1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
46 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
49 MPTCP_CMSG_TS = BIT(0),
50 MPTCP_CMSG_INQ = BIT(1),
53 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
55 static void __mptcp_destroy_sock(struct sock *sk);
56 static void __mptcp_check_send_data_fin(struct sock *sk);
58 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
59 static struct net_device mptcp_napi_dev;
61 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
62 * completed yet or has failed, return the subflow socket.
63 * Otherwise return NULL.
65 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
67 if (!msk->subflow || READ_ONCE(msk->can_ack))
73 /* Returns end sequence number of the receiver's advertised window */
74 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
76 return READ_ONCE(msk->wnd_end);
79 static bool mptcp_is_tcpsk(struct sock *sk)
81 struct socket *sock = sk->sk_socket;
83 if (unlikely(sk->sk_prot == &tcp_prot)) {
84 /* we are being invoked after mptcp_accept() has
85 * accepted a non-mp-capable flow: sk is a tcp_sk,
88 * Hand the socket over to tcp so all further socket ops
91 sock->ops = &inet_stream_ops;
93 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
94 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
95 sock->ops = &inet6_stream_ops;
103 static int __mptcp_socket_create(struct mptcp_sock *msk)
105 struct mptcp_subflow_context *subflow;
106 struct sock *sk = (struct sock *)msk;
107 struct socket *ssock;
110 err = mptcp_subflow_create_socket(sk, &ssock);
114 msk->first = ssock->sk;
115 msk->subflow = ssock;
116 subflow = mptcp_subflow_ctx(ssock->sk);
117 list_add(&subflow->node, &msk->conn_list);
118 sock_hold(ssock->sk);
119 subflow->request_mptcp = 1;
121 /* This is the first subflow, always with id 0 */
122 subflow->local_id_valid = 1;
123 mptcp_sock_graft(msk->first, sk->sk_socket);
128 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
130 sk_drops_add(sk, skb);
134 static void mptcp_rmem_charge(struct sock *sk, int size)
136 mptcp_sk(sk)->rmem_fwd_alloc -= size;
139 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
140 struct sk_buff *from)
145 if (MPTCP_SKB_CB(from)->offset ||
146 !skb_try_coalesce(to, from, &fragstolen, &delta))
149 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
150 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
151 to->len, MPTCP_SKB_CB(from)->end_seq);
152 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
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 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 bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
826 struct sock *sk = (struct sock *)msk;
828 if (sk->sk_state != TCP_ESTABLISHED)
831 /* attach to msk socket only after we are sure we will deal with it
834 if (sk->sk_socket && !ssk->sk_socket)
835 mptcp_sock_graft(ssk, sk->sk_socket);
837 mptcp_propagate_sndbuf((struct sock *)msk, ssk);
838 mptcp_sockopt_sync_locked(msk, ssk);
842 static void __mptcp_flush_join_list(struct sock *sk)
844 struct mptcp_subflow_context *tmp, *subflow;
845 struct mptcp_sock *msk = mptcp_sk(sk);
847 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
848 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
849 bool slow = lock_sock_fast(ssk);
851 list_move_tail(&subflow->node, &msk->conn_list);
852 if (!__mptcp_finish_join(msk, ssk))
853 mptcp_subflow_reset(ssk);
854 unlock_sock_fast(ssk, slow);
858 static bool mptcp_timer_pending(struct sock *sk)
860 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
863 static void mptcp_reset_timer(struct sock *sk)
865 struct inet_connection_sock *icsk = inet_csk(sk);
868 /* prevent rescheduling on close */
869 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
872 tout = mptcp_sk(sk)->timer_ival;
873 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
876 bool mptcp_schedule_work(struct sock *sk)
878 if (inet_sk_state_load(sk) != TCP_CLOSE &&
879 schedule_work(&mptcp_sk(sk)->work)) {
880 /* each subflow already holds a reference to the sk, and the
881 * workqueue is invoked by a subflow, so sk can't go away here.
889 void mptcp_subflow_eof(struct sock *sk)
891 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
892 mptcp_schedule_work(sk);
895 static void mptcp_check_for_eof(struct mptcp_sock *msk)
897 struct mptcp_subflow_context *subflow;
898 struct sock *sk = (struct sock *)msk;
901 mptcp_for_each_subflow(msk, subflow)
902 receivers += !subflow->rx_eof;
906 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
907 /* hopefully temporary hack: propagate shutdown status
908 * to msk, when all subflows agree on it
910 sk->sk_shutdown |= RCV_SHUTDOWN;
912 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
913 sk->sk_data_ready(sk);
916 switch (sk->sk_state) {
917 case TCP_ESTABLISHED:
918 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
921 inet_sk_state_store(sk, TCP_CLOSING);
924 inet_sk_state_store(sk, TCP_CLOSE);
929 mptcp_close_wake_up(sk);
932 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
934 struct mptcp_subflow_context *subflow;
935 struct sock *sk = (struct sock *)msk;
937 sock_owned_by_me(sk);
939 mptcp_for_each_subflow(msk, subflow) {
940 if (READ_ONCE(subflow->data_avail))
941 return mptcp_subflow_tcp_sock(subflow);
947 static bool mptcp_skb_can_collapse_to(u64 write_seq,
948 const struct sk_buff *skb,
949 const struct mptcp_ext *mpext)
951 if (!tcp_skb_can_collapse_to(skb))
954 /* can collapse only if MPTCP level sequence is in order and this
955 * mapping has not been xmitted yet
957 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
961 /* we can append data to the given data frag if:
962 * - there is space available in the backing page_frag
963 * - the data frag tail matches the current page_frag free offset
964 * - the data frag end sequence number matches the current write seq
966 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
967 const struct page_frag *pfrag,
968 const struct mptcp_data_frag *df)
970 return df && pfrag->page == df->page &&
971 pfrag->size - pfrag->offset > 0 &&
972 pfrag->offset == (df->offset + df->data_len) &&
973 df->data_seq + df->data_len == msk->write_seq;
976 static void dfrag_uncharge(struct sock *sk, int len)
978 sk_mem_uncharge(sk, len);
979 sk_wmem_queued_add(sk, -len);
982 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
984 int len = dfrag->data_len + dfrag->overhead;
986 list_del(&dfrag->list);
987 dfrag_uncharge(sk, len);
988 put_page(dfrag->page);
991 static void __mptcp_clean_una(struct sock *sk)
993 struct mptcp_sock *msk = mptcp_sk(sk);
994 struct mptcp_data_frag *dtmp, *dfrag;
997 /* on fallback we just need to ignore snd_una, as this is really
1000 if (__mptcp_check_fallback(msk))
1001 msk->snd_una = READ_ONCE(msk->snd_nxt);
1003 snd_una = msk->snd_una;
1004 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1005 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1008 if (unlikely(dfrag == msk->first_pending)) {
1009 /* in recovery mode can see ack after the current snd head */
1010 if (WARN_ON_ONCE(!msk->recovery))
1013 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1016 dfrag_clear(sk, dfrag);
1019 dfrag = mptcp_rtx_head(sk);
1020 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1021 u64 delta = snd_una - dfrag->data_seq;
1023 /* prevent wrap around in recovery mode */
1024 if (unlikely(delta > dfrag->already_sent)) {
1025 if (WARN_ON_ONCE(!msk->recovery))
1027 if (WARN_ON_ONCE(delta > dfrag->data_len))
1029 dfrag->already_sent += delta - dfrag->already_sent;
1032 dfrag->data_seq += delta;
1033 dfrag->offset += delta;
1034 dfrag->data_len -= delta;
1035 dfrag->already_sent -= delta;
1037 dfrag_uncharge(sk, delta);
1040 /* all retransmitted data acked, recovery completed */
1041 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1042 msk->recovery = false;
1045 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1046 snd_una == READ_ONCE(msk->write_seq)) {
1047 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1048 mptcp_stop_timer(sk);
1050 mptcp_reset_timer(sk);
1054 static void __mptcp_clean_una_wakeup(struct sock *sk)
1056 lockdep_assert_held_once(&sk->sk_lock.slock);
1058 __mptcp_clean_una(sk);
1059 mptcp_write_space(sk);
1062 static void mptcp_clean_una_wakeup(struct sock *sk)
1064 mptcp_data_lock(sk);
1065 __mptcp_clean_una_wakeup(sk);
1066 mptcp_data_unlock(sk);
1069 static void mptcp_enter_memory_pressure(struct sock *sk)
1071 struct mptcp_subflow_context *subflow;
1072 struct mptcp_sock *msk = mptcp_sk(sk);
1075 sk_stream_moderate_sndbuf(sk);
1076 mptcp_for_each_subflow(msk, subflow) {
1077 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1080 tcp_enter_memory_pressure(ssk);
1081 sk_stream_moderate_sndbuf(ssk);
1086 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1089 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1091 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1092 pfrag, sk->sk_allocation)))
1095 mptcp_enter_memory_pressure(sk);
1099 static struct mptcp_data_frag *
1100 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1103 int offset = ALIGN(orig_offset, sizeof(long));
1104 struct mptcp_data_frag *dfrag;
1106 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1107 dfrag->data_len = 0;
1108 dfrag->data_seq = msk->write_seq;
1109 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1110 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1111 dfrag->already_sent = 0;
1112 dfrag->page = pfrag->page;
1117 struct mptcp_sendmsg_info {
1123 bool data_lock_held;
1126 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1127 u64 data_seq, int avail_size)
1129 u64 window_end = mptcp_wnd_end(msk);
1132 if (__mptcp_check_fallback(msk))
1135 mptcp_snd_wnd = window_end - data_seq;
1136 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1138 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1139 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1140 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1146 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1148 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1152 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1156 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1158 struct sk_buff *skb;
1160 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1162 if (likely(__mptcp_add_ext(skb, gfp))) {
1163 skb_reserve(skb, MAX_TCP_HEADER);
1164 skb->ip_summed = CHECKSUM_PARTIAL;
1165 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1170 mptcp_enter_memory_pressure(sk);
1175 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1177 struct sk_buff *skb;
1179 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1183 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1184 tcp_skb_entail(ssk, skb);
1187 tcp_skb_tsorted_anchor_cleanup(skb);
1192 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1194 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1196 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1199 /* note: this always recompute the csum on the whole skb, even
1200 * if we just appended a single frag. More status info needed
1202 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1204 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1205 __wsum csum = ~csum_unfold(mpext->csum);
1206 int offset = skb->len - added;
1208 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1211 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1213 struct mptcp_ext *mpext)
1218 mpext->infinite_map = 1;
1219 mpext->data_len = 0;
1221 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1222 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1224 mptcp_do_fallback(ssk);
1227 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1228 struct mptcp_data_frag *dfrag,
1229 struct mptcp_sendmsg_info *info)
1231 u64 data_seq = dfrag->data_seq + info->sent;
1232 int offset = dfrag->offset + info->sent;
1233 struct mptcp_sock *msk = mptcp_sk(sk);
1234 bool zero_window_probe = false;
1235 struct mptcp_ext *mpext = NULL;
1236 bool can_coalesce = false;
1237 bool reuse_skb = true;
1238 struct sk_buff *skb;
1242 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1243 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1245 if (WARN_ON_ONCE(info->sent > info->limit ||
1246 info->limit > dfrag->data_len))
1249 if (unlikely(!__tcp_can_send(ssk)))
1252 /* compute send limit */
1253 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1254 copy = info->size_goal;
1256 skb = tcp_write_queue_tail(ssk);
1257 if (skb && copy > skb->len) {
1258 /* Limit the write to the size available in the
1259 * current skb, if any, so that we create at most a new skb.
1260 * Explicitly tells TCP internals to avoid collapsing on later
1261 * queue management operation, to avoid breaking the ext <->
1262 * SSN association set here
1264 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1265 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1266 TCP_SKB_CB(skb)->eor = 1;
1270 i = skb_shinfo(skb)->nr_frags;
1271 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1272 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1273 tcp_mark_push(tcp_sk(ssk), skb);
1280 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1284 i = skb_shinfo(skb)->nr_frags;
1286 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1289 /* Zero window and all data acked? Probe. */
1290 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1292 u64 snd_una = READ_ONCE(msk->snd_una);
1294 if (snd_una != msk->snd_nxt) {
1295 tcp_remove_empty_skb(ssk);
1299 zero_window_probe = true;
1300 data_seq = snd_una - 1;
1303 /* all mptcp-level data is acked, no skbs should be present into the
1306 WARN_ON_ONCE(reuse_skb);
1309 copy = min_t(size_t, copy, info->limit - info->sent);
1310 if (!sk_wmem_schedule(ssk, copy)) {
1311 tcp_remove_empty_skb(ssk);
1316 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1318 get_page(dfrag->page);
1319 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1323 skb->data_len += copy;
1324 skb->truesize += copy;
1325 sk_wmem_queued_add(ssk, copy);
1326 sk_mem_charge(ssk, copy);
1327 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1328 TCP_SKB_CB(skb)->end_seq += copy;
1329 tcp_skb_pcount_set(skb, 0);
1331 /* on skb reuse we just need to update the DSS len */
1333 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1334 mpext->data_len += copy;
1335 WARN_ON_ONCE(zero_window_probe);
1339 memset(mpext, 0, sizeof(*mpext));
1340 mpext->data_seq = data_seq;
1341 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1342 mpext->data_len = copy;
1346 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1347 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1350 if (zero_window_probe) {
1351 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1353 if (READ_ONCE(msk->csum_enabled))
1354 mptcp_update_data_checksum(skb, copy);
1355 tcp_push_pending_frames(ssk);
1359 if (READ_ONCE(msk->csum_enabled))
1360 mptcp_update_data_checksum(skb, copy);
1361 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1362 mptcp_update_infinite_map(msk, ssk, mpext);
1363 trace_mptcp_sendmsg_frag(mpext);
1364 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1368 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1369 sizeof(struct tcphdr) - \
1370 MAX_TCP_OPTION_SPACE - \
1371 sizeof(struct ipv6hdr) - \
1372 sizeof(struct frag_hdr))
1374 struct subflow_send_info {
1379 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1381 if (!subflow->stale)
1385 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1388 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1390 if (unlikely(subflow->stale)) {
1391 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1393 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1396 mptcp_subflow_set_active(subflow);
1398 return __mptcp_subflow_active(subflow);
1401 #define SSK_MODE_ACTIVE 0
1402 #define SSK_MODE_BACKUP 1
1403 #define SSK_MODE_MAX 2
1405 /* implement the mptcp packet scheduler;
1406 * returns the subflow that will transmit the next DSS
1407 * additionally updates the rtx timeout
1409 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1411 struct subflow_send_info send_info[SSK_MODE_MAX];
1412 struct mptcp_subflow_context *subflow;
1413 struct sock *sk = (struct sock *)msk;
1414 u32 pace, burst, wmem;
1415 int i, nr_active = 0;
1420 sock_owned_by_me(sk);
1422 if (__mptcp_check_fallback(msk)) {
1425 return __tcp_can_send(msk->first) &&
1426 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1429 /* re-use last subflow, if the burst allow that */
1430 if (msk->last_snd && msk->snd_burst > 0 &&
1431 sk_stream_memory_free(msk->last_snd) &&
1432 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1433 mptcp_set_timeout(sk);
1434 return msk->last_snd;
1437 /* pick the subflow with the lower wmem/wspace ratio */
1438 for (i = 0; i < SSK_MODE_MAX; ++i) {
1439 send_info[i].ssk = NULL;
1440 send_info[i].linger_time = -1;
1443 mptcp_for_each_subflow(msk, subflow) {
1444 trace_mptcp_subflow_get_send(subflow);
1445 ssk = mptcp_subflow_tcp_sock(subflow);
1446 if (!mptcp_subflow_active(subflow))
1449 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1450 nr_active += !subflow->backup;
1451 pace = subflow->avg_pacing_rate;
1452 if (unlikely(!pace)) {
1453 /* init pacing rate from socket */
1454 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1455 pace = subflow->avg_pacing_rate;
1460 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1461 if (linger_time < send_info[subflow->backup].linger_time) {
1462 send_info[subflow->backup].ssk = ssk;
1463 send_info[subflow->backup].linger_time = linger_time;
1466 __mptcp_set_timeout(sk, tout);
1468 /* pick the best backup if no other subflow is active */
1470 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1472 /* According to the blest algorithm, to avoid HoL blocking for the
1473 * faster flow, we need to:
1474 * - estimate the faster flow linger time
1475 * - use the above to estimate the amount of byte transferred
1476 * by the faster flow
1477 * - check that the amount of queued data is greter than the above,
1478 * otherwise do not use the picked, slower, subflow
1479 * We select the subflow with the shorter estimated time to flush
1480 * the queued mem, which basically ensure the above. We just need
1481 * to check that subflow has a non empty cwin.
1483 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1484 if (!ssk || !sk_stream_memory_free(ssk))
1487 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1488 wmem = READ_ONCE(ssk->sk_wmem_queued);
1490 msk->last_snd = NULL;
1494 subflow = mptcp_subflow_ctx(ssk);
1495 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1496 READ_ONCE(ssk->sk_pacing_rate) * burst,
1498 msk->last_snd = ssk;
1499 msk->snd_burst = burst;
1503 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1505 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1509 static void mptcp_update_post_push(struct mptcp_sock *msk,
1510 struct mptcp_data_frag *dfrag,
1513 u64 snd_nxt_new = dfrag->data_seq;
1515 dfrag->already_sent += sent;
1517 msk->snd_burst -= sent;
1519 snd_nxt_new += dfrag->already_sent;
1521 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1522 * is recovering after a failover. In that event, this re-sends
1525 * Thus compute snd_nxt_new candidate based on
1526 * the dfrag->data_seq that was sent and the data
1527 * that has been handed to the subflow for transmission
1528 * and skip update in case it was old dfrag.
1530 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1531 msk->snd_nxt = snd_nxt_new;
1534 void mptcp_check_and_set_pending(struct sock *sk)
1536 if (mptcp_send_head(sk))
1537 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1540 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1542 struct sock *prev_ssk = NULL, *ssk = NULL;
1543 struct mptcp_sock *msk = mptcp_sk(sk);
1544 struct mptcp_sendmsg_info info = {
1547 bool do_check_data_fin = false;
1548 struct mptcp_data_frag *dfrag;
1551 while ((dfrag = mptcp_send_head(sk))) {
1552 info.sent = dfrag->already_sent;
1553 info.limit = dfrag->data_len;
1554 len = dfrag->data_len - dfrag->already_sent;
1559 ssk = mptcp_subflow_get_send(msk);
1561 /* First check. If the ssk has changed since
1562 * the last round, release prev_ssk
1564 if (ssk != prev_ssk && prev_ssk)
1565 mptcp_push_release(prev_ssk, &info);
1569 /* Need to lock the new subflow only if different
1570 * from the previous one, otherwise we are still
1571 * helding the relevant lock
1573 if (ssk != prev_ssk)
1576 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1580 mptcp_push_release(ssk, &info);
1584 do_check_data_fin = true;
1588 mptcp_update_post_push(msk, dfrag, ret);
1590 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1593 /* at this point we held the socket lock for the last subflow we used */
1595 mptcp_push_release(ssk, &info);
1598 /* ensure the rtx timer is running */
1599 if (!mptcp_timer_pending(sk))
1600 mptcp_reset_timer(sk);
1601 if (do_check_data_fin)
1602 __mptcp_check_send_data_fin(sk);
1605 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1607 struct mptcp_sock *msk = mptcp_sk(sk);
1608 struct mptcp_sendmsg_info info = {
1609 .data_lock_held = true,
1611 struct mptcp_data_frag *dfrag;
1612 struct sock *xmit_ssk;
1613 int len, copied = 0;
1617 while ((dfrag = mptcp_send_head(sk))) {
1618 info.sent = dfrag->already_sent;
1619 info.limit = dfrag->data_len;
1620 len = dfrag->data_len - dfrag->already_sent;
1624 /* the caller already invoked the packet scheduler,
1625 * check for a different subflow usage only after
1626 * spooling the first chunk of data
1628 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1631 if (xmit_ssk != ssk) {
1632 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1633 MPTCP_DELEGATE_SEND);
1637 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1646 mptcp_update_post_push(msk, dfrag, ret);
1648 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1652 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1653 * not going to flush it via release_sock()
1656 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1658 if (!mptcp_timer_pending(sk))
1659 mptcp_reset_timer(sk);
1661 if (msk->snd_data_fin_enable &&
1662 msk->snd_nxt + 1 == msk->write_seq)
1663 mptcp_schedule_work(sk);
1667 static void mptcp_set_nospace(struct sock *sk)
1669 /* enable autotune */
1670 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1672 /* will be cleared on avail space */
1673 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1676 static int mptcp_sendmsg_fastopen(struct sock *sk, struct sock *ssk, struct msghdr *msg,
1677 size_t len, int *copied_syn)
1679 unsigned int saved_flags = msg->msg_flags;
1680 struct mptcp_sock *msk = mptcp_sk(sk);
1684 msg->msg_flags |= MSG_DONTWAIT;
1685 msk->connect_flags = O_NONBLOCK;
1686 msk->is_sendmsg = 1;
1687 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1688 msk->is_sendmsg = 0;
1689 msg->msg_flags = saved_flags;
1692 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1693 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1694 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1695 msg->msg_namelen, msg->msg_flags, 1);
1697 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1698 * case of any error, except timeout or signal
1700 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1707 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1709 struct mptcp_sock *msk = mptcp_sk(sk);
1710 struct page_frag *pfrag;
1711 struct socket *ssock;
1716 /* we don't support FASTOPEN yet */
1717 if (msg->msg_flags & MSG_FASTOPEN)
1720 /* silently ignore everything else */
1721 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1725 ssock = __mptcp_nmpc_socket(msk);
1726 if (unlikely(ssock && inet_sk(ssock->sk)->defer_connect)) {
1729 ret = mptcp_sendmsg_fastopen(sk, ssock->sk, msg, len, &copied_syn);
1730 copied += copied_syn;
1731 if (ret == -EINPROGRESS && copied_syn > 0)
1737 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1739 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1740 ret = sk_stream_wait_connect(sk, &timeo);
1746 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1749 pfrag = sk_page_frag(sk);
1751 while (msg_data_left(msg)) {
1752 int total_ts, frag_truesize = 0;
1753 struct mptcp_data_frag *dfrag;
1754 bool dfrag_collapsed;
1755 size_t psize, offset;
1757 /* reuse tail pfrag, if possible, or carve a new one from the
1760 dfrag = mptcp_pending_tail(sk);
1761 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1762 if (!dfrag_collapsed) {
1763 if (!sk_stream_memory_free(sk))
1764 goto wait_for_memory;
1766 if (!mptcp_page_frag_refill(sk, pfrag))
1767 goto wait_for_memory;
1769 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1770 frag_truesize = dfrag->overhead;
1773 /* we do not bound vs wspace, to allow a single packet.
1774 * memory accounting will prevent execessive memory usage
1777 offset = dfrag->offset + dfrag->data_len;
1778 psize = pfrag->size - offset;
1779 psize = min_t(size_t, psize, msg_data_left(msg));
1780 total_ts = psize + frag_truesize;
1782 if (!sk_wmem_schedule(sk, total_ts))
1783 goto wait_for_memory;
1785 if (copy_page_from_iter(dfrag->page, offset, psize,
1786 &msg->msg_iter) != psize) {
1791 /* data successfully copied into the write queue */
1792 sk->sk_forward_alloc -= total_ts;
1794 dfrag->data_len += psize;
1795 frag_truesize += psize;
1796 pfrag->offset += frag_truesize;
1797 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1799 /* charge data on mptcp pending queue to the msk socket
1800 * Note: we charge such data both to sk and ssk
1802 sk_wmem_queued_add(sk, frag_truesize);
1803 if (!dfrag_collapsed) {
1804 get_page(dfrag->page);
1805 list_add_tail(&dfrag->list, &msk->rtx_queue);
1806 if (!msk->first_pending)
1807 WRITE_ONCE(msk->first_pending, dfrag);
1809 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1810 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1816 mptcp_set_nospace(sk);
1817 __mptcp_push_pending(sk, msg->msg_flags);
1818 ret = sk_stream_wait_memory(sk, &timeo);
1824 __mptcp_push_pending(sk, msg->msg_flags);
1834 copied = sk_stream_error(sk, msg->msg_flags, ret);
1838 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1840 size_t len, int flags,
1841 struct scm_timestamping_internal *tss,
1844 struct sk_buff *skb, *tmp;
1847 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1848 u32 offset = MPTCP_SKB_CB(skb)->offset;
1849 u32 data_len = skb->len - offset;
1850 u32 count = min_t(size_t, len - copied, data_len);
1853 if (!(flags & MSG_TRUNC)) {
1854 err = skb_copy_datagram_msg(skb, offset, msg, count);
1855 if (unlikely(err < 0)) {
1862 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1863 tcp_update_recv_tstamps(skb, tss);
1864 *cmsg_flags |= MPTCP_CMSG_TS;
1869 if (count < data_len) {
1870 if (!(flags & MSG_PEEK)) {
1871 MPTCP_SKB_CB(skb)->offset += count;
1872 MPTCP_SKB_CB(skb)->map_seq += count;
1877 if (!(flags & MSG_PEEK)) {
1878 /* we will bulk release the skb memory later */
1879 skb->destructor = NULL;
1880 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1881 __skb_unlink(skb, &msk->receive_queue);
1892 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1894 * Only difference: Use highest rtt estimate of the subflows in use.
1896 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1898 struct mptcp_subflow_context *subflow;
1899 struct sock *sk = (struct sock *)msk;
1900 u32 time, advmss = 1;
1903 sock_owned_by_me(sk);
1908 msk->rcvq_space.copied += copied;
1910 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1911 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1913 rtt_us = msk->rcvq_space.rtt_us;
1914 if (rtt_us && time < (rtt_us >> 3))
1918 mptcp_for_each_subflow(msk, subflow) {
1919 const struct tcp_sock *tp;
1923 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1925 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1926 sf_advmss = READ_ONCE(tp->advmss);
1928 rtt_us = max(sf_rtt_us, rtt_us);
1929 advmss = max(sf_advmss, advmss);
1932 msk->rcvq_space.rtt_us = rtt_us;
1933 if (time < (rtt_us >> 3) || rtt_us == 0)
1936 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1939 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1940 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1944 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1946 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1948 do_div(grow, msk->rcvq_space.space);
1949 rcvwin += (grow << 1);
1951 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1952 while (tcp_win_from_space(sk, rcvmem) < advmss)
1955 do_div(rcvwin, advmss);
1956 rcvbuf = min_t(u64, rcvwin * rcvmem,
1957 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1959 if (rcvbuf > sk->sk_rcvbuf) {
1962 window_clamp = tcp_win_from_space(sk, rcvbuf);
1963 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1965 /* Make subflows follow along. If we do not do this, we
1966 * get drops at subflow level if skbs can't be moved to
1967 * the mptcp rx queue fast enough (announced rcv_win can
1968 * exceed ssk->sk_rcvbuf).
1970 mptcp_for_each_subflow(msk, subflow) {
1974 ssk = mptcp_subflow_tcp_sock(subflow);
1975 slow = lock_sock_fast(ssk);
1976 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1977 tcp_sk(ssk)->window_clamp = window_clamp;
1978 tcp_cleanup_rbuf(ssk, 1);
1979 unlock_sock_fast(ssk, slow);
1984 msk->rcvq_space.space = msk->rcvq_space.copied;
1986 msk->rcvq_space.copied = 0;
1987 msk->rcvq_space.time = mstamp;
1990 static void __mptcp_update_rmem(struct sock *sk)
1992 struct mptcp_sock *msk = mptcp_sk(sk);
1994 if (!msk->rmem_released)
1997 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1998 mptcp_rmem_uncharge(sk, msk->rmem_released);
1999 WRITE_ONCE(msk->rmem_released, 0);
2002 static void __mptcp_splice_receive_queue(struct sock *sk)
2004 struct mptcp_sock *msk = mptcp_sk(sk);
2006 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2009 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2011 struct sock *sk = (struct sock *)msk;
2012 unsigned int moved = 0;
2016 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2019 /* we can have data pending in the subflows only if the msk
2020 * receive buffer was full at subflow_data_ready() time,
2021 * that is an unlikely slow path.
2026 slowpath = lock_sock_fast(ssk);
2027 mptcp_data_lock(sk);
2028 __mptcp_update_rmem(sk);
2029 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2030 mptcp_data_unlock(sk);
2032 if (unlikely(ssk->sk_err))
2033 __mptcp_error_report(sk);
2034 unlock_sock_fast(ssk, slowpath);
2037 /* acquire the data lock only if some input data is pending */
2039 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2040 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2041 mptcp_data_lock(sk);
2042 __mptcp_update_rmem(sk);
2043 ret |= __mptcp_ofo_queue(msk);
2044 __mptcp_splice_receive_queue(sk);
2045 mptcp_data_unlock(sk);
2048 mptcp_check_data_fin((struct sock *)msk);
2049 return !skb_queue_empty(&msk->receive_queue);
2052 static unsigned int mptcp_inq_hint(const struct sock *sk)
2054 const struct mptcp_sock *msk = mptcp_sk(sk);
2055 const struct sk_buff *skb;
2057 skb = skb_peek(&msk->receive_queue);
2059 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2061 if (hint_val >= INT_MAX)
2064 return (unsigned int)hint_val;
2067 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2073 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2074 int flags, int *addr_len)
2076 struct mptcp_sock *msk = mptcp_sk(sk);
2077 struct scm_timestamping_internal tss;
2078 int copied = 0, cmsg_flags = 0;
2082 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2083 if (unlikely(flags & MSG_ERRQUEUE))
2084 return inet_recv_error(sk, msg, len, addr_len);
2087 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2092 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2094 len = min_t(size_t, len, INT_MAX);
2095 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2097 if (unlikely(msk->recvmsg_inq))
2098 cmsg_flags = MPTCP_CMSG_INQ;
2100 while (copied < len) {
2103 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2104 if (unlikely(bytes_read < 0)) {
2106 copied = bytes_read;
2110 copied += bytes_read;
2112 /* be sure to advertise window change */
2113 mptcp_cleanup_rbuf(msk);
2115 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2118 /* only the master socket status is relevant here. The exit
2119 * conditions mirror closely tcp_recvmsg()
2121 if (copied >= target)
2126 sk->sk_state == TCP_CLOSE ||
2127 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2129 signal_pending(current))
2133 copied = sock_error(sk);
2137 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2138 mptcp_check_for_eof(msk);
2140 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2141 /* race breaker: the shutdown could be after the
2142 * previous receive queue check
2144 if (__mptcp_move_skbs(msk))
2149 if (sk->sk_state == TCP_CLOSE) {
2159 if (signal_pending(current)) {
2160 copied = sock_intr_errno(timeo);
2165 pr_debug("block timeout %ld", timeo);
2166 sk_wait_data(sk, &timeo, NULL);
2170 if (cmsg_flags && copied >= 0) {
2171 if (cmsg_flags & MPTCP_CMSG_TS)
2172 tcp_recv_timestamp(msg, sk, &tss);
2174 if (cmsg_flags & MPTCP_CMSG_INQ) {
2175 unsigned int inq = mptcp_inq_hint(sk);
2177 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2181 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2182 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2183 skb_queue_empty(&msk->receive_queue), copied);
2184 if (!(flags & MSG_PEEK))
2185 mptcp_rcv_space_adjust(msk, copied);
2191 static void mptcp_retransmit_timer(struct timer_list *t)
2193 struct inet_connection_sock *icsk = from_timer(icsk, t,
2194 icsk_retransmit_timer);
2195 struct sock *sk = &icsk->icsk_inet.sk;
2196 struct mptcp_sock *msk = mptcp_sk(sk);
2199 if (!sock_owned_by_user(sk)) {
2200 /* we need a process context to retransmit */
2201 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2202 mptcp_schedule_work(sk);
2204 /* delegate our work to tcp_release_cb() */
2205 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2211 static void mptcp_timeout_timer(struct timer_list *t)
2213 struct sock *sk = from_timer(sk, t, sk_timer);
2215 mptcp_schedule_work(sk);
2219 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2222 * A backup subflow is returned only if that is the only kind available.
2224 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2226 struct sock *backup = NULL, *pick = NULL;
2227 struct mptcp_subflow_context *subflow;
2228 int min_stale_count = INT_MAX;
2230 sock_owned_by_me((const struct sock *)msk);
2232 if (__mptcp_check_fallback(msk))
2235 mptcp_for_each_subflow(msk, subflow) {
2236 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2238 if (!__mptcp_subflow_active(subflow))
2241 /* still data outstanding at TCP level? skip this */
2242 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2243 mptcp_pm_subflow_chk_stale(msk, ssk);
2244 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2248 if (subflow->backup) {
2261 /* use backup only if there are no progresses anywhere */
2262 return min_stale_count > 1 ? backup : NULL;
2265 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2268 iput(SOCK_INODE(msk->subflow));
2269 msk->subflow = NULL;
2273 bool __mptcp_retransmit_pending_data(struct sock *sk)
2275 struct mptcp_data_frag *cur, *rtx_head;
2276 struct mptcp_sock *msk = mptcp_sk(sk);
2278 if (__mptcp_check_fallback(mptcp_sk(sk)))
2281 if (tcp_rtx_and_write_queues_empty(sk))
2284 /* the closing socket has some data untransmitted and/or unacked:
2285 * some data in the mptcp rtx queue has not really xmitted yet.
2286 * keep it simple and re-inject the whole mptcp level rtx queue
2288 mptcp_data_lock(sk);
2289 __mptcp_clean_una_wakeup(sk);
2290 rtx_head = mptcp_rtx_head(sk);
2292 mptcp_data_unlock(sk);
2296 msk->recovery_snd_nxt = msk->snd_nxt;
2297 msk->recovery = true;
2298 mptcp_data_unlock(sk);
2300 msk->first_pending = rtx_head;
2303 /* be sure to clear the "sent status" on all re-injected fragments */
2304 list_for_each_entry(cur, &msk->rtx_queue, list) {
2305 if (!cur->already_sent)
2307 cur->already_sent = 0;
2313 /* flags for __mptcp_close_ssk() */
2314 #define MPTCP_CF_PUSH BIT(1)
2315 #define MPTCP_CF_FASTCLOSE BIT(2)
2317 /* subflow sockets can be either outgoing (connect) or incoming
2320 * Outgoing subflows use in-kernel sockets.
2321 * Incoming subflows do not have their own 'struct socket' allocated,
2322 * so we need to use tcp_close() after detaching them from the mptcp
2325 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2326 struct mptcp_subflow_context *subflow,
2329 struct mptcp_sock *msk = mptcp_sk(sk);
2330 bool need_push, dispose_it;
2332 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2334 list_del(&subflow->node);
2336 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2338 if (flags & MPTCP_CF_FASTCLOSE) {
2339 /* be sure to force the tcp_disconnect() path,
2340 * to generate the egress reset
2342 ssk->sk_lingertime = 0;
2343 sock_set_flag(ssk, SOCK_LINGER);
2344 subflow->send_fastclose = 1;
2347 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2349 tcp_disconnect(ssk, 0);
2350 msk->subflow->state = SS_UNCONNECTED;
2351 mptcp_subflow_ctx_reset(subflow);
2358 subflow->disposable = 1;
2360 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2361 * the ssk has been already destroyed, we just need to release the
2362 * reference owned by msk;
2364 if (!inet_csk(ssk)->icsk_ulp_ops) {
2365 kfree_rcu(subflow, rcu);
2367 /* otherwise tcp will dispose of the ssk and subflow ctx */
2368 if (ssk->sk_state == TCP_LISTEN) {
2369 tcp_set_state(ssk, TCP_CLOSE);
2370 mptcp_subflow_queue_clean(ssk);
2371 inet_csk_listen_stop(ssk);
2373 __tcp_close(ssk, 0);
2375 /* close acquired an extra ref */
2382 if (ssk == msk->first)
2386 if (ssk == msk->last_snd)
2387 msk->last_snd = NULL;
2390 __mptcp_push_pending(sk, 0);
2393 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2394 struct mptcp_subflow_context *subflow)
2396 if (sk->sk_state == TCP_ESTABLISHED)
2397 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2399 /* subflow aborted before reaching the fully_established status
2400 * attempt the creation of the next subflow
2402 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2404 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2407 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2412 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2414 struct mptcp_subflow_context *subflow, *tmp;
2418 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2419 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2421 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2424 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2425 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2428 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2432 static bool mptcp_check_close_timeout(const struct sock *sk)
2434 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2435 struct mptcp_subflow_context *subflow;
2437 if (delta >= TCP_TIMEWAIT_LEN)
2440 /* if all subflows are in closed status don't bother with additional
2443 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2444 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2451 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2453 struct mptcp_subflow_context *subflow, *tmp;
2454 struct sock *sk = &msk->sk.icsk_inet.sk;
2456 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2459 mptcp_token_destroy(msk);
2461 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2462 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2465 slow = lock_sock_fast(tcp_sk);
2466 if (tcp_sk->sk_state != TCP_CLOSE) {
2467 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2468 tcp_set_state(tcp_sk, TCP_CLOSE);
2470 unlock_sock_fast(tcp_sk, slow);
2473 /* Mirror the tcp_reset() error propagation */
2474 switch (sk->sk_state) {
2476 sk->sk_err = ECONNREFUSED;
2478 case TCP_CLOSE_WAIT:
2484 sk->sk_err = ECONNRESET;
2487 inet_sk_state_store(sk, TCP_CLOSE);
2488 sk->sk_shutdown = SHUTDOWN_MASK;
2489 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2490 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2492 /* the calling mptcp_worker will properly destroy the socket */
2493 if (sock_flag(sk, SOCK_DEAD))
2496 sk->sk_state_change(sk);
2497 sk_error_report(sk);
2500 static void __mptcp_retrans(struct sock *sk)
2502 struct mptcp_sock *msk = mptcp_sk(sk);
2503 struct mptcp_sendmsg_info info = {};
2504 struct mptcp_data_frag *dfrag;
2509 mptcp_clean_una_wakeup(sk);
2511 /* first check ssk: need to kick "stale" logic */
2512 ssk = mptcp_subflow_get_retrans(msk);
2513 dfrag = mptcp_rtx_head(sk);
2515 if (mptcp_data_fin_enabled(msk)) {
2516 struct inet_connection_sock *icsk = inet_csk(sk);
2518 icsk->icsk_retransmits++;
2519 mptcp_set_datafin_timeout(sk);
2520 mptcp_send_ack(msk);
2525 if (!mptcp_send_head(sk))
2536 /* limit retransmission to the bytes already sent on some subflows */
2538 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2539 while (info.sent < info.limit) {
2540 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2544 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2549 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2550 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2552 WRITE_ONCE(msk->allow_infinite_fallback, false);
2558 mptcp_check_and_set_pending(sk);
2560 if (!mptcp_timer_pending(sk))
2561 mptcp_reset_timer(sk);
2564 /* schedule the timeout timer for the relevant event: either close timeout
2565 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2567 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2569 struct sock *sk = (struct sock *)msk;
2570 unsigned long timeout, close_timeout;
2572 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2575 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2577 /* the close timeout takes precedence on the fail one, and here at least one of
2580 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2582 sk_reset_timer(sk, &sk->sk_timer, timeout);
2585 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2587 struct sock *ssk = msk->first;
2593 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2595 slow = lock_sock_fast(ssk);
2596 mptcp_subflow_reset(ssk);
2597 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2598 unlock_sock_fast(ssk, slow);
2600 mptcp_reset_timeout(msk, 0);
2603 static void mptcp_do_fastclose(struct sock *sk)
2605 struct mptcp_subflow_context *subflow, *tmp;
2606 struct mptcp_sock *msk = mptcp_sk(sk);
2608 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2609 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2610 subflow, MPTCP_CF_FASTCLOSE);
2613 static void mptcp_worker(struct work_struct *work)
2615 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2616 struct sock *sk = &msk->sk.icsk_inet.sk;
2617 unsigned long fail_tout;
2621 state = sk->sk_state;
2622 if (unlikely(state == TCP_CLOSE))
2625 mptcp_check_data_fin_ack(sk);
2627 mptcp_check_fastclose(msk);
2629 mptcp_pm_nl_work(msk);
2631 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2632 mptcp_check_for_eof(msk);
2634 __mptcp_check_send_data_fin(sk);
2635 mptcp_check_data_fin(sk);
2637 /* There is no point in keeping around an orphaned sk timedout or
2638 * closed, but we need the msk around to reply to incoming DATA_FIN,
2639 * even if it is orphaned and in FIN_WAIT2 state
2641 if (sock_flag(sk, SOCK_DEAD)) {
2642 if (mptcp_check_close_timeout(sk)) {
2643 inet_sk_state_store(sk, TCP_CLOSE);
2644 mptcp_do_fastclose(sk);
2646 if (sk->sk_state == TCP_CLOSE) {
2647 __mptcp_destroy_sock(sk);
2652 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2653 __mptcp_close_subflow(msk);
2655 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2656 __mptcp_retrans(sk);
2658 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2659 if (fail_tout && time_after(jiffies, fail_tout))
2660 mptcp_mp_fail_no_response(msk);
2667 static int __mptcp_init_sock(struct sock *sk)
2669 struct mptcp_sock *msk = mptcp_sk(sk);
2671 INIT_LIST_HEAD(&msk->conn_list);
2672 INIT_LIST_HEAD(&msk->join_list);
2673 INIT_LIST_HEAD(&msk->rtx_queue);
2674 INIT_WORK(&msk->work, mptcp_worker);
2675 __skb_queue_head_init(&msk->receive_queue);
2676 msk->out_of_order_queue = RB_ROOT;
2677 msk->first_pending = NULL;
2678 msk->rmem_fwd_alloc = 0;
2679 WRITE_ONCE(msk->rmem_released, 0);
2680 msk->timer_ival = TCP_RTO_MIN;
2683 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2684 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2685 WRITE_ONCE(msk->allow_infinite_fallback, true);
2686 msk->recovery = false;
2688 mptcp_pm_data_init(msk);
2690 /* re-use the csk retrans timer for MPTCP-level retrans */
2691 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2692 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2697 static void mptcp_ca_reset(struct sock *sk)
2699 struct inet_connection_sock *icsk = inet_csk(sk);
2701 tcp_assign_congestion_control(sk);
2702 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2704 /* no need to keep a reference to the ops, the name will suffice */
2705 tcp_cleanup_congestion_control(sk);
2706 icsk->icsk_ca_ops = NULL;
2709 static int mptcp_init_sock(struct sock *sk)
2711 struct net *net = sock_net(sk);
2714 ret = __mptcp_init_sock(sk);
2718 if (!mptcp_is_enabled(net))
2719 return -ENOPROTOOPT;
2721 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2724 ret = __mptcp_socket_create(mptcp_sk(sk));
2728 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2729 * propagate the correct value
2733 sk_sockets_allocated_inc(sk);
2734 sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2735 sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2740 static void __mptcp_clear_xmit(struct sock *sk)
2742 struct mptcp_sock *msk = mptcp_sk(sk);
2743 struct mptcp_data_frag *dtmp, *dfrag;
2745 WRITE_ONCE(msk->first_pending, NULL);
2746 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2747 dfrag_clear(sk, dfrag);
2750 void mptcp_cancel_work(struct sock *sk)
2752 struct mptcp_sock *msk = mptcp_sk(sk);
2754 if (cancel_work_sync(&msk->work))
2758 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2762 switch (ssk->sk_state) {
2764 if (!(how & RCV_SHUTDOWN))
2768 tcp_disconnect(ssk, O_NONBLOCK);
2771 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2772 pr_debug("Fallback");
2773 ssk->sk_shutdown |= how;
2774 tcp_shutdown(ssk, how);
2776 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2778 if (!mptcp_timer_pending(sk))
2779 mptcp_reset_timer(sk);
2787 static const unsigned char new_state[16] = {
2788 /* current state: new state: action: */
2789 [0 /* (Invalid) */] = TCP_CLOSE,
2790 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2791 [TCP_SYN_SENT] = TCP_CLOSE,
2792 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2793 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2794 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2795 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2796 [TCP_CLOSE] = TCP_CLOSE,
2797 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2798 [TCP_LAST_ACK] = TCP_LAST_ACK,
2799 [TCP_LISTEN] = TCP_CLOSE,
2800 [TCP_CLOSING] = TCP_CLOSING,
2801 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2804 static int mptcp_close_state(struct sock *sk)
2806 int next = (int)new_state[sk->sk_state];
2807 int ns = next & TCP_STATE_MASK;
2809 inet_sk_state_store(sk, ns);
2811 return next & TCP_ACTION_FIN;
2814 static void __mptcp_check_send_data_fin(struct sock *sk)
2816 struct mptcp_subflow_context *subflow;
2817 struct mptcp_sock *msk = mptcp_sk(sk);
2819 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2820 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2821 msk->snd_nxt, msk->write_seq);
2823 /* we still need to enqueue subflows or not really shutting down,
2826 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2827 mptcp_send_head(sk))
2830 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2832 /* fallback socket will not get data_fin/ack, can move to the next
2835 if (__mptcp_check_fallback(msk)) {
2836 WRITE_ONCE(msk->snd_una, msk->write_seq);
2837 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2838 inet_sk_state_store(sk, TCP_CLOSE);
2839 mptcp_close_wake_up(sk);
2840 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2841 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2845 mptcp_for_each_subflow(msk, subflow) {
2846 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2848 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2852 static void __mptcp_wr_shutdown(struct sock *sk)
2854 struct mptcp_sock *msk = mptcp_sk(sk);
2856 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2857 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2858 !!mptcp_send_head(sk));
2860 /* will be ignored by fallback sockets */
2861 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2862 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2864 __mptcp_check_send_data_fin(sk);
2867 static void __mptcp_destroy_sock(struct sock *sk)
2869 struct mptcp_sock *msk = mptcp_sk(sk);
2871 pr_debug("msk=%p", msk);
2875 mptcp_stop_timer(sk);
2876 sk_stop_timer(sk, &sk->sk_timer);
2879 sk->sk_prot->destroy(sk);
2881 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2882 WARN_ON_ONCE(msk->rmem_released);
2883 sk_stream_kill_queues(sk);
2884 xfrm_sk_free_policy(sk);
2886 sk_refcnt_debug_release(sk);
2890 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2892 /* Concurrent splices from sk_receive_queue into receive_queue will
2893 * always show at least one non-empty queue when checked in this order.
2895 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2896 skb_queue_empty_lockless(&msk->receive_queue))
2899 return EPOLLIN | EPOLLRDNORM;
2902 bool __mptcp_close(struct sock *sk, long timeout)
2904 struct mptcp_subflow_context *subflow;
2905 struct mptcp_sock *msk = mptcp_sk(sk);
2906 bool do_cancel_work = false;
2908 sk->sk_shutdown = SHUTDOWN_MASK;
2910 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2911 inet_sk_state_store(sk, TCP_CLOSE);
2915 if (mptcp_check_readable(msk)) {
2916 /* the msk has read data, do the MPTCP equivalent of TCP reset */
2917 inet_sk_state_store(sk, TCP_CLOSE);
2918 mptcp_do_fastclose(sk);
2919 } else if (mptcp_close_state(sk)) {
2920 __mptcp_wr_shutdown(sk);
2923 sk_stream_wait_close(sk, timeout);
2926 /* orphan all the subflows */
2927 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2928 mptcp_for_each_subflow(msk, subflow) {
2929 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2930 bool slow = lock_sock_fast_nested(ssk);
2932 /* since the close timeout takes precedence on the fail one,
2935 if (ssk == msk->first)
2936 subflow->fail_tout = 0;
2938 /* detach from the parent socket, but allow data_ready to
2939 * push incoming data into the mptcp stack, to properly ack it
2941 ssk->sk_socket = NULL;
2943 unlock_sock_fast(ssk, slow);
2948 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2949 if (mptcp_sk(sk)->token)
2950 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2952 if (sk->sk_state == TCP_CLOSE) {
2953 __mptcp_destroy_sock(sk);
2954 do_cancel_work = true;
2956 mptcp_reset_timeout(msk, 0);
2959 return do_cancel_work;
2962 static void mptcp_close(struct sock *sk, long timeout)
2964 bool do_cancel_work;
2968 do_cancel_work = __mptcp_close(sk, timeout);
2971 mptcp_cancel_work(sk);
2976 void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2978 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2979 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2980 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2982 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2983 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2986 msk6->saddr = ssk6->saddr;
2987 msk6->flow_label = ssk6->flow_label;
2991 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2992 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2993 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2994 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2995 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2996 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2999 static int mptcp_disconnect(struct sock *sk, int flags)
3001 struct mptcp_sock *msk = mptcp_sk(sk);
3003 inet_sk_state_store(sk, TCP_CLOSE);
3005 mptcp_stop_timer(sk);
3006 sk_stop_timer(sk, &sk->sk_timer);
3008 if (mptcp_sk(sk)->token)
3009 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
3011 /* msk->subflow is still intact, the following will not free the first
3014 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3015 msk->last_snd = NULL;
3016 WRITE_ONCE(msk->flags, 0);
3018 msk->push_pending = 0;
3019 msk->recovery = false;
3020 msk->can_ack = false;
3021 msk->fully_established = false;
3022 msk->rcv_data_fin = false;
3023 msk->snd_data_fin_enable = false;
3024 msk->rcv_fastclose = false;
3025 msk->use_64bit_ack = false;
3026 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3027 mptcp_pm_data_reset(msk);
3030 sk->sk_shutdown = 0;
3031 sk_error_report(sk);
3035 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3036 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3038 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3040 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3044 struct sock *mptcp_sk_clone(const struct sock *sk,
3045 const struct mptcp_options_received *mp_opt,
3046 struct request_sock *req)
3048 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3049 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3050 struct mptcp_sock *msk;
3056 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3057 if (nsk->sk_family == AF_INET6)
3058 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3061 __mptcp_init_sock(nsk);
3063 msk = mptcp_sk(nsk);
3064 msk->local_key = subflow_req->local_key;
3065 msk->token = subflow_req->token;
3066 msk->subflow = NULL;
3067 WRITE_ONCE(msk->fully_established, false);
3068 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3069 WRITE_ONCE(msk->csum_enabled, true);
3071 msk->write_seq = subflow_req->idsn + 1;
3072 msk->snd_nxt = msk->write_seq;
3073 msk->snd_una = msk->write_seq;
3074 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3075 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3077 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
3078 msk->can_ack = true;
3079 msk->remote_key = mp_opt->sndr_key;
3080 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
3082 WRITE_ONCE(msk->ack_seq, ack_seq);
3083 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3086 sock_reset_flag(nsk, SOCK_RCU_FREE);
3087 /* will be fully established after successful MPC subflow creation */
3088 inet_sk_state_store(nsk, TCP_SYN_RECV);
3090 security_inet_csk_clone(nsk, req);
3091 bh_unlock_sock(nsk);
3093 /* keep a single reference */
3098 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3100 const struct tcp_sock *tp = tcp_sk(ssk);
3102 msk->rcvq_space.copied = 0;
3103 msk->rcvq_space.rtt_us = 0;
3105 msk->rcvq_space.time = tp->tcp_mstamp;
3107 /* initial rcv_space offering made to peer */
3108 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3109 TCP_INIT_CWND * tp->advmss);
3110 if (msk->rcvq_space.space == 0)
3111 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3113 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3116 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3119 struct mptcp_sock *msk = mptcp_sk(sk);
3120 struct socket *listener;
3123 listener = __mptcp_nmpc_socket(msk);
3124 if (WARN_ON_ONCE(!listener)) {
3129 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3130 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3134 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3135 if (sk_is_mptcp(newsk)) {
3136 struct mptcp_subflow_context *subflow;
3137 struct sock *new_mptcp_sock;
3139 subflow = mptcp_subflow_ctx(newsk);
3140 new_mptcp_sock = subflow->conn;
3142 /* is_mptcp should be false if subflow->conn is missing, see
3143 * subflow_syn_recv_sock()
3145 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3146 tcp_sk(newsk)->is_mptcp = 0;
3150 /* acquire the 2nd reference for the owning socket */
3151 sock_hold(new_mptcp_sock);
3152 newsk = new_mptcp_sock;
3153 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3155 MPTCP_INC_STATS(sock_net(sk),
3156 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3160 newsk->sk_kern_sock = kern;
3164 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3166 struct mptcp_subflow_context *subflow, *tmp;
3167 struct sock *sk = (struct sock *)msk;
3169 __mptcp_clear_xmit(sk);
3171 /* join list will be eventually flushed (with rst) at sock lock release time */
3172 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3173 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3175 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3176 mptcp_data_lock(sk);
3177 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3178 __skb_queue_purge(&sk->sk_receive_queue);
3179 skb_rbtree_purge(&msk->out_of_order_queue);
3180 mptcp_data_unlock(sk);
3182 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3183 * inet_sock_destruct() will dispose it
3185 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3186 msk->rmem_fwd_alloc = 0;
3187 mptcp_token_destroy(msk);
3188 mptcp_pm_free_anno_list(msk);
3189 mptcp_free_local_addr_list(msk);
3192 static void mptcp_destroy(struct sock *sk)
3194 struct mptcp_sock *msk = mptcp_sk(sk);
3196 /* clears msk->subflow, allowing the following to close
3197 * even the initial subflow
3199 mptcp_dispose_initial_subflow(msk);
3200 mptcp_destroy_common(msk, 0);
3201 sk_sockets_allocated_dec(sk);
3204 void __mptcp_data_acked(struct sock *sk)
3206 if (!sock_owned_by_user(sk))
3207 __mptcp_clean_una(sk);
3209 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3211 if (mptcp_pending_data_fin_ack(sk))
3212 mptcp_schedule_work(sk);
3215 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3217 if (!mptcp_send_head(sk))
3220 if (!sock_owned_by_user(sk)) {
3221 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3223 if (xmit_ssk == ssk)
3224 __mptcp_subflow_push_pending(sk, ssk);
3226 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3228 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3232 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3233 BIT(MPTCP_RETRANSMIT) | \
3234 BIT(MPTCP_FLUSH_JOIN_LIST))
3236 /* processes deferred events and flush wmem */
3237 static void mptcp_release_cb(struct sock *sk)
3238 __must_hold(&sk->sk_lock.slock)
3240 struct mptcp_sock *msk = mptcp_sk(sk);
3243 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3248 /* the following actions acquire the subflow socket lock
3250 * 1) can't be invoked in atomic scope
3251 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3252 * datapath acquires the msk socket spinlock while helding
3253 * the subflow socket lock
3255 msk->push_pending = 0;
3256 msk->cb_flags &= ~flags;
3257 spin_unlock_bh(&sk->sk_lock.slock);
3258 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3259 __mptcp_flush_join_list(sk);
3260 if (flags & BIT(MPTCP_PUSH_PENDING))
3261 __mptcp_push_pending(sk, 0);
3262 if (flags & BIT(MPTCP_RETRANSMIT))
3263 __mptcp_retrans(sk);
3266 spin_lock_bh(&sk->sk_lock.slock);
3269 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3270 __mptcp_clean_una_wakeup(sk);
3271 if (unlikely(&msk->cb_flags)) {
3272 /* be sure to set the current sk state before tacking actions
3273 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3275 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3276 __mptcp_set_connected(sk);
3277 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3278 __mptcp_error_report(sk);
3279 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3280 msk->last_snd = NULL;
3283 __mptcp_update_rmem(sk);
3286 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3287 * TCP can't schedule delack timer before the subflow is fully established.
3288 * MPTCP uses the delack timer to do 3rd ack retransmissions
3290 static void schedule_3rdack_retransmission(struct sock *ssk)
3292 struct inet_connection_sock *icsk = inet_csk(ssk);
3293 struct tcp_sock *tp = tcp_sk(ssk);
3294 unsigned long timeout;
3296 if (mptcp_subflow_ctx(ssk)->fully_established)
3299 /* reschedule with a timeout above RTT, as we must look only for drop */
3301 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3303 timeout = TCP_TIMEOUT_INIT;
3306 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3307 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3308 icsk->icsk_ack.timeout = timeout;
3309 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3312 void mptcp_subflow_process_delegated(struct sock *ssk)
3314 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3315 struct sock *sk = subflow->conn;
3317 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3318 mptcp_data_lock(sk);
3319 if (!sock_owned_by_user(sk))
3320 __mptcp_subflow_push_pending(sk, ssk);
3322 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3323 mptcp_data_unlock(sk);
3324 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3326 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3327 schedule_3rdack_retransmission(ssk);
3328 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3332 static int mptcp_hash(struct sock *sk)
3334 /* should never be called,
3335 * we hash the TCP subflows not the master socket
3341 static void mptcp_unhash(struct sock *sk)
3343 /* called from sk_common_release(), but nothing to do here */
3346 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3348 struct mptcp_sock *msk = mptcp_sk(sk);
3349 struct socket *ssock;
3351 ssock = __mptcp_nmpc_socket(msk);
3352 pr_debug("msk=%p, subflow=%p", msk, ssock);
3353 if (WARN_ON_ONCE(!ssock))
3356 return inet_csk_get_port(ssock->sk, snum);
3359 void mptcp_finish_connect(struct sock *ssk)
3361 struct mptcp_subflow_context *subflow;
3362 struct mptcp_sock *msk;
3366 subflow = mptcp_subflow_ctx(ssk);
3370 pr_debug("msk=%p, token=%u", sk, subflow->token);
3372 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3374 subflow->map_seq = ack_seq;
3375 subflow->map_subflow_seq = 1;
3377 /* the socket is not connected yet, no msk/subflow ops can access/race
3378 * accessing the field below
3380 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3381 WRITE_ONCE(msk->local_key, subflow->local_key);
3382 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3383 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3384 WRITE_ONCE(msk->ack_seq, ack_seq);
3385 WRITE_ONCE(msk->can_ack, 1);
3386 WRITE_ONCE(msk->snd_una, msk->write_seq);
3387 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3389 mptcp_pm_new_connection(msk, ssk, 0);
3391 mptcp_rcv_space_init(msk, ssk);
3394 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3396 write_lock_bh(&sk->sk_callback_lock);
3397 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3398 sk_set_socket(sk, parent);
3399 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3400 write_unlock_bh(&sk->sk_callback_lock);
3403 bool mptcp_finish_join(struct sock *ssk)
3405 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3406 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3407 struct sock *parent = (void *)msk;
3410 pr_debug("msk=%p, subflow=%p", msk, subflow);
3412 /* mptcp socket already closing? */
3413 if (!mptcp_is_fully_established(parent)) {
3414 subflow->reset_reason = MPTCP_RST_EMPTCP;
3418 if (!list_empty(&subflow->node))
3421 if (!mptcp_pm_allow_new_subflow(msk))
3422 goto err_prohibited;
3424 /* active connections are already on conn_list.
3425 * If we can't acquire msk socket lock here, let the release callback
3428 mptcp_data_lock(parent);
3429 if (!sock_owned_by_user(parent)) {
3430 ret = __mptcp_finish_join(msk, ssk);
3433 list_add_tail(&subflow->node, &msk->conn_list);
3437 list_add_tail(&subflow->node, &msk->join_list);
3438 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3440 mptcp_data_unlock(parent);
3444 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3448 subflow->map_seq = READ_ONCE(msk->ack_seq);
3449 WRITE_ONCE(msk->allow_infinite_fallback, false);
3452 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3456 static void mptcp_shutdown(struct sock *sk, int how)
3458 pr_debug("sk=%p, how=%d", sk, how);
3460 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3461 __mptcp_wr_shutdown(sk);
3464 static int mptcp_forward_alloc_get(const struct sock *sk)
3466 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3469 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3471 const struct sock *sk = (void *)msk;
3474 if (sk->sk_state == TCP_LISTEN)
3477 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3480 delta = msk->write_seq - v;
3481 if (__mptcp_check_fallback(msk) && msk->first) {
3482 struct tcp_sock *tp = tcp_sk(msk->first);
3484 /* the first subflow is disconnected after close - see
3485 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3486 * so ignore that status, too.
3488 if (!((1 << msk->first->sk_state) &
3489 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3490 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3492 if (delta > INT_MAX)
3498 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3500 struct mptcp_sock *msk = mptcp_sk(sk);
3506 if (sk->sk_state == TCP_LISTEN)
3510 __mptcp_move_skbs(msk);
3511 answ = mptcp_inq_hint(sk);
3515 slow = lock_sock_fast(sk);
3516 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3517 unlock_sock_fast(sk, slow);
3520 slow = lock_sock_fast(sk);
3521 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3522 unlock_sock_fast(sk, slow);
3525 return -ENOIOCTLCMD;
3528 return put_user(answ, (int __user *)arg);
3531 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3532 struct mptcp_subflow_context *subflow)
3534 subflow->request_mptcp = 0;
3535 __mptcp_do_fallback(msk);
3538 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3540 struct mptcp_subflow_context *subflow;
3541 struct mptcp_sock *msk = mptcp_sk(sk);
3542 struct socket *ssock;
3545 ssock = __mptcp_nmpc_socket(msk);
3549 mptcp_token_destroy(msk);
3550 inet_sk_state_store(sk, TCP_SYN_SENT);
3551 subflow = mptcp_subflow_ctx(ssock->sk);
3552 #ifdef CONFIG_TCP_MD5SIG
3553 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3556 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3557 mptcp_subflow_early_fallback(msk, subflow);
3559 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3560 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3561 mptcp_subflow_early_fallback(msk, subflow);
3563 if (likely(!__mptcp_check_fallback(msk)))
3564 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3566 /* if reaching here via the fastopen/sendmsg path, the caller already
3567 * acquired the subflow socket lock, too.
3569 if (msk->is_sendmsg)
3570 err = __inet_stream_connect(ssock, uaddr, addr_len, msk->connect_flags, 1);
3572 err = inet_stream_connect(ssock, uaddr, addr_len, msk->connect_flags);
3573 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3575 /* on successful connect, the msk state will be moved to established by
3576 * subflow_finish_connect()
3578 if (unlikely(err && err != -EINPROGRESS)) {
3579 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3583 mptcp_copy_inaddrs(sk, ssock->sk);
3585 /* unblocking connect, mptcp-level inet_stream_connect will error out
3586 * without changing the socket state, update it here.
3588 if (err == -EINPROGRESS)
3589 sk->sk_socket->state = ssock->state;
3593 static struct proto mptcp_prot = {
3595 .owner = THIS_MODULE,
3596 .init = mptcp_init_sock,
3597 .connect = mptcp_connect,
3598 .disconnect = mptcp_disconnect,
3599 .close = mptcp_close,
3600 .accept = mptcp_accept,
3601 .setsockopt = mptcp_setsockopt,
3602 .getsockopt = mptcp_getsockopt,
3603 .shutdown = mptcp_shutdown,
3604 .destroy = mptcp_destroy,
3605 .sendmsg = mptcp_sendmsg,
3606 .ioctl = mptcp_ioctl,
3607 .recvmsg = mptcp_recvmsg,
3608 .release_cb = mptcp_release_cb,
3610 .unhash = mptcp_unhash,
3611 .get_port = mptcp_get_port,
3612 .forward_alloc_get = mptcp_forward_alloc_get,
3613 .sockets_allocated = &mptcp_sockets_allocated,
3615 .memory_allocated = &tcp_memory_allocated,
3616 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3618 .memory_pressure = &tcp_memory_pressure,
3619 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3620 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3621 .sysctl_mem = sysctl_tcp_mem,
3622 .obj_size = sizeof(struct mptcp_sock),
3623 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3624 .no_autobind = true,
3627 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3629 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3630 struct socket *ssock;
3633 lock_sock(sock->sk);
3634 ssock = __mptcp_nmpc_socket(msk);
3640 err = ssock->ops->bind(ssock, uaddr, addr_len);
3642 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3645 release_sock(sock->sk);
3649 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3650 int addr_len, int flags)
3654 lock_sock(sock->sk);
3655 mptcp_sk(sock->sk)->connect_flags = flags;
3656 ret = __inet_stream_connect(sock, uaddr, addr_len, flags, 0);
3657 release_sock(sock->sk);
3661 static int mptcp_listen(struct socket *sock, int backlog)
3663 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3664 struct socket *ssock;
3667 pr_debug("msk=%p", msk);
3669 lock_sock(sock->sk);
3670 ssock = __mptcp_nmpc_socket(msk);
3676 mptcp_token_destroy(msk);
3677 inet_sk_state_store(sock->sk, TCP_LISTEN);
3678 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3680 err = ssock->ops->listen(ssock, backlog);
3681 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3683 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3686 release_sock(sock->sk);
3690 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3691 int flags, bool kern)
3693 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3694 struct socket *ssock;
3697 pr_debug("msk=%p", msk);
3699 ssock = __mptcp_nmpc_socket(msk);
3703 err = ssock->ops->accept(sock, newsock, flags, kern);
3704 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3705 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3706 struct mptcp_subflow_context *subflow;
3707 struct sock *newsk = newsock->sk;
3711 /* PM/worker can now acquire the first subflow socket
3712 * lock without racing with listener queue cleanup,
3713 * we can notify it, if needed.
3715 * Even if remote has reset the initial subflow by now
3716 * the refcnt is still at least one.
3718 subflow = mptcp_subflow_ctx(msk->first);
3719 list_add(&subflow->node, &msk->conn_list);
3720 sock_hold(msk->first);
3721 if (mptcp_is_fully_established(newsk))
3722 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3724 mptcp_rcv_space_init(msk, msk->first);
3725 mptcp_propagate_sndbuf(newsk, msk->first);
3727 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3728 * This is needed so NOSPACE flag can be set from tcp stack.
3730 mptcp_for_each_subflow(msk, subflow) {
3731 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3733 if (!ssk->sk_socket)
3734 mptcp_sock_graft(ssk, newsock);
3736 release_sock(newsk);
3742 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3744 struct sock *sk = (struct sock *)msk;
3746 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3747 return EPOLLOUT | EPOLLWRNORM;
3749 if (sk_stream_is_writeable(sk))
3750 return EPOLLOUT | EPOLLWRNORM;
3752 mptcp_set_nospace(sk);
3753 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3754 if (sk_stream_is_writeable(sk))
3755 return EPOLLOUT | EPOLLWRNORM;
3760 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3761 struct poll_table_struct *wait)
3763 struct sock *sk = sock->sk;
3764 struct mptcp_sock *msk;
3769 sock_poll_wait(file, sock, wait);
3771 state = inet_sk_state_load(sk);
3772 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3773 if (state == TCP_LISTEN) {
3774 if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk))
3777 return inet_csk_listen_poll(msk->subflow->sk);
3780 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3781 mask |= mptcp_check_readable(msk);
3782 mask |= mptcp_check_writeable(msk);
3783 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3784 /* cf tcp_poll() note about TFO */
3785 mask |= EPOLLOUT | EPOLLWRNORM;
3787 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3789 if (sk->sk_shutdown & RCV_SHUTDOWN)
3790 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3792 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3800 static const struct proto_ops mptcp_stream_ops = {
3802 .owner = THIS_MODULE,
3803 .release = inet_release,
3805 .connect = mptcp_stream_connect,
3806 .socketpair = sock_no_socketpair,
3807 .accept = mptcp_stream_accept,
3808 .getname = inet_getname,
3810 .ioctl = inet_ioctl,
3811 .gettstamp = sock_gettstamp,
3812 .listen = mptcp_listen,
3813 .shutdown = inet_shutdown,
3814 .setsockopt = sock_common_setsockopt,
3815 .getsockopt = sock_common_getsockopt,
3816 .sendmsg = inet_sendmsg,
3817 .recvmsg = inet_recvmsg,
3818 .mmap = sock_no_mmap,
3819 .sendpage = inet_sendpage,
3822 static struct inet_protosw mptcp_protosw = {
3823 .type = SOCK_STREAM,
3824 .protocol = IPPROTO_MPTCP,
3825 .prot = &mptcp_prot,
3826 .ops = &mptcp_stream_ops,
3827 .flags = INET_PROTOSW_ICSK,
3830 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3832 struct mptcp_delegated_action *delegated;
3833 struct mptcp_subflow_context *subflow;
3836 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3837 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3838 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3840 bh_lock_sock_nested(ssk);
3841 if (!sock_owned_by_user(ssk) &&
3842 mptcp_subflow_has_delegated_action(subflow))
3843 mptcp_subflow_process_delegated(ssk);
3844 /* ... elsewhere tcp_release_cb_override already processed
3845 * the action or will do at next release_sock().
3846 * In both case must dequeue the subflow here - on the same
3847 * CPU that scheduled it.
3849 bh_unlock_sock(ssk);
3852 if (++work_done == budget)
3856 /* always provide a 0 'work_done' argument, so that napi_complete_done
3857 * will not try accessing the NULL napi->dev ptr
3859 napi_complete_done(napi, 0);
3863 void __init mptcp_proto_init(void)
3865 struct mptcp_delegated_action *delegated;
3868 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3870 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3871 panic("Failed to allocate MPTCP pcpu counter\n");
3873 init_dummy_netdev(&mptcp_napi_dev);
3874 for_each_possible_cpu(cpu) {
3875 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3876 INIT_LIST_HEAD(&delegated->head);
3877 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3879 napi_enable(&delegated->napi);
3882 mptcp_subflow_init();
3886 if (proto_register(&mptcp_prot, 1) != 0)
3887 panic("Failed to register MPTCP proto.\n");
3889 inet_register_protosw(&mptcp_protosw);
3891 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3894 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3895 static const struct proto_ops mptcp_v6_stream_ops = {
3897 .owner = THIS_MODULE,
3898 .release = inet6_release,
3900 .connect = mptcp_stream_connect,
3901 .socketpair = sock_no_socketpair,
3902 .accept = mptcp_stream_accept,
3903 .getname = inet6_getname,
3905 .ioctl = inet6_ioctl,
3906 .gettstamp = sock_gettstamp,
3907 .listen = mptcp_listen,
3908 .shutdown = inet_shutdown,
3909 .setsockopt = sock_common_setsockopt,
3910 .getsockopt = sock_common_getsockopt,
3911 .sendmsg = inet6_sendmsg,
3912 .recvmsg = inet6_recvmsg,
3913 .mmap = sock_no_mmap,
3914 .sendpage = inet_sendpage,
3915 #ifdef CONFIG_COMPAT
3916 .compat_ioctl = inet6_compat_ioctl,
3920 static struct proto mptcp_v6_prot;
3922 static void mptcp_v6_destroy(struct sock *sk)
3925 inet6_destroy_sock(sk);
3928 static struct inet_protosw mptcp_v6_protosw = {
3929 .type = SOCK_STREAM,
3930 .protocol = IPPROTO_MPTCP,
3931 .prot = &mptcp_v6_prot,
3932 .ops = &mptcp_v6_stream_ops,
3933 .flags = INET_PROTOSW_ICSK,
3936 int __init mptcp_proto_v6_init(void)
3940 mptcp_v6_prot = mptcp_prot;
3941 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3942 mptcp_v6_prot.slab = NULL;
3943 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3944 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3946 err = proto_register(&mptcp_v6_prot, 1);
3950 err = inet6_register_protosw(&mptcp_v6_protosw);
3952 proto_unregister(&mptcp_v6_prot);