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>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
46 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
49 MPTCP_CMSG_TS = BIT(0),
50 MPTCP_CMSG_INQ = BIT(1),
53 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
55 static void __mptcp_destroy_sock(struct sock *sk);
56 static void __mptcp_check_send_data_fin(struct sock *sk);
58 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
59 static struct net_device mptcp_napi_dev;
61 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
62 * completed yet or has failed, return the subflow socket.
63 * Otherwise return NULL.
65 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
67 if (!msk->subflow || READ_ONCE(msk->can_ack))
73 /* Returns end sequence number of the receiver's advertised window */
74 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
76 return READ_ONCE(msk->wnd_end);
79 static bool mptcp_is_tcpsk(struct sock *sk)
81 struct socket *sock = sk->sk_socket;
83 if (unlikely(sk->sk_prot == &tcp_prot)) {
84 /* we are being invoked after mptcp_accept() has
85 * accepted a non-mp-capable flow: sk is a tcp_sk,
88 * Hand the socket over to tcp so all further socket ops
91 sock->ops = &inet_stream_ops;
93 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
94 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
95 sock->ops = &inet6_stream_ops;
103 static int __mptcp_socket_create(struct mptcp_sock *msk)
105 struct mptcp_subflow_context *subflow;
106 struct sock *sk = (struct sock *)msk;
107 struct socket *ssock;
110 err = mptcp_subflow_create_socket(sk, &ssock);
114 msk->first = ssock->sk;
115 msk->subflow = ssock;
116 subflow = mptcp_subflow_ctx(ssock->sk);
117 list_add(&subflow->node, &msk->conn_list);
118 sock_hold(ssock->sk);
119 subflow->request_mptcp = 1;
121 /* This is the first subflow, always with id 0 */
122 subflow->local_id_valid = 1;
123 mptcp_sock_graft(msk->first, sk->sk_socket);
128 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
130 sk_drops_add(sk, skb);
134 static void mptcp_rmem_charge(struct sock *sk, int size)
136 mptcp_sk(sk)->rmem_fwd_alloc -= size;
139 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
140 struct sk_buff *from)
145 if (MPTCP_SKB_CB(from)->offset ||
146 !skb_try_coalesce(to, from, &fragstolen, &delta))
149 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
150 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
151 to->len, MPTCP_SKB_CB(from)->end_seq);
152 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
153 kfree_skb_partial(from, fragstolen);
154 atomic_add(delta, &sk->sk_rmem_alloc);
155 mptcp_rmem_charge(sk, delta);
159 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
160 struct sk_buff *from)
162 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
165 return mptcp_try_coalesce((struct sock *)msk, to, from);
168 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
170 amount >>= PAGE_SHIFT;
171 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
172 __sk_mem_reduce_allocated(sk, amount);
175 static void mptcp_rmem_uncharge(struct sock *sk, int size)
177 struct mptcp_sock *msk = mptcp_sk(sk);
180 msk->rmem_fwd_alloc += size;
181 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
183 /* see sk_mem_uncharge() for the rationale behind the following schema */
184 if (unlikely(reclaimable >= PAGE_SIZE))
185 __mptcp_rmem_reclaim(sk, reclaimable);
188 static void mptcp_rfree(struct sk_buff *skb)
190 unsigned int len = skb->truesize;
191 struct sock *sk = skb->sk;
193 atomic_sub(len, &sk->sk_rmem_alloc);
194 mptcp_rmem_uncharge(sk, len);
197 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
201 skb->destructor = mptcp_rfree;
202 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
203 mptcp_rmem_charge(sk, skb->truesize);
206 /* "inspired" by tcp_data_queue_ofo(), main differences:
208 * - don't cope with sacks
210 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
212 struct sock *sk = (struct sock *)msk;
213 struct rb_node **p, *parent;
214 u64 seq, end_seq, max_seq;
215 struct sk_buff *skb1;
217 seq = MPTCP_SKB_CB(skb)->map_seq;
218 end_seq = MPTCP_SKB_CB(skb)->end_seq;
219 max_seq = atomic64_read(&msk->rcv_wnd_sent);
221 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
222 RB_EMPTY_ROOT(&msk->out_of_order_queue));
223 if (after64(end_seq, max_seq)) {
226 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
227 (unsigned long long)end_seq - (unsigned long)max_seq,
228 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
229 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
233 p = &msk->out_of_order_queue.rb_node;
234 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
235 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
236 rb_link_node(&skb->rbnode, NULL, p);
237 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
238 msk->ooo_last_skb = skb;
242 /* with 2 subflows, adding at end of ooo queue is quite likely
243 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
245 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
251 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
252 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
253 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
254 parent = &msk->ooo_last_skb->rbnode;
255 p = &parent->rb_right;
259 /* Find place to insert this segment. Handle overlaps on the way. */
263 skb1 = rb_to_skb(parent);
264 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
265 p = &parent->rb_left;
268 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
269 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
270 /* All the bits are present. Drop. */
272 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
275 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
279 * continue traversing
282 /* skb's seq == skb1's seq and skb covers skb1.
283 * Replace skb1 with skb.
285 rb_replace_node(&skb1->rbnode, &skb->rbnode,
286 &msk->out_of_order_queue);
287 mptcp_drop(sk, skb1);
288 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
291 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
292 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
295 p = &parent->rb_right;
299 /* Insert segment into RB tree. */
300 rb_link_node(&skb->rbnode, parent, p);
301 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
304 /* Remove other segments covered by skb. */
305 while ((skb1 = skb_rb_next(skb)) != NULL) {
306 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
308 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
309 mptcp_drop(sk, skb1);
310 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
312 /* If there is no skb after us, we are the last_skb ! */
314 msk->ooo_last_skb = skb;
318 mptcp_set_owner_r(skb, sk);
321 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
323 struct mptcp_sock *msk = mptcp_sk(sk);
326 if (size <= msk->rmem_fwd_alloc)
329 size -= msk->rmem_fwd_alloc;
330 amt = sk_mem_pages(size);
331 amount = amt << PAGE_SHIFT;
332 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
335 msk->rmem_fwd_alloc += amount;
339 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
340 struct sk_buff *skb, unsigned int offset,
343 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
344 struct sock *sk = (struct sock *)msk;
345 struct sk_buff *tail;
348 __skb_unlink(skb, &ssk->sk_receive_queue);
353 /* try to fetch required memory from subflow */
354 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
357 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
359 /* the skb map_seq accounts for the skb offset:
360 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
363 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
364 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
365 MPTCP_SKB_CB(skb)->offset = offset;
366 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
368 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
370 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
371 tail = skb_peek_tail(&sk->sk_receive_queue);
372 if (tail && mptcp_try_coalesce(sk, tail, skb))
375 mptcp_set_owner_r(skb, sk);
376 __skb_queue_tail(&sk->sk_receive_queue, skb);
378 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
379 mptcp_data_queue_ofo(msk, skb);
383 /* old data, keep it simple and drop the whole pkt, sender
384 * will retransmit as needed, if needed.
386 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
392 static void mptcp_stop_timer(struct sock *sk)
394 struct inet_connection_sock *icsk = inet_csk(sk);
396 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
397 mptcp_sk(sk)->timer_ival = 0;
400 static void mptcp_close_wake_up(struct sock *sk)
402 if (sock_flag(sk, SOCK_DEAD))
405 sk->sk_state_change(sk);
406 if (sk->sk_shutdown == SHUTDOWN_MASK ||
407 sk->sk_state == TCP_CLOSE)
408 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
410 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
413 static bool mptcp_pending_data_fin_ack(struct sock *sk)
415 struct mptcp_sock *msk = mptcp_sk(sk);
417 return !__mptcp_check_fallback(msk) &&
418 ((1 << sk->sk_state) &
419 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
420 msk->write_seq == READ_ONCE(msk->snd_una);
423 static void mptcp_check_data_fin_ack(struct sock *sk)
425 struct mptcp_sock *msk = mptcp_sk(sk);
427 /* Look for an acknowledged DATA_FIN */
428 if (mptcp_pending_data_fin_ack(sk)) {
429 WRITE_ONCE(msk->snd_data_fin_enable, 0);
431 switch (sk->sk_state) {
433 inet_sk_state_store(sk, TCP_FIN_WAIT2);
437 inet_sk_state_store(sk, TCP_CLOSE);
441 mptcp_close_wake_up(sk);
445 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
447 struct mptcp_sock *msk = mptcp_sk(sk);
449 if (READ_ONCE(msk->rcv_data_fin) &&
450 ((1 << sk->sk_state) &
451 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
452 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
454 if (msk->ack_seq == rcv_data_fin_seq) {
456 *seq = rcv_data_fin_seq;
465 static void mptcp_set_datafin_timeout(const struct sock *sk)
467 struct inet_connection_sock *icsk = inet_csk(sk);
470 retransmits = min_t(u32, icsk->icsk_retransmits,
471 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
473 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
476 static void __mptcp_set_timeout(struct sock *sk, long tout)
478 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
481 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
483 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
485 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
486 inet_csk(ssk)->icsk_timeout - jiffies : 0;
489 static void mptcp_set_timeout(struct sock *sk)
491 struct mptcp_subflow_context *subflow;
494 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
495 tout = max(tout, mptcp_timeout_from_subflow(subflow));
496 __mptcp_set_timeout(sk, tout);
499 static inline bool tcp_can_send_ack(const struct sock *ssk)
501 return !((1 << inet_sk_state_load(ssk)) &
502 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
505 void __mptcp_subflow_send_ack(struct sock *ssk)
507 if (tcp_can_send_ack(ssk))
511 static void mptcp_subflow_send_ack(struct sock *ssk)
515 slow = lock_sock_fast(ssk);
516 __mptcp_subflow_send_ack(ssk);
517 unlock_sock_fast(ssk, slow);
520 static void mptcp_send_ack(struct mptcp_sock *msk)
522 struct mptcp_subflow_context *subflow;
524 mptcp_for_each_subflow(msk, subflow)
525 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
528 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
532 slow = lock_sock_fast(ssk);
533 if (tcp_can_send_ack(ssk))
534 tcp_cleanup_rbuf(ssk, 1);
535 unlock_sock_fast(ssk, slow);
538 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
540 const struct inet_connection_sock *icsk = inet_csk(ssk);
541 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
542 const struct tcp_sock *tp = tcp_sk(ssk);
544 return (ack_pending & ICSK_ACK_SCHED) &&
545 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
546 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
547 (rx_empty && ack_pending &
548 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
551 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
553 int old_space = READ_ONCE(msk->old_wspace);
554 struct mptcp_subflow_context *subflow;
555 struct sock *sk = (struct sock *)msk;
556 int space = __mptcp_space(sk);
557 bool cleanup, rx_empty;
559 cleanup = (space > 0) && (space >= (old_space << 1));
560 rx_empty = !__mptcp_rmem(sk);
562 mptcp_for_each_subflow(msk, subflow) {
563 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
565 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
566 mptcp_subflow_cleanup_rbuf(ssk);
570 static bool mptcp_check_data_fin(struct sock *sk)
572 struct mptcp_sock *msk = mptcp_sk(sk);
573 u64 rcv_data_fin_seq;
576 if (__mptcp_check_fallback(msk))
579 /* Need to ack a DATA_FIN received from a peer while this side
580 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
581 * msk->rcv_data_fin was set when parsing the incoming options
582 * at the subflow level and the msk lock was not held, so this
583 * is the first opportunity to act on the DATA_FIN and change
586 * If we are caught up to the sequence number of the incoming
587 * DATA_FIN, send the DATA_ACK now and do state transition. If
588 * not caught up, do nothing and let the recv code send DATA_ACK
592 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
593 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
594 WRITE_ONCE(msk->rcv_data_fin, 0);
596 sk->sk_shutdown |= RCV_SHUTDOWN;
597 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
599 switch (sk->sk_state) {
600 case TCP_ESTABLISHED:
601 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
604 inet_sk_state_store(sk, TCP_CLOSING);
607 inet_sk_state_store(sk, TCP_CLOSE);
610 /* Other states not expected */
617 mptcp_close_wake_up(sk);
622 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
626 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
627 struct sock *sk = (struct sock *)msk;
628 unsigned int moved = 0;
629 bool more_data_avail;
634 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
636 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
637 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
639 if (unlikely(ssk_rbuf > sk_rbuf)) {
640 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
645 pr_debug("msk=%p ssk=%p", msk, ssk);
648 u32 map_remaining, offset;
649 u32 seq = tp->copied_seq;
653 /* try to move as much data as available */
654 map_remaining = subflow->map_data_len -
655 mptcp_subflow_get_map_offset(subflow);
657 skb = skb_peek(&ssk->sk_receive_queue);
659 /* if no data is found, a racing workqueue/recvmsg
660 * already processed the new data, stop here or we
661 * can enter an infinite loop
668 if (__mptcp_check_fallback(msk)) {
669 /* if we are running under the workqueue, TCP could have
670 * collapsed skbs between dummy map creation and now
671 * be sure to adjust the size
673 map_remaining = skb->len;
674 subflow->map_data_len = skb->len;
677 offset = seq - TCP_SKB_CB(skb)->seq;
678 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
684 if (offset < skb->len) {
685 size_t len = skb->len - offset;
690 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
694 if (WARN_ON_ONCE(map_remaining < len))
698 sk_eat_skb(ssk, skb);
702 WRITE_ONCE(tp->copied_seq, seq);
703 more_data_avail = mptcp_subflow_data_available(ssk);
705 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
709 } while (more_data_avail);
715 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
717 struct sock *sk = (struct sock *)msk;
718 struct sk_buff *skb, *tail;
723 p = rb_first(&msk->out_of_order_queue);
724 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
727 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
731 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
733 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
736 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
740 end_seq = MPTCP_SKB_CB(skb)->end_seq;
741 tail = skb_peek_tail(&sk->sk_receive_queue);
742 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
743 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
745 /* skip overlapping data, if any */
746 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
747 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
749 MPTCP_SKB_CB(skb)->offset += delta;
750 MPTCP_SKB_CB(skb)->map_seq += delta;
751 __skb_queue_tail(&sk->sk_receive_queue, skb);
753 msk->ack_seq = end_seq;
759 /* In most cases we will be able to lock the mptcp socket. If its already
760 * owned, we need to defer to the work queue to avoid ABBA deadlock.
762 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
764 struct sock *sk = (struct sock *)msk;
765 unsigned int moved = 0;
767 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
768 __mptcp_ofo_queue(msk);
769 if (unlikely(ssk->sk_err)) {
770 if (!sock_owned_by_user(sk))
771 __mptcp_error_report(sk);
773 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
776 /* If the moves have caught up with the DATA_FIN sequence number
777 * it's time to ack the DATA_FIN and change socket state, but
778 * this is not a good place to change state. Let the workqueue
781 if (mptcp_pending_data_fin(sk, NULL))
782 mptcp_schedule_work(sk);
786 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
788 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
789 struct mptcp_sock *msk = mptcp_sk(sk);
790 int sk_rbuf, ssk_rbuf;
792 /* The peer can send data while we are shutting down this
793 * subflow at msk destruction time, but we must avoid enqueuing
794 * more data to the msk receive queue
796 if (unlikely(subflow->disposable))
799 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
800 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
801 if (unlikely(ssk_rbuf > sk_rbuf))
804 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
805 if (__mptcp_rmem(sk) > sk_rbuf) {
806 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
810 /* Wake-up the reader only for in-sequence data */
812 if (move_skbs_to_msk(msk, ssk))
813 sk->sk_data_ready(sk);
815 mptcp_data_unlock(sk);
818 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
820 struct sock *sk = (struct sock *)msk;
822 if (sk->sk_state != TCP_ESTABLISHED)
825 /* attach to msk socket only after we are sure we will deal with it
828 if (sk->sk_socket && !ssk->sk_socket)
829 mptcp_sock_graft(ssk, sk->sk_socket);
831 mptcp_propagate_sndbuf((struct sock *)msk, ssk);
832 mptcp_sockopt_sync_locked(msk, ssk);
836 static void __mptcp_flush_join_list(struct sock *sk)
838 struct mptcp_subflow_context *tmp, *subflow;
839 struct mptcp_sock *msk = mptcp_sk(sk);
841 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
842 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
843 bool slow = lock_sock_fast(ssk);
845 list_move_tail(&subflow->node, &msk->conn_list);
846 if (!__mptcp_finish_join(msk, ssk))
847 mptcp_subflow_reset(ssk);
848 unlock_sock_fast(ssk, slow);
852 static bool mptcp_timer_pending(struct sock *sk)
854 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
857 static void mptcp_reset_timer(struct sock *sk)
859 struct inet_connection_sock *icsk = inet_csk(sk);
862 /* prevent rescheduling on close */
863 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
866 tout = mptcp_sk(sk)->timer_ival;
867 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
870 bool mptcp_schedule_work(struct sock *sk)
872 if (inet_sk_state_load(sk) != TCP_CLOSE &&
873 schedule_work(&mptcp_sk(sk)->work)) {
874 /* each subflow already holds a reference to the sk, and the
875 * workqueue is invoked by a subflow, so sk can't go away here.
883 void mptcp_subflow_eof(struct sock *sk)
885 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
886 mptcp_schedule_work(sk);
889 static void mptcp_check_for_eof(struct mptcp_sock *msk)
891 struct mptcp_subflow_context *subflow;
892 struct sock *sk = (struct sock *)msk;
895 mptcp_for_each_subflow(msk, subflow)
896 receivers += !subflow->rx_eof;
900 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
901 /* hopefully temporary hack: propagate shutdown status
902 * to msk, when all subflows agree on it
904 sk->sk_shutdown |= RCV_SHUTDOWN;
906 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
907 sk->sk_data_ready(sk);
910 switch (sk->sk_state) {
911 case TCP_ESTABLISHED:
912 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
915 inet_sk_state_store(sk, TCP_CLOSING);
918 inet_sk_state_store(sk, TCP_CLOSE);
923 mptcp_close_wake_up(sk);
926 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
928 struct mptcp_subflow_context *subflow;
929 struct sock *sk = (struct sock *)msk;
931 sock_owned_by_me(sk);
933 mptcp_for_each_subflow(msk, subflow) {
934 if (READ_ONCE(subflow->data_avail))
935 return mptcp_subflow_tcp_sock(subflow);
941 static bool mptcp_skb_can_collapse_to(u64 write_seq,
942 const struct sk_buff *skb,
943 const struct mptcp_ext *mpext)
945 if (!tcp_skb_can_collapse_to(skb))
948 /* can collapse only if MPTCP level sequence is in order and this
949 * mapping has not been xmitted yet
951 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
955 /* we can append data to the given data frag if:
956 * - there is space available in the backing page_frag
957 * - the data frag tail matches the current page_frag free offset
958 * - the data frag end sequence number matches the current write seq
960 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
961 const struct page_frag *pfrag,
962 const struct mptcp_data_frag *df)
964 return df && pfrag->page == df->page &&
965 pfrag->size - pfrag->offset > 0 &&
966 pfrag->offset == (df->offset + df->data_len) &&
967 df->data_seq + df->data_len == msk->write_seq;
970 static void dfrag_uncharge(struct sock *sk, int len)
972 sk_mem_uncharge(sk, len);
973 sk_wmem_queued_add(sk, -len);
976 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
978 int len = dfrag->data_len + dfrag->overhead;
980 list_del(&dfrag->list);
981 dfrag_uncharge(sk, len);
982 put_page(dfrag->page);
985 static void __mptcp_clean_una(struct sock *sk)
987 struct mptcp_sock *msk = mptcp_sk(sk);
988 struct mptcp_data_frag *dtmp, *dfrag;
991 /* on fallback we just need to ignore snd_una, as this is really
994 if (__mptcp_check_fallback(msk))
995 msk->snd_una = READ_ONCE(msk->snd_nxt);
997 snd_una = msk->snd_una;
998 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
999 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1002 if (unlikely(dfrag == msk->first_pending)) {
1003 /* in recovery mode can see ack after the current snd head */
1004 if (WARN_ON_ONCE(!msk->recovery))
1007 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1010 dfrag_clear(sk, dfrag);
1013 dfrag = mptcp_rtx_head(sk);
1014 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1015 u64 delta = snd_una - dfrag->data_seq;
1017 /* prevent wrap around in recovery mode */
1018 if (unlikely(delta > dfrag->already_sent)) {
1019 if (WARN_ON_ONCE(!msk->recovery))
1021 if (WARN_ON_ONCE(delta > dfrag->data_len))
1023 dfrag->already_sent += delta - dfrag->already_sent;
1026 dfrag->data_seq += delta;
1027 dfrag->offset += delta;
1028 dfrag->data_len -= delta;
1029 dfrag->already_sent -= delta;
1031 dfrag_uncharge(sk, delta);
1034 /* all retransmitted data acked, recovery completed */
1035 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1036 msk->recovery = false;
1039 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1040 snd_una == READ_ONCE(msk->write_seq)) {
1041 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1042 mptcp_stop_timer(sk);
1044 mptcp_reset_timer(sk);
1048 static void __mptcp_clean_una_wakeup(struct sock *sk)
1050 lockdep_assert_held_once(&sk->sk_lock.slock);
1052 __mptcp_clean_una(sk);
1053 mptcp_write_space(sk);
1056 static void mptcp_clean_una_wakeup(struct sock *sk)
1058 mptcp_data_lock(sk);
1059 __mptcp_clean_una_wakeup(sk);
1060 mptcp_data_unlock(sk);
1063 static void mptcp_enter_memory_pressure(struct sock *sk)
1065 struct mptcp_subflow_context *subflow;
1066 struct mptcp_sock *msk = mptcp_sk(sk);
1069 sk_stream_moderate_sndbuf(sk);
1070 mptcp_for_each_subflow(msk, subflow) {
1071 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1074 tcp_enter_memory_pressure(ssk);
1075 sk_stream_moderate_sndbuf(ssk);
1080 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1083 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1085 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1086 pfrag, sk->sk_allocation)))
1089 mptcp_enter_memory_pressure(sk);
1093 static struct mptcp_data_frag *
1094 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1097 int offset = ALIGN(orig_offset, sizeof(long));
1098 struct mptcp_data_frag *dfrag;
1100 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1101 dfrag->data_len = 0;
1102 dfrag->data_seq = msk->write_seq;
1103 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1104 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1105 dfrag->already_sent = 0;
1106 dfrag->page = pfrag->page;
1111 struct mptcp_sendmsg_info {
1117 bool data_lock_held;
1120 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1121 u64 data_seq, int avail_size)
1123 u64 window_end = mptcp_wnd_end(msk);
1126 if (__mptcp_check_fallback(msk))
1129 mptcp_snd_wnd = window_end - data_seq;
1130 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1132 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1133 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1134 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1140 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1142 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1146 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1150 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1152 struct sk_buff *skb;
1154 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1156 if (likely(__mptcp_add_ext(skb, gfp))) {
1157 skb_reserve(skb, MAX_TCP_HEADER);
1158 skb->ip_summed = CHECKSUM_PARTIAL;
1159 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1164 mptcp_enter_memory_pressure(sk);
1169 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1171 struct sk_buff *skb;
1173 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1177 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1178 tcp_skb_entail(ssk, skb);
1181 tcp_skb_tsorted_anchor_cleanup(skb);
1186 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1188 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1190 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1193 /* note: this always recompute the csum on the whole skb, even
1194 * if we just appended a single frag. More status info needed
1196 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1198 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1199 __wsum csum = ~csum_unfold(mpext->csum);
1200 int offset = skb->len - added;
1202 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1205 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1207 struct mptcp_ext *mpext)
1212 mpext->infinite_map = 1;
1213 mpext->data_len = 0;
1215 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1216 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1218 mptcp_do_fallback(ssk);
1221 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1222 struct mptcp_data_frag *dfrag,
1223 struct mptcp_sendmsg_info *info)
1225 u64 data_seq = dfrag->data_seq + info->sent;
1226 int offset = dfrag->offset + info->sent;
1227 struct mptcp_sock *msk = mptcp_sk(sk);
1228 bool zero_window_probe = false;
1229 struct mptcp_ext *mpext = NULL;
1230 bool can_coalesce = false;
1231 bool reuse_skb = true;
1232 struct sk_buff *skb;
1236 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1237 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1239 if (WARN_ON_ONCE(info->sent > info->limit ||
1240 info->limit > dfrag->data_len))
1243 /* compute send limit */
1244 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1245 copy = info->size_goal;
1247 skb = tcp_write_queue_tail(ssk);
1248 if (skb && copy > skb->len) {
1249 /* Limit the write to the size available in the
1250 * current skb, if any, so that we create at most a new skb.
1251 * Explicitly tells TCP internals to avoid collapsing on later
1252 * queue management operation, to avoid breaking the ext <->
1253 * SSN association set here
1255 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1256 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1257 TCP_SKB_CB(skb)->eor = 1;
1261 i = skb_shinfo(skb)->nr_frags;
1262 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1263 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1264 tcp_mark_push(tcp_sk(ssk), skb);
1271 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1275 i = skb_shinfo(skb)->nr_frags;
1277 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1280 /* Zero window and all data acked? Probe. */
1281 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1283 u64 snd_una = READ_ONCE(msk->snd_una);
1285 if (snd_una != msk->snd_nxt) {
1286 tcp_remove_empty_skb(ssk);
1290 zero_window_probe = true;
1291 data_seq = snd_una - 1;
1294 /* all mptcp-level data is acked, no skbs should be present into the
1297 WARN_ON_ONCE(reuse_skb);
1300 copy = min_t(size_t, copy, info->limit - info->sent);
1301 if (!sk_wmem_schedule(ssk, copy)) {
1302 tcp_remove_empty_skb(ssk);
1307 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1309 get_page(dfrag->page);
1310 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1314 skb->data_len += copy;
1315 skb->truesize += copy;
1316 sk_wmem_queued_add(ssk, copy);
1317 sk_mem_charge(ssk, copy);
1318 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1319 TCP_SKB_CB(skb)->end_seq += copy;
1320 tcp_skb_pcount_set(skb, 0);
1322 /* on skb reuse we just need to update the DSS len */
1324 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1325 mpext->data_len += copy;
1326 WARN_ON_ONCE(zero_window_probe);
1330 memset(mpext, 0, sizeof(*mpext));
1331 mpext->data_seq = data_seq;
1332 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1333 mpext->data_len = copy;
1337 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1338 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1341 if (zero_window_probe) {
1342 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1344 if (READ_ONCE(msk->csum_enabled))
1345 mptcp_update_data_checksum(skb, copy);
1346 tcp_push_pending_frames(ssk);
1350 if (READ_ONCE(msk->csum_enabled))
1351 mptcp_update_data_checksum(skb, copy);
1352 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1353 mptcp_update_infinite_map(msk, ssk, mpext);
1354 trace_mptcp_sendmsg_frag(mpext);
1355 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1359 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1360 sizeof(struct tcphdr) - \
1361 MAX_TCP_OPTION_SPACE - \
1362 sizeof(struct ipv6hdr) - \
1363 sizeof(struct frag_hdr))
1365 struct subflow_send_info {
1370 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1372 if (!subflow->stale)
1376 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1379 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1381 if (unlikely(subflow->stale)) {
1382 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1384 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1387 mptcp_subflow_set_active(subflow);
1389 return __mptcp_subflow_active(subflow);
1392 #define SSK_MODE_ACTIVE 0
1393 #define SSK_MODE_BACKUP 1
1394 #define SSK_MODE_MAX 2
1396 /* implement the mptcp packet scheduler;
1397 * returns the subflow that will transmit the next DSS
1398 * additionally updates the rtx timeout
1400 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1402 struct subflow_send_info send_info[SSK_MODE_MAX];
1403 struct mptcp_subflow_context *subflow;
1404 struct sock *sk = (struct sock *)msk;
1405 u32 pace, burst, wmem;
1406 int i, nr_active = 0;
1411 sock_owned_by_me(sk);
1413 if (__mptcp_check_fallback(msk)) {
1416 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1419 /* re-use last subflow, if the burst allow that */
1420 if (msk->last_snd && msk->snd_burst > 0 &&
1421 sk_stream_memory_free(msk->last_snd) &&
1422 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1423 mptcp_set_timeout(sk);
1424 return msk->last_snd;
1427 /* pick the subflow with the lower wmem/wspace ratio */
1428 for (i = 0; i < SSK_MODE_MAX; ++i) {
1429 send_info[i].ssk = NULL;
1430 send_info[i].linger_time = -1;
1433 mptcp_for_each_subflow(msk, subflow) {
1434 trace_mptcp_subflow_get_send(subflow);
1435 ssk = mptcp_subflow_tcp_sock(subflow);
1436 if (!mptcp_subflow_active(subflow))
1439 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1440 nr_active += !subflow->backup;
1441 pace = subflow->avg_pacing_rate;
1442 if (unlikely(!pace)) {
1443 /* init pacing rate from socket */
1444 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1445 pace = subflow->avg_pacing_rate;
1450 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1451 if (linger_time < send_info[subflow->backup].linger_time) {
1452 send_info[subflow->backup].ssk = ssk;
1453 send_info[subflow->backup].linger_time = linger_time;
1456 __mptcp_set_timeout(sk, tout);
1458 /* pick the best backup if no other subflow is active */
1460 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1462 /* According to the blest algorithm, to avoid HoL blocking for the
1463 * faster flow, we need to:
1464 * - estimate the faster flow linger time
1465 * - use the above to estimate the amount of byte transferred
1466 * by the faster flow
1467 * - check that the amount of queued data is greter than the above,
1468 * otherwise do not use the picked, slower, subflow
1469 * We select the subflow with the shorter estimated time to flush
1470 * the queued mem, which basically ensure the above. We just need
1471 * to check that subflow has a non empty cwin.
1473 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1474 if (!ssk || !sk_stream_memory_free(ssk))
1477 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1478 wmem = READ_ONCE(ssk->sk_wmem_queued);
1480 msk->last_snd = NULL;
1484 subflow = mptcp_subflow_ctx(ssk);
1485 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1486 READ_ONCE(ssk->sk_pacing_rate) * burst,
1488 msk->last_snd = ssk;
1489 msk->snd_burst = burst;
1493 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1495 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1499 static void mptcp_update_post_push(struct mptcp_sock *msk,
1500 struct mptcp_data_frag *dfrag,
1503 u64 snd_nxt_new = dfrag->data_seq;
1505 dfrag->already_sent += sent;
1507 msk->snd_burst -= sent;
1509 snd_nxt_new += dfrag->already_sent;
1511 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1512 * is recovering after a failover. In that event, this re-sends
1515 * Thus compute snd_nxt_new candidate based on
1516 * the dfrag->data_seq that was sent and the data
1517 * that has been handed to the subflow for transmission
1518 * and skip update in case it was old dfrag.
1520 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1521 msk->snd_nxt = snd_nxt_new;
1524 void mptcp_check_and_set_pending(struct sock *sk)
1526 if (mptcp_send_head(sk))
1527 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1530 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1532 struct sock *prev_ssk = NULL, *ssk = NULL;
1533 struct mptcp_sock *msk = mptcp_sk(sk);
1534 struct mptcp_sendmsg_info info = {
1537 struct mptcp_data_frag *dfrag;
1538 int len, copied = 0;
1540 while ((dfrag = mptcp_send_head(sk))) {
1541 info.sent = dfrag->already_sent;
1542 info.limit = dfrag->data_len;
1543 len = dfrag->data_len - dfrag->already_sent;
1548 ssk = mptcp_subflow_get_send(msk);
1550 /* First check. If the ssk has changed since
1551 * the last round, release prev_ssk
1553 if (ssk != prev_ssk && prev_ssk)
1554 mptcp_push_release(prev_ssk, &info);
1558 /* Need to lock the new subflow only if different
1559 * from the previous one, otherwise we are still
1560 * helding the relevant lock
1562 if (ssk != prev_ssk)
1565 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1567 mptcp_push_release(ssk, &info);
1575 mptcp_update_post_push(msk, dfrag, ret);
1577 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1580 /* at this point we held the socket lock for the last subflow we used */
1582 mptcp_push_release(ssk, &info);
1585 /* ensure the rtx timer is running */
1586 if (!mptcp_timer_pending(sk))
1587 mptcp_reset_timer(sk);
1589 __mptcp_check_send_data_fin(sk);
1592 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1594 struct mptcp_sock *msk = mptcp_sk(sk);
1595 struct mptcp_sendmsg_info info = {
1596 .data_lock_held = true,
1598 struct mptcp_data_frag *dfrag;
1599 struct sock *xmit_ssk;
1600 int len, copied = 0;
1604 while ((dfrag = mptcp_send_head(sk))) {
1605 info.sent = dfrag->already_sent;
1606 info.limit = dfrag->data_len;
1607 len = dfrag->data_len - dfrag->already_sent;
1611 /* the caller already invoked the packet scheduler,
1612 * check for a different subflow usage only after
1613 * spooling the first chunk of data
1615 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1618 if (xmit_ssk != ssk) {
1619 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1620 MPTCP_DELEGATE_SEND);
1624 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1633 mptcp_update_post_push(msk, dfrag, ret);
1635 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1639 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1640 * not going to flush it via release_sock()
1643 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1645 if (!mptcp_timer_pending(sk))
1646 mptcp_reset_timer(sk);
1648 if (msk->snd_data_fin_enable &&
1649 msk->snd_nxt + 1 == msk->write_seq)
1650 mptcp_schedule_work(sk);
1654 static void mptcp_set_nospace(struct sock *sk)
1656 /* enable autotune */
1657 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1659 /* will be cleared on avail space */
1660 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1663 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1665 struct mptcp_sock *msk = mptcp_sk(sk);
1666 struct page_frag *pfrag;
1671 /* we don't support FASTOPEN yet */
1672 if (msg->msg_flags & MSG_FASTOPEN)
1675 /* silently ignore everything else */
1676 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1680 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1682 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1683 ret = sk_stream_wait_connect(sk, &timeo);
1688 pfrag = sk_page_frag(sk);
1690 while (msg_data_left(msg)) {
1691 int total_ts, frag_truesize = 0;
1692 struct mptcp_data_frag *dfrag;
1693 bool dfrag_collapsed;
1694 size_t psize, offset;
1696 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1701 /* reuse tail pfrag, if possible, or carve a new one from the
1704 dfrag = mptcp_pending_tail(sk);
1705 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1706 if (!dfrag_collapsed) {
1707 if (!sk_stream_memory_free(sk))
1708 goto wait_for_memory;
1710 if (!mptcp_page_frag_refill(sk, pfrag))
1711 goto wait_for_memory;
1713 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1714 frag_truesize = dfrag->overhead;
1717 /* we do not bound vs wspace, to allow a single packet.
1718 * memory accounting will prevent execessive memory usage
1721 offset = dfrag->offset + dfrag->data_len;
1722 psize = pfrag->size - offset;
1723 psize = min_t(size_t, psize, msg_data_left(msg));
1724 total_ts = psize + frag_truesize;
1726 if (!sk_wmem_schedule(sk, total_ts))
1727 goto wait_for_memory;
1729 if (copy_page_from_iter(dfrag->page, offset, psize,
1730 &msg->msg_iter) != psize) {
1735 /* data successfully copied into the write queue */
1736 sk->sk_forward_alloc -= total_ts;
1738 dfrag->data_len += psize;
1739 frag_truesize += psize;
1740 pfrag->offset += frag_truesize;
1741 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1743 /* charge data on mptcp pending queue to the msk socket
1744 * Note: we charge such data both to sk and ssk
1746 sk_wmem_queued_add(sk, frag_truesize);
1747 if (!dfrag_collapsed) {
1748 get_page(dfrag->page);
1749 list_add_tail(&dfrag->list, &msk->rtx_queue);
1750 if (!msk->first_pending)
1751 WRITE_ONCE(msk->first_pending, dfrag);
1753 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1754 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1760 mptcp_set_nospace(sk);
1761 __mptcp_push_pending(sk, msg->msg_flags);
1762 ret = sk_stream_wait_memory(sk, &timeo);
1768 __mptcp_push_pending(sk, msg->msg_flags);
1772 return copied ? : ret;
1775 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1777 size_t len, int flags,
1778 struct scm_timestamping_internal *tss,
1781 struct sk_buff *skb, *tmp;
1784 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1785 u32 offset = MPTCP_SKB_CB(skb)->offset;
1786 u32 data_len = skb->len - offset;
1787 u32 count = min_t(size_t, len - copied, data_len);
1790 if (!(flags & MSG_TRUNC)) {
1791 err = skb_copy_datagram_msg(skb, offset, msg, count);
1792 if (unlikely(err < 0)) {
1799 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1800 tcp_update_recv_tstamps(skb, tss);
1801 *cmsg_flags |= MPTCP_CMSG_TS;
1806 if (count < data_len) {
1807 if (!(flags & MSG_PEEK)) {
1808 MPTCP_SKB_CB(skb)->offset += count;
1809 MPTCP_SKB_CB(skb)->map_seq += 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 mptcp_rmem_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;
1953 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1956 /* we can have data pending in the subflows only if the msk
1957 * receive buffer was full at subflow_data_ready() time,
1958 * that is an unlikely slow path.
1963 slowpath = lock_sock_fast(ssk);
1964 mptcp_data_lock(sk);
1965 __mptcp_update_rmem(sk);
1966 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1967 mptcp_data_unlock(sk);
1969 if (unlikely(ssk->sk_err))
1970 __mptcp_error_report(sk);
1971 unlock_sock_fast(ssk, slowpath);
1974 /* acquire the data lock only if some input data is pending */
1976 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1977 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1978 mptcp_data_lock(sk);
1979 __mptcp_update_rmem(sk);
1980 ret |= __mptcp_ofo_queue(msk);
1981 __mptcp_splice_receive_queue(sk);
1982 mptcp_data_unlock(sk);
1985 mptcp_check_data_fin((struct sock *)msk);
1986 return !skb_queue_empty(&msk->receive_queue);
1989 static unsigned int mptcp_inq_hint(const struct sock *sk)
1991 const struct mptcp_sock *msk = mptcp_sk(sk);
1992 const struct sk_buff *skb;
1994 skb = skb_peek(&msk->receive_queue);
1996 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
1998 if (hint_val >= INT_MAX)
2001 return (unsigned int)hint_val;
2004 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2010 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2011 int flags, int *addr_len)
2013 struct mptcp_sock *msk = mptcp_sk(sk);
2014 struct scm_timestamping_internal tss;
2015 int copied = 0, cmsg_flags = 0;
2019 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2020 if (unlikely(flags & MSG_ERRQUEUE))
2021 return inet_recv_error(sk, msg, len, addr_len);
2024 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2029 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2031 len = min_t(size_t, len, INT_MAX);
2032 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2034 if (unlikely(msk->recvmsg_inq))
2035 cmsg_flags = MPTCP_CMSG_INQ;
2037 while (copied < len) {
2040 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2041 if (unlikely(bytes_read < 0)) {
2043 copied = bytes_read;
2047 copied += bytes_read;
2049 /* be sure to advertise window change */
2050 mptcp_cleanup_rbuf(msk);
2052 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2055 /* only the master socket status is relevant here. The exit
2056 * conditions mirror closely tcp_recvmsg()
2058 if (copied >= target)
2063 sk->sk_state == TCP_CLOSE ||
2064 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2066 signal_pending(current))
2070 copied = sock_error(sk);
2074 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2075 mptcp_check_for_eof(msk);
2077 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2078 /* race breaker: the shutdown could be after the
2079 * previous receive queue check
2081 if (__mptcp_move_skbs(msk))
2086 if (sk->sk_state == TCP_CLOSE) {
2096 if (signal_pending(current)) {
2097 copied = sock_intr_errno(timeo);
2102 pr_debug("block timeout %ld", timeo);
2103 sk_wait_data(sk, &timeo, NULL);
2107 if (cmsg_flags && copied >= 0) {
2108 if (cmsg_flags & MPTCP_CMSG_TS)
2109 tcp_recv_timestamp(msg, sk, &tss);
2111 if (cmsg_flags & MPTCP_CMSG_INQ) {
2112 unsigned int inq = mptcp_inq_hint(sk);
2114 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2118 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2119 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2120 skb_queue_empty(&msk->receive_queue), copied);
2121 if (!(flags & MSG_PEEK))
2122 mptcp_rcv_space_adjust(msk, copied);
2128 static void mptcp_retransmit_timer(struct timer_list *t)
2130 struct inet_connection_sock *icsk = from_timer(icsk, t,
2131 icsk_retransmit_timer);
2132 struct sock *sk = &icsk->icsk_inet.sk;
2133 struct mptcp_sock *msk = mptcp_sk(sk);
2136 if (!sock_owned_by_user(sk)) {
2137 /* we need a process context to retransmit */
2138 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2139 mptcp_schedule_work(sk);
2141 /* delegate our work to tcp_release_cb() */
2142 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2148 static void mptcp_timeout_timer(struct timer_list *t)
2150 struct sock *sk = from_timer(sk, t, sk_timer);
2152 mptcp_schedule_work(sk);
2156 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2159 * A backup subflow is returned only if that is the only kind available.
2161 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2163 struct sock *backup = NULL, *pick = NULL;
2164 struct mptcp_subflow_context *subflow;
2165 int min_stale_count = INT_MAX;
2167 sock_owned_by_me((const struct sock *)msk);
2169 if (__mptcp_check_fallback(msk))
2172 mptcp_for_each_subflow(msk, subflow) {
2173 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2175 if (!__mptcp_subflow_active(subflow))
2178 /* still data outstanding at TCP level? skip this */
2179 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2180 mptcp_pm_subflow_chk_stale(msk, ssk);
2181 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2185 if (subflow->backup) {
2198 /* use backup only if there are no progresses anywhere */
2199 return min_stale_count > 1 ? backup : NULL;
2202 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2205 iput(SOCK_INODE(msk->subflow));
2206 msk->subflow = NULL;
2210 bool __mptcp_retransmit_pending_data(struct sock *sk)
2212 struct mptcp_data_frag *cur, *rtx_head;
2213 struct mptcp_sock *msk = mptcp_sk(sk);
2215 if (__mptcp_check_fallback(mptcp_sk(sk)))
2218 if (tcp_rtx_and_write_queues_empty(sk))
2221 /* the closing socket has some data untransmitted and/or unacked:
2222 * some data in the mptcp rtx queue has not really xmitted yet.
2223 * keep it simple and re-inject the whole mptcp level rtx queue
2225 mptcp_data_lock(sk);
2226 __mptcp_clean_una_wakeup(sk);
2227 rtx_head = mptcp_rtx_head(sk);
2229 mptcp_data_unlock(sk);
2233 msk->recovery_snd_nxt = msk->snd_nxt;
2234 msk->recovery = true;
2235 mptcp_data_unlock(sk);
2237 msk->first_pending = rtx_head;
2240 /* be sure to clear the "sent status" on all re-injected fragments */
2241 list_for_each_entry(cur, &msk->rtx_queue, list) {
2242 if (!cur->already_sent)
2244 cur->already_sent = 0;
2250 /* flags for __mptcp_close_ssk() */
2251 #define MPTCP_CF_PUSH BIT(1)
2252 #define MPTCP_CF_FASTCLOSE BIT(2)
2254 /* subflow sockets can be either outgoing (connect) or incoming
2257 * Outgoing subflows use in-kernel sockets.
2258 * Incoming subflows do not have their own 'struct socket' allocated,
2259 * so we need to use tcp_close() after detaching them from the mptcp
2262 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2263 struct mptcp_subflow_context *subflow,
2266 struct mptcp_sock *msk = mptcp_sk(sk);
2267 bool need_push, dispose_it;
2269 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2271 list_del(&subflow->node);
2273 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2275 if (flags & MPTCP_CF_FASTCLOSE)
2276 subflow->send_fastclose = 1;
2278 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2280 tcp_disconnect(ssk, 0);
2281 msk->subflow->state = SS_UNCONNECTED;
2282 mptcp_subflow_ctx_reset(subflow);
2288 /* if we are invoked by the msk cleanup code, the subflow is
2294 subflow->disposable = 1;
2296 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2297 * the ssk has been already destroyed, we just need to release the
2298 * reference owned by msk;
2300 if (!inet_csk(ssk)->icsk_ulp_ops) {
2301 kfree_rcu(subflow, rcu);
2303 /* otherwise tcp will dispose of the ssk and subflow ctx */
2304 if (ssk->sk_state == TCP_LISTEN) {
2305 tcp_set_state(ssk, TCP_CLOSE);
2306 mptcp_subflow_queue_clean(ssk);
2307 inet_csk_listen_stop(ssk);
2309 __tcp_close(ssk, 0);
2311 /* close acquired an extra ref */
2318 if (ssk == msk->first)
2322 if (ssk == msk->last_snd)
2323 msk->last_snd = NULL;
2326 __mptcp_push_pending(sk, 0);
2329 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2330 struct mptcp_subflow_context *subflow)
2332 if (sk->sk_state == TCP_ESTABLISHED)
2333 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2335 /* subflow aborted before reaching the fully_established status
2336 * attempt the creation of the next subflow
2338 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2340 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2343 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2348 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2350 struct mptcp_subflow_context *subflow, *tmp;
2354 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2355 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2357 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2360 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2361 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2364 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2368 static bool mptcp_check_close_timeout(const struct sock *sk)
2370 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2371 struct mptcp_subflow_context *subflow;
2373 if (delta >= TCP_TIMEWAIT_LEN)
2376 /* if all subflows are in closed status don't bother with additional
2379 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2380 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2387 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2389 struct mptcp_subflow_context *subflow, *tmp;
2390 struct sock *sk = &msk->sk.icsk_inet.sk;
2392 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2395 mptcp_token_destroy(msk);
2397 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2398 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2401 slow = lock_sock_fast(tcp_sk);
2402 if (tcp_sk->sk_state != TCP_CLOSE) {
2403 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2404 tcp_set_state(tcp_sk, TCP_CLOSE);
2406 unlock_sock_fast(tcp_sk, slow);
2409 inet_sk_state_store(sk, TCP_CLOSE);
2410 sk->sk_shutdown = SHUTDOWN_MASK;
2411 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2412 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2414 mptcp_close_wake_up(sk);
2417 static void __mptcp_retrans(struct sock *sk)
2419 struct mptcp_sock *msk = mptcp_sk(sk);
2420 struct mptcp_sendmsg_info info = {};
2421 struct mptcp_data_frag *dfrag;
2426 mptcp_clean_una_wakeup(sk);
2428 /* first check ssk: need to kick "stale" logic */
2429 ssk = mptcp_subflow_get_retrans(msk);
2430 dfrag = mptcp_rtx_head(sk);
2432 if (mptcp_data_fin_enabled(msk)) {
2433 struct inet_connection_sock *icsk = inet_csk(sk);
2435 icsk->icsk_retransmits++;
2436 mptcp_set_datafin_timeout(sk);
2437 mptcp_send_ack(msk);
2442 if (!mptcp_send_head(sk))
2453 /* limit retransmission to the bytes already sent on some subflows */
2455 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2456 while (info.sent < info.limit) {
2457 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2461 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2466 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2467 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2469 WRITE_ONCE(msk->allow_infinite_fallback, false);
2475 mptcp_check_and_set_pending(sk);
2477 if (!mptcp_timer_pending(sk))
2478 mptcp_reset_timer(sk);
2481 /* schedule the timeout timer for the relevant event: either close timeout
2482 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2484 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2486 struct sock *sk = (struct sock *)msk;
2487 unsigned long timeout, close_timeout;
2489 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2492 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2494 /* the close timeout takes precedence on the fail one, and here at least one of
2497 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2499 sk_reset_timer(sk, &sk->sk_timer, timeout);
2502 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2504 struct sock *ssk = msk->first;
2510 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2512 slow = lock_sock_fast(ssk);
2513 mptcp_subflow_reset(ssk);
2514 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2515 unlock_sock_fast(ssk, slow);
2517 mptcp_reset_timeout(msk, 0);
2520 static void mptcp_worker(struct work_struct *work)
2522 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2523 struct sock *sk = &msk->sk.icsk_inet.sk;
2524 unsigned long fail_tout;
2528 state = sk->sk_state;
2529 if (unlikely(state == TCP_CLOSE))
2532 mptcp_check_data_fin_ack(sk);
2534 mptcp_check_fastclose(msk);
2536 mptcp_pm_nl_work(msk);
2538 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2539 mptcp_check_for_eof(msk);
2541 __mptcp_check_send_data_fin(sk);
2542 mptcp_check_data_fin(sk);
2544 /* There is no point in keeping around an orphaned sk timedout or
2545 * closed, but we need the msk around to reply to incoming DATA_FIN,
2546 * even if it is orphaned and in FIN_WAIT2 state
2548 if (sock_flag(sk, SOCK_DEAD) &&
2549 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2550 inet_sk_state_store(sk, TCP_CLOSE);
2551 __mptcp_destroy_sock(sk);
2555 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2556 __mptcp_close_subflow(msk);
2558 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2559 __mptcp_retrans(sk);
2561 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2562 if (fail_tout && time_after(jiffies, fail_tout))
2563 mptcp_mp_fail_no_response(msk);
2570 static int __mptcp_init_sock(struct sock *sk)
2572 struct mptcp_sock *msk = mptcp_sk(sk);
2574 INIT_LIST_HEAD(&msk->conn_list);
2575 INIT_LIST_HEAD(&msk->join_list);
2576 INIT_LIST_HEAD(&msk->rtx_queue);
2577 INIT_WORK(&msk->work, mptcp_worker);
2578 __skb_queue_head_init(&msk->receive_queue);
2579 msk->out_of_order_queue = RB_ROOT;
2580 msk->first_pending = NULL;
2581 msk->rmem_fwd_alloc = 0;
2582 WRITE_ONCE(msk->rmem_released, 0);
2583 msk->timer_ival = TCP_RTO_MIN;
2586 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2587 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2588 WRITE_ONCE(msk->allow_infinite_fallback, true);
2589 msk->recovery = false;
2591 mptcp_pm_data_init(msk);
2593 /* re-use the csk retrans timer for MPTCP-level retrans */
2594 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2595 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2600 static void mptcp_ca_reset(struct sock *sk)
2602 struct inet_connection_sock *icsk = inet_csk(sk);
2604 tcp_assign_congestion_control(sk);
2605 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2607 /* no need to keep a reference to the ops, the name will suffice */
2608 tcp_cleanup_congestion_control(sk);
2609 icsk->icsk_ca_ops = NULL;
2612 static int mptcp_init_sock(struct sock *sk)
2614 struct net *net = sock_net(sk);
2617 ret = __mptcp_init_sock(sk);
2621 if (!mptcp_is_enabled(net))
2622 return -ENOPROTOOPT;
2624 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2627 ret = __mptcp_socket_create(mptcp_sk(sk));
2631 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2632 * propagate the correct value
2636 sk_sockets_allocated_inc(sk);
2637 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2638 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2643 static void __mptcp_clear_xmit(struct sock *sk)
2645 struct mptcp_sock *msk = mptcp_sk(sk);
2646 struct mptcp_data_frag *dtmp, *dfrag;
2648 WRITE_ONCE(msk->first_pending, NULL);
2649 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2650 dfrag_clear(sk, dfrag);
2653 static void mptcp_cancel_work(struct sock *sk)
2655 struct mptcp_sock *msk = mptcp_sk(sk);
2657 if (cancel_work_sync(&msk->work))
2661 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2665 switch (ssk->sk_state) {
2667 if (!(how & RCV_SHUTDOWN))
2671 tcp_disconnect(ssk, O_NONBLOCK);
2674 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2675 pr_debug("Fallback");
2676 ssk->sk_shutdown |= how;
2677 tcp_shutdown(ssk, how);
2679 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2681 if (!mptcp_timer_pending(sk))
2682 mptcp_reset_timer(sk);
2690 static const unsigned char new_state[16] = {
2691 /* current state: new state: action: */
2692 [0 /* (Invalid) */] = TCP_CLOSE,
2693 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2694 [TCP_SYN_SENT] = TCP_CLOSE,
2695 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2696 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2697 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2698 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2699 [TCP_CLOSE] = TCP_CLOSE,
2700 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2701 [TCP_LAST_ACK] = TCP_LAST_ACK,
2702 [TCP_LISTEN] = TCP_CLOSE,
2703 [TCP_CLOSING] = TCP_CLOSING,
2704 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2707 static int mptcp_close_state(struct sock *sk)
2709 int next = (int)new_state[sk->sk_state];
2710 int ns = next & TCP_STATE_MASK;
2712 inet_sk_state_store(sk, ns);
2714 return next & TCP_ACTION_FIN;
2717 static void __mptcp_check_send_data_fin(struct sock *sk)
2719 struct mptcp_subflow_context *subflow;
2720 struct mptcp_sock *msk = mptcp_sk(sk);
2722 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2723 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2724 msk->snd_nxt, msk->write_seq);
2726 /* we still need to enqueue subflows or not really shutting down,
2729 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2730 mptcp_send_head(sk))
2733 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2735 /* fallback socket will not get data_fin/ack, can move to the next
2738 if (__mptcp_check_fallback(msk)) {
2739 WRITE_ONCE(msk->snd_una, msk->write_seq);
2740 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2741 inet_sk_state_store(sk, TCP_CLOSE);
2742 mptcp_close_wake_up(sk);
2743 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2744 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2748 mptcp_for_each_subflow(msk, subflow) {
2749 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2751 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2755 static void __mptcp_wr_shutdown(struct sock *sk)
2757 struct mptcp_sock *msk = mptcp_sk(sk);
2759 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2760 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2761 !!mptcp_send_head(sk));
2763 /* will be ignored by fallback sockets */
2764 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2765 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2767 __mptcp_check_send_data_fin(sk);
2770 static void __mptcp_destroy_sock(struct sock *sk)
2772 struct mptcp_subflow_context *subflow, *tmp;
2773 struct mptcp_sock *msk = mptcp_sk(sk);
2774 LIST_HEAD(conn_list);
2776 pr_debug("msk=%p", msk);
2780 /* join list will be eventually flushed (with rst) at sock lock release time*/
2781 list_splice_init(&msk->conn_list, &conn_list);
2783 mptcp_stop_timer(sk);
2784 sk_stop_timer(sk, &sk->sk_timer);
2787 /* clears msk->subflow, allowing the following loop to close
2788 * even the initial subflow
2790 mptcp_dispose_initial_subflow(msk);
2791 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2792 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2793 __mptcp_close_ssk(sk, ssk, subflow, 0);
2796 sk->sk_prot->destroy(sk);
2798 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2799 WARN_ON_ONCE(msk->rmem_released);
2800 sk_stream_kill_queues(sk);
2801 xfrm_sk_free_policy(sk);
2803 sk_refcnt_debug_release(sk);
2807 static void mptcp_close(struct sock *sk, long timeout)
2809 struct mptcp_subflow_context *subflow;
2810 struct mptcp_sock *msk = mptcp_sk(sk);
2811 bool do_cancel_work = false;
2814 sk->sk_shutdown = SHUTDOWN_MASK;
2816 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2817 inet_sk_state_store(sk, TCP_CLOSE);
2821 if (mptcp_close_state(sk))
2822 __mptcp_wr_shutdown(sk);
2824 sk_stream_wait_close(sk, timeout);
2827 /* orphan all the subflows */
2828 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2829 mptcp_for_each_subflow(msk, subflow) {
2830 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2831 bool slow = lock_sock_fast_nested(ssk);
2833 /* since the close timeout takes precedence on the fail one,
2836 if (ssk == msk->first)
2837 subflow->fail_tout = 0;
2840 unlock_sock_fast(ssk, slow);
2845 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2846 if (mptcp_sk(sk)->token)
2847 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2849 if (sk->sk_state == TCP_CLOSE) {
2850 __mptcp_destroy_sock(sk);
2851 do_cancel_work = true;
2853 mptcp_reset_timeout(msk, 0);
2857 mptcp_cancel_work(sk);
2862 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2864 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2865 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2866 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2868 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2869 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2872 msk6->saddr = ssk6->saddr;
2873 msk6->flow_label = ssk6->flow_label;
2877 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2878 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2879 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2880 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2881 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2882 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2885 static int mptcp_disconnect(struct sock *sk, int flags)
2887 struct mptcp_subflow_context *subflow, *tmp;
2888 struct mptcp_sock *msk = mptcp_sk(sk);
2890 inet_sk_state_store(sk, TCP_CLOSE);
2892 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2893 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2895 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_FASTCLOSE);
2898 mptcp_stop_timer(sk);
2899 sk_stop_timer(sk, &sk->sk_timer);
2901 if (mptcp_sk(sk)->token)
2902 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2904 mptcp_destroy_common(msk);
2905 msk->last_snd = NULL;
2906 WRITE_ONCE(msk->flags, 0);
2908 msk->push_pending = 0;
2909 msk->recovery = false;
2910 msk->can_ack = false;
2911 msk->fully_established = false;
2912 msk->rcv_data_fin = false;
2913 msk->snd_data_fin_enable = false;
2914 msk->rcv_fastclose = false;
2915 msk->use_64bit_ack = false;
2916 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2917 mptcp_pm_data_reset(msk);
2920 sk->sk_shutdown = 0;
2921 sk_error_report(sk);
2925 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2926 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2928 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2930 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2934 struct sock *mptcp_sk_clone(const struct sock *sk,
2935 const struct mptcp_options_received *mp_opt,
2936 struct request_sock *req)
2938 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2939 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2940 struct mptcp_sock *msk;
2946 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2947 if (nsk->sk_family == AF_INET6)
2948 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2951 __mptcp_init_sock(nsk);
2953 msk = mptcp_sk(nsk);
2954 msk->local_key = subflow_req->local_key;
2955 msk->token = subflow_req->token;
2956 msk->subflow = NULL;
2957 WRITE_ONCE(msk->fully_established, false);
2958 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
2959 WRITE_ONCE(msk->csum_enabled, true);
2961 msk->write_seq = subflow_req->idsn + 1;
2962 msk->snd_nxt = msk->write_seq;
2963 msk->snd_una = msk->write_seq;
2964 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2965 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2967 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
2968 msk->can_ack = true;
2969 msk->remote_key = mp_opt->sndr_key;
2970 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2972 WRITE_ONCE(msk->ack_seq, ack_seq);
2973 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
2976 sock_reset_flag(nsk, SOCK_RCU_FREE);
2977 /* will be fully established after successful MPC subflow creation */
2978 inet_sk_state_store(nsk, TCP_SYN_RECV);
2980 security_inet_csk_clone(nsk, req);
2981 bh_unlock_sock(nsk);
2983 /* keep a single reference */
2988 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2990 const struct tcp_sock *tp = tcp_sk(ssk);
2992 msk->rcvq_space.copied = 0;
2993 msk->rcvq_space.rtt_us = 0;
2995 msk->rcvq_space.time = tp->tcp_mstamp;
2997 /* initial rcv_space offering made to peer */
2998 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2999 TCP_INIT_CWND * tp->advmss);
3000 if (msk->rcvq_space.space == 0)
3001 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3003 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3006 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3009 struct mptcp_sock *msk = mptcp_sk(sk);
3010 struct socket *listener;
3013 listener = __mptcp_nmpc_socket(msk);
3014 if (WARN_ON_ONCE(!listener)) {
3019 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3020 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3024 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3025 if (sk_is_mptcp(newsk)) {
3026 struct mptcp_subflow_context *subflow;
3027 struct sock *new_mptcp_sock;
3029 subflow = mptcp_subflow_ctx(newsk);
3030 new_mptcp_sock = subflow->conn;
3032 /* is_mptcp should be false if subflow->conn is missing, see
3033 * subflow_syn_recv_sock()
3035 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3036 tcp_sk(newsk)->is_mptcp = 0;
3040 /* acquire the 2nd reference for the owning socket */
3041 sock_hold(new_mptcp_sock);
3042 newsk = new_mptcp_sock;
3043 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3045 MPTCP_INC_STATS(sock_net(sk),
3046 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3050 newsk->sk_kern_sock = kern;
3054 void mptcp_destroy_common(struct mptcp_sock *msk)
3056 struct sock *sk = (struct sock *)msk;
3058 __mptcp_clear_xmit(sk);
3060 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3061 mptcp_data_lock(sk);
3062 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3063 __skb_queue_purge(&sk->sk_receive_queue);
3064 skb_rbtree_purge(&msk->out_of_order_queue);
3065 mptcp_data_unlock(sk);
3067 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3068 * inet_sock_destruct() will dispose it
3070 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3071 msk->rmem_fwd_alloc = 0;
3072 mptcp_token_destroy(msk);
3073 mptcp_pm_free_anno_list(msk);
3074 mptcp_free_local_addr_list(msk);
3077 static void mptcp_destroy(struct sock *sk)
3079 struct mptcp_sock *msk = mptcp_sk(sk);
3081 mptcp_destroy_common(msk);
3082 sk_sockets_allocated_dec(sk);
3085 void __mptcp_data_acked(struct sock *sk)
3087 if (!sock_owned_by_user(sk))
3088 __mptcp_clean_una(sk);
3090 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3092 if (mptcp_pending_data_fin_ack(sk))
3093 mptcp_schedule_work(sk);
3096 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3098 if (!mptcp_send_head(sk))
3101 if (!sock_owned_by_user(sk)) {
3102 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3104 if (xmit_ssk == ssk)
3105 __mptcp_subflow_push_pending(sk, ssk);
3107 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3109 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3113 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3114 BIT(MPTCP_RETRANSMIT) | \
3115 BIT(MPTCP_FLUSH_JOIN_LIST))
3117 /* processes deferred events and flush wmem */
3118 static void mptcp_release_cb(struct sock *sk)
3119 __must_hold(&sk->sk_lock.slock)
3121 struct mptcp_sock *msk = mptcp_sk(sk);
3124 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3129 /* the following actions acquire the subflow socket lock
3131 * 1) can't be invoked in atomic scope
3132 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3133 * datapath acquires the msk socket spinlock while helding
3134 * the subflow socket lock
3136 msk->push_pending = 0;
3137 msk->cb_flags &= ~flags;
3138 spin_unlock_bh(&sk->sk_lock.slock);
3139 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3140 __mptcp_flush_join_list(sk);
3141 if (flags & BIT(MPTCP_PUSH_PENDING))
3142 __mptcp_push_pending(sk, 0);
3143 if (flags & BIT(MPTCP_RETRANSMIT))
3144 __mptcp_retrans(sk);
3147 spin_lock_bh(&sk->sk_lock.slock);
3150 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3151 __mptcp_clean_una_wakeup(sk);
3152 if (unlikely(&msk->cb_flags)) {
3153 /* be sure to set the current sk state before tacking actions
3154 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3156 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3157 __mptcp_set_connected(sk);
3158 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3159 __mptcp_error_report(sk);
3160 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3161 msk->last_snd = NULL;
3164 __mptcp_update_rmem(sk);
3167 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3168 * TCP can't schedule delack timer before the subflow is fully established.
3169 * MPTCP uses the delack timer to do 3rd ack retransmissions
3171 static void schedule_3rdack_retransmission(struct sock *ssk)
3173 struct inet_connection_sock *icsk = inet_csk(ssk);
3174 struct tcp_sock *tp = tcp_sk(ssk);
3175 unsigned long timeout;
3177 if (mptcp_subflow_ctx(ssk)->fully_established)
3180 /* reschedule with a timeout above RTT, as we must look only for drop */
3182 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3184 timeout = TCP_TIMEOUT_INIT;
3187 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3188 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3189 icsk->icsk_ack.timeout = timeout;
3190 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3193 void mptcp_subflow_process_delegated(struct sock *ssk)
3195 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3196 struct sock *sk = subflow->conn;
3198 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3199 mptcp_data_lock(sk);
3200 if (!sock_owned_by_user(sk))
3201 __mptcp_subflow_push_pending(sk, ssk);
3203 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3204 mptcp_data_unlock(sk);
3205 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3207 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3208 schedule_3rdack_retransmission(ssk);
3209 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3213 static int mptcp_hash(struct sock *sk)
3215 /* should never be called,
3216 * we hash the TCP subflows not the master socket
3222 static void mptcp_unhash(struct sock *sk)
3224 /* called from sk_common_release(), but nothing to do here */
3227 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3229 struct mptcp_sock *msk = mptcp_sk(sk);
3230 struct socket *ssock;
3232 ssock = __mptcp_nmpc_socket(msk);
3233 pr_debug("msk=%p, subflow=%p", msk, ssock);
3234 if (WARN_ON_ONCE(!ssock))
3237 return inet_csk_get_port(ssock->sk, snum);
3240 void mptcp_finish_connect(struct sock *ssk)
3242 struct mptcp_subflow_context *subflow;
3243 struct mptcp_sock *msk;
3247 subflow = mptcp_subflow_ctx(ssk);
3251 pr_debug("msk=%p, token=%u", sk, subflow->token);
3253 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3255 subflow->map_seq = ack_seq;
3256 subflow->map_subflow_seq = 1;
3258 /* the socket is not connected yet, no msk/subflow ops can access/race
3259 * accessing the field below
3261 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3262 WRITE_ONCE(msk->local_key, subflow->local_key);
3263 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3264 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3265 WRITE_ONCE(msk->ack_seq, ack_seq);
3266 WRITE_ONCE(msk->can_ack, 1);
3267 WRITE_ONCE(msk->snd_una, msk->write_seq);
3268 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3270 mptcp_pm_new_connection(msk, ssk, 0);
3272 mptcp_rcv_space_init(msk, ssk);
3275 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3277 write_lock_bh(&sk->sk_callback_lock);
3278 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3279 sk_set_socket(sk, parent);
3280 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3281 write_unlock_bh(&sk->sk_callback_lock);
3284 bool mptcp_finish_join(struct sock *ssk)
3286 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3287 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3288 struct sock *parent = (void *)msk;
3291 pr_debug("msk=%p, subflow=%p", msk, subflow);
3293 /* mptcp socket already closing? */
3294 if (!mptcp_is_fully_established(parent)) {
3295 subflow->reset_reason = MPTCP_RST_EMPTCP;
3299 if (!list_empty(&subflow->node))
3302 if (!mptcp_pm_allow_new_subflow(msk))
3303 goto err_prohibited;
3305 /* active connections are already on conn_list.
3306 * If we can't acquire msk socket lock here, let the release callback
3309 mptcp_data_lock(parent);
3310 if (!sock_owned_by_user(parent)) {
3311 ret = __mptcp_finish_join(msk, ssk);
3314 list_add_tail(&subflow->node, &msk->conn_list);
3318 list_add_tail(&subflow->node, &msk->join_list);
3319 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3321 mptcp_data_unlock(parent);
3325 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3329 subflow->map_seq = READ_ONCE(msk->ack_seq);
3330 WRITE_ONCE(msk->allow_infinite_fallback, false);
3333 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3337 static void mptcp_shutdown(struct sock *sk, int how)
3339 pr_debug("sk=%p, how=%d", sk, how);
3341 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3342 __mptcp_wr_shutdown(sk);
3345 static int mptcp_forward_alloc_get(const struct sock *sk)
3347 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3350 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3352 const struct sock *sk = (void *)msk;
3355 if (sk->sk_state == TCP_LISTEN)
3358 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3361 delta = msk->write_seq - v;
3362 if (__mptcp_check_fallback(msk) && msk->first) {
3363 struct tcp_sock *tp = tcp_sk(msk->first);
3365 /* the first subflow is disconnected after close - see
3366 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3367 * so ignore that status, too.
3369 if (!((1 << msk->first->sk_state) &
3370 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3371 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3373 if (delta > INT_MAX)
3379 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3381 struct mptcp_sock *msk = mptcp_sk(sk);
3387 if (sk->sk_state == TCP_LISTEN)
3391 __mptcp_move_skbs(msk);
3392 answ = mptcp_inq_hint(sk);
3396 slow = lock_sock_fast(sk);
3397 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3398 unlock_sock_fast(sk, slow);
3401 slow = lock_sock_fast(sk);
3402 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3403 unlock_sock_fast(sk, slow);
3406 return -ENOIOCTLCMD;
3409 return put_user(answ, (int __user *)arg);
3412 static struct proto mptcp_prot = {
3414 .owner = THIS_MODULE,
3415 .init = mptcp_init_sock,
3416 .disconnect = mptcp_disconnect,
3417 .close = mptcp_close,
3418 .accept = mptcp_accept,
3419 .setsockopt = mptcp_setsockopt,
3420 .getsockopt = mptcp_getsockopt,
3421 .shutdown = mptcp_shutdown,
3422 .destroy = mptcp_destroy,
3423 .sendmsg = mptcp_sendmsg,
3424 .ioctl = mptcp_ioctl,
3425 .recvmsg = mptcp_recvmsg,
3426 .release_cb = mptcp_release_cb,
3428 .unhash = mptcp_unhash,
3429 .get_port = mptcp_get_port,
3430 .forward_alloc_get = mptcp_forward_alloc_get,
3431 .sockets_allocated = &mptcp_sockets_allocated,
3433 .memory_allocated = &tcp_memory_allocated,
3434 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3436 .memory_pressure = &tcp_memory_pressure,
3437 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3438 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3439 .sysctl_mem = sysctl_tcp_mem,
3440 .obj_size = sizeof(struct mptcp_sock),
3441 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3442 .no_autobind = true,
3445 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3447 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3448 struct socket *ssock;
3451 lock_sock(sock->sk);
3452 ssock = __mptcp_nmpc_socket(msk);
3458 err = ssock->ops->bind(ssock, uaddr, addr_len);
3460 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3463 release_sock(sock->sk);
3467 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3468 struct mptcp_subflow_context *subflow)
3470 subflow->request_mptcp = 0;
3471 __mptcp_do_fallback(msk);
3474 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3475 int addr_len, int flags)
3477 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3478 struct mptcp_subflow_context *subflow;
3479 struct socket *ssock;
3482 lock_sock(sock->sk);
3484 if (addr_len < sizeof(uaddr->sa_family))
3487 if (uaddr->sa_family == AF_UNSPEC) {
3488 err = mptcp_disconnect(sock->sk, flags);
3489 sock->state = err ? SS_DISCONNECTING : SS_UNCONNECTED;
3494 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3495 /* pending connection or invalid state, let existing subflow
3498 ssock = msk->subflow;
3502 ssock = __mptcp_nmpc_socket(msk);
3506 mptcp_token_destroy(msk);
3507 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3508 subflow = mptcp_subflow_ctx(ssock->sk);
3509 #ifdef CONFIG_TCP_MD5SIG
3510 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3513 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3514 mptcp_subflow_early_fallback(msk, subflow);
3516 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3517 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3518 mptcp_subflow_early_fallback(msk, subflow);
3520 if (likely(!__mptcp_check_fallback(msk)))
3521 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3524 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3525 sock->state = ssock->state;
3527 /* on successful connect, the msk state will be moved to established by
3528 * subflow_finish_connect()
3530 if (!err || err == -EINPROGRESS)
3531 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3533 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3536 release_sock(sock->sk);
3540 static int mptcp_listen(struct socket *sock, int backlog)
3542 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3543 struct socket *ssock;
3546 pr_debug("msk=%p", msk);
3548 lock_sock(sock->sk);
3549 ssock = __mptcp_nmpc_socket(msk);
3555 mptcp_token_destroy(msk);
3556 inet_sk_state_store(sock->sk, TCP_LISTEN);
3557 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3559 err = ssock->ops->listen(ssock, backlog);
3560 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3562 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3565 release_sock(sock->sk);
3569 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3570 int flags, bool kern)
3572 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3573 struct socket *ssock;
3576 pr_debug("msk=%p", msk);
3578 ssock = __mptcp_nmpc_socket(msk);
3582 err = ssock->ops->accept(sock, newsock, flags, kern);
3583 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3584 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3585 struct mptcp_subflow_context *subflow;
3586 struct sock *newsk = newsock->sk;
3590 /* PM/worker can now acquire the first subflow socket
3591 * lock without racing with listener queue cleanup,
3592 * we can notify it, if needed.
3594 * Even if remote has reset the initial subflow by now
3595 * the refcnt is still at least one.
3597 subflow = mptcp_subflow_ctx(msk->first);
3598 list_add(&subflow->node, &msk->conn_list);
3599 sock_hold(msk->first);
3600 if (mptcp_is_fully_established(newsk))
3601 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3603 mptcp_copy_inaddrs(newsk, msk->first);
3604 mptcp_rcv_space_init(msk, msk->first);
3605 mptcp_propagate_sndbuf(newsk, msk->first);
3607 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3608 * This is needed so NOSPACE flag can be set from tcp stack.
3610 mptcp_for_each_subflow(msk, subflow) {
3611 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3613 if (!ssk->sk_socket)
3614 mptcp_sock_graft(ssk, newsock);
3616 release_sock(newsk);
3622 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3624 /* Concurrent splices from sk_receive_queue into receive_queue will
3625 * always show at least one non-empty queue when checked in this order.
3627 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
3628 skb_queue_empty_lockless(&msk->receive_queue))
3631 return EPOLLIN | EPOLLRDNORM;
3634 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3636 struct sock *sk = (struct sock *)msk;
3638 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3639 return EPOLLOUT | EPOLLWRNORM;
3641 if (sk_stream_is_writeable(sk))
3642 return EPOLLOUT | EPOLLWRNORM;
3644 mptcp_set_nospace(sk);
3645 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3646 if (sk_stream_is_writeable(sk))
3647 return EPOLLOUT | EPOLLWRNORM;
3652 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3653 struct poll_table_struct *wait)
3655 struct sock *sk = sock->sk;
3656 struct mptcp_sock *msk;
3661 sock_poll_wait(file, sock, wait);
3663 state = inet_sk_state_load(sk);
3664 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3665 if (state == TCP_LISTEN) {
3666 if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk))
3669 return inet_csk_listen_poll(msk->subflow->sk);
3672 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3673 mask |= mptcp_check_readable(msk);
3674 mask |= mptcp_check_writeable(msk);
3676 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3678 if (sk->sk_shutdown & RCV_SHUTDOWN)
3679 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3681 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3689 static const struct proto_ops mptcp_stream_ops = {
3691 .owner = THIS_MODULE,
3692 .release = inet_release,
3694 .connect = mptcp_stream_connect,
3695 .socketpair = sock_no_socketpair,
3696 .accept = mptcp_stream_accept,
3697 .getname = inet_getname,
3699 .ioctl = inet_ioctl,
3700 .gettstamp = sock_gettstamp,
3701 .listen = mptcp_listen,
3702 .shutdown = inet_shutdown,
3703 .setsockopt = sock_common_setsockopt,
3704 .getsockopt = sock_common_getsockopt,
3705 .sendmsg = inet_sendmsg,
3706 .recvmsg = inet_recvmsg,
3707 .mmap = sock_no_mmap,
3708 .sendpage = inet_sendpage,
3711 static struct inet_protosw mptcp_protosw = {
3712 .type = SOCK_STREAM,
3713 .protocol = IPPROTO_MPTCP,
3714 .prot = &mptcp_prot,
3715 .ops = &mptcp_stream_ops,
3716 .flags = INET_PROTOSW_ICSK,
3719 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3721 struct mptcp_delegated_action *delegated;
3722 struct mptcp_subflow_context *subflow;
3725 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3726 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3727 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3729 bh_lock_sock_nested(ssk);
3730 if (!sock_owned_by_user(ssk) &&
3731 mptcp_subflow_has_delegated_action(subflow))
3732 mptcp_subflow_process_delegated(ssk);
3733 /* ... elsewhere tcp_release_cb_override already processed
3734 * the action or will do at next release_sock().
3735 * In both case must dequeue the subflow here - on the same
3736 * CPU that scheduled it.
3738 bh_unlock_sock(ssk);
3741 if (++work_done == budget)
3745 /* always provide a 0 'work_done' argument, so that napi_complete_done
3746 * will not try accessing the NULL napi->dev ptr
3748 napi_complete_done(napi, 0);
3752 void __init mptcp_proto_init(void)
3754 struct mptcp_delegated_action *delegated;
3757 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3759 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3760 panic("Failed to allocate MPTCP pcpu counter\n");
3762 init_dummy_netdev(&mptcp_napi_dev);
3763 for_each_possible_cpu(cpu) {
3764 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3765 INIT_LIST_HEAD(&delegated->head);
3766 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3768 napi_enable(&delegated->napi);
3771 mptcp_subflow_init();
3775 if (proto_register(&mptcp_prot, 1) != 0)
3776 panic("Failed to register MPTCP proto.\n");
3778 inet_register_protosw(&mptcp_protosw);
3780 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3783 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3784 static const struct proto_ops mptcp_v6_stream_ops = {
3786 .owner = THIS_MODULE,
3787 .release = inet6_release,
3789 .connect = mptcp_stream_connect,
3790 .socketpair = sock_no_socketpair,
3791 .accept = mptcp_stream_accept,
3792 .getname = inet6_getname,
3794 .ioctl = inet6_ioctl,
3795 .gettstamp = sock_gettstamp,
3796 .listen = mptcp_listen,
3797 .shutdown = inet_shutdown,
3798 .setsockopt = sock_common_setsockopt,
3799 .getsockopt = sock_common_getsockopt,
3800 .sendmsg = inet6_sendmsg,
3801 .recvmsg = inet6_recvmsg,
3802 .mmap = sock_no_mmap,
3803 .sendpage = inet_sendpage,
3804 #ifdef CONFIG_COMPAT
3805 .compat_ioctl = inet6_compat_ioctl,
3809 static struct proto mptcp_v6_prot;
3811 static void mptcp_v6_destroy(struct sock *sk)
3814 inet6_destroy_sock(sk);
3817 static struct inet_protosw mptcp_v6_protosw = {
3818 .type = SOCK_STREAM,
3819 .protocol = IPPROTO_MPTCP,
3820 .prot = &mptcp_v6_prot,
3821 .ops = &mptcp_v6_stream_ops,
3822 .flags = INET_PROTOSW_ICSK,
3825 int __init mptcp_proto_v6_init(void)
3829 mptcp_v6_prot = mptcp_prot;
3830 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3831 mptcp_v6_prot.slab = NULL;
3832 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3833 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3835 err = proto_register(&mptcp_v6_prot, 1);
3839 err = inet6_register_protosw(&mptcp_v6_protosw);
3841 proto_unregister(&mptcp_v6_prot);