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);
411 retransmits = min_t(u32, icsk->icsk_retransmits,
412 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
414 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
417 static void __mptcp_set_timeout(struct sock *sk, long tout)
419 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
422 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
424 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
426 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
427 inet_csk(ssk)->icsk_timeout - jiffies : 0;
430 static void mptcp_set_timeout(struct sock *sk)
432 struct mptcp_subflow_context *subflow;
435 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
436 tout = max(tout, mptcp_timeout_from_subflow(subflow));
437 __mptcp_set_timeout(sk, tout);
440 static bool tcp_can_send_ack(const struct sock *ssk)
442 return !((1 << inet_sk_state_load(ssk)) &
443 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
446 void mptcp_subflow_send_ack(struct sock *ssk)
450 slow = lock_sock_fast(ssk);
451 if (tcp_can_send_ack(ssk))
453 unlock_sock_fast(ssk, slow);
456 static void mptcp_send_ack(struct mptcp_sock *msk)
458 struct mptcp_subflow_context *subflow;
460 mptcp_for_each_subflow(msk, subflow)
461 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
464 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
468 slow = lock_sock_fast(ssk);
469 if (tcp_can_send_ack(ssk))
470 tcp_cleanup_rbuf(ssk, 1);
471 unlock_sock_fast(ssk, slow);
474 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
476 const struct inet_connection_sock *icsk = inet_csk(ssk);
477 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
478 const struct tcp_sock *tp = tcp_sk(ssk);
480 return (ack_pending & ICSK_ACK_SCHED) &&
481 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
482 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
483 (rx_empty && ack_pending &
484 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
487 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
489 int old_space = READ_ONCE(msk->old_wspace);
490 struct mptcp_subflow_context *subflow;
491 struct sock *sk = (struct sock *)msk;
492 int space = __mptcp_space(sk);
493 bool cleanup, rx_empty;
495 cleanup = (space > 0) && (space >= (old_space << 1));
496 rx_empty = !__mptcp_rmem(sk);
498 mptcp_for_each_subflow(msk, subflow) {
499 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
501 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
502 mptcp_subflow_cleanup_rbuf(ssk);
506 static bool mptcp_check_data_fin(struct sock *sk)
508 struct mptcp_sock *msk = mptcp_sk(sk);
509 u64 rcv_data_fin_seq;
512 if (__mptcp_check_fallback(msk))
515 /* Need to ack a DATA_FIN received from a peer while this side
516 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
517 * msk->rcv_data_fin was set when parsing the incoming options
518 * at the subflow level and the msk lock was not held, so this
519 * is the first opportunity to act on the DATA_FIN and change
522 * If we are caught up to the sequence number of the incoming
523 * DATA_FIN, send the DATA_ACK now and do state transition. If
524 * not caught up, do nothing and let the recv code send DATA_ACK
528 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
529 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
530 WRITE_ONCE(msk->rcv_data_fin, 0);
532 sk->sk_shutdown |= RCV_SHUTDOWN;
533 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
535 switch (sk->sk_state) {
536 case TCP_ESTABLISHED:
537 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
540 inet_sk_state_store(sk, TCP_CLOSING);
543 inet_sk_state_store(sk, TCP_CLOSE);
546 /* Other states not expected */
553 mptcp_close_wake_up(sk);
558 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
562 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
563 struct sock *sk = (struct sock *)msk;
564 unsigned int moved = 0;
565 bool more_data_avail;
570 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
572 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
573 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
575 if (unlikely(ssk_rbuf > sk_rbuf)) {
576 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
581 pr_debug("msk=%p ssk=%p", msk, ssk);
584 u32 map_remaining, offset;
585 u32 seq = tp->copied_seq;
589 /* try to move as much data as available */
590 map_remaining = subflow->map_data_len -
591 mptcp_subflow_get_map_offset(subflow);
593 skb = skb_peek(&ssk->sk_receive_queue);
595 /* if no data is found, a racing workqueue/recvmsg
596 * already processed the new data, stop here or we
597 * can enter an infinite loop
604 if (__mptcp_check_fallback(msk)) {
605 /* if we are running under the workqueue, TCP could have
606 * collapsed skbs between dummy map creation and now
607 * be sure to adjust the size
609 map_remaining = skb->len;
610 subflow->map_data_len = skb->len;
613 offset = seq - TCP_SKB_CB(skb)->seq;
614 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
620 if (offset < skb->len) {
621 size_t len = skb->len - offset;
626 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
630 if (WARN_ON_ONCE(map_remaining < len))
634 sk_eat_skb(ssk, skb);
638 WRITE_ONCE(tp->copied_seq, seq);
639 more_data_avail = mptcp_subflow_data_available(ssk);
641 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
645 } while (more_data_avail);
651 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
653 struct sock *sk = (struct sock *)msk;
654 struct sk_buff *skb, *tail;
659 p = rb_first(&msk->out_of_order_queue);
660 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
663 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
667 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
669 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
672 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
676 end_seq = MPTCP_SKB_CB(skb)->end_seq;
677 tail = skb_peek_tail(&sk->sk_receive_queue);
678 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
679 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
681 /* skip overlapping data, if any */
682 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
683 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
685 MPTCP_SKB_CB(skb)->offset += delta;
686 __skb_queue_tail(&sk->sk_receive_queue, skb);
688 msk->ack_seq = end_seq;
694 /* In most cases we will be able to lock the mptcp socket. If its already
695 * owned, we need to defer to the work queue to avoid ABBA deadlock.
697 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
699 struct sock *sk = (struct sock *)msk;
700 unsigned int moved = 0;
702 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
703 __mptcp_ofo_queue(msk);
704 if (unlikely(ssk->sk_err)) {
705 if (!sock_owned_by_user(sk))
706 __mptcp_error_report(sk);
708 set_bit(MPTCP_ERROR_REPORT, &msk->flags);
711 /* If the moves have caught up with the DATA_FIN sequence number
712 * it's time to ack the DATA_FIN and change socket state, but
713 * this is not a good place to change state. Let the workqueue
716 if (mptcp_pending_data_fin(sk, NULL))
717 mptcp_schedule_work(sk);
721 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
723 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
724 struct mptcp_sock *msk = mptcp_sk(sk);
725 int sk_rbuf, ssk_rbuf;
727 /* The peer can send data while we are shutting down this
728 * subflow at msk destruction time, but we must avoid enqueuing
729 * more data to the msk receive queue
731 if (unlikely(subflow->disposable))
734 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
735 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
736 if (unlikely(ssk_rbuf > sk_rbuf))
739 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
740 if (__mptcp_rmem(sk) > sk_rbuf) {
741 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
745 /* Wake-up the reader only for in-sequence data */
747 if (move_skbs_to_msk(msk, ssk))
748 sk->sk_data_ready(sk);
750 mptcp_data_unlock(sk);
753 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
755 struct mptcp_subflow_context *subflow;
758 if (likely(list_empty(&msk->join_list)))
761 spin_lock_bh(&msk->join_list_lock);
762 list_for_each_entry(subflow, &msk->join_list, node) {
763 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
765 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
766 if (READ_ONCE(msk->setsockopt_seq) != sseq)
769 list_splice_tail_init(&msk->join_list, &msk->conn_list);
770 spin_unlock_bh(&msk->join_list_lock);
775 void __mptcp_flush_join_list(struct mptcp_sock *msk)
777 if (likely(!mptcp_do_flush_join_list(msk)))
780 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
781 mptcp_schedule_work((struct sock *)msk);
784 static void mptcp_flush_join_list(struct mptcp_sock *msk)
786 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
790 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
793 mptcp_sockopt_sync_all(msk);
796 static bool mptcp_timer_pending(struct sock *sk)
798 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
801 static void mptcp_reset_timer(struct sock *sk)
803 struct inet_connection_sock *icsk = inet_csk(sk);
806 /* prevent rescheduling on close */
807 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
810 tout = mptcp_sk(sk)->timer_ival;
811 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
814 bool mptcp_schedule_work(struct sock *sk)
816 if (inet_sk_state_load(sk) != TCP_CLOSE &&
817 schedule_work(&mptcp_sk(sk)->work)) {
818 /* each subflow already holds a reference to the sk, and the
819 * workqueue is invoked by a subflow, so sk can't go away here.
827 void mptcp_subflow_eof(struct sock *sk)
829 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
830 mptcp_schedule_work(sk);
833 static void mptcp_check_for_eof(struct mptcp_sock *msk)
835 struct mptcp_subflow_context *subflow;
836 struct sock *sk = (struct sock *)msk;
839 mptcp_for_each_subflow(msk, subflow)
840 receivers += !subflow->rx_eof;
844 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
845 /* hopefully temporary hack: propagate shutdown status
846 * to msk, when all subflows agree on it
848 sk->sk_shutdown |= RCV_SHUTDOWN;
850 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
851 sk->sk_data_ready(sk);
854 switch (sk->sk_state) {
855 case TCP_ESTABLISHED:
856 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
859 inet_sk_state_store(sk, TCP_CLOSING);
862 inet_sk_state_store(sk, TCP_CLOSE);
867 mptcp_close_wake_up(sk);
870 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
872 struct mptcp_subflow_context *subflow;
873 struct sock *sk = (struct sock *)msk;
875 sock_owned_by_me(sk);
877 mptcp_for_each_subflow(msk, subflow) {
878 if (READ_ONCE(subflow->data_avail))
879 return mptcp_subflow_tcp_sock(subflow);
885 static bool mptcp_skb_can_collapse_to(u64 write_seq,
886 const struct sk_buff *skb,
887 const struct mptcp_ext *mpext)
889 if (!tcp_skb_can_collapse_to(skb))
892 /* can collapse only if MPTCP level sequence is in order and this
893 * mapping has not been xmitted yet
895 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
899 /* we can append data to the given data frag if:
900 * - there is space available in the backing page_frag
901 * - the data frag tail matches the current page_frag free offset
902 * - the data frag end sequence number matches the current write seq
904 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
905 const struct page_frag *pfrag,
906 const struct mptcp_data_frag *df)
908 return df && pfrag->page == df->page &&
909 pfrag->size - pfrag->offset > 0 &&
910 pfrag->offset == (df->offset + df->data_len) &&
911 df->data_seq + df->data_len == msk->write_seq;
914 static int mptcp_wmem_with_overhead(int size)
916 return size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
919 static void __mptcp_wmem_reserve(struct sock *sk, int size)
921 int amount = mptcp_wmem_with_overhead(size);
922 struct mptcp_sock *msk = mptcp_sk(sk);
924 WARN_ON_ONCE(msk->wmem_reserved);
925 if (WARN_ON_ONCE(amount < 0))
928 if (amount <= sk->sk_forward_alloc)
931 /* under memory pressure try to reserve at most a single page
932 * otherwise try to reserve the full estimate and fallback
933 * to a single page before entering the error path
935 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
936 !sk_wmem_schedule(sk, amount)) {
937 if (amount <= PAGE_SIZE)
941 if (!sk_wmem_schedule(sk, amount))
946 msk->wmem_reserved = amount;
947 sk->sk_forward_alloc -= amount;
951 /* we will wait for memory on next allocation */
952 msk->wmem_reserved = -1;
955 static void __mptcp_update_wmem(struct sock *sk)
957 struct mptcp_sock *msk = mptcp_sk(sk);
959 #ifdef CONFIG_LOCKDEP
960 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
963 if (!msk->wmem_reserved)
966 if (msk->wmem_reserved < 0)
967 msk->wmem_reserved = 0;
968 if (msk->wmem_reserved > 0) {
969 sk->sk_forward_alloc += msk->wmem_reserved;
970 msk->wmem_reserved = 0;
974 static bool mptcp_wmem_alloc(struct sock *sk, int size)
976 struct mptcp_sock *msk = mptcp_sk(sk);
978 /* check for pre-existing error condition */
979 if (msk->wmem_reserved < 0)
982 if (msk->wmem_reserved >= size)
986 if (!sk_wmem_schedule(sk, size)) {
987 mptcp_data_unlock(sk);
991 sk->sk_forward_alloc -= size;
992 msk->wmem_reserved += size;
993 mptcp_data_unlock(sk);
996 msk->wmem_reserved -= size;
1000 static void mptcp_wmem_uncharge(struct sock *sk, int size)
1002 struct mptcp_sock *msk = mptcp_sk(sk);
1004 if (msk->wmem_reserved < 0)
1005 msk->wmem_reserved = 0;
1006 msk->wmem_reserved += size;
1009 static void __mptcp_mem_reclaim_partial(struct sock *sk)
1011 lockdep_assert_held_once(&sk->sk_lock.slock);
1012 __mptcp_update_wmem(sk);
1013 sk_mem_reclaim_partial(sk);
1016 static void mptcp_mem_reclaim_partial(struct sock *sk)
1018 struct mptcp_sock *msk = mptcp_sk(sk);
1020 /* if we are experiencing a transint allocation error,
1021 * the forward allocation memory has been already
1024 if (msk->wmem_reserved < 0)
1027 mptcp_data_lock(sk);
1028 sk->sk_forward_alloc += msk->wmem_reserved;
1029 sk_mem_reclaim_partial(sk);
1030 msk->wmem_reserved = sk->sk_forward_alloc;
1031 sk->sk_forward_alloc = 0;
1032 mptcp_data_unlock(sk);
1035 static void dfrag_uncharge(struct sock *sk, int len)
1037 sk_mem_uncharge(sk, len);
1038 sk_wmem_queued_add(sk, -len);
1041 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1043 int len = dfrag->data_len + dfrag->overhead;
1045 list_del(&dfrag->list);
1046 dfrag_uncharge(sk, len);
1047 put_page(dfrag->page);
1050 static void __mptcp_clean_una(struct sock *sk)
1052 struct mptcp_sock *msk = mptcp_sk(sk);
1053 struct mptcp_data_frag *dtmp, *dfrag;
1054 bool cleaned = false;
1057 /* on fallback we just need to ignore snd_una, as this is really
1060 if (__mptcp_check_fallback(msk))
1061 msk->snd_una = READ_ONCE(msk->snd_nxt);
1063 snd_una = msk->snd_una;
1064 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1065 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1068 if (unlikely(dfrag == msk->first_pending)) {
1069 /* in recovery mode can see ack after the current snd head */
1070 if (WARN_ON_ONCE(!msk->recovery))
1073 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1076 dfrag_clear(sk, dfrag);
1080 dfrag = mptcp_rtx_head(sk);
1081 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1082 u64 delta = snd_una - dfrag->data_seq;
1084 /* prevent wrap around in recovery mode */
1085 if (unlikely(delta > dfrag->already_sent)) {
1086 if (WARN_ON_ONCE(!msk->recovery))
1088 if (WARN_ON_ONCE(delta > dfrag->data_len))
1090 dfrag->already_sent += delta - dfrag->already_sent;
1093 dfrag->data_seq += delta;
1094 dfrag->offset += delta;
1095 dfrag->data_len -= delta;
1096 dfrag->already_sent -= delta;
1098 dfrag_uncharge(sk, delta);
1102 /* all retransmitted data acked, recovery completed */
1103 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1104 msk->recovery = false;
1107 if (cleaned && tcp_under_memory_pressure(sk))
1108 __mptcp_mem_reclaim_partial(sk);
1110 if (snd_una == READ_ONCE(msk->snd_nxt) && !msk->recovery) {
1111 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1112 mptcp_stop_timer(sk);
1114 mptcp_reset_timer(sk);
1118 static void __mptcp_clean_una_wakeup(struct sock *sk)
1120 #ifdef CONFIG_LOCKDEP
1121 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
1123 __mptcp_clean_una(sk);
1124 mptcp_write_space(sk);
1127 static void mptcp_clean_una_wakeup(struct sock *sk)
1129 mptcp_data_lock(sk);
1130 __mptcp_clean_una_wakeup(sk);
1131 mptcp_data_unlock(sk);
1134 static void mptcp_enter_memory_pressure(struct sock *sk)
1136 struct mptcp_subflow_context *subflow;
1137 struct mptcp_sock *msk = mptcp_sk(sk);
1140 sk_stream_moderate_sndbuf(sk);
1141 mptcp_for_each_subflow(msk, subflow) {
1142 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1145 tcp_enter_memory_pressure(ssk);
1146 sk_stream_moderate_sndbuf(ssk);
1151 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1154 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1156 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1157 pfrag, sk->sk_allocation)))
1160 mptcp_enter_memory_pressure(sk);
1164 static struct mptcp_data_frag *
1165 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1168 int offset = ALIGN(orig_offset, sizeof(long));
1169 struct mptcp_data_frag *dfrag;
1171 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1172 dfrag->data_len = 0;
1173 dfrag->data_seq = msk->write_seq;
1174 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1175 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1176 dfrag->already_sent = 0;
1177 dfrag->page = pfrag->page;
1182 struct mptcp_sendmsg_info {
1188 bool data_lock_held;
1191 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1194 u64 window_end = mptcp_wnd_end(msk);
1196 if (__mptcp_check_fallback(msk))
1199 if (!before64(data_seq + avail_size, window_end)) {
1200 u64 allowed_size = window_end - data_seq;
1202 return min_t(unsigned int, allowed_size, avail_size);
1208 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1210 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1214 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1218 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1220 struct sk_buff *skb;
1222 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1224 if (likely(__mptcp_add_ext(skb, gfp))) {
1225 skb_reserve(skb, MAX_TCP_HEADER);
1226 skb->reserved_tailroom = skb->end - skb->tail;
1227 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1232 mptcp_enter_memory_pressure(sk);
1237 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1239 struct sk_buff *skb;
1241 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1245 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1246 tcp_skb_entail(ssk, skb);
1249 tcp_skb_tsorted_anchor_cleanup(skb);
1254 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1256 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1258 if (unlikely(tcp_under_memory_pressure(sk))) {
1260 __mptcp_mem_reclaim_partial(sk);
1262 mptcp_mem_reclaim_partial(sk);
1264 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1267 /* note: this always recompute the csum on the whole skb, even
1268 * if we just appended a single frag. More status info needed
1270 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1272 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1273 __wsum csum = ~csum_unfold(mpext->csum);
1274 int offset = skb->len - added;
1276 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1279 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1280 struct mptcp_data_frag *dfrag,
1281 struct mptcp_sendmsg_info *info)
1283 u64 data_seq = dfrag->data_seq + info->sent;
1284 int offset = dfrag->offset + info->sent;
1285 struct mptcp_sock *msk = mptcp_sk(sk);
1286 bool zero_window_probe = false;
1287 struct mptcp_ext *mpext = NULL;
1288 bool can_coalesce = false;
1289 bool reuse_skb = true;
1290 struct sk_buff *skb;
1294 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1295 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1297 if (WARN_ON_ONCE(info->sent > info->limit ||
1298 info->limit > dfrag->data_len))
1301 /* compute send limit */
1302 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1303 copy = info->size_goal;
1305 skb = tcp_write_queue_tail(ssk);
1306 if (skb && copy > skb->len) {
1307 /* Limit the write to the size available in the
1308 * current skb, if any, so that we create at most a new skb.
1309 * Explicitly tells TCP internals to avoid collapsing on later
1310 * queue management operation, to avoid breaking the ext <->
1311 * SSN association set here
1313 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1314 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1315 TCP_SKB_CB(skb)->eor = 1;
1319 i = skb_shinfo(skb)->nr_frags;
1320 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1321 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1322 tcp_mark_push(tcp_sk(ssk), skb);
1329 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1333 i = skb_shinfo(skb)->nr_frags;
1335 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1338 /* Zero window and all data acked? Probe. */
1339 copy = mptcp_check_allowed_size(msk, data_seq, copy);
1341 u64 snd_una = READ_ONCE(msk->snd_una);
1343 if (snd_una != msk->snd_nxt) {
1344 tcp_remove_empty_skb(ssk, tcp_write_queue_tail(ssk));
1348 zero_window_probe = true;
1349 data_seq = snd_una - 1;
1352 /* all mptcp-level data is acked, no skbs should be present into the
1355 WARN_ON_ONCE(reuse_skb);
1358 copy = min_t(size_t, copy, info->limit - info->sent);
1359 if (!sk_wmem_schedule(ssk, copy)) {
1360 tcp_remove_empty_skb(ssk, tcp_write_queue_tail(ssk));
1365 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1367 get_page(dfrag->page);
1368 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1372 skb->data_len += copy;
1373 skb->truesize += copy;
1374 sk_wmem_queued_add(ssk, copy);
1375 sk_mem_charge(ssk, copy);
1376 skb->ip_summed = CHECKSUM_PARTIAL;
1377 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1378 TCP_SKB_CB(skb)->end_seq += copy;
1379 tcp_skb_pcount_set(skb, 0);
1381 /* on skb reuse we just need to update the DSS len */
1383 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1384 mpext->data_len += copy;
1385 WARN_ON_ONCE(zero_window_probe);
1389 memset(mpext, 0, sizeof(*mpext));
1390 mpext->data_seq = data_seq;
1391 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1392 mpext->data_len = copy;
1396 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1397 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1400 if (zero_window_probe) {
1401 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1403 if (READ_ONCE(msk->csum_enabled))
1404 mptcp_update_data_checksum(skb, copy);
1405 tcp_push_pending_frames(ssk);
1409 if (READ_ONCE(msk->csum_enabled))
1410 mptcp_update_data_checksum(skb, copy);
1411 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1415 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1416 sizeof(struct tcphdr) - \
1417 MAX_TCP_OPTION_SPACE - \
1418 sizeof(struct ipv6hdr) - \
1419 sizeof(struct frag_hdr))
1421 struct subflow_send_info {
1426 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1428 if (!subflow->stale)
1432 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1435 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1437 if (unlikely(subflow->stale)) {
1438 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1440 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1443 mptcp_subflow_set_active(subflow);
1445 return __mptcp_subflow_active(subflow);
1448 /* implement the mptcp packet scheduler;
1449 * returns the subflow that will transmit the next DSS
1450 * additionally updates the rtx timeout
1452 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1454 struct subflow_send_info send_info[2];
1455 struct mptcp_subflow_context *subflow;
1456 struct sock *sk = (struct sock *)msk;
1457 int i, nr_active = 0;
1463 sock_owned_by_me(sk);
1465 if (__mptcp_check_fallback(msk)) {
1468 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1471 /* re-use last subflow, if the burst allow that */
1472 if (msk->last_snd && msk->snd_burst > 0 &&
1473 sk_stream_memory_free(msk->last_snd) &&
1474 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1475 mptcp_set_timeout(sk);
1476 return msk->last_snd;
1479 /* pick the subflow with the lower wmem/wspace ratio */
1480 for (i = 0; i < 2; ++i) {
1481 send_info[i].ssk = NULL;
1482 send_info[i].ratio = -1;
1484 mptcp_for_each_subflow(msk, subflow) {
1485 trace_mptcp_subflow_get_send(subflow);
1486 ssk = mptcp_subflow_tcp_sock(subflow);
1487 if (!mptcp_subflow_active(subflow))
1490 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1491 nr_active += !subflow->backup;
1492 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1495 pace = READ_ONCE(ssk->sk_pacing_rate);
1499 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1501 if (ratio < send_info[subflow->backup].ratio) {
1502 send_info[subflow->backup].ssk = ssk;
1503 send_info[subflow->backup].ratio = ratio;
1506 __mptcp_set_timeout(sk, tout);
1508 /* pick the best backup if no other subflow is active */
1510 send_info[0].ssk = send_info[1].ssk;
1512 if (send_info[0].ssk) {
1513 msk->last_snd = send_info[0].ssk;
1514 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1515 tcp_sk(msk->last_snd)->snd_wnd);
1516 return msk->last_snd;
1522 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1523 struct mptcp_sendmsg_info *info)
1525 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1529 static void mptcp_update_post_push(struct mptcp_sock *msk,
1530 struct mptcp_data_frag *dfrag,
1533 u64 snd_nxt_new = dfrag->data_seq;
1535 dfrag->already_sent += sent;
1537 msk->snd_burst -= sent;
1538 msk->tx_pending_data -= sent;
1540 snd_nxt_new += dfrag->already_sent;
1542 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1543 * is recovering after a failover. In that event, this re-sends
1546 * Thus compute snd_nxt_new candidate based on
1547 * the dfrag->data_seq that was sent and the data
1548 * that has been handed to the subflow for transmission
1549 * and skip update in case it was old dfrag.
1551 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1552 msk->snd_nxt = snd_nxt_new;
1555 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1557 struct sock *prev_ssk = NULL, *ssk = NULL;
1558 struct mptcp_sock *msk = mptcp_sk(sk);
1559 struct mptcp_sendmsg_info info = {
1562 struct mptcp_data_frag *dfrag;
1563 int len, copied = 0;
1565 while ((dfrag = mptcp_send_head(sk))) {
1566 info.sent = dfrag->already_sent;
1567 info.limit = dfrag->data_len;
1568 len = dfrag->data_len - dfrag->already_sent;
1573 __mptcp_flush_join_list(msk);
1574 ssk = mptcp_subflow_get_send(msk);
1576 /* First check. If the ssk has changed since
1577 * the last round, release prev_ssk
1579 if (ssk != prev_ssk && prev_ssk)
1580 mptcp_push_release(sk, prev_ssk, &info);
1584 /* Need to lock the new subflow only if different
1585 * from the previous one, otherwise we are still
1586 * helding the relevant lock
1588 if (ssk != prev_ssk)
1591 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1593 mptcp_push_release(sk, ssk, &info);
1601 mptcp_update_post_push(msk, dfrag, ret);
1603 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1606 /* at this point we held the socket lock for the last subflow we used */
1608 mptcp_push_release(sk, ssk, &info);
1611 /* ensure the rtx timer is running */
1612 if (!mptcp_timer_pending(sk))
1613 mptcp_reset_timer(sk);
1615 __mptcp_check_send_data_fin(sk);
1618 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1620 struct mptcp_sock *msk = mptcp_sk(sk);
1621 struct mptcp_sendmsg_info info = {
1622 .data_lock_held = true,
1624 struct mptcp_data_frag *dfrag;
1625 struct sock *xmit_ssk;
1626 int len, copied = 0;
1630 while ((dfrag = mptcp_send_head(sk))) {
1631 info.sent = dfrag->already_sent;
1632 info.limit = dfrag->data_len;
1633 len = dfrag->data_len - dfrag->already_sent;
1637 /* the caller already invoked the packet scheduler,
1638 * check for a different subflow usage only after
1639 * spooling the first chunk of data
1641 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1644 if (xmit_ssk != ssk) {
1645 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1646 MPTCP_DELEGATE_SEND);
1650 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1659 mptcp_update_post_push(msk, dfrag, ret);
1661 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1665 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1666 * not going to flush it via release_sock()
1668 __mptcp_update_wmem(sk);
1670 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1672 if (!mptcp_timer_pending(sk))
1673 mptcp_reset_timer(sk);
1675 if (msk->snd_data_fin_enable &&
1676 msk->snd_nxt + 1 == msk->write_seq)
1677 mptcp_schedule_work(sk);
1681 static void mptcp_set_nospace(struct sock *sk)
1683 /* enable autotune */
1684 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1686 /* will be cleared on avail space */
1687 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1690 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1692 struct mptcp_sock *msk = mptcp_sk(sk);
1693 struct page_frag *pfrag;
1698 /* we don't support FASTOPEN yet */
1699 if (msg->msg_flags & MSG_FASTOPEN)
1702 /* silently ignore everything else */
1703 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1705 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1707 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1709 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1710 ret = sk_stream_wait_connect(sk, &timeo);
1715 pfrag = sk_page_frag(sk);
1717 while (msg_data_left(msg)) {
1718 int total_ts, frag_truesize = 0;
1719 struct mptcp_data_frag *dfrag;
1720 bool dfrag_collapsed;
1721 size_t psize, offset;
1723 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1728 /* reuse tail pfrag, if possible, or carve a new one from the
1731 dfrag = mptcp_pending_tail(sk);
1732 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1733 if (!dfrag_collapsed) {
1734 if (!sk_stream_memory_free(sk))
1735 goto wait_for_memory;
1737 if (!mptcp_page_frag_refill(sk, pfrag))
1738 goto wait_for_memory;
1740 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1741 frag_truesize = dfrag->overhead;
1744 /* we do not bound vs wspace, to allow a single packet.
1745 * memory accounting will prevent execessive memory usage
1748 offset = dfrag->offset + dfrag->data_len;
1749 psize = pfrag->size - offset;
1750 psize = min_t(size_t, psize, msg_data_left(msg));
1751 total_ts = psize + frag_truesize;
1753 if (!mptcp_wmem_alloc(sk, total_ts))
1754 goto wait_for_memory;
1756 if (copy_page_from_iter(dfrag->page, offset, psize,
1757 &msg->msg_iter) != psize) {
1758 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1763 /* data successfully copied into the write queue */
1765 dfrag->data_len += psize;
1766 frag_truesize += psize;
1767 pfrag->offset += frag_truesize;
1768 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1769 msk->tx_pending_data += psize;
1771 /* charge data on mptcp pending queue to the msk socket
1772 * Note: we charge such data both to sk and ssk
1774 sk_wmem_queued_add(sk, frag_truesize);
1775 if (!dfrag_collapsed) {
1776 get_page(dfrag->page);
1777 list_add_tail(&dfrag->list, &msk->rtx_queue);
1778 if (!msk->first_pending)
1779 WRITE_ONCE(msk->first_pending, dfrag);
1781 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1782 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1788 mptcp_set_nospace(sk);
1789 __mptcp_push_pending(sk, msg->msg_flags);
1790 ret = sk_stream_wait_memory(sk, &timeo);
1796 __mptcp_push_pending(sk, msg->msg_flags);
1800 return copied ? : ret;
1803 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1805 size_t len, int flags,
1806 struct scm_timestamping_internal *tss,
1809 struct sk_buff *skb, *tmp;
1812 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1813 u32 offset = MPTCP_SKB_CB(skb)->offset;
1814 u32 data_len = skb->len - offset;
1815 u32 count = min_t(size_t, len - copied, data_len);
1818 if (!(flags & MSG_TRUNC)) {
1819 err = skb_copy_datagram_msg(skb, offset, msg, count);
1820 if (unlikely(err < 0)) {
1827 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1828 tcp_update_recv_tstamps(skb, tss);
1829 *cmsg_flags |= MPTCP_CMSG_TS;
1834 if (count < data_len) {
1835 if (!(flags & MSG_PEEK))
1836 MPTCP_SKB_CB(skb)->offset += count;
1840 if (!(flags & MSG_PEEK)) {
1841 /* we will bulk release the skb memory later */
1842 skb->destructor = NULL;
1843 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1844 __skb_unlink(skb, &msk->receive_queue);
1855 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1857 * Only difference: Use highest rtt estimate of the subflows in use.
1859 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1861 struct mptcp_subflow_context *subflow;
1862 struct sock *sk = (struct sock *)msk;
1863 u32 time, advmss = 1;
1866 sock_owned_by_me(sk);
1871 msk->rcvq_space.copied += copied;
1873 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1874 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1876 rtt_us = msk->rcvq_space.rtt_us;
1877 if (rtt_us && time < (rtt_us >> 3))
1881 mptcp_for_each_subflow(msk, subflow) {
1882 const struct tcp_sock *tp;
1886 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1888 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1889 sf_advmss = READ_ONCE(tp->advmss);
1891 rtt_us = max(sf_rtt_us, rtt_us);
1892 advmss = max(sf_advmss, advmss);
1895 msk->rcvq_space.rtt_us = rtt_us;
1896 if (time < (rtt_us >> 3) || rtt_us == 0)
1899 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1902 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1903 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1907 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1909 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1911 do_div(grow, msk->rcvq_space.space);
1912 rcvwin += (grow << 1);
1914 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1915 while (tcp_win_from_space(sk, rcvmem) < advmss)
1918 do_div(rcvwin, advmss);
1919 rcvbuf = min_t(u64, rcvwin * rcvmem,
1920 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1922 if (rcvbuf > sk->sk_rcvbuf) {
1925 window_clamp = tcp_win_from_space(sk, rcvbuf);
1926 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1928 /* Make subflows follow along. If we do not do this, we
1929 * get drops at subflow level if skbs can't be moved to
1930 * the mptcp rx queue fast enough (announced rcv_win can
1931 * exceed ssk->sk_rcvbuf).
1933 mptcp_for_each_subflow(msk, subflow) {
1937 ssk = mptcp_subflow_tcp_sock(subflow);
1938 slow = lock_sock_fast(ssk);
1939 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1940 tcp_sk(ssk)->window_clamp = window_clamp;
1941 tcp_cleanup_rbuf(ssk, 1);
1942 unlock_sock_fast(ssk, slow);
1947 msk->rcvq_space.space = msk->rcvq_space.copied;
1949 msk->rcvq_space.copied = 0;
1950 msk->rcvq_space.time = mstamp;
1953 static void __mptcp_update_rmem(struct sock *sk)
1955 struct mptcp_sock *msk = mptcp_sk(sk);
1957 if (!msk->rmem_released)
1960 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1961 sk_mem_uncharge(sk, msk->rmem_released);
1962 WRITE_ONCE(msk->rmem_released, 0);
1965 static void __mptcp_splice_receive_queue(struct sock *sk)
1967 struct mptcp_sock *msk = mptcp_sk(sk);
1969 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1972 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1974 struct sock *sk = (struct sock *)msk;
1975 unsigned int moved = 0;
1978 mptcp_flush_join_list(msk);
1980 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1983 /* we can have data pending in the subflows only if the msk
1984 * receive buffer was full at subflow_data_ready() time,
1985 * that is an unlikely slow path.
1990 slowpath = lock_sock_fast(ssk);
1991 mptcp_data_lock(sk);
1992 __mptcp_update_rmem(sk);
1993 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1994 mptcp_data_unlock(sk);
1996 if (unlikely(ssk->sk_err))
1997 __mptcp_error_report(sk);
1998 unlock_sock_fast(ssk, slowpath);
2001 /* acquire the data lock only if some input data is pending */
2003 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2004 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2005 mptcp_data_lock(sk);
2006 __mptcp_update_rmem(sk);
2007 ret |= __mptcp_ofo_queue(msk);
2008 __mptcp_splice_receive_queue(sk);
2009 mptcp_data_unlock(sk);
2012 mptcp_check_data_fin((struct sock *)msk);
2013 return !skb_queue_empty(&msk->receive_queue);
2016 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2017 int nonblock, int flags, int *addr_len)
2019 struct mptcp_sock *msk = mptcp_sk(sk);
2020 struct scm_timestamping_internal tss;
2021 int copied = 0, cmsg_flags = 0;
2025 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2026 if (unlikely(flags & MSG_ERRQUEUE))
2027 return inet_recv_error(sk, msg, len, addr_len);
2029 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
2030 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2035 timeo = sock_rcvtimeo(sk, nonblock);
2037 len = min_t(size_t, len, INT_MAX);
2038 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2040 while (copied < len) {
2043 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2044 if (unlikely(bytes_read < 0)) {
2046 copied = bytes_read;
2050 copied += bytes_read;
2052 /* be sure to advertise window change */
2053 mptcp_cleanup_rbuf(msk);
2055 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2058 /* only the master socket status is relevant here. The exit
2059 * conditions mirror closely tcp_recvmsg()
2061 if (copied >= target)
2066 sk->sk_state == TCP_CLOSE ||
2067 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2069 signal_pending(current))
2073 copied = sock_error(sk);
2077 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2078 mptcp_check_for_eof(msk);
2080 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2081 /* race breaker: the shutdown could be after the
2082 * previous receive queue check
2084 if (__mptcp_move_skbs(msk))
2089 if (sk->sk_state == TCP_CLOSE) {
2099 if (signal_pending(current)) {
2100 copied = sock_intr_errno(timeo);
2105 pr_debug("block timeout %ld", timeo);
2106 sk_wait_data(sk, &timeo, NULL);
2110 if (cmsg_flags && copied >= 0) {
2111 if (cmsg_flags & MPTCP_CMSG_TS)
2112 tcp_recv_timestamp(msg, sk, &tss);
2115 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2116 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2117 skb_queue_empty(&msk->receive_queue), copied);
2118 if (!(flags & MSG_PEEK))
2119 mptcp_rcv_space_adjust(msk, copied);
2125 static void mptcp_retransmit_timer(struct timer_list *t)
2127 struct inet_connection_sock *icsk = from_timer(icsk, t,
2128 icsk_retransmit_timer);
2129 struct sock *sk = &icsk->icsk_inet.sk;
2130 struct mptcp_sock *msk = mptcp_sk(sk);
2133 if (!sock_owned_by_user(sk)) {
2134 /* we need a process context to retransmit */
2135 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2136 mptcp_schedule_work(sk);
2138 /* delegate our work to tcp_release_cb() */
2139 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2145 static void mptcp_timeout_timer(struct timer_list *t)
2147 struct sock *sk = from_timer(sk, t, sk_timer);
2149 mptcp_schedule_work(sk);
2153 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2156 * A backup subflow is returned only if that is the only kind available.
2158 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2160 struct sock *backup = NULL, *pick = NULL;
2161 struct mptcp_subflow_context *subflow;
2162 int min_stale_count = INT_MAX;
2164 sock_owned_by_me((const struct sock *)msk);
2166 if (__mptcp_check_fallback(msk))
2169 mptcp_for_each_subflow(msk, subflow) {
2170 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2172 if (!__mptcp_subflow_active(subflow))
2175 /* still data outstanding at TCP level? skip this */
2176 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2177 mptcp_pm_subflow_chk_stale(msk, ssk);
2178 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2182 if (subflow->backup) {
2195 /* use backup only if there are no progresses anywhere */
2196 return min_stale_count > 1 ? backup : NULL;
2199 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2202 iput(SOCK_INODE(msk->subflow));
2203 msk->subflow = NULL;
2207 bool __mptcp_retransmit_pending_data(struct sock *sk)
2209 struct mptcp_data_frag *cur, *rtx_head;
2210 struct mptcp_sock *msk = mptcp_sk(sk);
2212 if (__mptcp_check_fallback(mptcp_sk(sk)))
2215 if (tcp_rtx_and_write_queues_empty(sk))
2218 /* the closing socket has some data untransmitted and/or unacked:
2219 * some data in the mptcp rtx queue has not really xmitted yet.
2220 * keep it simple and re-inject the whole mptcp level rtx queue
2222 mptcp_data_lock(sk);
2223 __mptcp_clean_una_wakeup(sk);
2224 rtx_head = mptcp_rtx_head(sk);
2226 mptcp_data_unlock(sk);
2230 msk->recovery_snd_nxt = msk->snd_nxt;
2231 msk->recovery = true;
2232 mptcp_data_unlock(sk);
2234 msk->first_pending = rtx_head;
2235 msk->tx_pending_data += msk->snd_nxt - rtx_head->data_seq;
2238 /* be sure to clear the "sent status" on all re-injected fragments */
2239 list_for_each_entry(cur, &msk->rtx_queue, list) {
2240 if (!cur->already_sent)
2242 cur->already_sent = 0;
2248 /* subflow sockets can be either outgoing (connect) or incoming
2251 * Outgoing subflows use in-kernel sockets.
2252 * Incoming subflows do not have their own 'struct socket' allocated,
2253 * so we need to use tcp_close() after detaching them from the mptcp
2256 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2257 struct mptcp_subflow_context *subflow)
2259 struct mptcp_sock *msk = mptcp_sk(sk);
2262 list_del(&subflow->node);
2264 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2266 /* if we are invoked by the msk cleanup code, the subflow is
2272 need_push = __mptcp_retransmit_pending_data(sk);
2273 subflow->disposable = 1;
2275 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2276 * the ssk has been already destroyed, we just need to release the
2277 * reference owned by msk;
2279 if (!inet_csk(ssk)->icsk_ulp_ops) {
2280 kfree_rcu(subflow, rcu);
2282 /* otherwise tcp will dispose of the ssk and subflow ctx */
2283 __tcp_close(ssk, 0);
2285 /* close acquired an extra ref */
2292 if (ssk == msk->last_snd)
2293 msk->last_snd = NULL;
2295 if (ssk == msk->first)
2298 if (msk->subflow && ssk == msk->subflow->sk)
2299 mptcp_dispose_initial_subflow(msk);
2302 __mptcp_push_pending(sk, 0);
2305 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2306 struct mptcp_subflow_context *subflow)
2308 if (sk->sk_state == TCP_ESTABLISHED)
2309 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2310 __mptcp_close_ssk(sk, ssk, subflow);
2313 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2318 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2320 struct mptcp_subflow_context *subflow, *tmp;
2324 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2325 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2327 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2330 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2331 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2334 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2338 static bool mptcp_check_close_timeout(const struct sock *sk)
2340 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2341 struct mptcp_subflow_context *subflow;
2343 if (delta >= TCP_TIMEWAIT_LEN)
2346 /* if all subflows are in closed status don't bother with additional
2349 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2350 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2357 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2359 struct mptcp_subflow_context *subflow, *tmp;
2360 struct sock *sk = &msk->sk.icsk_inet.sk;
2362 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2365 mptcp_token_destroy(msk);
2367 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2368 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2371 slow = lock_sock_fast(tcp_sk);
2372 if (tcp_sk->sk_state != TCP_CLOSE) {
2373 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2374 tcp_set_state(tcp_sk, TCP_CLOSE);
2376 unlock_sock_fast(tcp_sk, slow);
2379 inet_sk_state_store(sk, TCP_CLOSE);
2380 sk->sk_shutdown = SHUTDOWN_MASK;
2381 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2382 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2384 mptcp_close_wake_up(sk);
2387 static void __mptcp_retrans(struct sock *sk)
2389 struct mptcp_sock *msk = mptcp_sk(sk);
2390 struct mptcp_sendmsg_info info = {};
2391 struct mptcp_data_frag *dfrag;
2396 mptcp_clean_una_wakeup(sk);
2397 dfrag = mptcp_rtx_head(sk);
2399 if (mptcp_data_fin_enabled(msk)) {
2400 struct inet_connection_sock *icsk = inet_csk(sk);
2402 icsk->icsk_retransmits++;
2403 mptcp_set_datafin_timeout(sk);
2404 mptcp_send_ack(msk);
2412 ssk = mptcp_subflow_get_retrans(msk);
2418 /* limit retransmission to the bytes already sent on some subflows */
2420 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2421 while (info.sent < info.limit) {
2422 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2426 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2431 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2432 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2439 if (!mptcp_timer_pending(sk))
2440 mptcp_reset_timer(sk);
2443 static void mptcp_worker(struct work_struct *work)
2445 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2446 struct sock *sk = &msk->sk.icsk_inet.sk;
2450 state = sk->sk_state;
2451 if (unlikely(state == TCP_CLOSE))
2454 mptcp_check_data_fin_ack(sk);
2455 mptcp_flush_join_list(msk);
2457 mptcp_check_fastclose(msk);
2460 mptcp_pm_nl_work(msk);
2462 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2463 mptcp_check_for_eof(msk);
2465 __mptcp_check_send_data_fin(sk);
2466 mptcp_check_data_fin(sk);
2468 /* There is no point in keeping around an orphaned sk timedout or
2469 * closed, but we need the msk around to reply to incoming DATA_FIN,
2470 * even if it is orphaned and in FIN_WAIT2 state
2472 if (sock_flag(sk, SOCK_DEAD) &&
2473 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2474 inet_sk_state_store(sk, TCP_CLOSE);
2475 __mptcp_destroy_sock(sk);
2479 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2480 __mptcp_close_subflow(msk);
2482 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2483 __mptcp_retrans(sk);
2490 static int __mptcp_init_sock(struct sock *sk)
2492 struct mptcp_sock *msk = mptcp_sk(sk);
2494 spin_lock_init(&msk->join_list_lock);
2496 INIT_LIST_HEAD(&msk->conn_list);
2497 INIT_LIST_HEAD(&msk->join_list);
2498 INIT_LIST_HEAD(&msk->rtx_queue);
2499 INIT_WORK(&msk->work, mptcp_worker);
2500 __skb_queue_head_init(&msk->receive_queue);
2501 msk->out_of_order_queue = RB_ROOT;
2502 msk->first_pending = NULL;
2503 msk->wmem_reserved = 0;
2504 WRITE_ONCE(msk->rmem_released, 0);
2505 msk->tx_pending_data = 0;
2506 msk->timer_ival = TCP_RTO_MIN;
2509 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2510 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2511 msk->recovery = false;
2513 mptcp_pm_data_init(msk);
2515 /* re-use the csk retrans timer for MPTCP-level retrans */
2516 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2517 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2522 static int mptcp_init_sock(struct sock *sk)
2524 struct inet_connection_sock *icsk = inet_csk(sk);
2525 struct net *net = sock_net(sk);
2528 ret = __mptcp_init_sock(sk);
2532 if (!mptcp_is_enabled(net))
2533 return -ENOPROTOOPT;
2535 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2538 ret = __mptcp_socket_create(mptcp_sk(sk));
2542 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2543 * propagate the correct value
2545 tcp_assign_congestion_control(sk);
2546 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2548 /* no need to keep a reference to the ops, the name will suffice */
2549 tcp_cleanup_congestion_control(sk);
2550 icsk->icsk_ca_ops = NULL;
2552 sk_sockets_allocated_inc(sk);
2553 sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2554 sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2559 static void __mptcp_clear_xmit(struct sock *sk)
2561 struct mptcp_sock *msk = mptcp_sk(sk);
2562 struct mptcp_data_frag *dtmp, *dfrag;
2564 WRITE_ONCE(msk->first_pending, NULL);
2565 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2566 dfrag_clear(sk, dfrag);
2569 static void mptcp_cancel_work(struct sock *sk)
2571 struct mptcp_sock *msk = mptcp_sk(sk);
2573 if (cancel_work_sync(&msk->work))
2577 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2581 switch (ssk->sk_state) {
2583 if (!(how & RCV_SHUTDOWN))
2587 tcp_disconnect(ssk, O_NONBLOCK);
2590 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2591 pr_debug("Fallback");
2592 ssk->sk_shutdown |= how;
2593 tcp_shutdown(ssk, how);
2595 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2597 if (!mptcp_timer_pending(sk))
2598 mptcp_reset_timer(sk);
2606 static const unsigned char new_state[16] = {
2607 /* current state: new state: action: */
2608 [0 /* (Invalid) */] = TCP_CLOSE,
2609 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2610 [TCP_SYN_SENT] = TCP_CLOSE,
2611 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2612 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2613 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2614 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2615 [TCP_CLOSE] = TCP_CLOSE,
2616 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2617 [TCP_LAST_ACK] = TCP_LAST_ACK,
2618 [TCP_LISTEN] = TCP_CLOSE,
2619 [TCP_CLOSING] = TCP_CLOSING,
2620 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2623 static int mptcp_close_state(struct sock *sk)
2625 int next = (int)new_state[sk->sk_state];
2626 int ns = next & TCP_STATE_MASK;
2628 inet_sk_state_store(sk, ns);
2630 return next & TCP_ACTION_FIN;
2633 static void __mptcp_check_send_data_fin(struct sock *sk)
2635 struct mptcp_subflow_context *subflow;
2636 struct mptcp_sock *msk = mptcp_sk(sk);
2638 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2639 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2640 msk->snd_nxt, msk->write_seq);
2642 /* we still need to enqueue subflows or not really shutting down,
2645 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2646 mptcp_send_head(sk))
2649 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2651 /* fallback socket will not get data_fin/ack, can move to the next
2654 if (__mptcp_check_fallback(msk)) {
2655 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2656 inet_sk_state_store(sk, TCP_CLOSE);
2657 mptcp_close_wake_up(sk);
2658 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2659 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2663 mptcp_flush_join_list(msk);
2664 mptcp_for_each_subflow(msk, subflow) {
2665 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2667 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2671 static void __mptcp_wr_shutdown(struct sock *sk)
2673 struct mptcp_sock *msk = mptcp_sk(sk);
2675 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2676 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2677 !!mptcp_send_head(sk));
2679 /* will be ignored by fallback sockets */
2680 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2681 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2683 __mptcp_check_send_data_fin(sk);
2686 static void __mptcp_destroy_sock(struct sock *sk)
2688 struct mptcp_subflow_context *subflow, *tmp;
2689 struct mptcp_sock *msk = mptcp_sk(sk);
2690 LIST_HEAD(conn_list);
2692 pr_debug("msk=%p", msk);
2696 /* be sure to always acquire the join list lock, to sync vs
2697 * mptcp_finish_join().
2699 spin_lock_bh(&msk->join_list_lock);
2700 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2701 spin_unlock_bh(&msk->join_list_lock);
2702 list_splice_init(&msk->conn_list, &conn_list);
2704 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2705 sk_stop_timer(sk, &sk->sk_timer);
2708 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2709 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2710 __mptcp_close_ssk(sk, ssk, subflow);
2713 sk->sk_prot->destroy(sk);
2715 WARN_ON_ONCE(msk->wmem_reserved);
2716 WARN_ON_ONCE(msk->rmem_released);
2717 sk_stream_kill_queues(sk);
2718 xfrm_sk_free_policy(sk);
2720 sk_refcnt_debug_release(sk);
2721 mptcp_dispose_initial_subflow(msk);
2725 static void mptcp_close(struct sock *sk, long timeout)
2727 struct mptcp_subflow_context *subflow;
2728 bool do_cancel_work = false;
2731 sk->sk_shutdown = SHUTDOWN_MASK;
2733 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2734 inet_sk_state_store(sk, TCP_CLOSE);
2738 if (mptcp_close_state(sk))
2739 __mptcp_wr_shutdown(sk);
2741 sk_stream_wait_close(sk, timeout);
2744 /* orphan all the subflows */
2745 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2746 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2747 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2748 bool slow = lock_sock_fast_nested(ssk);
2751 unlock_sock_fast(ssk, slow);
2756 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2757 if (sk->sk_state == TCP_CLOSE) {
2758 __mptcp_destroy_sock(sk);
2759 do_cancel_work = true;
2761 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2765 mptcp_cancel_work(sk);
2767 if (mptcp_sk(sk)->token)
2768 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2773 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2775 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2776 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2777 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2779 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2780 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2783 msk6->saddr = ssk6->saddr;
2784 msk6->flow_label = ssk6->flow_label;
2788 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2789 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2790 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2791 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2792 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2793 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2796 static int mptcp_disconnect(struct sock *sk, int flags)
2798 struct mptcp_subflow_context *subflow;
2799 struct mptcp_sock *msk = mptcp_sk(sk);
2801 mptcp_do_flush_join_list(msk);
2803 mptcp_for_each_subflow(msk, subflow) {
2804 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2807 tcp_disconnect(ssk, flags);
2813 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2814 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2816 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2818 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2822 struct sock *mptcp_sk_clone(const struct sock *sk,
2823 const struct mptcp_options_received *mp_opt,
2824 struct request_sock *req)
2826 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2827 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2828 struct mptcp_sock *msk;
2834 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2835 if (nsk->sk_family == AF_INET6)
2836 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2839 __mptcp_init_sock(nsk);
2841 msk = mptcp_sk(nsk);
2842 msk->local_key = subflow_req->local_key;
2843 msk->token = subflow_req->token;
2844 msk->subflow = NULL;
2845 WRITE_ONCE(msk->fully_established, false);
2846 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
2847 WRITE_ONCE(msk->csum_enabled, true);
2849 msk->write_seq = subflow_req->idsn + 1;
2850 msk->snd_nxt = msk->write_seq;
2851 msk->snd_una = msk->write_seq;
2852 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2853 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2855 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
2856 msk->can_ack = true;
2857 msk->remote_key = mp_opt->sndr_key;
2858 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2860 WRITE_ONCE(msk->ack_seq, ack_seq);
2861 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2864 sock_reset_flag(nsk, SOCK_RCU_FREE);
2865 /* will be fully established after successful MPC subflow creation */
2866 inet_sk_state_store(nsk, TCP_SYN_RECV);
2868 security_inet_csk_clone(nsk, req);
2869 bh_unlock_sock(nsk);
2871 /* keep a single reference */
2876 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2878 const struct tcp_sock *tp = tcp_sk(ssk);
2880 msk->rcvq_space.copied = 0;
2881 msk->rcvq_space.rtt_us = 0;
2883 msk->rcvq_space.time = tp->tcp_mstamp;
2885 /* initial rcv_space offering made to peer */
2886 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2887 TCP_INIT_CWND * tp->advmss);
2888 if (msk->rcvq_space.space == 0)
2889 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2891 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2894 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2897 struct mptcp_sock *msk = mptcp_sk(sk);
2898 struct socket *listener;
2901 listener = __mptcp_nmpc_socket(msk);
2902 if (WARN_ON_ONCE(!listener)) {
2907 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2908 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2912 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2913 if (sk_is_mptcp(newsk)) {
2914 struct mptcp_subflow_context *subflow;
2915 struct sock *new_mptcp_sock;
2917 subflow = mptcp_subflow_ctx(newsk);
2918 new_mptcp_sock = subflow->conn;
2920 /* is_mptcp should be false if subflow->conn is missing, see
2921 * subflow_syn_recv_sock()
2923 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2924 tcp_sk(newsk)->is_mptcp = 0;
2928 /* acquire the 2nd reference for the owning socket */
2929 sock_hold(new_mptcp_sock);
2930 newsk = new_mptcp_sock;
2931 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2933 MPTCP_INC_STATS(sock_net(sk),
2934 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2938 newsk->sk_kern_sock = kern;
2942 void mptcp_destroy_common(struct mptcp_sock *msk)
2944 struct sock *sk = (struct sock *)msk;
2946 __mptcp_clear_xmit(sk);
2948 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2949 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2951 skb_rbtree_purge(&msk->out_of_order_queue);
2952 mptcp_token_destroy(msk);
2953 mptcp_pm_free_anno_list(msk);
2956 static void mptcp_destroy(struct sock *sk)
2958 struct mptcp_sock *msk = mptcp_sk(sk);
2960 mptcp_destroy_common(msk);
2961 sk_sockets_allocated_dec(sk);
2964 void __mptcp_data_acked(struct sock *sk)
2966 if (!sock_owned_by_user(sk))
2967 __mptcp_clean_una(sk);
2969 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2971 if (mptcp_pending_data_fin_ack(sk))
2972 mptcp_schedule_work(sk);
2975 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2977 if (!mptcp_send_head(sk))
2980 if (!sock_owned_by_user(sk)) {
2981 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2983 if (xmit_ssk == ssk)
2984 __mptcp_subflow_push_pending(sk, ssk);
2986 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
2988 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2992 /* processes deferred events and flush wmem */
2993 static void mptcp_release_cb(struct sock *sk)
2996 unsigned long flags = 0;
2998 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2999 flags |= BIT(MPTCP_PUSH_PENDING);
3000 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
3001 flags |= BIT(MPTCP_RETRANSMIT);
3005 /* the following actions acquire the subflow socket lock
3007 * 1) can't be invoked in atomic scope
3008 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3009 * datapath acquires the msk socket spinlock while helding
3010 * the subflow socket lock
3013 spin_unlock_bh(&sk->sk_lock.slock);
3014 if (flags & BIT(MPTCP_PUSH_PENDING))
3015 __mptcp_push_pending(sk, 0);
3016 if (flags & BIT(MPTCP_RETRANSMIT))
3017 __mptcp_retrans(sk);
3020 spin_lock_bh(&sk->sk_lock.slock);
3023 /* be sure to set the current sk state before tacking actions
3024 * depending on sk_state
3026 if (test_and_clear_bit(MPTCP_CONNECTED, &mptcp_sk(sk)->flags))
3027 __mptcp_set_connected(sk);
3028 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
3029 __mptcp_clean_una_wakeup(sk);
3030 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
3031 __mptcp_error_report(sk);
3033 /* push_pending may touch wmem_reserved, ensure we do the cleanup
3036 __mptcp_update_wmem(sk);
3037 __mptcp_update_rmem(sk);
3040 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3041 * TCP can't schedule delack timer before the subflow is fully established.
3042 * MPTCP uses the delack timer to do 3rd ack retransmissions
3044 static void schedule_3rdack_retransmission(struct sock *ssk)
3046 struct inet_connection_sock *icsk = inet_csk(ssk);
3047 struct tcp_sock *tp = tcp_sk(ssk);
3048 unsigned long timeout;
3050 if (mptcp_subflow_ctx(ssk)->fully_established)
3053 /* reschedule with a timeout above RTT, as we must look only for drop */
3055 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3057 timeout = TCP_TIMEOUT_INIT;
3060 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3061 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3062 icsk->icsk_ack.timeout = timeout;
3063 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3066 void mptcp_subflow_process_delegated(struct sock *ssk)
3068 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3069 struct sock *sk = subflow->conn;
3071 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3072 mptcp_data_lock(sk);
3073 if (!sock_owned_by_user(sk))
3074 __mptcp_subflow_push_pending(sk, ssk);
3076 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
3077 mptcp_data_unlock(sk);
3078 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3080 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3081 schedule_3rdack_retransmission(ssk);
3082 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3086 static int mptcp_hash(struct sock *sk)
3088 /* should never be called,
3089 * we hash the TCP subflows not the master socket
3095 static void mptcp_unhash(struct sock *sk)
3097 /* called from sk_common_release(), but nothing to do here */
3100 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3102 struct mptcp_sock *msk = mptcp_sk(sk);
3103 struct socket *ssock;
3105 ssock = __mptcp_nmpc_socket(msk);
3106 pr_debug("msk=%p, subflow=%p", msk, ssock);
3107 if (WARN_ON_ONCE(!ssock))
3110 return inet_csk_get_port(ssock->sk, snum);
3113 void mptcp_finish_connect(struct sock *ssk)
3115 struct mptcp_subflow_context *subflow;
3116 struct mptcp_sock *msk;
3120 subflow = mptcp_subflow_ctx(ssk);
3124 pr_debug("msk=%p, token=%u", sk, subflow->token);
3126 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3128 subflow->map_seq = ack_seq;
3129 subflow->map_subflow_seq = 1;
3131 /* the socket is not connected yet, no msk/subflow ops can access/race
3132 * accessing the field below
3134 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3135 WRITE_ONCE(msk->local_key, subflow->local_key);
3136 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3137 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3138 WRITE_ONCE(msk->ack_seq, ack_seq);
3139 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3140 WRITE_ONCE(msk->can_ack, 1);
3141 WRITE_ONCE(msk->snd_una, msk->write_seq);
3143 mptcp_pm_new_connection(msk, ssk, 0);
3145 mptcp_rcv_space_init(msk, ssk);
3148 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3150 write_lock_bh(&sk->sk_callback_lock);
3151 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3152 sk_set_socket(sk, parent);
3153 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3154 write_unlock_bh(&sk->sk_callback_lock);
3157 bool mptcp_finish_join(struct sock *ssk)
3159 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3160 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3161 struct sock *parent = (void *)msk;
3162 struct socket *parent_sock;
3165 pr_debug("msk=%p, subflow=%p", msk, subflow);
3167 /* mptcp socket already closing? */
3168 if (!mptcp_is_fully_established(parent)) {
3169 subflow->reset_reason = MPTCP_RST_EMPTCP;
3173 if (!msk->pm.server_side)
3176 if (!mptcp_pm_allow_new_subflow(msk)) {
3177 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3181 /* active connections are already on conn_list, and we can't acquire
3183 * use the join list lock as synchronization point and double-check
3184 * msk status to avoid racing with __mptcp_destroy_sock()
3186 spin_lock_bh(&msk->join_list_lock);
3187 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3188 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3189 list_add_tail(&subflow->node, &msk->join_list);
3192 spin_unlock_bh(&msk->join_list_lock);
3194 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3198 /* attach to msk socket only after we are sure he will deal with us
3201 parent_sock = READ_ONCE(parent->sk_socket);
3202 if (parent_sock && !ssk->sk_socket)
3203 mptcp_sock_graft(ssk, parent_sock);
3204 subflow->map_seq = READ_ONCE(msk->ack_seq);
3206 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3210 static void mptcp_shutdown(struct sock *sk, int how)
3212 pr_debug("sk=%p, how=%d", sk, how);
3214 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3215 __mptcp_wr_shutdown(sk);
3218 static struct proto mptcp_prot = {
3220 .owner = THIS_MODULE,
3221 .init = mptcp_init_sock,
3222 .disconnect = mptcp_disconnect,
3223 .close = mptcp_close,
3224 .accept = mptcp_accept,
3225 .setsockopt = mptcp_setsockopt,
3226 .getsockopt = mptcp_getsockopt,
3227 .shutdown = mptcp_shutdown,
3228 .destroy = mptcp_destroy,
3229 .sendmsg = mptcp_sendmsg,
3230 .recvmsg = mptcp_recvmsg,
3231 .release_cb = mptcp_release_cb,
3233 .unhash = mptcp_unhash,
3234 .get_port = mptcp_get_port,
3235 .sockets_allocated = &mptcp_sockets_allocated,
3236 .memory_allocated = &tcp_memory_allocated,
3237 .memory_pressure = &tcp_memory_pressure,
3238 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3239 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3240 .sysctl_mem = sysctl_tcp_mem,
3241 .obj_size = sizeof(struct mptcp_sock),
3242 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3243 .no_autobind = true,
3246 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3248 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3249 struct socket *ssock;
3252 lock_sock(sock->sk);
3253 ssock = __mptcp_nmpc_socket(msk);
3259 err = ssock->ops->bind(ssock, uaddr, addr_len);
3261 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3264 release_sock(sock->sk);
3268 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3269 struct mptcp_subflow_context *subflow)
3271 subflow->request_mptcp = 0;
3272 __mptcp_do_fallback(msk);
3275 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3276 int addr_len, int flags)
3278 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3279 struct mptcp_subflow_context *subflow;
3280 struct socket *ssock;
3283 lock_sock(sock->sk);
3284 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3285 /* pending connection or invalid state, let existing subflow
3288 ssock = msk->subflow;
3292 ssock = __mptcp_nmpc_socket(msk);
3298 mptcp_token_destroy(msk);
3299 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3300 subflow = mptcp_subflow_ctx(ssock->sk);
3301 #ifdef CONFIG_TCP_MD5SIG
3302 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3305 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3306 mptcp_subflow_early_fallback(msk, subflow);
3308 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3309 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3310 mptcp_subflow_early_fallback(msk, subflow);
3312 if (likely(!__mptcp_check_fallback(msk)))
3313 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3316 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3317 sock->state = ssock->state;
3319 /* on successful connect, the msk state will be moved to established by
3320 * subflow_finish_connect()
3322 if (!err || err == -EINPROGRESS)
3323 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3325 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3328 release_sock(sock->sk);
3332 static int mptcp_listen(struct socket *sock, int backlog)
3334 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3335 struct socket *ssock;
3338 pr_debug("msk=%p", msk);
3340 lock_sock(sock->sk);
3341 ssock = __mptcp_nmpc_socket(msk);
3347 mptcp_token_destroy(msk);
3348 inet_sk_state_store(sock->sk, TCP_LISTEN);
3349 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3351 err = ssock->ops->listen(ssock, backlog);
3352 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3354 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3357 release_sock(sock->sk);
3361 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3362 int flags, bool kern)
3364 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3365 struct socket *ssock;
3368 pr_debug("msk=%p", msk);
3370 lock_sock(sock->sk);
3371 if (sock->sk->sk_state != TCP_LISTEN)
3374 ssock = __mptcp_nmpc_socket(msk);
3378 clear_bit(MPTCP_DATA_READY, &msk->flags);
3379 sock_hold(ssock->sk);
3380 release_sock(sock->sk);
3382 err = ssock->ops->accept(sock, newsock, flags, kern);
3383 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3384 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3385 struct mptcp_subflow_context *subflow;
3386 struct sock *newsk = newsock->sk;
3390 /* PM/worker can now acquire the first subflow socket
3391 * lock without racing with listener queue cleanup,
3392 * we can notify it, if needed.
3394 * Even if remote has reset the initial subflow by now
3395 * the refcnt is still at least one.
3397 subflow = mptcp_subflow_ctx(msk->first);
3398 list_add(&subflow->node, &msk->conn_list);
3399 sock_hold(msk->first);
3400 if (mptcp_is_fully_established(newsk))
3401 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3403 mptcp_copy_inaddrs(newsk, msk->first);
3404 mptcp_rcv_space_init(msk, msk->first);
3405 mptcp_propagate_sndbuf(newsk, msk->first);
3407 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3408 * This is needed so NOSPACE flag can be set from tcp stack.
3410 mptcp_flush_join_list(msk);
3411 mptcp_for_each_subflow(msk, subflow) {
3412 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3414 if (!ssk->sk_socket)
3415 mptcp_sock_graft(ssk, newsock);
3417 release_sock(newsk);
3420 if (inet_csk_listen_poll(ssock->sk))
3421 set_bit(MPTCP_DATA_READY, &msk->flags);
3422 sock_put(ssock->sk);
3426 release_sock(sock->sk);
3430 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3432 /* Concurrent splices from sk_receive_queue into receive_queue will
3433 * always show at least one non-empty queue when checked in this order.
3435 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
3436 skb_queue_empty_lockless(&msk->receive_queue))
3439 return EPOLLIN | EPOLLRDNORM;
3442 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3444 struct sock *sk = (struct sock *)msk;
3446 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3447 return EPOLLOUT | EPOLLWRNORM;
3449 if (sk_stream_is_writeable(sk))
3450 return EPOLLOUT | EPOLLWRNORM;
3452 mptcp_set_nospace(sk);
3453 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3454 if (sk_stream_is_writeable(sk))
3455 return EPOLLOUT | EPOLLWRNORM;
3460 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3461 struct poll_table_struct *wait)
3463 struct sock *sk = sock->sk;
3464 struct mptcp_sock *msk;
3469 sock_poll_wait(file, sock, wait);
3471 state = inet_sk_state_load(sk);
3472 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3473 if (state == TCP_LISTEN)
3474 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM : 0;
3476 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3477 mask |= mptcp_check_readable(msk);
3478 mask |= mptcp_check_writeable(msk);
3480 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3482 if (sk->sk_shutdown & RCV_SHUTDOWN)
3483 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3485 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3493 static const struct proto_ops mptcp_stream_ops = {
3495 .owner = THIS_MODULE,
3496 .release = inet_release,
3498 .connect = mptcp_stream_connect,
3499 .socketpair = sock_no_socketpair,
3500 .accept = mptcp_stream_accept,
3501 .getname = inet_getname,
3503 .ioctl = inet_ioctl,
3504 .gettstamp = sock_gettstamp,
3505 .listen = mptcp_listen,
3506 .shutdown = inet_shutdown,
3507 .setsockopt = sock_common_setsockopt,
3508 .getsockopt = sock_common_getsockopt,
3509 .sendmsg = inet_sendmsg,
3510 .recvmsg = inet_recvmsg,
3511 .mmap = sock_no_mmap,
3512 .sendpage = inet_sendpage,
3515 static struct inet_protosw mptcp_protosw = {
3516 .type = SOCK_STREAM,
3517 .protocol = IPPROTO_MPTCP,
3518 .prot = &mptcp_prot,
3519 .ops = &mptcp_stream_ops,
3520 .flags = INET_PROTOSW_ICSK,
3523 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3525 struct mptcp_delegated_action *delegated;
3526 struct mptcp_subflow_context *subflow;
3529 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3530 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3531 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3533 bh_lock_sock_nested(ssk);
3534 if (!sock_owned_by_user(ssk) &&
3535 mptcp_subflow_has_delegated_action(subflow))
3536 mptcp_subflow_process_delegated(ssk);
3537 /* ... elsewhere tcp_release_cb_override already processed
3538 * the action or will do at next release_sock().
3539 * In both case must dequeue the subflow here - on the same
3540 * CPU that scheduled it.
3542 bh_unlock_sock(ssk);
3545 if (++work_done == budget)
3549 /* always provide a 0 'work_done' argument, so that napi_complete_done
3550 * will not try accessing the NULL napi->dev ptr
3552 napi_complete_done(napi, 0);
3556 void __init mptcp_proto_init(void)
3558 struct mptcp_delegated_action *delegated;
3561 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3563 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3564 panic("Failed to allocate MPTCP pcpu counter\n");
3566 init_dummy_netdev(&mptcp_napi_dev);
3567 for_each_possible_cpu(cpu) {
3568 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3569 INIT_LIST_HEAD(&delegated->head);
3570 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3572 napi_enable(&delegated->napi);
3575 mptcp_subflow_init();
3579 if (proto_register(&mptcp_prot, 1) != 0)
3580 panic("Failed to register MPTCP proto.\n");
3582 inet_register_protosw(&mptcp_protosw);
3584 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3587 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3588 static const struct proto_ops mptcp_v6_stream_ops = {
3590 .owner = THIS_MODULE,
3591 .release = inet6_release,
3593 .connect = mptcp_stream_connect,
3594 .socketpair = sock_no_socketpair,
3595 .accept = mptcp_stream_accept,
3596 .getname = inet6_getname,
3598 .ioctl = inet6_ioctl,
3599 .gettstamp = sock_gettstamp,
3600 .listen = mptcp_listen,
3601 .shutdown = inet_shutdown,
3602 .setsockopt = sock_common_setsockopt,
3603 .getsockopt = sock_common_getsockopt,
3604 .sendmsg = inet6_sendmsg,
3605 .recvmsg = inet6_recvmsg,
3606 .mmap = sock_no_mmap,
3607 .sendpage = inet_sendpage,
3608 #ifdef CONFIG_COMPAT
3609 .compat_ioctl = inet6_compat_ioctl,
3613 static struct proto mptcp_v6_prot;
3615 static void mptcp_v6_destroy(struct sock *sk)
3618 inet6_destroy_sock(sk);
3621 static struct inet_protosw mptcp_v6_protosw = {
3622 .type = SOCK_STREAM,
3623 .protocol = IPPROTO_MPTCP,
3624 .prot = &mptcp_v6_prot,
3625 .ops = &mptcp_v6_stream_ops,
3626 .flags = INET_PROTOSW_ICSK,
3629 int __init mptcp_proto_v6_init(void)
3633 mptcp_v6_prot = mptcp_prot;
3634 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3635 mptcp_v6_prot.slab = NULL;
3636 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3637 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3639 err = proto_register(&mptcp_v6_prot, 1);
3643 err = inet6_register_protosw(&mptcp_v6_protosw);
3645 proto_unregister(&mptcp_v6_prot);