newsk, msk->first);
3248 mptcp_rcv_space_init(msk, msk->first);
3249 mptcp_propagate_sndbuf(newsk, msk->first);
3251 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3252 * This is needed so NOSPACE flag can be set from tcp stack.
3254 mptcp_flush_join_list(msk);
3255 mptcp_for_each_subflow(msk, subflow) {
3256 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3258 if (!ssk->sk_socket)
3259 mptcp_sock_graft(ssk, newsock);
3261 release_sock(newsk);
3264 if (inet_csk_listen_poll(ssock->sk))
3265 set_bit(MPTCP_DATA_READY, &msk->flags);
3266 sock_put(ssock->sk);
3270 release_sock(sock->sk);
3274 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3276 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3280 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3282 struct sock *sk = (struct sock *)msk;
3284 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3285 return EPOLLOUT | EPOLLWRNORM;
3287 if (sk_stream_is_writeable(sk))
3288 return EPOLLOUT | EPOLLWRNORM;
3290 mptcp_set_nospace(sk);
3291 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3292 if (sk_stream_is_writeable(sk))
3293 return EPOLLOUT | EPOLLWRNORM;
3298 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3299 struct poll_table_struct *wait)
3301 struct sock *sk = sock->sk;
3302 struct mptcp_sock *msk;
3307 sock_poll_wait(file, sock, wait);
3309 state = inet_sk_state_load(sk);
3310 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3311 if (state == TCP_LISTEN)
3312 return mptcp_check_readable(msk);
3314 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3315 mask |= mptcp_check_readable(msk);
3316 mask |= mptcp_check_writeable(msk);
3318 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3320 if (sk->sk_shutdown & RCV_SHUTDOWN)
3321 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3323 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3331 static const struct proto_ops mptcp_stream_ops = {
3333 .owner = THIS_MODULE,
3334 .release = inet_release,
3336 .connect = mptcp_stream_connect,
3337 .socketpair = sock_no_socketpair,
3338 .accept = mptcp_stream_accept,
3339 .getname = inet_getname,
3341 .ioctl = inet_ioctl,
3342 .gettstamp = sock_gettstamp,
3343 .listen = mptcp_listen,
3344 .shutdown = inet_shutdown,
3345 .setsockopt = sock_common_setsockopt,
3346 .getsockopt = sock_common_getsockopt,
3347 .sendmsg = inet_sendmsg,
3348 .recvmsg = inet_recvmsg,
3349 .mmap = sock_no_mmap,
3350 .sendpage = inet_sendpage,
3353 static struct inet_protosw mptcp_protosw = {
3354 .type = SOCK_STREAM,
3355 .protocol = IPPROTO_MPTCP,
3356 .prot = &mptcp_prot,
3357 .ops = &mptcp_stream_ops,
3358 .flags = INET_PROTOSW_ICSK,
3361 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3363 struct mptcp_delegated_action *delegated;
3364 struct mptcp_subflow_context *subflow;
3367 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3368 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3369 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3371 bh_lock_sock_nested(ssk);
3372 if (!sock_owned_by_user(ssk) &&
3373 mptcp_subflow_has_delegated_action(subflow))
3374 mptcp_subflow_process_delegated(ssk);
3375 /* ... elsewhere tcp_release_cb_override already processed
3376 * the action or will do at next release_sock().
3377 * In both case must dequeue the subflow here - on the same
3378 * CPU that scheduled it.
3380 bh_unlock_sock(ssk);
3383 if (++work_done == budget)
3387 /* always provide a 0 'work_done' argument, so that napi_complete_done
3388 * will not try accessing the NULL napi->dev ptr
3390 napi_complete_done(napi, 0);
3394 void __init mptcp_proto_init(void)
3396 struct mptcp_delegated_action *delegated;
3399 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3401 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3402 panic("Failed to allocate MPTCP pcpu counter\n");
3404 init_dummy_netdev(&mptcp_napi_dev);
3405 for_each_possible_cpu(cpu) {
3406 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3407 INIT_LIST_HEAD(&delegated->head);
3408 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3410 napi_enable(&delegated->napi);
3413 mptcp_subflow_init();
3417 if (proto_register(&mptcp_prot, 1) != 0)
3418 panic("Failed to register MPTCP proto.\n");
3420 inet_register_protosw(&mptcp_protosw);
3422 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3425 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3426 static const struct proto_ops mptcp_v6_stream_ops = {
3428 .owner = THIS_MODULE,
3429 .release = inet6_release,
3431 .connect = mptcp_stream_connect,
3432 .socketpair = sock_no_socketpair,
3433 .accept = mptcp_stream_accept,
3434 .getname = inet6_getname,
3436 .ioctl = inet6_ioctl,
3437 .gettstamp = sock_gettstamp,
3438 .listen = mptcp_listen,
3439 .shutdown = inet_shutdown,
3440 .setsockopt = sock_common_setsockopt,
3441 .getsockopt = sock_common_getsockopt,
3442 .sendmsg = inet6_sendmsg,
3443 .recvmsg = inet6_recvmsg,
3444 .mmap = sock_no_mmap,
3445 .sendpage = inet_sendpage,
3446 #ifdef CONFIG_COMPAT
3447 .compat_ioctl = inet6_compat_ioctl,
3451 static struct proto mptcp_v6_prot;
3453 static void mptcp_v6_destroy(struct sock *sk)
3456 inet6_destroy_sock(sk);
3459 static struct inet_protosw mptcp_v6_protosw = {
3460 .type = SOCK_STREAM,
3461 .protocol = IPPROTO_MPTCP,
3462 .prot = &mptcp_v6_prot,
3463 .ops = &mptcp_v6_stream_ops,
3464 .flags = INET_PROTOSW_ICSK,
3467 int __init mptcp_proto_v6_init(void)
3471 mptcp_v6_prot = mptcp_prot;
3472 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3473 mptcp_v6_prot.slab = NULL;
3474 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3475 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3477 err = proto_register(&mptcp_v6_prot, 1);
3481 err = inet6_register_protosw(&mptcp_v6_protosw);
3483 proto_unregister(&mptcp_v6_prot);