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;
100 /* This is the first subflow, always with id 0 */
101 subflow->local_id_valid = 1;
102 mptcp_sock_graft(msk->first, sk->sk_socket);
107 /* If the MPC handshake is not started, returns the first subflow,
108 * eventually allocating it.
110 struct socket *__mptcp_nmpc_socket(struct mptcp_sock *msk)
112 struct sock *sk = (struct sock *)msk;
115 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
116 return ERR_PTR(-EINVAL);
120 return ERR_PTR(-EINVAL);
122 ret = __mptcp_socket_create(msk);
126 mptcp_sockopt_sync(msk, msk->first);
132 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
134 sk_drops_add(sk, skb);
138 static void mptcp_rmem_charge(struct sock *sk, int size)
140 mptcp_sk(sk)->rmem_fwd_alloc -= size;
143 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
144 struct sk_buff *from)
149 if (MPTCP_SKB_CB(from)->offset ||
150 !skb_try_coalesce(to, from, &fragstolen, &delta))
153 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
154 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
155 to->len, MPTCP_SKB_CB(from)->end_seq);
156 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
158 /* note the fwd memory can reach a negative value after accounting
159 * for the delta, but the later skb free will restore a non
162 atomic_add(delta, &sk->sk_rmem_alloc);
163 mptcp_rmem_charge(sk, delta);
164 kfree_skb_partial(from, fragstolen);
169 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
170 struct sk_buff *from)
172 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
175 return mptcp_try_coalesce((struct sock *)msk, to, from);
178 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
180 amount >>= PAGE_SHIFT;
181 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
182 __sk_mem_reduce_allocated(sk, amount);
185 static void mptcp_rmem_uncharge(struct sock *sk, int size)
187 struct mptcp_sock *msk = mptcp_sk(sk);
190 msk->rmem_fwd_alloc += size;
191 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
193 /* see sk_mem_uncharge() for the rationale behind the following schema */
194 if (unlikely(reclaimable >= PAGE_SIZE))
195 __mptcp_rmem_reclaim(sk, reclaimable);
198 static void mptcp_rfree(struct sk_buff *skb)
200 unsigned int len = skb->truesize;
201 struct sock *sk = skb->sk;
203 atomic_sub(len, &sk->sk_rmem_alloc);
204 mptcp_rmem_uncharge(sk, len);
207 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
211 skb->destructor = mptcp_rfree;
212 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
213 mptcp_rmem_charge(sk, skb->truesize);
216 /* "inspired" by tcp_data_queue_ofo(), main differences:
218 * - don't cope with sacks
220 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
222 struct sock *sk = (struct sock *)msk;
223 struct rb_node **p, *parent;
224 u64 seq, end_seq, max_seq;
225 struct sk_buff *skb1;
227 seq = MPTCP_SKB_CB(skb)->map_seq;
228 end_seq = MPTCP_SKB_CB(skb)->end_seq;
229 max_seq = atomic64_read(&msk->rcv_wnd_sent);
231 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
232 RB_EMPTY_ROOT(&msk->out_of_order_queue));
233 if (after64(end_seq, max_seq)) {
236 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
237 (unsigned long long)end_seq - (unsigned long)max_seq,
238 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
239 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
243 p = &msk->out_of_order_queue.rb_node;
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
245 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
246 rb_link_node(&skb->rbnode, NULL, p);
247 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
248 msk->ooo_last_skb = skb;
252 /* with 2 subflows, adding at end of ooo queue is quite likely
253 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
255 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
256 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
257 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
261 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
262 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
263 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
264 parent = &msk->ooo_last_skb->rbnode;
265 p = &parent->rb_right;
269 /* Find place to insert this segment. Handle overlaps on the way. */
273 skb1 = rb_to_skb(parent);
274 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
275 p = &parent->rb_left;
278 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
279 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
280 /* All the bits are present. Drop. */
282 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
285 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
289 * continue traversing
292 /* skb's seq == skb1's seq and skb covers skb1.
293 * Replace skb1 with skb.
295 rb_replace_node(&skb1->rbnode, &skb->rbnode,
296 &msk->out_of_order_queue);
297 mptcp_drop(sk, skb1);
298 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
301 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
302 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
305 p = &parent->rb_right;
309 /* Insert segment into RB tree. */
310 rb_link_node(&skb->rbnode, parent, p);
311 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
314 /* Remove other segments covered by skb. */
315 while ((skb1 = skb_rb_next(skb)) != NULL) {
316 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
318 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
319 mptcp_drop(sk, skb1);
320 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
322 /* If there is no skb after us, we are the last_skb ! */
324 msk->ooo_last_skb = skb;
328 mptcp_set_owner_r(skb, sk);
331 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
333 struct mptcp_sock *msk = mptcp_sk(sk);
336 if (size <= msk->rmem_fwd_alloc)
339 size -= msk->rmem_fwd_alloc;
340 amt = sk_mem_pages(size);
341 amount = amt << PAGE_SHIFT;
342 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
345 msk->rmem_fwd_alloc += amount;
349 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
350 struct sk_buff *skb, unsigned int offset,
353 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
354 struct sock *sk = (struct sock *)msk;
355 struct sk_buff *tail;
358 __skb_unlink(skb, &ssk->sk_receive_queue);
363 /* try to fetch required memory from subflow */
364 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
367 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
369 /* the skb map_seq accounts for the skb offset:
370 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
373 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
374 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
375 MPTCP_SKB_CB(skb)->offset = offset;
376 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
378 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
380 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
381 tail = skb_peek_tail(&sk->sk_receive_queue);
382 if (tail && mptcp_try_coalesce(sk, tail, skb))
385 mptcp_set_owner_r(skb, sk);
386 __skb_queue_tail(&sk->sk_receive_queue, skb);
388 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
389 mptcp_data_queue_ofo(msk, skb);
393 /* old data, keep it simple and drop the whole pkt, sender
394 * will retransmit as needed, if needed.
396 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
402 static void mptcp_stop_timer(struct sock *sk)
404 struct inet_connection_sock *icsk = inet_csk(sk);
406 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
407 mptcp_sk(sk)->timer_ival = 0;
410 static void mptcp_close_wake_up(struct sock *sk)
412 if (sock_flag(sk, SOCK_DEAD))
415 sk->sk_state_change(sk);
416 if (sk->sk_shutdown == SHUTDOWN_MASK ||
417 sk->sk_state == TCP_CLOSE)
418 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
420 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
423 static bool mptcp_pending_data_fin_ack(struct sock *sk)
425 struct mptcp_sock *msk = mptcp_sk(sk);
427 return ((1 << sk->sk_state) &
428 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
429 msk->write_seq == READ_ONCE(msk->snd_una);
432 static void mptcp_check_data_fin_ack(struct sock *sk)
434 struct mptcp_sock *msk = mptcp_sk(sk);
436 /* Look for an acknowledged DATA_FIN */
437 if (mptcp_pending_data_fin_ack(sk)) {
438 WRITE_ONCE(msk->snd_data_fin_enable, 0);
440 switch (sk->sk_state) {
442 inet_sk_state_store(sk, TCP_FIN_WAIT2);
446 inet_sk_state_store(sk, TCP_CLOSE);
450 mptcp_close_wake_up(sk);
454 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
456 struct mptcp_sock *msk = mptcp_sk(sk);
458 if (READ_ONCE(msk->rcv_data_fin) &&
459 ((1 << sk->sk_state) &
460 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
461 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
463 if (msk->ack_seq == rcv_data_fin_seq) {
465 *seq = rcv_data_fin_seq;
474 static void mptcp_set_datafin_timeout(struct sock *sk)
476 struct inet_connection_sock *icsk = inet_csk(sk);
479 retransmits = min_t(u32, icsk->icsk_retransmits,
480 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
482 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
485 static void __mptcp_set_timeout(struct sock *sk, long tout)
487 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
490 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
492 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
494 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
495 inet_csk(ssk)->icsk_timeout - jiffies : 0;
498 static void mptcp_set_timeout(struct sock *sk)
500 struct mptcp_subflow_context *subflow;
503 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
504 tout = max(tout, mptcp_timeout_from_subflow(subflow));
505 __mptcp_set_timeout(sk, tout);
508 static inline bool tcp_can_send_ack(const struct sock *ssk)
510 return !((1 << inet_sk_state_load(ssk)) &
511 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
514 void __mptcp_subflow_send_ack(struct sock *ssk)
516 if (tcp_can_send_ack(ssk))
520 static void mptcp_subflow_send_ack(struct sock *ssk)
524 slow = lock_sock_fast(ssk);
525 __mptcp_subflow_send_ack(ssk);
526 unlock_sock_fast(ssk, slow);
529 static void mptcp_send_ack(struct mptcp_sock *msk)
531 struct mptcp_subflow_context *subflow;
533 mptcp_for_each_subflow(msk, subflow)
534 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
537 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
541 slow = lock_sock_fast(ssk);
542 if (tcp_can_send_ack(ssk))
543 tcp_cleanup_rbuf(ssk, 1);
544 unlock_sock_fast(ssk, slow);
547 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
549 const struct inet_connection_sock *icsk = inet_csk(ssk);
550 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
551 const struct tcp_sock *tp = tcp_sk(ssk);
553 return (ack_pending & ICSK_ACK_SCHED) &&
554 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
555 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
556 (rx_empty && ack_pending &
557 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
560 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
562 int old_space = READ_ONCE(msk->old_wspace);
563 struct mptcp_subflow_context *subflow;
564 struct sock *sk = (struct sock *)msk;
565 int space = __mptcp_space(sk);
566 bool cleanup, rx_empty;
568 cleanup = (space > 0) && (space >= (old_space << 1));
569 rx_empty = !__mptcp_rmem(sk);
571 mptcp_for_each_subflow(msk, subflow) {
572 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
574 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
575 mptcp_subflow_cleanup_rbuf(ssk);
579 static bool mptcp_check_data_fin(struct sock *sk)
581 struct mptcp_sock *msk = mptcp_sk(sk);
582 u64 rcv_data_fin_seq;
585 /* Need to ack a DATA_FIN received from a peer while this side
586 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
587 * msk->rcv_data_fin was set when parsing the incoming options
588 * at the subflow level and the msk lock was not held, so this
589 * is the first opportunity to act on the DATA_FIN and change
592 * If we are caught up to the sequence number of the incoming
593 * DATA_FIN, send the DATA_ACK now and do state transition. If
594 * not caught up, do nothing and let the recv code send DATA_ACK
598 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
599 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
600 WRITE_ONCE(msk->rcv_data_fin, 0);
602 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
603 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
605 switch (sk->sk_state) {
606 case TCP_ESTABLISHED:
607 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
610 inet_sk_state_store(sk, TCP_CLOSING);
613 inet_sk_state_store(sk, TCP_CLOSE);
616 /* Other states not expected */
622 if (!__mptcp_check_fallback(msk))
624 mptcp_close_wake_up(sk);
629 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
633 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
634 struct sock *sk = (struct sock *)msk;
635 unsigned int moved = 0;
636 bool more_data_avail;
641 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
643 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
644 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
646 if (unlikely(ssk_rbuf > sk_rbuf)) {
647 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
652 pr_debug("msk=%p ssk=%p", msk, ssk);
655 u32 map_remaining, offset;
656 u32 seq = tp->copied_seq;
660 /* try to move as much data as available */
661 map_remaining = subflow->map_data_len -
662 mptcp_subflow_get_map_offset(subflow);
664 skb = skb_peek(&ssk->sk_receive_queue);
666 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
667 * a different CPU can have already processed the pending
668 * data, stop here or we can enter an infinite loop
675 if (__mptcp_check_fallback(msk)) {
676 /* Under fallback skbs have no MPTCP extension and TCP could
677 * collapse them between the dummy map creation and the
678 * current dequeue. Be sure to adjust the map size.
680 map_remaining = skb->len;
681 subflow->map_data_len = skb->len;
684 offset = seq - TCP_SKB_CB(skb)->seq;
685 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
691 if (offset < skb->len) {
692 size_t len = skb->len - offset;
697 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
701 if (WARN_ON_ONCE(map_remaining < len))
705 sk_eat_skb(ssk, skb);
709 WRITE_ONCE(tp->copied_seq, seq);
710 more_data_avail = mptcp_subflow_data_available(ssk);
712 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
716 } while (more_data_avail);
722 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
724 struct sock *sk = (struct sock *)msk;
725 struct sk_buff *skb, *tail;
730 p = rb_first(&msk->out_of_order_queue);
731 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
734 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
738 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
740 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
743 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
747 end_seq = MPTCP_SKB_CB(skb)->end_seq;
748 tail = skb_peek_tail(&sk->sk_receive_queue);
749 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
750 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
752 /* skip overlapping data, if any */
753 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
754 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
756 MPTCP_SKB_CB(skb)->offset += delta;
757 MPTCP_SKB_CB(skb)->map_seq += delta;
758 __skb_queue_tail(&sk->sk_receive_queue, skb);
760 msk->ack_seq = end_seq;
766 /* In most cases we will be able to lock the mptcp socket. If its already
767 * owned, we need to defer to the work queue to avoid ABBA deadlock.
769 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
771 struct sock *sk = (struct sock *)msk;
772 unsigned int moved = 0;
774 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
775 __mptcp_ofo_queue(msk);
776 if (unlikely(ssk->sk_err)) {
777 if (!sock_owned_by_user(sk))
778 __mptcp_error_report(sk);
780 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
783 /* If the moves have caught up with the DATA_FIN sequence number
784 * it's time to ack the DATA_FIN and change socket state, but
785 * this is not a good place to change state. Let the workqueue
788 if (mptcp_pending_data_fin(sk, NULL))
789 mptcp_schedule_work(sk);
793 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
795 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
796 struct mptcp_sock *msk = mptcp_sk(sk);
797 int sk_rbuf, ssk_rbuf;
799 /* The peer can send data while we are shutting down this
800 * subflow at msk destruction time, but we must avoid enqueuing
801 * more data to the msk receive queue
803 if (unlikely(subflow->disposable))
806 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
807 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
808 if (unlikely(ssk_rbuf > sk_rbuf))
811 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
812 if (__mptcp_rmem(sk) > sk_rbuf) {
813 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
817 /* Wake-up the reader only for in-sequence data */
819 if (move_skbs_to_msk(msk, ssk))
820 sk->sk_data_ready(sk);
822 mptcp_data_unlock(sk);
825 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
827 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
828 WRITE_ONCE(msk->allow_infinite_fallback, false);
829 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
832 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
834 struct sock *sk = (struct sock *)msk;
836 if (sk->sk_state != TCP_ESTABLISHED)
839 /* attach to msk socket only after we are sure we will deal with it
842 if (sk->sk_socket && !ssk->sk_socket)
843 mptcp_sock_graft(ssk, sk->sk_socket);
845 mptcp_sockopt_sync_locked(msk, ssk);
846 mptcp_subflow_joined(msk, ssk);
850 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
852 struct mptcp_subflow_context *tmp, *subflow;
853 struct mptcp_sock *msk = mptcp_sk(sk);
855 list_for_each_entry_safe(subflow, tmp, join_list, node) {
856 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
857 bool slow = lock_sock_fast(ssk);
859 list_move_tail(&subflow->node, &msk->conn_list);
860 if (!__mptcp_finish_join(msk, ssk))
861 mptcp_subflow_reset(ssk);
862 unlock_sock_fast(ssk, slow);
866 static bool mptcp_timer_pending(struct sock *sk)
868 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
871 static void mptcp_reset_timer(struct sock *sk)
873 struct inet_connection_sock *icsk = inet_csk(sk);
876 /* prevent rescheduling on close */
877 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
880 tout = mptcp_sk(sk)->timer_ival;
881 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
884 bool mptcp_schedule_work(struct sock *sk)
886 if (inet_sk_state_load(sk) != TCP_CLOSE &&
887 schedule_work(&mptcp_sk(sk)->work)) {
888 /* each subflow already holds a reference to the sk, and the
889 * workqueue is invoked by a subflow, so sk can't go away here.
897 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
899 struct mptcp_subflow_context *subflow;
901 msk_owned_by_me(msk);
903 mptcp_for_each_subflow(msk, subflow) {
904 if (READ_ONCE(subflow->data_avail))
905 return mptcp_subflow_tcp_sock(subflow);
911 static bool mptcp_skb_can_collapse_to(u64 write_seq,
912 const struct sk_buff *skb,
913 const struct mptcp_ext *mpext)
915 if (!tcp_skb_can_collapse_to(skb))
918 /* can collapse only if MPTCP level sequence is in order and this
919 * mapping has not been xmitted yet
921 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
925 /* we can append data to the given data frag if:
926 * - there is space available in the backing page_frag
927 * - the data frag tail matches the current page_frag free offset
928 * - the data frag end sequence number matches the current write seq
930 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
931 const struct page_frag *pfrag,
932 const struct mptcp_data_frag *df)
934 return df && pfrag->page == df->page &&
935 pfrag->size - pfrag->offset > 0 &&
936 pfrag->offset == (df->offset + df->data_len) &&
937 df->data_seq + df->data_len == msk->write_seq;
940 static void dfrag_uncharge(struct sock *sk, int len)
942 sk_mem_uncharge(sk, len);
943 sk_wmem_queued_add(sk, -len);
946 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
948 int len = dfrag->data_len + dfrag->overhead;
950 list_del(&dfrag->list);
951 dfrag_uncharge(sk, len);
952 put_page(dfrag->page);
955 static void __mptcp_clean_una(struct sock *sk)
957 struct mptcp_sock *msk = mptcp_sk(sk);
958 struct mptcp_data_frag *dtmp, *dfrag;
961 /* on fallback we just need to ignore snd_una, as this is really
964 if (__mptcp_check_fallback(msk))
965 msk->snd_una = READ_ONCE(msk->snd_nxt);
967 snd_una = msk->snd_una;
968 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
969 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
972 if (unlikely(dfrag == msk->first_pending)) {
973 /* in recovery mode can see ack after the current snd head */
974 if (WARN_ON_ONCE(!msk->recovery))
977 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
980 dfrag_clear(sk, dfrag);
983 dfrag = mptcp_rtx_head(sk);
984 if (dfrag && after64(snd_una, dfrag->data_seq)) {
985 u64 delta = snd_una - dfrag->data_seq;
987 /* prevent wrap around in recovery mode */
988 if (unlikely(delta > dfrag->already_sent)) {
989 if (WARN_ON_ONCE(!msk->recovery))
991 if (WARN_ON_ONCE(delta > dfrag->data_len))
993 dfrag->already_sent += delta - dfrag->already_sent;
996 dfrag->data_seq += delta;
997 dfrag->offset += delta;
998 dfrag->data_len -= delta;
999 dfrag->already_sent -= delta;
1001 dfrag_uncharge(sk, delta);
1004 /* all retransmitted data acked, recovery completed */
1005 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1006 msk->recovery = false;
1009 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1010 snd_una == READ_ONCE(msk->write_seq)) {
1011 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1012 mptcp_stop_timer(sk);
1014 mptcp_reset_timer(sk);
1018 static void __mptcp_clean_una_wakeup(struct sock *sk)
1020 lockdep_assert_held_once(&sk->sk_lock.slock);
1022 __mptcp_clean_una(sk);
1023 mptcp_write_space(sk);
1026 static void mptcp_clean_una_wakeup(struct sock *sk)
1028 mptcp_data_lock(sk);
1029 __mptcp_clean_una_wakeup(sk);
1030 mptcp_data_unlock(sk);
1033 static void mptcp_enter_memory_pressure(struct sock *sk)
1035 struct mptcp_subflow_context *subflow;
1036 struct mptcp_sock *msk = mptcp_sk(sk);
1039 sk_stream_moderate_sndbuf(sk);
1040 mptcp_for_each_subflow(msk, subflow) {
1041 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1044 tcp_enter_memory_pressure(ssk);
1045 sk_stream_moderate_sndbuf(ssk);
1050 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1053 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1055 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1056 pfrag, sk->sk_allocation)))
1059 mptcp_enter_memory_pressure(sk);
1063 static struct mptcp_data_frag *
1064 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1067 int offset = ALIGN(orig_offset, sizeof(long));
1068 struct mptcp_data_frag *dfrag;
1070 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1071 dfrag->data_len = 0;
1072 dfrag->data_seq = msk->write_seq;
1073 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1074 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1075 dfrag->already_sent = 0;
1076 dfrag->page = pfrag->page;
1081 struct mptcp_sendmsg_info {
1087 bool data_lock_held;
1090 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1091 u64 data_seq, int avail_size)
1093 u64 window_end = mptcp_wnd_end(msk);
1096 if (__mptcp_check_fallback(msk))
1099 mptcp_snd_wnd = window_end - data_seq;
1100 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1102 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1103 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1104 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1110 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1112 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1116 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1120 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1122 struct sk_buff *skb;
1124 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1126 if (likely(__mptcp_add_ext(skb, gfp))) {
1127 skb_reserve(skb, MAX_TCP_HEADER);
1128 skb->ip_summed = CHECKSUM_PARTIAL;
1129 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1134 mptcp_enter_memory_pressure(sk);
1139 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1141 struct sk_buff *skb;
1143 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1147 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1148 tcp_skb_entail(ssk, skb);
1151 tcp_skb_tsorted_anchor_cleanup(skb);
1156 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1158 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1160 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1163 /* note: this always recompute the csum on the whole skb, even
1164 * if we just appended a single frag. More status info needed
1166 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1168 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1169 __wsum csum = ~csum_unfold(mpext->csum);
1170 int offset = skb->len - added;
1172 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1175 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1177 struct mptcp_ext *mpext)
1182 mpext->infinite_map = 1;
1183 mpext->data_len = 0;
1185 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1186 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1188 mptcp_do_fallback(ssk);
1191 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1192 struct mptcp_data_frag *dfrag,
1193 struct mptcp_sendmsg_info *info)
1195 u64 data_seq = dfrag->data_seq + info->sent;
1196 int offset = dfrag->offset + info->sent;
1197 struct mptcp_sock *msk = mptcp_sk(sk);
1198 bool zero_window_probe = false;
1199 struct mptcp_ext *mpext = NULL;
1200 bool can_coalesce = false;
1201 bool reuse_skb = true;
1202 struct sk_buff *skb;
1206 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1207 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1209 if (WARN_ON_ONCE(info->sent > info->limit ||
1210 info->limit > dfrag->data_len))
1213 if (unlikely(!__tcp_can_send(ssk)))
1216 /* compute send limit */
1217 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1218 copy = info->size_goal;
1220 skb = tcp_write_queue_tail(ssk);
1221 if (skb && copy > skb->len) {
1222 /* Limit the write to the size available in the
1223 * current skb, if any, so that we create at most a new skb.
1224 * Explicitly tells TCP internals to avoid collapsing on later
1225 * queue management operation, to avoid breaking the ext <->
1226 * SSN association set here
1228 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1229 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1230 TCP_SKB_CB(skb)->eor = 1;
1234 i = skb_shinfo(skb)->nr_frags;
1235 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1236 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1237 tcp_mark_push(tcp_sk(ssk), skb);
1244 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1248 i = skb_shinfo(skb)->nr_frags;
1250 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1253 /* Zero window and all data acked? Probe. */
1254 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1256 u64 snd_una = READ_ONCE(msk->snd_una);
1258 if (snd_una != msk->snd_nxt) {
1259 tcp_remove_empty_skb(ssk);
1263 zero_window_probe = true;
1264 data_seq = snd_una - 1;
1267 /* all mptcp-level data is acked, no skbs should be present into the
1270 WARN_ON_ONCE(reuse_skb);
1273 copy = min_t(size_t, copy, info->limit - info->sent);
1274 if (!sk_wmem_schedule(ssk, copy)) {
1275 tcp_remove_empty_skb(ssk);
1280 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1282 get_page(dfrag->page);
1283 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1287 skb->data_len += copy;
1288 skb->truesize += copy;
1289 sk_wmem_queued_add(ssk, copy);
1290 sk_mem_charge(ssk, copy);
1291 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1292 TCP_SKB_CB(skb)->end_seq += copy;
1293 tcp_skb_pcount_set(skb, 0);
1295 /* on skb reuse we just need to update the DSS len */
1297 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1298 mpext->data_len += copy;
1299 WARN_ON_ONCE(zero_window_probe);
1303 memset(mpext, 0, sizeof(*mpext));
1304 mpext->data_seq = data_seq;
1305 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1306 mpext->data_len = copy;
1310 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1311 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1314 if (zero_window_probe) {
1315 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1317 if (READ_ONCE(msk->csum_enabled))
1318 mptcp_update_data_checksum(skb, copy);
1319 tcp_push_pending_frames(ssk);
1323 if (READ_ONCE(msk->csum_enabled))
1324 mptcp_update_data_checksum(skb, copy);
1325 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1326 mptcp_update_infinite_map(msk, ssk, mpext);
1327 trace_mptcp_sendmsg_frag(mpext);
1328 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1332 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1333 sizeof(struct tcphdr) - \
1334 MAX_TCP_OPTION_SPACE - \
1335 sizeof(struct ipv6hdr) - \
1336 sizeof(struct frag_hdr))
1338 struct subflow_send_info {
1343 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1345 if (!subflow->stale)
1349 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1352 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1354 if (unlikely(subflow->stale)) {
1355 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1357 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1360 mptcp_subflow_set_active(subflow);
1362 return __mptcp_subflow_active(subflow);
1365 #define SSK_MODE_ACTIVE 0
1366 #define SSK_MODE_BACKUP 1
1367 #define SSK_MODE_MAX 2
1369 /* implement the mptcp packet scheduler;
1370 * returns the subflow that will transmit the next DSS
1371 * additionally updates the rtx timeout
1373 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1375 struct subflow_send_info send_info[SSK_MODE_MAX];
1376 struct mptcp_subflow_context *subflow;
1377 struct sock *sk = (struct sock *)msk;
1378 u32 pace, burst, wmem;
1379 int i, nr_active = 0;
1384 msk_owned_by_me(msk);
1386 if (__mptcp_check_fallback(msk)) {
1389 return __tcp_can_send(msk->first) &&
1390 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1393 /* re-use last subflow, if the burst allow that */
1394 if (msk->last_snd && msk->snd_burst > 0 &&
1395 sk_stream_memory_free(msk->last_snd) &&
1396 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1397 mptcp_set_timeout(sk);
1398 return msk->last_snd;
1401 /* pick the subflow with the lower wmem/wspace ratio */
1402 for (i = 0; i < SSK_MODE_MAX; ++i) {
1403 send_info[i].ssk = NULL;
1404 send_info[i].linger_time = -1;
1407 mptcp_for_each_subflow(msk, subflow) {
1408 trace_mptcp_subflow_get_send(subflow);
1409 ssk = mptcp_subflow_tcp_sock(subflow);
1410 if (!mptcp_subflow_active(subflow))
1413 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1414 nr_active += !subflow->backup;
1415 pace = subflow->avg_pacing_rate;
1416 if (unlikely(!pace)) {
1417 /* init pacing rate from socket */
1418 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1419 pace = subflow->avg_pacing_rate;
1424 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1425 if (linger_time < send_info[subflow->backup].linger_time) {
1426 send_info[subflow->backup].ssk = ssk;
1427 send_info[subflow->backup].linger_time = linger_time;
1430 __mptcp_set_timeout(sk, tout);
1432 /* pick the best backup if no other subflow is active */
1434 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1436 /* According to the blest algorithm, to avoid HoL blocking for the
1437 * faster flow, we need to:
1438 * - estimate the faster flow linger time
1439 * - use the above to estimate the amount of byte transferred
1440 * by the faster flow
1441 * - check that the amount of queued data is greter than the above,
1442 * otherwise do not use the picked, slower, subflow
1443 * We select the subflow with the shorter estimated time to flush
1444 * the queued mem, which basically ensure the above. We just need
1445 * to check that subflow has a non empty cwin.
1447 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1448 if (!ssk || !sk_stream_memory_free(ssk))
1451 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1452 wmem = READ_ONCE(ssk->sk_wmem_queued);
1454 msk->last_snd = NULL;
1458 subflow = mptcp_subflow_ctx(ssk);
1459 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1460 READ_ONCE(ssk->sk_pacing_rate) * burst,
1462 msk->last_snd = ssk;
1463 msk->snd_burst = burst;
1467 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1469 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1473 static void mptcp_update_post_push(struct mptcp_sock *msk,
1474 struct mptcp_data_frag *dfrag,
1477 u64 snd_nxt_new = dfrag->data_seq;
1479 dfrag->already_sent += sent;
1481 msk->snd_burst -= sent;
1483 snd_nxt_new += dfrag->already_sent;
1485 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1486 * is recovering after a failover. In that event, this re-sends
1489 * Thus compute snd_nxt_new candidate based on
1490 * the dfrag->data_seq that was sent and the data
1491 * that has been handed to the subflow for transmission
1492 * and skip update in case it was old dfrag.
1494 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1495 msk->snd_nxt = snd_nxt_new;
1498 void mptcp_check_and_set_pending(struct sock *sk)
1500 if (mptcp_send_head(sk))
1501 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1504 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1506 struct sock *prev_ssk = NULL, *ssk = NULL;
1507 struct mptcp_sock *msk = mptcp_sk(sk);
1508 struct mptcp_sendmsg_info info = {
1511 bool do_check_data_fin = false;
1512 struct mptcp_data_frag *dfrag;
1515 while ((dfrag = mptcp_send_head(sk))) {
1516 info.sent = dfrag->already_sent;
1517 info.limit = dfrag->data_len;
1518 len = dfrag->data_len - dfrag->already_sent;
1523 ssk = mptcp_subflow_get_send(msk);
1525 /* First check. If the ssk has changed since
1526 * the last round, release prev_ssk
1528 if (ssk != prev_ssk && prev_ssk)
1529 mptcp_push_release(prev_ssk, &info);
1533 /* Need to lock the new subflow only if different
1534 * from the previous one, otherwise we are still
1535 * helding the relevant lock
1537 if (ssk != prev_ssk)
1540 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1544 mptcp_push_release(ssk, &info);
1548 do_check_data_fin = true;
1552 mptcp_update_post_push(msk, dfrag, ret);
1554 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1557 /* at this point we held the socket lock for the last subflow we used */
1559 mptcp_push_release(ssk, &info);
1562 /* ensure the rtx timer is running */
1563 if (!mptcp_timer_pending(sk))
1564 mptcp_reset_timer(sk);
1565 if (do_check_data_fin)
1566 mptcp_check_send_data_fin(sk);
1569 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1571 struct mptcp_sock *msk = mptcp_sk(sk);
1572 struct mptcp_sendmsg_info info = {
1573 .data_lock_held = true,
1575 struct mptcp_data_frag *dfrag;
1576 struct sock *xmit_ssk;
1577 int len, copied = 0;
1580 while ((dfrag = mptcp_send_head(sk))) {
1581 info.sent = dfrag->already_sent;
1582 info.limit = dfrag->data_len;
1583 len = dfrag->data_len - dfrag->already_sent;
1587 /* check for a different subflow usage only after
1588 * spooling the first chunk of data
1590 xmit_ssk = first ? ssk : mptcp_subflow_get_send(msk);
1593 if (xmit_ssk != ssk) {
1594 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1595 MPTCP_DELEGATE_SEND);
1599 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1608 mptcp_update_post_push(msk, dfrag, ret);
1610 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1614 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1615 * not going to flush it via release_sock()
1618 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1620 if (!mptcp_timer_pending(sk))
1621 mptcp_reset_timer(sk);
1623 if (msk->snd_data_fin_enable &&
1624 msk->snd_nxt + 1 == msk->write_seq)
1625 mptcp_schedule_work(sk);
1629 static void mptcp_set_nospace(struct sock *sk)
1631 /* enable autotune */
1632 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1634 /* will be cleared on avail space */
1635 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1638 static int mptcp_disconnect(struct sock *sk, int flags);
1640 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1641 size_t len, int *copied_syn)
1643 unsigned int saved_flags = msg->msg_flags;
1644 struct mptcp_sock *msk = mptcp_sk(sk);
1645 struct socket *ssock;
1649 /* on flags based fastopen the mptcp is supposed to create the
1650 * first subflow right now. Otherwise we are in the defer_connect
1651 * path, and the first subflow must be already present.
1652 * Since the defer_connect flag is cleared after the first succsful
1653 * fastopen attempt, no need to check for additional subflow status.
1655 if (msg->msg_flags & MSG_FASTOPEN) {
1656 ssock = __mptcp_nmpc_socket(msk);
1658 return PTR_ERR(ssock);
1666 msg->msg_flags |= MSG_DONTWAIT;
1667 msk->fastopening = 1;
1668 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1669 msk->fastopening = 0;
1670 msg->msg_flags = saved_flags;
1673 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1674 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1675 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1676 msg->msg_namelen, msg->msg_flags, 1);
1678 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1679 * case of any error, except timeout or signal
1681 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1683 } else if (ret && ret != -EINPROGRESS) {
1684 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1685 * __inet_stream_connect() can fail, due to looking check,
1686 * see mptcp_disconnect().
1687 * Attempt it again outside the problematic scope.
1689 if (!mptcp_disconnect(sk, 0))
1690 sk->sk_socket->state = SS_UNCONNECTED;
1692 inet_sk(sk)->defer_connect = 0;
1697 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1699 struct mptcp_sock *msk = mptcp_sk(sk);
1700 struct page_frag *pfrag;
1705 /* silently ignore everything else */
1706 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1710 if (unlikely(inet_sk(sk)->defer_connect || msg->msg_flags & MSG_FASTOPEN)) {
1713 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1714 copied += copied_syn;
1715 if (ret == -EINPROGRESS && copied_syn > 0)
1721 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1723 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1724 ret = sk_stream_wait_connect(sk, &timeo);
1730 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1733 pfrag = sk_page_frag(sk);
1735 while (msg_data_left(msg)) {
1736 int total_ts, frag_truesize = 0;
1737 struct mptcp_data_frag *dfrag;
1738 bool dfrag_collapsed;
1739 size_t psize, offset;
1741 /* reuse tail pfrag, if possible, or carve a new one from the
1744 dfrag = mptcp_pending_tail(sk);
1745 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1746 if (!dfrag_collapsed) {
1747 if (!sk_stream_memory_free(sk))
1748 goto wait_for_memory;
1750 if (!mptcp_page_frag_refill(sk, pfrag))
1751 goto wait_for_memory;
1753 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1754 frag_truesize = dfrag->overhead;
1757 /* we do not bound vs wspace, to allow a single packet.
1758 * memory accounting will prevent execessive memory usage
1761 offset = dfrag->offset + dfrag->data_len;
1762 psize = pfrag->size - offset;
1763 psize = min_t(size_t, psize, msg_data_left(msg));
1764 total_ts = psize + frag_truesize;
1766 if (!sk_wmem_schedule(sk, total_ts))
1767 goto wait_for_memory;
1769 if (copy_page_from_iter(dfrag->page, offset, psize,
1770 &msg->msg_iter) != psize) {
1775 /* data successfully copied into the write queue */
1776 sk->sk_forward_alloc -= total_ts;
1778 dfrag->data_len += psize;
1779 frag_truesize += psize;
1780 pfrag->offset += frag_truesize;
1781 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1783 /* charge data on mptcp pending queue to the msk socket
1784 * Note: we charge such data both to sk and ssk
1786 sk_wmem_queued_add(sk, frag_truesize);
1787 if (!dfrag_collapsed) {
1788 get_page(dfrag->page);
1789 list_add_tail(&dfrag->list, &msk->rtx_queue);
1790 if (!msk->first_pending)
1791 WRITE_ONCE(msk->first_pending, dfrag);
1793 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1794 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1800 mptcp_set_nospace(sk);
1801 __mptcp_push_pending(sk, msg->msg_flags);
1802 ret = sk_stream_wait_memory(sk, &timeo);
1808 __mptcp_push_pending(sk, msg->msg_flags);
1818 copied = sk_stream_error(sk, msg->msg_flags, ret);
1822 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1824 size_t len, int flags,
1825 struct scm_timestamping_internal *tss,
1828 struct sk_buff *skb, *tmp;
1831 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1832 u32 offset = MPTCP_SKB_CB(skb)->offset;
1833 u32 data_len = skb->len - offset;
1834 u32 count = min_t(size_t, len - copied, data_len);
1837 if (!(flags & MSG_TRUNC)) {
1838 err = skb_copy_datagram_msg(skb, offset, msg, count);
1839 if (unlikely(err < 0)) {
1846 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1847 tcp_update_recv_tstamps(skb, tss);
1848 *cmsg_flags |= MPTCP_CMSG_TS;
1853 if (count < data_len) {
1854 if (!(flags & MSG_PEEK)) {
1855 MPTCP_SKB_CB(skb)->offset += count;
1856 MPTCP_SKB_CB(skb)->map_seq += count;
1861 if (!(flags & MSG_PEEK)) {
1862 /* we will bulk release the skb memory later */
1863 skb->destructor = NULL;
1864 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1865 __skb_unlink(skb, &msk->receive_queue);
1876 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1878 * Only difference: Use highest rtt estimate of the subflows in use.
1880 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1882 struct mptcp_subflow_context *subflow;
1883 struct sock *sk = (struct sock *)msk;
1884 u32 time, advmss = 1;
1887 msk_owned_by_me(msk);
1892 msk->rcvq_space.copied += copied;
1894 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1895 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1897 rtt_us = msk->rcvq_space.rtt_us;
1898 if (rtt_us && time < (rtt_us >> 3))
1902 mptcp_for_each_subflow(msk, subflow) {
1903 const struct tcp_sock *tp;
1907 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1909 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1910 sf_advmss = READ_ONCE(tp->advmss);
1912 rtt_us = max(sf_rtt_us, rtt_us);
1913 advmss = max(sf_advmss, advmss);
1916 msk->rcvq_space.rtt_us = rtt_us;
1917 if (time < (rtt_us >> 3) || rtt_us == 0)
1920 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1923 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1924 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1928 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1930 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1932 do_div(grow, msk->rcvq_space.space);
1933 rcvwin += (grow << 1);
1935 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1936 while (tcp_win_from_space(sk, rcvmem) < advmss)
1939 do_div(rcvwin, advmss);
1940 rcvbuf = min_t(u64, rcvwin * rcvmem,
1941 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1943 if (rcvbuf > sk->sk_rcvbuf) {
1946 window_clamp = tcp_win_from_space(sk, rcvbuf);
1947 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1949 /* Make subflows follow along. If we do not do this, we
1950 * get drops at subflow level if skbs can't be moved to
1951 * the mptcp rx queue fast enough (announced rcv_win can
1952 * exceed ssk->sk_rcvbuf).
1954 mptcp_for_each_subflow(msk, subflow) {
1958 ssk = mptcp_subflow_tcp_sock(subflow);
1959 slow = lock_sock_fast(ssk);
1960 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1961 tcp_sk(ssk)->window_clamp = window_clamp;
1962 tcp_cleanup_rbuf(ssk, 1);
1963 unlock_sock_fast(ssk, slow);
1968 msk->rcvq_space.space = msk->rcvq_space.copied;
1970 msk->rcvq_space.copied = 0;
1971 msk->rcvq_space.time = mstamp;
1974 static void __mptcp_update_rmem(struct sock *sk)
1976 struct mptcp_sock *msk = mptcp_sk(sk);
1978 if (!msk->rmem_released)
1981 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1982 mptcp_rmem_uncharge(sk, msk->rmem_released);
1983 WRITE_ONCE(msk->rmem_released, 0);
1986 static void __mptcp_splice_receive_queue(struct sock *sk)
1988 struct mptcp_sock *msk = mptcp_sk(sk);
1990 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1993 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1995 struct sock *sk = (struct sock *)msk;
1996 unsigned int moved = 0;
2000 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2003 /* we can have data pending in the subflows only if the msk
2004 * receive buffer was full at subflow_data_ready() time,
2005 * that is an unlikely slow path.
2010 slowpath = lock_sock_fast(ssk);
2011 mptcp_data_lock(sk);
2012 __mptcp_update_rmem(sk);
2013 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2014 mptcp_data_unlock(sk);
2016 if (unlikely(ssk->sk_err))
2017 __mptcp_error_report(sk);
2018 unlock_sock_fast(ssk, slowpath);
2021 /* acquire the data lock only if some input data is pending */
2023 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2024 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2025 mptcp_data_lock(sk);
2026 __mptcp_update_rmem(sk);
2027 ret |= __mptcp_ofo_queue(msk);
2028 __mptcp_splice_receive_queue(sk);
2029 mptcp_data_unlock(sk);
2032 mptcp_check_data_fin((struct sock *)msk);
2033 return !skb_queue_empty(&msk->receive_queue);
2036 static unsigned int mptcp_inq_hint(const struct sock *sk)
2038 const struct mptcp_sock *msk = mptcp_sk(sk);
2039 const struct sk_buff *skb;
2041 skb = skb_peek(&msk->receive_queue);
2043 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2045 if (hint_val >= INT_MAX)
2048 return (unsigned int)hint_val;
2051 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2057 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2058 int flags, int *addr_len)
2060 struct mptcp_sock *msk = mptcp_sk(sk);
2061 struct scm_timestamping_internal tss;
2062 int copied = 0, cmsg_flags = 0;
2066 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2067 if (unlikely(flags & MSG_ERRQUEUE))
2068 return inet_recv_error(sk, msg, len, addr_len);
2071 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2076 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2078 len = min_t(size_t, len, INT_MAX);
2079 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2081 if (unlikely(msk->recvmsg_inq))
2082 cmsg_flags = MPTCP_CMSG_INQ;
2084 while (copied < len) {
2087 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2088 if (unlikely(bytes_read < 0)) {
2090 copied = bytes_read;
2094 copied += bytes_read;
2096 /* be sure to advertise window change */
2097 mptcp_cleanup_rbuf(msk);
2099 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2102 /* only the master socket status is relevant here. The exit
2103 * conditions mirror closely tcp_recvmsg()
2105 if (copied >= target)
2110 sk->sk_state == TCP_CLOSE ||
2111 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2113 signal_pending(current))
2117 copied = sock_error(sk);
2121 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2122 /* race breaker: the shutdown could be after the
2123 * previous receive queue check
2125 if (__mptcp_move_skbs(msk))
2130 if (sk->sk_state == TCP_CLOSE) {
2140 if (signal_pending(current)) {
2141 copied = sock_intr_errno(timeo);
2146 pr_debug("block timeout %ld", timeo);
2147 sk_wait_data(sk, &timeo, NULL);
2151 if (cmsg_flags && copied >= 0) {
2152 if (cmsg_flags & MPTCP_CMSG_TS)
2153 tcp_recv_timestamp(msg, sk, &tss);
2155 if (cmsg_flags & MPTCP_CMSG_INQ) {
2156 unsigned int inq = mptcp_inq_hint(sk);
2158 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2162 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2163 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2164 skb_queue_empty(&msk->receive_queue), copied);
2165 if (!(flags & MSG_PEEK))
2166 mptcp_rcv_space_adjust(msk, copied);
2172 static void mptcp_retransmit_timer(struct timer_list *t)
2174 struct inet_connection_sock *icsk = from_timer(icsk, t,
2175 icsk_retransmit_timer);
2176 struct sock *sk = &icsk->icsk_inet.sk;
2177 struct mptcp_sock *msk = mptcp_sk(sk);
2180 if (!sock_owned_by_user(sk)) {
2181 /* we need a process context to retransmit */
2182 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2183 mptcp_schedule_work(sk);
2185 /* delegate our work to tcp_release_cb() */
2186 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2192 static void mptcp_timeout_timer(struct timer_list *t)
2194 struct sock *sk = from_timer(sk, t, sk_timer);
2196 mptcp_schedule_work(sk);
2200 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2203 * A backup subflow is returned only if that is the only kind available.
2205 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2207 struct sock *backup = NULL, *pick = NULL;
2208 struct mptcp_subflow_context *subflow;
2209 int min_stale_count = INT_MAX;
2211 msk_owned_by_me(msk);
2213 if (__mptcp_check_fallback(msk))
2216 mptcp_for_each_subflow(msk, subflow) {
2217 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2219 if (!__mptcp_subflow_active(subflow))
2222 /* still data outstanding at TCP level? skip this */
2223 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2224 mptcp_pm_subflow_chk_stale(msk, ssk);
2225 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2229 if (subflow->backup) {
2242 /* use backup only if there are no progresses anywhere */
2243 return min_stale_count > 1 ? backup : NULL;
2246 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2249 iput(SOCK_INODE(msk->subflow));
2250 WRITE_ONCE(msk->subflow, NULL);
2254 bool __mptcp_retransmit_pending_data(struct sock *sk)
2256 struct mptcp_data_frag *cur, *rtx_head;
2257 struct mptcp_sock *msk = mptcp_sk(sk);
2259 if (__mptcp_check_fallback(msk))
2262 if (tcp_rtx_and_write_queues_empty(sk))
2265 /* the closing socket has some data untransmitted and/or unacked:
2266 * some data in the mptcp rtx queue has not really xmitted yet.
2267 * keep it simple and re-inject the whole mptcp level rtx queue
2269 mptcp_data_lock(sk);
2270 __mptcp_clean_una_wakeup(sk);
2271 rtx_head = mptcp_rtx_head(sk);
2273 mptcp_data_unlock(sk);
2277 msk->recovery_snd_nxt = msk->snd_nxt;
2278 msk->recovery = true;
2279 mptcp_data_unlock(sk);
2281 msk->first_pending = rtx_head;
2284 /* be sure to clear the "sent status" on all re-injected fragments */
2285 list_for_each_entry(cur, &msk->rtx_queue, list) {
2286 if (!cur->already_sent)
2288 cur->already_sent = 0;
2294 /* flags for __mptcp_close_ssk() */
2295 #define MPTCP_CF_PUSH BIT(1)
2296 #define MPTCP_CF_FASTCLOSE BIT(2)
2298 /* subflow sockets can be either outgoing (connect) or incoming
2301 * Outgoing subflows use in-kernel sockets.
2302 * Incoming subflows do not have their own 'struct socket' allocated,
2303 * so we need to use tcp_close() after detaching them from the mptcp
2306 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2307 struct mptcp_subflow_context *subflow,
2310 struct mptcp_sock *msk = mptcp_sk(sk);
2311 bool dispose_it, need_push = false;
2313 /* If the first subflow moved to a close state before accept, e.g. due
2314 * to an incoming reset, mptcp either:
2315 * - if either the subflow or the msk are dead, destroy the context
2316 * (the subflow socket is deleted by inet_child_forget) and the msk
2317 * - otherwise do nothing at the moment and take action at accept and/or
2318 * listener shutdown - user-space must be able to accept() the closed
2321 if (msk->in_accept_queue && msk->first == ssk) {
2322 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2325 /* ensure later check in mptcp_worker() will dispose the msk */
2326 sock_set_flag(sk, SOCK_DEAD);
2327 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2328 mptcp_subflow_drop_ctx(ssk);
2332 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2334 list_del(&subflow->node);
2336 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2338 if (flags & MPTCP_CF_FASTCLOSE) {
2339 /* be sure to force the tcp_disconnect() path,
2340 * to generate the egress reset
2342 ssk->sk_lingertime = 0;
2343 sock_set_flag(ssk, SOCK_LINGER);
2344 subflow->send_fastclose = 1;
2347 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2349 /* The MPTCP code never wait on the subflow sockets, TCP-level
2350 * disconnect should never fail
2352 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2353 msk->subflow->state = SS_UNCONNECTED;
2354 mptcp_subflow_ctx_reset(subflow);
2360 subflow->disposable = 1;
2362 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2363 * the ssk has been already destroyed, we just need to release the
2364 * reference owned by msk;
2366 if (!inet_csk(ssk)->icsk_ulp_ops) {
2367 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2368 kfree_rcu(subflow, rcu);
2370 /* otherwise tcp will dispose of the ssk and subflow ctx */
2371 __tcp_close(ssk, 0);
2373 /* close acquired an extra ref */
2382 if (ssk == msk->first)
2383 WRITE_ONCE(msk->first, NULL);
2386 if (ssk == msk->last_snd)
2387 msk->last_snd = NULL;
2390 __mptcp_push_pending(sk, 0);
2393 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2394 struct mptcp_subflow_context *subflow)
2396 if (sk->sk_state == TCP_ESTABLISHED)
2397 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2399 /* subflow aborted before reaching the fully_established status
2400 * attempt the creation of the next subflow
2402 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2404 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2407 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2412 static void __mptcp_close_subflow(struct sock *sk)
2414 struct mptcp_subflow_context *subflow, *tmp;
2415 struct mptcp_sock *msk = mptcp_sk(sk);
2419 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2420 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2422 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2425 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2426 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2429 mptcp_close_ssk(sk, ssk, subflow);
2434 static bool mptcp_should_close(const struct sock *sk)
2436 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2437 struct mptcp_subflow_context *subflow;
2439 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2442 /* if all subflows are in closed status don't bother with additional
2445 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2446 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2453 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2455 struct mptcp_subflow_context *subflow, *tmp;
2456 struct sock *sk = (struct sock *)msk;
2458 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2461 mptcp_token_destroy(msk);
2463 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2464 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2467 slow = lock_sock_fast(tcp_sk);
2468 if (tcp_sk->sk_state != TCP_CLOSE) {
2469 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2470 tcp_set_state(tcp_sk, TCP_CLOSE);
2472 unlock_sock_fast(tcp_sk, slow);
2475 /* Mirror the tcp_reset() error propagation */
2476 switch (sk->sk_state) {
2478 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2480 case TCP_CLOSE_WAIT:
2481 WRITE_ONCE(sk->sk_err, EPIPE);
2486 WRITE_ONCE(sk->sk_err, ECONNRESET);
2489 inet_sk_state_store(sk, TCP_CLOSE);
2490 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2491 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2492 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2494 /* the calling mptcp_worker will properly destroy the socket */
2495 if (sock_flag(sk, SOCK_DEAD))
2498 sk->sk_state_change(sk);
2499 sk_error_report(sk);
2502 static void __mptcp_retrans(struct sock *sk)
2504 struct mptcp_sock *msk = mptcp_sk(sk);
2505 struct mptcp_sendmsg_info info = {};
2506 struct mptcp_data_frag *dfrag;
2511 mptcp_clean_una_wakeup(sk);
2513 /* first check ssk: need to kick "stale" logic */
2514 ssk = mptcp_subflow_get_retrans(msk);
2515 dfrag = mptcp_rtx_head(sk);
2517 if (mptcp_data_fin_enabled(msk)) {
2518 struct inet_connection_sock *icsk = inet_csk(sk);
2520 icsk->icsk_retransmits++;
2521 mptcp_set_datafin_timeout(sk);
2522 mptcp_send_ack(msk);
2527 if (!mptcp_send_head(sk))
2538 /* limit retransmission to the bytes already sent on some subflows */
2540 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2541 while (info.sent < info.limit) {
2542 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2546 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2551 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2552 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2554 WRITE_ONCE(msk->allow_infinite_fallback, false);
2560 mptcp_check_and_set_pending(sk);
2562 if (!mptcp_timer_pending(sk))
2563 mptcp_reset_timer(sk);
2566 /* schedule the timeout timer for the relevant event: either close timeout
2567 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2569 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2571 struct sock *sk = (struct sock *)msk;
2572 unsigned long timeout, close_timeout;
2574 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2577 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2579 /* the close timeout takes precedence on the fail one, and here at least one of
2582 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2584 sk_reset_timer(sk, &sk->sk_timer, timeout);
2587 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2589 struct sock *ssk = msk->first;
2595 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2597 slow = lock_sock_fast(ssk);
2598 mptcp_subflow_reset(ssk);
2599 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2600 unlock_sock_fast(ssk, slow);
2602 mptcp_reset_timeout(msk, 0);
2605 static void mptcp_do_fastclose(struct sock *sk)
2607 struct mptcp_subflow_context *subflow, *tmp;
2608 struct mptcp_sock *msk = mptcp_sk(sk);
2610 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2611 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2612 subflow, MPTCP_CF_FASTCLOSE);
2615 static void mptcp_worker(struct work_struct *work)
2617 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2618 struct sock *sk = (struct sock *)msk;
2619 unsigned long fail_tout;
2623 state = sk->sk_state;
2624 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2627 mptcp_check_fastclose(msk);
2629 mptcp_pm_nl_work(msk);
2631 mptcp_check_send_data_fin(sk);
2632 mptcp_check_data_fin_ack(sk);
2633 mptcp_check_data_fin(sk);
2635 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2636 __mptcp_close_subflow(sk);
2638 /* There is no point in keeping around an orphaned sk timedout or
2639 * closed, but we need the msk around to reply to incoming DATA_FIN,
2640 * even if it is orphaned and in FIN_WAIT2 state
2642 if (sock_flag(sk, SOCK_DEAD)) {
2643 if (mptcp_should_close(sk)) {
2644 inet_sk_state_store(sk, TCP_CLOSE);
2645 mptcp_do_fastclose(sk);
2647 if (sk->sk_state == TCP_CLOSE) {
2648 __mptcp_destroy_sock(sk);
2653 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2654 __mptcp_retrans(sk);
2656 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2657 if (fail_tout && time_after(jiffies, fail_tout))
2658 mptcp_mp_fail_no_response(msk);
2665 static int __mptcp_init_sock(struct sock *sk)
2667 struct mptcp_sock *msk = mptcp_sk(sk);
2669 INIT_LIST_HEAD(&msk->conn_list);
2670 INIT_LIST_HEAD(&msk->join_list);
2671 INIT_LIST_HEAD(&msk->rtx_queue);
2672 INIT_WORK(&msk->work, mptcp_worker);
2673 __skb_queue_head_init(&msk->receive_queue);
2674 msk->out_of_order_queue = RB_ROOT;
2675 msk->first_pending = NULL;
2676 msk->rmem_fwd_alloc = 0;
2677 WRITE_ONCE(msk->rmem_released, 0);
2678 msk->timer_ival = TCP_RTO_MIN;
2680 WRITE_ONCE(msk->first, NULL);
2681 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2682 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2683 WRITE_ONCE(msk->allow_infinite_fallback, true);
2684 msk->recovery = false;
2686 mptcp_pm_data_init(msk);
2688 /* re-use the csk retrans timer for MPTCP-level retrans */
2689 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2690 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2695 static void mptcp_ca_reset(struct sock *sk)
2697 struct inet_connection_sock *icsk = inet_csk(sk);
2699 tcp_assign_congestion_control(sk);
2700 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2702 /* no need to keep a reference to the ops, the name will suffice */
2703 tcp_cleanup_congestion_control(sk);
2704 icsk->icsk_ca_ops = NULL;
2707 static int mptcp_init_sock(struct sock *sk)
2709 struct net *net = sock_net(sk);
2712 ret = __mptcp_init_sock(sk);
2716 if (!mptcp_is_enabled(net))
2717 return -ENOPROTOOPT;
2719 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2722 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2724 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2725 * propagate the correct value
2729 sk_sockets_allocated_inc(sk);
2730 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2731 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2736 static void __mptcp_clear_xmit(struct sock *sk)
2738 struct mptcp_sock *msk = mptcp_sk(sk);
2739 struct mptcp_data_frag *dtmp, *dfrag;
2741 WRITE_ONCE(msk->first_pending, NULL);
2742 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2743 dfrag_clear(sk, dfrag);
2746 void mptcp_cancel_work(struct sock *sk)
2748 struct mptcp_sock *msk = mptcp_sk(sk);
2750 if (cancel_work_sync(&msk->work))
2754 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2758 switch (ssk->sk_state) {
2760 if (!(how & RCV_SHUTDOWN))
2764 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2767 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2768 pr_debug("Fallback");
2769 ssk->sk_shutdown |= how;
2770 tcp_shutdown(ssk, how);
2772 /* simulate the data_fin ack reception to let the state
2773 * machine move forward
2775 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2776 mptcp_schedule_work(sk);
2778 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2780 if (!mptcp_timer_pending(sk))
2781 mptcp_reset_timer(sk);
2789 static const unsigned char new_state[16] = {
2790 /* current state: new state: action: */
2791 [0 /* (Invalid) */] = TCP_CLOSE,
2792 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2793 [TCP_SYN_SENT] = TCP_CLOSE,
2794 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2795 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2796 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2797 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2798 [TCP_CLOSE] = TCP_CLOSE,
2799 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2800 [TCP_LAST_ACK] = TCP_LAST_ACK,
2801 [TCP_LISTEN] = TCP_CLOSE,
2802 [TCP_CLOSING] = TCP_CLOSING,
2803 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2806 static int mptcp_close_state(struct sock *sk)
2808 int next = (int)new_state[sk->sk_state];
2809 int ns = next & TCP_STATE_MASK;
2811 inet_sk_state_store(sk, ns);
2813 return next & TCP_ACTION_FIN;
2816 static void mptcp_check_send_data_fin(struct sock *sk)
2818 struct mptcp_subflow_context *subflow;
2819 struct mptcp_sock *msk = mptcp_sk(sk);
2821 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2822 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2823 msk->snd_nxt, msk->write_seq);
2825 /* we still need to enqueue subflows or not really shutting down,
2828 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2829 mptcp_send_head(sk))
2832 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2834 mptcp_for_each_subflow(msk, subflow) {
2835 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2837 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2841 static void __mptcp_wr_shutdown(struct sock *sk)
2843 struct mptcp_sock *msk = mptcp_sk(sk);
2845 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2846 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2847 !!mptcp_send_head(sk));
2849 /* will be ignored by fallback sockets */
2850 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2851 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2853 mptcp_check_send_data_fin(sk);
2856 static void __mptcp_destroy_sock(struct sock *sk)
2858 struct mptcp_sock *msk = mptcp_sk(sk);
2860 pr_debug("msk=%p", msk);
2864 mptcp_stop_timer(sk);
2865 sk_stop_timer(sk, &sk->sk_timer);
2868 sk->sk_prot->destroy(sk);
2870 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2871 WARN_ON_ONCE(msk->rmem_released);
2872 sk_stream_kill_queues(sk);
2873 xfrm_sk_free_policy(sk);
2878 void __mptcp_unaccepted_force_close(struct sock *sk)
2880 sock_set_flag(sk, SOCK_DEAD);
2881 inet_sk_state_store(sk, TCP_CLOSE);
2882 mptcp_do_fastclose(sk);
2883 __mptcp_destroy_sock(sk);
2886 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2888 /* Concurrent splices from sk_receive_queue into receive_queue will
2889 * always show at least one non-empty queue when checked in this order.
2891 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2892 skb_queue_empty_lockless(&msk->receive_queue))
2895 return EPOLLIN | EPOLLRDNORM;
2898 static void mptcp_check_listen_stop(struct sock *sk)
2902 if (inet_sk_state_load(sk) != TCP_LISTEN)
2905 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2906 ssk = mptcp_sk(sk)->first;
2907 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2910 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2911 mptcp_subflow_queue_clean(sk, ssk);
2912 inet_csk_listen_stop(ssk);
2913 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2914 tcp_set_state(ssk, TCP_CLOSE);
2918 bool __mptcp_close(struct sock *sk, long timeout)
2920 struct mptcp_subflow_context *subflow;
2921 struct mptcp_sock *msk = mptcp_sk(sk);
2922 bool do_cancel_work = false;
2923 int subflows_alive = 0;
2925 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2927 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2928 mptcp_check_listen_stop(sk);
2929 inet_sk_state_store(sk, TCP_CLOSE);
2933 if (mptcp_check_readable(msk) || timeout < 0) {
2934 /* If the msk has read data, or the caller explicitly ask it,
2935 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
2937 inet_sk_state_store(sk, TCP_CLOSE);
2938 mptcp_do_fastclose(sk);
2940 } else if (mptcp_close_state(sk)) {
2941 __mptcp_wr_shutdown(sk);
2944 sk_stream_wait_close(sk, timeout);
2947 /* orphan all the subflows */
2948 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2949 mptcp_for_each_subflow(msk, subflow) {
2950 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2951 bool slow = lock_sock_fast_nested(ssk);
2953 subflows_alive += ssk->sk_state != TCP_CLOSE;
2955 /* since the close timeout takes precedence on the fail one,
2958 if (ssk == msk->first)
2959 subflow->fail_tout = 0;
2961 /* detach from the parent socket, but allow data_ready to
2962 * push incoming data into the mptcp stack, to properly ack it
2964 ssk->sk_socket = NULL;
2966 unlock_sock_fast(ssk, slow);
2970 /* all the subflows are closed, only timeout can change the msk
2971 * state, let's not keep resources busy for no reasons
2973 if (subflows_alive == 0)
2974 inet_sk_state_store(sk, TCP_CLOSE);
2977 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2979 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2981 if (sk->sk_state == TCP_CLOSE) {
2982 __mptcp_destroy_sock(sk);
2983 do_cancel_work = true;
2985 mptcp_reset_timeout(msk, 0);
2988 return do_cancel_work;
2991 static void mptcp_close(struct sock *sk, long timeout)
2993 bool do_cancel_work;
2997 do_cancel_work = __mptcp_close(sk, timeout);
3000 mptcp_cancel_work(sk);
3005 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3007 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3008 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3009 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3011 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3012 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3015 msk6->saddr = ssk6->saddr;
3016 msk6->flow_label = ssk6->flow_label;
3020 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3021 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3022 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3023 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3024 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3025 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3028 static int mptcp_disconnect(struct sock *sk, int flags)
3030 struct mptcp_sock *msk = mptcp_sk(sk);
3032 /* Deny disconnect if other threads are blocked in sk_wait_event()
3033 * or inet_wait_for_connect().
3035 if (sk->sk_wait_pending)
3038 /* We are on the fastopen error path. We can't call straight into the
3039 * subflows cleanup code due to lock nesting (we are already under
3040 * msk->firstsocket lock).
3042 if (msk->fastopening)
3045 mptcp_check_listen_stop(sk);
3046 inet_sk_state_store(sk, TCP_CLOSE);
3048 mptcp_stop_timer(sk);
3049 sk_stop_timer(sk, &sk->sk_timer);
3052 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3054 /* msk->subflow is still intact, the following will not free the first
3057 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3058 msk->last_snd = NULL;
3059 WRITE_ONCE(msk->flags, 0);
3061 msk->push_pending = 0;
3062 msk->recovery = false;
3063 msk->can_ack = false;
3064 msk->fully_established = false;
3065 msk->rcv_data_fin = false;
3066 msk->snd_data_fin_enable = false;
3067 msk->rcv_fastclose = false;
3068 msk->use_64bit_ack = false;
3069 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3070 mptcp_pm_data_reset(msk);
3073 WRITE_ONCE(sk->sk_shutdown, 0);
3074 sk_error_report(sk);
3078 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3079 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3081 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3083 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3087 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3088 const struct mptcp_options_received *mp_opt,
3090 struct request_sock *req)
3092 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3093 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3094 struct mptcp_sock *msk;
3099 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3100 if (nsk->sk_family == AF_INET6)
3101 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3104 nsk->sk_wait_pending = 0;
3105 __mptcp_init_sock(nsk);
3107 msk = mptcp_sk(nsk);
3108 msk->local_key = subflow_req->local_key;
3109 msk->token = subflow_req->token;
3110 WRITE_ONCE(msk->subflow, NULL);
3111 msk->in_accept_queue = 1;
3112 WRITE_ONCE(msk->fully_established, false);
3113 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3114 WRITE_ONCE(msk->csum_enabled, true);
3116 msk->write_seq = subflow_req->idsn + 1;
3117 msk->snd_nxt = msk->write_seq;
3118 msk->snd_una = msk->write_seq;
3119 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3120 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3122 sock_reset_flag(nsk, SOCK_RCU_FREE);
3123 security_inet_csk_clone(nsk, req);
3125 /* this can't race with mptcp_close(), as the msk is
3126 * not yet exposted to user-space
3128 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3130 /* The msk maintain a ref to each subflow in the connections list */
3131 WRITE_ONCE(msk->first, ssk);
3132 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3135 /* new mpc subflow takes ownership of the newly
3136 * created mptcp socket
3138 mptcp_token_accept(subflow_req, msk);
3140 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3141 * uses the correct data
3143 mptcp_copy_inaddrs(nsk, ssk);
3144 mptcp_propagate_sndbuf(nsk, ssk);
3146 mptcp_rcv_space_init(msk, ssk);
3147 bh_unlock_sock(nsk);
3149 /* note: the newly allocated socket refcount is 2 now */
3153 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3155 const struct tcp_sock *tp = tcp_sk(ssk);
3157 msk->rcvq_space.copied = 0;
3158 msk->rcvq_space.rtt_us = 0;
3160 msk->rcvq_space.time = tp->tcp_mstamp;
3162 /* initial rcv_space offering made to peer */
3163 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3164 TCP_INIT_CWND * tp->advmss);
3165 if (msk->rcvq_space.space == 0)
3166 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3168 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3171 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3174 struct mptcp_sock *msk = mptcp_sk(sk);
3175 struct socket *listener;
3178 listener = READ_ONCE(msk->subflow);
3179 if (WARN_ON_ONCE(!listener)) {
3184 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3185 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3189 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3190 if (sk_is_mptcp(newsk)) {
3191 struct mptcp_subflow_context *subflow;
3192 struct sock *new_mptcp_sock;
3194 subflow = mptcp_subflow_ctx(newsk);
3195 new_mptcp_sock = subflow->conn;
3197 /* is_mptcp should be false if subflow->conn is missing, see
3198 * subflow_syn_recv_sock()
3200 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3201 tcp_sk(newsk)->is_mptcp = 0;
3205 newsk = new_mptcp_sock;
3206 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3208 MPTCP_INC_STATS(sock_net(sk),
3209 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3213 newsk->sk_kern_sock = kern;
3217 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3219 struct mptcp_subflow_context *subflow, *tmp;
3220 struct sock *sk = (struct sock *)msk;
3222 __mptcp_clear_xmit(sk);
3224 /* join list will be eventually flushed (with rst) at sock lock release time */
3225 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3226 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3228 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3229 mptcp_data_lock(sk);
3230 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3231 __skb_queue_purge(&sk->sk_receive_queue);
3232 skb_rbtree_purge(&msk->out_of_order_queue);
3233 mptcp_data_unlock(sk);
3235 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3236 * inet_sock_destruct() will dispose it
3238 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3239 msk->rmem_fwd_alloc = 0;
3240 mptcp_token_destroy(msk);
3241 mptcp_pm_free_anno_list(msk);
3242 mptcp_free_local_addr_list(msk);
3245 static void mptcp_destroy(struct sock *sk)
3247 struct mptcp_sock *msk = mptcp_sk(sk);
3249 /* clears msk->subflow, allowing the following to close
3250 * even the initial subflow
3252 mptcp_dispose_initial_subflow(msk);
3253 mptcp_destroy_common(msk, 0);
3254 sk_sockets_allocated_dec(sk);
3257 void __mptcp_data_acked(struct sock *sk)
3259 if (!sock_owned_by_user(sk))
3260 __mptcp_clean_una(sk);
3262 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3264 if (mptcp_pending_data_fin_ack(sk))
3265 mptcp_schedule_work(sk);
3268 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3270 if (!mptcp_send_head(sk))
3273 if (!sock_owned_by_user(sk))
3274 __mptcp_subflow_push_pending(sk, ssk, false);
3276 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3279 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3280 BIT(MPTCP_RETRANSMIT) | \
3281 BIT(MPTCP_FLUSH_JOIN_LIST))
3283 /* processes deferred events and flush wmem */
3284 static void mptcp_release_cb(struct sock *sk)
3285 __must_hold(&sk->sk_lock.slock)
3287 struct mptcp_sock *msk = mptcp_sk(sk);
3290 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3292 struct list_head join_list;
3297 INIT_LIST_HEAD(&join_list);
3298 list_splice_init(&msk->join_list, &join_list);
3300 /* the following actions acquire the subflow socket lock
3302 * 1) can't be invoked in atomic scope
3303 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3304 * datapath acquires the msk socket spinlock while helding
3305 * the subflow socket lock
3307 msk->push_pending = 0;
3308 msk->cb_flags &= ~flags;
3309 spin_unlock_bh(&sk->sk_lock.slock);
3311 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3312 __mptcp_flush_join_list(sk, &join_list);
3313 if (flags & BIT(MPTCP_PUSH_PENDING))
3314 __mptcp_push_pending(sk, 0);
3315 if (flags & BIT(MPTCP_RETRANSMIT))
3316 __mptcp_retrans(sk);
3319 spin_lock_bh(&sk->sk_lock.slock);
3322 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3323 __mptcp_clean_una_wakeup(sk);
3324 if (unlikely(&msk->cb_flags)) {
3325 /* be sure to set the current sk state before tacking actions
3326 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3328 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3329 __mptcp_set_connected(sk);
3330 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3331 __mptcp_error_report(sk);
3332 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3333 msk->last_snd = NULL;
3336 __mptcp_update_rmem(sk);
3339 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3340 * TCP can't schedule delack timer before the subflow is fully established.
3341 * MPTCP uses the delack timer to do 3rd ack retransmissions
3343 static void schedule_3rdack_retransmission(struct sock *ssk)
3345 struct inet_connection_sock *icsk = inet_csk(ssk);
3346 struct tcp_sock *tp = tcp_sk(ssk);
3347 unsigned long timeout;
3349 if (mptcp_subflow_ctx(ssk)->fully_established)
3352 /* reschedule with a timeout above RTT, as we must look only for drop */
3354 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3356 timeout = TCP_TIMEOUT_INIT;
3359 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3360 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3361 icsk->icsk_ack.timeout = timeout;
3362 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3365 void mptcp_subflow_process_delegated(struct sock *ssk)
3367 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3368 struct sock *sk = subflow->conn;
3370 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3371 mptcp_data_lock(sk);
3372 if (!sock_owned_by_user(sk))
3373 __mptcp_subflow_push_pending(sk, ssk, true);
3375 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3376 mptcp_data_unlock(sk);
3377 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3379 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3380 schedule_3rdack_retransmission(ssk);
3381 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3385 static int mptcp_hash(struct sock *sk)
3387 /* should never be called,
3388 * we hash the TCP subflows not the master socket
3394 static void mptcp_unhash(struct sock *sk)
3396 /* called from sk_common_release(), but nothing to do here */
3399 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3401 struct mptcp_sock *msk = mptcp_sk(sk);
3402 struct socket *ssock;
3404 ssock = msk->subflow;
3405 pr_debug("msk=%p, subflow=%p", msk, ssock);
3406 if (WARN_ON_ONCE(!ssock))
3409 return inet_csk_get_port(ssock->sk, snum);
3412 void mptcp_finish_connect(struct sock *ssk)
3414 struct mptcp_subflow_context *subflow;
3415 struct mptcp_sock *msk;
3418 subflow = mptcp_subflow_ctx(ssk);
3422 pr_debug("msk=%p, token=%u", sk, subflow->token);
3424 subflow->map_seq = subflow->iasn;
3425 subflow->map_subflow_seq = 1;
3427 /* the socket is not connected yet, no msk/subflow ops can access/race
3428 * accessing the field below
3430 WRITE_ONCE(msk->local_key, subflow->local_key);
3431 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3432 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3433 WRITE_ONCE(msk->snd_una, msk->write_seq);
3435 mptcp_pm_new_connection(msk, ssk, 0);
3437 mptcp_rcv_space_init(msk, ssk);
3440 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3442 write_lock_bh(&sk->sk_callback_lock);
3443 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3444 sk_set_socket(sk, parent);
3445 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3446 write_unlock_bh(&sk->sk_callback_lock);
3449 bool mptcp_finish_join(struct sock *ssk)
3451 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3452 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3453 struct sock *parent = (void *)msk;
3456 pr_debug("msk=%p, subflow=%p", msk, subflow);
3458 /* mptcp socket already closing? */
3459 if (!mptcp_is_fully_established(parent)) {
3460 subflow->reset_reason = MPTCP_RST_EMPTCP;
3464 /* active subflow, already present inside the conn_list */
3465 if (!list_empty(&subflow->node)) {
3466 mptcp_subflow_joined(msk, ssk);
3470 if (!mptcp_pm_allow_new_subflow(msk))
3471 goto err_prohibited;
3473 /* If we can't acquire msk socket lock here, let the release callback
3476 mptcp_data_lock(parent);
3477 if (!sock_owned_by_user(parent)) {
3478 ret = __mptcp_finish_join(msk, ssk);
3481 list_add_tail(&subflow->node, &msk->conn_list);
3485 list_add_tail(&subflow->node, &msk->join_list);
3486 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3488 mptcp_data_unlock(parent);
3492 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3499 static void mptcp_shutdown(struct sock *sk, int how)
3501 pr_debug("sk=%p, how=%d", sk, how);
3503 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3504 __mptcp_wr_shutdown(sk);
3507 static int mptcp_forward_alloc_get(const struct sock *sk)
3509 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3512 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3514 const struct sock *sk = (void *)msk;
3517 if (sk->sk_state == TCP_LISTEN)
3520 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3523 delta = msk->write_seq - v;
3524 if (__mptcp_check_fallback(msk) && msk->first) {
3525 struct tcp_sock *tp = tcp_sk(msk->first);
3527 /* the first subflow is disconnected after close - see
3528 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3529 * so ignore that status, too.
3531 if (!((1 << msk->first->sk_state) &
3532 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3533 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3535 if (delta > INT_MAX)
3541 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3543 struct mptcp_sock *msk = mptcp_sk(sk);
3549 if (sk->sk_state == TCP_LISTEN)
3553 __mptcp_move_skbs(msk);
3554 answ = mptcp_inq_hint(sk);
3558 slow = lock_sock_fast(sk);
3559 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3560 unlock_sock_fast(sk, slow);
3563 slow = lock_sock_fast(sk);
3564 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3565 unlock_sock_fast(sk, slow);
3568 return -ENOIOCTLCMD;
3571 return put_user(answ, (int __user *)arg);
3574 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3575 struct mptcp_subflow_context *subflow)
3577 subflow->request_mptcp = 0;
3578 __mptcp_do_fallback(msk);
3581 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3583 struct mptcp_subflow_context *subflow;
3584 struct mptcp_sock *msk = mptcp_sk(sk);
3585 struct socket *ssock;
3588 ssock = __mptcp_nmpc_socket(msk);
3590 return PTR_ERR(ssock);
3592 mptcp_token_destroy(msk);
3593 inet_sk_state_store(sk, TCP_SYN_SENT);
3594 subflow = mptcp_subflow_ctx(ssock->sk);
3595 #ifdef CONFIG_TCP_MD5SIG
3596 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3599 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3600 mptcp_subflow_early_fallback(msk, subflow);
3602 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3603 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3604 mptcp_subflow_early_fallback(msk, subflow);
3606 if (likely(!__mptcp_check_fallback(msk)))
3607 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3609 /* if reaching here via the fastopen/sendmsg path, the caller already
3610 * acquired the subflow socket lock, too.
3612 if (msk->fastopening)
3613 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3615 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3616 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3618 /* on successful connect, the msk state will be moved to established by
3619 * subflow_finish_connect()
3621 if (unlikely(err && err != -EINPROGRESS)) {
3622 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3626 mptcp_copy_inaddrs(sk, ssock->sk);
3628 /* silence EINPROGRESS and let the caller inet_stream_connect
3629 * handle the connection in progress
3634 static struct proto mptcp_prot = {
3636 .owner = THIS_MODULE,
3637 .init = mptcp_init_sock,
3638 .connect = mptcp_connect,
3639 .disconnect = mptcp_disconnect,
3640 .close = mptcp_close,
3641 .accept = mptcp_accept,
3642 .setsockopt = mptcp_setsockopt,
3643 .getsockopt = mptcp_getsockopt,
3644 .shutdown = mptcp_shutdown,
3645 .destroy = mptcp_destroy,
3646 .sendmsg = mptcp_sendmsg,
3647 .ioctl = mptcp_ioctl,
3648 .recvmsg = mptcp_recvmsg,
3649 .release_cb = mptcp_release_cb,
3651 .unhash = mptcp_unhash,
3652 .get_port = mptcp_get_port,
3653 .forward_alloc_get = mptcp_forward_alloc_get,
3654 .sockets_allocated = &mptcp_sockets_allocated,
3656 .memory_allocated = &tcp_memory_allocated,
3657 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3659 .memory_pressure = &tcp_memory_pressure,
3660 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3661 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3662 .sysctl_mem = sysctl_tcp_mem,
3663 .obj_size = sizeof(struct mptcp_sock),
3664 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3665 .no_autobind = true,
3668 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3670 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3671 struct socket *ssock;
3674 lock_sock(sock->sk);
3675 ssock = __mptcp_nmpc_socket(msk);
3676 if (IS_ERR(ssock)) {
3677 err = PTR_ERR(ssock);
3681 err = ssock->ops->bind(ssock, uaddr, addr_len);
3683 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3686 release_sock(sock->sk);
3690 static int mptcp_listen(struct socket *sock, int backlog)
3692 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3693 struct sock *sk = sock->sk;
3694 struct socket *ssock;
3697 pr_debug("msk=%p", msk);
3700 ssock = __mptcp_nmpc_socket(msk);
3701 if (IS_ERR(ssock)) {
3702 err = PTR_ERR(ssock);
3706 mptcp_token_destroy(msk);
3707 inet_sk_state_store(sk, TCP_LISTEN);
3708 sock_set_flag(sk, SOCK_RCU_FREE);
3710 err = ssock->ops->listen(ssock, backlog);
3711 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3713 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3714 mptcp_copy_inaddrs(sk, ssock->sk);
3717 mptcp_event_pm_listener(ssock->sk, MPTCP_EVENT_LISTENER_CREATED);
3724 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3725 int flags, bool kern)
3727 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3728 struct socket *ssock;
3731 pr_debug("msk=%p", msk);
3733 /* Buggy applications can call accept on socket states other then LISTEN
3734 * but no need to allocate the first subflow just to error out.
3736 ssock = READ_ONCE(msk->subflow);
3740 err = ssock->ops->accept(sock, newsock, flags, kern);
3741 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3742 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3743 struct mptcp_subflow_context *subflow;
3744 struct sock *newsk = newsock->sk;
3746 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3747 msk->in_accept_queue = 0;
3751 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3752 * This is needed so NOSPACE flag can be set from tcp stack.
3754 mptcp_for_each_subflow(msk, subflow) {
3755 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3757 if (!ssk->sk_socket)
3758 mptcp_sock_graft(ssk, newsock);
3761 /* Do late cleanup for the first subflow as necessary. Also
3762 * deal with bad peers not doing a complete shutdown.
3765 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3766 __mptcp_close_ssk(newsk, msk->first,
3767 mptcp_subflow_ctx(msk->first), 0);
3768 if (unlikely(list_empty(&msk->conn_list)))
3769 inet_sk_state_store(newsk, TCP_CLOSE);
3772 release_sock(newsk);
3778 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3780 struct sock *sk = (struct sock *)msk;
3782 if (sk_stream_is_writeable(sk))
3783 return EPOLLOUT | EPOLLWRNORM;
3785 mptcp_set_nospace(sk);
3786 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3787 if (sk_stream_is_writeable(sk))
3788 return EPOLLOUT | EPOLLWRNORM;
3793 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3794 struct poll_table_struct *wait)
3796 struct sock *sk = sock->sk;
3797 struct mptcp_sock *msk;
3803 sock_poll_wait(file, sock, wait);
3805 state = inet_sk_state_load(sk);
3806 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3807 if (state == TCP_LISTEN) {
3808 struct socket *ssock = READ_ONCE(msk->subflow);
3810 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3813 return inet_csk_listen_poll(ssock->sk);
3816 shutdown = READ_ONCE(sk->sk_shutdown);
3817 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3819 if (shutdown & RCV_SHUTDOWN)
3820 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3822 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3823 mask |= mptcp_check_readable(msk);
3824 if (shutdown & SEND_SHUTDOWN)
3825 mask |= EPOLLOUT | EPOLLWRNORM;
3827 mask |= mptcp_check_writeable(msk);
3828 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3829 /* cf tcp_poll() note about TFO */
3830 mask |= EPOLLOUT | EPOLLWRNORM;
3833 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3835 if (READ_ONCE(sk->sk_err))
3841 static const struct proto_ops mptcp_stream_ops = {
3843 .owner = THIS_MODULE,
3844 .release = inet_release,
3846 .connect = inet_stream_connect,
3847 .socketpair = sock_no_socketpair,
3848 .accept = mptcp_stream_accept,
3849 .getname = inet_getname,
3851 .ioctl = inet_ioctl,
3852 .gettstamp = sock_gettstamp,
3853 .listen = mptcp_listen,
3854 .shutdown = inet_shutdown,
3855 .setsockopt = sock_common_setsockopt,
3856 .getsockopt = sock_common_getsockopt,
3857 .sendmsg = inet_sendmsg,
3858 .recvmsg = inet_recvmsg,
3859 .mmap = sock_no_mmap,
3860 .sendpage = inet_sendpage,
3863 static struct inet_protosw mptcp_protosw = {
3864 .type = SOCK_STREAM,
3865 .protocol = IPPROTO_MPTCP,
3866 .prot = &mptcp_prot,
3867 .ops = &mptcp_stream_ops,
3868 .flags = INET_PROTOSW_ICSK,
3871 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3873 struct mptcp_delegated_action *delegated;
3874 struct mptcp_subflow_context *subflow;
3877 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3878 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3879 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3881 bh_lock_sock_nested(ssk);
3882 if (!sock_owned_by_user(ssk) &&
3883 mptcp_subflow_has_delegated_action(subflow))
3884 mptcp_subflow_process_delegated(ssk);
3885 /* ... elsewhere tcp_release_cb_override already processed
3886 * the action or will do at next release_sock().
3887 * In both case must dequeue the subflow here - on the same
3888 * CPU that scheduled it.
3890 bh_unlock_sock(ssk);
3893 if (++work_done == budget)
3897 /* always provide a 0 'work_done' argument, so that napi_complete_done
3898 * will not try accessing the NULL napi->dev ptr
3900 napi_complete_done(napi, 0);
3904 void __init mptcp_proto_init(void)
3906 struct mptcp_delegated_action *delegated;
3909 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3911 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3912 panic("Failed to allocate MPTCP pcpu counter\n");
3914 init_dummy_netdev(&mptcp_napi_dev);
3915 for_each_possible_cpu(cpu) {
3916 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3917 INIT_LIST_HEAD(&delegated->head);
3918 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3920 napi_enable(&delegated->napi);
3923 mptcp_subflow_init();
3927 if (proto_register(&mptcp_prot, 1) != 0)
3928 panic("Failed to register MPTCP proto.\n");
3930 inet_register_protosw(&mptcp_protosw);
3932 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3935 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3936 static const struct proto_ops mptcp_v6_stream_ops = {
3938 .owner = THIS_MODULE,
3939 .release = inet6_release,
3941 .connect = inet_stream_connect,
3942 .socketpair = sock_no_socketpair,
3943 .accept = mptcp_stream_accept,
3944 .getname = inet6_getname,
3946 .ioctl = inet6_ioctl,
3947 .gettstamp = sock_gettstamp,
3948 .listen = mptcp_listen,
3949 .shutdown = inet_shutdown,
3950 .setsockopt = sock_common_setsockopt,
3951 .getsockopt = sock_common_getsockopt,
3952 .sendmsg = inet6_sendmsg,
3953 .recvmsg = inet6_recvmsg,
3954 .mmap = sock_no_mmap,
3955 .sendpage = inet_sendpage,
3956 #ifdef CONFIG_COMPAT
3957 .compat_ioctl = inet6_compat_ioctl,
3961 static struct proto mptcp_v6_prot;
3963 static struct inet_protosw mptcp_v6_protosw = {
3964 .type = SOCK_STREAM,
3965 .protocol = IPPROTO_MPTCP,
3966 .prot = &mptcp_v6_prot,
3967 .ops = &mptcp_v6_stream_ops,
3968 .flags = INET_PROTOSW_ICSK,
3971 int __init mptcp_proto_v6_init(void)
3975 mptcp_v6_prot = mptcp_prot;
3976 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3977 mptcp_v6_prot.slab = NULL;
3978 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3980 err = proto_register(&mptcp_v6_prot, 1);
3984 err = inet6_register_protosw(&mptcp_v6_protosw);
3986 proto_unregister(&mptcp_v6_prot);