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(struct sock *sk, long tout)
416 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
419 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
421 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
423 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
424 inet_csk(ssk)->icsk_timeout - jiffies : 0;
427 static void mptcp_set_timeout(struct sock *sk)
429 struct mptcp_subflow_context *subflow;
432 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
433 tout = max(tout, mptcp_timeout_from_subflow(subflow));
434 __mptcp_set_timeout(sk, tout);
437 static bool tcp_can_send_ack(const struct sock *ssk)
439 return !((1 << inet_sk_state_load(ssk)) &
440 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
443 static void mptcp_send_ack(struct mptcp_sock *msk)
445 struct mptcp_subflow_context *subflow;
447 mptcp_for_each_subflow(msk, subflow) {
448 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
451 slow = lock_sock_fast(ssk);
452 if (tcp_can_send_ack(ssk))
454 unlock_sock_fast(ssk, slow);
458 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
462 slow = lock_sock_fast(ssk);
463 if (tcp_can_send_ack(ssk))
464 tcp_cleanup_rbuf(ssk, 1);
465 unlock_sock_fast(ssk, slow);
468 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
470 const struct inet_connection_sock *icsk = inet_csk(ssk);
471 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
472 const struct tcp_sock *tp = tcp_sk(ssk);
474 return (ack_pending & ICSK_ACK_SCHED) &&
475 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
476 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
477 (rx_empty && ack_pending &
478 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
481 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
483 int old_space = READ_ONCE(msk->old_wspace);
484 struct mptcp_subflow_context *subflow;
485 struct sock *sk = (struct sock *)msk;
486 int space = __mptcp_space(sk);
487 bool cleanup, rx_empty;
489 cleanup = (space > 0) && (space >= (old_space << 1));
490 rx_empty = !__mptcp_rmem(sk);
492 mptcp_for_each_subflow(msk, subflow) {
493 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
495 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
496 mptcp_subflow_cleanup_rbuf(ssk);
500 static bool mptcp_check_data_fin(struct sock *sk)
502 struct mptcp_sock *msk = mptcp_sk(sk);
503 u64 rcv_data_fin_seq;
506 if (__mptcp_check_fallback(msk))
509 /* Need to ack a DATA_FIN received from a peer while this side
510 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
511 * msk->rcv_data_fin was set when parsing the incoming options
512 * at the subflow level and the msk lock was not held, so this
513 * is the first opportunity to act on the DATA_FIN and change
516 * If we are caught up to the sequence number of the incoming
517 * DATA_FIN, send the DATA_ACK now and do state transition. If
518 * not caught up, do nothing and let the recv code send DATA_ACK
522 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
523 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
524 WRITE_ONCE(msk->rcv_data_fin, 0);
526 sk->sk_shutdown |= RCV_SHUTDOWN;
527 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
528 set_bit(MPTCP_DATA_READY, &msk->flags);
530 switch (sk->sk_state) {
531 case TCP_ESTABLISHED:
532 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
535 inet_sk_state_store(sk, TCP_CLOSING);
538 inet_sk_state_store(sk, TCP_CLOSE);
541 /* Other states not expected */
548 mptcp_close_wake_up(sk);
553 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
557 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
558 struct sock *sk = (struct sock *)msk;
559 unsigned int moved = 0;
560 bool more_data_avail;
565 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
567 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
568 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
570 if (unlikely(ssk_rbuf > sk_rbuf)) {
571 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
576 pr_debug("msk=%p ssk=%p", msk, ssk);
579 u32 map_remaining, offset;
580 u32 seq = tp->copied_seq;
584 /* try to move as much data as available */
585 map_remaining = subflow->map_data_len -
586 mptcp_subflow_get_map_offset(subflow);
588 skb = skb_peek(&ssk->sk_receive_queue);
590 /* if no data is found, a racing workqueue/recvmsg
591 * already processed the new data, stop here or we
592 * can enter an infinite loop
599 if (__mptcp_check_fallback(msk)) {
600 /* if we are running under the workqueue, TCP could have
601 * collapsed skbs between dummy map creation and now
602 * be sure to adjust the size
604 map_remaining = skb->len;
605 subflow->map_data_len = skb->len;
608 offset = seq - TCP_SKB_CB(skb)->seq;
609 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
615 if (offset < skb->len) {
616 size_t len = skb->len - offset;
621 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
625 if (WARN_ON_ONCE(map_remaining < len))
629 sk_eat_skb(ssk, skb);
633 WRITE_ONCE(tp->copied_seq, seq);
634 more_data_avail = mptcp_subflow_data_available(ssk);
636 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
640 } while (more_data_avail);
646 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
648 struct sock *sk = (struct sock *)msk;
649 struct sk_buff *skb, *tail;
654 p = rb_first(&msk->out_of_order_queue);
655 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
658 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
662 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
664 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
667 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
671 end_seq = MPTCP_SKB_CB(skb)->end_seq;
672 tail = skb_peek_tail(&sk->sk_receive_queue);
673 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
674 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
676 /* skip overlapping data, if any */
677 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
678 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
680 MPTCP_SKB_CB(skb)->offset += delta;
681 __skb_queue_tail(&sk->sk_receive_queue, skb);
683 msk->ack_seq = end_seq;
689 /* In most cases we will be able to lock the mptcp socket. If its already
690 * owned, we need to defer to the work queue to avoid ABBA deadlock.
692 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
694 struct sock *sk = (struct sock *)msk;
695 unsigned int moved = 0;
697 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
698 __mptcp_ofo_queue(msk);
699 if (unlikely(ssk->sk_err)) {
700 if (!sock_owned_by_user(sk))
701 __mptcp_error_report(sk);
703 set_bit(MPTCP_ERROR_REPORT, &msk->flags);
706 /* If the moves have caught up with the DATA_FIN sequence number
707 * it's time to ack the DATA_FIN and change socket state, but
708 * this is not a good place to change state. Let the workqueue
711 if (mptcp_pending_data_fin(sk, NULL))
712 mptcp_schedule_work(sk);
716 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
718 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
719 struct mptcp_sock *msk = mptcp_sk(sk);
720 int sk_rbuf, ssk_rbuf;
722 /* The peer can send data while we are shutting down this
723 * subflow at msk destruction time, but we must avoid enqueuing
724 * more data to the msk receive queue
726 if (unlikely(subflow->disposable))
729 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
730 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
731 if (unlikely(ssk_rbuf > sk_rbuf))
734 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
735 if (__mptcp_rmem(sk) > sk_rbuf) {
736 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
740 /* Wake-up the reader only for in-sequence data */
742 if (move_skbs_to_msk(msk, ssk)) {
743 set_bit(MPTCP_DATA_READY, &msk->flags);
744 sk->sk_data_ready(sk);
746 mptcp_data_unlock(sk);
749 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
751 struct mptcp_subflow_context *subflow;
754 if (likely(list_empty(&msk->join_list)))
757 spin_lock_bh(&msk->join_list_lock);
758 list_for_each_entry(subflow, &msk->join_list, node) {
759 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
761 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
762 if (READ_ONCE(msk->setsockopt_seq) != sseq)
765 list_splice_tail_init(&msk->join_list, &msk->conn_list);
766 spin_unlock_bh(&msk->join_list_lock);
771 void __mptcp_flush_join_list(struct mptcp_sock *msk)
773 if (likely(!mptcp_do_flush_join_list(msk)))
776 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
777 mptcp_schedule_work((struct sock *)msk);
780 static void mptcp_flush_join_list(struct mptcp_sock *msk)
782 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
786 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
789 mptcp_sockopt_sync_all(msk);
792 static bool mptcp_timer_pending(struct sock *sk)
794 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
797 static void mptcp_reset_timer(struct sock *sk)
799 struct inet_connection_sock *icsk = inet_csk(sk);
802 /* prevent rescheduling on close */
803 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
806 tout = mptcp_sk(sk)->timer_ival;
807 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
810 bool mptcp_schedule_work(struct sock *sk)
812 if (inet_sk_state_load(sk) != TCP_CLOSE &&
813 schedule_work(&mptcp_sk(sk)->work)) {
814 /* each subflow already holds a reference to the sk, and the
815 * workqueue is invoked by a subflow, so sk can't go away here.
823 void mptcp_subflow_eof(struct sock *sk)
825 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
826 mptcp_schedule_work(sk);
829 static void mptcp_check_for_eof(struct mptcp_sock *msk)
831 struct mptcp_subflow_context *subflow;
832 struct sock *sk = (struct sock *)msk;
835 mptcp_for_each_subflow(msk, subflow)
836 receivers += !subflow->rx_eof;
840 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
841 /* hopefully temporary hack: propagate shutdown status
842 * to msk, when all subflows agree on it
844 sk->sk_shutdown |= RCV_SHUTDOWN;
846 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
847 set_bit(MPTCP_DATA_READY, &msk->flags);
848 sk->sk_data_ready(sk);
851 switch (sk->sk_state) {
852 case TCP_ESTABLISHED:
853 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
856 inet_sk_state_store(sk, TCP_CLOSING);
859 inet_sk_state_store(sk, TCP_CLOSE);
864 mptcp_close_wake_up(sk);
867 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
869 struct mptcp_subflow_context *subflow;
870 struct sock *sk = (struct sock *)msk;
872 sock_owned_by_me(sk);
874 mptcp_for_each_subflow(msk, subflow) {
875 if (READ_ONCE(subflow->data_avail))
876 return mptcp_subflow_tcp_sock(subflow);
882 static bool mptcp_skb_can_collapse_to(u64 write_seq,
883 const struct sk_buff *skb,
884 const struct mptcp_ext *mpext)
886 if (!tcp_skb_can_collapse_to(skb))
889 /* can collapse only if MPTCP level sequence is in order and this
890 * mapping has not been xmitted yet
892 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
896 /* we can append data to the given data frag if:
897 * - there is space available in the backing page_frag
898 * - the data frag tail matches the current page_frag free offset
899 * - the data frag end sequence number matches the current write seq
901 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
902 const struct page_frag *pfrag,
903 const struct mptcp_data_frag *df)
905 return df && pfrag->page == df->page &&
906 pfrag->size - pfrag->offset > 0 &&
907 pfrag->offset == (df->offset + df->data_len) &&
908 df->data_seq + df->data_len == msk->write_seq;
911 static int mptcp_wmem_with_overhead(int size)
913 return size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
916 static void __mptcp_wmem_reserve(struct sock *sk, int size)
918 int amount = mptcp_wmem_with_overhead(size);
919 struct mptcp_sock *msk = mptcp_sk(sk);
921 WARN_ON_ONCE(msk->wmem_reserved);
922 if (WARN_ON_ONCE(amount < 0))
925 if (amount <= sk->sk_forward_alloc)
928 /* under memory pressure try to reserve at most a single page
929 * otherwise try to reserve the full estimate and fallback
930 * to a single page before entering the error path
932 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
933 !sk_wmem_schedule(sk, amount)) {
934 if (amount <= PAGE_SIZE)
938 if (!sk_wmem_schedule(sk, amount))
943 msk->wmem_reserved = amount;
944 sk->sk_forward_alloc -= amount;
948 /* we will wait for memory on next allocation */
949 msk->wmem_reserved = -1;
952 static void __mptcp_update_wmem(struct sock *sk)
954 struct mptcp_sock *msk = mptcp_sk(sk);
956 #ifdef CONFIG_LOCKDEP
957 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
960 if (!msk->wmem_reserved)
963 if (msk->wmem_reserved < 0)
964 msk->wmem_reserved = 0;
965 if (msk->wmem_reserved > 0) {
966 sk->sk_forward_alloc += msk->wmem_reserved;
967 msk->wmem_reserved = 0;
971 static bool mptcp_wmem_alloc(struct sock *sk, int size)
973 struct mptcp_sock *msk = mptcp_sk(sk);
975 /* check for pre-existing error condition */
976 if (msk->wmem_reserved < 0)
979 if (msk->wmem_reserved >= size)
983 if (!sk_wmem_schedule(sk, size)) {
984 mptcp_data_unlock(sk);
988 sk->sk_forward_alloc -= size;
989 msk->wmem_reserved += size;
990 mptcp_data_unlock(sk);
993 msk->wmem_reserved -= size;
997 static void mptcp_wmem_uncharge(struct sock *sk, int size)
999 struct mptcp_sock *msk = mptcp_sk(sk);
1001 if (msk->wmem_reserved < 0)
1002 msk->wmem_reserved = 0;
1003 msk->wmem_reserved += size;
1006 static void mptcp_mem_reclaim_partial(struct sock *sk)
1008 struct mptcp_sock *msk = mptcp_sk(sk);
1010 /* if we are experiencing a transint allocation error,
1011 * the forward allocation memory has been already
1014 if (msk->wmem_reserved < 0)
1017 mptcp_data_lock(sk);
1018 sk->sk_forward_alloc += msk->wmem_reserved;
1019 sk_mem_reclaim_partial(sk);
1020 msk->wmem_reserved = sk->sk_forward_alloc;
1021 sk->sk_forward_alloc = 0;
1022 mptcp_data_unlock(sk);
1025 static void dfrag_uncharge(struct sock *sk, int len)
1027 sk_mem_uncharge(sk, len);
1028 sk_wmem_queued_add(sk, -len);
1031 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1033 int len = dfrag->data_len + dfrag->overhead;
1035 list_del(&dfrag->list);
1036 dfrag_uncharge(sk, len);
1037 put_page(dfrag->page);
1040 static void __mptcp_clean_una(struct sock *sk)
1042 struct mptcp_sock *msk = mptcp_sk(sk);
1043 struct mptcp_data_frag *dtmp, *dfrag;
1044 bool cleaned = false;
1047 /* on fallback we just need to ignore snd_una, as this is really
1050 if (__mptcp_check_fallback(msk))
1051 msk->snd_una = READ_ONCE(msk->snd_nxt);
1053 snd_una = msk->snd_una;
1054 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1055 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1058 if (unlikely(dfrag == msk->first_pending)) {
1059 /* in recovery mode can see ack after the current snd head */
1060 if (WARN_ON_ONCE(!msk->recovery))
1063 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1066 dfrag_clear(sk, dfrag);
1070 dfrag = mptcp_rtx_head(sk);
1071 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1072 u64 delta = snd_una - dfrag->data_seq;
1074 /* prevent wrap around in recovery mode */
1075 if (unlikely(delta > dfrag->already_sent)) {
1076 if (WARN_ON_ONCE(!msk->recovery))
1078 if (WARN_ON_ONCE(delta > dfrag->data_len))
1080 dfrag->already_sent += delta - dfrag->already_sent;
1083 dfrag->data_seq += delta;
1084 dfrag->offset += delta;
1085 dfrag->data_len -= delta;
1086 dfrag->already_sent -= delta;
1088 dfrag_uncharge(sk, delta);
1092 /* all retransmitted data acked, recovery completed */
1093 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1094 msk->recovery = false;
1098 if (tcp_under_memory_pressure(sk)) {
1099 __mptcp_update_wmem(sk);
1100 sk_mem_reclaim_partial(sk);
1104 if (snd_una == READ_ONCE(msk->snd_nxt) && !msk->recovery) {
1105 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1106 mptcp_stop_timer(sk);
1108 mptcp_reset_timer(sk);
1112 static void __mptcp_clean_una_wakeup(struct sock *sk)
1114 #ifdef CONFIG_LOCKDEP
1115 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
1117 __mptcp_clean_una(sk);
1118 mptcp_write_space(sk);
1121 static void mptcp_clean_una_wakeup(struct sock *sk)
1123 mptcp_data_lock(sk);
1124 __mptcp_clean_una_wakeup(sk);
1125 mptcp_data_unlock(sk);
1128 static void mptcp_enter_memory_pressure(struct sock *sk)
1130 struct mptcp_subflow_context *subflow;
1131 struct mptcp_sock *msk = mptcp_sk(sk);
1134 sk_stream_moderate_sndbuf(sk);
1135 mptcp_for_each_subflow(msk, subflow) {
1136 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1139 tcp_enter_memory_pressure(ssk);
1140 sk_stream_moderate_sndbuf(ssk);
1145 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1148 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1150 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1151 pfrag, sk->sk_allocation)))
1154 mptcp_enter_memory_pressure(sk);
1158 static struct mptcp_data_frag *
1159 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1162 int offset = ALIGN(orig_offset, sizeof(long));
1163 struct mptcp_data_frag *dfrag;
1165 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1166 dfrag->data_len = 0;
1167 dfrag->data_seq = msk->write_seq;
1168 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1169 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1170 dfrag->already_sent = 0;
1171 dfrag->page = pfrag->page;
1176 struct mptcp_sendmsg_info {
1184 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1187 u64 window_end = mptcp_wnd_end(msk);
1189 if (__mptcp_check_fallback(msk))
1192 if (!before64(data_seq + avail_size, window_end)) {
1193 u64 allowed_size = window_end - data_seq;
1195 return min_t(unsigned int, allowed_size, avail_size);
1201 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1203 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1207 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1211 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1213 struct sk_buff *skb;
1215 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1217 if (likely(__mptcp_add_ext(skb, gfp))) {
1218 skb_reserve(skb, MAX_TCP_HEADER);
1219 skb->reserved_tailroom = skb->end - skb->tail;
1224 mptcp_enter_memory_pressure(sk);
1229 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1231 struct sk_buff *skb;
1233 if (ssk->sk_tx_skb_cache) {
1234 skb = ssk->sk_tx_skb_cache;
1235 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1236 !__mptcp_add_ext(skb, gfp)))
1241 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1245 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1246 ssk->sk_tx_skb_cache = skb;
1253 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1255 return !ssk->sk_tx_skb_cache &&
1256 tcp_under_memory_pressure(sk);
1259 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1261 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1262 mptcp_mem_reclaim_partial(sk);
1263 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1266 /* note: this always recompute the csum on the whole skb, even
1267 * if we just appended a single frag. More status info needed
1269 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1271 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1272 __wsum csum = ~csum_unfold(mpext->csum);
1273 int offset = skb->len - added;
1275 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1278 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1279 struct mptcp_data_frag *dfrag,
1280 struct mptcp_sendmsg_info *info)
1282 u64 data_seq = dfrag->data_seq + info->sent;
1283 struct mptcp_sock *msk = mptcp_sk(sk);
1284 bool zero_window_probe = false;
1285 struct mptcp_ext *mpext = NULL;
1286 struct sk_buff *skb, *tail;
1287 bool can_collapse = false;
1292 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1293 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1295 /* compute send limit */
1296 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1297 avail_size = info->size_goal;
1298 skb = tcp_write_queue_tail(ssk);
1300 /* Limit the write to the size available in the
1301 * current skb, if any, so that we create at most a new skb.
1302 * Explicitly tells TCP internals to avoid collapsing on later
1303 * queue management operation, to avoid breaking the ext <->
1304 * SSN association set here
1306 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1307 can_collapse = (info->size_goal - skb->len > 0) &&
1308 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1309 if (!can_collapse) {
1310 TCP_SKB_CB(skb)->eor = 1;
1312 size_bias = skb->len;
1313 avail_size = info->size_goal - skb->len;
1317 /* Zero window and all data acked? Probe. */
1318 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1319 if (avail_size == 0) {
1320 u64 snd_una = READ_ONCE(msk->snd_una);
1322 if (skb || snd_una != msk->snd_nxt)
1324 zero_window_probe = true;
1325 data_seq = snd_una - 1;
1329 if (WARN_ON_ONCE(info->sent > info->limit ||
1330 info->limit > dfrag->data_len))
1333 ret = info->limit - info->sent;
1334 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1335 dfrag->page, dfrag->offset + info->sent, &ret);
1337 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1341 /* if the tail skb is still the cached one, collapsing really happened.
1344 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1345 mpext->data_len += ret;
1346 WARN_ON_ONCE(!can_collapse);
1347 WARN_ON_ONCE(zero_window_probe);
1351 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1352 if (WARN_ON_ONCE(!mpext)) {
1353 /* should never reach here, stream corrupted */
1357 memset(mpext, 0, sizeof(*mpext));
1358 mpext->data_seq = data_seq;
1359 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1360 mpext->data_len = ret;
1364 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1365 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1368 if (zero_window_probe) {
1369 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1371 if (READ_ONCE(msk->csum_enabled))
1372 mptcp_update_data_checksum(tail, ret);
1373 tcp_push_pending_frames(ssk);
1377 if (READ_ONCE(msk->csum_enabled))
1378 mptcp_update_data_checksum(tail, ret);
1379 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1383 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1384 sizeof(struct tcphdr) - \
1385 MAX_TCP_OPTION_SPACE - \
1386 sizeof(struct ipv6hdr) - \
1387 sizeof(struct frag_hdr))
1389 struct subflow_send_info {
1394 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1396 if (!subflow->stale)
1400 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1403 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1405 if (unlikely(subflow->stale)) {
1406 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1408 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1411 mptcp_subflow_set_active(subflow);
1413 return __mptcp_subflow_active(subflow);
1416 /* implement the mptcp packet scheduler;
1417 * returns the subflow that will transmit the next DSS
1418 * additionally updates the rtx timeout
1420 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1422 struct subflow_send_info send_info[2];
1423 struct mptcp_subflow_context *subflow;
1424 struct sock *sk = (struct sock *)msk;
1425 int i, nr_active = 0;
1431 sock_owned_by_me(sk);
1433 if (__mptcp_check_fallback(msk)) {
1436 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1439 /* re-use last subflow, if the burst allow that */
1440 if (msk->last_snd && msk->snd_burst > 0 &&
1441 sk_stream_memory_free(msk->last_snd) &&
1442 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1443 mptcp_set_timeout(sk);
1444 return msk->last_snd;
1447 /* pick the subflow with the lower wmem/wspace ratio */
1448 for (i = 0; i < 2; ++i) {
1449 send_info[i].ssk = NULL;
1450 send_info[i].ratio = -1;
1452 mptcp_for_each_subflow(msk, subflow) {
1453 trace_mptcp_subflow_get_send(subflow);
1454 ssk = mptcp_subflow_tcp_sock(subflow);
1455 if (!mptcp_subflow_active(subflow))
1458 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1459 nr_active += !subflow->backup;
1460 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1463 pace = READ_ONCE(ssk->sk_pacing_rate);
1467 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1469 if (ratio < send_info[subflow->backup].ratio) {
1470 send_info[subflow->backup].ssk = ssk;
1471 send_info[subflow->backup].ratio = ratio;
1474 __mptcp_set_timeout(sk, tout);
1476 /* pick the best backup if no other subflow is active */
1478 send_info[0].ssk = send_info[1].ssk;
1480 if (send_info[0].ssk) {
1481 msk->last_snd = send_info[0].ssk;
1482 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1483 tcp_sk(msk->last_snd)->snd_wnd);
1484 return msk->last_snd;
1490 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1491 struct mptcp_sendmsg_info *info)
1493 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1497 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1499 struct sock *prev_ssk = NULL, *ssk = NULL;
1500 struct mptcp_sock *msk = mptcp_sk(sk);
1501 struct mptcp_sendmsg_info info = {
1504 struct mptcp_data_frag *dfrag;
1505 int len, copied = 0;
1507 while ((dfrag = mptcp_send_head(sk))) {
1508 info.sent = dfrag->already_sent;
1509 info.limit = dfrag->data_len;
1510 len = dfrag->data_len - dfrag->already_sent;
1515 mptcp_flush_join_list(msk);
1516 ssk = mptcp_subflow_get_send(msk);
1518 /* First check. If the ssk has changed since
1519 * the last round, release prev_ssk
1521 if (ssk != prev_ssk && prev_ssk)
1522 mptcp_push_release(sk, prev_ssk, &info);
1526 /* Need to lock the new subflow only if different
1527 * from the previous one, otherwise we are still
1528 * helding the relevant lock
1530 if (ssk != prev_ssk)
1533 /* keep it simple and always provide a new skb for the
1534 * subflow, even if we will not use it when collapsing
1535 * on the pending one
1537 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1538 mptcp_push_release(sk, ssk, &info);
1542 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1544 mptcp_push_release(sk, ssk, &info);
1549 dfrag->already_sent += ret;
1550 msk->snd_nxt += ret;
1551 msk->snd_burst -= ret;
1552 msk->tx_pending_data -= ret;
1556 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1559 /* at this point we held the socket lock for the last subflow we used */
1561 mptcp_push_release(sk, ssk, &info);
1564 /* ensure the rtx timer is running */
1565 if (!mptcp_timer_pending(sk))
1566 mptcp_reset_timer(sk);
1568 __mptcp_check_send_data_fin(sk);
1571 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1573 struct mptcp_sock *msk = mptcp_sk(sk);
1574 struct mptcp_sendmsg_info info;
1575 struct mptcp_data_frag *dfrag;
1576 struct sock *xmit_ssk;
1577 int len, copied = 0;
1581 while ((dfrag = mptcp_send_head(sk))) {
1582 info.sent = dfrag->already_sent;
1583 info.limit = dfrag->data_len;
1584 len = dfrag->data_len - dfrag->already_sent;
1588 /* the caller already invoked the packet scheduler,
1589 * check for a different subflow usage only after
1590 * spooling the first chunk of data
1592 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1595 if (xmit_ssk != ssk) {
1596 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1600 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1601 __mptcp_update_wmem(sk);
1602 sk_mem_reclaim_partial(sk);
1604 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1607 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1612 dfrag->already_sent += ret;
1613 msk->snd_nxt += ret;
1614 msk->snd_burst -= ret;
1615 msk->tx_pending_data -= ret;
1620 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1624 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1625 * not going to flush it via release_sock()
1627 __mptcp_update_wmem(sk);
1629 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1631 if (!mptcp_timer_pending(sk))
1632 mptcp_reset_timer(sk);
1634 if (msk->snd_data_fin_enable &&
1635 msk->snd_nxt + 1 == msk->write_seq)
1636 mptcp_schedule_work(sk);
1640 static void mptcp_set_nospace(struct sock *sk)
1642 /* enable autotune */
1643 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1645 /* will be cleared on avail space */
1646 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1649 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1651 struct mptcp_sock *msk = mptcp_sk(sk);
1652 struct page_frag *pfrag;
1657 /* we don't support FASTOPEN yet */
1658 if (msg->msg_flags & MSG_FASTOPEN)
1661 /* silently ignore everything else */
1662 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1664 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1666 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1668 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1669 ret = sk_stream_wait_connect(sk, &timeo);
1674 pfrag = sk_page_frag(sk);
1676 while (msg_data_left(msg)) {
1677 int total_ts, frag_truesize = 0;
1678 struct mptcp_data_frag *dfrag;
1679 bool dfrag_collapsed;
1680 size_t psize, offset;
1682 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1687 /* reuse tail pfrag, if possible, or carve a new one from the
1690 dfrag = mptcp_pending_tail(sk);
1691 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1692 if (!dfrag_collapsed) {
1693 if (!sk_stream_memory_free(sk))
1694 goto wait_for_memory;
1696 if (!mptcp_page_frag_refill(sk, pfrag))
1697 goto wait_for_memory;
1699 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1700 frag_truesize = dfrag->overhead;
1703 /* we do not bound vs wspace, to allow a single packet.
1704 * memory accounting will prevent execessive memory usage
1707 offset = dfrag->offset + dfrag->data_len;
1708 psize = pfrag->size - offset;
1709 psize = min_t(size_t, psize, msg_data_left(msg));
1710 total_ts = psize + frag_truesize;
1712 if (!mptcp_wmem_alloc(sk, total_ts))
1713 goto wait_for_memory;
1715 if (copy_page_from_iter(dfrag->page, offset, psize,
1716 &msg->msg_iter) != psize) {
1717 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1722 /* data successfully copied into the write queue */
1724 dfrag->data_len += psize;
1725 frag_truesize += psize;
1726 pfrag->offset += frag_truesize;
1727 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1728 msk->tx_pending_data += psize;
1730 /* charge data on mptcp pending queue to the msk socket
1731 * Note: we charge such data both to sk and ssk
1733 sk_wmem_queued_add(sk, frag_truesize);
1734 if (!dfrag_collapsed) {
1735 get_page(dfrag->page);
1736 list_add_tail(&dfrag->list, &msk->rtx_queue);
1737 if (!msk->first_pending)
1738 WRITE_ONCE(msk->first_pending, dfrag);
1740 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1741 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1747 mptcp_set_nospace(sk);
1748 __mptcp_push_pending(sk, msg->msg_flags);
1749 ret = sk_stream_wait_memory(sk, &timeo);
1755 __mptcp_push_pending(sk, msg->msg_flags);
1759 return copied ? : ret;
1762 static void mptcp_wait_data(struct sock *sk, long *timeo)
1764 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1765 struct mptcp_sock *msk = mptcp_sk(sk);
1767 add_wait_queue(sk_sleep(sk), &wait);
1768 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1770 sk_wait_event(sk, timeo,
1771 test_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1773 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1774 remove_wait_queue(sk_sleep(sk), &wait);
1777 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1779 size_t len, int flags,
1780 struct scm_timestamping_internal *tss,
1783 struct sk_buff *skb, *tmp;
1786 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1787 u32 offset = MPTCP_SKB_CB(skb)->offset;
1788 u32 data_len = skb->len - offset;
1789 u32 count = min_t(size_t, len - copied, data_len);
1792 if (!(flags & MSG_TRUNC)) {
1793 err = skb_copy_datagram_msg(skb, offset, msg, count);
1794 if (unlikely(err < 0)) {
1801 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1802 tcp_update_recv_tstamps(skb, tss);
1803 *cmsg_flags |= MPTCP_CMSG_TS;
1808 if (count < data_len) {
1809 if (!(flags & MSG_PEEK))
1810 MPTCP_SKB_CB(skb)->offset += count;
1814 if (!(flags & MSG_PEEK)) {
1815 /* we will bulk release the skb memory later */
1816 skb->destructor = NULL;
1817 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1818 __skb_unlink(skb, &msk->receive_queue);
1829 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1831 * Only difference: Use highest rtt estimate of the subflows in use.
1833 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1835 struct mptcp_subflow_context *subflow;
1836 struct sock *sk = (struct sock *)msk;
1837 u32 time, advmss = 1;
1840 sock_owned_by_me(sk);
1845 msk->rcvq_space.copied += copied;
1847 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1848 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1850 rtt_us = msk->rcvq_space.rtt_us;
1851 if (rtt_us && time < (rtt_us >> 3))
1855 mptcp_for_each_subflow(msk, subflow) {
1856 const struct tcp_sock *tp;
1860 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1862 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1863 sf_advmss = READ_ONCE(tp->advmss);
1865 rtt_us = max(sf_rtt_us, rtt_us);
1866 advmss = max(sf_advmss, advmss);
1869 msk->rcvq_space.rtt_us = rtt_us;
1870 if (time < (rtt_us >> 3) || rtt_us == 0)
1873 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1876 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1877 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1881 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1883 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1885 do_div(grow, msk->rcvq_space.space);
1886 rcvwin += (grow << 1);
1888 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1889 while (tcp_win_from_space(sk, rcvmem) < advmss)
1892 do_div(rcvwin, advmss);
1893 rcvbuf = min_t(u64, rcvwin * rcvmem,
1894 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1896 if (rcvbuf > sk->sk_rcvbuf) {
1899 window_clamp = tcp_win_from_space(sk, rcvbuf);
1900 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1902 /* Make subflows follow along. If we do not do this, we
1903 * get drops at subflow level if skbs can't be moved to
1904 * the mptcp rx queue fast enough (announced rcv_win can
1905 * exceed ssk->sk_rcvbuf).
1907 mptcp_for_each_subflow(msk, subflow) {
1911 ssk = mptcp_subflow_tcp_sock(subflow);
1912 slow = lock_sock_fast(ssk);
1913 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1914 tcp_sk(ssk)->window_clamp = window_clamp;
1915 tcp_cleanup_rbuf(ssk, 1);
1916 unlock_sock_fast(ssk, slow);
1921 msk->rcvq_space.space = msk->rcvq_space.copied;
1923 msk->rcvq_space.copied = 0;
1924 msk->rcvq_space.time = mstamp;
1927 static void __mptcp_update_rmem(struct sock *sk)
1929 struct mptcp_sock *msk = mptcp_sk(sk);
1931 if (!msk->rmem_released)
1934 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1935 sk_mem_uncharge(sk, msk->rmem_released);
1936 WRITE_ONCE(msk->rmem_released, 0);
1939 static void __mptcp_splice_receive_queue(struct sock *sk)
1941 struct mptcp_sock *msk = mptcp_sk(sk);
1943 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1946 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1948 struct sock *sk = (struct sock *)msk;
1949 unsigned int moved = 0;
1952 mptcp_flush_join_list(msk);
1954 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1957 /* we can have data pending in the subflows only if the msk
1958 * receive buffer was full at subflow_data_ready() time,
1959 * that is an unlikely slow path.
1964 slowpath = lock_sock_fast(ssk);
1965 mptcp_data_lock(sk);
1966 __mptcp_update_rmem(sk);
1967 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1968 mptcp_data_unlock(sk);
1970 if (unlikely(ssk->sk_err))
1971 __mptcp_error_report(sk);
1972 unlock_sock_fast(ssk, slowpath);
1975 /* acquire the data lock only if some input data is pending */
1977 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1978 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1979 mptcp_data_lock(sk);
1980 __mptcp_update_rmem(sk);
1981 ret |= __mptcp_ofo_queue(msk);
1982 __mptcp_splice_receive_queue(sk);
1983 mptcp_data_unlock(sk);
1986 mptcp_check_data_fin((struct sock *)msk);
1987 return !skb_queue_empty(&msk->receive_queue);
1990 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1991 int nonblock, int flags, int *addr_len)
1993 struct mptcp_sock *msk = mptcp_sk(sk);
1994 struct scm_timestamping_internal tss;
1995 int copied = 0, cmsg_flags = 0;
1999 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2000 if (unlikely(flags & MSG_ERRQUEUE))
2001 return inet_recv_error(sk, msg, len, addr_len);
2003 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
2004 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2009 timeo = sock_rcvtimeo(sk, nonblock);
2011 len = min_t(size_t, len, INT_MAX);
2012 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2014 while (copied < len) {
2017 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2018 if (unlikely(bytes_read < 0)) {
2020 copied = bytes_read;
2024 copied += bytes_read;
2026 /* be sure to advertise window change */
2027 mptcp_cleanup_rbuf(msk);
2029 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2032 /* only the master socket status is relevant here. The exit
2033 * conditions mirror closely tcp_recvmsg()
2035 if (copied >= target)
2040 sk->sk_state == TCP_CLOSE ||
2041 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2043 signal_pending(current))
2047 copied = sock_error(sk);
2051 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2052 mptcp_check_for_eof(msk);
2054 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2055 /* race breaker: the shutdown could be after the
2056 * previous receive queue check
2058 if (__mptcp_move_skbs(msk))
2063 if (sk->sk_state == TCP_CLOSE) {
2073 if (signal_pending(current)) {
2074 copied = sock_intr_errno(timeo);
2079 pr_debug("block timeout %ld", timeo);
2080 mptcp_wait_data(sk, &timeo);
2083 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2084 skb_queue_empty(&msk->receive_queue)) {
2085 /* entire backlog drained, clear DATA_READY. */
2086 clear_bit(MPTCP_DATA_READY, &msk->flags);
2088 /* .. race-breaker: ssk might have gotten new data
2089 * after last __mptcp_move_skbs() returned false.
2091 if (unlikely(__mptcp_move_skbs(msk)))
2092 set_bit(MPTCP_DATA_READY, &msk->flags);
2096 if (cmsg_flags && copied >= 0) {
2097 if (cmsg_flags & MPTCP_CMSG_TS)
2098 tcp_recv_timestamp(msg, sk, &tss);
2101 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2102 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2103 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2104 if (!(flags & MSG_PEEK))
2105 mptcp_rcv_space_adjust(msk, copied);
2111 static void mptcp_retransmit_timer(struct timer_list *t)
2113 struct inet_connection_sock *icsk = from_timer(icsk, t,
2114 icsk_retransmit_timer);
2115 struct sock *sk = &icsk->icsk_inet.sk;
2116 struct mptcp_sock *msk = mptcp_sk(sk);
2119 if (!sock_owned_by_user(sk)) {
2120 /* we need a process context to retransmit */
2121 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2122 mptcp_schedule_work(sk);
2124 /* delegate our work to tcp_release_cb() */
2125 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2131 static void mptcp_timeout_timer(struct timer_list *t)
2133 struct sock *sk = from_timer(sk, t, sk_timer);
2135 mptcp_schedule_work(sk);
2139 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2142 * A backup subflow is returned only if that is the only kind available.
2144 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2146 struct sock *backup = NULL, *pick = NULL;
2147 struct mptcp_subflow_context *subflow;
2148 int min_stale_count = INT_MAX;
2150 sock_owned_by_me((const struct sock *)msk);
2152 if (__mptcp_check_fallback(msk))
2155 mptcp_for_each_subflow(msk, subflow) {
2156 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2158 if (!__mptcp_subflow_active(subflow))
2161 /* still data outstanding at TCP level? skip this */
2162 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2163 mptcp_pm_subflow_chk_stale(msk, ssk);
2164 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2168 if (subflow->backup) {
2181 /* use backup only if there are no progresses anywhere */
2182 return min_stale_count > 1 ? backup : NULL;
2185 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2188 iput(SOCK_INODE(msk->subflow));
2189 msk->subflow = NULL;
2193 bool __mptcp_retransmit_pending_data(struct sock *sk)
2195 struct mptcp_data_frag *cur, *rtx_head;
2196 struct mptcp_sock *msk = mptcp_sk(sk);
2198 if (__mptcp_check_fallback(mptcp_sk(sk)))
2201 if (tcp_rtx_and_write_queues_empty(sk))
2204 /* the closing socket has some data untransmitted and/or unacked:
2205 * some data in the mptcp rtx queue has not really xmitted yet.
2206 * keep it simple and re-inject the whole mptcp level rtx queue
2208 mptcp_data_lock(sk);
2209 __mptcp_clean_una_wakeup(sk);
2210 rtx_head = mptcp_rtx_head(sk);
2212 mptcp_data_unlock(sk);
2216 /* will accept ack for reijected data before re-sending them */
2217 if (!msk->recovery || after64(msk->snd_nxt, msk->recovery_snd_nxt))
2218 msk->recovery_snd_nxt = msk->snd_nxt;
2219 msk->recovery = true;
2220 mptcp_data_unlock(sk);
2222 msk->first_pending = rtx_head;
2223 msk->tx_pending_data += msk->snd_nxt - rtx_head->data_seq;
2224 msk->snd_nxt = rtx_head->data_seq;
2227 /* be sure to clear the "sent status" on all re-injected fragments */
2228 list_for_each_entry(cur, &msk->rtx_queue, list) {
2229 if (!cur->already_sent)
2231 cur->already_sent = 0;
2237 /* subflow sockets can be either outgoing (connect) or incoming
2240 * Outgoing subflows use in-kernel sockets.
2241 * Incoming subflows do not have their own 'struct socket' allocated,
2242 * so we need to use tcp_close() after detaching them from the mptcp
2245 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2246 struct mptcp_subflow_context *subflow)
2248 struct mptcp_sock *msk = mptcp_sk(sk);
2251 list_del(&subflow->node);
2253 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2255 /* if we are invoked by the msk cleanup code, the subflow is
2261 need_push = __mptcp_retransmit_pending_data(sk);
2262 subflow->disposable = 1;
2264 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2265 * the ssk has been already destroyed, we just need to release the
2266 * reference owned by msk;
2268 if (!inet_csk(ssk)->icsk_ulp_ops) {
2269 kfree_rcu(subflow, rcu);
2271 /* otherwise tcp will dispose of the ssk and subflow ctx */
2272 __tcp_close(ssk, 0);
2274 /* close acquired an extra ref */
2281 if (ssk == msk->last_snd)
2282 msk->last_snd = NULL;
2284 if (ssk == msk->first)
2287 if (msk->subflow && ssk == msk->subflow->sk)
2288 mptcp_dispose_initial_subflow(msk);
2291 __mptcp_push_pending(sk, 0);
2294 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2295 struct mptcp_subflow_context *subflow)
2297 if (sk->sk_state == TCP_ESTABLISHED)
2298 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2299 __mptcp_close_ssk(sk, ssk, subflow);
2302 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2307 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2309 struct mptcp_subflow_context *subflow, *tmp;
2313 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2314 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2316 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2319 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2320 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2323 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2327 static bool mptcp_check_close_timeout(const struct sock *sk)
2329 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2330 struct mptcp_subflow_context *subflow;
2332 if (delta >= TCP_TIMEWAIT_LEN)
2335 /* if all subflows are in closed status don't bother with additional
2338 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2339 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2346 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2348 struct mptcp_subflow_context *subflow, *tmp;
2349 struct sock *sk = &msk->sk.icsk_inet.sk;
2351 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2354 mptcp_token_destroy(msk);
2356 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2357 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2360 slow = lock_sock_fast(tcp_sk);
2361 if (tcp_sk->sk_state != TCP_CLOSE) {
2362 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2363 tcp_set_state(tcp_sk, TCP_CLOSE);
2365 unlock_sock_fast(tcp_sk, slow);
2368 inet_sk_state_store(sk, TCP_CLOSE);
2369 sk->sk_shutdown = SHUTDOWN_MASK;
2370 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2371 set_bit(MPTCP_DATA_READY, &msk->flags);
2372 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2374 mptcp_close_wake_up(sk);
2377 static void __mptcp_retrans(struct sock *sk)
2379 struct mptcp_sock *msk = mptcp_sk(sk);
2380 struct mptcp_sendmsg_info info = {};
2381 struct mptcp_data_frag *dfrag;
2386 mptcp_clean_una_wakeup(sk);
2387 dfrag = mptcp_rtx_head(sk);
2389 if (mptcp_data_fin_enabled(msk)) {
2390 struct inet_connection_sock *icsk = inet_csk(sk);
2392 icsk->icsk_retransmits++;
2393 mptcp_set_datafin_timeout(sk);
2394 mptcp_send_ack(msk);
2402 ssk = mptcp_subflow_get_retrans(msk);
2408 /* limit retransmission to the bytes already sent on some subflows */
2410 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2411 while (info.sent < info.limit) {
2412 if (!mptcp_alloc_tx_skb(sk, ssk))
2415 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2419 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2424 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2425 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2432 if (!mptcp_timer_pending(sk))
2433 mptcp_reset_timer(sk);
2436 static void mptcp_worker(struct work_struct *work)
2438 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2439 struct sock *sk = &msk->sk.icsk_inet.sk;
2443 state = sk->sk_state;
2444 if (unlikely(state == TCP_CLOSE))
2447 mptcp_check_data_fin_ack(sk);
2448 mptcp_flush_join_list(msk);
2450 mptcp_check_fastclose(msk);
2453 mptcp_pm_nl_work(msk);
2455 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2456 mptcp_check_for_eof(msk);
2458 __mptcp_check_send_data_fin(sk);
2459 mptcp_check_data_fin(sk);
2461 /* There is no point in keeping around an orphaned sk timedout or
2462 * closed, but we need the msk around to reply to incoming DATA_FIN,
2463 * even if it is orphaned and in FIN_WAIT2 state
2465 if (sock_flag(sk, SOCK_DEAD) &&
2466 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2467 inet_sk_state_store(sk, TCP_CLOSE);
2468 __mptcp_destroy_sock(sk);
2472 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2473 __mptcp_close_subflow(msk);
2475 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2476 __mptcp_retrans(sk);
2483 static int __mptcp_init_sock(struct sock *sk)
2485 struct mptcp_sock *msk = mptcp_sk(sk);
2487 spin_lock_init(&msk->join_list_lock);
2489 INIT_LIST_HEAD(&msk->conn_list);
2490 INIT_LIST_HEAD(&msk->join_list);
2491 INIT_LIST_HEAD(&msk->rtx_queue);
2492 INIT_WORK(&msk->work, mptcp_worker);
2493 __skb_queue_head_init(&msk->receive_queue);
2494 msk->out_of_order_queue = RB_ROOT;
2495 msk->first_pending = NULL;
2496 msk->wmem_reserved = 0;
2497 WRITE_ONCE(msk->rmem_released, 0);
2498 msk->tx_pending_data = 0;
2499 msk->timer_ival = TCP_RTO_MIN;
2502 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2503 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2504 msk->recovery = false;
2506 mptcp_pm_data_init(msk);
2508 /* re-use the csk retrans timer for MPTCP-level retrans */
2509 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2510 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2515 static int mptcp_init_sock(struct sock *sk)
2517 struct inet_connection_sock *icsk = inet_csk(sk);
2518 struct net *net = sock_net(sk);
2521 ret = __mptcp_init_sock(sk);
2525 if (!mptcp_is_enabled(net))
2526 return -ENOPROTOOPT;
2528 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2531 ret = __mptcp_socket_create(mptcp_sk(sk));
2535 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2536 * propagate the correct value
2538 tcp_assign_congestion_control(sk);
2539 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2541 /* no need to keep a reference to the ops, the name will suffice */
2542 tcp_cleanup_congestion_control(sk);
2543 icsk->icsk_ca_ops = NULL;
2545 sk_sockets_allocated_inc(sk);
2546 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2547 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2552 static void __mptcp_clear_xmit(struct sock *sk)
2554 struct mptcp_sock *msk = mptcp_sk(sk);
2555 struct mptcp_data_frag *dtmp, *dfrag;
2557 WRITE_ONCE(msk->first_pending, NULL);
2558 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2559 dfrag_clear(sk, dfrag);
2562 static void mptcp_cancel_work(struct sock *sk)
2564 struct mptcp_sock *msk = mptcp_sk(sk);
2566 if (cancel_work_sync(&msk->work))
2570 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2574 switch (ssk->sk_state) {
2576 if (!(how & RCV_SHUTDOWN))
2580 tcp_disconnect(ssk, O_NONBLOCK);
2583 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2584 pr_debug("Fallback");
2585 ssk->sk_shutdown |= how;
2586 tcp_shutdown(ssk, how);
2588 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2590 if (!mptcp_timer_pending(sk))
2591 mptcp_reset_timer(sk);
2599 static const unsigned char new_state[16] = {
2600 /* current state: new state: action: */
2601 [0 /* (Invalid) */] = TCP_CLOSE,
2602 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2603 [TCP_SYN_SENT] = TCP_CLOSE,
2604 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2605 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2606 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2607 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2608 [TCP_CLOSE] = TCP_CLOSE,
2609 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2610 [TCP_LAST_ACK] = TCP_LAST_ACK,
2611 [TCP_LISTEN] = TCP_CLOSE,
2612 [TCP_CLOSING] = TCP_CLOSING,
2613 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2616 static int mptcp_close_state(struct sock *sk)
2618 int next = (int)new_state[sk->sk_state];
2619 int ns = next & TCP_STATE_MASK;
2621 inet_sk_state_store(sk, ns);
2623 return next & TCP_ACTION_FIN;
2626 static void __mptcp_check_send_data_fin(struct sock *sk)
2628 struct mptcp_subflow_context *subflow;
2629 struct mptcp_sock *msk = mptcp_sk(sk);
2631 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2632 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2633 msk->snd_nxt, msk->write_seq);
2635 /* we still need to enqueue subflows or not really shutting down,
2638 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2639 mptcp_send_head(sk))
2642 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2644 /* fallback socket will not get data_fin/ack, can move to the next
2647 if (__mptcp_check_fallback(msk)) {
2648 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2649 inet_sk_state_store(sk, TCP_CLOSE);
2650 mptcp_close_wake_up(sk);
2651 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2652 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2656 mptcp_flush_join_list(msk);
2657 mptcp_for_each_subflow(msk, subflow) {
2658 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2660 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2664 static void __mptcp_wr_shutdown(struct sock *sk)
2666 struct mptcp_sock *msk = mptcp_sk(sk);
2668 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2669 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2670 !!mptcp_send_head(sk));
2672 /* will be ignored by fallback sockets */
2673 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2674 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2676 __mptcp_check_send_data_fin(sk);
2679 static void __mptcp_destroy_sock(struct sock *sk)
2681 struct mptcp_subflow_context *subflow, *tmp;
2682 struct mptcp_sock *msk = mptcp_sk(sk);
2683 LIST_HEAD(conn_list);
2685 pr_debug("msk=%p", msk);
2689 /* be sure to always acquire the join list lock, to sync vs
2690 * mptcp_finish_join().
2692 spin_lock_bh(&msk->join_list_lock);
2693 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2694 spin_unlock_bh(&msk->join_list_lock);
2695 list_splice_init(&msk->conn_list, &conn_list);
2697 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2698 sk_stop_timer(sk, &sk->sk_timer);
2701 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2702 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2703 __mptcp_close_ssk(sk, ssk, subflow);
2706 sk->sk_prot->destroy(sk);
2708 WARN_ON_ONCE(msk->wmem_reserved);
2709 WARN_ON_ONCE(msk->rmem_released);
2710 sk_stream_kill_queues(sk);
2711 xfrm_sk_free_policy(sk);
2713 sk_refcnt_debug_release(sk);
2714 mptcp_dispose_initial_subflow(msk);
2718 static void mptcp_close(struct sock *sk, long timeout)
2720 struct mptcp_subflow_context *subflow;
2721 bool do_cancel_work = false;
2724 sk->sk_shutdown = SHUTDOWN_MASK;
2726 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2727 inet_sk_state_store(sk, TCP_CLOSE);
2731 if (mptcp_close_state(sk))
2732 __mptcp_wr_shutdown(sk);
2734 sk_stream_wait_close(sk, timeout);
2737 /* orphan all the subflows */
2738 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2739 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2740 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2741 bool slow = lock_sock_fast(ssk);
2744 unlock_sock_fast(ssk, slow);
2749 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2750 if (sk->sk_state == TCP_CLOSE) {
2751 __mptcp_destroy_sock(sk);
2752 do_cancel_work = true;
2754 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2758 mptcp_cancel_work(sk);
2760 if (mptcp_sk(sk)->token)
2761 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2766 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2768 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2769 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2770 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2772 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2773 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2776 msk6->saddr = ssk6->saddr;
2777 msk6->flow_label = ssk6->flow_label;
2781 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2782 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2783 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2784 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2785 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2786 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2789 static int mptcp_disconnect(struct sock *sk, int flags)
2791 struct mptcp_subflow_context *subflow;
2792 struct mptcp_sock *msk = mptcp_sk(sk);
2794 mptcp_do_flush_join_list(msk);
2796 mptcp_for_each_subflow(msk, subflow) {
2797 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2800 tcp_disconnect(ssk, flags);
2806 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2807 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2809 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2811 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2815 struct sock *mptcp_sk_clone(const struct sock *sk,
2816 const struct mptcp_options_received *mp_opt,
2817 struct request_sock *req)
2819 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2820 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2821 struct mptcp_sock *msk;
2827 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2828 if (nsk->sk_family == AF_INET6)
2829 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2832 __mptcp_init_sock(nsk);
2834 msk = mptcp_sk(nsk);
2835 msk->local_key = subflow_req->local_key;
2836 msk->token = subflow_req->token;
2837 msk->subflow = NULL;
2838 WRITE_ONCE(msk->fully_established, false);
2839 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
2840 WRITE_ONCE(msk->csum_enabled, true);
2842 msk->write_seq = subflow_req->idsn + 1;
2843 msk->snd_nxt = msk->write_seq;
2844 msk->snd_una = msk->write_seq;
2845 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2846 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2848 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
2849 msk->can_ack = true;
2850 msk->remote_key = mp_opt->sndr_key;
2851 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2853 WRITE_ONCE(msk->ack_seq, ack_seq);
2854 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2857 sock_reset_flag(nsk, SOCK_RCU_FREE);
2858 /* will be fully established after successful MPC subflow creation */
2859 inet_sk_state_store(nsk, TCP_SYN_RECV);
2861 security_inet_csk_clone(nsk, req);
2862 bh_unlock_sock(nsk);
2864 /* keep a single reference */
2869 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2871 const struct tcp_sock *tp = tcp_sk(ssk);
2873 msk->rcvq_space.copied = 0;
2874 msk->rcvq_space.rtt_us = 0;
2876 msk->rcvq_space.time = tp->tcp_mstamp;
2878 /* initial rcv_space offering made to peer */
2879 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2880 TCP_INIT_CWND * tp->advmss);
2881 if (msk->rcvq_space.space == 0)
2882 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2884 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2887 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2890 struct mptcp_sock *msk = mptcp_sk(sk);
2891 struct socket *listener;
2894 listener = __mptcp_nmpc_socket(msk);
2895 if (WARN_ON_ONCE(!listener)) {
2900 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2901 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2905 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2906 if (sk_is_mptcp(newsk)) {
2907 struct mptcp_subflow_context *subflow;
2908 struct sock *new_mptcp_sock;
2910 subflow = mptcp_subflow_ctx(newsk);
2911 new_mptcp_sock = subflow->conn;
2913 /* is_mptcp should be false if subflow->conn is missing, see
2914 * subflow_syn_recv_sock()
2916 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2917 tcp_sk(newsk)->is_mptcp = 0;
2921 /* acquire the 2nd reference for the owning socket */
2922 sock_hold(new_mptcp_sock);
2923 newsk = new_mptcp_sock;
2924 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2926 MPTCP_INC_STATS(sock_net(sk),
2927 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2933 void mptcp_destroy_common(struct mptcp_sock *msk)
2935 struct sock *sk = (struct sock *)msk;
2937 __mptcp_clear_xmit(sk);
2939 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2940 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2942 skb_rbtree_purge(&msk->out_of_order_queue);
2943 mptcp_token_destroy(msk);
2944 mptcp_pm_free_anno_list(msk);
2947 static void mptcp_destroy(struct sock *sk)
2949 struct mptcp_sock *msk = mptcp_sk(sk);
2951 mptcp_destroy_common(msk);
2952 sk_sockets_allocated_dec(sk);
2955 void __mptcp_data_acked(struct sock *sk)
2957 if (!sock_owned_by_user(sk))
2958 __mptcp_clean_una(sk);
2960 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2962 if (mptcp_pending_data_fin_ack(sk))
2963 mptcp_schedule_work(sk);
2966 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2968 if (!mptcp_send_head(sk))
2971 if (!sock_owned_by_user(sk)) {
2972 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2974 if (xmit_ssk == ssk)
2975 __mptcp_subflow_push_pending(sk, ssk);
2977 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
2979 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2983 /* processes deferred events and flush wmem */
2984 static void mptcp_release_cb(struct sock *sk)
2987 unsigned long flags = 0;
2989 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2990 flags |= BIT(MPTCP_PUSH_PENDING);
2991 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2992 flags |= BIT(MPTCP_RETRANSMIT);
2996 /* the following actions acquire the subflow socket lock
2998 * 1) can't be invoked in atomic scope
2999 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3000 * datapath acquires the msk socket spinlock while helding
3001 * the subflow socket lock
3004 spin_unlock_bh(&sk->sk_lock.slock);
3005 if (flags & BIT(MPTCP_PUSH_PENDING))
3006 __mptcp_push_pending(sk, 0);
3007 if (flags & BIT(MPTCP_RETRANSMIT))
3008 __mptcp_retrans(sk);
3011 spin_lock_bh(&sk->sk_lock.slock);
3014 /* be sure to set the current sk state before tacking actions
3015 * depending on sk_state
3017 if (test_and_clear_bit(MPTCP_CONNECTED, &mptcp_sk(sk)->flags))
3018 __mptcp_set_connected(sk);
3019 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
3020 __mptcp_clean_una_wakeup(sk);
3021 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
3022 __mptcp_error_report(sk);
3024 /* push_pending may touch wmem_reserved, ensure we do the cleanup
3027 __mptcp_update_wmem(sk);
3028 __mptcp_update_rmem(sk);
3031 void mptcp_subflow_process_delegated(struct sock *ssk)
3033 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3034 struct sock *sk = subflow->conn;
3036 mptcp_data_lock(sk);
3037 if (!sock_owned_by_user(sk))
3038 __mptcp_subflow_push_pending(sk, ssk);
3040 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
3041 mptcp_data_unlock(sk);
3042 mptcp_subflow_delegated_done(subflow);
3045 static int mptcp_hash(struct sock *sk)
3047 /* should never be called,
3048 * we hash the TCP subflows not the master socket
3054 static void mptcp_unhash(struct sock *sk)
3056 /* called from sk_common_release(), but nothing to do here */
3059 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3061 struct mptcp_sock *msk = mptcp_sk(sk);
3062 struct socket *ssock;
3064 ssock = __mptcp_nmpc_socket(msk);
3065 pr_debug("msk=%p, subflow=%p", msk, ssock);
3066 if (WARN_ON_ONCE(!ssock))
3069 return inet_csk_get_port(ssock->sk, snum);
3072 void mptcp_finish_connect(struct sock *ssk)
3074 struct mptcp_subflow_context *subflow;
3075 struct mptcp_sock *msk;
3079 subflow = mptcp_subflow_ctx(ssk);
3083 pr_debug("msk=%p, token=%u", sk, subflow->token);
3085 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3087 subflow->map_seq = ack_seq;
3088 subflow->map_subflow_seq = 1;
3090 /* the socket is not connected yet, no msk/subflow ops can access/race
3091 * accessing the field below
3093 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3094 WRITE_ONCE(msk->local_key, subflow->local_key);
3095 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3096 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3097 WRITE_ONCE(msk->ack_seq, ack_seq);
3098 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3099 WRITE_ONCE(msk->can_ack, 1);
3100 WRITE_ONCE(msk->snd_una, msk->write_seq);
3102 mptcp_pm_new_connection(msk, ssk, 0);
3104 mptcp_rcv_space_init(msk, ssk);
3107 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3109 write_lock_bh(&sk->sk_callback_lock);
3110 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3111 sk_set_socket(sk, parent);
3112 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3113 write_unlock_bh(&sk->sk_callback_lock);
3116 bool mptcp_finish_join(struct sock *ssk)
3118 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3119 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3120 struct sock *parent = (void *)msk;
3121 struct socket *parent_sock;
3124 pr_debug("msk=%p, subflow=%p", msk, subflow);
3126 /* mptcp socket already closing? */
3127 if (!mptcp_is_fully_established(parent)) {
3128 subflow->reset_reason = MPTCP_RST_EMPTCP;
3132 if (!msk->pm.server_side)
3135 if (!mptcp_pm_allow_new_subflow(msk)) {
3136 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3140 /* active connections are already on conn_list, and we can't acquire
3142 * use the join list lock as synchronization point and double-check
3143 * msk status to avoid racing with __mptcp_destroy_sock()
3145 spin_lock_bh(&msk->join_list_lock);
3146 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3147 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3148 list_add_tail(&subflow->node, &msk->join_list);
3151 spin_unlock_bh(&msk->join_list_lock);
3153 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3157 /* attach to msk socket only after we are sure he will deal with us
3160 parent_sock = READ_ONCE(parent->sk_socket);
3161 if (parent_sock && !ssk->sk_socket)
3162 mptcp_sock_graft(ssk, parent_sock);
3163 subflow->map_seq = READ_ONCE(msk->ack_seq);
3165 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3169 static void mptcp_shutdown(struct sock *sk, int how)
3171 pr_debug("sk=%p, how=%d", sk, how);
3173 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3174 __mptcp_wr_shutdown(sk);
3177 static struct proto mptcp_prot = {
3179 .owner = THIS_MODULE,
3180 .init = mptcp_init_sock,
3181 .disconnect = mptcp_disconnect,
3182 .close = mptcp_close,
3183 .accept = mptcp_accept,
3184 .setsockopt = mptcp_setsockopt,
3185 .getsockopt = mptcp_getsockopt,
3186 .shutdown = mptcp_shutdown,
3187 .destroy = mptcp_destroy,
3188 .sendmsg = mptcp_sendmsg,
3189 .recvmsg = mptcp_recvmsg,
3190 .release_cb = mptcp_release_cb,
3192 .unhash = mptcp_unhash,
3193 .get_port = mptcp_get_port,
3194 .sockets_allocated = &mptcp_sockets_allocated,
3195 .memory_allocated = &tcp_memory_allocated,
3196 .memory_pressure = &tcp_memory_pressure,
3197 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3198 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3199 .sysctl_mem = sysctl_tcp_mem,
3200 .obj_size = sizeof(struct mptcp_sock),
3201 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3202 .no_autobind = true,
3205 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3207 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3208 struct socket *ssock;
3211 lock_sock(sock->sk);
3212 ssock = __mptcp_nmpc_socket(msk);
3218 err = ssock->ops->bind(ssock, uaddr, addr_len);
3220 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3223 release_sock(sock->sk);
3227 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3228 struct mptcp_subflow_context *subflow)
3230 subflow->request_mptcp = 0;
3231 __mptcp_do_fallback(msk);
3234 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3235 int addr_len, int flags)
3237 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3238 struct mptcp_subflow_context *subflow;
3239 struct socket *ssock;
3242 lock_sock(sock->sk);
3243 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3244 /* pending connection or invalid state, let existing subflow
3247 ssock = msk->subflow;
3251 ssock = __mptcp_nmpc_socket(msk);
3257 mptcp_token_destroy(msk);
3258 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3259 subflow = mptcp_subflow_ctx(ssock->sk);
3260 #ifdef CONFIG_TCP_MD5SIG
3261 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3264 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3265 mptcp_subflow_early_fallback(msk, subflow);
3267 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3268 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3269 mptcp_subflow_early_fallback(msk, subflow);
3271 if (likely(!__mptcp_check_fallback(msk)))
3272 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3275 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3276 sock->state = ssock->state;
3278 /* on successful connect, the msk state will be moved to established by
3279 * subflow_finish_connect()
3281 if (!err || err == -EINPROGRESS)
3282 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3284 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3287 release_sock(sock->sk);
3291 static int mptcp_listen(struct socket *sock, int backlog)
3293 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3294 struct socket *ssock;
3297 pr_debug("msk=%p", msk);
3299 lock_sock(sock->sk);
3300 ssock = __mptcp_nmpc_socket(msk);
3306 mptcp_token_destroy(msk);
3307 inet_sk_state_store(sock->sk, TCP_LISTEN);
3308 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3310 err = ssock->ops->listen(ssock, backlog);
3311 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3313 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3316 release_sock(sock->sk);
3320 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3321 int flags, bool kern)
3323 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3324 struct socket *ssock;
3327 pr_debug("msk=%p", msk);
3329 lock_sock(sock->sk);
3330 if (sock->sk->sk_state != TCP_LISTEN)
3333 ssock = __mptcp_nmpc_socket(msk);
3337 clear_bit(MPTCP_DATA_READY, &msk->flags);
3338 sock_hold(ssock->sk);
3339 release_sock(sock->sk);
3341 err = ssock->ops->accept(sock, newsock, flags, kern);
3342 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3343 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3344 struct mptcp_subflow_context *subflow;
3345 struct sock *newsk = newsock->sk;
3349 /* PM/worker can now acquire the first subflow socket
3350 * lock without racing with listener queue cleanup,
3351 * we can notify it, if needed.
3353 * Even if remote has reset the initial subflow by now
3354 * the refcnt is still at least one.
3356 subflow = mptcp_subflow_ctx(msk->first);
3357 list_add(&subflow->node, &msk->conn_list);
3358 sock_hold(msk->first);
3359 if (mptcp_is_fully_established(newsk))
3360 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3362 mptcp_copy_inaddrs(newsk, msk->first);
3363 mptcp_rcv_space_init(msk, msk->first);
3364 mptcp_propagate_sndbuf(newsk, msk->first);
3366 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3367 * This is needed so NOSPACE flag can be set from tcp stack.
3369 mptcp_flush_join_list(msk);
3370 mptcp_for_each_subflow(msk, subflow) {
3371 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3373 if (!ssk->sk_socket)
3374 mptcp_sock_graft(ssk, newsock);
3376 release_sock(newsk);
3379 if (inet_csk_listen_poll(ssock->sk))
3380 set_bit(MPTCP_DATA_READY, &msk->flags);
3381 sock_put(ssock->sk);
3385 release_sock(sock->sk);
3389 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3391 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3395 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3397 struct sock *sk = (struct sock *)msk;
3399 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3400 return EPOLLOUT | EPOLLWRNORM;
3402 if (sk_stream_is_writeable(sk))
3403 return EPOLLOUT | EPOLLWRNORM;
3405 mptcp_set_nospace(sk);
3406 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3407 if (sk_stream_is_writeable(sk))
3408 return EPOLLOUT | EPOLLWRNORM;
3413 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3414 struct poll_table_struct *wait)
3416 struct sock *sk = sock->sk;
3417 struct mptcp_sock *msk;
3422 sock_poll_wait(file, sock, wait);
3424 state = inet_sk_state_load(sk);
3425 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3426 if (state == TCP_LISTEN)
3427 return mptcp_check_readable(msk);
3429 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3430 mask |= mptcp_check_readable(msk);
3431 mask |= mptcp_check_writeable(msk);
3433 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3435 if (sk->sk_shutdown & RCV_SHUTDOWN)
3436 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3438 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3446 static const struct proto_ops mptcp_stream_ops = {
3448 .owner = THIS_MODULE,
3449 .release = inet_release,
3451 .connect = mptcp_stream_connect,
3452 .socketpair = sock_no_socketpair,
3453 .accept = mptcp_stream_accept,
3454 .getname = inet_getname,
3456 .ioctl = inet_ioctl,
3457 .gettstamp = sock_gettstamp,
3458 .listen = mptcp_listen,
3459 .shutdown = inet_shutdown,
3460 .setsockopt = sock_common_setsockopt,
3461 .getsockopt = sock_common_getsockopt,
3462 .sendmsg = inet_sendmsg,
3463 .recvmsg = inet_recvmsg,
3464 .mmap = sock_no_mmap,
3465 .sendpage = inet_sendpage,
3468 static struct inet_protosw mptcp_protosw = {
3469 .type = SOCK_STREAM,
3470 .protocol = IPPROTO_MPTCP,
3471 .prot = &mptcp_prot,
3472 .ops = &mptcp_stream_ops,
3473 .flags = INET_PROTOSW_ICSK,
3476 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3478 struct mptcp_delegated_action *delegated;
3479 struct mptcp_subflow_context *subflow;
3482 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3483 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3484 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3486 bh_lock_sock_nested(ssk);
3487 if (!sock_owned_by_user(ssk) &&
3488 mptcp_subflow_has_delegated_action(subflow))
3489 mptcp_subflow_process_delegated(ssk);
3490 /* ... elsewhere tcp_release_cb_override already processed
3491 * the action or will do at next release_sock().
3492 * In both case must dequeue the subflow here - on the same
3493 * CPU that scheduled it.
3495 bh_unlock_sock(ssk);
3498 if (++work_done == budget)
3502 /* always provide a 0 'work_done' argument, so that napi_complete_done
3503 * will not try accessing the NULL napi->dev ptr
3505 napi_complete_done(napi, 0);
3509 void __init mptcp_proto_init(void)
3511 struct mptcp_delegated_action *delegated;
3514 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3516 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3517 panic("Failed to allocate MPTCP pcpu counter\n");
3519 init_dummy_netdev(&mptcp_napi_dev);
3520 for_each_possible_cpu(cpu) {
3521 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3522 INIT_LIST_HEAD(&delegated->head);
3523 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3525 napi_enable(&delegated->napi);
3528 mptcp_subflow_init();
3532 if (proto_register(&mptcp_prot, 1) != 0)
3533 panic("Failed to register MPTCP proto.\n");
3535 inet_register_protosw(&mptcp_protosw);
3537 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3540 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3541 static const struct proto_ops mptcp_v6_stream_ops = {
3543 .owner = THIS_MODULE,
3544 .release = inet6_release,
3546 .connect = mptcp_stream_connect,
3547 .socketpair = sock_no_socketpair,
3548 .accept = mptcp_stream_accept,
3549 .getname = inet6_getname,
3551 .ioctl = inet6_ioctl,
3552 .gettstamp = sock_gettstamp,
3553 .listen = mptcp_listen,
3554 .shutdown = inet_shutdown,
3555 .setsockopt = sock_common_setsockopt,
3556 .getsockopt = sock_common_getsockopt,
3557 .sendmsg = inet6_sendmsg,
3558 .recvmsg = inet6_recvmsg,
3559 .mmap = sock_no_mmap,
3560 .sendpage = inet_sendpage,
3561 #ifdef CONFIG_COMPAT
3562 .compat_ioctl = inet6_compat_ioctl,
3566 static struct proto mptcp_v6_prot;
3568 static void mptcp_v6_destroy(struct sock *sk)
3571 inet6_destroy_sock(sk);
3574 static struct inet_protosw mptcp_v6_protosw = {
3575 .type = SOCK_STREAM,
3576 .protocol = IPPROTO_MPTCP,
3577 .prot = &mptcp_v6_prot,
3578 .ops = &mptcp_v6_stream_ops,
3579 .flags = INET_PROTOSW_ICSK,
3582 int __init mptcp_proto_v6_init(void)
3586 mptcp_v6_prot = mptcp_prot;
3587 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3588 mptcp_v6_prot.slab = NULL;
3589 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3590 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3592 err = proto_register(&mptcp_v6_prot, 1);
3596 err = inet6_register_protosw(&mptcp_v6_protosw);
3598 proto_unregister(&mptcp_v6_prot);