1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
40 MPTCP_CMSG_TS = BIT(0),
41 MPTCP_CMSG_INQ = BIT(1),
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void __mptcp_check_send_data_fin(struct sock *sk);
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
52 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
53 * completed yet or has failed, return the subflow socket.
54 * Otherwise return NULL.
56 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
58 if (!msk->subflow || READ_ONCE(msk->can_ack))
64 /* Returns end sequence number of the receiver's advertised window */
65 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
67 return READ_ONCE(msk->wnd_end);
70 static bool mptcp_is_tcpsk(struct sock *sk)
72 struct socket *sock = sk->sk_socket;
74 if (unlikely(sk->sk_prot == &tcp_prot)) {
75 /* we are being invoked after mptcp_accept() has
76 * accepted a non-mp-capable flow: sk is a tcp_sk,
79 * Hand the socket over to tcp so all further socket ops
82 sock->ops = &inet_stream_ops;
84 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
85 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
86 sock->ops = &inet6_stream_ops;
94 static int __mptcp_socket_create(struct mptcp_sock *msk)
96 struct mptcp_subflow_context *subflow;
97 struct sock *sk = (struct sock *)msk;
101 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
105 msk->first = ssock->sk;
106 msk->subflow = ssock;
107 subflow = mptcp_subflow_ctx(ssock->sk);
108 list_add(&subflow->node, &msk->conn_list);
109 sock_hold(ssock->sk);
110 subflow->request_mptcp = 1;
112 /* This is the first subflow, always with id 0 */
113 subflow->local_id_valid = 1;
114 mptcp_sock_graft(msk->first, sk->sk_socket);
119 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
121 sk_drops_add(sk, skb);
125 static void mptcp_rmem_charge(struct sock *sk, int size)
127 mptcp_sk(sk)->rmem_fwd_alloc -= size;
130 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
131 struct sk_buff *from)
136 if (MPTCP_SKB_CB(from)->offset ||
137 !skb_try_coalesce(to, from, &fragstolen, &delta))
140 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
141 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
142 to->len, MPTCP_SKB_CB(from)->end_seq);
143 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
145 /* note the fwd memory can reach a negative value after accounting
146 * for the delta, but the later skb free will restore a non
149 atomic_add(delta, &sk->sk_rmem_alloc);
150 mptcp_rmem_charge(sk, delta);
151 kfree_skb_partial(from, fragstolen);
156 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
157 struct sk_buff *from)
159 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
162 return mptcp_try_coalesce((struct sock *)msk, to, from);
165 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
167 amount >>= PAGE_SHIFT;
168 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
169 __sk_mem_reduce_allocated(sk, amount);
172 static void mptcp_rmem_uncharge(struct sock *sk, int size)
174 struct mptcp_sock *msk = mptcp_sk(sk);
177 msk->rmem_fwd_alloc += size;
178 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
180 /* see sk_mem_uncharge() for the rationale behind the following schema */
181 if (unlikely(reclaimable >= PAGE_SIZE))
182 __mptcp_rmem_reclaim(sk, reclaimable);
185 static void mptcp_rfree(struct sk_buff *skb)
187 unsigned int len = skb->truesize;
188 struct sock *sk = skb->sk;
190 atomic_sub(len, &sk->sk_rmem_alloc);
191 mptcp_rmem_uncharge(sk, len);
194 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
198 skb->destructor = mptcp_rfree;
199 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
200 mptcp_rmem_charge(sk, skb->truesize);
203 /* "inspired" by tcp_data_queue_ofo(), main differences:
205 * - don't cope with sacks
207 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
209 struct sock *sk = (struct sock *)msk;
210 struct rb_node **p, *parent;
211 u64 seq, end_seq, max_seq;
212 struct sk_buff *skb1;
214 seq = MPTCP_SKB_CB(skb)->map_seq;
215 end_seq = MPTCP_SKB_CB(skb)->end_seq;
216 max_seq = atomic64_read(&msk->rcv_wnd_sent);
218 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
219 RB_EMPTY_ROOT(&msk->out_of_order_queue));
220 if (after64(end_seq, max_seq)) {
223 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
224 (unsigned long long)end_seq - (unsigned long)max_seq,
225 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
226 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
230 p = &msk->out_of_order_queue.rb_node;
231 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
232 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
233 rb_link_node(&skb->rbnode, NULL, p);
234 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
235 msk->ooo_last_skb = skb;
239 /* with 2 subflows, adding at end of ooo queue is quite likely
240 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
242 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
248 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
249 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
250 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
251 parent = &msk->ooo_last_skb->rbnode;
252 p = &parent->rb_right;
256 /* Find place to insert this segment. Handle overlaps on the way. */
260 skb1 = rb_to_skb(parent);
261 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
262 p = &parent->rb_left;
265 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
266 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
267 /* All the bits are present. Drop. */
269 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
272 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
276 * continue traversing
279 /* skb's seq == skb1's seq and skb covers skb1.
280 * Replace skb1 with skb.
282 rb_replace_node(&skb1->rbnode, &skb->rbnode,
283 &msk->out_of_order_queue);
284 mptcp_drop(sk, skb1);
285 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
288 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
289 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
292 p = &parent->rb_right;
296 /* Insert segment into RB tree. */
297 rb_link_node(&skb->rbnode, parent, p);
298 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
301 /* Remove other segments covered by skb. */
302 while ((skb1 = skb_rb_next(skb)) != NULL) {
303 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
305 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
306 mptcp_drop(sk, skb1);
307 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
309 /* If there is no skb after us, we are the last_skb ! */
311 msk->ooo_last_skb = skb;
315 mptcp_set_owner_r(skb, sk);
318 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
320 struct mptcp_sock *msk = mptcp_sk(sk);
323 if (size <= msk->rmem_fwd_alloc)
326 size -= msk->rmem_fwd_alloc;
327 amt = sk_mem_pages(size);
328 amount = amt << PAGE_SHIFT;
329 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
332 msk->rmem_fwd_alloc += amount;
336 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
337 struct sk_buff *skb, unsigned int offset,
340 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
341 struct sock *sk = (struct sock *)msk;
342 struct sk_buff *tail;
345 __skb_unlink(skb, &ssk->sk_receive_queue);
350 /* try to fetch required memory from subflow */
351 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
354 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
356 /* the skb map_seq accounts for the skb offset:
357 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
360 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
361 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
362 MPTCP_SKB_CB(skb)->offset = offset;
363 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
365 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
367 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
368 tail = skb_peek_tail(&sk->sk_receive_queue);
369 if (tail && mptcp_try_coalesce(sk, tail, skb))
372 mptcp_set_owner_r(skb, sk);
373 __skb_queue_tail(&sk->sk_receive_queue, skb);
375 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
376 mptcp_data_queue_ofo(msk, skb);
380 /* old data, keep it simple and drop the whole pkt, sender
381 * will retransmit as needed, if needed.
383 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
389 static void mptcp_stop_timer(struct sock *sk)
391 struct inet_connection_sock *icsk = inet_csk(sk);
393 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
394 mptcp_sk(sk)->timer_ival = 0;
397 static void mptcp_close_wake_up(struct sock *sk)
399 if (sock_flag(sk, SOCK_DEAD))
402 sk->sk_state_change(sk);
403 if (sk->sk_shutdown == SHUTDOWN_MASK ||
404 sk->sk_state == TCP_CLOSE)
405 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
407 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
410 static bool mptcp_pending_data_fin_ack(struct sock *sk)
412 struct mptcp_sock *msk = mptcp_sk(sk);
414 return !__mptcp_check_fallback(msk) &&
415 ((1 << sk->sk_state) &
416 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
417 msk->write_seq == READ_ONCE(msk->snd_una);
420 static void mptcp_check_data_fin_ack(struct sock *sk)
422 struct mptcp_sock *msk = mptcp_sk(sk);
424 /* Look for an acknowledged DATA_FIN */
425 if (mptcp_pending_data_fin_ack(sk)) {
426 WRITE_ONCE(msk->snd_data_fin_enable, 0);
428 switch (sk->sk_state) {
430 inet_sk_state_store(sk, TCP_FIN_WAIT2);
434 inet_sk_state_store(sk, TCP_CLOSE);
438 mptcp_close_wake_up(sk);
442 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
444 struct mptcp_sock *msk = mptcp_sk(sk);
446 if (READ_ONCE(msk->rcv_data_fin) &&
447 ((1 << sk->sk_state) &
448 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
449 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
451 if (msk->ack_seq == rcv_data_fin_seq) {
453 *seq = rcv_data_fin_seq;
462 static void mptcp_set_datafin_timeout(const struct sock *sk)
464 struct inet_connection_sock *icsk = inet_csk(sk);
467 retransmits = min_t(u32, icsk->icsk_retransmits,
468 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
470 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
473 static void __mptcp_set_timeout(struct sock *sk, long tout)
475 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
478 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
480 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
482 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
483 inet_csk(ssk)->icsk_timeout - jiffies : 0;
486 static void mptcp_set_timeout(struct sock *sk)
488 struct mptcp_subflow_context *subflow;
491 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
492 tout = max(tout, mptcp_timeout_from_subflow(subflow));
493 __mptcp_set_timeout(sk, tout);
496 static inline bool tcp_can_send_ack(const struct sock *ssk)
498 return !((1 << inet_sk_state_load(ssk)) &
499 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
502 void __mptcp_subflow_send_ack(struct sock *ssk)
504 if (tcp_can_send_ack(ssk))
508 static void mptcp_subflow_send_ack(struct sock *ssk)
512 slow = lock_sock_fast(ssk);
513 __mptcp_subflow_send_ack(ssk);
514 unlock_sock_fast(ssk, slow);
517 static void mptcp_send_ack(struct mptcp_sock *msk)
519 struct mptcp_subflow_context *subflow;
521 mptcp_for_each_subflow(msk, subflow)
522 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
525 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
529 slow = lock_sock_fast(ssk);
530 if (tcp_can_send_ack(ssk))
531 tcp_cleanup_rbuf(ssk, 1);
532 unlock_sock_fast(ssk, slow);
535 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
537 const struct inet_connection_sock *icsk = inet_csk(ssk);
538 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
539 const struct tcp_sock *tp = tcp_sk(ssk);
541 return (ack_pending & ICSK_ACK_SCHED) &&
542 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
543 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
544 (rx_empty && ack_pending &
545 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
548 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
550 int old_space = READ_ONCE(msk->old_wspace);
551 struct mptcp_subflow_context *subflow;
552 struct sock *sk = (struct sock *)msk;
553 int space = __mptcp_space(sk);
554 bool cleanup, rx_empty;
556 cleanup = (space > 0) && (space >= (old_space << 1));
557 rx_empty = !__mptcp_rmem(sk);
559 mptcp_for_each_subflow(msk, subflow) {
560 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
562 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
563 mptcp_subflow_cleanup_rbuf(ssk);
567 static bool mptcp_check_data_fin(struct sock *sk)
569 struct mptcp_sock *msk = mptcp_sk(sk);
570 u64 rcv_data_fin_seq;
573 if (__mptcp_check_fallback(msk))
576 /* Need to ack a DATA_FIN received from a peer while this side
577 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
578 * msk->rcv_data_fin was set when parsing the incoming options
579 * at the subflow level and the msk lock was not held, so this
580 * is the first opportunity to act on the DATA_FIN and change
583 * If we are caught up to the sequence number of the incoming
584 * DATA_FIN, send the DATA_ACK now and do state transition. If
585 * not caught up, do nothing and let the recv code send DATA_ACK
589 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
590 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
591 WRITE_ONCE(msk->rcv_data_fin, 0);
593 sk->sk_shutdown |= RCV_SHUTDOWN;
594 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
596 switch (sk->sk_state) {
597 case TCP_ESTABLISHED:
598 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
601 inet_sk_state_store(sk, TCP_CLOSING);
604 inet_sk_state_store(sk, TCP_CLOSE);
607 /* Other states not expected */
614 mptcp_close_wake_up(sk);
619 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
623 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
624 struct sock *sk = (struct sock *)msk;
625 unsigned int moved = 0;
626 bool more_data_avail;
631 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
633 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
634 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
636 if (unlikely(ssk_rbuf > sk_rbuf)) {
637 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
642 pr_debug("msk=%p ssk=%p", msk, ssk);
645 u32 map_remaining, offset;
646 u32 seq = tp->copied_seq;
650 /* try to move as much data as available */
651 map_remaining = subflow->map_data_len -
652 mptcp_subflow_get_map_offset(subflow);
654 skb = skb_peek(&ssk->sk_receive_queue);
656 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
657 * a different CPU can have already processed the pending
658 * data, stop here or we can enter an infinite loop
665 if (__mptcp_check_fallback(msk)) {
666 /* Under fallback skbs have no MPTCP extension and TCP could
667 * collapse them between the dummy map creation and the
668 * current dequeue. Be sure to adjust the map size.
670 map_remaining = skb->len;
671 subflow->map_data_len = skb->len;
674 offset = seq - TCP_SKB_CB(skb)->seq;
675 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
681 if (offset < skb->len) {
682 size_t len = skb->len - offset;
687 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
691 if (WARN_ON_ONCE(map_remaining < len))
695 sk_eat_skb(ssk, skb);
699 WRITE_ONCE(tp->copied_seq, seq);
700 more_data_avail = mptcp_subflow_data_available(ssk);
702 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
706 } while (more_data_avail);
712 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
714 struct sock *sk = (struct sock *)msk;
715 struct sk_buff *skb, *tail;
720 p = rb_first(&msk->out_of_order_queue);
721 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
724 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
728 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
730 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
733 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
737 end_seq = MPTCP_SKB_CB(skb)->end_seq;
738 tail = skb_peek_tail(&sk->sk_receive_queue);
739 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
740 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
742 /* skip overlapping data, if any */
743 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
744 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
746 MPTCP_SKB_CB(skb)->offset += delta;
747 MPTCP_SKB_CB(skb)->map_seq += delta;
748 __skb_queue_tail(&sk->sk_receive_queue, skb);
750 msk->ack_seq = end_seq;
756 /* In most cases we will be able to lock the mptcp socket. If its already
757 * owned, we need to defer to the work queue to avoid ABBA deadlock.
759 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
761 struct sock *sk = (struct sock *)msk;
762 unsigned int moved = 0;
764 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
765 __mptcp_ofo_queue(msk);
766 if (unlikely(ssk->sk_err)) {
767 if (!sock_owned_by_user(sk))
768 __mptcp_error_report(sk);
770 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
773 /* If the moves have caught up with the DATA_FIN sequence number
774 * it's time to ack the DATA_FIN and change socket state, but
775 * this is not a good place to change state. Let the workqueue
778 if (mptcp_pending_data_fin(sk, NULL))
779 mptcp_schedule_work(sk);
783 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
785 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
786 struct mptcp_sock *msk = mptcp_sk(sk);
787 int sk_rbuf, ssk_rbuf;
789 /* The peer can send data while we are shutting down this
790 * subflow at msk destruction time, but we must avoid enqueuing
791 * more data to the msk receive queue
793 if (unlikely(subflow->disposable))
796 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
797 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
798 if (unlikely(ssk_rbuf > sk_rbuf))
801 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
802 if (__mptcp_rmem(sk) > sk_rbuf) {
803 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
807 /* Wake-up the reader only for in-sequence data */
809 if (move_skbs_to_msk(msk, ssk))
810 sk->sk_data_ready(sk);
812 mptcp_data_unlock(sk);
815 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
817 struct sock *sk = (struct sock *)msk;
819 if (sk->sk_state != TCP_ESTABLISHED)
822 /* attach to msk socket only after we are sure we will deal with it
825 if (sk->sk_socket && !ssk->sk_socket)
826 mptcp_sock_graft(ssk, sk->sk_socket);
828 mptcp_sockopt_sync_locked(msk, ssk);
832 static void __mptcp_flush_join_list(struct sock *sk)
834 struct mptcp_subflow_context *tmp, *subflow;
835 struct mptcp_sock *msk = mptcp_sk(sk);
837 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
838 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
839 bool slow = lock_sock_fast(ssk);
841 list_move_tail(&subflow->node, &msk->conn_list);
842 if (!__mptcp_finish_join(msk, ssk))
843 mptcp_subflow_reset(ssk);
844 unlock_sock_fast(ssk, slow);
848 static bool mptcp_timer_pending(struct sock *sk)
850 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
853 static void mptcp_reset_timer(struct sock *sk)
855 struct inet_connection_sock *icsk = inet_csk(sk);
858 /* prevent rescheduling on close */
859 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
862 tout = mptcp_sk(sk)->timer_ival;
863 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
866 bool mptcp_schedule_work(struct sock *sk)
868 if (inet_sk_state_load(sk) != TCP_CLOSE &&
869 schedule_work(&mptcp_sk(sk)->work)) {
870 /* each subflow already holds a reference to the sk, and the
871 * workqueue is invoked by a subflow, so sk can't go away here.
879 void mptcp_subflow_eof(struct sock *sk)
881 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
882 mptcp_schedule_work(sk);
885 static void mptcp_check_for_eof(struct mptcp_sock *msk)
887 struct mptcp_subflow_context *subflow;
888 struct sock *sk = (struct sock *)msk;
891 mptcp_for_each_subflow(msk, subflow)
892 receivers += !subflow->rx_eof;
896 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
897 /* hopefully temporary hack: propagate shutdown status
898 * to msk, when all subflows agree on it
900 sk->sk_shutdown |= RCV_SHUTDOWN;
902 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
903 sk->sk_data_ready(sk);
906 switch (sk->sk_state) {
907 case TCP_ESTABLISHED:
908 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
911 inet_sk_state_store(sk, TCP_CLOSING);
914 inet_sk_state_store(sk, TCP_CLOSE);
919 mptcp_close_wake_up(sk);
922 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
924 struct mptcp_subflow_context *subflow;
926 msk_owned_by_me(msk);
928 mptcp_for_each_subflow(msk, subflow) {
929 if (READ_ONCE(subflow->data_avail))
930 return mptcp_subflow_tcp_sock(subflow);
936 static bool mptcp_skb_can_collapse_to(u64 write_seq,
937 const struct sk_buff *skb,
938 const struct mptcp_ext *mpext)
940 if (!tcp_skb_can_collapse_to(skb))
943 /* can collapse only if MPTCP level sequence is in order and this
944 * mapping has not been xmitted yet
946 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
950 /* we can append data to the given data frag if:
951 * - there is space available in the backing page_frag
952 * - the data frag tail matches the current page_frag free offset
953 * - the data frag end sequence number matches the current write seq
955 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
956 const struct page_frag *pfrag,
957 const struct mptcp_data_frag *df)
959 return df && pfrag->page == df->page &&
960 pfrag->size - pfrag->offset > 0 &&
961 pfrag->offset == (df->offset + df->data_len) &&
962 df->data_seq + df->data_len == msk->write_seq;
965 static void dfrag_uncharge(struct sock *sk, int len)
967 sk_mem_uncharge(sk, len);
968 sk_wmem_queued_add(sk, -len);
971 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
973 int len = dfrag->data_len + dfrag->overhead;
975 list_del(&dfrag->list);
976 dfrag_uncharge(sk, len);
977 put_page(dfrag->page);
980 static void __mptcp_clean_una(struct sock *sk)
982 struct mptcp_sock *msk = mptcp_sk(sk);
983 struct mptcp_data_frag *dtmp, *dfrag;
986 /* on fallback we just need to ignore snd_una, as this is really
989 if (__mptcp_check_fallback(msk))
990 msk->snd_una = READ_ONCE(msk->snd_nxt);
992 snd_una = msk->snd_una;
993 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
994 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
997 if (unlikely(dfrag == msk->first_pending)) {
998 /* in recovery mode can see ack after the current snd head */
999 if (WARN_ON_ONCE(!msk->recovery))
1002 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1005 dfrag_clear(sk, dfrag);
1008 dfrag = mptcp_rtx_head(sk);
1009 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1010 u64 delta = snd_una - dfrag->data_seq;
1012 /* prevent wrap around in recovery mode */
1013 if (unlikely(delta > dfrag->already_sent)) {
1014 if (WARN_ON_ONCE(!msk->recovery))
1016 if (WARN_ON_ONCE(delta > dfrag->data_len))
1018 dfrag->already_sent += delta - dfrag->already_sent;
1021 dfrag->data_seq += delta;
1022 dfrag->offset += delta;
1023 dfrag->data_len -= delta;
1024 dfrag->already_sent -= delta;
1026 dfrag_uncharge(sk, delta);
1029 /* all retransmitted data acked, recovery completed */
1030 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1031 msk->recovery = false;
1034 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1035 snd_una == READ_ONCE(msk->write_seq)) {
1036 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1037 mptcp_stop_timer(sk);
1039 mptcp_reset_timer(sk);
1043 static void __mptcp_clean_una_wakeup(struct sock *sk)
1045 lockdep_assert_held_once(&sk->sk_lock.slock);
1047 __mptcp_clean_una(sk);
1048 mptcp_write_space(sk);
1051 static void mptcp_clean_una_wakeup(struct sock *sk)
1053 mptcp_data_lock(sk);
1054 __mptcp_clean_una_wakeup(sk);
1055 mptcp_data_unlock(sk);
1058 static void mptcp_enter_memory_pressure(struct sock *sk)
1060 struct mptcp_subflow_context *subflow;
1061 struct mptcp_sock *msk = mptcp_sk(sk);
1064 sk_stream_moderate_sndbuf(sk);
1065 mptcp_for_each_subflow(msk, subflow) {
1066 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1069 tcp_enter_memory_pressure(ssk);
1070 sk_stream_moderate_sndbuf(ssk);
1075 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1078 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1080 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1081 pfrag, sk->sk_allocation)))
1084 mptcp_enter_memory_pressure(sk);
1088 static struct mptcp_data_frag *
1089 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1092 int offset = ALIGN(orig_offset, sizeof(long));
1093 struct mptcp_data_frag *dfrag;
1095 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1096 dfrag->data_len = 0;
1097 dfrag->data_seq = msk->write_seq;
1098 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1099 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1100 dfrag->already_sent = 0;
1101 dfrag->page = pfrag->page;
1106 struct mptcp_sendmsg_info {
1112 bool data_lock_held;
1115 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1116 u64 data_seq, int avail_size)
1118 u64 window_end = mptcp_wnd_end(msk);
1121 if (__mptcp_check_fallback(msk))
1124 mptcp_snd_wnd = window_end - data_seq;
1125 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1127 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1128 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1129 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1135 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1137 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1141 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1145 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1147 struct sk_buff *skb;
1149 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1151 if (likely(__mptcp_add_ext(skb, gfp))) {
1152 skb_reserve(skb, MAX_TCP_HEADER);
1153 skb->ip_summed = CHECKSUM_PARTIAL;
1154 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1159 mptcp_enter_memory_pressure(sk);
1164 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1166 struct sk_buff *skb;
1168 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1172 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1173 tcp_skb_entail(ssk, skb);
1176 tcp_skb_tsorted_anchor_cleanup(skb);
1181 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1183 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1185 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1188 /* note: this always recompute the csum on the whole skb, even
1189 * if we just appended a single frag. More status info needed
1191 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1193 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1194 __wsum csum = ~csum_unfold(mpext->csum);
1195 int offset = skb->len - added;
1197 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1200 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1202 struct mptcp_ext *mpext)
1207 mpext->infinite_map = 1;
1208 mpext->data_len = 0;
1210 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1211 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1213 mptcp_do_fallback(ssk);
1216 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1217 struct mptcp_data_frag *dfrag,
1218 struct mptcp_sendmsg_info *info)
1220 u64 data_seq = dfrag->data_seq + info->sent;
1221 int offset = dfrag->offset + info->sent;
1222 struct mptcp_sock *msk = mptcp_sk(sk);
1223 bool zero_window_probe = false;
1224 struct mptcp_ext *mpext = NULL;
1225 bool can_coalesce = false;
1226 bool reuse_skb = true;
1227 struct sk_buff *skb;
1231 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1232 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1234 if (WARN_ON_ONCE(info->sent > info->limit ||
1235 info->limit > dfrag->data_len))
1238 if (unlikely(!__tcp_can_send(ssk)))
1241 /* compute send limit */
1242 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1243 copy = info->size_goal;
1245 skb = tcp_write_queue_tail(ssk);
1246 if (skb && copy > skb->len) {
1247 /* Limit the write to the size available in the
1248 * current skb, if any, so that we create at most a new skb.
1249 * Explicitly tells TCP internals to avoid collapsing on later
1250 * queue management operation, to avoid breaking the ext <->
1251 * SSN association set here
1253 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1254 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1255 TCP_SKB_CB(skb)->eor = 1;
1259 i = skb_shinfo(skb)->nr_frags;
1260 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1261 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1262 tcp_mark_push(tcp_sk(ssk), skb);
1269 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1273 i = skb_shinfo(skb)->nr_frags;
1275 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1278 /* Zero window and all data acked? Probe. */
1279 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1281 u64 snd_una = READ_ONCE(msk->snd_una);
1283 if (snd_una != msk->snd_nxt) {
1284 tcp_remove_empty_skb(ssk);
1288 zero_window_probe = true;
1289 data_seq = snd_una - 1;
1292 /* all mptcp-level data is acked, no skbs should be present into the
1295 WARN_ON_ONCE(reuse_skb);
1298 copy = min_t(size_t, copy, info->limit - info->sent);
1299 if (!sk_wmem_schedule(ssk, copy)) {
1300 tcp_remove_empty_skb(ssk);
1305 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1307 get_page(dfrag->page);
1308 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1312 skb->data_len += copy;
1313 skb->truesize += copy;
1314 sk_wmem_queued_add(ssk, copy);
1315 sk_mem_charge(ssk, copy);
1316 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1317 TCP_SKB_CB(skb)->end_seq += copy;
1318 tcp_skb_pcount_set(skb, 0);
1320 /* on skb reuse we just need to update the DSS len */
1322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1323 mpext->data_len += copy;
1324 WARN_ON_ONCE(zero_window_probe);
1328 memset(mpext, 0, sizeof(*mpext));
1329 mpext->data_seq = data_seq;
1330 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1331 mpext->data_len = copy;
1335 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1336 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1339 if (zero_window_probe) {
1340 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1342 if (READ_ONCE(msk->csum_enabled))
1343 mptcp_update_data_checksum(skb, copy);
1344 tcp_push_pending_frames(ssk);
1348 if (READ_ONCE(msk->csum_enabled))
1349 mptcp_update_data_checksum(skb, copy);
1350 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1351 mptcp_update_infinite_map(msk, ssk, mpext);
1352 trace_mptcp_sendmsg_frag(mpext);
1353 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1357 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1358 sizeof(struct tcphdr) - \
1359 MAX_TCP_OPTION_SPACE - \
1360 sizeof(struct ipv6hdr) - \
1361 sizeof(struct frag_hdr))
1363 struct subflow_send_info {
1368 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1370 if (!subflow->stale)
1374 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1377 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1379 if (unlikely(subflow->stale)) {
1380 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1382 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1385 mptcp_subflow_set_active(subflow);
1387 return __mptcp_subflow_active(subflow);
1390 #define SSK_MODE_ACTIVE 0
1391 #define SSK_MODE_BACKUP 1
1392 #define SSK_MODE_MAX 2
1394 /* implement the mptcp packet scheduler;
1395 * returns the subflow that will transmit the next DSS
1396 * additionally updates the rtx timeout
1398 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1400 struct subflow_send_info send_info[SSK_MODE_MAX];
1401 struct mptcp_subflow_context *subflow;
1402 struct sock *sk = (struct sock *)msk;
1403 u32 pace, burst, wmem;
1404 int i, nr_active = 0;
1409 msk_owned_by_me(msk);
1411 if (__mptcp_check_fallback(msk)) {
1414 return __tcp_can_send(msk->first) &&
1415 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1418 /* re-use last subflow, if the burst allow that */
1419 if (msk->last_snd && msk->snd_burst > 0 &&
1420 sk_stream_memory_free(msk->last_snd) &&
1421 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1422 mptcp_set_timeout(sk);
1423 return msk->last_snd;
1426 /* pick the subflow with the lower wmem/wspace ratio */
1427 for (i = 0; i < SSK_MODE_MAX; ++i) {
1428 send_info[i].ssk = NULL;
1429 send_info[i].linger_time = -1;
1432 mptcp_for_each_subflow(msk, subflow) {
1433 trace_mptcp_subflow_get_send(subflow);
1434 ssk = mptcp_subflow_tcp_sock(subflow);
1435 if (!mptcp_subflow_active(subflow))
1438 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1439 nr_active += !subflow->backup;
1440 pace = subflow->avg_pacing_rate;
1441 if (unlikely(!pace)) {
1442 /* init pacing rate from socket */
1443 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1444 pace = subflow->avg_pacing_rate;
1449 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1450 if (linger_time < send_info[subflow->backup].linger_time) {
1451 send_info[subflow->backup].ssk = ssk;
1452 send_info[subflow->backup].linger_time = linger_time;
1455 __mptcp_set_timeout(sk, tout);
1457 /* pick the best backup if no other subflow is active */
1459 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1461 /* According to the blest algorithm, to avoid HoL blocking for the
1462 * faster flow, we need to:
1463 * - estimate the faster flow linger time
1464 * - use the above to estimate the amount of byte transferred
1465 * by the faster flow
1466 * - check that the amount of queued data is greter than the above,
1467 * otherwise do not use the picked, slower, subflow
1468 * We select the subflow with the shorter estimated time to flush
1469 * the queued mem, which basically ensure the above. We just need
1470 * to check that subflow has a non empty cwin.
1472 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1473 if (!ssk || !sk_stream_memory_free(ssk))
1476 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1477 wmem = READ_ONCE(ssk->sk_wmem_queued);
1479 msk->last_snd = NULL;
1483 subflow = mptcp_subflow_ctx(ssk);
1484 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1485 READ_ONCE(ssk->sk_pacing_rate) * burst,
1487 msk->last_snd = ssk;
1488 msk->snd_burst = burst;
1492 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1494 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1498 static void mptcp_update_post_push(struct mptcp_sock *msk,
1499 struct mptcp_data_frag *dfrag,
1502 u64 snd_nxt_new = dfrag->data_seq;
1504 dfrag->already_sent += sent;
1506 msk->snd_burst -= sent;
1508 snd_nxt_new += dfrag->already_sent;
1510 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1511 * is recovering after a failover. In that event, this re-sends
1514 * Thus compute snd_nxt_new candidate based on
1515 * the dfrag->data_seq that was sent and the data
1516 * that has been handed to the subflow for transmission
1517 * and skip update in case it was old dfrag.
1519 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1520 msk->snd_nxt = snd_nxt_new;
1523 void mptcp_check_and_set_pending(struct sock *sk)
1525 if (mptcp_send_head(sk))
1526 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1529 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1531 struct sock *prev_ssk = NULL, *ssk = NULL;
1532 struct mptcp_sock *msk = mptcp_sk(sk);
1533 struct mptcp_sendmsg_info info = {
1536 bool do_check_data_fin = false;
1537 struct mptcp_data_frag *dfrag;
1540 while ((dfrag = mptcp_send_head(sk))) {
1541 info.sent = dfrag->already_sent;
1542 info.limit = dfrag->data_len;
1543 len = dfrag->data_len - dfrag->already_sent;
1548 ssk = mptcp_subflow_get_send(msk);
1550 /* First check. If the ssk has changed since
1551 * the last round, release prev_ssk
1553 if (ssk != prev_ssk && prev_ssk)
1554 mptcp_push_release(prev_ssk, &info);
1558 /* Need to lock the new subflow only if different
1559 * from the previous one, otherwise we are still
1560 * helding the relevant lock
1562 if (ssk != prev_ssk)
1565 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1569 mptcp_push_release(ssk, &info);
1573 do_check_data_fin = true;
1577 mptcp_update_post_push(msk, dfrag, ret);
1579 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1582 /* at this point we held the socket lock for the last subflow we used */
1584 mptcp_push_release(ssk, &info);
1587 /* ensure the rtx timer is running */
1588 if (!mptcp_timer_pending(sk))
1589 mptcp_reset_timer(sk);
1590 if (do_check_data_fin)
1591 __mptcp_check_send_data_fin(sk);
1594 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1596 struct mptcp_sock *msk = mptcp_sk(sk);
1597 struct mptcp_sendmsg_info info = {
1598 .data_lock_held = true,
1600 struct mptcp_data_frag *dfrag;
1601 struct sock *xmit_ssk;
1602 int len, copied = 0;
1605 while ((dfrag = mptcp_send_head(sk))) {
1606 info.sent = dfrag->already_sent;
1607 info.limit = dfrag->data_len;
1608 len = dfrag->data_len - dfrag->already_sent;
1612 /* check for a different subflow usage only after
1613 * spooling the first chunk of data
1615 xmit_ssk = first ? ssk : mptcp_subflow_get_send(msk);
1618 if (xmit_ssk != ssk) {
1619 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1620 MPTCP_DELEGATE_SEND);
1624 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1633 mptcp_update_post_push(msk, dfrag, ret);
1635 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1639 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1640 * not going to flush it via release_sock()
1643 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1645 if (!mptcp_timer_pending(sk))
1646 mptcp_reset_timer(sk);
1648 if (msk->snd_data_fin_enable &&
1649 msk->snd_nxt + 1 == msk->write_seq)
1650 mptcp_schedule_work(sk);
1654 static void mptcp_set_nospace(struct sock *sk)
1656 /* enable autotune */
1657 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1659 /* will be cleared on avail space */
1660 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1663 static int mptcp_disconnect(struct sock *sk, int flags);
1665 static int mptcp_sendmsg_fastopen(struct sock *sk, struct sock *ssk, struct msghdr *msg,
1666 size_t len, int *copied_syn)
1668 unsigned int saved_flags = msg->msg_flags;
1669 struct mptcp_sock *msk = mptcp_sk(sk);
1673 msg->msg_flags |= MSG_DONTWAIT;
1674 msk->connect_flags = O_NONBLOCK;
1675 msk->fastopening = 1;
1676 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1677 msk->fastopening = 0;
1678 msg->msg_flags = saved_flags;
1681 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1682 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1683 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1684 msg->msg_namelen, msg->msg_flags, 1);
1686 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1687 * case of any error, except timeout or signal
1689 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1691 } else if (ret && ret != -EINPROGRESS) {
1692 mptcp_disconnect(sk, 0);
1698 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1700 struct mptcp_sock *msk = mptcp_sk(sk);
1701 struct page_frag *pfrag;
1702 struct socket *ssock;
1707 /* silently ignore everything else */
1708 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1712 ssock = __mptcp_nmpc_socket(msk);
1713 if (unlikely(ssock && (inet_sk(ssock->sk)->defer_connect ||
1714 msg->msg_flags & MSG_FASTOPEN))) {
1717 ret = mptcp_sendmsg_fastopen(sk, ssock->sk, msg, len, &copied_syn);
1718 copied += copied_syn;
1719 if (ret == -EINPROGRESS && copied_syn > 0)
1725 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1727 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1728 ret = sk_stream_wait_connect(sk, &timeo);
1734 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1737 pfrag = sk_page_frag(sk);
1739 while (msg_data_left(msg)) {
1740 int total_ts, frag_truesize = 0;
1741 struct mptcp_data_frag *dfrag;
1742 bool dfrag_collapsed;
1743 size_t psize, offset;
1745 /* reuse tail pfrag, if possible, or carve a new one from the
1748 dfrag = mptcp_pending_tail(sk);
1749 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1750 if (!dfrag_collapsed) {
1751 if (!sk_stream_memory_free(sk))
1752 goto wait_for_memory;
1754 if (!mptcp_page_frag_refill(sk, pfrag))
1755 goto wait_for_memory;
1757 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1758 frag_truesize = dfrag->overhead;
1761 /* we do not bound vs wspace, to allow a single packet.
1762 * memory accounting will prevent execessive memory usage
1765 offset = dfrag->offset + dfrag->data_len;
1766 psize = pfrag->size - offset;
1767 psize = min_t(size_t, psize, msg_data_left(msg));
1768 total_ts = psize + frag_truesize;
1770 if (!sk_wmem_schedule(sk, total_ts))
1771 goto wait_for_memory;
1773 if (copy_page_from_iter(dfrag->page, offset, psize,
1774 &msg->msg_iter) != psize) {
1779 /* data successfully copied into the write queue */
1780 sk->sk_forward_alloc -= total_ts;
1782 dfrag->data_len += psize;
1783 frag_truesize += psize;
1784 pfrag->offset += frag_truesize;
1785 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1787 /* charge data on mptcp pending queue to the msk socket
1788 * Note: we charge such data both to sk and ssk
1790 sk_wmem_queued_add(sk, frag_truesize);
1791 if (!dfrag_collapsed) {
1792 get_page(dfrag->page);
1793 list_add_tail(&dfrag->list, &msk->rtx_queue);
1794 if (!msk->first_pending)
1795 WRITE_ONCE(msk->first_pending, dfrag);
1797 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1798 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1804 mptcp_set_nospace(sk);
1805 __mptcp_push_pending(sk, msg->msg_flags);
1806 ret = sk_stream_wait_memory(sk, &timeo);
1812 __mptcp_push_pending(sk, msg->msg_flags);
1822 copied = sk_stream_error(sk, msg->msg_flags, ret);
1826 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1828 size_t len, int flags,
1829 struct scm_timestamping_internal *tss,
1832 struct sk_buff *skb, *tmp;
1835 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1836 u32 offset = MPTCP_SKB_CB(skb)->offset;
1837 u32 data_len = skb->len - offset;
1838 u32 count = min_t(size_t, len - copied, data_len);
1841 if (!(flags & MSG_TRUNC)) {
1842 err = skb_copy_datagram_msg(skb, offset, msg, count);
1843 if (unlikely(err < 0)) {
1850 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1851 tcp_update_recv_tstamps(skb, tss);
1852 *cmsg_flags |= MPTCP_CMSG_TS;
1857 if (count < data_len) {
1858 if (!(flags & MSG_PEEK)) {
1859 MPTCP_SKB_CB(skb)->offset += count;
1860 MPTCP_SKB_CB(skb)->map_seq += count;
1865 if (!(flags & MSG_PEEK)) {
1866 /* we will bulk release the skb memory later */
1867 skb->destructor = NULL;
1868 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1869 __skb_unlink(skb, &msk->receive_queue);
1880 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1882 * Only difference: Use highest rtt estimate of the subflows in use.
1884 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1886 struct mptcp_subflow_context *subflow;
1887 struct sock *sk = (struct sock *)msk;
1888 u32 time, advmss = 1;
1891 msk_owned_by_me(msk);
1896 msk->rcvq_space.copied += copied;
1898 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1899 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1901 rtt_us = msk->rcvq_space.rtt_us;
1902 if (rtt_us && time < (rtt_us >> 3))
1906 mptcp_for_each_subflow(msk, subflow) {
1907 const struct tcp_sock *tp;
1911 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1913 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1914 sf_advmss = READ_ONCE(tp->advmss);
1916 rtt_us = max(sf_rtt_us, rtt_us);
1917 advmss = max(sf_advmss, advmss);
1920 msk->rcvq_space.rtt_us = rtt_us;
1921 if (time < (rtt_us >> 3) || rtt_us == 0)
1924 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1927 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1928 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1932 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1934 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1936 do_div(grow, msk->rcvq_space.space);
1937 rcvwin += (grow << 1);
1939 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1940 while (tcp_win_from_space(sk, rcvmem) < advmss)
1943 do_div(rcvwin, advmss);
1944 rcvbuf = min_t(u64, rcvwin * rcvmem,
1945 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1947 if (rcvbuf > sk->sk_rcvbuf) {
1950 window_clamp = tcp_win_from_space(sk, rcvbuf);
1951 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1953 /* Make subflows follow along. If we do not do this, we
1954 * get drops at subflow level if skbs can't be moved to
1955 * the mptcp rx queue fast enough (announced rcv_win can
1956 * exceed ssk->sk_rcvbuf).
1958 mptcp_for_each_subflow(msk, subflow) {
1962 ssk = mptcp_subflow_tcp_sock(subflow);
1963 slow = lock_sock_fast(ssk);
1964 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1965 tcp_sk(ssk)->window_clamp = window_clamp;
1966 tcp_cleanup_rbuf(ssk, 1);
1967 unlock_sock_fast(ssk, slow);
1972 msk->rcvq_space.space = msk->rcvq_space.copied;
1974 msk->rcvq_space.copied = 0;
1975 msk->rcvq_space.time = mstamp;
1978 static void __mptcp_update_rmem(struct sock *sk)
1980 struct mptcp_sock *msk = mptcp_sk(sk);
1982 if (!msk->rmem_released)
1985 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1986 mptcp_rmem_uncharge(sk, msk->rmem_released);
1987 WRITE_ONCE(msk->rmem_released, 0);
1990 static void __mptcp_splice_receive_queue(struct sock *sk)
1992 struct mptcp_sock *msk = mptcp_sk(sk);
1994 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1997 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1999 struct sock *sk = (struct sock *)msk;
2000 unsigned int moved = 0;
2004 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2007 /* we can have data pending in the subflows only if the msk
2008 * receive buffer was full at subflow_data_ready() time,
2009 * that is an unlikely slow path.
2014 slowpath = lock_sock_fast(ssk);
2015 mptcp_data_lock(sk);
2016 __mptcp_update_rmem(sk);
2017 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2018 mptcp_data_unlock(sk);
2020 if (unlikely(ssk->sk_err))
2021 __mptcp_error_report(sk);
2022 unlock_sock_fast(ssk, slowpath);
2025 /* acquire the data lock only if some input data is pending */
2027 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2028 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2029 mptcp_data_lock(sk);
2030 __mptcp_update_rmem(sk);
2031 ret |= __mptcp_ofo_queue(msk);
2032 __mptcp_splice_receive_queue(sk);
2033 mptcp_data_unlock(sk);
2036 mptcp_check_data_fin((struct sock *)msk);
2037 return !skb_queue_empty(&msk->receive_queue);
2040 static unsigned int mptcp_inq_hint(const struct sock *sk)
2042 const struct mptcp_sock *msk = mptcp_sk(sk);
2043 const struct sk_buff *skb;
2045 skb = skb_peek(&msk->receive_queue);
2047 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2049 if (hint_val >= INT_MAX)
2052 return (unsigned int)hint_val;
2055 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2061 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2062 int flags, int *addr_len)
2064 struct mptcp_sock *msk = mptcp_sk(sk);
2065 struct scm_timestamping_internal tss;
2066 int copied = 0, cmsg_flags = 0;
2070 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2071 if (unlikely(flags & MSG_ERRQUEUE))
2072 return inet_recv_error(sk, msg, len, addr_len);
2075 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2080 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2082 len = min_t(size_t, len, INT_MAX);
2083 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2085 if (unlikely(msk->recvmsg_inq))
2086 cmsg_flags = MPTCP_CMSG_INQ;
2088 while (copied < len) {
2091 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2092 if (unlikely(bytes_read < 0)) {
2094 copied = bytes_read;
2098 copied += bytes_read;
2100 /* be sure to advertise window change */
2101 mptcp_cleanup_rbuf(msk);
2103 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2106 /* only the master socket status is relevant here. The exit
2107 * conditions mirror closely tcp_recvmsg()
2109 if (copied >= target)
2114 sk->sk_state == TCP_CLOSE ||
2115 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2117 signal_pending(current))
2121 copied = sock_error(sk);
2125 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2126 mptcp_check_for_eof(msk);
2128 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2129 /* race breaker: the shutdown could be after the
2130 * previous receive queue check
2132 if (__mptcp_move_skbs(msk))
2137 if (sk->sk_state == TCP_CLOSE) {
2147 if (signal_pending(current)) {
2148 copied = sock_intr_errno(timeo);
2153 pr_debug("block timeout %ld", timeo);
2154 sk_wait_data(sk, &timeo, NULL);
2158 if (cmsg_flags && copied >= 0) {
2159 if (cmsg_flags & MPTCP_CMSG_TS)
2160 tcp_recv_timestamp(msg, sk, &tss);
2162 if (cmsg_flags & MPTCP_CMSG_INQ) {
2163 unsigned int inq = mptcp_inq_hint(sk);
2165 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2169 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2170 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2171 skb_queue_empty(&msk->receive_queue), copied);
2172 if (!(flags & MSG_PEEK))
2173 mptcp_rcv_space_adjust(msk, copied);
2179 static void mptcp_retransmit_timer(struct timer_list *t)
2181 struct inet_connection_sock *icsk = from_timer(icsk, t,
2182 icsk_retransmit_timer);
2183 struct sock *sk = &icsk->icsk_inet.sk;
2184 struct mptcp_sock *msk = mptcp_sk(sk);
2187 if (!sock_owned_by_user(sk)) {
2188 /* we need a process context to retransmit */
2189 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2190 mptcp_schedule_work(sk);
2192 /* delegate our work to tcp_release_cb() */
2193 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2199 static void mptcp_timeout_timer(struct timer_list *t)
2201 struct sock *sk = from_timer(sk, t, sk_timer);
2203 mptcp_schedule_work(sk);
2207 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2210 * A backup subflow is returned only if that is the only kind available.
2212 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2214 struct sock *backup = NULL, *pick = NULL;
2215 struct mptcp_subflow_context *subflow;
2216 int min_stale_count = INT_MAX;
2218 msk_owned_by_me(msk);
2220 if (__mptcp_check_fallback(msk))
2223 mptcp_for_each_subflow(msk, subflow) {
2224 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2226 if (!__mptcp_subflow_active(subflow))
2229 /* still data outstanding at TCP level? skip this */
2230 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2231 mptcp_pm_subflow_chk_stale(msk, ssk);
2232 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2236 if (subflow->backup) {
2249 /* use backup only if there are no progresses anywhere */
2250 return min_stale_count > 1 ? backup : NULL;
2253 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2256 iput(SOCK_INODE(msk->subflow));
2257 msk->subflow = NULL;
2261 bool __mptcp_retransmit_pending_data(struct sock *sk)
2263 struct mptcp_data_frag *cur, *rtx_head;
2264 struct mptcp_sock *msk = mptcp_sk(sk);
2266 if (__mptcp_check_fallback(msk))
2269 if (tcp_rtx_and_write_queues_empty(sk))
2272 /* the closing socket has some data untransmitted and/or unacked:
2273 * some data in the mptcp rtx queue has not really xmitted yet.
2274 * keep it simple and re-inject the whole mptcp level rtx queue
2276 mptcp_data_lock(sk);
2277 __mptcp_clean_una_wakeup(sk);
2278 rtx_head = mptcp_rtx_head(sk);
2280 mptcp_data_unlock(sk);
2284 msk->recovery_snd_nxt = msk->snd_nxt;
2285 msk->recovery = true;
2286 mptcp_data_unlock(sk);
2288 msk->first_pending = rtx_head;
2291 /* be sure to clear the "sent status" on all re-injected fragments */
2292 list_for_each_entry(cur, &msk->rtx_queue, list) {
2293 if (!cur->already_sent)
2295 cur->already_sent = 0;
2301 /* flags for __mptcp_close_ssk() */
2302 #define MPTCP_CF_PUSH BIT(1)
2303 #define MPTCP_CF_FASTCLOSE BIT(2)
2305 /* subflow sockets can be either outgoing (connect) or incoming
2308 * Outgoing subflows use in-kernel sockets.
2309 * Incoming subflows do not have their own 'struct socket' allocated,
2310 * so we need to use tcp_close() after detaching them from the mptcp
2313 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2314 struct mptcp_subflow_context *subflow,
2317 struct mptcp_sock *msk = mptcp_sk(sk);
2318 bool need_push, dispose_it;
2320 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2322 list_del(&subflow->node);
2324 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2326 if (flags & MPTCP_CF_FASTCLOSE) {
2327 /* be sure to force the tcp_disconnect() path,
2328 * to generate the egress reset
2330 ssk->sk_lingertime = 0;
2331 sock_set_flag(ssk, SOCK_LINGER);
2332 subflow->send_fastclose = 1;
2335 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2337 tcp_disconnect(ssk, 0);
2338 msk->subflow->state = SS_UNCONNECTED;
2339 mptcp_subflow_ctx_reset(subflow);
2345 subflow->disposable = 1;
2347 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2348 * the ssk has been already destroyed, we just need to release the
2349 * reference owned by msk;
2351 if (!inet_csk(ssk)->icsk_ulp_ops) {
2352 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2353 kfree_rcu(subflow, rcu);
2354 } else if (msk->in_accept_queue && msk->first == ssk) {
2355 /* if the first subflow moved to a close state, e.g. due to
2356 * incoming reset and we reach here before inet_child_forget()
2357 * the TCP stack could later try to close it via
2358 * inet_csk_listen_stop(), or deliver it to the user space via
2360 * We can't delete the subflow - or risk a double free - nor let
2361 * the msk survive - or will be leaked in the non accept scenario:
2362 * fallback and let TCP cope with the subflow cleanup.
2364 WARN_ON_ONCE(sock_flag(ssk, SOCK_DEAD));
2365 mptcp_subflow_drop_ctx(ssk);
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(sk, ssk);
2371 inet_csk_listen_stop(ssk);
2372 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2374 __tcp_close(ssk, 0);
2376 /* close acquired an extra ref */
2383 if (ssk == msk->first)
2387 if (ssk == msk->last_snd)
2388 msk->last_snd = NULL;
2391 __mptcp_push_pending(sk, 0);
2394 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2395 struct mptcp_subflow_context *subflow)
2397 if (sk->sk_state == TCP_ESTABLISHED)
2398 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2400 /* subflow aborted before reaching the fully_established status
2401 * attempt the creation of the next subflow
2403 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2405 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2408 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2413 static void __mptcp_close_subflow(struct sock *sk)
2415 struct mptcp_subflow_context *subflow, *tmp;
2416 struct mptcp_sock *msk = mptcp_sk(sk);
2420 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2421 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2423 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2426 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2427 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2430 mptcp_close_ssk(sk, ssk, subflow);
2433 /* if the MPC subflow has been closed before the msk is accepted,
2434 * msk will never be accept-ed, close it now
2436 if (!msk->first && msk->in_accept_queue) {
2437 sock_set_flag(sk, SOCK_DEAD);
2438 inet_sk_state_store(sk, TCP_CLOSE);
2442 static bool mptcp_check_close_timeout(const struct sock *sk)
2444 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2445 struct mptcp_subflow_context *subflow;
2447 if (delta >= TCP_TIMEWAIT_LEN)
2450 /* if all subflows are in closed status don't bother with additional
2453 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2454 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2461 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2463 struct mptcp_subflow_context *subflow, *tmp;
2464 struct sock *sk = (struct sock *)msk;
2466 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2469 mptcp_token_destroy(msk);
2471 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2472 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2475 slow = lock_sock_fast(tcp_sk);
2476 if (tcp_sk->sk_state != TCP_CLOSE) {
2477 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2478 tcp_set_state(tcp_sk, TCP_CLOSE);
2480 unlock_sock_fast(tcp_sk, slow);
2483 /* Mirror the tcp_reset() error propagation */
2484 switch (sk->sk_state) {
2486 sk->sk_err = ECONNREFUSED;
2488 case TCP_CLOSE_WAIT:
2494 sk->sk_err = ECONNRESET;
2497 inet_sk_state_store(sk, TCP_CLOSE);
2498 sk->sk_shutdown = SHUTDOWN_MASK;
2499 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2500 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2502 /* the calling mptcp_worker will properly destroy the socket */
2503 if (sock_flag(sk, SOCK_DEAD))
2506 sk->sk_state_change(sk);
2507 sk_error_report(sk);
2510 static void __mptcp_retrans(struct sock *sk)
2512 struct mptcp_sock *msk = mptcp_sk(sk);
2513 struct mptcp_sendmsg_info info = {};
2514 struct mptcp_data_frag *dfrag;
2519 mptcp_clean_una_wakeup(sk);
2521 /* first check ssk: need to kick "stale" logic */
2522 ssk = mptcp_subflow_get_retrans(msk);
2523 dfrag = mptcp_rtx_head(sk);
2525 if (mptcp_data_fin_enabled(msk)) {
2526 struct inet_connection_sock *icsk = inet_csk(sk);
2528 icsk->icsk_retransmits++;
2529 mptcp_set_datafin_timeout(sk);
2530 mptcp_send_ack(msk);
2535 if (!mptcp_send_head(sk))
2546 /* limit retransmission to the bytes already sent on some subflows */
2548 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2549 while (info.sent < info.limit) {
2550 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2554 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2559 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2560 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2562 WRITE_ONCE(msk->allow_infinite_fallback, false);
2568 mptcp_check_and_set_pending(sk);
2570 if (!mptcp_timer_pending(sk))
2571 mptcp_reset_timer(sk);
2574 /* schedule the timeout timer for the relevant event: either close timeout
2575 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2577 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2579 struct sock *sk = (struct sock *)msk;
2580 unsigned long timeout, close_timeout;
2582 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2585 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2587 /* the close timeout takes precedence on the fail one, and here at least one of
2590 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2592 sk_reset_timer(sk, &sk->sk_timer, timeout);
2595 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2597 struct sock *ssk = msk->first;
2603 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2605 slow = lock_sock_fast(ssk);
2606 mptcp_subflow_reset(ssk);
2607 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2608 unlock_sock_fast(ssk, slow);
2610 mptcp_reset_timeout(msk, 0);
2613 static void mptcp_do_fastclose(struct sock *sk)
2615 struct mptcp_subflow_context *subflow, *tmp;
2616 struct mptcp_sock *msk = mptcp_sk(sk);
2618 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2619 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2620 subflow, MPTCP_CF_FASTCLOSE);
2623 static void mptcp_worker(struct work_struct *work)
2625 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2626 struct sock *sk = (struct sock *)msk;
2627 unsigned long fail_tout;
2631 state = sk->sk_state;
2632 if (unlikely(state == TCP_CLOSE))
2635 mptcp_check_data_fin_ack(sk);
2637 mptcp_check_fastclose(msk);
2639 mptcp_pm_nl_work(msk);
2641 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2642 mptcp_check_for_eof(msk);
2644 __mptcp_check_send_data_fin(sk);
2645 mptcp_check_data_fin(sk);
2647 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2648 __mptcp_close_subflow(sk);
2650 /* There is no point in keeping around an orphaned sk timedout or
2651 * closed, but we need the msk around to reply to incoming DATA_FIN,
2652 * even if it is orphaned and in FIN_WAIT2 state
2654 if (sock_flag(sk, SOCK_DEAD)) {
2655 if (mptcp_check_close_timeout(sk)) {
2656 inet_sk_state_store(sk, TCP_CLOSE);
2657 mptcp_do_fastclose(sk);
2659 if (sk->sk_state == TCP_CLOSE) {
2660 __mptcp_destroy_sock(sk);
2665 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2666 __mptcp_retrans(sk);
2668 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2669 if (fail_tout && time_after(jiffies, fail_tout))
2670 mptcp_mp_fail_no_response(msk);
2677 static int __mptcp_init_sock(struct sock *sk)
2679 struct mptcp_sock *msk = mptcp_sk(sk);
2681 INIT_LIST_HEAD(&msk->conn_list);
2682 INIT_LIST_HEAD(&msk->join_list);
2683 INIT_LIST_HEAD(&msk->rtx_queue);
2684 INIT_WORK(&msk->work, mptcp_worker);
2685 __skb_queue_head_init(&msk->receive_queue);
2686 msk->out_of_order_queue = RB_ROOT;
2687 msk->first_pending = NULL;
2688 msk->rmem_fwd_alloc = 0;
2689 WRITE_ONCE(msk->rmem_released, 0);
2690 msk->timer_ival = TCP_RTO_MIN;
2693 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2694 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2695 WRITE_ONCE(msk->allow_infinite_fallback, true);
2696 msk->recovery = false;
2698 mptcp_pm_data_init(msk);
2700 /* re-use the csk retrans timer for MPTCP-level retrans */
2701 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2702 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2707 static void mptcp_ca_reset(struct sock *sk)
2709 struct inet_connection_sock *icsk = inet_csk(sk);
2711 tcp_assign_congestion_control(sk);
2712 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2714 /* no need to keep a reference to the ops, the name will suffice */
2715 tcp_cleanup_congestion_control(sk);
2716 icsk->icsk_ca_ops = NULL;
2719 static int mptcp_init_sock(struct sock *sk)
2721 struct net *net = sock_net(sk);
2724 ret = __mptcp_init_sock(sk);
2728 if (!mptcp_is_enabled(net))
2729 return -ENOPROTOOPT;
2731 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2734 ret = __mptcp_socket_create(mptcp_sk(sk));
2738 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2740 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2741 * propagate the correct value
2745 sk_sockets_allocated_inc(sk);
2746 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2747 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2752 static void __mptcp_clear_xmit(struct sock *sk)
2754 struct mptcp_sock *msk = mptcp_sk(sk);
2755 struct mptcp_data_frag *dtmp, *dfrag;
2757 WRITE_ONCE(msk->first_pending, NULL);
2758 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2759 dfrag_clear(sk, dfrag);
2762 void mptcp_cancel_work(struct sock *sk)
2764 struct mptcp_sock *msk = mptcp_sk(sk);
2766 if (cancel_work_sync(&msk->work))
2770 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2774 switch (ssk->sk_state) {
2776 if (!(how & RCV_SHUTDOWN))
2780 tcp_disconnect(ssk, O_NONBLOCK);
2783 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2784 pr_debug("Fallback");
2785 ssk->sk_shutdown |= how;
2786 tcp_shutdown(ssk, how);
2788 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2790 if (!mptcp_timer_pending(sk))
2791 mptcp_reset_timer(sk);
2799 static const unsigned char new_state[16] = {
2800 /* current state: new state: action: */
2801 [0 /* (Invalid) */] = TCP_CLOSE,
2802 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2803 [TCP_SYN_SENT] = TCP_CLOSE,
2804 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2805 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2806 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2807 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2808 [TCP_CLOSE] = TCP_CLOSE,
2809 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2810 [TCP_LAST_ACK] = TCP_LAST_ACK,
2811 [TCP_LISTEN] = TCP_CLOSE,
2812 [TCP_CLOSING] = TCP_CLOSING,
2813 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2816 static int mptcp_close_state(struct sock *sk)
2818 int next = (int)new_state[sk->sk_state];
2819 int ns = next & TCP_STATE_MASK;
2821 inet_sk_state_store(sk, ns);
2823 return next & TCP_ACTION_FIN;
2826 static void __mptcp_check_send_data_fin(struct sock *sk)
2828 struct mptcp_subflow_context *subflow;
2829 struct mptcp_sock *msk = mptcp_sk(sk);
2831 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2832 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2833 msk->snd_nxt, msk->write_seq);
2835 /* we still need to enqueue subflows or not really shutting down,
2838 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2839 mptcp_send_head(sk))
2842 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2844 /* fallback socket will not get data_fin/ack, can move to the next
2847 if (__mptcp_check_fallback(msk)) {
2848 WRITE_ONCE(msk->snd_una, msk->write_seq);
2849 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2850 inet_sk_state_store(sk, TCP_CLOSE);
2851 mptcp_close_wake_up(sk);
2852 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2853 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2857 mptcp_for_each_subflow(msk, subflow) {
2858 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2860 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2864 static void __mptcp_wr_shutdown(struct sock *sk)
2866 struct mptcp_sock *msk = mptcp_sk(sk);
2868 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2869 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2870 !!mptcp_send_head(sk));
2872 /* will be ignored by fallback sockets */
2873 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2874 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2876 __mptcp_check_send_data_fin(sk);
2879 static void __mptcp_destroy_sock(struct sock *sk)
2881 struct mptcp_sock *msk = mptcp_sk(sk);
2883 pr_debug("msk=%p", msk);
2887 mptcp_stop_timer(sk);
2888 sk_stop_timer(sk, &sk->sk_timer);
2891 sk->sk_prot->destroy(sk);
2893 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2894 WARN_ON_ONCE(msk->rmem_released);
2895 sk_stream_kill_queues(sk);
2896 xfrm_sk_free_policy(sk);
2901 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2903 /* Concurrent splices from sk_receive_queue into receive_queue will
2904 * always show at least one non-empty queue when checked in this order.
2906 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2907 skb_queue_empty_lockless(&msk->receive_queue))
2910 return EPOLLIN | EPOLLRDNORM;
2913 static void mptcp_listen_inuse_dec(struct sock *sk)
2915 if (inet_sk_state_load(sk) == TCP_LISTEN)
2916 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2919 bool __mptcp_close(struct sock *sk, long timeout)
2921 struct mptcp_subflow_context *subflow;
2922 struct mptcp_sock *msk = mptcp_sk(sk);
2923 bool do_cancel_work = false;
2924 int subflows_alive = 0;
2926 sk->sk_shutdown = SHUTDOWN_MASK;
2928 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2929 mptcp_listen_inuse_dec(sk);
2930 inet_sk_state_store(sk, TCP_CLOSE);
2934 if (mptcp_check_readable(msk)) {
2935 /* the msk has read data, do the MPTCP equivalent of TCP reset */
2936 inet_sk_state_store(sk, TCP_CLOSE);
2937 mptcp_do_fastclose(sk);
2938 } else if (mptcp_close_state(sk)) {
2939 __mptcp_wr_shutdown(sk);
2942 sk_stream_wait_close(sk, timeout);
2945 /* orphan all the subflows */
2946 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2947 mptcp_for_each_subflow(msk, subflow) {
2948 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2949 bool slow = lock_sock_fast_nested(ssk);
2951 subflows_alive += ssk->sk_state != TCP_CLOSE;
2953 /* since the close timeout takes precedence on the fail one,
2956 if (ssk == msk->first)
2957 subflow->fail_tout = 0;
2959 /* detach from the parent socket, but allow data_ready to
2960 * push incoming data into the mptcp stack, to properly ack it
2962 ssk->sk_socket = NULL;
2964 unlock_sock_fast(ssk, slow);
2968 /* all the subflows are closed, only timeout can change the msk
2969 * state, let's not keep resources busy for no reasons
2971 if (subflows_alive == 0)
2972 inet_sk_state_store(sk, TCP_CLOSE);
2975 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2977 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2979 if (sk->sk_state == TCP_CLOSE) {
2980 __mptcp_destroy_sock(sk);
2981 do_cancel_work = true;
2983 mptcp_reset_timeout(msk, 0);
2986 return do_cancel_work;
2989 static void mptcp_close(struct sock *sk, long timeout)
2991 bool do_cancel_work;
2995 do_cancel_work = __mptcp_close(sk, timeout);
2998 mptcp_cancel_work(sk);
3003 void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3005 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3006 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3007 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3009 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3010 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3013 msk6->saddr = ssk6->saddr;
3014 msk6->flow_label = ssk6->flow_label;
3018 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3019 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3020 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3021 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3022 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3023 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3026 static int mptcp_disconnect(struct sock *sk, int flags)
3028 struct mptcp_sock *msk = mptcp_sk(sk);
3030 /* We are on the fastopen error path. We can't call straight into the
3031 * subflows cleanup code due to lock nesting (we are already under
3032 * msk->firstsocket lock). Do nothing and leave the cleanup to the
3035 if (msk->fastopening)
3038 mptcp_listen_inuse_dec(sk);
3039 inet_sk_state_store(sk, TCP_CLOSE);
3041 mptcp_stop_timer(sk);
3042 sk_stop_timer(sk, &sk->sk_timer);
3045 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3047 /* msk->subflow is still intact, the following will not free the first
3050 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3051 msk->last_snd = NULL;
3052 WRITE_ONCE(msk->flags, 0);
3054 msk->push_pending = 0;
3055 msk->recovery = false;
3056 msk->can_ack = false;
3057 msk->fully_established = false;
3058 msk->rcv_data_fin = false;
3059 msk->snd_data_fin_enable = false;
3060 msk->rcv_fastclose = false;
3061 msk->use_64bit_ack = false;
3062 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3063 mptcp_pm_data_reset(msk);
3066 sk->sk_shutdown = 0;
3067 sk_error_report(sk);
3071 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3072 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3074 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3076 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3080 struct sock *mptcp_sk_clone(const struct sock *sk,
3081 const struct mptcp_options_received *mp_opt,
3082 struct request_sock *req)
3084 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3085 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3086 struct mptcp_sock *msk;
3091 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3092 if (nsk->sk_family == AF_INET6)
3093 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3096 __mptcp_init_sock(nsk);
3098 msk = mptcp_sk(nsk);
3099 msk->local_key = subflow_req->local_key;
3100 msk->token = subflow_req->token;
3101 msk->subflow = NULL;
3102 msk->in_accept_queue = 1;
3103 WRITE_ONCE(msk->fully_established, false);
3104 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3105 WRITE_ONCE(msk->csum_enabled, true);
3107 msk->write_seq = subflow_req->idsn + 1;
3108 msk->snd_nxt = msk->write_seq;
3109 msk->snd_una = msk->write_seq;
3110 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3111 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3113 sock_reset_flag(nsk, SOCK_RCU_FREE);
3114 /* will be fully established after successful MPC subflow creation */
3115 inet_sk_state_store(nsk, TCP_SYN_RECV);
3117 security_inet_csk_clone(nsk, req);
3118 bh_unlock_sock(nsk);
3120 /* note: the newly allocated socket refcount is 2 now */
3124 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3126 const struct tcp_sock *tp = tcp_sk(ssk);
3128 msk->rcvq_space.copied = 0;
3129 msk->rcvq_space.rtt_us = 0;
3131 msk->rcvq_space.time = tp->tcp_mstamp;
3133 /* initial rcv_space offering made to peer */
3134 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3135 TCP_INIT_CWND * tp->advmss);
3136 if (msk->rcvq_space.space == 0)
3137 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3139 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3142 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3145 struct mptcp_sock *msk = mptcp_sk(sk);
3146 struct socket *listener;
3149 listener = __mptcp_nmpc_socket(msk);
3150 if (WARN_ON_ONCE(!listener)) {
3155 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3156 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3160 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3161 if (sk_is_mptcp(newsk)) {
3162 struct mptcp_subflow_context *subflow;
3163 struct sock *new_mptcp_sock;
3165 subflow = mptcp_subflow_ctx(newsk);
3166 new_mptcp_sock = subflow->conn;
3168 /* is_mptcp should be false if subflow->conn is missing, see
3169 * subflow_syn_recv_sock()
3171 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3172 tcp_sk(newsk)->is_mptcp = 0;
3176 newsk = new_mptcp_sock;
3177 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3179 MPTCP_INC_STATS(sock_net(sk),
3180 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3184 newsk->sk_kern_sock = kern;
3188 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3190 struct mptcp_subflow_context *subflow, *tmp;
3191 struct sock *sk = (struct sock *)msk;
3193 __mptcp_clear_xmit(sk);
3195 /* join list will be eventually flushed (with rst) at sock lock release time */
3196 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3197 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3199 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3200 mptcp_data_lock(sk);
3201 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3202 __skb_queue_purge(&sk->sk_receive_queue);
3203 skb_rbtree_purge(&msk->out_of_order_queue);
3204 mptcp_data_unlock(sk);
3206 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3207 * inet_sock_destruct() will dispose it
3209 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3210 msk->rmem_fwd_alloc = 0;
3211 mptcp_token_destroy(msk);
3212 mptcp_pm_free_anno_list(msk);
3213 mptcp_free_local_addr_list(msk);
3216 static void mptcp_destroy(struct sock *sk)
3218 struct mptcp_sock *msk = mptcp_sk(sk);
3220 /* clears msk->subflow, allowing the following to close
3221 * even the initial subflow
3223 mptcp_dispose_initial_subflow(msk);
3224 mptcp_destroy_common(msk, 0);
3225 sk_sockets_allocated_dec(sk);
3228 void __mptcp_data_acked(struct sock *sk)
3230 if (!sock_owned_by_user(sk))
3231 __mptcp_clean_una(sk);
3233 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3235 if (mptcp_pending_data_fin_ack(sk))
3236 mptcp_schedule_work(sk);
3239 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3241 if (!mptcp_send_head(sk))
3244 if (!sock_owned_by_user(sk))
3245 __mptcp_subflow_push_pending(sk, ssk, false);
3247 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3250 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3251 BIT(MPTCP_RETRANSMIT) | \
3252 BIT(MPTCP_FLUSH_JOIN_LIST))
3254 /* processes deferred events and flush wmem */
3255 static void mptcp_release_cb(struct sock *sk)
3256 __must_hold(&sk->sk_lock.slock)
3258 struct mptcp_sock *msk = mptcp_sk(sk);
3261 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3266 /* the following actions acquire the subflow socket lock
3268 * 1) can't be invoked in atomic scope
3269 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3270 * datapath acquires the msk socket spinlock while helding
3271 * the subflow socket lock
3273 msk->push_pending = 0;
3274 msk->cb_flags &= ~flags;
3275 spin_unlock_bh(&sk->sk_lock.slock);
3276 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3277 __mptcp_flush_join_list(sk);
3278 if (flags & BIT(MPTCP_PUSH_PENDING))
3279 __mptcp_push_pending(sk, 0);
3280 if (flags & BIT(MPTCP_RETRANSMIT))
3281 __mptcp_retrans(sk);
3284 spin_lock_bh(&sk->sk_lock.slock);
3287 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3288 __mptcp_clean_una_wakeup(sk);
3289 if (unlikely(&msk->cb_flags)) {
3290 /* be sure to set the current sk state before tacking actions
3291 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3293 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3294 __mptcp_set_connected(sk);
3295 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3296 __mptcp_error_report(sk);
3297 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3298 msk->last_snd = NULL;
3301 __mptcp_update_rmem(sk);
3304 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3305 * TCP can't schedule delack timer before the subflow is fully established.
3306 * MPTCP uses the delack timer to do 3rd ack retransmissions
3308 static void schedule_3rdack_retransmission(struct sock *ssk)
3310 struct inet_connection_sock *icsk = inet_csk(ssk);
3311 struct tcp_sock *tp = tcp_sk(ssk);
3312 unsigned long timeout;
3314 if (mptcp_subflow_ctx(ssk)->fully_established)
3317 /* reschedule with a timeout above RTT, as we must look only for drop */
3319 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3321 timeout = TCP_TIMEOUT_INIT;
3324 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3325 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3326 icsk->icsk_ack.timeout = timeout;
3327 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3330 void mptcp_subflow_process_delegated(struct sock *ssk)
3332 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3333 struct sock *sk = subflow->conn;
3335 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3336 mptcp_data_lock(sk);
3337 if (!sock_owned_by_user(sk))
3338 __mptcp_subflow_push_pending(sk, ssk, true);
3340 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3341 mptcp_data_unlock(sk);
3342 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3344 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3345 schedule_3rdack_retransmission(ssk);
3346 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3350 static int mptcp_hash(struct sock *sk)
3352 /* should never be called,
3353 * we hash the TCP subflows not the master socket
3359 static void mptcp_unhash(struct sock *sk)
3361 /* called from sk_common_release(), but nothing to do here */
3364 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3366 struct mptcp_sock *msk = mptcp_sk(sk);
3367 struct socket *ssock;
3369 ssock = __mptcp_nmpc_socket(msk);
3370 pr_debug("msk=%p, subflow=%p", msk, ssock);
3371 if (WARN_ON_ONCE(!ssock))
3374 return inet_csk_get_port(ssock->sk, snum);
3377 void mptcp_finish_connect(struct sock *ssk)
3379 struct mptcp_subflow_context *subflow;
3380 struct mptcp_sock *msk;
3383 subflow = mptcp_subflow_ctx(ssk);
3387 pr_debug("msk=%p, token=%u", sk, subflow->token);
3389 subflow->map_seq = subflow->iasn;
3390 subflow->map_subflow_seq = 1;
3392 /* the socket is not connected yet, no msk/subflow ops can access/race
3393 * accessing the field below
3395 WRITE_ONCE(msk->local_key, subflow->local_key);
3396 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3397 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3398 WRITE_ONCE(msk->snd_una, msk->write_seq);
3400 mptcp_pm_new_connection(msk, ssk, 0);
3402 mptcp_rcv_space_init(msk, ssk);
3405 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3407 write_lock_bh(&sk->sk_callback_lock);
3408 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3409 sk_set_socket(sk, parent);
3410 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3411 write_unlock_bh(&sk->sk_callback_lock);
3414 bool mptcp_finish_join(struct sock *ssk)
3416 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3417 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3418 struct sock *parent = (void *)msk;
3421 pr_debug("msk=%p, subflow=%p", msk, subflow);
3423 /* mptcp socket already closing? */
3424 if (!mptcp_is_fully_established(parent)) {
3425 subflow->reset_reason = MPTCP_RST_EMPTCP;
3429 if (!list_empty(&subflow->node))
3432 if (!mptcp_pm_allow_new_subflow(msk))
3433 goto err_prohibited;
3435 /* active connections are already on conn_list.
3436 * If we can't acquire msk socket lock here, let the release callback
3439 mptcp_data_lock(parent);
3440 if (!sock_owned_by_user(parent)) {
3441 ret = __mptcp_finish_join(msk, ssk);
3444 list_add_tail(&subflow->node, &msk->conn_list);
3448 list_add_tail(&subflow->node, &msk->join_list);
3449 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3451 mptcp_data_unlock(parent);
3455 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3459 subflow->map_seq = READ_ONCE(msk->ack_seq);
3460 WRITE_ONCE(msk->allow_infinite_fallback, false);
3463 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3467 static void mptcp_shutdown(struct sock *sk, int how)
3469 pr_debug("sk=%p, how=%d", sk, how);
3471 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3472 __mptcp_wr_shutdown(sk);
3475 static int mptcp_forward_alloc_get(const struct sock *sk)
3477 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3480 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3482 const struct sock *sk = (void *)msk;
3485 if (sk->sk_state == TCP_LISTEN)
3488 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3491 delta = msk->write_seq - v;
3492 if (__mptcp_check_fallback(msk) && msk->first) {
3493 struct tcp_sock *tp = tcp_sk(msk->first);
3495 /* the first subflow is disconnected after close - see
3496 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3497 * so ignore that status, too.
3499 if (!((1 << msk->first->sk_state) &
3500 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3501 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3503 if (delta > INT_MAX)
3509 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3511 struct mptcp_sock *msk = mptcp_sk(sk);
3517 if (sk->sk_state == TCP_LISTEN)
3521 __mptcp_move_skbs(msk);
3522 answ = mptcp_inq_hint(sk);
3526 slow = lock_sock_fast(sk);
3527 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3528 unlock_sock_fast(sk, slow);
3531 slow = lock_sock_fast(sk);
3532 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3533 unlock_sock_fast(sk, slow);
3536 return -ENOIOCTLCMD;
3539 return put_user(answ, (int __user *)arg);
3542 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3543 struct mptcp_subflow_context *subflow)
3545 subflow->request_mptcp = 0;
3546 __mptcp_do_fallback(msk);
3549 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3551 struct mptcp_subflow_context *subflow;
3552 struct mptcp_sock *msk = mptcp_sk(sk);
3553 struct socket *ssock;
3556 ssock = __mptcp_nmpc_socket(msk);
3560 mptcp_token_destroy(msk);
3561 inet_sk_state_store(sk, TCP_SYN_SENT);
3562 subflow = mptcp_subflow_ctx(ssock->sk);
3563 #ifdef CONFIG_TCP_MD5SIG
3564 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3567 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3568 mptcp_subflow_early_fallback(msk, subflow);
3570 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3571 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3572 mptcp_subflow_early_fallback(msk, subflow);
3574 if (likely(!__mptcp_check_fallback(msk)))
3575 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3577 /* if reaching here via the fastopen/sendmsg path, the caller already
3578 * acquired the subflow socket lock, too.
3580 if (msk->fastopening)
3581 err = __inet_stream_connect(ssock, uaddr, addr_len, msk->connect_flags, 1);
3583 err = inet_stream_connect(ssock, uaddr, addr_len, msk->connect_flags);
3584 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3586 /* on successful connect, the msk state will be moved to established by
3587 * subflow_finish_connect()
3589 if (unlikely(err && err != -EINPROGRESS)) {
3590 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3594 mptcp_copy_inaddrs(sk, ssock->sk);
3596 /* unblocking connect, mptcp-level inet_stream_connect will error out
3597 * without changing the socket state, update it here.
3599 if (err == -EINPROGRESS)
3600 sk->sk_socket->state = ssock->state;
3604 static struct proto mptcp_prot = {
3606 .owner = THIS_MODULE,
3607 .init = mptcp_init_sock,
3608 .connect = mptcp_connect,
3609 .disconnect = mptcp_disconnect,
3610 .close = mptcp_close,
3611 .accept = mptcp_accept,
3612 .setsockopt = mptcp_setsockopt,
3613 .getsockopt = mptcp_getsockopt,
3614 .shutdown = mptcp_shutdown,
3615 .destroy = mptcp_destroy,
3616 .sendmsg = mptcp_sendmsg,
3617 .ioctl = mptcp_ioctl,
3618 .recvmsg = mptcp_recvmsg,
3619 .release_cb = mptcp_release_cb,
3621 .unhash = mptcp_unhash,
3622 .get_port = mptcp_get_port,
3623 .forward_alloc_get = mptcp_forward_alloc_get,
3624 .sockets_allocated = &mptcp_sockets_allocated,
3626 .memory_allocated = &tcp_memory_allocated,
3627 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3629 .memory_pressure = &tcp_memory_pressure,
3630 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3631 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3632 .sysctl_mem = sysctl_tcp_mem,
3633 .obj_size = sizeof(struct mptcp_sock),
3634 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3635 .no_autobind = true,
3638 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3640 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3641 struct socket *ssock;
3644 lock_sock(sock->sk);
3645 ssock = __mptcp_nmpc_socket(msk);
3651 err = ssock->ops->bind(ssock, uaddr, addr_len);
3653 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3656 release_sock(sock->sk);
3660 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3661 int addr_len, int flags)
3665 lock_sock(sock->sk);
3666 mptcp_sk(sock->sk)->connect_flags = flags;
3667 ret = __inet_stream_connect(sock, uaddr, addr_len, flags, 0);
3668 release_sock(sock->sk);
3672 static int mptcp_listen(struct socket *sock, int backlog)
3674 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3675 struct sock *sk = sock->sk;
3676 struct socket *ssock;
3679 pr_debug("msk=%p", msk);
3682 ssock = __mptcp_nmpc_socket(msk);
3688 mptcp_token_destroy(msk);
3689 inet_sk_state_store(sk, TCP_LISTEN);
3690 sock_set_flag(sk, SOCK_RCU_FREE);
3692 err = ssock->ops->listen(ssock, backlog);
3693 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3695 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3696 mptcp_copy_inaddrs(sk, ssock->sk);
3699 mptcp_event_pm_listener(ssock->sk, MPTCP_EVENT_LISTENER_CREATED);
3706 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3707 int flags, bool kern)
3709 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3710 struct socket *ssock;
3713 pr_debug("msk=%p", msk);
3715 ssock = __mptcp_nmpc_socket(msk);
3719 err = ssock->ops->accept(sock, newsock, flags, kern);
3720 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3721 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3722 struct mptcp_subflow_context *subflow;
3723 struct sock *newsk = newsock->sk;
3725 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3726 msk->in_accept_queue = 0;
3730 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3731 * This is needed so NOSPACE flag can be set from tcp stack.
3733 mptcp_for_each_subflow(msk, subflow) {
3734 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3736 if (!ssk->sk_socket)
3737 mptcp_sock_graft(ssk, newsock);
3739 release_sock(newsk);
3745 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3747 struct sock *sk = (struct sock *)msk;
3749 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3750 return EPOLLOUT | EPOLLWRNORM;
3752 if (sk_stream_is_writeable(sk))
3753 return EPOLLOUT | EPOLLWRNORM;
3755 mptcp_set_nospace(sk);
3756 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3757 if (sk_stream_is_writeable(sk))
3758 return EPOLLOUT | EPOLLWRNORM;
3763 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3764 struct poll_table_struct *wait)
3766 struct sock *sk = sock->sk;
3767 struct mptcp_sock *msk;
3772 sock_poll_wait(file, sock, wait);
3774 state = inet_sk_state_load(sk);
3775 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3776 if (state == TCP_LISTEN) {
3777 if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk))
3780 return inet_csk_listen_poll(msk->subflow->sk);
3783 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3784 mask |= mptcp_check_readable(msk);
3785 mask |= mptcp_check_writeable(msk);
3786 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3787 /* cf tcp_poll() note about TFO */
3788 mask |= EPOLLOUT | EPOLLWRNORM;
3790 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3792 if (sk->sk_shutdown & RCV_SHUTDOWN)
3793 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3795 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3803 static const struct proto_ops mptcp_stream_ops = {
3805 .owner = THIS_MODULE,
3806 .release = inet_release,
3808 .connect = mptcp_stream_connect,
3809 .socketpair = sock_no_socketpair,
3810 .accept = mptcp_stream_accept,
3811 .getname = inet_getname,
3813 .ioctl = inet_ioctl,
3814 .gettstamp = sock_gettstamp,
3815 .listen = mptcp_listen,
3816 .shutdown = inet_shutdown,
3817 .setsockopt = sock_common_setsockopt,
3818 .getsockopt = sock_common_getsockopt,
3819 .sendmsg = inet_sendmsg,
3820 .recvmsg = inet_recvmsg,
3821 .mmap = sock_no_mmap,
3822 .sendpage = inet_sendpage,
3825 static struct inet_protosw mptcp_protosw = {
3826 .type = SOCK_STREAM,
3827 .protocol = IPPROTO_MPTCP,
3828 .prot = &mptcp_prot,
3829 .ops = &mptcp_stream_ops,
3830 .flags = INET_PROTOSW_ICSK,
3833 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3835 struct mptcp_delegated_action *delegated;
3836 struct mptcp_subflow_context *subflow;
3839 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3840 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3841 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3843 bh_lock_sock_nested(ssk);
3844 if (!sock_owned_by_user(ssk) &&
3845 mptcp_subflow_has_delegated_action(subflow))
3846 mptcp_subflow_process_delegated(ssk);
3847 /* ... elsewhere tcp_release_cb_override already processed
3848 * the action or will do at next release_sock().
3849 * In both case must dequeue the subflow here - on the same
3850 * CPU that scheduled it.
3852 bh_unlock_sock(ssk);
3855 if (++work_done == budget)
3859 /* always provide a 0 'work_done' argument, so that napi_complete_done
3860 * will not try accessing the NULL napi->dev ptr
3862 napi_complete_done(napi, 0);
3866 void __init mptcp_proto_init(void)
3868 struct mptcp_delegated_action *delegated;
3871 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3873 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3874 panic("Failed to allocate MPTCP pcpu counter\n");
3876 init_dummy_netdev(&mptcp_napi_dev);
3877 for_each_possible_cpu(cpu) {
3878 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3879 INIT_LIST_HEAD(&delegated->head);
3880 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3882 napi_enable(&delegated->napi);
3885 mptcp_subflow_init();
3889 if (proto_register(&mptcp_prot, 1) != 0)
3890 panic("Failed to register MPTCP proto.\n");
3892 inet_register_protosw(&mptcp_protosw);
3894 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3897 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3898 static const struct proto_ops mptcp_v6_stream_ops = {
3900 .owner = THIS_MODULE,
3901 .release = inet6_release,
3903 .connect = mptcp_stream_connect,
3904 .socketpair = sock_no_socketpair,
3905 .accept = mptcp_stream_accept,
3906 .getname = inet6_getname,
3908 .ioctl = inet6_ioctl,
3909 .gettstamp = sock_gettstamp,
3910 .listen = mptcp_listen,
3911 .shutdown = inet_shutdown,
3912 .setsockopt = sock_common_setsockopt,
3913 .getsockopt = sock_common_getsockopt,
3914 .sendmsg = inet6_sendmsg,
3915 .recvmsg = inet6_recvmsg,
3916 .mmap = sock_no_mmap,
3917 .sendpage = inet_sendpage,
3918 #ifdef CONFIG_COMPAT
3919 .compat_ioctl = inet6_compat_ioctl,
3923 static struct proto mptcp_v6_prot;
3925 static struct inet_protosw mptcp_v6_protosw = {
3926 .type = SOCK_STREAM,
3927 .protocol = IPPROTO_MPTCP,
3928 .prot = &mptcp_v6_prot,
3929 .ops = &mptcp_v6_stream_ops,
3930 .flags = INET_PROTOSW_ICSK,
3933 int __init mptcp_proto_v6_init(void)
3937 mptcp_v6_prot = mptcp_prot;
3938 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3939 mptcp_v6_prot.slab = NULL;
3940 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3942 err = proto_register(&mptcp_v6_prot, 1);
3946 err = inet6_register_protosw(&mptcp_v6_protosw);
3948 proto_unregister(&mptcp_v6_prot);