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;
40 MPTCP_CMSG_TS = BIT(0),
41 MPTCP_CMSG_INQ = BIT(1),
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void __mptcp_check_send_data_fin(struct sock *sk);
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
52 /* Returns end sequence number of the receiver's advertised window */
53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
55 return READ_ONCE(msk->wnd_end);
58 static bool mptcp_is_tcpsk(struct sock *sk)
60 struct socket *sock = sk->sk_socket;
62 if (unlikely(sk->sk_prot == &tcp_prot)) {
63 /* we are being invoked after mptcp_accept() has
64 * accepted a non-mp-capable flow: sk is a tcp_sk,
67 * Hand the socket over to tcp so all further socket ops
70 sock->ops = &inet_stream_ops;
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 sock->ops = &inet6_stream_ops;
82 static int __mptcp_socket_create(struct mptcp_sock *msk)
84 struct mptcp_subflow_context *subflow;
85 struct sock *sk = (struct sock *)msk;
89 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
93 WRITE_ONCE(msk->first, ssock->sk);
94 WRITE_ONCE(msk->subflow, ssock);
95 subflow = mptcp_subflow_ctx(ssock->sk);
96 list_add(&subflow->node, &msk->conn_list);
98 subflow->request_mptcp = 1;
99 subflow->subflow_id = msk->subflow_id++;
101 /* This is the first subflow, always with id 0 */
102 subflow->local_id_valid = 1;
103 mptcp_sock_graft(msk->first, sk->sk_socket);
108 /* If the MPC handshake is not started, returns the first subflow,
109 * eventually allocating it.
111 struct socket *__mptcp_nmpc_socket(struct mptcp_sock *msk)
113 struct sock *sk = (struct sock *)msk;
116 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
117 return ERR_PTR(-EINVAL);
121 return ERR_PTR(-EINVAL);
123 ret = __mptcp_socket_create(msk);
127 mptcp_sockopt_sync(msk, msk->first);
133 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
135 sk_drops_add(sk, skb);
139 static void mptcp_rmem_charge(struct sock *sk, int size)
141 mptcp_sk(sk)->rmem_fwd_alloc -= size;
144 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
145 struct sk_buff *from)
150 if (MPTCP_SKB_CB(from)->offset ||
151 !skb_try_coalesce(to, from, &fragstolen, &delta))
154 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
155 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
156 to->len, MPTCP_SKB_CB(from)->end_seq);
157 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
159 /* note the fwd memory can reach a negative value after accounting
160 * for the delta, but the later skb free will restore a non
163 atomic_add(delta, &sk->sk_rmem_alloc);
164 mptcp_rmem_charge(sk, delta);
165 kfree_skb_partial(from, fragstolen);
170 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
171 struct sk_buff *from)
173 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
176 return mptcp_try_coalesce((struct sock *)msk, to, from);
179 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
181 amount >>= PAGE_SHIFT;
182 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
183 __sk_mem_reduce_allocated(sk, amount);
186 static void mptcp_rmem_uncharge(struct sock *sk, int size)
188 struct mptcp_sock *msk = mptcp_sk(sk);
191 msk->rmem_fwd_alloc += size;
192 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
194 /* see sk_mem_uncharge() for the rationale behind the following schema */
195 if (unlikely(reclaimable >= PAGE_SIZE))
196 __mptcp_rmem_reclaim(sk, reclaimable);
199 static void mptcp_rfree(struct sk_buff *skb)
201 unsigned int len = skb->truesize;
202 struct sock *sk = skb->sk;
204 atomic_sub(len, &sk->sk_rmem_alloc);
205 mptcp_rmem_uncharge(sk, len);
208 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
212 skb->destructor = mptcp_rfree;
213 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
214 mptcp_rmem_charge(sk, skb->truesize);
217 /* "inspired" by tcp_data_queue_ofo(), main differences:
219 * - don't cope with sacks
221 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
223 struct sock *sk = (struct sock *)msk;
224 struct rb_node **p, *parent;
225 u64 seq, end_seq, max_seq;
226 struct sk_buff *skb1;
228 seq = MPTCP_SKB_CB(skb)->map_seq;
229 end_seq = MPTCP_SKB_CB(skb)->end_seq;
230 max_seq = atomic64_read(&msk->rcv_wnd_sent);
232 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
233 RB_EMPTY_ROOT(&msk->out_of_order_queue));
234 if (after64(end_seq, max_seq)) {
237 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
238 (unsigned long long)end_seq - (unsigned long)max_seq,
239 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
244 p = &msk->out_of_order_queue.rb_node;
245 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
246 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
247 rb_link_node(&skb->rbnode, NULL, p);
248 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
249 msk->ooo_last_skb = skb;
253 /* with 2 subflows, adding at end of ooo queue is quite likely
254 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
256 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
257 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
258 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
262 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
263 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
264 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
265 parent = &msk->ooo_last_skb->rbnode;
266 p = &parent->rb_right;
270 /* Find place to insert this segment. Handle overlaps on the way. */
274 skb1 = rb_to_skb(parent);
275 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
276 p = &parent->rb_left;
279 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
280 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
281 /* All the bits are present. Drop. */
283 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
286 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
290 * continue traversing
293 /* skb's seq == skb1's seq and skb covers skb1.
294 * Replace skb1 with skb.
296 rb_replace_node(&skb1->rbnode, &skb->rbnode,
297 &msk->out_of_order_queue);
298 mptcp_drop(sk, skb1);
299 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
302 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
303 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
306 p = &parent->rb_right;
310 /* Insert segment into RB tree. */
311 rb_link_node(&skb->rbnode, parent, p);
312 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
315 /* Remove other segments covered by skb. */
316 while ((skb1 = skb_rb_next(skb)) != NULL) {
317 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
319 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
320 mptcp_drop(sk, skb1);
321 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
323 /* If there is no skb after us, we are the last_skb ! */
325 msk->ooo_last_skb = skb;
329 mptcp_set_owner_r(skb, sk);
332 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
334 struct mptcp_sock *msk = mptcp_sk(sk);
337 if (size <= msk->rmem_fwd_alloc)
340 size -= msk->rmem_fwd_alloc;
341 amt = sk_mem_pages(size);
342 amount = amt << PAGE_SHIFT;
343 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
346 msk->rmem_fwd_alloc += amount;
350 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
351 struct sk_buff *skb, unsigned int offset,
354 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
355 struct sock *sk = (struct sock *)msk;
356 struct sk_buff *tail;
359 __skb_unlink(skb, &ssk->sk_receive_queue);
364 /* try to fetch required memory from subflow */
365 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
368 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
370 /* the skb map_seq accounts for the skb offset:
371 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
374 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
375 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
376 MPTCP_SKB_CB(skb)->offset = offset;
377 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
379 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
381 msk->bytes_received += copy_len;
382 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
383 tail = skb_peek_tail(&sk->sk_receive_queue);
384 if (tail && mptcp_try_coalesce(sk, tail, skb))
387 mptcp_set_owner_r(skb, sk);
388 __skb_queue_tail(&sk->sk_receive_queue, skb);
390 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
391 mptcp_data_queue_ofo(msk, skb);
395 /* old data, keep it simple and drop the whole pkt, sender
396 * will retransmit as needed, if needed.
398 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
404 static void mptcp_stop_timer(struct sock *sk)
406 struct inet_connection_sock *icsk = inet_csk(sk);
408 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
409 mptcp_sk(sk)->timer_ival = 0;
412 static void mptcp_close_wake_up(struct sock *sk)
414 if (sock_flag(sk, SOCK_DEAD))
417 sk->sk_state_change(sk);
418 if (sk->sk_shutdown == SHUTDOWN_MASK ||
419 sk->sk_state == TCP_CLOSE)
420 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
422 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
425 static bool mptcp_pending_data_fin_ack(struct sock *sk)
427 struct mptcp_sock *msk = mptcp_sk(sk);
429 return !__mptcp_check_fallback(msk) &&
430 ((1 << sk->sk_state) &
431 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
432 msk->write_seq == READ_ONCE(msk->snd_una);
435 static void mptcp_check_data_fin_ack(struct sock *sk)
437 struct mptcp_sock *msk = mptcp_sk(sk);
439 /* Look for an acknowledged DATA_FIN */
440 if (mptcp_pending_data_fin_ack(sk)) {
441 WRITE_ONCE(msk->snd_data_fin_enable, 0);
443 switch (sk->sk_state) {
445 inet_sk_state_store(sk, TCP_FIN_WAIT2);
449 inet_sk_state_store(sk, TCP_CLOSE);
453 mptcp_close_wake_up(sk);
457 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
459 struct mptcp_sock *msk = mptcp_sk(sk);
461 if (READ_ONCE(msk->rcv_data_fin) &&
462 ((1 << sk->sk_state) &
463 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
464 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
466 if (msk->ack_seq == rcv_data_fin_seq) {
468 *seq = rcv_data_fin_seq;
477 static void mptcp_set_datafin_timeout(struct sock *sk)
479 struct inet_connection_sock *icsk = inet_csk(sk);
482 retransmits = min_t(u32, icsk->icsk_retransmits,
483 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
485 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
488 static void __mptcp_set_timeout(struct sock *sk, long tout)
490 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
493 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
495 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
497 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
498 inet_csk(ssk)->icsk_timeout - jiffies : 0;
501 static void mptcp_set_timeout(struct sock *sk)
503 struct mptcp_subflow_context *subflow;
506 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
507 tout = max(tout, mptcp_timeout_from_subflow(subflow));
508 __mptcp_set_timeout(sk, tout);
511 static inline bool tcp_can_send_ack(const struct sock *ssk)
513 return !((1 << inet_sk_state_load(ssk)) &
514 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
517 void __mptcp_subflow_send_ack(struct sock *ssk)
519 if (tcp_can_send_ack(ssk))
523 static void mptcp_subflow_send_ack(struct sock *ssk)
527 slow = lock_sock_fast(ssk);
528 __mptcp_subflow_send_ack(ssk);
529 unlock_sock_fast(ssk, slow);
532 static void mptcp_send_ack(struct mptcp_sock *msk)
534 struct mptcp_subflow_context *subflow;
536 mptcp_for_each_subflow(msk, subflow)
537 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
540 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
544 slow = lock_sock_fast(ssk);
545 if (tcp_can_send_ack(ssk))
546 tcp_cleanup_rbuf(ssk, 1);
547 unlock_sock_fast(ssk, slow);
550 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
552 const struct inet_connection_sock *icsk = inet_csk(ssk);
553 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
554 const struct tcp_sock *tp = tcp_sk(ssk);
556 return (ack_pending & ICSK_ACK_SCHED) &&
557 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
558 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
559 (rx_empty && ack_pending &
560 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
563 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
565 int old_space = READ_ONCE(msk->old_wspace);
566 struct mptcp_subflow_context *subflow;
567 struct sock *sk = (struct sock *)msk;
568 int space = __mptcp_space(sk);
569 bool cleanup, rx_empty;
571 cleanup = (space > 0) && (space >= (old_space << 1));
572 rx_empty = !__mptcp_rmem(sk);
574 mptcp_for_each_subflow(msk, subflow) {
575 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
577 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
578 mptcp_subflow_cleanup_rbuf(ssk);
582 static bool mptcp_check_data_fin(struct sock *sk)
584 struct mptcp_sock *msk = mptcp_sk(sk);
585 u64 rcv_data_fin_seq;
588 if (__mptcp_check_fallback(msk))
591 /* Need to ack a DATA_FIN received from a peer while this side
592 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
593 * msk->rcv_data_fin was set when parsing the incoming options
594 * at the subflow level and the msk lock was not held, so this
595 * is the first opportunity to act on the DATA_FIN and change
598 * If we are caught up to the sequence number of the incoming
599 * DATA_FIN, send the DATA_ACK now and do state transition. If
600 * not caught up, do nothing and let the recv code send DATA_ACK
604 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
605 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
606 WRITE_ONCE(msk->rcv_data_fin, 0);
608 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
609 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
611 switch (sk->sk_state) {
612 case TCP_ESTABLISHED:
613 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
616 inet_sk_state_store(sk, TCP_CLOSING);
619 inet_sk_state_store(sk, TCP_CLOSE);
622 /* Other states not expected */
629 mptcp_close_wake_up(sk);
634 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
638 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
639 struct sock *sk = (struct sock *)msk;
640 unsigned int moved = 0;
641 bool more_data_avail;
646 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
648 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
649 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
651 if (unlikely(ssk_rbuf > sk_rbuf)) {
652 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
657 pr_debug("msk=%p ssk=%p", msk, ssk);
660 u32 map_remaining, offset;
661 u32 seq = tp->copied_seq;
665 /* try to move as much data as available */
666 map_remaining = subflow->map_data_len -
667 mptcp_subflow_get_map_offset(subflow);
669 skb = skb_peek(&ssk->sk_receive_queue);
671 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
672 * a different CPU can have already processed the pending
673 * data, stop here or we can enter an infinite loop
680 if (__mptcp_check_fallback(msk)) {
681 /* Under fallback skbs have no MPTCP extension and TCP could
682 * collapse them between the dummy map creation and the
683 * current dequeue. Be sure to adjust the map size.
685 map_remaining = skb->len;
686 subflow->map_data_len = skb->len;
689 offset = seq - TCP_SKB_CB(skb)->seq;
690 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
696 if (offset < skb->len) {
697 size_t len = skb->len - offset;
702 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
706 if (WARN_ON_ONCE(map_remaining < len))
710 sk_eat_skb(ssk, skb);
714 WRITE_ONCE(tp->copied_seq, seq);
715 more_data_avail = mptcp_subflow_data_available(ssk);
717 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
721 } while (more_data_avail);
727 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
729 struct sock *sk = (struct sock *)msk;
730 struct sk_buff *skb, *tail;
735 p = rb_first(&msk->out_of_order_queue);
736 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
739 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
743 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
745 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
748 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
752 end_seq = MPTCP_SKB_CB(skb)->end_seq;
753 tail = skb_peek_tail(&sk->sk_receive_queue);
754 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
755 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
757 /* skip overlapping data, if any */
758 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
759 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
761 MPTCP_SKB_CB(skb)->offset += delta;
762 MPTCP_SKB_CB(skb)->map_seq += delta;
763 __skb_queue_tail(&sk->sk_receive_queue, skb);
765 msk->bytes_received += end_seq - msk->ack_seq;
766 msk->ack_seq = end_seq;
772 /* In most cases we will be able to lock the mptcp socket. If its already
773 * owned, we need to defer to the work queue to avoid ABBA deadlock.
775 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
777 struct sock *sk = (struct sock *)msk;
778 unsigned int moved = 0;
780 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
781 __mptcp_ofo_queue(msk);
782 if (unlikely(ssk->sk_err)) {
783 if (!sock_owned_by_user(sk))
784 __mptcp_error_report(sk);
786 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
789 /* If the moves have caught up with the DATA_FIN sequence number
790 * it's time to ack the DATA_FIN and change socket state, but
791 * this is not a good place to change state. Let the workqueue
794 if (mptcp_pending_data_fin(sk, NULL))
795 mptcp_schedule_work(sk);
799 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
801 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
802 struct mptcp_sock *msk = mptcp_sk(sk);
803 int sk_rbuf, ssk_rbuf;
805 /* The peer can send data while we are shutting down this
806 * subflow at msk destruction time, but we must avoid enqueuing
807 * more data to the msk receive queue
809 if (unlikely(subflow->disposable))
812 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
813 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
814 if (unlikely(ssk_rbuf > sk_rbuf))
817 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
818 if (__mptcp_rmem(sk) > sk_rbuf) {
819 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
823 /* Wake-up the reader only for in-sequence data */
825 if (move_skbs_to_msk(msk, ssk))
826 sk->sk_data_ready(sk);
828 mptcp_data_unlock(sk);
831 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
833 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
834 WRITE_ONCE(msk->allow_infinite_fallback, false);
835 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
838 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
840 struct sock *sk = (struct sock *)msk;
842 if (sk->sk_state != TCP_ESTABLISHED)
845 /* attach to msk socket only after we are sure we will deal with it
848 if (sk->sk_socket && !ssk->sk_socket)
849 mptcp_sock_graft(ssk, sk->sk_socket);
851 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
852 mptcp_sockopt_sync_locked(msk, ssk);
853 mptcp_subflow_joined(msk, ssk);
857 static void __mptcp_flush_join_list(struct sock *sk)
859 struct mptcp_subflow_context *tmp, *subflow;
860 struct mptcp_sock *msk = mptcp_sk(sk);
862 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
863 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
864 bool slow = lock_sock_fast(ssk);
866 list_move_tail(&subflow->node, &msk->conn_list);
867 if (!__mptcp_finish_join(msk, ssk))
868 mptcp_subflow_reset(ssk);
869 unlock_sock_fast(ssk, slow);
873 static bool mptcp_timer_pending(struct sock *sk)
875 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
878 static void mptcp_reset_timer(struct sock *sk)
880 struct inet_connection_sock *icsk = inet_csk(sk);
883 /* prevent rescheduling on close */
884 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
887 tout = mptcp_sk(sk)->timer_ival;
888 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
891 bool mptcp_schedule_work(struct sock *sk)
893 if (inet_sk_state_load(sk) != TCP_CLOSE &&
894 schedule_work(&mptcp_sk(sk)->work)) {
895 /* each subflow already holds a reference to the sk, and the
896 * workqueue is invoked by a subflow, so sk can't go away here.
904 void mptcp_subflow_eof(struct sock *sk)
906 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
907 mptcp_schedule_work(sk);
910 static void mptcp_check_for_eof(struct mptcp_sock *msk)
912 struct mptcp_subflow_context *subflow;
913 struct sock *sk = (struct sock *)msk;
916 mptcp_for_each_subflow(msk, subflow)
917 receivers += !subflow->rx_eof;
921 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
922 /* hopefully temporary hack: propagate shutdown status
923 * to msk, when all subflows agree on it
925 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
927 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
928 sk->sk_data_ready(sk);
931 switch (sk->sk_state) {
932 case TCP_ESTABLISHED:
933 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
936 inet_sk_state_store(sk, TCP_CLOSING);
939 inet_sk_state_store(sk, TCP_CLOSE);
944 mptcp_close_wake_up(sk);
947 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
949 struct mptcp_subflow_context *subflow;
951 msk_owned_by_me(msk);
953 mptcp_for_each_subflow(msk, subflow) {
954 if (READ_ONCE(subflow->data_avail))
955 return mptcp_subflow_tcp_sock(subflow);
961 static bool mptcp_skb_can_collapse_to(u64 write_seq,
962 const struct sk_buff *skb,
963 const struct mptcp_ext *mpext)
965 if (!tcp_skb_can_collapse_to(skb))
968 /* can collapse only if MPTCP level sequence is in order and this
969 * mapping has not been xmitted yet
971 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
975 /* we can append data to the given data frag if:
976 * - there is space available in the backing page_frag
977 * - the data frag tail matches the current page_frag free offset
978 * - the data frag end sequence number matches the current write seq
980 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
981 const struct page_frag *pfrag,
982 const struct mptcp_data_frag *df)
984 return df && pfrag->page == df->page &&
985 pfrag->size - pfrag->offset > 0 &&
986 pfrag->offset == (df->offset + df->data_len) &&
987 df->data_seq + df->data_len == msk->write_seq;
990 static void dfrag_uncharge(struct sock *sk, int len)
992 sk_mem_uncharge(sk, len);
993 sk_wmem_queued_add(sk, -len);
996 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
998 int len = dfrag->data_len + dfrag->overhead;
1000 list_del(&dfrag->list);
1001 dfrag_uncharge(sk, len);
1002 put_page(dfrag->page);
1005 static void __mptcp_clean_una(struct sock *sk)
1007 struct mptcp_sock *msk = mptcp_sk(sk);
1008 struct mptcp_data_frag *dtmp, *dfrag;
1011 snd_una = msk->snd_una;
1012 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1013 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1016 if (unlikely(dfrag == msk->first_pending)) {
1017 /* in recovery mode can see ack after the current snd head */
1018 if (WARN_ON_ONCE(!msk->recovery))
1021 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1024 dfrag_clear(sk, dfrag);
1027 dfrag = mptcp_rtx_head(sk);
1028 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1029 u64 delta = snd_una - dfrag->data_seq;
1031 /* prevent wrap around in recovery mode */
1032 if (unlikely(delta > dfrag->already_sent)) {
1033 if (WARN_ON_ONCE(!msk->recovery))
1035 if (WARN_ON_ONCE(delta > dfrag->data_len))
1037 dfrag->already_sent += delta - dfrag->already_sent;
1040 dfrag->data_seq += delta;
1041 dfrag->offset += delta;
1042 dfrag->data_len -= delta;
1043 dfrag->already_sent -= delta;
1045 dfrag_uncharge(sk, delta);
1048 /* all retransmitted data acked, recovery completed */
1049 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1050 msk->recovery = false;
1053 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1054 snd_una == READ_ONCE(msk->write_seq)) {
1055 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1056 mptcp_stop_timer(sk);
1058 mptcp_reset_timer(sk);
1062 static void __mptcp_clean_una_wakeup(struct sock *sk)
1064 lockdep_assert_held_once(&sk->sk_lock.slock);
1066 __mptcp_clean_una(sk);
1067 mptcp_write_space(sk);
1070 static void mptcp_clean_una_wakeup(struct sock *sk)
1072 mptcp_data_lock(sk);
1073 __mptcp_clean_una_wakeup(sk);
1074 mptcp_data_unlock(sk);
1077 static void mptcp_enter_memory_pressure(struct sock *sk)
1079 struct mptcp_subflow_context *subflow;
1080 struct mptcp_sock *msk = mptcp_sk(sk);
1083 sk_stream_moderate_sndbuf(sk);
1084 mptcp_for_each_subflow(msk, subflow) {
1085 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1088 tcp_enter_memory_pressure(ssk);
1089 sk_stream_moderate_sndbuf(ssk);
1094 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1097 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1099 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1100 pfrag, sk->sk_allocation)))
1103 mptcp_enter_memory_pressure(sk);
1107 static struct mptcp_data_frag *
1108 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1111 int offset = ALIGN(orig_offset, sizeof(long));
1112 struct mptcp_data_frag *dfrag;
1114 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1115 dfrag->data_len = 0;
1116 dfrag->data_seq = msk->write_seq;
1117 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1118 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1119 dfrag->already_sent = 0;
1120 dfrag->page = pfrag->page;
1125 struct mptcp_sendmsg_info {
1131 bool data_lock_held;
1134 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1135 u64 data_seq, int avail_size)
1137 u64 window_end = mptcp_wnd_end(msk);
1140 if (__mptcp_check_fallback(msk))
1143 mptcp_snd_wnd = window_end - data_seq;
1144 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1146 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1147 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1148 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1154 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1156 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1160 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1164 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1166 struct sk_buff *skb;
1168 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1170 if (likely(__mptcp_add_ext(skb, gfp))) {
1171 skb_reserve(skb, MAX_TCP_HEADER);
1172 skb->ip_summed = CHECKSUM_PARTIAL;
1173 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1178 mptcp_enter_memory_pressure(sk);
1183 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1185 struct sk_buff *skb;
1187 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1191 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1192 tcp_skb_entail(ssk, skb);
1195 tcp_skb_tsorted_anchor_cleanup(skb);
1200 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1202 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1204 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1207 /* note: this always recompute the csum on the whole skb, even
1208 * if we just appended a single frag. More status info needed
1210 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1212 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1213 __wsum csum = ~csum_unfold(mpext->csum);
1214 int offset = skb->len - added;
1216 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1219 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1221 struct mptcp_ext *mpext)
1226 mpext->infinite_map = 1;
1227 mpext->data_len = 0;
1229 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1230 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1232 mptcp_do_fallback(ssk);
1235 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1236 struct mptcp_data_frag *dfrag,
1237 struct mptcp_sendmsg_info *info)
1239 u64 data_seq = dfrag->data_seq + info->sent;
1240 int offset = dfrag->offset + info->sent;
1241 struct mptcp_sock *msk = mptcp_sk(sk);
1242 bool zero_window_probe = false;
1243 struct mptcp_ext *mpext = NULL;
1244 bool can_coalesce = false;
1245 bool reuse_skb = true;
1246 struct sk_buff *skb;
1250 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1251 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1253 if (WARN_ON_ONCE(info->sent > info->limit ||
1254 info->limit > dfrag->data_len))
1257 if (unlikely(!__tcp_can_send(ssk)))
1260 /* compute send limit */
1261 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1262 copy = info->size_goal;
1264 skb = tcp_write_queue_tail(ssk);
1265 if (skb && copy > skb->len) {
1266 /* Limit the write to the size available in the
1267 * current skb, if any, so that we create at most a new skb.
1268 * Explicitly tells TCP internals to avoid collapsing on later
1269 * queue management operation, to avoid breaking the ext <->
1270 * SSN association set here
1272 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1273 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1274 TCP_SKB_CB(skb)->eor = 1;
1278 i = skb_shinfo(skb)->nr_frags;
1279 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1280 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1281 tcp_mark_push(tcp_sk(ssk), skb);
1288 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1292 i = skb_shinfo(skb)->nr_frags;
1294 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1297 /* Zero window and all data acked? Probe. */
1298 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1300 u64 snd_una = READ_ONCE(msk->snd_una);
1302 if (snd_una != msk->snd_nxt) {
1303 tcp_remove_empty_skb(ssk);
1307 zero_window_probe = true;
1308 data_seq = snd_una - 1;
1311 /* all mptcp-level data is acked, no skbs should be present into the
1314 WARN_ON_ONCE(reuse_skb);
1317 copy = min_t(size_t, copy, info->limit - info->sent);
1318 if (!sk_wmem_schedule(ssk, copy)) {
1319 tcp_remove_empty_skb(ssk);
1324 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1326 get_page(dfrag->page);
1327 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1331 skb->data_len += copy;
1332 skb->truesize += copy;
1333 sk_wmem_queued_add(ssk, copy);
1334 sk_mem_charge(ssk, copy);
1335 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1336 TCP_SKB_CB(skb)->end_seq += copy;
1337 tcp_skb_pcount_set(skb, 0);
1339 /* on skb reuse we just need to update the DSS len */
1341 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1342 mpext->data_len += copy;
1343 WARN_ON_ONCE(zero_window_probe);
1347 memset(mpext, 0, sizeof(*mpext));
1348 mpext->data_seq = data_seq;
1349 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1350 mpext->data_len = copy;
1354 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1355 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1358 if (zero_window_probe) {
1359 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1361 if (READ_ONCE(msk->csum_enabled))
1362 mptcp_update_data_checksum(skb, copy);
1363 tcp_push_pending_frames(ssk);
1367 if (READ_ONCE(msk->csum_enabled))
1368 mptcp_update_data_checksum(skb, copy);
1369 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1370 mptcp_update_infinite_map(msk, ssk, mpext);
1371 trace_mptcp_sendmsg_frag(mpext);
1372 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1376 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1377 sizeof(struct tcphdr) - \
1378 MAX_TCP_OPTION_SPACE - \
1379 sizeof(struct ipv6hdr) - \
1380 sizeof(struct frag_hdr))
1382 struct subflow_send_info {
1387 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1389 if (!subflow->stale)
1393 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1396 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1398 if (unlikely(subflow->stale)) {
1399 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1401 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1404 mptcp_subflow_set_active(subflow);
1406 return __mptcp_subflow_active(subflow);
1409 #define SSK_MODE_ACTIVE 0
1410 #define SSK_MODE_BACKUP 1
1411 #define SSK_MODE_MAX 2
1413 /* implement the mptcp packet scheduler;
1414 * returns the subflow that will transmit the next DSS
1415 * additionally updates the rtx timeout
1417 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1419 struct subflow_send_info send_info[SSK_MODE_MAX];
1420 struct mptcp_subflow_context *subflow;
1421 struct sock *sk = (struct sock *)msk;
1422 u32 pace, burst, wmem;
1423 int i, nr_active = 0;
1428 msk_owned_by_me(msk);
1430 if (__mptcp_check_fallback(msk)) {
1433 return __tcp_can_send(msk->first) &&
1434 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1437 /* re-use last subflow, if the burst allow that */
1438 if (msk->last_snd && msk->snd_burst > 0 &&
1439 sk_stream_memory_free(msk->last_snd) &&
1440 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1441 mptcp_set_timeout(sk);
1442 return msk->last_snd;
1445 /* pick the subflow with the lower wmem/wspace ratio */
1446 for (i = 0; i < SSK_MODE_MAX; ++i) {
1447 send_info[i].ssk = NULL;
1448 send_info[i].linger_time = -1;
1451 mptcp_for_each_subflow(msk, subflow) {
1452 trace_mptcp_subflow_get_send(subflow);
1453 ssk = mptcp_subflow_tcp_sock(subflow);
1454 if (!mptcp_subflow_active(subflow))
1457 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1458 nr_active += !subflow->backup;
1459 pace = subflow->avg_pacing_rate;
1460 if (unlikely(!pace)) {
1461 /* init pacing rate from socket */
1462 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1463 pace = subflow->avg_pacing_rate;
1468 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1469 if (linger_time < send_info[subflow->backup].linger_time) {
1470 send_info[subflow->backup].ssk = ssk;
1471 send_info[subflow->backup].linger_time = linger_time;
1474 __mptcp_set_timeout(sk, tout);
1476 /* pick the best backup if no other subflow is active */
1478 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1480 /* According to the blest algorithm, to avoid HoL blocking for the
1481 * faster flow, we need to:
1482 * - estimate the faster flow linger time
1483 * - use the above to estimate the amount of byte transferred
1484 * by the faster flow
1485 * - check that the amount of queued data is greter than the above,
1486 * otherwise do not use the picked, slower, subflow
1487 * We select the subflow with the shorter estimated time to flush
1488 * the queued mem, which basically ensure the above. We just need
1489 * to check that subflow has a non empty cwin.
1491 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1492 if (!ssk || !sk_stream_memory_free(ssk))
1495 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1496 wmem = READ_ONCE(ssk->sk_wmem_queued);
1498 msk->last_snd = NULL;
1502 subflow = mptcp_subflow_ctx(ssk);
1503 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1504 READ_ONCE(ssk->sk_pacing_rate) * burst,
1506 msk->last_snd = ssk;
1507 msk->snd_burst = burst;
1511 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1513 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1517 static void mptcp_update_post_push(struct mptcp_sock *msk,
1518 struct mptcp_data_frag *dfrag,
1521 u64 snd_nxt_new = dfrag->data_seq;
1523 dfrag->already_sent += sent;
1525 msk->snd_burst -= sent;
1527 snd_nxt_new += dfrag->already_sent;
1529 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1530 * is recovering after a failover. In that event, this re-sends
1533 * Thus compute snd_nxt_new candidate based on
1534 * the dfrag->data_seq that was sent and the data
1535 * that has been handed to the subflow for transmission
1536 * and skip update in case it was old dfrag.
1538 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1539 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1540 msk->snd_nxt = snd_nxt_new;
1544 void mptcp_check_and_set_pending(struct sock *sk)
1546 if (mptcp_send_head(sk))
1547 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1550 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1552 struct sock *prev_ssk = NULL, *ssk = NULL;
1553 struct mptcp_sock *msk = mptcp_sk(sk);
1554 struct mptcp_sendmsg_info info = {
1557 bool do_check_data_fin = false;
1558 struct mptcp_data_frag *dfrag;
1561 while ((dfrag = mptcp_send_head(sk))) {
1562 info.sent = dfrag->already_sent;
1563 info.limit = dfrag->data_len;
1564 len = dfrag->data_len - dfrag->already_sent;
1569 ssk = mptcp_subflow_get_send(msk);
1571 /* First check. If the ssk has changed since
1572 * the last round, release prev_ssk
1574 if (ssk != prev_ssk && prev_ssk)
1575 mptcp_push_release(prev_ssk, &info);
1579 /* Need to lock the new subflow only if different
1580 * from the previous one, otherwise we are still
1581 * helding the relevant lock
1583 if (ssk != prev_ssk)
1586 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1590 mptcp_push_release(ssk, &info);
1594 do_check_data_fin = true;
1598 mptcp_update_post_push(msk, dfrag, ret);
1600 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1603 /* at this point we held the socket lock for the last subflow we used */
1605 mptcp_push_release(ssk, &info);
1608 /* ensure the rtx timer is running */
1609 if (!mptcp_timer_pending(sk))
1610 mptcp_reset_timer(sk);
1611 if (do_check_data_fin)
1612 __mptcp_check_send_data_fin(sk);
1615 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1617 struct mptcp_sock *msk = mptcp_sk(sk);
1618 struct mptcp_sendmsg_info info = {
1619 .data_lock_held = true,
1621 struct mptcp_data_frag *dfrag;
1622 struct sock *xmit_ssk;
1623 int len, copied = 0;
1626 while ((dfrag = mptcp_send_head(sk))) {
1627 info.sent = dfrag->already_sent;
1628 info.limit = dfrag->data_len;
1629 len = dfrag->data_len - dfrag->already_sent;
1633 /* check for a different subflow usage only after
1634 * spooling the first chunk of data
1636 xmit_ssk = first ? ssk : mptcp_subflow_get_send(msk);
1639 if (xmit_ssk != ssk) {
1640 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1641 MPTCP_DELEGATE_SEND);
1645 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1654 mptcp_update_post_push(msk, dfrag, ret);
1656 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1660 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1661 * not going to flush it via release_sock()
1664 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1666 if (!mptcp_timer_pending(sk))
1667 mptcp_reset_timer(sk);
1669 if (msk->snd_data_fin_enable &&
1670 msk->snd_nxt + 1 == msk->write_seq)
1671 mptcp_schedule_work(sk);
1675 static void mptcp_set_nospace(struct sock *sk)
1677 /* enable autotune */
1678 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1680 /* will be cleared on avail space */
1681 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1684 static int mptcp_disconnect(struct sock *sk, int flags);
1686 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1687 size_t len, int *copied_syn)
1689 unsigned int saved_flags = msg->msg_flags;
1690 struct mptcp_sock *msk = mptcp_sk(sk);
1691 struct socket *ssock;
1695 /* on flags based fastopen the mptcp is supposed to create the
1696 * first subflow right now. Otherwise we are in the defer_connect
1697 * path, and the first subflow must be already present.
1698 * Since the defer_connect flag is cleared after the first succsful
1699 * fastopen attempt, no need to check for additional subflow status.
1701 if (msg->msg_flags & MSG_FASTOPEN) {
1702 ssock = __mptcp_nmpc_socket(msk);
1704 return PTR_ERR(ssock);
1712 msg->msg_flags |= MSG_DONTWAIT;
1713 msk->fastopening = 1;
1714 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1715 msk->fastopening = 0;
1716 msg->msg_flags = saved_flags;
1719 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1720 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1721 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1722 msg->msg_namelen, msg->msg_flags, 1);
1724 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1725 * case of any error, except timeout or signal
1727 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1729 } else if (ret && ret != -EINPROGRESS) {
1730 mptcp_disconnect(sk, 0);
1732 inet_sk(sk)->defer_connect = 0;
1737 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1739 struct mptcp_sock *msk = mptcp_sk(sk);
1740 struct page_frag *pfrag;
1745 /* silently ignore everything else */
1746 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1750 if (unlikely(inet_sk(sk)->defer_connect || msg->msg_flags & MSG_FASTOPEN)) {
1753 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1754 copied += copied_syn;
1755 if (ret == -EINPROGRESS && copied_syn > 0)
1761 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1763 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1764 ret = sk_stream_wait_connect(sk, &timeo);
1770 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1773 pfrag = sk_page_frag(sk);
1775 while (msg_data_left(msg)) {
1776 int total_ts, frag_truesize = 0;
1777 struct mptcp_data_frag *dfrag;
1778 bool dfrag_collapsed;
1779 size_t psize, offset;
1781 /* reuse tail pfrag, if possible, or carve a new one from the
1784 dfrag = mptcp_pending_tail(sk);
1785 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1786 if (!dfrag_collapsed) {
1787 if (!sk_stream_memory_free(sk))
1788 goto wait_for_memory;
1790 if (!mptcp_page_frag_refill(sk, pfrag))
1791 goto wait_for_memory;
1793 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1794 frag_truesize = dfrag->overhead;
1797 /* we do not bound vs wspace, to allow a single packet.
1798 * memory accounting will prevent execessive memory usage
1801 offset = dfrag->offset + dfrag->data_len;
1802 psize = pfrag->size - offset;
1803 psize = min_t(size_t, psize, msg_data_left(msg));
1804 total_ts = psize + frag_truesize;
1806 if (!sk_wmem_schedule(sk, total_ts))
1807 goto wait_for_memory;
1809 if (copy_page_from_iter(dfrag->page, offset, psize,
1810 &msg->msg_iter) != psize) {
1815 /* data successfully copied into the write queue */
1816 sk->sk_forward_alloc -= total_ts;
1818 dfrag->data_len += psize;
1819 frag_truesize += psize;
1820 pfrag->offset += frag_truesize;
1821 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1823 /* charge data on mptcp pending queue to the msk socket
1824 * Note: we charge such data both to sk and ssk
1826 sk_wmem_queued_add(sk, frag_truesize);
1827 if (!dfrag_collapsed) {
1828 get_page(dfrag->page);
1829 list_add_tail(&dfrag->list, &msk->rtx_queue);
1830 if (!msk->first_pending)
1831 WRITE_ONCE(msk->first_pending, dfrag);
1833 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1834 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1840 mptcp_set_nospace(sk);
1841 __mptcp_push_pending(sk, msg->msg_flags);
1842 ret = sk_stream_wait_memory(sk, &timeo);
1848 __mptcp_push_pending(sk, msg->msg_flags);
1858 copied = sk_stream_error(sk, msg->msg_flags, ret);
1862 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1864 size_t len, int flags,
1865 struct scm_timestamping_internal *tss,
1868 struct sk_buff *skb, *tmp;
1871 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1872 u32 offset = MPTCP_SKB_CB(skb)->offset;
1873 u32 data_len = skb->len - offset;
1874 u32 count = min_t(size_t, len - copied, data_len);
1877 if (!(flags & MSG_TRUNC)) {
1878 err = skb_copy_datagram_msg(skb, offset, msg, count);
1879 if (unlikely(err < 0)) {
1886 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1887 tcp_update_recv_tstamps(skb, tss);
1888 *cmsg_flags |= MPTCP_CMSG_TS;
1893 if (count < data_len) {
1894 if (!(flags & MSG_PEEK)) {
1895 MPTCP_SKB_CB(skb)->offset += count;
1896 MPTCP_SKB_CB(skb)->map_seq += count;
1901 if (!(flags & MSG_PEEK)) {
1902 /* we will bulk release the skb memory later */
1903 skb->destructor = NULL;
1904 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1905 __skb_unlink(skb, &msk->receive_queue);
1916 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1918 * Only difference: Use highest rtt estimate of the subflows in use.
1920 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1922 struct mptcp_subflow_context *subflow;
1923 struct sock *sk = (struct sock *)msk;
1924 u32 time, advmss = 1;
1927 msk_owned_by_me(msk);
1932 msk->rcvq_space.copied += copied;
1934 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1935 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1937 rtt_us = msk->rcvq_space.rtt_us;
1938 if (rtt_us && time < (rtt_us >> 3))
1942 mptcp_for_each_subflow(msk, subflow) {
1943 const struct tcp_sock *tp;
1947 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1949 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1950 sf_advmss = READ_ONCE(tp->advmss);
1952 rtt_us = max(sf_rtt_us, rtt_us);
1953 advmss = max(sf_advmss, advmss);
1956 msk->rcvq_space.rtt_us = rtt_us;
1957 if (time < (rtt_us >> 3) || rtt_us == 0)
1960 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1963 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1964 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1968 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1970 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1972 do_div(grow, msk->rcvq_space.space);
1973 rcvwin += (grow << 1);
1975 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1976 while (tcp_win_from_space(sk, rcvmem) < advmss)
1979 do_div(rcvwin, advmss);
1980 rcvbuf = min_t(u64, rcvwin * rcvmem,
1981 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1983 if (rcvbuf > sk->sk_rcvbuf) {
1986 window_clamp = tcp_win_from_space(sk, rcvbuf);
1987 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1989 /* Make subflows follow along. If we do not do this, we
1990 * get drops at subflow level if skbs can't be moved to
1991 * the mptcp rx queue fast enough (announced rcv_win can
1992 * exceed ssk->sk_rcvbuf).
1994 mptcp_for_each_subflow(msk, subflow) {
1998 ssk = mptcp_subflow_tcp_sock(subflow);
1999 slow = lock_sock_fast(ssk);
2000 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2001 tcp_sk(ssk)->window_clamp = window_clamp;
2002 tcp_cleanup_rbuf(ssk, 1);
2003 unlock_sock_fast(ssk, slow);
2008 msk->rcvq_space.space = msk->rcvq_space.copied;
2010 msk->rcvq_space.copied = 0;
2011 msk->rcvq_space.time = mstamp;
2014 static void __mptcp_update_rmem(struct sock *sk)
2016 struct mptcp_sock *msk = mptcp_sk(sk);
2018 if (!msk->rmem_released)
2021 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2022 mptcp_rmem_uncharge(sk, msk->rmem_released);
2023 WRITE_ONCE(msk->rmem_released, 0);
2026 static void __mptcp_splice_receive_queue(struct sock *sk)
2028 struct mptcp_sock *msk = mptcp_sk(sk);
2030 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2033 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2035 struct sock *sk = (struct sock *)msk;
2036 unsigned int moved = 0;
2040 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2043 /* we can have data pending in the subflows only if the msk
2044 * receive buffer was full at subflow_data_ready() time,
2045 * that is an unlikely slow path.
2050 slowpath = lock_sock_fast(ssk);
2051 mptcp_data_lock(sk);
2052 __mptcp_update_rmem(sk);
2053 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2054 mptcp_data_unlock(sk);
2056 if (unlikely(ssk->sk_err))
2057 __mptcp_error_report(sk);
2058 unlock_sock_fast(ssk, slowpath);
2061 /* acquire the data lock only if some input data is pending */
2063 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2064 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2065 mptcp_data_lock(sk);
2066 __mptcp_update_rmem(sk);
2067 ret |= __mptcp_ofo_queue(msk);
2068 __mptcp_splice_receive_queue(sk);
2069 mptcp_data_unlock(sk);
2072 mptcp_check_data_fin((struct sock *)msk);
2073 return !skb_queue_empty(&msk->receive_queue);
2076 static unsigned int mptcp_inq_hint(const struct sock *sk)
2078 const struct mptcp_sock *msk = mptcp_sk(sk);
2079 const struct sk_buff *skb;
2081 skb = skb_peek(&msk->receive_queue);
2083 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2085 if (hint_val >= INT_MAX)
2088 return (unsigned int)hint_val;
2091 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2097 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2098 int flags, int *addr_len)
2100 struct mptcp_sock *msk = mptcp_sk(sk);
2101 struct scm_timestamping_internal tss;
2102 int copied = 0, cmsg_flags = 0;
2106 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2107 if (unlikely(flags & MSG_ERRQUEUE))
2108 return inet_recv_error(sk, msg, len, addr_len);
2111 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2116 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2118 len = min_t(size_t, len, INT_MAX);
2119 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2121 if (unlikely(msk->recvmsg_inq))
2122 cmsg_flags = MPTCP_CMSG_INQ;
2124 while (copied < len) {
2127 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2128 if (unlikely(bytes_read < 0)) {
2130 copied = bytes_read;
2134 copied += bytes_read;
2136 /* be sure to advertise window change */
2137 mptcp_cleanup_rbuf(msk);
2139 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2142 /* only the master socket status is relevant here. The exit
2143 * conditions mirror closely tcp_recvmsg()
2145 if (copied >= target)
2150 sk->sk_state == TCP_CLOSE ||
2151 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2153 signal_pending(current))
2157 copied = sock_error(sk);
2161 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2162 mptcp_check_for_eof(msk);
2164 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2165 /* race breaker: the shutdown could be after the
2166 * previous receive queue check
2168 if (__mptcp_move_skbs(msk))
2173 if (sk->sk_state == TCP_CLOSE) {
2183 if (signal_pending(current)) {
2184 copied = sock_intr_errno(timeo);
2189 pr_debug("block timeout %ld", timeo);
2190 sk_wait_data(sk, &timeo, NULL);
2194 if (cmsg_flags && copied >= 0) {
2195 if (cmsg_flags & MPTCP_CMSG_TS)
2196 tcp_recv_timestamp(msg, sk, &tss);
2198 if (cmsg_flags & MPTCP_CMSG_INQ) {
2199 unsigned int inq = mptcp_inq_hint(sk);
2201 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2205 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2206 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2207 skb_queue_empty(&msk->receive_queue), copied);
2208 if (!(flags & MSG_PEEK))
2209 mptcp_rcv_space_adjust(msk, copied);
2215 static void mptcp_retransmit_timer(struct timer_list *t)
2217 struct inet_connection_sock *icsk = from_timer(icsk, t,
2218 icsk_retransmit_timer);
2219 struct sock *sk = &icsk->icsk_inet.sk;
2220 struct mptcp_sock *msk = mptcp_sk(sk);
2223 if (!sock_owned_by_user(sk)) {
2224 /* we need a process context to retransmit */
2225 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2226 mptcp_schedule_work(sk);
2228 /* delegate our work to tcp_release_cb() */
2229 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2235 static void mptcp_timeout_timer(struct timer_list *t)
2237 struct sock *sk = from_timer(sk, t, sk_timer);
2239 mptcp_schedule_work(sk);
2243 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2246 * A backup subflow is returned only if that is the only kind available.
2248 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2250 struct sock *backup = NULL, *pick = NULL;
2251 struct mptcp_subflow_context *subflow;
2252 int min_stale_count = INT_MAX;
2254 msk_owned_by_me(msk);
2256 if (__mptcp_check_fallback(msk))
2259 mptcp_for_each_subflow(msk, subflow) {
2260 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2262 if (!__mptcp_subflow_active(subflow))
2265 /* still data outstanding at TCP level? skip this */
2266 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2267 mptcp_pm_subflow_chk_stale(msk, ssk);
2268 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2272 if (subflow->backup) {
2285 /* use backup only if there are no progresses anywhere */
2286 return min_stale_count > 1 ? backup : NULL;
2289 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2292 iput(SOCK_INODE(msk->subflow));
2293 WRITE_ONCE(msk->subflow, NULL);
2297 bool __mptcp_retransmit_pending_data(struct sock *sk)
2299 struct mptcp_data_frag *cur, *rtx_head;
2300 struct mptcp_sock *msk = mptcp_sk(sk);
2302 if (__mptcp_check_fallback(msk))
2305 if (tcp_rtx_and_write_queues_empty(sk))
2308 /* the closing socket has some data untransmitted and/or unacked:
2309 * some data in the mptcp rtx queue has not really xmitted yet.
2310 * keep it simple and re-inject the whole mptcp level rtx queue
2312 mptcp_data_lock(sk);
2313 __mptcp_clean_una_wakeup(sk);
2314 rtx_head = mptcp_rtx_head(sk);
2316 mptcp_data_unlock(sk);
2320 msk->recovery_snd_nxt = msk->snd_nxt;
2321 msk->recovery = true;
2322 mptcp_data_unlock(sk);
2324 msk->first_pending = rtx_head;
2327 /* be sure to clear the "sent status" on all re-injected fragments */
2328 list_for_each_entry(cur, &msk->rtx_queue, list) {
2329 if (!cur->already_sent)
2331 cur->already_sent = 0;
2337 /* flags for __mptcp_close_ssk() */
2338 #define MPTCP_CF_PUSH BIT(1)
2339 #define MPTCP_CF_FASTCLOSE BIT(2)
2341 /* subflow sockets can be either outgoing (connect) or incoming
2344 * Outgoing subflows use in-kernel sockets.
2345 * Incoming subflows do not have their own 'struct socket' allocated,
2346 * so we need to use tcp_close() after detaching them from the mptcp
2349 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2350 struct mptcp_subflow_context *subflow,
2353 struct mptcp_sock *msk = mptcp_sk(sk);
2354 bool dispose_it, need_push = false;
2356 /* If the first subflow moved to a close state before accept, e.g. due
2357 * to an incoming reset, mptcp either:
2358 * - if either the subflow or the msk are dead, destroy the context
2359 * (the subflow socket is deleted by inet_child_forget) and the msk
2360 * - otherwise do nothing at the moment and take action at accept and/or
2361 * listener shutdown - user-space must be able to accept() the closed
2364 if (msk->in_accept_queue && msk->first == ssk) {
2365 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2368 /* ensure later check in mptcp_worker() will dispose the msk */
2369 sock_set_flag(sk, SOCK_DEAD);
2370 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2371 mptcp_subflow_drop_ctx(ssk);
2375 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2377 list_del(&subflow->node);
2379 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2381 if (flags & MPTCP_CF_FASTCLOSE) {
2382 /* be sure to force the tcp_disconnect() path,
2383 * to generate the egress reset
2385 ssk->sk_lingertime = 0;
2386 sock_set_flag(ssk, SOCK_LINGER);
2387 subflow->send_fastclose = 1;
2390 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2392 tcp_disconnect(ssk, 0);
2393 msk->subflow->state = SS_UNCONNECTED;
2394 mptcp_subflow_ctx_reset(subflow);
2400 subflow->disposable = 1;
2402 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2403 * the ssk has been already destroyed, we just need to release the
2404 * reference owned by msk;
2406 if (!inet_csk(ssk)->icsk_ulp_ops) {
2407 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2408 kfree_rcu(subflow, rcu);
2410 /* otherwise tcp will dispose of the ssk and subflow ctx */
2411 if (ssk->sk_state == TCP_LISTEN) {
2412 tcp_set_state(ssk, TCP_CLOSE);
2413 mptcp_subflow_queue_clean(sk, ssk);
2414 inet_csk_listen_stop(ssk);
2415 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2418 __tcp_close(ssk, 0);
2420 /* close acquired an extra ref */
2429 if (ssk == msk->first)
2430 WRITE_ONCE(msk->first, NULL);
2433 if (ssk == msk->last_snd)
2434 msk->last_snd = NULL;
2437 __mptcp_push_pending(sk, 0);
2440 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2441 struct mptcp_subflow_context *subflow)
2443 if (sk->sk_state == TCP_ESTABLISHED)
2444 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2446 /* subflow aborted before reaching the fully_established status
2447 * attempt the creation of the next subflow
2449 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2451 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2454 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2459 static void __mptcp_close_subflow(struct sock *sk)
2461 struct mptcp_subflow_context *subflow, *tmp;
2462 struct mptcp_sock *msk = mptcp_sk(sk);
2466 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2467 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2469 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2472 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2473 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2476 mptcp_close_ssk(sk, ssk, subflow);
2481 static bool mptcp_should_close(const struct sock *sk)
2483 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2484 struct mptcp_subflow_context *subflow;
2486 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2489 /* if all subflows are in closed status don't bother with additional
2492 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2493 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2500 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2502 struct mptcp_subflow_context *subflow, *tmp;
2503 struct sock *sk = (struct sock *)msk;
2505 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2508 mptcp_token_destroy(msk);
2510 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2511 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2514 slow = lock_sock_fast(tcp_sk);
2515 if (tcp_sk->sk_state != TCP_CLOSE) {
2516 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2517 tcp_set_state(tcp_sk, TCP_CLOSE);
2519 unlock_sock_fast(tcp_sk, slow);
2522 /* Mirror the tcp_reset() error propagation */
2523 switch (sk->sk_state) {
2525 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2527 case TCP_CLOSE_WAIT:
2528 WRITE_ONCE(sk->sk_err, EPIPE);
2533 WRITE_ONCE(sk->sk_err, ECONNRESET);
2536 inet_sk_state_store(sk, TCP_CLOSE);
2537 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2538 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2539 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2541 /* the calling mptcp_worker will properly destroy the socket */
2542 if (sock_flag(sk, SOCK_DEAD))
2545 sk->sk_state_change(sk);
2546 sk_error_report(sk);
2549 static void __mptcp_retrans(struct sock *sk)
2551 struct mptcp_sock *msk = mptcp_sk(sk);
2552 struct mptcp_sendmsg_info info = {};
2553 struct mptcp_data_frag *dfrag;
2558 mptcp_clean_una_wakeup(sk);
2560 /* first check ssk: need to kick "stale" logic */
2561 ssk = mptcp_subflow_get_retrans(msk);
2562 dfrag = mptcp_rtx_head(sk);
2564 if (mptcp_data_fin_enabled(msk)) {
2565 struct inet_connection_sock *icsk = inet_csk(sk);
2567 icsk->icsk_retransmits++;
2568 mptcp_set_datafin_timeout(sk);
2569 mptcp_send_ack(msk);
2574 if (!mptcp_send_head(sk))
2585 /* limit retransmission to the bytes already sent on some subflows */
2587 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2588 while (info.sent < info.limit) {
2589 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2593 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2598 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2599 msk->bytes_retrans += copied;
2600 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2602 WRITE_ONCE(msk->allow_infinite_fallback, false);
2608 mptcp_check_and_set_pending(sk);
2610 if (!mptcp_timer_pending(sk))
2611 mptcp_reset_timer(sk);
2614 /* schedule the timeout timer for the relevant event: either close timeout
2615 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2617 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2619 struct sock *sk = (struct sock *)msk;
2620 unsigned long timeout, close_timeout;
2622 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2625 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2627 /* the close timeout takes precedence on the fail one, and here at least one of
2630 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2632 sk_reset_timer(sk, &sk->sk_timer, timeout);
2635 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2637 struct sock *ssk = msk->first;
2643 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2645 slow = lock_sock_fast(ssk);
2646 mptcp_subflow_reset(ssk);
2647 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2648 unlock_sock_fast(ssk, slow);
2650 mptcp_reset_timeout(msk, 0);
2653 static void mptcp_do_fastclose(struct sock *sk)
2655 struct mptcp_subflow_context *subflow, *tmp;
2656 struct mptcp_sock *msk = mptcp_sk(sk);
2658 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2659 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2660 subflow, MPTCP_CF_FASTCLOSE);
2663 static void mptcp_worker(struct work_struct *work)
2665 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2666 struct sock *sk = (struct sock *)msk;
2667 unsigned long fail_tout;
2671 state = sk->sk_state;
2672 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2675 mptcp_check_data_fin_ack(sk);
2677 mptcp_check_fastclose(msk);
2679 mptcp_pm_nl_work(msk);
2681 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2682 mptcp_check_for_eof(msk);
2684 __mptcp_check_send_data_fin(sk);
2685 mptcp_check_data_fin(sk);
2687 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2688 __mptcp_close_subflow(sk);
2690 /* There is no point in keeping around an orphaned sk timedout or
2691 * closed, but we need the msk around to reply to incoming DATA_FIN,
2692 * even if it is orphaned and in FIN_WAIT2 state
2694 if (sock_flag(sk, SOCK_DEAD)) {
2695 if (mptcp_should_close(sk)) {
2696 inet_sk_state_store(sk, TCP_CLOSE);
2697 mptcp_do_fastclose(sk);
2699 if (sk->sk_state == TCP_CLOSE) {
2700 __mptcp_destroy_sock(sk);
2705 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2706 __mptcp_retrans(sk);
2708 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2709 if (fail_tout && time_after(jiffies, fail_tout))
2710 mptcp_mp_fail_no_response(msk);
2717 static int __mptcp_init_sock(struct sock *sk)
2719 struct mptcp_sock *msk = mptcp_sk(sk);
2721 INIT_LIST_HEAD(&msk->conn_list);
2722 INIT_LIST_HEAD(&msk->join_list);
2723 INIT_LIST_HEAD(&msk->rtx_queue);
2724 INIT_WORK(&msk->work, mptcp_worker);
2725 __skb_queue_head_init(&msk->receive_queue);
2726 msk->out_of_order_queue = RB_ROOT;
2727 msk->first_pending = NULL;
2728 msk->rmem_fwd_alloc = 0;
2729 WRITE_ONCE(msk->rmem_released, 0);
2730 msk->timer_ival = TCP_RTO_MIN;
2732 WRITE_ONCE(msk->first, NULL);
2733 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2734 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2735 WRITE_ONCE(msk->allow_infinite_fallback, true);
2736 msk->recovery = false;
2737 msk->subflow_id = 1;
2739 mptcp_pm_data_init(msk);
2741 /* re-use the csk retrans timer for MPTCP-level retrans */
2742 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2743 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2748 static void mptcp_ca_reset(struct sock *sk)
2750 struct inet_connection_sock *icsk = inet_csk(sk);
2752 tcp_assign_congestion_control(sk);
2753 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2755 /* no need to keep a reference to the ops, the name will suffice */
2756 tcp_cleanup_congestion_control(sk);
2757 icsk->icsk_ca_ops = NULL;
2760 static int mptcp_init_sock(struct sock *sk)
2762 struct net *net = sock_net(sk);
2765 ret = __mptcp_init_sock(sk);
2769 if (!mptcp_is_enabled(net))
2770 return -ENOPROTOOPT;
2772 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2775 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2777 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2778 * propagate the correct value
2782 sk_sockets_allocated_inc(sk);
2783 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2784 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2789 static void __mptcp_clear_xmit(struct sock *sk)
2791 struct mptcp_sock *msk = mptcp_sk(sk);
2792 struct mptcp_data_frag *dtmp, *dfrag;
2794 WRITE_ONCE(msk->first_pending, NULL);
2795 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2796 dfrag_clear(sk, dfrag);
2799 void mptcp_cancel_work(struct sock *sk)
2801 struct mptcp_sock *msk = mptcp_sk(sk);
2803 if (cancel_work_sync(&msk->work))
2807 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2811 switch (ssk->sk_state) {
2813 if (!(how & RCV_SHUTDOWN))
2817 tcp_disconnect(ssk, O_NONBLOCK);
2820 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2821 pr_debug("Fallback");
2822 ssk->sk_shutdown |= how;
2823 tcp_shutdown(ssk, how);
2825 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2827 if (!mptcp_timer_pending(sk))
2828 mptcp_reset_timer(sk);
2836 static const unsigned char new_state[16] = {
2837 /* current state: new state: action: */
2838 [0 /* (Invalid) */] = TCP_CLOSE,
2839 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2840 [TCP_SYN_SENT] = TCP_CLOSE,
2841 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2842 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2843 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2844 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2845 [TCP_CLOSE] = TCP_CLOSE,
2846 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2847 [TCP_LAST_ACK] = TCP_LAST_ACK,
2848 [TCP_LISTEN] = TCP_CLOSE,
2849 [TCP_CLOSING] = TCP_CLOSING,
2850 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2853 static int mptcp_close_state(struct sock *sk)
2855 int next = (int)new_state[sk->sk_state];
2856 int ns = next & TCP_STATE_MASK;
2858 inet_sk_state_store(sk, ns);
2860 return next & TCP_ACTION_FIN;
2863 static void __mptcp_check_send_data_fin(struct sock *sk)
2865 struct mptcp_subflow_context *subflow;
2866 struct mptcp_sock *msk = mptcp_sk(sk);
2868 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2869 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2870 msk->snd_nxt, msk->write_seq);
2872 /* we still need to enqueue subflows or not really shutting down,
2875 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2876 mptcp_send_head(sk))
2879 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2881 /* fallback socket will not get data_fin/ack, can move to the next
2884 if (__mptcp_check_fallback(msk)) {
2885 WRITE_ONCE(msk->snd_una, msk->write_seq);
2886 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2887 inet_sk_state_store(sk, TCP_CLOSE);
2888 mptcp_close_wake_up(sk);
2889 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2890 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2894 mptcp_for_each_subflow(msk, subflow) {
2895 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2897 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2901 static void __mptcp_wr_shutdown(struct sock *sk)
2903 struct mptcp_sock *msk = mptcp_sk(sk);
2905 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2906 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2907 !!mptcp_send_head(sk));
2909 /* will be ignored by fallback sockets */
2910 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2911 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2913 __mptcp_check_send_data_fin(sk);
2916 static void __mptcp_destroy_sock(struct sock *sk)
2918 struct mptcp_sock *msk = mptcp_sk(sk);
2920 pr_debug("msk=%p", msk);
2924 mptcp_stop_timer(sk);
2925 sk_stop_timer(sk, &sk->sk_timer);
2928 sk->sk_prot->destroy(sk);
2930 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2931 WARN_ON_ONCE(msk->rmem_released);
2932 sk_stream_kill_queues(sk);
2933 xfrm_sk_free_policy(sk);
2938 void __mptcp_unaccepted_force_close(struct sock *sk)
2940 sock_set_flag(sk, SOCK_DEAD);
2941 inet_sk_state_store(sk, TCP_CLOSE);
2942 mptcp_do_fastclose(sk);
2943 __mptcp_destroy_sock(sk);
2946 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2948 /* Concurrent splices from sk_receive_queue into receive_queue will
2949 * always show at least one non-empty queue when checked in this order.
2951 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2952 skb_queue_empty_lockless(&msk->receive_queue))
2955 return EPOLLIN | EPOLLRDNORM;
2958 static void mptcp_listen_inuse_dec(struct sock *sk)
2960 if (inet_sk_state_load(sk) == TCP_LISTEN)
2961 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2964 bool __mptcp_close(struct sock *sk, long timeout)
2966 struct mptcp_subflow_context *subflow;
2967 struct mptcp_sock *msk = mptcp_sk(sk);
2968 bool do_cancel_work = false;
2969 int subflows_alive = 0;
2971 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2973 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2974 mptcp_listen_inuse_dec(sk);
2975 inet_sk_state_store(sk, TCP_CLOSE);
2979 if (mptcp_check_readable(msk) || timeout < 0) {
2980 /* If the msk has read data, or the caller explicitly ask it,
2981 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
2983 inet_sk_state_store(sk, TCP_CLOSE);
2984 mptcp_do_fastclose(sk);
2986 } else if (mptcp_close_state(sk)) {
2987 __mptcp_wr_shutdown(sk);
2990 sk_stream_wait_close(sk, timeout);
2993 /* orphan all the subflows */
2994 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2995 mptcp_for_each_subflow(msk, subflow) {
2996 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2997 bool slow = lock_sock_fast_nested(ssk);
2999 subflows_alive += ssk->sk_state != TCP_CLOSE;
3001 /* since the close timeout takes precedence on the fail one,
3004 if (ssk == msk->first)
3005 subflow->fail_tout = 0;
3007 /* detach from the parent socket, but allow data_ready to
3008 * push incoming data into the mptcp stack, to properly ack it
3010 ssk->sk_socket = NULL;
3012 unlock_sock_fast(ssk, slow);
3016 /* all the subflows are closed, only timeout can change the msk
3017 * state, let's not keep resources busy for no reasons
3019 if (subflows_alive == 0)
3020 inet_sk_state_store(sk, TCP_CLOSE);
3023 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3025 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3027 if (sk->sk_state == TCP_CLOSE) {
3028 __mptcp_destroy_sock(sk);
3029 do_cancel_work = true;
3031 mptcp_reset_timeout(msk, 0);
3034 return do_cancel_work;
3037 static void mptcp_close(struct sock *sk, long timeout)
3039 bool do_cancel_work;
3043 do_cancel_work = __mptcp_close(sk, timeout);
3046 mptcp_cancel_work(sk);
3051 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3053 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3054 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3055 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3057 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3058 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3061 msk6->saddr = ssk6->saddr;
3062 msk6->flow_label = ssk6->flow_label;
3066 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3067 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3068 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3069 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3070 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3071 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3074 static int mptcp_disconnect(struct sock *sk, int flags)
3076 struct mptcp_sock *msk = mptcp_sk(sk);
3078 /* We are on the fastopen error path. We can't call straight into the
3079 * subflows cleanup code due to lock nesting (we are already under
3080 * msk->firstsocket lock). Do nothing and leave the cleanup to the
3083 if (msk->fastopening)
3086 mptcp_listen_inuse_dec(sk);
3087 inet_sk_state_store(sk, TCP_CLOSE);
3089 mptcp_stop_timer(sk);
3090 sk_stop_timer(sk, &sk->sk_timer);
3093 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3095 /* msk->subflow is still intact, the following will not free the first
3098 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3099 msk->last_snd = NULL;
3100 WRITE_ONCE(msk->flags, 0);
3102 msk->push_pending = 0;
3103 msk->recovery = false;
3104 msk->can_ack = false;
3105 msk->fully_established = false;
3106 msk->rcv_data_fin = false;
3107 msk->snd_data_fin_enable = false;
3108 msk->rcv_fastclose = false;
3109 msk->use_64bit_ack = false;
3110 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3111 mptcp_pm_data_reset(msk);
3113 msk->bytes_acked = 0;
3114 msk->bytes_received = 0;
3115 msk->bytes_sent = 0;
3116 msk->bytes_retrans = 0;
3118 WRITE_ONCE(sk->sk_shutdown, 0);
3119 sk_error_report(sk);
3123 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3124 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3126 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3128 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3132 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3133 const struct mptcp_options_received *mp_opt,
3135 struct request_sock *req)
3137 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3138 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3139 struct mptcp_sock *msk;
3144 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3145 if (nsk->sk_family == AF_INET6)
3146 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3149 __mptcp_init_sock(nsk);
3151 msk = mptcp_sk(nsk);
3152 msk->local_key = subflow_req->local_key;
3153 msk->token = subflow_req->token;
3154 WRITE_ONCE(msk->subflow, NULL);
3155 msk->in_accept_queue = 1;
3156 WRITE_ONCE(msk->fully_established, false);
3157 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3158 WRITE_ONCE(msk->csum_enabled, true);
3160 msk->write_seq = subflow_req->idsn + 1;
3161 msk->snd_nxt = msk->write_seq;
3162 msk->snd_una = msk->write_seq;
3163 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3164 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3166 /* passive msk is created after the first/MPC subflow */
3167 msk->subflow_id = 2;
3169 sock_reset_flag(nsk, SOCK_RCU_FREE);
3170 security_inet_csk_clone(nsk, req);
3172 /* this can't race with mptcp_close(), as the msk is
3173 * not yet exposted to user-space
3175 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3177 /* The msk maintain a ref to each subflow in the connections list */
3178 WRITE_ONCE(msk->first, ssk);
3179 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3182 /* new mpc subflow takes ownership of the newly
3183 * created mptcp socket
3185 mptcp_token_accept(subflow_req, msk);
3187 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3188 * uses the correct data
3190 mptcp_copy_inaddrs(nsk, ssk);
3191 mptcp_propagate_sndbuf(nsk, ssk);
3193 mptcp_rcv_space_init(msk, ssk);
3194 bh_unlock_sock(nsk);
3196 /* note: the newly allocated socket refcount is 2 now */
3200 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3202 const struct tcp_sock *tp = tcp_sk(ssk);
3204 msk->rcvq_space.copied = 0;
3205 msk->rcvq_space.rtt_us = 0;
3207 msk->rcvq_space.time = tp->tcp_mstamp;
3209 /* initial rcv_space offering made to peer */
3210 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3211 TCP_INIT_CWND * tp->advmss);
3212 if (msk->rcvq_space.space == 0)
3213 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3215 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3218 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3221 struct mptcp_sock *msk = mptcp_sk(sk);
3222 struct socket *listener;
3225 listener = READ_ONCE(msk->subflow);
3226 if (WARN_ON_ONCE(!listener)) {
3231 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3232 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3236 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3237 if (sk_is_mptcp(newsk)) {
3238 struct mptcp_subflow_context *subflow;
3239 struct sock *new_mptcp_sock;
3241 subflow = mptcp_subflow_ctx(newsk);
3242 new_mptcp_sock = subflow->conn;
3244 /* is_mptcp should be false if subflow->conn is missing, see
3245 * subflow_syn_recv_sock()
3247 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3248 tcp_sk(newsk)->is_mptcp = 0;
3252 newsk = new_mptcp_sock;
3253 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3255 MPTCP_INC_STATS(sock_net(sk),
3256 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3260 newsk->sk_kern_sock = kern;
3264 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3266 struct mptcp_subflow_context *subflow, *tmp;
3267 struct sock *sk = (struct sock *)msk;
3269 __mptcp_clear_xmit(sk);
3271 /* join list will be eventually flushed (with rst) at sock lock release time */
3272 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3273 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3275 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3276 mptcp_data_lock(sk);
3277 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3278 __skb_queue_purge(&sk->sk_receive_queue);
3279 skb_rbtree_purge(&msk->out_of_order_queue);
3280 mptcp_data_unlock(sk);
3282 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3283 * inet_sock_destruct() will dispose it
3285 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3286 msk->rmem_fwd_alloc = 0;
3287 mptcp_token_destroy(msk);
3288 mptcp_pm_free_anno_list(msk);
3289 mptcp_free_local_addr_list(msk);
3292 static void mptcp_destroy(struct sock *sk)
3294 struct mptcp_sock *msk = mptcp_sk(sk);
3296 /* clears msk->subflow, allowing the following to close
3297 * even the initial subflow
3299 mptcp_dispose_initial_subflow(msk);
3300 mptcp_destroy_common(msk, 0);
3301 sk_sockets_allocated_dec(sk);
3304 void __mptcp_data_acked(struct sock *sk)
3306 if (!sock_owned_by_user(sk))
3307 __mptcp_clean_una(sk);
3309 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3311 if (mptcp_pending_data_fin_ack(sk))
3312 mptcp_schedule_work(sk);
3315 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3317 if (!mptcp_send_head(sk))
3320 if (!sock_owned_by_user(sk))
3321 __mptcp_subflow_push_pending(sk, ssk, false);
3323 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3326 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3327 BIT(MPTCP_RETRANSMIT) | \
3328 BIT(MPTCP_FLUSH_JOIN_LIST))
3330 /* processes deferred events and flush wmem */
3331 static void mptcp_release_cb(struct sock *sk)
3332 __must_hold(&sk->sk_lock.slock)
3334 struct mptcp_sock *msk = mptcp_sk(sk);
3337 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3342 /* the following actions acquire the subflow socket lock
3344 * 1) can't be invoked in atomic scope
3345 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3346 * datapath acquires the msk socket spinlock while helding
3347 * the subflow socket lock
3349 msk->push_pending = 0;
3350 msk->cb_flags &= ~flags;
3351 spin_unlock_bh(&sk->sk_lock.slock);
3352 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3353 __mptcp_flush_join_list(sk);
3354 if (flags & BIT(MPTCP_PUSH_PENDING))
3355 __mptcp_push_pending(sk, 0);
3356 if (flags & BIT(MPTCP_RETRANSMIT))
3357 __mptcp_retrans(sk);
3360 spin_lock_bh(&sk->sk_lock.slock);
3363 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3364 __mptcp_clean_una_wakeup(sk);
3365 if (unlikely(&msk->cb_flags)) {
3366 /* be sure to set the current sk state before tacking actions
3367 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3369 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3370 __mptcp_set_connected(sk);
3371 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3372 __mptcp_error_report(sk);
3373 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3374 msk->last_snd = NULL;
3377 __mptcp_update_rmem(sk);
3380 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3381 * TCP can't schedule delack timer before the subflow is fully established.
3382 * MPTCP uses the delack timer to do 3rd ack retransmissions
3384 static void schedule_3rdack_retransmission(struct sock *ssk)
3386 struct inet_connection_sock *icsk = inet_csk(ssk);
3387 struct tcp_sock *tp = tcp_sk(ssk);
3388 unsigned long timeout;
3390 if (mptcp_subflow_ctx(ssk)->fully_established)
3393 /* reschedule with a timeout above RTT, as we must look only for drop */
3395 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3397 timeout = TCP_TIMEOUT_INIT;
3400 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3401 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3402 icsk->icsk_ack.timeout = timeout;
3403 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3406 void mptcp_subflow_process_delegated(struct sock *ssk)
3408 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3409 struct sock *sk = subflow->conn;
3411 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3412 mptcp_data_lock(sk);
3413 if (!sock_owned_by_user(sk))
3414 __mptcp_subflow_push_pending(sk, ssk, true);
3416 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3417 mptcp_data_unlock(sk);
3418 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3420 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3421 schedule_3rdack_retransmission(ssk);
3422 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3426 static int mptcp_hash(struct sock *sk)
3428 /* should never be called,
3429 * we hash the TCP subflows not the master socket
3435 static void mptcp_unhash(struct sock *sk)
3437 /* called from sk_common_release(), but nothing to do here */
3440 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3442 struct mptcp_sock *msk = mptcp_sk(sk);
3443 struct socket *ssock;
3445 ssock = msk->subflow;
3446 pr_debug("msk=%p, subflow=%p", msk, ssock);
3447 if (WARN_ON_ONCE(!ssock))
3450 return inet_csk_get_port(ssock->sk, snum);
3453 void mptcp_finish_connect(struct sock *ssk)
3455 struct mptcp_subflow_context *subflow;
3456 struct mptcp_sock *msk;
3459 subflow = mptcp_subflow_ctx(ssk);
3463 pr_debug("msk=%p, token=%u", sk, subflow->token);
3465 subflow->map_seq = subflow->iasn;
3466 subflow->map_subflow_seq = 1;
3468 /* the socket is not connected yet, no msk/subflow ops can access/race
3469 * accessing the field below
3471 WRITE_ONCE(msk->local_key, subflow->local_key);
3472 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3473 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3474 WRITE_ONCE(msk->snd_una, msk->write_seq);
3476 mptcp_pm_new_connection(msk, ssk, 0);
3478 mptcp_rcv_space_init(msk, ssk);
3481 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3483 write_lock_bh(&sk->sk_callback_lock);
3484 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3485 sk_set_socket(sk, parent);
3486 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3487 write_unlock_bh(&sk->sk_callback_lock);
3490 bool mptcp_finish_join(struct sock *ssk)
3492 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3493 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3494 struct sock *parent = (void *)msk;
3497 pr_debug("msk=%p, subflow=%p", msk, subflow);
3499 /* mptcp socket already closing? */
3500 if (!mptcp_is_fully_established(parent)) {
3501 subflow->reset_reason = MPTCP_RST_EMPTCP;
3505 /* active subflow, already present inside the conn_list */
3506 if (!list_empty(&subflow->node)) {
3507 mptcp_subflow_joined(msk, ssk);
3511 if (!mptcp_pm_allow_new_subflow(msk))
3512 goto err_prohibited;
3514 /* If we can't acquire msk socket lock here, let the release callback
3517 mptcp_data_lock(parent);
3518 if (!sock_owned_by_user(parent)) {
3519 ret = __mptcp_finish_join(msk, ssk);
3522 list_add_tail(&subflow->node, &msk->conn_list);
3526 list_add_tail(&subflow->node, &msk->join_list);
3527 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3529 mptcp_data_unlock(parent);
3533 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3540 static void mptcp_shutdown(struct sock *sk, int how)
3542 pr_debug("sk=%p, how=%d", sk, how);
3544 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3545 __mptcp_wr_shutdown(sk);
3548 static int mptcp_forward_alloc_get(const struct sock *sk)
3550 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3553 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3555 const struct sock *sk = (void *)msk;
3558 if (sk->sk_state == TCP_LISTEN)
3561 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3564 delta = msk->write_seq - v;
3565 if (__mptcp_check_fallback(msk) && msk->first) {
3566 struct tcp_sock *tp = tcp_sk(msk->first);
3568 /* the first subflow is disconnected after close - see
3569 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3570 * so ignore that status, too.
3572 if (!((1 << msk->first->sk_state) &
3573 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3574 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3576 if (delta > INT_MAX)
3582 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3584 struct mptcp_sock *msk = mptcp_sk(sk);
3589 if (sk->sk_state == TCP_LISTEN)
3593 __mptcp_move_skbs(msk);
3594 *karg = mptcp_inq_hint(sk);
3598 slow = lock_sock_fast(sk);
3599 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3600 unlock_sock_fast(sk, slow);
3603 slow = lock_sock_fast(sk);
3604 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3605 unlock_sock_fast(sk, slow);
3608 return -ENOIOCTLCMD;
3614 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3615 struct mptcp_subflow_context *subflow)
3617 subflow->request_mptcp = 0;
3618 __mptcp_do_fallback(msk);
3621 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3623 struct mptcp_subflow_context *subflow;
3624 struct mptcp_sock *msk = mptcp_sk(sk);
3625 struct socket *ssock;
3628 ssock = __mptcp_nmpc_socket(msk);
3630 return PTR_ERR(ssock);
3632 mptcp_token_destroy(msk);
3633 inet_sk_state_store(sk, TCP_SYN_SENT);
3634 subflow = mptcp_subflow_ctx(ssock->sk);
3635 #ifdef CONFIG_TCP_MD5SIG
3636 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3639 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3640 mptcp_subflow_early_fallback(msk, subflow);
3642 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3643 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3644 mptcp_subflow_early_fallback(msk, subflow);
3646 if (likely(!__mptcp_check_fallback(msk)))
3647 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3649 /* if reaching here via the fastopen/sendmsg path, the caller already
3650 * acquired the subflow socket lock, too.
3652 if (msk->fastopening)
3653 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3655 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3656 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3658 /* on successful connect, the msk state will be moved to established by
3659 * subflow_finish_connect()
3661 if (unlikely(err && err != -EINPROGRESS)) {
3662 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3666 mptcp_copy_inaddrs(sk, ssock->sk);
3668 /* silence EINPROGRESS and let the caller inet_stream_connect
3669 * handle the connection in progress
3674 static struct proto mptcp_prot = {
3676 .owner = THIS_MODULE,
3677 .init = mptcp_init_sock,
3678 .connect = mptcp_connect,
3679 .disconnect = mptcp_disconnect,
3680 .close = mptcp_close,
3681 .accept = mptcp_accept,
3682 .setsockopt = mptcp_setsockopt,
3683 .getsockopt = mptcp_getsockopt,
3684 .shutdown = mptcp_shutdown,
3685 .destroy = mptcp_destroy,
3686 .sendmsg = mptcp_sendmsg,
3687 .ioctl = mptcp_ioctl,
3688 .recvmsg = mptcp_recvmsg,
3689 .release_cb = mptcp_release_cb,
3691 .unhash = mptcp_unhash,
3692 .get_port = mptcp_get_port,
3693 .forward_alloc_get = mptcp_forward_alloc_get,
3694 .sockets_allocated = &mptcp_sockets_allocated,
3696 .memory_allocated = &tcp_memory_allocated,
3697 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3699 .memory_pressure = &tcp_memory_pressure,
3700 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3701 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3702 .sysctl_mem = sysctl_tcp_mem,
3703 .obj_size = sizeof(struct mptcp_sock),
3704 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3705 .no_autobind = true,
3708 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3710 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3711 struct socket *ssock;
3714 lock_sock(sock->sk);
3715 ssock = __mptcp_nmpc_socket(msk);
3716 if (IS_ERR(ssock)) {
3717 err = PTR_ERR(ssock);
3721 err = ssock->ops->bind(ssock, uaddr, addr_len);
3723 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3726 release_sock(sock->sk);
3730 static int mptcp_listen(struct socket *sock, int backlog)
3732 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3733 struct sock *sk = sock->sk;
3734 struct socket *ssock;
3737 pr_debug("msk=%p", msk);
3740 ssock = __mptcp_nmpc_socket(msk);
3741 if (IS_ERR(ssock)) {
3742 err = PTR_ERR(ssock);
3746 mptcp_token_destroy(msk);
3747 inet_sk_state_store(sk, TCP_LISTEN);
3748 sock_set_flag(sk, SOCK_RCU_FREE);
3750 err = ssock->ops->listen(ssock, backlog);
3751 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3753 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3754 mptcp_copy_inaddrs(sk, ssock->sk);
3757 mptcp_event_pm_listener(ssock->sk, MPTCP_EVENT_LISTENER_CREATED);
3764 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3765 int flags, bool kern)
3767 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3768 struct socket *ssock;
3772 pr_debug("msk=%p", msk);
3774 /* Buggy applications can call accept on socket states other then LISTEN
3775 * but no need to allocate the first subflow just to error out.
3777 ssock = READ_ONCE(msk->subflow);
3781 newsk = mptcp_accept(sock->sk, flags, &err, kern);
3787 __inet_accept(sock, newsock, newsk);
3788 if (!mptcp_is_tcpsk(newsock->sk)) {
3789 struct mptcp_sock *msk = mptcp_sk(newsk);
3790 struct mptcp_subflow_context *subflow;
3792 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3793 msk->in_accept_queue = 0;
3795 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3796 * This is needed so NOSPACE flag can be set from tcp stack.
3798 mptcp_for_each_subflow(msk, subflow) {
3799 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3801 if (!ssk->sk_socket)
3802 mptcp_sock_graft(ssk, newsock);
3805 /* Do late cleanup for the first subflow as necessary. Also
3806 * deal with bad peers not doing a complete shutdown.
3809 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3810 __mptcp_close_ssk(newsk, msk->first,
3811 mptcp_subflow_ctx(msk->first), 0);
3812 if (unlikely(list_empty(&msk->conn_list)))
3813 inet_sk_state_store(newsk, TCP_CLOSE);
3816 release_sock(newsk);
3821 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3823 struct sock *sk = (struct sock *)msk;
3825 if (sk_stream_is_writeable(sk))
3826 return EPOLLOUT | EPOLLWRNORM;
3828 mptcp_set_nospace(sk);
3829 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3830 if (sk_stream_is_writeable(sk))
3831 return EPOLLOUT | EPOLLWRNORM;
3836 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3837 struct poll_table_struct *wait)
3839 struct sock *sk = sock->sk;
3840 struct mptcp_sock *msk;
3846 sock_poll_wait(file, sock, wait);
3848 state = inet_sk_state_load(sk);
3849 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3850 if (state == TCP_LISTEN) {
3851 struct socket *ssock = READ_ONCE(msk->subflow);
3853 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3856 return inet_csk_listen_poll(ssock->sk);
3859 shutdown = READ_ONCE(sk->sk_shutdown);
3860 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3862 if (shutdown & RCV_SHUTDOWN)
3863 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3865 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3866 mask |= mptcp_check_readable(msk);
3867 if (shutdown & SEND_SHUTDOWN)
3868 mask |= EPOLLOUT | EPOLLWRNORM;
3870 mask |= mptcp_check_writeable(msk);
3871 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3872 /* cf tcp_poll() note about TFO */
3873 mask |= EPOLLOUT | EPOLLWRNORM;
3876 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3878 if (READ_ONCE(sk->sk_err))
3884 static const struct proto_ops mptcp_stream_ops = {
3886 .owner = THIS_MODULE,
3887 .release = inet_release,
3889 .connect = inet_stream_connect,
3890 .socketpair = sock_no_socketpair,
3891 .accept = mptcp_stream_accept,
3892 .getname = inet_getname,
3894 .ioctl = inet_ioctl,
3895 .gettstamp = sock_gettstamp,
3896 .listen = mptcp_listen,
3897 .shutdown = inet_shutdown,
3898 .setsockopt = sock_common_setsockopt,
3899 .getsockopt = sock_common_getsockopt,
3900 .sendmsg = inet_sendmsg,
3901 .recvmsg = inet_recvmsg,
3902 .mmap = sock_no_mmap,
3903 .sendpage = inet_sendpage,
3906 static struct inet_protosw mptcp_protosw = {
3907 .type = SOCK_STREAM,
3908 .protocol = IPPROTO_MPTCP,
3909 .prot = &mptcp_prot,
3910 .ops = &mptcp_stream_ops,
3911 .flags = INET_PROTOSW_ICSK,
3914 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3916 struct mptcp_delegated_action *delegated;
3917 struct mptcp_subflow_context *subflow;
3920 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3921 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3922 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3924 bh_lock_sock_nested(ssk);
3925 if (!sock_owned_by_user(ssk) &&
3926 mptcp_subflow_has_delegated_action(subflow))
3927 mptcp_subflow_process_delegated(ssk);
3928 /* ... elsewhere tcp_release_cb_override already processed
3929 * the action or will do at next release_sock().
3930 * In both case must dequeue the subflow here - on the same
3931 * CPU that scheduled it.
3933 bh_unlock_sock(ssk);
3936 if (++work_done == budget)
3940 /* always provide a 0 'work_done' argument, so that napi_complete_done
3941 * will not try accessing the NULL napi->dev ptr
3943 napi_complete_done(napi, 0);
3947 void __init mptcp_proto_init(void)
3949 struct mptcp_delegated_action *delegated;
3952 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3954 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3955 panic("Failed to allocate MPTCP pcpu counter\n");
3957 init_dummy_netdev(&mptcp_napi_dev);
3958 for_each_possible_cpu(cpu) {
3959 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3960 INIT_LIST_HEAD(&delegated->head);
3961 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3963 napi_enable(&delegated->napi);
3966 mptcp_subflow_init();
3970 if (proto_register(&mptcp_prot, 1) != 0)
3971 panic("Failed to register MPTCP proto.\n");
3973 inet_register_protosw(&mptcp_protosw);
3975 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3978 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3979 static const struct proto_ops mptcp_v6_stream_ops = {
3981 .owner = THIS_MODULE,
3982 .release = inet6_release,
3984 .connect = inet_stream_connect,
3985 .socketpair = sock_no_socketpair,
3986 .accept = mptcp_stream_accept,
3987 .getname = inet6_getname,
3989 .ioctl = inet6_ioctl,
3990 .gettstamp = sock_gettstamp,
3991 .listen = mptcp_listen,
3992 .shutdown = inet_shutdown,
3993 .setsockopt = sock_common_setsockopt,
3994 .getsockopt = sock_common_getsockopt,
3995 .sendmsg = inet6_sendmsg,
3996 .recvmsg = inet6_recvmsg,
3997 .mmap = sock_no_mmap,
3998 .sendpage = inet_sendpage,
3999 #ifdef CONFIG_COMPAT
4000 .compat_ioctl = inet6_compat_ioctl,
4004 static struct proto mptcp_v6_prot;
4006 static struct inet_protosw mptcp_v6_protosw = {
4007 .type = SOCK_STREAM,
4008 .protocol = IPPROTO_MPTCP,
4009 .prot = &mptcp_v6_prot,
4010 .ops = &mptcp_v6_stream_ops,
4011 .flags = INET_PROTOSW_ICSK,
4014 int __init mptcp_proto_v6_init(void)
4018 mptcp_v6_prot = mptcp_prot;
4019 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4020 mptcp_v6_prot.slab = NULL;
4021 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4023 err = proto_register(&mptcp_v6_prot, 1);
4027 err = inet6_register_protosw(&mptcp_v6_protosw);
4029 proto_unregister(&mptcp_v6_prot);