1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
46 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
49 MPTCP_CMSG_TS = BIT(0),
50 MPTCP_CMSG_INQ = BIT(1),
53 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
55 static void __mptcp_destroy_sock(struct sock *sk);
56 static void __mptcp_check_send_data_fin(struct sock *sk);
58 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
59 static struct net_device mptcp_napi_dev;
61 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
62 * completed yet or has failed, return the subflow socket.
63 * Otherwise return NULL.
65 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
67 if (!msk->subflow || READ_ONCE(msk->can_ack))
73 /* Returns end sequence number of the receiver's advertised window */
74 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
76 return READ_ONCE(msk->wnd_end);
79 static bool mptcp_is_tcpsk(struct sock *sk)
81 struct socket *sock = sk->sk_socket;
83 if (unlikely(sk->sk_prot == &tcp_prot)) {
84 /* we are being invoked after mptcp_accept() has
85 * accepted a non-mp-capable flow: sk is a tcp_sk,
88 * Hand the socket over to tcp so all further socket ops
91 sock->ops = &inet_stream_ops;
93 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
94 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
95 sock->ops = &inet6_stream_ops;
103 static int __mptcp_socket_create(struct mptcp_sock *msk)
105 struct mptcp_subflow_context *subflow;
106 struct sock *sk = (struct sock *)msk;
107 struct socket *ssock;
110 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
114 WRITE_ONCE(msk->first, ssock->sk);
115 WRITE_ONCE(msk->subflow, ssock);
116 subflow = mptcp_subflow_ctx(ssock->sk);
117 list_add(&subflow->node, &msk->conn_list);
118 sock_hold(ssock->sk);
119 subflow->request_mptcp = 1;
121 /* This is the first subflow, always with id 0 */
122 subflow->local_id_valid = 1;
123 mptcp_sock_graft(msk->first, sk->sk_socket);
128 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
130 sk_drops_add(sk, skb);
134 static void mptcp_rmem_charge(struct sock *sk, int size)
136 mptcp_sk(sk)->rmem_fwd_alloc -= size;
139 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
140 struct sk_buff *from)
145 if (MPTCP_SKB_CB(from)->offset ||
146 !skb_try_coalesce(to, from, &fragstolen, &delta))
149 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
150 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
151 to->len, MPTCP_SKB_CB(from)->end_seq);
152 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
154 /* note the fwd memory can reach a negative value after accounting
155 * for the delta, but the later skb free will restore a non
158 atomic_add(delta, &sk->sk_rmem_alloc);
159 mptcp_rmem_charge(sk, delta);
160 kfree_skb_partial(from, fragstolen);
165 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
166 struct sk_buff *from)
168 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
171 return mptcp_try_coalesce((struct sock *)msk, to, from);
174 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
176 amount >>= PAGE_SHIFT;
177 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
178 __sk_mem_reduce_allocated(sk, amount);
181 static void mptcp_rmem_uncharge(struct sock *sk, int size)
183 struct mptcp_sock *msk = mptcp_sk(sk);
186 msk->rmem_fwd_alloc += size;
187 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
189 /* see sk_mem_uncharge() for the rationale behind the following schema */
190 if (unlikely(reclaimable >= PAGE_SIZE))
191 __mptcp_rmem_reclaim(sk, reclaimable);
194 static void mptcp_rfree(struct sk_buff *skb)
196 unsigned int len = skb->truesize;
197 struct sock *sk = skb->sk;
199 atomic_sub(len, &sk->sk_rmem_alloc);
200 mptcp_rmem_uncharge(sk, len);
203 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
207 skb->destructor = mptcp_rfree;
208 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
209 mptcp_rmem_charge(sk, skb->truesize);
212 /* "inspired" by tcp_data_queue_ofo(), main differences:
214 * - don't cope with sacks
216 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
218 struct sock *sk = (struct sock *)msk;
219 struct rb_node **p, *parent;
220 u64 seq, end_seq, max_seq;
221 struct sk_buff *skb1;
223 seq = MPTCP_SKB_CB(skb)->map_seq;
224 end_seq = MPTCP_SKB_CB(skb)->end_seq;
225 max_seq = atomic64_read(&msk->rcv_wnd_sent);
227 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
228 RB_EMPTY_ROOT(&msk->out_of_order_queue));
229 if (after64(end_seq, max_seq)) {
232 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
233 (unsigned long long)end_seq - (unsigned long)max_seq,
234 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
235 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
239 p = &msk->out_of_order_queue.rb_node;
240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
241 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
242 rb_link_node(&skb->rbnode, NULL, p);
243 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
244 msk->ooo_last_skb = skb;
248 /* with 2 subflows, adding at end of ooo queue is quite likely
249 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
251 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
252 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
253 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
257 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
258 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
259 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
260 parent = &msk->ooo_last_skb->rbnode;
261 p = &parent->rb_right;
265 /* Find place to insert this segment. Handle overlaps on the way. */
269 skb1 = rb_to_skb(parent);
270 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
271 p = &parent->rb_left;
274 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
275 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
276 /* All the bits are present. Drop. */
278 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
281 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
285 * continue traversing
288 /* skb's seq == skb1's seq and skb covers skb1.
289 * Replace skb1 with skb.
291 rb_replace_node(&skb1->rbnode, &skb->rbnode,
292 &msk->out_of_order_queue);
293 mptcp_drop(sk, skb1);
294 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
297 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
298 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
301 p = &parent->rb_right;
305 /* Insert segment into RB tree. */
306 rb_link_node(&skb->rbnode, parent, p);
307 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
310 /* Remove other segments covered by skb. */
311 while ((skb1 = skb_rb_next(skb)) != NULL) {
312 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
314 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
315 mptcp_drop(sk, skb1);
316 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
318 /* If there is no skb after us, we are the last_skb ! */
320 msk->ooo_last_skb = skb;
324 mptcp_set_owner_r(skb, sk);
327 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
329 struct mptcp_sock *msk = mptcp_sk(sk);
332 if (size <= msk->rmem_fwd_alloc)
335 size -= msk->rmem_fwd_alloc;
336 amt = sk_mem_pages(size);
337 amount = amt << PAGE_SHIFT;
338 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
341 msk->rmem_fwd_alloc += amount;
345 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
346 struct sk_buff *skb, unsigned int offset,
349 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
350 struct sock *sk = (struct sock *)msk;
351 struct sk_buff *tail;
354 __skb_unlink(skb, &ssk->sk_receive_queue);
359 /* try to fetch required memory from subflow */
360 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
363 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
365 /* the skb map_seq accounts for the skb offset:
366 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
369 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
370 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
371 MPTCP_SKB_CB(skb)->offset = offset;
372 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
374 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
376 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
377 tail = skb_peek_tail(&sk->sk_receive_queue);
378 if (tail && mptcp_try_coalesce(sk, tail, skb))
381 mptcp_set_owner_r(skb, sk);
382 __skb_queue_tail(&sk->sk_receive_queue, skb);
384 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
385 mptcp_data_queue_ofo(msk, skb);
389 /* old data, keep it simple and drop the whole pkt, sender
390 * will retransmit as needed, if needed.
392 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
398 static void mptcp_stop_timer(struct sock *sk)
400 struct inet_connection_sock *icsk = inet_csk(sk);
402 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
403 mptcp_sk(sk)->timer_ival = 0;
406 static void mptcp_close_wake_up(struct sock *sk)
408 if (sock_flag(sk, SOCK_DEAD))
411 sk->sk_state_change(sk);
412 if (sk->sk_shutdown == SHUTDOWN_MASK ||
413 sk->sk_state == TCP_CLOSE)
414 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
416 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
419 static bool mptcp_pending_data_fin_ack(struct sock *sk)
421 struct mptcp_sock *msk = mptcp_sk(sk);
423 return !__mptcp_check_fallback(msk) &&
424 ((1 << sk->sk_state) &
425 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
426 msk->write_seq == READ_ONCE(msk->snd_una);
429 static void mptcp_check_data_fin_ack(struct sock *sk)
431 struct mptcp_sock *msk = mptcp_sk(sk);
433 /* Look for an acknowledged DATA_FIN */
434 if (mptcp_pending_data_fin_ack(sk)) {
435 WRITE_ONCE(msk->snd_data_fin_enable, 0);
437 switch (sk->sk_state) {
439 inet_sk_state_store(sk, TCP_FIN_WAIT2);
443 inet_sk_state_store(sk, TCP_CLOSE);
447 mptcp_close_wake_up(sk);
451 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
453 struct mptcp_sock *msk = mptcp_sk(sk);
455 if (READ_ONCE(msk->rcv_data_fin) &&
456 ((1 << sk->sk_state) &
457 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
458 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
460 if (msk->ack_seq == rcv_data_fin_seq) {
462 *seq = rcv_data_fin_seq;
471 static void mptcp_set_datafin_timeout(const struct sock *sk)
473 struct inet_connection_sock *icsk = inet_csk(sk);
476 retransmits = min_t(u32, icsk->icsk_retransmits,
477 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
479 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
482 static void __mptcp_set_timeout(struct sock *sk, long tout)
484 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
487 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
489 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
491 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
492 inet_csk(ssk)->icsk_timeout - jiffies : 0;
495 static void mptcp_set_timeout(struct sock *sk)
497 struct mptcp_subflow_context *subflow;
500 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
501 tout = max(tout, mptcp_timeout_from_subflow(subflow));
502 __mptcp_set_timeout(sk, tout);
505 static inline bool tcp_can_send_ack(const struct sock *ssk)
507 return !((1 << inet_sk_state_load(ssk)) &
508 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
511 void __mptcp_subflow_send_ack(struct sock *ssk)
513 if (tcp_can_send_ack(ssk))
517 static void mptcp_subflow_send_ack(struct sock *ssk)
521 slow = lock_sock_fast(ssk);
522 __mptcp_subflow_send_ack(ssk);
523 unlock_sock_fast(ssk, slow);
526 static void mptcp_send_ack(struct mptcp_sock *msk)
528 struct mptcp_subflow_context *subflow;
530 mptcp_for_each_subflow(msk, subflow)
531 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
534 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
538 slow = lock_sock_fast(ssk);
539 if (tcp_can_send_ack(ssk))
540 tcp_cleanup_rbuf(ssk, 1);
541 unlock_sock_fast(ssk, slow);
544 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
546 const struct inet_connection_sock *icsk = inet_csk(ssk);
547 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
548 const struct tcp_sock *tp = tcp_sk(ssk);
550 return (ack_pending & ICSK_ACK_SCHED) &&
551 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
552 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
553 (rx_empty && ack_pending &
554 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
557 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
559 int old_space = READ_ONCE(msk->old_wspace);
560 struct mptcp_subflow_context *subflow;
561 struct sock *sk = (struct sock *)msk;
562 int space = __mptcp_space(sk);
563 bool cleanup, rx_empty;
565 cleanup = (space > 0) && (space >= (old_space << 1));
566 rx_empty = !__mptcp_rmem(sk);
568 mptcp_for_each_subflow(msk, subflow) {
569 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
571 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
572 mptcp_subflow_cleanup_rbuf(ssk);
576 static bool mptcp_check_data_fin(struct sock *sk)
578 struct mptcp_sock *msk = mptcp_sk(sk);
579 u64 rcv_data_fin_seq;
582 if (__mptcp_check_fallback(msk))
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 */
623 mptcp_close_wake_up(sk);
628 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
632 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
633 struct sock *sk = (struct sock *)msk;
634 unsigned int moved = 0;
635 bool more_data_avail;
640 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
642 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
643 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
645 if (unlikely(ssk_rbuf > sk_rbuf)) {
646 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
651 pr_debug("msk=%p ssk=%p", msk, ssk);
654 u32 map_remaining, offset;
655 u32 seq = tp->copied_seq;
659 /* try to move as much data as available */
660 map_remaining = subflow->map_data_len -
661 mptcp_subflow_get_map_offset(subflow);
663 skb = skb_peek(&ssk->sk_receive_queue);
665 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
666 * a different CPU can have already processed the pending
667 * data, stop here or we can enter an infinite loop
674 if (__mptcp_check_fallback(msk)) {
675 /* Under fallback skbs have no MPTCP extension and TCP could
676 * collapse them between the dummy map creation and the
677 * current dequeue. Be sure to adjust the map size.
679 map_remaining = skb->len;
680 subflow->map_data_len = skb->len;
683 offset = seq - TCP_SKB_CB(skb)->seq;
684 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
690 if (offset < skb->len) {
691 size_t len = skb->len - offset;
696 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
700 if (WARN_ON_ONCE(map_remaining < len))
704 sk_eat_skb(ssk, skb);
708 WRITE_ONCE(tp->copied_seq, seq);
709 more_data_avail = mptcp_subflow_data_available(ssk);
711 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
715 } while (more_data_avail);
721 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
723 struct sock *sk = (struct sock *)msk;
724 struct sk_buff *skb, *tail;
729 p = rb_first(&msk->out_of_order_queue);
730 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
733 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
737 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
739 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
742 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
746 end_seq = MPTCP_SKB_CB(skb)->end_seq;
747 tail = skb_peek_tail(&sk->sk_receive_queue);
748 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
749 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
751 /* skip overlapping data, if any */
752 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
753 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
755 MPTCP_SKB_CB(skb)->offset += delta;
756 MPTCP_SKB_CB(skb)->map_seq += delta;
757 __skb_queue_tail(&sk->sk_receive_queue, skb);
759 msk->ack_seq = end_seq;
765 /* In most cases we will be able to lock the mptcp socket. If its already
766 * owned, we need to defer to the work queue to avoid ABBA deadlock.
768 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
770 struct sock *sk = (struct sock *)msk;
771 unsigned int moved = 0;
773 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
774 __mptcp_ofo_queue(msk);
775 if (unlikely(ssk->sk_err)) {
776 if (!sock_owned_by_user(sk))
777 __mptcp_error_report(sk);
779 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
782 /* If the moves have caught up with the DATA_FIN sequence number
783 * it's time to ack the DATA_FIN and change socket state, but
784 * this is not a good place to change state. Let the workqueue
787 if (mptcp_pending_data_fin(sk, NULL))
788 mptcp_schedule_work(sk);
792 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
794 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
795 struct mptcp_sock *msk = mptcp_sk(sk);
796 int sk_rbuf, ssk_rbuf;
798 /* The peer can send data while we are shutting down this
799 * subflow at msk destruction time, but we must avoid enqueuing
800 * more data to the msk receive queue
802 if (unlikely(subflow->disposable))
805 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
806 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
807 if (unlikely(ssk_rbuf > sk_rbuf))
810 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
811 if (__mptcp_rmem(sk) > sk_rbuf) {
812 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
816 /* Wake-up the reader only for in-sequence data */
818 if (move_skbs_to_msk(msk, ssk))
819 sk->sk_data_ready(sk);
821 mptcp_data_unlock(sk);
824 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
826 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
827 WRITE_ONCE(msk->allow_infinite_fallback, false);
828 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
831 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
833 struct sock *sk = (struct sock *)msk;
835 if (sk->sk_state != TCP_ESTABLISHED)
838 /* attach to msk socket only after we are sure we will deal with it
841 if (sk->sk_socket && !ssk->sk_socket)
842 mptcp_sock_graft(ssk, sk->sk_socket);
844 mptcp_sockopt_sync_locked(msk, ssk);
845 mptcp_subflow_joined(msk, ssk);
849 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
851 struct mptcp_subflow_context *tmp, *subflow;
852 struct mptcp_sock *msk = mptcp_sk(sk);
854 list_for_each_entry_safe(subflow, tmp, join_list, node) {
855 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
856 bool slow = lock_sock_fast(ssk);
858 list_move_tail(&subflow->node, &msk->conn_list);
859 if (!__mptcp_finish_join(msk, ssk))
860 mptcp_subflow_reset(ssk);
861 unlock_sock_fast(ssk, slow);
865 static bool mptcp_timer_pending(struct sock *sk)
867 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
870 static void mptcp_reset_timer(struct sock *sk)
872 struct inet_connection_sock *icsk = inet_csk(sk);
875 /* prevent rescheduling on close */
876 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
879 tout = mptcp_sk(sk)->timer_ival;
880 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
883 bool mptcp_schedule_work(struct sock *sk)
885 if (inet_sk_state_load(sk) != TCP_CLOSE &&
886 schedule_work(&mptcp_sk(sk)->work)) {
887 /* each subflow already holds a reference to the sk, and the
888 * workqueue is invoked by a subflow, so sk can't go away here.
896 void mptcp_subflow_eof(struct sock *sk)
898 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
899 mptcp_schedule_work(sk);
902 static void mptcp_check_for_eof(struct mptcp_sock *msk)
904 struct mptcp_subflow_context *subflow;
905 struct sock *sk = (struct sock *)msk;
908 mptcp_for_each_subflow(msk, subflow)
909 receivers += !subflow->rx_eof;
913 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
914 /* hopefully temporary hack: propagate shutdown status
915 * to msk, when all subflows agree on it
917 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
919 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
920 sk->sk_data_ready(sk);
923 switch (sk->sk_state) {
924 case TCP_ESTABLISHED:
925 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
928 inet_sk_state_store(sk, TCP_CLOSING);
931 inet_sk_state_store(sk, TCP_CLOSE);
936 mptcp_close_wake_up(sk);
939 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
941 struct mptcp_subflow_context *subflow;
942 struct sock *sk = (struct sock *)msk;
944 sock_owned_by_me(sk);
946 mptcp_for_each_subflow(msk, subflow) {
947 if (READ_ONCE(subflow->data_avail))
948 return mptcp_subflow_tcp_sock(subflow);
954 static bool mptcp_skb_can_collapse_to(u64 write_seq,
955 const struct sk_buff *skb,
956 const struct mptcp_ext *mpext)
958 if (!tcp_skb_can_collapse_to(skb))
961 /* can collapse only if MPTCP level sequence is in order and this
962 * mapping has not been xmitted yet
964 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
968 /* we can append data to the given data frag if:
969 * - there is space available in the backing page_frag
970 * - the data frag tail matches the current page_frag free offset
971 * - the data frag end sequence number matches the current write seq
973 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
974 const struct page_frag *pfrag,
975 const struct mptcp_data_frag *df)
977 return df && pfrag->page == df->page &&
978 pfrag->size - pfrag->offset > 0 &&
979 pfrag->offset == (df->offset + df->data_len) &&
980 df->data_seq + df->data_len == msk->write_seq;
983 static void dfrag_uncharge(struct sock *sk, int len)
985 sk_mem_uncharge(sk, len);
986 sk_wmem_queued_add(sk, -len);
989 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
991 int len = dfrag->data_len + dfrag->overhead;
993 list_del(&dfrag->list);
994 dfrag_uncharge(sk, len);
995 put_page(dfrag->page);
998 static void __mptcp_clean_una(struct sock *sk)
1000 struct mptcp_sock *msk = mptcp_sk(sk);
1001 struct mptcp_data_frag *dtmp, *dfrag;
1004 /* on fallback we just need to ignore snd_una, as this is really
1007 if (__mptcp_check_fallback(msk))
1008 msk->snd_una = READ_ONCE(msk->snd_nxt);
1010 snd_una = msk->snd_una;
1011 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1012 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1015 if (unlikely(dfrag == msk->first_pending)) {
1016 /* in recovery mode can see ack after the current snd head */
1017 if (WARN_ON_ONCE(!msk->recovery))
1020 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1023 dfrag_clear(sk, dfrag);
1026 dfrag = mptcp_rtx_head(sk);
1027 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1028 u64 delta = snd_una - dfrag->data_seq;
1030 /* prevent wrap around in recovery mode */
1031 if (unlikely(delta > dfrag->already_sent)) {
1032 if (WARN_ON_ONCE(!msk->recovery))
1034 if (WARN_ON_ONCE(delta > dfrag->data_len))
1036 dfrag->already_sent += delta - dfrag->already_sent;
1039 dfrag->data_seq += delta;
1040 dfrag->offset += delta;
1041 dfrag->data_len -= delta;
1042 dfrag->already_sent -= delta;
1044 dfrag_uncharge(sk, delta);
1047 /* all retransmitted data acked, recovery completed */
1048 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1049 msk->recovery = false;
1052 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1053 snd_una == READ_ONCE(msk->write_seq)) {
1054 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1055 mptcp_stop_timer(sk);
1057 mptcp_reset_timer(sk);
1061 static void __mptcp_clean_una_wakeup(struct sock *sk)
1063 lockdep_assert_held_once(&sk->sk_lock.slock);
1065 __mptcp_clean_una(sk);
1066 mptcp_write_space(sk);
1069 static void mptcp_clean_una_wakeup(struct sock *sk)
1071 mptcp_data_lock(sk);
1072 __mptcp_clean_una_wakeup(sk);
1073 mptcp_data_unlock(sk);
1076 static void mptcp_enter_memory_pressure(struct sock *sk)
1078 struct mptcp_subflow_context *subflow;
1079 struct mptcp_sock *msk = mptcp_sk(sk);
1082 sk_stream_moderate_sndbuf(sk);
1083 mptcp_for_each_subflow(msk, subflow) {
1084 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1087 tcp_enter_memory_pressure(ssk);
1088 sk_stream_moderate_sndbuf(ssk);
1093 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1096 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1098 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1099 pfrag, sk->sk_allocation)))
1102 mptcp_enter_memory_pressure(sk);
1106 static struct mptcp_data_frag *
1107 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1110 int offset = ALIGN(orig_offset, sizeof(long));
1111 struct mptcp_data_frag *dfrag;
1113 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1114 dfrag->data_len = 0;
1115 dfrag->data_seq = msk->write_seq;
1116 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1117 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1118 dfrag->already_sent = 0;
1119 dfrag->page = pfrag->page;
1124 struct mptcp_sendmsg_info {
1130 bool data_lock_held;
1133 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1134 u64 data_seq, int avail_size)
1136 u64 window_end = mptcp_wnd_end(msk);
1139 if (__mptcp_check_fallback(msk))
1142 mptcp_snd_wnd = window_end - data_seq;
1143 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1145 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1146 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1147 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1153 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1155 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1159 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1163 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1165 struct sk_buff *skb;
1167 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1169 if (likely(__mptcp_add_ext(skb, gfp))) {
1170 skb_reserve(skb, MAX_TCP_HEADER);
1171 skb->ip_summed = CHECKSUM_PARTIAL;
1172 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1177 mptcp_enter_memory_pressure(sk);
1182 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1184 struct sk_buff *skb;
1186 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1190 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1191 tcp_skb_entail(ssk, skb);
1194 tcp_skb_tsorted_anchor_cleanup(skb);
1199 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1201 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1203 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1206 /* note: this always recompute the csum on the whole skb, even
1207 * if we just appended a single frag. More status info needed
1209 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1211 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1212 __wsum csum = ~csum_unfold(mpext->csum);
1213 int offset = skb->len - added;
1215 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1218 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1220 struct mptcp_ext *mpext)
1225 mpext->infinite_map = 1;
1226 mpext->data_len = 0;
1228 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1229 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1231 mptcp_do_fallback(ssk);
1234 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1235 struct mptcp_data_frag *dfrag,
1236 struct mptcp_sendmsg_info *info)
1238 u64 data_seq = dfrag->data_seq + info->sent;
1239 int offset = dfrag->offset + info->sent;
1240 struct mptcp_sock *msk = mptcp_sk(sk);
1241 bool zero_window_probe = false;
1242 struct mptcp_ext *mpext = NULL;
1243 bool can_coalesce = false;
1244 bool reuse_skb = true;
1245 struct sk_buff *skb;
1249 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1250 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1252 if (WARN_ON_ONCE(info->sent > info->limit ||
1253 info->limit > dfrag->data_len))
1256 if (unlikely(!__tcp_can_send(ssk)))
1259 /* compute send limit */
1260 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1261 copy = info->size_goal;
1263 skb = tcp_write_queue_tail(ssk);
1264 if (skb && copy > skb->len) {
1265 /* Limit the write to the size available in the
1266 * current skb, if any, so that we create at most a new skb.
1267 * Explicitly tells TCP internals to avoid collapsing on later
1268 * queue management operation, to avoid breaking the ext <->
1269 * SSN association set here
1271 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1272 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1273 TCP_SKB_CB(skb)->eor = 1;
1277 i = skb_shinfo(skb)->nr_frags;
1278 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1279 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1280 tcp_mark_push(tcp_sk(ssk), skb);
1287 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1291 i = skb_shinfo(skb)->nr_frags;
1293 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1296 /* Zero window and all data acked? Probe. */
1297 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1299 u64 snd_una = READ_ONCE(msk->snd_una);
1301 if (snd_una != msk->snd_nxt) {
1302 tcp_remove_empty_skb(ssk);
1306 zero_window_probe = true;
1307 data_seq = snd_una - 1;
1310 /* all mptcp-level data is acked, no skbs should be present into the
1313 WARN_ON_ONCE(reuse_skb);
1316 copy = min_t(size_t, copy, info->limit - info->sent);
1317 if (!sk_wmem_schedule(ssk, copy)) {
1318 tcp_remove_empty_skb(ssk);
1323 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1325 get_page(dfrag->page);
1326 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1330 skb->data_len += copy;
1331 skb->truesize += copy;
1332 sk_wmem_queued_add(ssk, copy);
1333 sk_mem_charge(ssk, copy);
1334 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1335 TCP_SKB_CB(skb)->end_seq += copy;
1336 tcp_skb_pcount_set(skb, 0);
1338 /* on skb reuse we just need to update the DSS len */
1340 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1341 mpext->data_len += copy;
1342 WARN_ON_ONCE(zero_window_probe);
1346 memset(mpext, 0, sizeof(*mpext));
1347 mpext->data_seq = data_seq;
1348 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1349 mpext->data_len = copy;
1353 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1354 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1357 if (zero_window_probe) {
1358 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1360 if (READ_ONCE(msk->csum_enabled))
1361 mptcp_update_data_checksum(skb, copy);
1362 tcp_push_pending_frames(ssk);
1366 if (READ_ONCE(msk->csum_enabled))
1367 mptcp_update_data_checksum(skb, copy);
1368 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1369 mptcp_update_infinite_map(msk, ssk, mpext);
1370 trace_mptcp_sendmsg_frag(mpext);
1371 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1375 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1376 sizeof(struct tcphdr) - \
1377 MAX_TCP_OPTION_SPACE - \
1378 sizeof(struct ipv6hdr) - \
1379 sizeof(struct frag_hdr))
1381 struct subflow_send_info {
1386 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1388 if (!subflow->stale)
1392 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1395 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1397 if (unlikely(subflow->stale)) {
1398 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1400 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1403 mptcp_subflow_set_active(subflow);
1405 return __mptcp_subflow_active(subflow);
1408 #define SSK_MODE_ACTIVE 0
1409 #define SSK_MODE_BACKUP 1
1410 #define SSK_MODE_MAX 2
1412 /* implement the mptcp packet scheduler;
1413 * returns the subflow that will transmit the next DSS
1414 * additionally updates the rtx timeout
1416 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1418 struct subflow_send_info send_info[SSK_MODE_MAX];
1419 struct mptcp_subflow_context *subflow;
1420 struct sock *sk = (struct sock *)msk;
1421 u32 pace, burst, wmem;
1422 int i, nr_active = 0;
1427 sock_owned_by_me(sk);
1429 if (__mptcp_check_fallback(msk)) {
1432 return __tcp_can_send(msk->first) &&
1433 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1436 /* re-use last subflow, if the burst allow that */
1437 if (msk->last_snd && msk->snd_burst > 0 &&
1438 sk_stream_memory_free(msk->last_snd) &&
1439 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1440 mptcp_set_timeout(sk);
1441 return msk->last_snd;
1444 /* pick the subflow with the lower wmem/wspace ratio */
1445 for (i = 0; i < SSK_MODE_MAX; ++i) {
1446 send_info[i].ssk = NULL;
1447 send_info[i].linger_time = -1;
1450 mptcp_for_each_subflow(msk, subflow) {
1451 trace_mptcp_subflow_get_send(subflow);
1452 ssk = mptcp_subflow_tcp_sock(subflow);
1453 if (!mptcp_subflow_active(subflow))
1456 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1457 nr_active += !subflow->backup;
1458 pace = subflow->avg_pacing_rate;
1459 if (unlikely(!pace)) {
1460 /* init pacing rate from socket */
1461 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1462 pace = subflow->avg_pacing_rate;
1467 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1468 if (linger_time < send_info[subflow->backup].linger_time) {
1469 send_info[subflow->backup].ssk = ssk;
1470 send_info[subflow->backup].linger_time = linger_time;
1473 __mptcp_set_timeout(sk, tout);
1475 /* pick the best backup if no other subflow is active */
1477 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1479 /* According to the blest algorithm, to avoid HoL blocking for the
1480 * faster flow, we need to:
1481 * - estimate the faster flow linger time
1482 * - use the above to estimate the amount of byte transferred
1483 * by the faster flow
1484 * - check that the amount of queued data is greter than the above,
1485 * otherwise do not use the picked, slower, subflow
1486 * We select the subflow with the shorter estimated time to flush
1487 * the queued mem, which basically ensure the above. We just need
1488 * to check that subflow has a non empty cwin.
1490 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1491 if (!ssk || !sk_stream_memory_free(ssk))
1494 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1495 wmem = READ_ONCE(ssk->sk_wmem_queued);
1497 msk->last_snd = NULL;
1501 subflow = mptcp_subflow_ctx(ssk);
1502 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1503 READ_ONCE(ssk->sk_pacing_rate) * burst,
1505 msk->last_snd = ssk;
1506 msk->snd_burst = burst;
1510 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1512 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1516 static void mptcp_update_post_push(struct mptcp_sock *msk,
1517 struct mptcp_data_frag *dfrag,
1520 u64 snd_nxt_new = dfrag->data_seq;
1522 dfrag->already_sent += sent;
1524 msk->snd_burst -= sent;
1526 snd_nxt_new += dfrag->already_sent;
1528 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1529 * is recovering after a failover. In that event, this re-sends
1532 * Thus compute snd_nxt_new candidate based on
1533 * the dfrag->data_seq that was sent and the data
1534 * that has been handed to the subflow for transmission
1535 * and skip update in case it was old dfrag.
1537 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1538 msk->snd_nxt = snd_nxt_new;
1541 void mptcp_check_and_set_pending(struct sock *sk)
1543 if (mptcp_send_head(sk))
1544 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1547 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1549 struct sock *prev_ssk = NULL, *ssk = NULL;
1550 struct mptcp_sock *msk = mptcp_sk(sk);
1551 struct mptcp_sendmsg_info info = {
1554 bool do_check_data_fin = false;
1555 struct mptcp_data_frag *dfrag;
1558 while ((dfrag = mptcp_send_head(sk))) {
1559 info.sent = dfrag->already_sent;
1560 info.limit = dfrag->data_len;
1561 len = dfrag->data_len - dfrag->already_sent;
1566 ssk = mptcp_subflow_get_send(msk);
1568 /* First check. If the ssk has changed since
1569 * the last round, release prev_ssk
1571 if (ssk != prev_ssk && prev_ssk)
1572 mptcp_push_release(prev_ssk, &info);
1576 /* Need to lock the new subflow only if different
1577 * from the previous one, otherwise we are still
1578 * helding the relevant lock
1580 if (ssk != prev_ssk)
1583 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1587 mptcp_push_release(ssk, &info);
1591 do_check_data_fin = true;
1595 mptcp_update_post_push(msk, dfrag, ret);
1597 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1600 /* at this point we held the socket lock for the last subflow we used */
1602 mptcp_push_release(ssk, &info);
1605 /* ensure the rtx timer is running */
1606 if (!mptcp_timer_pending(sk))
1607 mptcp_reset_timer(sk);
1608 if (do_check_data_fin)
1609 __mptcp_check_send_data_fin(sk);
1612 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1614 struct mptcp_sock *msk = mptcp_sk(sk);
1615 struct mptcp_sendmsg_info info = {
1616 .data_lock_held = true,
1618 struct mptcp_data_frag *dfrag;
1619 struct sock *xmit_ssk;
1620 int len, copied = 0;
1624 while ((dfrag = mptcp_send_head(sk))) {
1625 info.sent = dfrag->already_sent;
1626 info.limit = dfrag->data_len;
1627 len = dfrag->data_len - dfrag->already_sent;
1631 /* the caller already invoked the packet scheduler,
1632 * check for a different subflow usage only after
1633 * spooling the first chunk of data
1635 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1638 if (xmit_ssk != ssk) {
1639 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1640 MPTCP_DELEGATE_SEND);
1644 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1653 mptcp_update_post_push(msk, dfrag, ret);
1655 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1659 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1660 * not going to flush it via release_sock()
1663 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1665 if (!mptcp_timer_pending(sk))
1666 mptcp_reset_timer(sk);
1668 if (msk->snd_data_fin_enable &&
1669 msk->snd_nxt + 1 == msk->write_seq)
1670 mptcp_schedule_work(sk);
1674 static void mptcp_set_nospace(struct sock *sk)
1676 /* enable autotune */
1677 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1679 /* will be cleared on avail space */
1680 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1683 static int mptcp_disconnect(struct sock *sk, int flags);
1685 static int mptcp_sendmsg_fastopen(struct sock *sk, struct sock *ssk, struct msghdr *msg,
1686 size_t len, int *copied_syn)
1688 unsigned int saved_flags = msg->msg_flags;
1689 struct mptcp_sock *msk = mptcp_sk(sk);
1693 msg->msg_flags |= MSG_DONTWAIT;
1694 msk->fastopening = 1;
1695 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1696 msk->fastopening = 0;
1697 msg->msg_flags = saved_flags;
1700 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1701 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1702 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1703 msg->msg_namelen, msg->msg_flags, 1);
1705 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1706 * case of any error, except timeout or signal
1708 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1710 } else if (ret && ret != -EINPROGRESS) {
1711 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1712 * __inet_stream_connect() can fail, due to looking check,
1713 * see mptcp_disconnect().
1714 * Attempt it again outside the problematic scope.
1716 if (!mptcp_disconnect(sk, 0))
1717 sk->sk_socket->state = SS_UNCONNECTED;
1723 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1725 struct mptcp_sock *msk = mptcp_sk(sk);
1726 struct page_frag *pfrag;
1727 struct socket *ssock;
1732 /* we don't support FASTOPEN yet */
1733 if (msg->msg_flags & MSG_FASTOPEN)
1736 /* silently ignore everything else */
1737 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1741 ssock = __mptcp_nmpc_socket(msk);
1742 if (unlikely(ssock && inet_sk(ssock->sk)->defer_connect)) {
1745 ret = mptcp_sendmsg_fastopen(sk, ssock->sk, msg, len, &copied_syn);
1746 copied += copied_syn;
1747 if (ret == -EINPROGRESS && copied_syn > 0)
1753 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1755 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1756 ret = sk_stream_wait_connect(sk, &timeo);
1762 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1765 pfrag = sk_page_frag(sk);
1767 while (msg_data_left(msg)) {
1768 int total_ts, frag_truesize = 0;
1769 struct mptcp_data_frag *dfrag;
1770 bool dfrag_collapsed;
1771 size_t psize, offset;
1773 /* reuse tail pfrag, if possible, or carve a new one from the
1776 dfrag = mptcp_pending_tail(sk);
1777 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1778 if (!dfrag_collapsed) {
1779 if (!sk_stream_memory_free(sk))
1780 goto wait_for_memory;
1782 if (!mptcp_page_frag_refill(sk, pfrag))
1783 goto wait_for_memory;
1785 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1786 frag_truesize = dfrag->overhead;
1789 /* we do not bound vs wspace, to allow a single packet.
1790 * memory accounting will prevent execessive memory usage
1793 offset = dfrag->offset + dfrag->data_len;
1794 psize = pfrag->size - offset;
1795 psize = min_t(size_t, psize, msg_data_left(msg));
1796 total_ts = psize + frag_truesize;
1798 if (!sk_wmem_schedule(sk, total_ts))
1799 goto wait_for_memory;
1801 if (copy_page_from_iter(dfrag->page, offset, psize,
1802 &msg->msg_iter) != psize) {
1807 /* data successfully copied into the write queue */
1808 sk->sk_forward_alloc -= total_ts;
1810 dfrag->data_len += psize;
1811 frag_truesize += psize;
1812 pfrag->offset += frag_truesize;
1813 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1815 /* charge data on mptcp pending queue to the msk socket
1816 * Note: we charge such data both to sk and ssk
1818 sk_wmem_queued_add(sk, frag_truesize);
1819 if (!dfrag_collapsed) {
1820 get_page(dfrag->page);
1821 list_add_tail(&dfrag->list, &msk->rtx_queue);
1822 if (!msk->first_pending)
1823 WRITE_ONCE(msk->first_pending, dfrag);
1825 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1826 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1832 mptcp_set_nospace(sk);
1833 __mptcp_push_pending(sk, msg->msg_flags);
1834 ret = sk_stream_wait_memory(sk, &timeo);
1840 __mptcp_push_pending(sk, msg->msg_flags);
1850 copied = sk_stream_error(sk, msg->msg_flags, ret);
1854 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1856 size_t len, int flags,
1857 struct scm_timestamping_internal *tss,
1860 struct sk_buff *skb, *tmp;
1863 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1864 u32 offset = MPTCP_SKB_CB(skb)->offset;
1865 u32 data_len = skb->len - offset;
1866 u32 count = min_t(size_t, len - copied, data_len);
1869 if (!(flags & MSG_TRUNC)) {
1870 err = skb_copy_datagram_msg(skb, offset, msg, count);
1871 if (unlikely(err < 0)) {
1878 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1879 tcp_update_recv_tstamps(skb, tss);
1880 *cmsg_flags |= MPTCP_CMSG_TS;
1885 if (count < data_len) {
1886 if (!(flags & MSG_PEEK)) {
1887 MPTCP_SKB_CB(skb)->offset += count;
1888 MPTCP_SKB_CB(skb)->map_seq += count;
1893 if (!(flags & MSG_PEEK)) {
1894 /* we will bulk release the skb memory later */
1895 skb->destructor = NULL;
1896 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1897 __skb_unlink(skb, &msk->receive_queue);
1908 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1910 * Only difference: Use highest rtt estimate of the subflows in use.
1912 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1914 struct mptcp_subflow_context *subflow;
1915 struct sock *sk = (struct sock *)msk;
1916 u32 time, advmss = 1;
1919 sock_owned_by_me(sk);
1924 msk->rcvq_space.copied += copied;
1926 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1927 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1929 rtt_us = msk->rcvq_space.rtt_us;
1930 if (rtt_us && time < (rtt_us >> 3))
1934 mptcp_for_each_subflow(msk, subflow) {
1935 const struct tcp_sock *tp;
1939 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1941 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1942 sf_advmss = READ_ONCE(tp->advmss);
1944 rtt_us = max(sf_rtt_us, rtt_us);
1945 advmss = max(sf_advmss, advmss);
1948 msk->rcvq_space.rtt_us = rtt_us;
1949 if (time < (rtt_us >> 3) || rtt_us == 0)
1952 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1955 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1956 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1960 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1962 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1964 do_div(grow, msk->rcvq_space.space);
1965 rcvwin += (grow << 1);
1967 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1968 while (tcp_win_from_space(sk, rcvmem) < advmss)
1971 do_div(rcvwin, advmss);
1972 rcvbuf = min_t(u64, rcvwin * rcvmem,
1973 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1975 if (rcvbuf > sk->sk_rcvbuf) {
1978 window_clamp = tcp_win_from_space(sk, rcvbuf);
1979 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1981 /* Make subflows follow along. If we do not do this, we
1982 * get drops at subflow level if skbs can't be moved to
1983 * the mptcp rx queue fast enough (announced rcv_win can
1984 * exceed ssk->sk_rcvbuf).
1986 mptcp_for_each_subflow(msk, subflow) {
1990 ssk = mptcp_subflow_tcp_sock(subflow);
1991 slow = lock_sock_fast(ssk);
1992 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1993 tcp_sk(ssk)->window_clamp = window_clamp;
1994 tcp_cleanup_rbuf(ssk, 1);
1995 unlock_sock_fast(ssk, slow);
2000 msk->rcvq_space.space = msk->rcvq_space.copied;
2002 msk->rcvq_space.copied = 0;
2003 msk->rcvq_space.time = mstamp;
2006 static void __mptcp_update_rmem(struct sock *sk)
2008 struct mptcp_sock *msk = mptcp_sk(sk);
2010 if (!msk->rmem_released)
2013 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2014 mptcp_rmem_uncharge(sk, msk->rmem_released);
2015 WRITE_ONCE(msk->rmem_released, 0);
2018 static void __mptcp_splice_receive_queue(struct sock *sk)
2020 struct mptcp_sock *msk = mptcp_sk(sk);
2022 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2025 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2027 struct sock *sk = (struct sock *)msk;
2028 unsigned int moved = 0;
2032 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2035 /* we can have data pending in the subflows only if the msk
2036 * receive buffer was full at subflow_data_ready() time,
2037 * that is an unlikely slow path.
2042 slowpath = lock_sock_fast(ssk);
2043 mptcp_data_lock(sk);
2044 __mptcp_update_rmem(sk);
2045 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2046 mptcp_data_unlock(sk);
2048 if (unlikely(ssk->sk_err))
2049 __mptcp_error_report(sk);
2050 unlock_sock_fast(ssk, slowpath);
2053 /* acquire the data lock only if some input data is pending */
2055 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2056 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2057 mptcp_data_lock(sk);
2058 __mptcp_update_rmem(sk);
2059 ret |= __mptcp_ofo_queue(msk);
2060 __mptcp_splice_receive_queue(sk);
2061 mptcp_data_unlock(sk);
2064 mptcp_check_data_fin((struct sock *)msk);
2065 return !skb_queue_empty(&msk->receive_queue);
2068 static unsigned int mptcp_inq_hint(const struct sock *sk)
2070 const struct mptcp_sock *msk = mptcp_sk(sk);
2071 const struct sk_buff *skb;
2073 skb = skb_peek(&msk->receive_queue);
2075 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2077 if (hint_val >= INT_MAX)
2080 return (unsigned int)hint_val;
2083 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2089 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2090 int flags, int *addr_len)
2092 struct mptcp_sock *msk = mptcp_sk(sk);
2093 struct scm_timestamping_internal tss;
2094 int copied = 0, cmsg_flags = 0;
2098 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2099 if (unlikely(flags & MSG_ERRQUEUE))
2100 return inet_recv_error(sk, msg, len, addr_len);
2103 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2108 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2110 len = min_t(size_t, len, INT_MAX);
2111 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2113 if (unlikely(msk->recvmsg_inq))
2114 cmsg_flags = MPTCP_CMSG_INQ;
2116 while (copied < len) {
2119 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2120 if (unlikely(bytes_read < 0)) {
2122 copied = bytes_read;
2126 copied += bytes_read;
2128 /* be sure to advertise window change */
2129 mptcp_cleanup_rbuf(msk);
2131 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2134 /* only the master socket status is relevant here. The exit
2135 * conditions mirror closely tcp_recvmsg()
2137 if (copied >= target)
2142 sk->sk_state == TCP_CLOSE ||
2143 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2145 signal_pending(current))
2149 copied = sock_error(sk);
2153 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2154 mptcp_check_for_eof(msk);
2156 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2157 /* race breaker: the shutdown could be after the
2158 * previous receive queue check
2160 if (__mptcp_move_skbs(msk))
2165 if (sk->sk_state == TCP_CLOSE) {
2175 if (signal_pending(current)) {
2176 copied = sock_intr_errno(timeo);
2181 pr_debug("block timeout %ld", timeo);
2182 sk_wait_data(sk, &timeo, NULL);
2186 if (cmsg_flags && copied >= 0) {
2187 if (cmsg_flags & MPTCP_CMSG_TS)
2188 tcp_recv_timestamp(msg, sk, &tss);
2190 if (cmsg_flags & MPTCP_CMSG_INQ) {
2191 unsigned int inq = mptcp_inq_hint(sk);
2193 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2197 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2198 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2199 skb_queue_empty(&msk->receive_queue), copied);
2200 if (!(flags & MSG_PEEK))
2201 mptcp_rcv_space_adjust(msk, copied);
2207 static void mptcp_retransmit_timer(struct timer_list *t)
2209 struct inet_connection_sock *icsk = from_timer(icsk, t,
2210 icsk_retransmit_timer);
2211 struct sock *sk = &icsk->icsk_inet.sk;
2212 struct mptcp_sock *msk = mptcp_sk(sk);
2215 if (!sock_owned_by_user(sk)) {
2216 /* we need a process context to retransmit */
2217 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2218 mptcp_schedule_work(sk);
2220 /* delegate our work to tcp_release_cb() */
2221 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2227 static void mptcp_timeout_timer(struct timer_list *t)
2229 struct sock *sk = from_timer(sk, t, sk_timer);
2231 mptcp_schedule_work(sk);
2235 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2238 * A backup subflow is returned only if that is the only kind available.
2240 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2242 struct sock *backup = NULL, *pick = NULL;
2243 struct mptcp_subflow_context *subflow;
2244 int min_stale_count = INT_MAX;
2246 sock_owned_by_me((const struct sock *)msk);
2248 if (__mptcp_check_fallback(msk))
2251 mptcp_for_each_subflow(msk, subflow) {
2252 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2254 if (!__mptcp_subflow_active(subflow))
2257 /* still data outstanding at TCP level? skip this */
2258 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2259 mptcp_pm_subflow_chk_stale(msk, ssk);
2260 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2264 if (subflow->backup) {
2277 /* use backup only if there are no progresses anywhere */
2278 return min_stale_count > 1 ? backup : NULL;
2281 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2284 iput(SOCK_INODE(msk->subflow));
2285 WRITE_ONCE(msk->subflow, NULL);
2289 bool __mptcp_retransmit_pending_data(struct sock *sk)
2291 struct mptcp_data_frag *cur, *rtx_head;
2292 struct mptcp_sock *msk = mptcp_sk(sk);
2294 if (__mptcp_check_fallback(mptcp_sk(sk)))
2297 if (tcp_rtx_and_write_queues_empty(sk))
2300 /* the closing socket has some data untransmitted and/or unacked:
2301 * some data in the mptcp rtx queue has not really xmitted yet.
2302 * keep it simple and re-inject the whole mptcp level rtx queue
2304 mptcp_data_lock(sk);
2305 __mptcp_clean_una_wakeup(sk);
2306 rtx_head = mptcp_rtx_head(sk);
2308 mptcp_data_unlock(sk);
2312 msk->recovery_snd_nxt = msk->snd_nxt;
2313 msk->recovery = true;
2314 mptcp_data_unlock(sk);
2316 msk->first_pending = rtx_head;
2319 /* be sure to clear the "sent status" on all re-injected fragments */
2320 list_for_each_entry(cur, &msk->rtx_queue, list) {
2321 if (!cur->already_sent)
2323 cur->already_sent = 0;
2329 /* flags for __mptcp_close_ssk() */
2330 #define MPTCP_CF_PUSH BIT(1)
2331 #define MPTCP_CF_FASTCLOSE BIT(2)
2333 /* subflow sockets can be either outgoing (connect) or incoming
2336 * Outgoing subflows use in-kernel sockets.
2337 * Incoming subflows do not have their own 'struct socket' allocated,
2338 * so we need to use tcp_close() after detaching them from the mptcp
2341 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2342 struct mptcp_subflow_context *subflow,
2345 struct mptcp_sock *msk = mptcp_sk(sk);
2346 bool dispose_it, need_push = false;
2348 /* If the first subflow moved to a close state before accept, e.g. due
2349 * to an incoming reset, mptcp either:
2350 * - if either the subflow or the msk are dead, destroy the context
2351 * (the subflow socket is deleted by inet_child_forget) and the msk
2352 * - otherwise do nothing at the moment and take action at accept and/or
2353 * listener shutdown - user-space must be able to accept() the closed
2356 if (msk->in_accept_queue && msk->first == ssk) {
2357 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2360 /* ensure later check in mptcp_worker() will dispose the msk */
2361 sock_set_flag(sk, SOCK_DEAD);
2362 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2363 mptcp_subflow_drop_ctx(ssk);
2367 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2369 list_del(&subflow->node);
2371 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2373 if (flags & MPTCP_CF_FASTCLOSE) {
2374 /* be sure to force the tcp_disconnect() path,
2375 * to generate the egress reset
2377 ssk->sk_lingertime = 0;
2378 sock_set_flag(ssk, SOCK_LINGER);
2379 subflow->send_fastclose = 1;
2382 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2384 /* The MPTCP code never wait on the subflow sockets, TCP-level
2385 * disconnect should never fail
2387 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2388 msk->subflow->state = SS_UNCONNECTED;
2389 mptcp_subflow_ctx_reset(subflow);
2395 subflow->disposable = 1;
2397 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2398 * the ssk has been already destroyed, we just need to release the
2399 * reference owned by msk;
2401 if (!inet_csk(ssk)->icsk_ulp_ops) {
2402 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2403 kfree_rcu(subflow, rcu);
2405 /* otherwise tcp will dispose of the ssk and subflow ctx */
2406 if (ssk->sk_state == TCP_LISTEN) {
2407 tcp_set_state(ssk, TCP_CLOSE);
2408 mptcp_subflow_queue_clean(sk, ssk);
2409 inet_csk_listen_stop(ssk);
2412 __tcp_close(ssk, 0);
2414 /* close acquired an extra ref */
2423 if (ssk == msk->first)
2424 WRITE_ONCE(msk->first, NULL);
2427 if (ssk == msk->last_snd)
2428 msk->last_snd = NULL;
2431 __mptcp_push_pending(sk, 0);
2434 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2435 struct mptcp_subflow_context *subflow)
2437 if (sk->sk_state == TCP_ESTABLISHED)
2438 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2440 /* subflow aborted before reaching the fully_established status
2441 * attempt the creation of the next subflow
2443 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2445 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2448 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2453 static void __mptcp_close_subflow(struct sock *sk)
2455 struct mptcp_subflow_context *subflow, *tmp;
2456 struct mptcp_sock *msk = mptcp_sk(sk);
2460 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2461 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2463 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2466 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2467 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2470 mptcp_close_ssk(sk, ssk, subflow);
2475 static bool mptcp_should_close(const struct sock *sk)
2477 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2478 struct mptcp_subflow_context *subflow;
2480 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2483 /* if all subflows are in closed status don't bother with additional
2486 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2487 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2494 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2496 struct mptcp_subflow_context *subflow, *tmp;
2497 struct sock *sk = &msk->sk.icsk_inet.sk;
2499 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2502 mptcp_token_destroy(msk);
2504 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2505 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2508 slow = lock_sock_fast(tcp_sk);
2509 if (tcp_sk->sk_state != TCP_CLOSE) {
2510 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2511 tcp_set_state(tcp_sk, TCP_CLOSE);
2513 unlock_sock_fast(tcp_sk, slow);
2516 /* Mirror the tcp_reset() error propagation */
2517 switch (sk->sk_state) {
2519 sk->sk_err = ECONNREFUSED;
2521 case TCP_CLOSE_WAIT:
2527 sk->sk_err = ECONNRESET;
2530 inet_sk_state_store(sk, TCP_CLOSE);
2531 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2532 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2533 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2535 /* the calling mptcp_worker will properly destroy the socket */
2536 if (sock_flag(sk, SOCK_DEAD))
2539 sk->sk_state_change(sk);
2540 sk_error_report(sk);
2543 static void __mptcp_retrans(struct sock *sk)
2545 struct mptcp_sock *msk = mptcp_sk(sk);
2546 struct mptcp_sendmsg_info info = {};
2547 struct mptcp_data_frag *dfrag;
2552 mptcp_clean_una_wakeup(sk);
2554 /* first check ssk: need to kick "stale" logic */
2555 ssk = mptcp_subflow_get_retrans(msk);
2556 dfrag = mptcp_rtx_head(sk);
2558 if (mptcp_data_fin_enabled(msk)) {
2559 struct inet_connection_sock *icsk = inet_csk(sk);
2561 icsk->icsk_retransmits++;
2562 mptcp_set_datafin_timeout(sk);
2563 mptcp_send_ack(msk);
2568 if (!mptcp_send_head(sk))
2579 /* limit retransmission to the bytes already sent on some subflows */
2581 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2582 while (info.sent < info.limit) {
2583 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2587 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2592 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2593 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2595 WRITE_ONCE(msk->allow_infinite_fallback, false);
2601 mptcp_check_and_set_pending(sk);
2603 if (!mptcp_timer_pending(sk))
2604 mptcp_reset_timer(sk);
2607 /* schedule the timeout timer for the relevant event: either close timeout
2608 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2610 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2612 struct sock *sk = (struct sock *)msk;
2613 unsigned long timeout, close_timeout;
2615 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2618 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2620 /* the close timeout takes precedence on the fail one, and here at least one of
2623 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2625 sk_reset_timer(sk, &sk->sk_timer, timeout);
2628 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2630 struct sock *ssk = msk->first;
2636 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2638 slow = lock_sock_fast(ssk);
2639 mptcp_subflow_reset(ssk);
2640 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2641 unlock_sock_fast(ssk, slow);
2643 mptcp_reset_timeout(msk, 0);
2646 static void mptcp_do_fastclose(struct sock *sk)
2648 struct mptcp_subflow_context *subflow, *tmp;
2649 struct mptcp_sock *msk = mptcp_sk(sk);
2651 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2652 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2653 subflow, MPTCP_CF_FASTCLOSE);
2656 static void mptcp_worker(struct work_struct *work)
2658 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2659 struct sock *sk = &msk->sk.icsk_inet.sk;
2660 unsigned long fail_tout;
2664 state = sk->sk_state;
2665 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2668 mptcp_check_data_fin_ack(sk);
2670 mptcp_check_fastclose(msk);
2672 mptcp_pm_nl_work(msk);
2674 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2675 mptcp_check_for_eof(msk);
2677 __mptcp_check_send_data_fin(sk);
2678 mptcp_check_data_fin(sk);
2680 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2681 __mptcp_close_subflow(sk);
2683 /* There is no point in keeping around an orphaned sk timedout or
2684 * closed, but we need the msk around to reply to incoming DATA_FIN,
2685 * even if it is orphaned and in FIN_WAIT2 state
2687 if (sock_flag(sk, SOCK_DEAD)) {
2688 if (mptcp_should_close(sk)) {
2689 inet_sk_state_store(sk, TCP_CLOSE);
2690 mptcp_do_fastclose(sk);
2692 if (sk->sk_state == TCP_CLOSE) {
2693 __mptcp_destroy_sock(sk);
2698 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2699 __mptcp_retrans(sk);
2701 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2702 if (fail_tout && time_after(jiffies, fail_tout))
2703 mptcp_mp_fail_no_response(msk);
2710 static int __mptcp_init_sock(struct sock *sk)
2712 struct mptcp_sock *msk = mptcp_sk(sk);
2714 INIT_LIST_HEAD(&msk->conn_list);
2715 INIT_LIST_HEAD(&msk->join_list);
2716 INIT_LIST_HEAD(&msk->rtx_queue);
2717 INIT_WORK(&msk->work, mptcp_worker);
2718 __skb_queue_head_init(&msk->receive_queue);
2719 msk->out_of_order_queue = RB_ROOT;
2720 msk->first_pending = NULL;
2721 msk->rmem_fwd_alloc = 0;
2722 WRITE_ONCE(msk->rmem_released, 0);
2723 msk->timer_ival = TCP_RTO_MIN;
2725 WRITE_ONCE(msk->first, NULL);
2726 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2727 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2728 WRITE_ONCE(msk->allow_infinite_fallback, true);
2729 msk->recovery = false;
2731 mptcp_pm_data_init(msk);
2733 /* re-use the csk retrans timer for MPTCP-level retrans */
2734 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2735 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2740 static void mptcp_ca_reset(struct sock *sk)
2742 struct inet_connection_sock *icsk = inet_csk(sk);
2744 tcp_assign_congestion_control(sk);
2745 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2747 /* no need to keep a reference to the ops, the name will suffice */
2748 tcp_cleanup_congestion_control(sk);
2749 icsk->icsk_ca_ops = NULL;
2752 static int mptcp_init_sock(struct sock *sk)
2754 struct net *net = sock_net(sk);
2757 ret = __mptcp_init_sock(sk);
2761 if (!mptcp_is_enabled(net))
2762 return -ENOPROTOOPT;
2764 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2767 ret = __mptcp_socket_create(mptcp_sk(sk));
2771 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2772 * propagate the correct value
2776 sk_sockets_allocated_inc(sk);
2777 sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2778 sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2783 static void __mptcp_clear_xmit(struct sock *sk)
2785 struct mptcp_sock *msk = mptcp_sk(sk);
2786 struct mptcp_data_frag *dtmp, *dfrag;
2788 WRITE_ONCE(msk->first_pending, NULL);
2789 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2790 dfrag_clear(sk, dfrag);
2793 void mptcp_cancel_work(struct sock *sk)
2795 struct mptcp_sock *msk = mptcp_sk(sk);
2797 if (cancel_work_sync(&msk->work))
2801 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2805 switch (ssk->sk_state) {
2807 if (!(how & RCV_SHUTDOWN))
2811 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2814 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2815 pr_debug("Fallback");
2816 ssk->sk_shutdown |= how;
2817 tcp_shutdown(ssk, how);
2819 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2821 if (!mptcp_timer_pending(sk))
2822 mptcp_reset_timer(sk);
2830 static const unsigned char new_state[16] = {
2831 /* current state: new state: action: */
2832 [0 /* (Invalid) */] = TCP_CLOSE,
2833 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2834 [TCP_SYN_SENT] = TCP_CLOSE,
2835 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2836 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2837 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2838 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2839 [TCP_CLOSE] = TCP_CLOSE,
2840 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2841 [TCP_LAST_ACK] = TCP_LAST_ACK,
2842 [TCP_LISTEN] = TCP_CLOSE,
2843 [TCP_CLOSING] = TCP_CLOSING,
2844 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2847 static int mptcp_close_state(struct sock *sk)
2849 int next = (int)new_state[sk->sk_state];
2850 int ns = next & TCP_STATE_MASK;
2852 inet_sk_state_store(sk, ns);
2854 return next & TCP_ACTION_FIN;
2857 static void __mptcp_check_send_data_fin(struct sock *sk)
2859 struct mptcp_subflow_context *subflow;
2860 struct mptcp_sock *msk = mptcp_sk(sk);
2862 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2863 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2864 msk->snd_nxt, msk->write_seq);
2866 /* we still need to enqueue subflows or not really shutting down,
2869 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2870 mptcp_send_head(sk))
2873 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2875 /* fallback socket will not get data_fin/ack, can move to the next
2878 if (__mptcp_check_fallback(msk)) {
2879 WRITE_ONCE(msk->snd_una, msk->write_seq);
2880 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2881 inet_sk_state_store(sk, TCP_CLOSE);
2882 mptcp_close_wake_up(sk);
2883 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2884 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2888 mptcp_for_each_subflow(msk, subflow) {
2889 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2891 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2895 static void __mptcp_wr_shutdown(struct sock *sk)
2897 struct mptcp_sock *msk = mptcp_sk(sk);
2899 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2900 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2901 !!mptcp_send_head(sk));
2903 /* will be ignored by fallback sockets */
2904 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2905 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2907 __mptcp_check_send_data_fin(sk);
2910 static void __mptcp_destroy_sock(struct sock *sk)
2912 struct mptcp_sock *msk = mptcp_sk(sk);
2914 pr_debug("msk=%p", msk);
2918 mptcp_stop_timer(sk);
2919 sk_stop_timer(sk, &sk->sk_timer);
2922 sk->sk_prot->destroy(sk);
2924 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2925 WARN_ON_ONCE(msk->rmem_released);
2926 sk_stream_kill_queues(sk);
2927 xfrm_sk_free_policy(sk);
2929 sk_refcnt_debug_release(sk);
2933 void __mptcp_unaccepted_force_close(struct sock *sk)
2935 sock_set_flag(sk, SOCK_DEAD);
2936 inet_sk_state_store(sk, TCP_CLOSE);
2937 mptcp_do_fastclose(sk);
2938 __mptcp_destroy_sock(sk);
2941 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2943 /* Concurrent splices from sk_receive_queue into receive_queue will
2944 * always show at least one non-empty queue when checked in this order.
2946 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2947 skb_queue_empty_lockless(&msk->receive_queue))
2950 return EPOLLIN | EPOLLRDNORM;
2953 bool __mptcp_close(struct sock *sk, long timeout)
2955 struct mptcp_subflow_context *subflow;
2956 struct mptcp_sock *msk = mptcp_sk(sk);
2957 bool do_cancel_work = false;
2958 int subflows_alive = 0;
2960 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2962 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2963 inet_sk_state_store(sk, TCP_CLOSE);
2967 if (mptcp_check_readable(msk)) {
2968 /* the msk has read data, do the MPTCP equivalent of TCP reset */
2969 inet_sk_state_store(sk, TCP_CLOSE);
2970 mptcp_do_fastclose(sk);
2971 } else if (mptcp_close_state(sk)) {
2972 __mptcp_wr_shutdown(sk);
2975 sk_stream_wait_close(sk, timeout);
2978 /* orphan all the subflows */
2979 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2980 mptcp_for_each_subflow(msk, subflow) {
2981 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2982 bool slow = lock_sock_fast_nested(ssk);
2984 subflows_alive += ssk->sk_state != TCP_CLOSE;
2986 /* since the close timeout takes precedence on the fail one,
2989 if (ssk == msk->first)
2990 subflow->fail_tout = 0;
2992 /* detach from the parent socket, but allow data_ready to
2993 * push incoming data into the mptcp stack, to properly ack it
2995 ssk->sk_socket = NULL;
2997 unlock_sock_fast(ssk, slow);
3001 /* all the subflows are closed, only timeout can change the msk
3002 * state, let's not keep resources busy for no reasons
3004 if (subflows_alive == 0)
3005 inet_sk_state_store(sk, TCP_CLOSE);
3008 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3009 if (mptcp_sk(sk)->token)
3010 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3012 if (sk->sk_state == TCP_CLOSE) {
3013 __mptcp_destroy_sock(sk);
3014 do_cancel_work = true;
3016 mptcp_reset_timeout(msk, 0);
3019 return do_cancel_work;
3022 static void mptcp_close(struct sock *sk, long timeout)
3024 bool do_cancel_work;
3028 do_cancel_work = __mptcp_close(sk, timeout);
3031 mptcp_cancel_work(sk);
3036 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3038 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3039 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3040 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3042 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3043 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3046 msk6->saddr = ssk6->saddr;
3047 msk6->flow_label = ssk6->flow_label;
3051 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3052 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3053 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3054 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3055 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3056 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3059 static int mptcp_disconnect(struct sock *sk, int flags)
3061 struct mptcp_sock *msk = mptcp_sk(sk);
3063 /* Deny disconnect if other threads are blocked in sk_wait_event()
3064 * or inet_wait_for_connect().
3066 if (sk->sk_wait_pending)
3069 /* We are on the fastopen error path. We can't call straight into the
3070 * subflows cleanup code due to lock nesting (we are already under
3071 * msk->firstsocket lock).
3073 if (msk->fastopening)
3076 inet_sk_state_store(sk, TCP_CLOSE);
3078 mptcp_stop_timer(sk);
3079 sk_stop_timer(sk, &sk->sk_timer);
3081 if (mptcp_sk(sk)->token)
3082 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
3084 /* msk->subflow is still intact, the following will not free the first
3087 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3088 msk->last_snd = NULL;
3089 WRITE_ONCE(msk->flags, 0);
3091 msk->push_pending = 0;
3092 msk->recovery = false;
3093 msk->can_ack = false;
3094 msk->fully_established = false;
3095 msk->rcv_data_fin = false;
3096 msk->snd_data_fin_enable = false;
3097 msk->rcv_fastclose = false;
3098 msk->use_64bit_ack = false;
3099 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3100 mptcp_pm_data_reset(msk);
3103 WRITE_ONCE(sk->sk_shutdown, 0);
3104 sk_error_report(sk);
3108 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3109 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3111 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3113 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3117 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3118 const struct mptcp_options_received *mp_opt,
3120 struct request_sock *req)
3122 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3123 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3124 struct mptcp_sock *msk;
3130 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3131 if (nsk->sk_family == AF_INET6)
3132 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3135 nsk->sk_wait_pending = 0;
3136 __mptcp_init_sock(nsk);
3138 msk = mptcp_sk(nsk);
3139 msk->local_key = subflow_req->local_key;
3140 msk->token = subflow_req->token;
3141 WRITE_ONCE(msk->subflow, NULL);
3142 msk->in_accept_queue = 1;
3143 WRITE_ONCE(msk->fully_established, false);
3144 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3145 WRITE_ONCE(msk->csum_enabled, true);
3147 msk->write_seq = subflow_req->idsn + 1;
3148 msk->snd_nxt = msk->write_seq;
3149 msk->snd_una = msk->write_seq;
3150 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3151 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3153 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
3154 msk->can_ack = true;
3155 msk->remote_key = mp_opt->sndr_key;
3156 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
3158 WRITE_ONCE(msk->ack_seq, ack_seq);
3159 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3162 sock_reset_flag(nsk, SOCK_RCU_FREE);
3163 security_inet_csk_clone(nsk, req);
3165 /* this can't race with mptcp_close(), as the msk is
3166 * not yet exposted to user-space
3168 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3170 /* The msk maintain a ref to each subflow in the connections list */
3171 WRITE_ONCE(msk->first, ssk);
3172 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3175 /* new mpc subflow takes ownership of the newly
3176 * created mptcp socket
3178 mptcp_token_accept(subflow_req, msk);
3180 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3181 * uses the correct data
3183 mptcp_copy_inaddrs(nsk, ssk);
3184 mptcp_propagate_sndbuf(nsk, ssk);
3186 mptcp_rcv_space_init(msk, ssk);
3187 bh_unlock_sock(nsk);
3189 /* note: the newly allocated socket refcount is 2 now */
3193 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3195 const struct tcp_sock *tp = tcp_sk(ssk);
3197 msk->rcvq_space.copied = 0;
3198 msk->rcvq_space.rtt_us = 0;
3200 msk->rcvq_space.time = tp->tcp_mstamp;
3202 /* initial rcv_space offering made to peer */
3203 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3204 TCP_INIT_CWND * tp->advmss);
3205 if (msk->rcvq_space.space == 0)
3206 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3208 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3211 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3214 struct mptcp_sock *msk = mptcp_sk(sk);
3215 struct socket *listener;
3218 listener = READ_ONCE(msk->subflow);
3219 if (WARN_ON_ONCE(!listener)) {
3224 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3225 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3229 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3230 if (sk_is_mptcp(newsk)) {
3231 struct mptcp_subflow_context *subflow;
3232 struct sock *new_mptcp_sock;
3234 subflow = mptcp_subflow_ctx(newsk);
3235 new_mptcp_sock = subflow->conn;
3237 /* is_mptcp should be false if subflow->conn is missing, see
3238 * subflow_syn_recv_sock()
3240 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3241 tcp_sk(newsk)->is_mptcp = 0;
3245 newsk = new_mptcp_sock;
3246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3248 MPTCP_INC_STATS(sock_net(sk),
3249 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3253 newsk->sk_kern_sock = kern;
3257 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3259 struct mptcp_subflow_context *subflow, *tmp;
3260 struct sock *sk = (struct sock *)msk;
3262 __mptcp_clear_xmit(sk);
3264 /* join list will be eventually flushed (with rst) at sock lock release time */
3265 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3266 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3268 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3269 mptcp_data_lock(sk);
3270 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3271 __skb_queue_purge(&sk->sk_receive_queue);
3272 skb_rbtree_purge(&msk->out_of_order_queue);
3273 mptcp_data_unlock(sk);
3275 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3276 * inet_sock_destruct() will dispose it
3278 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3279 msk->rmem_fwd_alloc = 0;
3280 mptcp_token_destroy(msk);
3281 mptcp_pm_free_anno_list(msk);
3282 mptcp_free_local_addr_list(msk);
3285 static void mptcp_destroy(struct sock *sk)
3287 struct mptcp_sock *msk = mptcp_sk(sk);
3289 /* clears msk->subflow, allowing the following to close
3290 * even the initial subflow
3292 mptcp_dispose_initial_subflow(msk);
3293 mptcp_destroy_common(msk, 0);
3294 sk_sockets_allocated_dec(sk);
3297 void __mptcp_data_acked(struct sock *sk)
3299 if (!sock_owned_by_user(sk))
3300 __mptcp_clean_una(sk);
3302 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3304 if (mptcp_pending_data_fin_ack(sk))
3305 mptcp_schedule_work(sk);
3308 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3310 if (!mptcp_send_head(sk))
3313 if (!sock_owned_by_user(sk)) {
3314 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3316 if (xmit_ssk == ssk)
3317 __mptcp_subflow_push_pending(sk, ssk);
3319 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3321 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3325 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3326 BIT(MPTCP_RETRANSMIT) | \
3327 BIT(MPTCP_FLUSH_JOIN_LIST))
3329 /* processes deferred events and flush wmem */
3330 static void mptcp_release_cb(struct sock *sk)
3331 __must_hold(&sk->sk_lock.slock)
3333 struct mptcp_sock *msk = mptcp_sk(sk);
3336 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3338 struct list_head join_list;
3343 INIT_LIST_HEAD(&join_list);
3344 list_splice_init(&msk->join_list, &join_list);
3346 /* the following actions acquire the subflow socket lock
3348 * 1) can't be invoked in atomic scope
3349 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3350 * datapath acquires the msk socket spinlock while helding
3351 * the subflow socket lock
3353 msk->push_pending = 0;
3354 msk->cb_flags &= ~flags;
3355 spin_unlock_bh(&sk->sk_lock.slock);
3357 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3358 __mptcp_flush_join_list(sk, &join_list);
3359 if (flags & BIT(MPTCP_PUSH_PENDING))
3360 __mptcp_push_pending(sk, 0);
3361 if (flags & BIT(MPTCP_RETRANSMIT))
3362 __mptcp_retrans(sk);
3365 spin_lock_bh(&sk->sk_lock.slock);
3368 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3369 __mptcp_clean_una_wakeup(sk);
3370 if (unlikely(&msk->cb_flags)) {
3371 /* be sure to set the current sk state before tacking actions
3372 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3374 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3375 __mptcp_set_connected(sk);
3376 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3377 __mptcp_error_report(sk);
3378 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3379 msk->last_snd = NULL;
3382 __mptcp_update_rmem(sk);
3385 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3386 * TCP can't schedule delack timer before the subflow is fully established.
3387 * MPTCP uses the delack timer to do 3rd ack retransmissions
3389 static void schedule_3rdack_retransmission(struct sock *ssk)
3391 struct inet_connection_sock *icsk = inet_csk(ssk);
3392 struct tcp_sock *tp = tcp_sk(ssk);
3393 unsigned long timeout;
3395 if (mptcp_subflow_ctx(ssk)->fully_established)
3398 /* reschedule with a timeout above RTT, as we must look only for drop */
3400 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3402 timeout = TCP_TIMEOUT_INIT;
3405 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3406 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3407 icsk->icsk_ack.timeout = timeout;
3408 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3411 void mptcp_subflow_process_delegated(struct sock *ssk)
3413 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3414 struct sock *sk = subflow->conn;
3416 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3417 mptcp_data_lock(sk);
3418 if (!sock_owned_by_user(sk))
3419 __mptcp_subflow_push_pending(sk, ssk);
3421 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3422 mptcp_data_unlock(sk);
3423 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3425 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3426 schedule_3rdack_retransmission(ssk);
3427 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3431 static int mptcp_hash(struct sock *sk)
3433 /* should never be called,
3434 * we hash the TCP subflows not the master socket
3440 static void mptcp_unhash(struct sock *sk)
3442 /* called from sk_common_release(), but nothing to do here */
3445 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3447 struct mptcp_sock *msk = mptcp_sk(sk);
3448 struct socket *ssock;
3450 ssock = msk->subflow;
3451 pr_debug("msk=%p, subflow=%p", msk, ssock);
3452 if (WARN_ON_ONCE(!ssock))
3455 return inet_csk_get_port(ssock->sk, snum);
3458 void mptcp_finish_connect(struct sock *ssk)
3460 struct mptcp_subflow_context *subflow;
3461 struct mptcp_sock *msk;
3465 subflow = mptcp_subflow_ctx(ssk);
3469 pr_debug("msk=%p, token=%u", sk, subflow->token);
3471 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3473 subflow->map_seq = ack_seq;
3474 subflow->map_subflow_seq = 1;
3476 /* the socket is not connected yet, no msk/subflow ops can access/race
3477 * accessing the field below
3479 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3480 WRITE_ONCE(msk->local_key, subflow->local_key);
3481 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3482 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3483 WRITE_ONCE(msk->ack_seq, ack_seq);
3484 WRITE_ONCE(msk->can_ack, 1);
3485 WRITE_ONCE(msk->snd_una, msk->write_seq);
3486 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3488 mptcp_pm_new_connection(msk, ssk, 0);
3490 mptcp_rcv_space_init(msk, ssk);
3493 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3495 write_lock_bh(&sk->sk_callback_lock);
3496 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3497 sk_set_socket(sk, parent);
3498 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3499 write_unlock_bh(&sk->sk_callback_lock);
3502 bool mptcp_finish_join(struct sock *ssk)
3504 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3505 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3506 struct sock *parent = (void *)msk;
3509 pr_debug("msk=%p, subflow=%p", msk, subflow);
3511 /* mptcp socket already closing? */
3512 if (!mptcp_is_fully_established(parent)) {
3513 subflow->reset_reason = MPTCP_RST_EMPTCP;
3517 /* active subflow, already present inside the conn_list */
3518 if (!list_empty(&subflow->node)) {
3519 mptcp_subflow_joined(msk, ssk);
3523 if (!mptcp_pm_allow_new_subflow(msk))
3524 goto err_prohibited;
3526 /* If we can't acquire msk socket lock here, let the release callback
3529 mptcp_data_lock(parent);
3530 if (!sock_owned_by_user(parent)) {
3531 ret = __mptcp_finish_join(msk, ssk);
3534 list_add_tail(&subflow->node, &msk->conn_list);
3538 list_add_tail(&subflow->node, &msk->join_list);
3539 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3541 mptcp_data_unlock(parent);
3545 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3552 static void mptcp_shutdown(struct sock *sk, int how)
3554 pr_debug("sk=%p, how=%d", sk, how);
3556 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3557 __mptcp_wr_shutdown(sk);
3560 static int mptcp_forward_alloc_get(const struct sock *sk)
3562 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3565 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3567 const struct sock *sk = (void *)msk;
3570 if (sk->sk_state == TCP_LISTEN)
3573 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3576 delta = msk->write_seq - v;
3577 if (__mptcp_check_fallback(msk) && msk->first) {
3578 struct tcp_sock *tp = tcp_sk(msk->first);
3580 /* the first subflow is disconnected after close - see
3581 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3582 * so ignore that status, too.
3584 if (!((1 << msk->first->sk_state) &
3585 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3586 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3588 if (delta > INT_MAX)
3594 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3596 struct mptcp_sock *msk = mptcp_sk(sk);
3602 if (sk->sk_state == TCP_LISTEN)
3606 __mptcp_move_skbs(msk);
3607 answ = mptcp_inq_hint(sk);
3611 slow = lock_sock_fast(sk);
3612 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3613 unlock_sock_fast(sk, slow);
3616 slow = lock_sock_fast(sk);
3617 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3618 unlock_sock_fast(sk, slow);
3621 return -ENOIOCTLCMD;
3624 return put_user(answ, (int __user *)arg);
3627 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3628 struct mptcp_subflow_context *subflow)
3630 subflow->request_mptcp = 0;
3631 __mptcp_do_fallback(msk);
3634 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3636 struct mptcp_subflow_context *subflow;
3637 struct mptcp_sock *msk = mptcp_sk(sk);
3638 struct socket *ssock;
3641 ssock = __mptcp_nmpc_socket(msk);
3645 mptcp_token_destroy(msk);
3646 inet_sk_state_store(sk, TCP_SYN_SENT);
3647 subflow = mptcp_subflow_ctx(ssock->sk);
3648 #ifdef CONFIG_TCP_MD5SIG
3649 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3652 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3653 mptcp_subflow_early_fallback(msk, subflow);
3655 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3656 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3657 mptcp_subflow_early_fallback(msk, subflow);
3659 if (likely(!__mptcp_check_fallback(msk)))
3660 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3662 /* if reaching here via the fastopen/sendmsg path, the caller already
3663 * acquired the subflow socket lock, too.
3665 if (msk->fastopening)
3666 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3668 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3669 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3671 /* on successful connect, the msk state will be moved to established by
3672 * subflow_finish_connect()
3674 if (unlikely(err && err != -EINPROGRESS)) {
3675 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3679 mptcp_copy_inaddrs(sk, ssock->sk);
3681 /* silence EINPROGRESS and let the caller inet_stream_connect
3682 * handle the connection in progress
3687 static struct proto mptcp_prot = {
3689 .owner = THIS_MODULE,
3690 .init = mptcp_init_sock,
3691 .connect = mptcp_connect,
3692 .disconnect = mptcp_disconnect,
3693 .close = mptcp_close,
3694 .accept = mptcp_accept,
3695 .setsockopt = mptcp_setsockopt,
3696 .getsockopt = mptcp_getsockopt,
3697 .shutdown = mptcp_shutdown,
3698 .destroy = mptcp_destroy,
3699 .sendmsg = mptcp_sendmsg,
3700 .ioctl = mptcp_ioctl,
3701 .recvmsg = mptcp_recvmsg,
3702 .release_cb = mptcp_release_cb,
3704 .unhash = mptcp_unhash,
3705 .get_port = mptcp_get_port,
3706 .forward_alloc_get = mptcp_forward_alloc_get,
3707 .sockets_allocated = &mptcp_sockets_allocated,
3709 .memory_allocated = &tcp_memory_allocated,
3710 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3712 .memory_pressure = &tcp_memory_pressure,
3713 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3714 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3715 .sysctl_mem = sysctl_tcp_mem,
3716 .obj_size = sizeof(struct mptcp_sock),
3717 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3718 .no_autobind = true,
3721 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3723 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3724 struct socket *ssock;
3727 lock_sock(sock->sk);
3728 ssock = __mptcp_nmpc_socket(msk);
3734 err = ssock->ops->bind(ssock, uaddr, addr_len);
3736 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3739 release_sock(sock->sk);
3743 static int mptcp_listen(struct socket *sock, int backlog)
3745 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3746 struct socket *ssock;
3749 pr_debug("msk=%p", msk);
3751 lock_sock(sock->sk);
3752 ssock = __mptcp_nmpc_socket(msk);
3758 mptcp_token_destroy(msk);
3759 inet_sk_state_store(sock->sk, TCP_LISTEN);
3760 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3762 err = ssock->ops->listen(ssock, backlog);
3763 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3765 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3768 release_sock(sock->sk);
3772 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3773 int flags, bool kern)
3775 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3776 struct socket *ssock;
3779 pr_debug("msk=%p", msk);
3781 /* Buggy applications can call accept on socket states other then LISTEN
3782 * but no need to allocate the first subflow just to error out.
3784 ssock = READ_ONCE(msk->subflow);
3788 err = ssock->ops->accept(sock, newsock, flags, kern);
3789 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3790 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3791 struct mptcp_subflow_context *subflow;
3792 struct sock *newsk = newsock->sk;
3794 msk->in_accept_queue = 0;
3798 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3799 * This is needed so NOSPACE flag can be set from tcp stack.
3801 mptcp_for_each_subflow(msk, subflow) {
3802 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3804 if (!ssk->sk_socket)
3805 mptcp_sock_graft(ssk, newsock);
3808 /* Do late cleanup for the first subflow as necessary. Also
3809 * deal with bad peers not doing a complete shutdown.
3812 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3813 __mptcp_close_ssk(newsk, msk->first,
3814 mptcp_subflow_ctx(msk->first), 0);
3815 if (unlikely(list_empty(&msk->conn_list)))
3816 inet_sk_state_store(newsk, TCP_CLOSE);
3819 release_sock(newsk);
3825 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3827 struct sock *sk = (struct sock *)msk;
3829 if (sk_stream_is_writeable(sk))
3830 return EPOLLOUT | EPOLLWRNORM;
3832 mptcp_set_nospace(sk);
3833 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3834 if (sk_stream_is_writeable(sk))
3835 return EPOLLOUT | EPOLLWRNORM;
3840 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3841 struct poll_table_struct *wait)
3843 struct sock *sk = sock->sk;
3844 struct mptcp_sock *msk;
3850 sock_poll_wait(file, sock, wait);
3852 state = inet_sk_state_load(sk);
3853 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3854 if (state == TCP_LISTEN) {
3855 struct socket *ssock = READ_ONCE(msk->subflow);
3857 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3860 return inet_csk_listen_poll(ssock->sk);
3863 shutdown = READ_ONCE(sk->sk_shutdown);
3864 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3866 if (shutdown & RCV_SHUTDOWN)
3867 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3869 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3870 mask |= mptcp_check_readable(msk);
3871 if (shutdown & SEND_SHUTDOWN)
3872 mask |= EPOLLOUT | EPOLLWRNORM;
3874 mask |= mptcp_check_writeable(msk);
3875 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3876 /* cf tcp_poll() note about TFO */
3877 mask |= EPOLLOUT | EPOLLWRNORM;
3880 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3888 static const struct proto_ops mptcp_stream_ops = {
3890 .owner = THIS_MODULE,
3891 .release = inet_release,
3893 .connect = inet_stream_connect,
3894 .socketpair = sock_no_socketpair,
3895 .accept = mptcp_stream_accept,
3896 .getname = inet_getname,
3898 .ioctl = inet_ioctl,
3899 .gettstamp = sock_gettstamp,
3900 .listen = mptcp_listen,
3901 .shutdown = inet_shutdown,
3902 .setsockopt = sock_common_setsockopt,
3903 .getsockopt = sock_common_getsockopt,
3904 .sendmsg = inet_sendmsg,
3905 .recvmsg = inet_recvmsg,
3906 .mmap = sock_no_mmap,
3907 .sendpage = inet_sendpage,
3910 static struct inet_protosw mptcp_protosw = {
3911 .type = SOCK_STREAM,
3912 .protocol = IPPROTO_MPTCP,
3913 .prot = &mptcp_prot,
3914 .ops = &mptcp_stream_ops,
3915 .flags = INET_PROTOSW_ICSK,
3918 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3920 struct mptcp_delegated_action *delegated;
3921 struct mptcp_subflow_context *subflow;
3924 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3925 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3926 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3928 bh_lock_sock_nested(ssk);
3929 if (!sock_owned_by_user(ssk) &&
3930 mptcp_subflow_has_delegated_action(subflow))
3931 mptcp_subflow_process_delegated(ssk);
3932 /* ... elsewhere tcp_release_cb_override already processed
3933 * the action or will do at next release_sock().
3934 * In both case must dequeue the subflow here - on the same
3935 * CPU that scheduled it.
3937 bh_unlock_sock(ssk);
3940 if (++work_done == budget)
3944 /* always provide a 0 'work_done' argument, so that napi_complete_done
3945 * will not try accessing the NULL napi->dev ptr
3947 napi_complete_done(napi, 0);
3951 void __init mptcp_proto_init(void)
3953 struct mptcp_delegated_action *delegated;
3956 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3958 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3959 panic("Failed to allocate MPTCP pcpu counter\n");
3961 init_dummy_netdev(&mptcp_napi_dev);
3962 for_each_possible_cpu(cpu) {
3963 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3964 INIT_LIST_HEAD(&delegated->head);
3965 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3967 napi_enable(&delegated->napi);
3970 mptcp_subflow_init();
3974 if (proto_register(&mptcp_prot, 1) != 0)
3975 panic("Failed to register MPTCP proto.\n");
3977 inet_register_protosw(&mptcp_protosw);
3979 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3982 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3983 static const struct proto_ops mptcp_v6_stream_ops = {
3985 .owner = THIS_MODULE,
3986 .release = inet6_release,
3988 .connect = inet_stream_connect,
3989 .socketpair = sock_no_socketpair,
3990 .accept = mptcp_stream_accept,
3991 .getname = inet6_getname,
3993 .ioctl = inet6_ioctl,
3994 .gettstamp = sock_gettstamp,
3995 .listen = mptcp_listen,
3996 .shutdown = inet_shutdown,
3997 .setsockopt = sock_common_setsockopt,
3998 .getsockopt = sock_common_getsockopt,
3999 .sendmsg = inet6_sendmsg,
4000 .recvmsg = inet6_recvmsg,
4001 .mmap = sock_no_mmap,
4002 .sendpage = inet_sendpage,
4003 #ifdef CONFIG_COMPAT
4004 .compat_ioctl = inet6_compat_ioctl,
4008 static struct proto mptcp_v6_prot;
4010 static struct inet_protosw mptcp_v6_protosw = {
4011 .type = SOCK_STREAM,
4012 .protocol = IPPROTO_MPTCP,
4013 .prot = &mptcp_v6_prot,
4014 .ops = &mptcp_v6_stream_ops,
4015 .flags = INET_PROTOSW_ICSK,
4018 int __init mptcp_proto_v6_init(void)
4022 mptcp_v6_prot = mptcp_prot;
4023 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4024 mptcp_v6_prot.slab = NULL;
4025 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4027 err = proto_register(&mptcp_v6_prot, 1);
4031 err = inet6_register_protosw(&mptcp_v6_protosw);
4033 proto_unregister(&mptcp_v6_prot);