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>
28 #define CREATE_TRACE_POINTS
29 #include <trace/events/mptcp.h>
31 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
33 struct mptcp_sock msk;
45 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
48 MPTCP_CMSG_TS = BIT(0),
51 static struct percpu_counter mptcp_sockets_allocated;
53 static void __mptcp_destroy_sock(struct sock *sk);
54 static void __mptcp_check_send_data_fin(struct sock *sk);
56 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
57 static struct net_device mptcp_napi_dev;
59 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
60 * completed yet or has failed, return the subflow socket.
61 * Otherwise return NULL.
63 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
65 if (!msk->subflow || READ_ONCE(msk->can_ack))
71 /* Returns end sequence number of the receiver's advertised window */
72 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
74 return READ_ONCE(msk->wnd_end);
77 static bool mptcp_is_tcpsk(struct sock *sk)
79 struct socket *sock = sk->sk_socket;
81 if (unlikely(sk->sk_prot == &tcp_prot)) {
82 /* we are being invoked after mptcp_accept() has
83 * accepted a non-mp-capable flow: sk is a tcp_sk,
86 * Hand the socket over to tcp so all further socket ops
89 sock->ops = &inet_stream_ops;
91 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
92 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
93 sock->ops = &inet6_stream_ops;
101 static int __mptcp_socket_create(struct mptcp_sock *msk)
103 struct mptcp_subflow_context *subflow;
104 struct sock *sk = (struct sock *)msk;
105 struct socket *ssock;
108 err = mptcp_subflow_create_socket(sk, &ssock);
112 msk->first = ssock->sk;
113 msk->subflow = ssock;
114 subflow = mptcp_subflow_ctx(ssock->sk);
115 list_add(&subflow->node, &msk->conn_list);
116 sock_hold(ssock->sk);
117 subflow->request_mptcp = 1;
118 mptcp_sock_graft(msk->first, sk->sk_socket);
123 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
125 sk_drops_add(sk, skb);
129 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
130 struct sk_buff *from)
135 if (MPTCP_SKB_CB(from)->offset ||
136 !skb_try_coalesce(to, from, &fragstolen, &delta))
139 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
140 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
141 to->len, MPTCP_SKB_CB(from)->end_seq);
142 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
143 kfree_skb_partial(from, fragstolen);
144 atomic_add(delta, &sk->sk_rmem_alloc);
145 sk_mem_charge(sk, delta);
149 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
150 struct sk_buff *from)
152 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
155 return mptcp_try_coalesce((struct sock *)msk, to, from);
158 /* "inspired" by tcp_data_queue_ofo(), main differences:
160 * - don't cope with sacks
162 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
164 struct sock *sk = (struct sock *)msk;
165 struct rb_node **p, *parent;
166 u64 seq, end_seq, max_seq;
167 struct sk_buff *skb1;
169 seq = MPTCP_SKB_CB(skb)->map_seq;
170 end_seq = MPTCP_SKB_CB(skb)->end_seq;
171 max_seq = READ_ONCE(msk->rcv_wnd_sent);
173 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
174 RB_EMPTY_ROOT(&msk->out_of_order_queue));
175 if (after64(end_seq, max_seq)) {
178 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
179 (unsigned long long)end_seq - (unsigned long)max_seq,
180 (unsigned long long)msk->rcv_wnd_sent);
181 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
185 p = &msk->out_of_order_queue.rb_node;
186 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
187 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
188 rb_link_node(&skb->rbnode, NULL, p);
189 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
190 msk->ooo_last_skb = skb;
194 /* with 2 subflows, adding at end of ooo queue is quite likely
195 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
197 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
198 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
199 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
203 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
204 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
205 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
206 parent = &msk->ooo_last_skb->rbnode;
207 p = &parent->rb_right;
211 /* Find place to insert this segment. Handle overlaps on the way. */
215 skb1 = rb_to_skb(parent);
216 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
217 p = &parent->rb_left;
220 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
221 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
222 /* All the bits are present. Drop. */
224 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
227 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
231 * continue traversing
234 /* skb's seq == skb1's seq and skb covers skb1.
235 * Replace skb1 with skb.
237 rb_replace_node(&skb1->rbnode, &skb->rbnode,
238 &msk->out_of_order_queue);
239 mptcp_drop(sk, skb1);
240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
243 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
247 p = &parent->rb_right;
251 /* Insert segment into RB tree. */
252 rb_link_node(&skb->rbnode, parent, p);
253 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
256 /* Remove other segments covered by skb. */
257 while ((skb1 = skb_rb_next(skb)) != NULL) {
258 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
260 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
261 mptcp_drop(sk, skb1);
262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
264 /* If there is no skb after us, we are the last_skb ! */
266 msk->ooo_last_skb = skb;
270 skb_set_owner_r(skb, sk);
273 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
274 struct sk_buff *skb, unsigned int offset,
277 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
278 struct sock *sk = (struct sock *)msk;
279 struct sk_buff *tail;
282 __skb_unlink(skb, &ssk->sk_receive_queue);
287 /* try to fetch required memory from subflow */
288 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
289 int amount = sk_mem_pages(skb->truesize) << SK_MEM_QUANTUM_SHIFT;
291 if (ssk->sk_forward_alloc < amount)
294 ssk->sk_forward_alloc -= amount;
295 sk->sk_forward_alloc += amount;
298 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
300 /* the skb map_seq accounts for the skb offset:
301 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
304 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
305 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
306 MPTCP_SKB_CB(skb)->offset = offset;
307 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
309 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
311 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
312 tail = skb_peek_tail(&sk->sk_receive_queue);
313 if (tail && mptcp_try_coalesce(sk, tail, skb))
316 skb_set_owner_r(skb, sk);
317 __skb_queue_tail(&sk->sk_receive_queue, skb);
319 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
320 mptcp_data_queue_ofo(msk, skb);
324 /* old data, keep it simple and drop the whole pkt, sender
325 * will retransmit as needed, if needed.
327 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
333 static void mptcp_stop_timer(struct sock *sk)
335 struct inet_connection_sock *icsk = inet_csk(sk);
337 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
338 mptcp_sk(sk)->timer_ival = 0;
341 static void mptcp_close_wake_up(struct sock *sk)
343 if (sock_flag(sk, SOCK_DEAD))
346 sk->sk_state_change(sk);
347 if (sk->sk_shutdown == SHUTDOWN_MASK ||
348 sk->sk_state == TCP_CLOSE)
349 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
351 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
354 static bool mptcp_pending_data_fin_ack(struct sock *sk)
356 struct mptcp_sock *msk = mptcp_sk(sk);
358 return !__mptcp_check_fallback(msk) &&
359 ((1 << sk->sk_state) &
360 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
361 msk->write_seq == READ_ONCE(msk->snd_una);
364 static void mptcp_check_data_fin_ack(struct sock *sk)
366 struct mptcp_sock *msk = mptcp_sk(sk);
368 /* Look for an acknowledged DATA_FIN */
369 if (mptcp_pending_data_fin_ack(sk)) {
370 WRITE_ONCE(msk->snd_data_fin_enable, 0);
372 switch (sk->sk_state) {
374 inet_sk_state_store(sk, TCP_FIN_WAIT2);
378 inet_sk_state_store(sk, TCP_CLOSE);
382 mptcp_close_wake_up(sk);
386 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
388 struct mptcp_sock *msk = mptcp_sk(sk);
390 if (READ_ONCE(msk->rcv_data_fin) &&
391 ((1 << sk->sk_state) &
392 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
393 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
395 if (msk->ack_seq == rcv_data_fin_seq) {
397 *seq = rcv_data_fin_seq;
406 static void mptcp_set_datafin_timeout(const struct sock *sk)
408 struct inet_connection_sock *icsk = inet_csk(sk);
410 mptcp_sk(sk)->timer_ival = min(TCP_RTO_MAX,
411 TCP_RTO_MIN << icsk->icsk_retransmits);
414 static void __mptcp_set_timeout(struct sock *sk, long tout)
416 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
419 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
421 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
423 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
424 inet_csk(ssk)->icsk_timeout - jiffies : 0;
427 static void mptcp_set_timeout(struct sock *sk)
429 struct mptcp_subflow_context *subflow;
432 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
433 tout = max(tout, mptcp_timeout_from_subflow(subflow));
434 __mptcp_set_timeout(sk, tout);
437 static bool tcp_can_send_ack(const struct sock *ssk)
439 return !((1 << inet_sk_state_load(ssk)) &
440 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
443 void mptcp_subflow_send_ack(struct sock *ssk)
447 slow = lock_sock_fast(ssk);
448 if (tcp_can_send_ack(ssk))
450 unlock_sock_fast(ssk, slow);
453 static void mptcp_send_ack(struct mptcp_sock *msk)
455 struct mptcp_subflow_context *subflow;
457 mptcp_for_each_subflow(msk, subflow)
458 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
461 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
465 slow = lock_sock_fast(ssk);
466 if (tcp_can_send_ack(ssk))
467 tcp_cleanup_rbuf(ssk, 1);
468 unlock_sock_fast(ssk, slow);
471 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
473 const struct inet_connection_sock *icsk = inet_csk(ssk);
474 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
475 const struct tcp_sock *tp = tcp_sk(ssk);
477 return (ack_pending & ICSK_ACK_SCHED) &&
478 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
479 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
480 (rx_empty && ack_pending &
481 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
484 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
486 int old_space = READ_ONCE(msk->old_wspace);
487 struct mptcp_subflow_context *subflow;
488 struct sock *sk = (struct sock *)msk;
489 int space = __mptcp_space(sk);
490 bool cleanup, rx_empty;
492 cleanup = (space > 0) && (space >= (old_space << 1));
493 rx_empty = !__mptcp_rmem(sk);
495 mptcp_for_each_subflow(msk, subflow) {
496 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
498 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
499 mptcp_subflow_cleanup_rbuf(ssk);
503 static bool mptcp_check_data_fin(struct sock *sk)
505 struct mptcp_sock *msk = mptcp_sk(sk);
506 u64 rcv_data_fin_seq;
509 if (__mptcp_check_fallback(msk))
512 /* Need to ack a DATA_FIN received from a peer while this side
513 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
514 * msk->rcv_data_fin was set when parsing the incoming options
515 * at the subflow level and the msk lock was not held, so this
516 * is the first opportunity to act on the DATA_FIN and change
519 * If we are caught up to the sequence number of the incoming
520 * DATA_FIN, send the DATA_ACK now and do state transition. If
521 * not caught up, do nothing and let the recv code send DATA_ACK
525 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
526 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
527 WRITE_ONCE(msk->rcv_data_fin, 0);
529 sk->sk_shutdown |= RCV_SHUTDOWN;
530 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
532 switch (sk->sk_state) {
533 case TCP_ESTABLISHED:
534 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
537 inet_sk_state_store(sk, TCP_CLOSING);
540 inet_sk_state_store(sk, TCP_CLOSE);
543 /* Other states not expected */
550 mptcp_close_wake_up(sk);
555 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
559 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
560 struct sock *sk = (struct sock *)msk;
561 unsigned int moved = 0;
562 bool more_data_avail;
567 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
569 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
570 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
572 if (unlikely(ssk_rbuf > sk_rbuf)) {
573 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
578 pr_debug("msk=%p ssk=%p", msk, ssk);
581 u32 map_remaining, offset;
582 u32 seq = tp->copied_seq;
586 /* try to move as much data as available */
587 map_remaining = subflow->map_data_len -
588 mptcp_subflow_get_map_offset(subflow);
590 skb = skb_peek(&ssk->sk_receive_queue);
592 /* if no data is found, a racing workqueue/recvmsg
593 * already processed the new data, stop here or we
594 * can enter an infinite loop
601 if (__mptcp_check_fallback(msk)) {
602 /* if we are running under the workqueue, TCP could have
603 * collapsed skbs between dummy map creation and now
604 * be sure to adjust the size
606 map_remaining = skb->len;
607 subflow->map_data_len = skb->len;
610 offset = seq - TCP_SKB_CB(skb)->seq;
611 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
617 if (offset < skb->len) {
618 size_t len = skb->len - offset;
623 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
627 if (WARN_ON_ONCE(map_remaining < len))
631 sk_eat_skb(ssk, skb);
635 WRITE_ONCE(tp->copied_seq, seq);
636 more_data_avail = mptcp_subflow_data_available(ssk);
638 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
642 } while (more_data_avail);
648 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
650 struct sock *sk = (struct sock *)msk;
651 struct sk_buff *skb, *tail;
656 p = rb_first(&msk->out_of_order_queue);
657 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
660 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
664 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
666 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
669 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
673 end_seq = MPTCP_SKB_CB(skb)->end_seq;
674 tail = skb_peek_tail(&sk->sk_receive_queue);
675 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
676 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
678 /* skip overlapping data, if any */
679 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
680 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
682 MPTCP_SKB_CB(skb)->offset += delta;
683 __skb_queue_tail(&sk->sk_receive_queue, skb);
685 msk->ack_seq = end_seq;
691 /* In most cases we will be able to lock the mptcp socket. If its already
692 * owned, we need to defer to the work queue to avoid ABBA deadlock.
694 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
696 struct sock *sk = (struct sock *)msk;
697 unsigned int moved = 0;
699 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
700 __mptcp_ofo_queue(msk);
701 if (unlikely(ssk->sk_err)) {
702 if (!sock_owned_by_user(sk))
703 __mptcp_error_report(sk);
705 set_bit(MPTCP_ERROR_REPORT, &msk->flags);
708 /* If the moves have caught up with the DATA_FIN sequence number
709 * it's time to ack the DATA_FIN and change socket state, but
710 * this is not a good place to change state. Let the workqueue
713 if (mptcp_pending_data_fin(sk, NULL))
714 mptcp_schedule_work(sk);
718 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
720 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
721 struct mptcp_sock *msk = mptcp_sk(sk);
722 int sk_rbuf, ssk_rbuf;
724 /* The peer can send data while we are shutting down this
725 * subflow at msk destruction time, but we must avoid enqueuing
726 * more data to the msk receive queue
728 if (unlikely(subflow->disposable))
731 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
732 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
733 if (unlikely(ssk_rbuf > sk_rbuf))
736 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
737 if (__mptcp_rmem(sk) > sk_rbuf) {
738 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
742 /* Wake-up the reader only for in-sequence data */
744 if (move_skbs_to_msk(msk, ssk))
745 sk->sk_data_ready(sk);
747 mptcp_data_unlock(sk);
750 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
752 struct mptcp_subflow_context *subflow;
755 if (likely(list_empty(&msk->join_list)))
758 spin_lock_bh(&msk->join_list_lock);
759 list_for_each_entry(subflow, &msk->join_list, node) {
760 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
762 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
763 if (READ_ONCE(msk->setsockopt_seq) != sseq)
766 list_splice_tail_init(&msk->join_list, &msk->conn_list);
767 spin_unlock_bh(&msk->join_list_lock);
772 void __mptcp_flush_join_list(struct mptcp_sock *msk)
774 if (likely(!mptcp_do_flush_join_list(msk)))
777 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
778 mptcp_schedule_work((struct sock *)msk);
781 static void mptcp_flush_join_list(struct mptcp_sock *msk)
783 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
787 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
790 mptcp_sockopt_sync_all(msk);
793 static bool mptcp_timer_pending(struct sock *sk)
795 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
798 static void mptcp_reset_timer(struct sock *sk)
800 struct inet_connection_sock *icsk = inet_csk(sk);
803 /* prevent rescheduling on close */
804 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
807 tout = mptcp_sk(sk)->timer_ival;
808 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
811 bool mptcp_schedule_work(struct sock *sk)
813 if (inet_sk_state_load(sk) != TCP_CLOSE &&
814 schedule_work(&mptcp_sk(sk)->work)) {
815 /* each subflow already holds a reference to the sk, and the
816 * workqueue is invoked by a subflow, so sk can't go away here.
824 void mptcp_subflow_eof(struct sock *sk)
826 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
827 mptcp_schedule_work(sk);
830 static void mptcp_check_for_eof(struct mptcp_sock *msk)
832 struct mptcp_subflow_context *subflow;
833 struct sock *sk = (struct sock *)msk;
836 mptcp_for_each_subflow(msk, subflow)
837 receivers += !subflow->rx_eof;
841 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
842 /* hopefully temporary hack: propagate shutdown status
843 * to msk, when all subflows agree on it
845 sk->sk_shutdown |= RCV_SHUTDOWN;
847 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
848 sk->sk_data_ready(sk);
851 switch (sk->sk_state) {
852 case TCP_ESTABLISHED:
853 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
856 inet_sk_state_store(sk, TCP_CLOSING);
859 inet_sk_state_store(sk, TCP_CLOSE);
864 mptcp_close_wake_up(sk);
867 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
869 struct mptcp_subflow_context *subflow;
870 struct sock *sk = (struct sock *)msk;
872 sock_owned_by_me(sk);
874 mptcp_for_each_subflow(msk, subflow) {
875 if (READ_ONCE(subflow->data_avail))
876 return mptcp_subflow_tcp_sock(subflow);
882 static bool mptcp_skb_can_collapse_to(u64 write_seq,
883 const struct sk_buff *skb,
884 const struct mptcp_ext *mpext)
886 if (!tcp_skb_can_collapse_to(skb))
889 /* can collapse only if MPTCP level sequence is in order and this
890 * mapping has not been xmitted yet
892 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
896 /* we can append data to the given data frag if:
897 * - there is space available in the backing page_frag
898 * - the data frag tail matches the current page_frag free offset
899 * - the data frag end sequence number matches the current write seq
901 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
902 const struct page_frag *pfrag,
903 const struct mptcp_data_frag *df)
905 return df && pfrag->page == df->page &&
906 pfrag->size - pfrag->offset > 0 &&
907 pfrag->offset == (df->offset + df->data_len) &&
908 df->data_seq + df->data_len == msk->write_seq;
911 static int mptcp_wmem_with_overhead(int size)
913 return size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
916 static void __mptcp_wmem_reserve(struct sock *sk, int size)
918 int amount = mptcp_wmem_with_overhead(size);
919 struct mptcp_sock *msk = mptcp_sk(sk);
921 WARN_ON_ONCE(msk->wmem_reserved);
922 if (WARN_ON_ONCE(amount < 0))
925 if (amount <= sk->sk_forward_alloc)
928 /* under memory pressure try to reserve at most a single page
929 * otherwise try to reserve the full estimate and fallback
930 * to a single page before entering the error path
932 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
933 !sk_wmem_schedule(sk, amount)) {
934 if (amount <= PAGE_SIZE)
938 if (!sk_wmem_schedule(sk, amount))
943 msk->wmem_reserved = amount;
944 sk->sk_forward_alloc -= amount;
948 /* we will wait for memory on next allocation */
949 msk->wmem_reserved = -1;
952 static void __mptcp_update_wmem(struct sock *sk)
954 struct mptcp_sock *msk = mptcp_sk(sk);
956 #ifdef CONFIG_LOCKDEP
957 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
960 if (!msk->wmem_reserved)
963 if (msk->wmem_reserved < 0)
964 msk->wmem_reserved = 0;
965 if (msk->wmem_reserved > 0) {
966 sk->sk_forward_alloc += msk->wmem_reserved;
967 msk->wmem_reserved = 0;
971 static bool mptcp_wmem_alloc(struct sock *sk, int size)
973 struct mptcp_sock *msk = mptcp_sk(sk);
975 /* check for pre-existing error condition */
976 if (msk->wmem_reserved < 0)
979 if (msk->wmem_reserved >= size)
983 if (!sk_wmem_schedule(sk, size)) {
984 mptcp_data_unlock(sk);
988 sk->sk_forward_alloc -= size;
989 msk->wmem_reserved += size;
990 mptcp_data_unlock(sk);
993 msk->wmem_reserved -= size;
997 static void mptcp_wmem_uncharge(struct sock *sk, int size)
999 struct mptcp_sock *msk = mptcp_sk(sk);
1001 if (msk->wmem_reserved < 0)
1002 msk->wmem_reserved = 0;
1003 msk->wmem_reserved += size;
1006 static void __mptcp_mem_reclaim_partial(struct sock *sk)
1008 lockdep_assert_held_once(&sk->sk_lock.slock);
1009 __mptcp_update_wmem(sk);
1010 sk_mem_reclaim_partial(sk);
1013 static void mptcp_mem_reclaim_partial(struct sock *sk)
1015 struct mptcp_sock *msk = mptcp_sk(sk);
1017 /* if we are experiencing a transint allocation error,
1018 * the forward allocation memory has been already
1021 if (msk->wmem_reserved < 0)
1024 mptcp_data_lock(sk);
1025 sk->sk_forward_alloc += msk->wmem_reserved;
1026 sk_mem_reclaim_partial(sk);
1027 msk->wmem_reserved = sk->sk_forward_alloc;
1028 sk->sk_forward_alloc = 0;
1029 mptcp_data_unlock(sk);
1032 static void dfrag_uncharge(struct sock *sk, int len)
1034 sk_mem_uncharge(sk, len);
1035 sk_wmem_queued_add(sk, -len);
1038 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1040 int len = dfrag->data_len + dfrag->overhead;
1042 list_del(&dfrag->list);
1043 dfrag_uncharge(sk, len);
1044 put_page(dfrag->page);
1047 static void __mptcp_clean_una(struct sock *sk)
1049 struct mptcp_sock *msk = mptcp_sk(sk);
1050 struct mptcp_data_frag *dtmp, *dfrag;
1051 bool cleaned = false;
1054 /* on fallback we just need to ignore snd_una, as this is really
1057 if (__mptcp_check_fallback(msk))
1058 msk->snd_una = READ_ONCE(msk->snd_nxt);
1060 snd_una = msk->snd_una;
1061 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1062 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1065 if (unlikely(dfrag == msk->first_pending)) {
1066 /* in recovery mode can see ack after the current snd head */
1067 if (WARN_ON_ONCE(!msk->recovery))
1070 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1073 dfrag_clear(sk, dfrag);
1077 dfrag = mptcp_rtx_head(sk);
1078 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1079 u64 delta = snd_una - dfrag->data_seq;
1081 /* prevent wrap around in recovery mode */
1082 if (unlikely(delta > dfrag->already_sent)) {
1083 if (WARN_ON_ONCE(!msk->recovery))
1085 if (WARN_ON_ONCE(delta > dfrag->data_len))
1087 dfrag->already_sent += delta - dfrag->already_sent;
1090 dfrag->data_seq += delta;
1091 dfrag->offset += delta;
1092 dfrag->data_len -= delta;
1093 dfrag->already_sent -= delta;
1095 dfrag_uncharge(sk, delta);
1099 /* all retransmitted data acked, recovery completed */
1100 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1101 msk->recovery = false;
1104 if (cleaned && tcp_under_memory_pressure(sk))
1105 __mptcp_mem_reclaim_partial(sk);
1107 if (snd_una == READ_ONCE(msk->snd_nxt) && !msk->recovery) {
1108 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1109 mptcp_stop_timer(sk);
1111 mptcp_reset_timer(sk);
1115 static void __mptcp_clean_una_wakeup(struct sock *sk)
1117 #ifdef CONFIG_LOCKDEP
1118 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
1120 __mptcp_clean_una(sk);
1121 mptcp_write_space(sk);
1124 static void mptcp_clean_una_wakeup(struct sock *sk)
1126 mptcp_data_lock(sk);
1127 __mptcp_clean_una_wakeup(sk);
1128 mptcp_data_unlock(sk);
1131 static void mptcp_enter_memory_pressure(struct sock *sk)
1133 struct mptcp_subflow_context *subflow;
1134 struct mptcp_sock *msk = mptcp_sk(sk);
1137 sk_stream_moderate_sndbuf(sk);
1138 mptcp_for_each_subflow(msk, subflow) {
1139 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1142 tcp_enter_memory_pressure(ssk);
1143 sk_stream_moderate_sndbuf(ssk);
1148 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1151 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1153 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1154 pfrag, sk->sk_allocation)))
1157 mptcp_enter_memory_pressure(sk);
1161 static struct mptcp_data_frag *
1162 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1165 int offset = ALIGN(orig_offset, sizeof(long));
1166 struct mptcp_data_frag *dfrag;
1168 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1169 dfrag->data_len = 0;
1170 dfrag->data_seq = msk->write_seq;
1171 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1172 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1173 dfrag->already_sent = 0;
1174 dfrag->page = pfrag->page;
1179 struct mptcp_sendmsg_info {
1185 bool data_lock_held;
1188 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1191 u64 window_end = mptcp_wnd_end(msk);
1193 if (__mptcp_check_fallback(msk))
1196 if (!before64(data_seq + avail_size, window_end)) {
1197 u64 allowed_size = window_end - data_seq;
1199 return min_t(unsigned int, allowed_size, avail_size);
1205 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1207 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1211 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1215 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1217 struct sk_buff *skb;
1219 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1221 if (likely(__mptcp_add_ext(skb, gfp))) {
1222 skb_reserve(skb, MAX_TCP_HEADER);
1223 skb->reserved_tailroom = skb->end - skb->tail;
1228 mptcp_enter_memory_pressure(sk);
1233 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1235 struct sk_buff *skb;
1237 if (ssk->sk_tx_skb_cache) {
1238 skb = ssk->sk_tx_skb_cache;
1239 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1240 !__mptcp_add_ext(skb, gfp)))
1245 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1249 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1250 ssk->sk_tx_skb_cache = skb;
1257 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1259 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1261 if (unlikely(tcp_under_memory_pressure(sk))) {
1263 __mptcp_mem_reclaim_partial(sk);
1265 mptcp_mem_reclaim_partial(sk);
1267 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1270 /* note: this always recompute the csum on the whole skb, even
1271 * if we just appended a single frag. More status info needed
1273 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1275 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1276 __wsum csum = ~csum_unfold(mpext->csum);
1277 int offset = skb->len - added;
1279 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1282 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1283 struct mptcp_data_frag *dfrag,
1284 struct mptcp_sendmsg_info *info)
1286 u64 data_seq = dfrag->data_seq + info->sent;
1287 struct mptcp_sock *msk = mptcp_sk(sk);
1288 bool zero_window_probe = false;
1289 struct mptcp_ext *mpext = NULL;
1290 struct sk_buff *skb, *tail;
1291 bool must_collapse = false;
1296 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1297 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1299 /* compute send limit */
1300 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1301 avail_size = info->size_goal;
1302 skb = tcp_write_queue_tail(ssk);
1304 /* Limit the write to the size available in the
1305 * current skb, if any, so that we create at most a new skb.
1306 * Explicitly tells TCP internals to avoid collapsing on later
1307 * queue management operation, to avoid breaking the ext <->
1308 * SSN association set here
1310 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1311 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1312 TCP_SKB_CB(skb)->eor = 1;
1316 must_collapse = (info->size_goal > skb->len) &&
1317 (skb_shinfo(skb)->nr_frags < sysctl_max_skb_frags);
1318 if (must_collapse) {
1319 size_bias = skb->len;
1320 avail_size = info->size_goal - skb->len;
1325 if (!must_collapse &&
1326 !mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held))
1329 /* Zero window and all data acked? Probe. */
1330 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1331 if (avail_size == 0) {
1332 u64 snd_una = READ_ONCE(msk->snd_una);
1334 if (skb || snd_una != msk->snd_nxt)
1336 zero_window_probe = true;
1337 data_seq = snd_una - 1;
1341 if (WARN_ON_ONCE(info->sent > info->limit ||
1342 info->limit > dfrag->data_len))
1345 ret = info->limit - info->sent;
1346 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1347 dfrag->page, dfrag->offset + info->sent, &ret);
1349 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1353 /* if the tail skb is still the cached one, collapsing really happened.
1356 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1357 mpext->data_len += ret;
1358 WARN_ON_ONCE(zero_window_probe);
1362 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1363 if (WARN_ON_ONCE(!mpext)) {
1364 /* should never reach here, stream corrupted */
1368 memset(mpext, 0, sizeof(*mpext));
1369 mpext->data_seq = data_seq;
1370 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1371 mpext->data_len = ret;
1375 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1376 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1379 if (zero_window_probe) {
1380 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1382 if (READ_ONCE(msk->csum_enabled))
1383 mptcp_update_data_checksum(tail, ret);
1384 tcp_push_pending_frames(ssk);
1388 if (READ_ONCE(msk->csum_enabled))
1389 mptcp_update_data_checksum(tail, ret);
1390 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1394 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1395 sizeof(struct tcphdr) - \
1396 MAX_TCP_OPTION_SPACE - \
1397 sizeof(struct ipv6hdr) - \
1398 sizeof(struct frag_hdr))
1400 struct subflow_send_info {
1405 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1407 if (!subflow->stale)
1411 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1414 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1416 if (unlikely(subflow->stale)) {
1417 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1419 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1422 mptcp_subflow_set_active(subflow);
1424 return __mptcp_subflow_active(subflow);
1427 /* implement the mptcp packet scheduler;
1428 * returns the subflow that will transmit the next DSS
1429 * additionally updates the rtx timeout
1431 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1433 struct subflow_send_info send_info[2];
1434 struct mptcp_subflow_context *subflow;
1435 struct sock *sk = (struct sock *)msk;
1436 int i, nr_active = 0;
1442 sock_owned_by_me(sk);
1444 if (__mptcp_check_fallback(msk)) {
1447 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1450 /* re-use last subflow, if the burst allow that */
1451 if (msk->last_snd && msk->snd_burst > 0 &&
1452 sk_stream_memory_free(msk->last_snd) &&
1453 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1454 mptcp_set_timeout(sk);
1455 return msk->last_snd;
1458 /* pick the subflow with the lower wmem/wspace ratio */
1459 for (i = 0; i < 2; ++i) {
1460 send_info[i].ssk = NULL;
1461 send_info[i].ratio = -1;
1463 mptcp_for_each_subflow(msk, subflow) {
1464 trace_mptcp_subflow_get_send(subflow);
1465 ssk = mptcp_subflow_tcp_sock(subflow);
1466 if (!mptcp_subflow_active(subflow))
1469 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1470 nr_active += !subflow->backup;
1471 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1474 pace = READ_ONCE(ssk->sk_pacing_rate);
1478 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1480 if (ratio < send_info[subflow->backup].ratio) {
1481 send_info[subflow->backup].ssk = ssk;
1482 send_info[subflow->backup].ratio = ratio;
1485 __mptcp_set_timeout(sk, tout);
1487 /* pick the best backup if no other subflow is active */
1489 send_info[0].ssk = send_info[1].ssk;
1491 if (send_info[0].ssk) {
1492 msk->last_snd = send_info[0].ssk;
1493 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1494 tcp_sk(msk->last_snd)->snd_wnd);
1495 return msk->last_snd;
1501 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1502 struct mptcp_sendmsg_info *info)
1504 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1508 static void mptcp_update_post_push(struct mptcp_sock *msk,
1509 struct mptcp_data_frag *dfrag,
1512 u64 snd_nxt_new = dfrag->data_seq;
1514 dfrag->already_sent += sent;
1516 msk->snd_burst -= sent;
1517 msk->tx_pending_data -= sent;
1519 snd_nxt_new += dfrag->already_sent;
1521 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1522 * is recovering after a failover. In that event, this re-sends
1525 * Thus compute snd_nxt_new candidate based on
1526 * the dfrag->data_seq that was sent and the data
1527 * that has been handed to the subflow for transmission
1528 * and skip update in case it was old dfrag.
1530 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1531 msk->snd_nxt = snd_nxt_new;
1534 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1536 struct sock *prev_ssk = NULL, *ssk = NULL;
1537 struct mptcp_sock *msk = mptcp_sk(sk);
1538 struct mptcp_sendmsg_info info = {
1541 struct mptcp_data_frag *dfrag;
1542 int len, copied = 0;
1544 while ((dfrag = mptcp_send_head(sk))) {
1545 info.sent = dfrag->already_sent;
1546 info.limit = dfrag->data_len;
1547 len = dfrag->data_len - dfrag->already_sent;
1552 __mptcp_flush_join_list(msk);
1553 ssk = mptcp_subflow_get_send(msk);
1555 /* First check. If the ssk has changed since
1556 * the last round, release prev_ssk
1558 if (ssk != prev_ssk && prev_ssk)
1559 mptcp_push_release(sk, prev_ssk, &info);
1563 /* Need to lock the new subflow only if different
1564 * from the previous one, otherwise we are still
1565 * helding the relevant lock
1567 if (ssk != prev_ssk)
1570 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1572 mptcp_push_release(sk, ssk, &info);
1580 mptcp_update_post_push(msk, dfrag, ret);
1582 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1585 /* at this point we held the socket lock for the last subflow we used */
1587 mptcp_push_release(sk, ssk, &info);
1590 /* ensure the rtx timer is running */
1591 if (!mptcp_timer_pending(sk))
1592 mptcp_reset_timer(sk);
1594 __mptcp_check_send_data_fin(sk);
1597 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1599 struct mptcp_sock *msk = mptcp_sk(sk);
1600 struct mptcp_sendmsg_info info = {
1601 .data_lock_held = true,
1603 struct mptcp_data_frag *dfrag;
1604 struct sock *xmit_ssk;
1605 int len, copied = 0;
1609 while ((dfrag = mptcp_send_head(sk))) {
1610 info.sent = dfrag->already_sent;
1611 info.limit = dfrag->data_len;
1612 len = dfrag->data_len - dfrag->already_sent;
1616 /* the caller already invoked the packet scheduler,
1617 * check for a different subflow usage only after
1618 * spooling the first chunk of data
1620 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1623 if (xmit_ssk != ssk) {
1624 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1625 MPTCP_DELEGATE_SEND);
1629 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1638 mptcp_update_post_push(msk, dfrag, ret);
1640 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1644 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1645 * not going to flush it via release_sock()
1647 __mptcp_update_wmem(sk);
1649 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1651 if (!mptcp_timer_pending(sk))
1652 mptcp_reset_timer(sk);
1654 if (msk->snd_data_fin_enable &&
1655 msk->snd_nxt + 1 == msk->write_seq)
1656 mptcp_schedule_work(sk);
1660 static void mptcp_set_nospace(struct sock *sk)
1662 /* enable autotune */
1663 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1665 /* will be cleared on avail space */
1666 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1669 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1671 struct mptcp_sock *msk = mptcp_sk(sk);
1672 struct page_frag *pfrag;
1677 /* we don't support FASTOPEN yet */
1678 if (msg->msg_flags & MSG_FASTOPEN)
1681 /* silently ignore everything else */
1682 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1684 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1686 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1688 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1689 ret = sk_stream_wait_connect(sk, &timeo);
1694 pfrag = sk_page_frag(sk);
1696 while (msg_data_left(msg)) {
1697 int total_ts, frag_truesize = 0;
1698 struct mptcp_data_frag *dfrag;
1699 bool dfrag_collapsed;
1700 size_t psize, offset;
1702 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1707 /* reuse tail pfrag, if possible, or carve a new one from the
1710 dfrag = mptcp_pending_tail(sk);
1711 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1712 if (!dfrag_collapsed) {
1713 if (!sk_stream_memory_free(sk))
1714 goto wait_for_memory;
1716 if (!mptcp_page_frag_refill(sk, pfrag))
1717 goto wait_for_memory;
1719 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1720 frag_truesize = dfrag->overhead;
1723 /* we do not bound vs wspace, to allow a single packet.
1724 * memory accounting will prevent execessive memory usage
1727 offset = dfrag->offset + dfrag->data_len;
1728 psize = pfrag->size - offset;
1729 psize = min_t(size_t, psize, msg_data_left(msg));
1730 total_ts = psize + frag_truesize;
1732 if (!mptcp_wmem_alloc(sk, total_ts))
1733 goto wait_for_memory;
1735 if (copy_page_from_iter(dfrag->page, offset, psize,
1736 &msg->msg_iter) != psize) {
1737 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1742 /* data successfully copied into the write queue */
1744 dfrag->data_len += psize;
1745 frag_truesize += psize;
1746 pfrag->offset += frag_truesize;
1747 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1748 msk->tx_pending_data += psize;
1750 /* charge data on mptcp pending queue to the msk socket
1751 * Note: we charge such data both to sk and ssk
1753 sk_wmem_queued_add(sk, frag_truesize);
1754 if (!dfrag_collapsed) {
1755 get_page(dfrag->page);
1756 list_add_tail(&dfrag->list, &msk->rtx_queue);
1757 if (!msk->first_pending)
1758 WRITE_ONCE(msk->first_pending, dfrag);
1760 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1761 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1767 mptcp_set_nospace(sk);
1768 __mptcp_push_pending(sk, msg->msg_flags);
1769 ret = sk_stream_wait_memory(sk, &timeo);
1775 __mptcp_push_pending(sk, msg->msg_flags);
1779 return copied ? : ret;
1782 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1784 size_t len, int flags,
1785 struct scm_timestamping_internal *tss,
1788 struct sk_buff *skb, *tmp;
1791 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1792 u32 offset = MPTCP_SKB_CB(skb)->offset;
1793 u32 data_len = skb->len - offset;
1794 u32 count = min_t(size_t, len - copied, data_len);
1797 if (!(flags & MSG_TRUNC)) {
1798 err = skb_copy_datagram_msg(skb, offset, msg, count);
1799 if (unlikely(err < 0)) {
1806 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1807 tcp_update_recv_tstamps(skb, tss);
1808 *cmsg_flags |= MPTCP_CMSG_TS;
1813 if (count < data_len) {
1814 if (!(flags & MSG_PEEK))
1815 MPTCP_SKB_CB(skb)->offset += count;
1819 if (!(flags & MSG_PEEK)) {
1820 /* we will bulk release the skb memory later */
1821 skb->destructor = NULL;
1822 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1823 __skb_unlink(skb, &msk->receive_queue);
1834 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1836 * Only difference: Use highest rtt estimate of the subflows in use.
1838 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1840 struct mptcp_subflow_context *subflow;
1841 struct sock *sk = (struct sock *)msk;
1842 u32 time, advmss = 1;
1845 sock_owned_by_me(sk);
1850 msk->rcvq_space.copied += copied;
1852 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1853 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1855 rtt_us = msk->rcvq_space.rtt_us;
1856 if (rtt_us && time < (rtt_us >> 3))
1860 mptcp_for_each_subflow(msk, subflow) {
1861 const struct tcp_sock *tp;
1865 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1867 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1868 sf_advmss = READ_ONCE(tp->advmss);
1870 rtt_us = max(sf_rtt_us, rtt_us);
1871 advmss = max(sf_advmss, advmss);
1874 msk->rcvq_space.rtt_us = rtt_us;
1875 if (time < (rtt_us >> 3) || rtt_us == 0)
1878 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1881 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1882 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1886 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1888 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1890 do_div(grow, msk->rcvq_space.space);
1891 rcvwin += (grow << 1);
1893 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1894 while (tcp_win_from_space(sk, rcvmem) < advmss)
1897 do_div(rcvwin, advmss);
1898 rcvbuf = min_t(u64, rcvwin * rcvmem,
1899 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1901 if (rcvbuf > sk->sk_rcvbuf) {
1904 window_clamp = tcp_win_from_space(sk, rcvbuf);
1905 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1907 /* Make subflows follow along. If we do not do this, we
1908 * get drops at subflow level if skbs can't be moved to
1909 * the mptcp rx queue fast enough (announced rcv_win can
1910 * exceed ssk->sk_rcvbuf).
1912 mptcp_for_each_subflow(msk, subflow) {
1916 ssk = mptcp_subflow_tcp_sock(subflow);
1917 slow = lock_sock_fast(ssk);
1918 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1919 tcp_sk(ssk)->window_clamp = window_clamp;
1920 tcp_cleanup_rbuf(ssk, 1);
1921 unlock_sock_fast(ssk, slow);
1926 msk->rcvq_space.space = msk->rcvq_space.copied;
1928 msk->rcvq_space.copied = 0;
1929 msk->rcvq_space.time = mstamp;
1932 static void __mptcp_update_rmem(struct sock *sk)
1934 struct mptcp_sock *msk = mptcp_sk(sk);
1936 if (!msk->rmem_released)
1939 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1940 sk_mem_uncharge(sk, msk->rmem_released);
1941 WRITE_ONCE(msk->rmem_released, 0);
1944 static void __mptcp_splice_receive_queue(struct sock *sk)
1946 struct mptcp_sock *msk = mptcp_sk(sk);
1948 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1951 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1953 struct sock *sk = (struct sock *)msk;
1954 unsigned int moved = 0;
1957 mptcp_flush_join_list(msk);
1959 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1962 /* we can have data pending in the subflows only if the msk
1963 * receive buffer was full at subflow_data_ready() time,
1964 * that is an unlikely slow path.
1969 slowpath = lock_sock_fast(ssk);
1970 mptcp_data_lock(sk);
1971 __mptcp_update_rmem(sk);
1972 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1973 mptcp_data_unlock(sk);
1975 if (unlikely(ssk->sk_err))
1976 __mptcp_error_report(sk);
1977 unlock_sock_fast(ssk, slowpath);
1980 /* acquire the data lock only if some input data is pending */
1982 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1983 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1984 mptcp_data_lock(sk);
1985 __mptcp_update_rmem(sk);
1986 ret |= __mptcp_ofo_queue(msk);
1987 __mptcp_splice_receive_queue(sk);
1988 mptcp_data_unlock(sk);
1991 mptcp_check_data_fin((struct sock *)msk);
1992 return !skb_queue_empty(&msk->receive_queue);
1995 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1996 int nonblock, int flags, int *addr_len)
1998 struct mptcp_sock *msk = mptcp_sk(sk);
1999 struct scm_timestamping_internal tss;
2000 int copied = 0, cmsg_flags = 0;
2004 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2005 if (unlikely(flags & MSG_ERRQUEUE))
2006 return inet_recv_error(sk, msg, len, addr_len);
2008 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
2009 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2014 timeo = sock_rcvtimeo(sk, nonblock);
2016 len = min_t(size_t, len, INT_MAX);
2017 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2019 while (copied < len) {
2022 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2023 if (unlikely(bytes_read < 0)) {
2025 copied = bytes_read;
2029 copied += bytes_read;
2031 /* be sure to advertise window change */
2032 mptcp_cleanup_rbuf(msk);
2034 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2037 /* only the master socket status is relevant here. The exit
2038 * conditions mirror closely tcp_recvmsg()
2040 if (copied >= target)
2045 sk->sk_state == TCP_CLOSE ||
2046 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2048 signal_pending(current))
2052 copied = sock_error(sk);
2056 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2057 mptcp_check_for_eof(msk);
2059 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2060 /* race breaker: the shutdown could be after the
2061 * previous receive queue check
2063 if (__mptcp_move_skbs(msk))
2068 if (sk->sk_state == TCP_CLOSE) {
2078 if (signal_pending(current)) {
2079 copied = sock_intr_errno(timeo);
2084 pr_debug("block timeout %ld", timeo);
2085 sk_wait_data(sk, &timeo, NULL);
2089 if (cmsg_flags && copied >= 0) {
2090 if (cmsg_flags & MPTCP_CMSG_TS)
2091 tcp_recv_timestamp(msg, sk, &tss);
2094 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2095 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2096 skb_queue_empty(&msk->receive_queue), copied);
2097 if (!(flags & MSG_PEEK))
2098 mptcp_rcv_space_adjust(msk, copied);
2104 static void mptcp_retransmit_timer(struct timer_list *t)
2106 struct inet_connection_sock *icsk = from_timer(icsk, t,
2107 icsk_retransmit_timer);
2108 struct sock *sk = &icsk->icsk_inet.sk;
2109 struct mptcp_sock *msk = mptcp_sk(sk);
2112 if (!sock_owned_by_user(sk)) {
2113 /* we need a process context to retransmit */
2114 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2115 mptcp_schedule_work(sk);
2117 /* delegate our work to tcp_release_cb() */
2118 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2124 static void mptcp_timeout_timer(struct timer_list *t)
2126 struct sock *sk = from_timer(sk, t, sk_timer);
2128 mptcp_schedule_work(sk);
2132 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2135 * A backup subflow is returned only if that is the only kind available.
2137 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2139 struct sock *backup = NULL, *pick = NULL;
2140 struct mptcp_subflow_context *subflow;
2141 int min_stale_count = INT_MAX;
2143 sock_owned_by_me((const struct sock *)msk);
2145 if (__mptcp_check_fallback(msk))
2148 mptcp_for_each_subflow(msk, subflow) {
2149 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2151 if (!__mptcp_subflow_active(subflow))
2154 /* still data outstanding at TCP level? skip this */
2155 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2156 mptcp_pm_subflow_chk_stale(msk, ssk);
2157 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2161 if (subflow->backup) {
2174 /* use backup only if there are no progresses anywhere */
2175 return min_stale_count > 1 ? backup : NULL;
2178 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2181 iput(SOCK_INODE(msk->subflow));
2182 msk->subflow = NULL;
2186 bool __mptcp_retransmit_pending_data(struct sock *sk)
2188 struct mptcp_data_frag *cur, *rtx_head;
2189 struct mptcp_sock *msk = mptcp_sk(sk);
2191 if (__mptcp_check_fallback(mptcp_sk(sk)))
2194 if (tcp_rtx_and_write_queues_empty(sk))
2197 /* the closing socket has some data untransmitted and/or unacked:
2198 * some data in the mptcp rtx queue has not really xmitted yet.
2199 * keep it simple and re-inject the whole mptcp level rtx queue
2201 mptcp_data_lock(sk);
2202 __mptcp_clean_una_wakeup(sk);
2203 rtx_head = mptcp_rtx_head(sk);
2205 mptcp_data_unlock(sk);
2209 msk->recovery_snd_nxt = msk->snd_nxt;
2210 msk->recovery = true;
2211 mptcp_data_unlock(sk);
2213 msk->first_pending = rtx_head;
2214 msk->tx_pending_data += msk->snd_nxt - rtx_head->data_seq;
2217 /* be sure to clear the "sent status" on all re-injected fragments */
2218 list_for_each_entry(cur, &msk->rtx_queue, list) {
2219 if (!cur->already_sent)
2221 cur->already_sent = 0;
2227 /* subflow sockets can be either outgoing (connect) or incoming
2230 * Outgoing subflows use in-kernel sockets.
2231 * Incoming subflows do not have their own 'struct socket' allocated,
2232 * so we need to use tcp_close() after detaching them from the mptcp
2235 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2236 struct mptcp_subflow_context *subflow)
2238 struct mptcp_sock *msk = mptcp_sk(sk);
2241 list_del(&subflow->node);
2243 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2245 /* if we are invoked by the msk cleanup code, the subflow is
2251 need_push = __mptcp_retransmit_pending_data(sk);
2252 subflow->disposable = 1;
2254 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2255 * the ssk has been already destroyed, we just need to release the
2256 * reference owned by msk;
2258 if (!inet_csk(ssk)->icsk_ulp_ops) {
2259 kfree_rcu(subflow, rcu);
2261 /* otherwise tcp will dispose of the ssk and subflow ctx */
2262 __tcp_close(ssk, 0);
2264 /* close acquired an extra ref */
2271 if (ssk == msk->last_snd)
2272 msk->last_snd = NULL;
2274 if (ssk == msk->first)
2277 if (msk->subflow && ssk == msk->subflow->sk)
2278 mptcp_dispose_initial_subflow(msk);
2281 __mptcp_push_pending(sk, 0);
2284 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2285 struct mptcp_subflow_context *subflow)
2287 if (sk->sk_state == TCP_ESTABLISHED)
2288 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2289 __mptcp_close_ssk(sk, ssk, subflow);
2292 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2297 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2299 struct mptcp_subflow_context *subflow, *tmp;
2303 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2304 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2306 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2309 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2310 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2313 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2317 static bool mptcp_check_close_timeout(const struct sock *sk)
2319 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2320 struct mptcp_subflow_context *subflow;
2322 if (delta >= TCP_TIMEWAIT_LEN)
2325 /* if all subflows are in closed status don't bother with additional
2328 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2329 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2336 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2338 struct mptcp_subflow_context *subflow, *tmp;
2339 struct sock *sk = &msk->sk.icsk_inet.sk;
2341 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2344 mptcp_token_destroy(msk);
2346 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2347 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2350 slow = lock_sock_fast(tcp_sk);
2351 if (tcp_sk->sk_state != TCP_CLOSE) {
2352 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2353 tcp_set_state(tcp_sk, TCP_CLOSE);
2355 unlock_sock_fast(tcp_sk, slow);
2358 inet_sk_state_store(sk, TCP_CLOSE);
2359 sk->sk_shutdown = SHUTDOWN_MASK;
2360 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2361 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2363 mptcp_close_wake_up(sk);
2366 static void __mptcp_retrans(struct sock *sk)
2368 struct mptcp_sock *msk = mptcp_sk(sk);
2369 struct mptcp_sendmsg_info info = {};
2370 struct mptcp_data_frag *dfrag;
2375 mptcp_clean_una_wakeup(sk);
2376 dfrag = mptcp_rtx_head(sk);
2378 if (mptcp_data_fin_enabled(msk)) {
2379 struct inet_connection_sock *icsk = inet_csk(sk);
2381 icsk->icsk_retransmits++;
2382 mptcp_set_datafin_timeout(sk);
2383 mptcp_send_ack(msk);
2391 ssk = mptcp_subflow_get_retrans(msk);
2397 /* limit retransmission to the bytes already sent on some subflows */
2399 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2400 while (info.sent < info.limit) {
2401 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2405 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2410 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2411 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2418 if (!mptcp_timer_pending(sk))
2419 mptcp_reset_timer(sk);
2422 static void mptcp_worker(struct work_struct *work)
2424 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2425 struct sock *sk = &msk->sk.icsk_inet.sk;
2429 state = sk->sk_state;
2430 if (unlikely(state == TCP_CLOSE))
2433 mptcp_check_data_fin_ack(sk);
2434 mptcp_flush_join_list(msk);
2436 mptcp_check_fastclose(msk);
2439 mptcp_pm_nl_work(msk);
2441 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2442 mptcp_check_for_eof(msk);
2444 __mptcp_check_send_data_fin(sk);
2445 mptcp_check_data_fin(sk);
2447 /* There is no point in keeping around an orphaned sk timedout or
2448 * closed, but we need the msk around to reply to incoming DATA_FIN,
2449 * even if it is orphaned and in FIN_WAIT2 state
2451 if (sock_flag(sk, SOCK_DEAD) &&
2452 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2453 inet_sk_state_store(sk, TCP_CLOSE);
2454 __mptcp_destroy_sock(sk);
2458 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2459 __mptcp_close_subflow(msk);
2461 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2462 __mptcp_retrans(sk);
2469 static int __mptcp_init_sock(struct sock *sk)
2471 struct mptcp_sock *msk = mptcp_sk(sk);
2473 spin_lock_init(&msk->join_list_lock);
2475 INIT_LIST_HEAD(&msk->conn_list);
2476 INIT_LIST_HEAD(&msk->join_list);
2477 INIT_LIST_HEAD(&msk->rtx_queue);
2478 INIT_WORK(&msk->work, mptcp_worker);
2479 __skb_queue_head_init(&msk->receive_queue);
2480 msk->out_of_order_queue = RB_ROOT;
2481 msk->first_pending = NULL;
2482 msk->wmem_reserved = 0;
2483 WRITE_ONCE(msk->rmem_released, 0);
2484 msk->tx_pending_data = 0;
2485 msk->timer_ival = TCP_RTO_MIN;
2488 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2489 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2490 msk->recovery = false;
2492 mptcp_pm_data_init(msk);
2494 /* re-use the csk retrans timer for MPTCP-level retrans */
2495 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2496 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2501 static int mptcp_init_sock(struct sock *sk)
2503 struct inet_connection_sock *icsk = inet_csk(sk);
2504 struct net *net = sock_net(sk);
2507 ret = __mptcp_init_sock(sk);
2511 if (!mptcp_is_enabled(net))
2512 return -ENOPROTOOPT;
2514 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2517 ret = __mptcp_socket_create(mptcp_sk(sk));
2521 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2522 * propagate the correct value
2524 tcp_assign_congestion_control(sk);
2525 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2527 /* no need to keep a reference to the ops, the name will suffice */
2528 tcp_cleanup_congestion_control(sk);
2529 icsk->icsk_ca_ops = NULL;
2531 sk_sockets_allocated_inc(sk);
2532 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2533 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2538 static void __mptcp_clear_xmit(struct sock *sk)
2540 struct mptcp_sock *msk = mptcp_sk(sk);
2541 struct mptcp_data_frag *dtmp, *dfrag;
2543 WRITE_ONCE(msk->first_pending, NULL);
2544 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2545 dfrag_clear(sk, dfrag);
2548 static void mptcp_cancel_work(struct sock *sk)
2550 struct mptcp_sock *msk = mptcp_sk(sk);
2552 if (cancel_work_sync(&msk->work))
2556 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2560 switch (ssk->sk_state) {
2562 if (!(how & RCV_SHUTDOWN))
2566 tcp_disconnect(ssk, O_NONBLOCK);
2569 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2570 pr_debug("Fallback");
2571 ssk->sk_shutdown |= how;
2572 tcp_shutdown(ssk, how);
2574 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2576 if (!mptcp_timer_pending(sk))
2577 mptcp_reset_timer(sk);
2585 static const unsigned char new_state[16] = {
2586 /* current state: new state: action: */
2587 [0 /* (Invalid) */] = TCP_CLOSE,
2588 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2589 [TCP_SYN_SENT] = TCP_CLOSE,
2590 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2591 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2592 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2593 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2594 [TCP_CLOSE] = TCP_CLOSE,
2595 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2596 [TCP_LAST_ACK] = TCP_LAST_ACK,
2597 [TCP_LISTEN] = TCP_CLOSE,
2598 [TCP_CLOSING] = TCP_CLOSING,
2599 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2602 static int mptcp_close_state(struct sock *sk)
2604 int next = (int)new_state[sk->sk_state];
2605 int ns = next & TCP_STATE_MASK;
2607 inet_sk_state_store(sk, ns);
2609 return next & TCP_ACTION_FIN;
2612 static void __mptcp_check_send_data_fin(struct sock *sk)
2614 struct mptcp_subflow_context *subflow;
2615 struct mptcp_sock *msk = mptcp_sk(sk);
2617 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2618 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2619 msk->snd_nxt, msk->write_seq);
2621 /* we still need to enqueue subflows or not really shutting down,
2624 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2625 mptcp_send_head(sk))
2628 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2630 /* fallback socket will not get data_fin/ack, can move to the next
2633 if (__mptcp_check_fallback(msk)) {
2634 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2635 inet_sk_state_store(sk, TCP_CLOSE);
2636 mptcp_close_wake_up(sk);
2637 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2638 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2642 mptcp_flush_join_list(msk);
2643 mptcp_for_each_subflow(msk, subflow) {
2644 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2646 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2650 static void __mptcp_wr_shutdown(struct sock *sk)
2652 struct mptcp_sock *msk = mptcp_sk(sk);
2654 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2655 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2656 !!mptcp_send_head(sk));
2658 /* will be ignored by fallback sockets */
2659 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2660 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2662 __mptcp_check_send_data_fin(sk);
2665 static void __mptcp_destroy_sock(struct sock *sk)
2667 struct mptcp_subflow_context *subflow, *tmp;
2668 struct mptcp_sock *msk = mptcp_sk(sk);
2669 LIST_HEAD(conn_list);
2671 pr_debug("msk=%p", msk);
2675 /* be sure to always acquire the join list lock, to sync vs
2676 * mptcp_finish_join().
2678 spin_lock_bh(&msk->join_list_lock);
2679 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2680 spin_unlock_bh(&msk->join_list_lock);
2681 list_splice_init(&msk->conn_list, &conn_list);
2683 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2684 sk_stop_timer(sk, &sk->sk_timer);
2687 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2688 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2689 __mptcp_close_ssk(sk, ssk, subflow);
2692 sk->sk_prot->destroy(sk);
2694 WARN_ON_ONCE(msk->wmem_reserved);
2695 WARN_ON_ONCE(msk->rmem_released);
2696 sk_stream_kill_queues(sk);
2697 xfrm_sk_free_policy(sk);
2699 sk_refcnt_debug_release(sk);
2700 mptcp_dispose_initial_subflow(msk);
2704 static void mptcp_close(struct sock *sk, long timeout)
2706 struct mptcp_subflow_context *subflow;
2707 bool do_cancel_work = false;
2710 sk->sk_shutdown = SHUTDOWN_MASK;
2712 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2713 inet_sk_state_store(sk, TCP_CLOSE);
2717 if (mptcp_close_state(sk))
2718 __mptcp_wr_shutdown(sk);
2720 sk_stream_wait_close(sk, timeout);
2723 /* orphan all the subflows */
2724 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2725 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2726 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2727 bool slow = lock_sock_fast_nested(ssk);
2730 unlock_sock_fast(ssk, slow);
2735 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2736 if (sk->sk_state == TCP_CLOSE) {
2737 __mptcp_destroy_sock(sk);
2738 do_cancel_work = true;
2740 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2744 mptcp_cancel_work(sk);
2746 if (mptcp_sk(sk)->token)
2747 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2752 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2754 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2755 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2756 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2758 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2759 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2762 msk6->saddr = ssk6->saddr;
2763 msk6->flow_label = ssk6->flow_label;
2767 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2768 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2769 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2770 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2771 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2772 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2775 static int mptcp_disconnect(struct sock *sk, int flags)
2777 struct mptcp_subflow_context *subflow;
2778 struct mptcp_sock *msk = mptcp_sk(sk);
2780 mptcp_do_flush_join_list(msk);
2782 mptcp_for_each_subflow(msk, subflow) {
2783 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2786 tcp_disconnect(ssk, flags);
2792 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2793 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2795 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2797 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2801 struct sock *mptcp_sk_clone(const struct sock *sk,
2802 const struct mptcp_options_received *mp_opt,
2803 struct request_sock *req)
2805 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2806 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2807 struct mptcp_sock *msk;
2813 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2814 if (nsk->sk_family == AF_INET6)
2815 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2818 __mptcp_init_sock(nsk);
2820 msk = mptcp_sk(nsk);
2821 msk->local_key = subflow_req->local_key;
2822 msk->token = subflow_req->token;
2823 msk->subflow = NULL;
2824 WRITE_ONCE(msk->fully_established, false);
2825 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
2826 WRITE_ONCE(msk->csum_enabled, true);
2828 msk->write_seq = subflow_req->idsn + 1;
2829 msk->snd_nxt = msk->write_seq;
2830 msk->snd_una = msk->write_seq;
2831 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2832 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2834 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
2835 msk->can_ack = true;
2836 msk->remote_key = mp_opt->sndr_key;
2837 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2839 WRITE_ONCE(msk->ack_seq, ack_seq);
2840 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2843 sock_reset_flag(nsk, SOCK_RCU_FREE);
2844 /* will be fully established after successful MPC subflow creation */
2845 inet_sk_state_store(nsk, TCP_SYN_RECV);
2847 security_inet_csk_clone(nsk, req);
2848 bh_unlock_sock(nsk);
2850 /* keep a single reference */
2855 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2857 const struct tcp_sock *tp = tcp_sk(ssk);
2859 msk->rcvq_space.copied = 0;
2860 msk->rcvq_space.rtt_us = 0;
2862 msk->rcvq_space.time = tp->tcp_mstamp;
2864 /* initial rcv_space offering made to peer */
2865 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2866 TCP_INIT_CWND * tp->advmss);
2867 if (msk->rcvq_space.space == 0)
2868 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2870 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2873 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2876 struct mptcp_sock *msk = mptcp_sk(sk);
2877 struct socket *listener;
2880 listener = __mptcp_nmpc_socket(msk);
2881 if (WARN_ON_ONCE(!listener)) {
2886 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2887 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2891 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2892 if (sk_is_mptcp(newsk)) {
2893 struct mptcp_subflow_context *subflow;
2894 struct sock *new_mptcp_sock;
2896 subflow = mptcp_subflow_ctx(newsk);
2897 new_mptcp_sock = subflow->conn;
2899 /* is_mptcp should be false if subflow->conn is missing, see
2900 * subflow_syn_recv_sock()
2902 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2903 tcp_sk(newsk)->is_mptcp = 0;
2907 /* acquire the 2nd reference for the owning socket */
2908 sock_hold(new_mptcp_sock);
2909 newsk = new_mptcp_sock;
2910 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2912 MPTCP_INC_STATS(sock_net(sk),
2913 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2917 newsk->sk_kern_sock = kern;
2921 void mptcp_destroy_common(struct mptcp_sock *msk)
2923 struct sock *sk = (struct sock *)msk;
2925 __mptcp_clear_xmit(sk);
2927 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2928 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2930 skb_rbtree_purge(&msk->out_of_order_queue);
2931 mptcp_token_destroy(msk);
2932 mptcp_pm_free_anno_list(msk);
2935 static void mptcp_destroy(struct sock *sk)
2937 struct mptcp_sock *msk = mptcp_sk(sk);
2939 mptcp_destroy_common(msk);
2940 sk_sockets_allocated_dec(sk);
2943 void __mptcp_data_acked(struct sock *sk)
2945 if (!sock_owned_by_user(sk))
2946 __mptcp_clean_una(sk);
2948 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2950 if (mptcp_pending_data_fin_ack(sk))
2951 mptcp_schedule_work(sk);
2954 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2956 if (!mptcp_send_head(sk))
2959 if (!sock_owned_by_user(sk)) {
2960 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2962 if (xmit_ssk == ssk)
2963 __mptcp_subflow_push_pending(sk, ssk);
2965 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
2967 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2971 /* processes deferred events and flush wmem */
2972 static void mptcp_release_cb(struct sock *sk)
2975 unsigned long flags = 0;
2977 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2978 flags |= BIT(MPTCP_PUSH_PENDING);
2979 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2980 flags |= BIT(MPTCP_RETRANSMIT);
2984 /* the following actions acquire the subflow socket lock
2986 * 1) can't be invoked in atomic scope
2987 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2988 * datapath acquires the msk socket spinlock while helding
2989 * the subflow socket lock
2992 spin_unlock_bh(&sk->sk_lock.slock);
2993 if (flags & BIT(MPTCP_PUSH_PENDING))
2994 __mptcp_push_pending(sk, 0);
2995 if (flags & BIT(MPTCP_RETRANSMIT))
2996 __mptcp_retrans(sk);
2999 spin_lock_bh(&sk->sk_lock.slock);
3002 /* be sure to set the current sk state before tacking actions
3003 * depending on sk_state
3005 if (test_and_clear_bit(MPTCP_CONNECTED, &mptcp_sk(sk)->flags))
3006 __mptcp_set_connected(sk);
3007 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
3008 __mptcp_clean_una_wakeup(sk);
3009 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
3010 __mptcp_error_report(sk);
3012 /* push_pending may touch wmem_reserved, ensure we do the cleanup
3015 __mptcp_update_wmem(sk);
3016 __mptcp_update_rmem(sk);
3019 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3020 * TCP can't schedule delack timer before the subflow is fully established.
3021 * MPTCP uses the delack timer to do 3rd ack retransmissions
3023 static void schedule_3rdack_retransmission(struct sock *ssk)
3025 struct inet_connection_sock *icsk = inet_csk(ssk);
3026 struct tcp_sock *tp = tcp_sk(ssk);
3027 unsigned long timeout;
3029 if (mptcp_subflow_ctx(ssk)->fully_established)
3032 /* reschedule with a timeout above RTT, as we must look only for drop */
3034 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3036 timeout = TCP_TIMEOUT_INIT;
3039 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3040 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3041 icsk->icsk_ack.timeout = timeout;
3042 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3045 void mptcp_subflow_process_delegated(struct sock *ssk)
3047 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3048 struct sock *sk = subflow->conn;
3050 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3051 mptcp_data_lock(sk);
3052 if (!sock_owned_by_user(sk))
3053 __mptcp_subflow_push_pending(sk, ssk);
3055 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
3056 mptcp_data_unlock(sk);
3057 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3059 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3060 schedule_3rdack_retransmission(ssk);
3061 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3065 static int mptcp_hash(struct sock *sk)
3067 /* should never be called,
3068 * we hash the TCP subflows not the master socket
3074 static void mptcp_unhash(struct sock *sk)
3076 /* called from sk_common_release(), but nothing to do here */
3079 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3081 struct mptcp_sock *msk = mptcp_sk(sk);
3082 struct socket *ssock;
3084 ssock = __mptcp_nmpc_socket(msk);
3085 pr_debug("msk=%p, subflow=%p", msk, ssock);
3086 if (WARN_ON_ONCE(!ssock))
3089 return inet_csk_get_port(ssock->sk, snum);
3092 void mptcp_finish_connect(struct sock *ssk)
3094 struct mptcp_subflow_context *subflow;
3095 struct mptcp_sock *msk;
3099 subflow = mptcp_subflow_ctx(ssk);
3103 pr_debug("msk=%p, token=%u", sk, subflow->token);
3105 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3107 subflow->map_seq = ack_seq;
3108 subflow->map_subflow_seq = 1;
3110 /* the socket is not connected yet, no msk/subflow ops can access/race
3111 * accessing the field below
3113 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3114 WRITE_ONCE(msk->local_key, subflow->local_key);
3115 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3116 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3117 WRITE_ONCE(msk->ack_seq, ack_seq);
3118 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3119 WRITE_ONCE(msk->can_ack, 1);
3120 WRITE_ONCE(msk->snd_una, msk->write_seq);
3122 mptcp_pm_new_connection(msk, ssk, 0);
3124 mptcp_rcv_space_init(msk, ssk);
3127 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3129 write_lock_bh(&sk->sk_callback_lock);
3130 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3131 sk_set_socket(sk, parent);
3132 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3133 write_unlock_bh(&sk->sk_callback_lock);
3136 bool mptcp_finish_join(struct sock *ssk)
3138 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3139 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3140 struct sock *parent = (void *)msk;
3141 struct socket *parent_sock;
3144 pr_debug("msk=%p, subflow=%p", msk, subflow);
3146 /* mptcp socket already closing? */
3147 if (!mptcp_is_fully_established(parent)) {
3148 subflow->reset_reason = MPTCP_RST_EMPTCP;
3152 if (!msk->pm.server_side)
3155 if (!mptcp_pm_allow_new_subflow(msk)) {
3156 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3160 /* active connections are already on conn_list, and we can't acquire
3162 * use the join list lock as synchronization point and double-check
3163 * msk status to avoid racing with __mptcp_destroy_sock()
3165 spin_lock_bh(&msk->join_list_lock);
3166 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3167 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3168 list_add_tail(&subflow->node, &msk->join_list);
3171 spin_unlock_bh(&msk->join_list_lock);
3173 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3177 /* attach to msk socket only after we are sure he will deal with us
3180 parent_sock = READ_ONCE(parent->sk_socket);
3181 if (parent_sock && !ssk->sk_socket)
3182 mptcp_sock_graft(ssk, parent_sock);
3183 subflow->map_seq = READ_ONCE(msk->ack_seq);
3185 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3189 static void mptcp_shutdown(struct sock *sk, int how)
3191 pr_debug("sk=%p, how=%d", sk, how);
3193 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3194 __mptcp_wr_shutdown(sk);
3197 static struct proto mptcp_prot = {
3199 .owner = THIS_MODULE,
3200 .init = mptcp_init_sock,
3201 .disconnect = mptcp_disconnect,
3202 .close = mptcp_close,
3203 .accept = mptcp_accept,
3204 .setsockopt = mptcp_setsockopt,
3205 .getsockopt = mptcp_getsockopt,
3206 .shutdown = mptcp_shutdown,
3207 .destroy = mptcp_destroy,
3208 .sendmsg = mptcp_sendmsg,
3209 .recvmsg = mptcp_recvmsg,
3210 .release_cb = mptcp_release_cb,
3212 .unhash = mptcp_unhash,
3213 .get_port = mptcp_get_port,
3214 .sockets_allocated = &mptcp_sockets_allocated,
3215 .memory_allocated = &tcp_memory_allocated,
3216 .memory_pressure = &tcp_memory_pressure,
3217 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3218 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3219 .sysctl_mem = sysctl_tcp_mem,
3220 .obj_size = sizeof(struct mptcp_sock),
3221 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3222 .no_autobind = true,
3225 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3227 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3228 struct socket *ssock;
3231 lock_sock(sock->sk);
3232 ssock = __mptcp_nmpc_socket(msk);
3238 err = ssock->ops->bind(ssock, uaddr, addr_len);
3240 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3243 release_sock(sock->sk);
3247 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3248 struct mptcp_subflow_context *subflow)
3250 subflow->request_mptcp = 0;
3251 __mptcp_do_fallback(msk);
3254 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3255 int addr_len, int flags)
3257 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3258 struct mptcp_subflow_context *subflow;
3259 struct socket *ssock;
3262 lock_sock(sock->sk);
3263 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3264 /* pending connection or invalid state, let existing subflow
3267 ssock = msk->subflow;
3271 ssock = __mptcp_nmpc_socket(msk);
3277 mptcp_token_destroy(msk);
3278 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3279 subflow = mptcp_subflow_ctx(ssock->sk);
3280 #ifdef CONFIG_TCP_MD5SIG
3281 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3284 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3285 mptcp_subflow_early_fallback(msk, subflow);
3287 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3288 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3289 mptcp_subflow_early_fallback(msk, subflow);
3291 if (likely(!__mptcp_check_fallback(msk)))
3292 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3295 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3296 sock->state = ssock->state;
3298 /* on successful connect, the msk state will be moved to established by
3299 * subflow_finish_connect()
3301 if (!err || err == -EINPROGRESS)
3302 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3304 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3307 release_sock(sock->sk);
3311 static int mptcp_listen(struct socket *sock, int backlog)
3313 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3314 struct socket *ssock;
3317 pr_debug("msk=%p", msk);
3319 lock_sock(sock->sk);
3320 ssock = __mptcp_nmpc_socket(msk);
3326 mptcp_token_destroy(msk);
3327 inet_sk_state_store(sock->sk, TCP_LISTEN);
3328 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3330 err = ssock->ops->listen(ssock, backlog);
3331 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3333 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3336 release_sock(sock->sk);
3340 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3341 int flags, bool kern)
3343 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3344 struct socket *ssock;
3347 pr_debug("msk=%p", msk);
3349 lock_sock(sock->sk);
3350 if (sock->sk->sk_state != TCP_LISTEN)
3353 ssock = __mptcp_nmpc_socket(msk);
3357 clear_bit(MPTCP_DATA_READY, &msk->flags);
3358 sock_hold(ssock->sk);
3359 release_sock(sock->sk);
3361 err = ssock->ops->accept(sock, newsock, flags, kern);
3362 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3363 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3364 struct mptcp_subflow_context *subflow;
3365 struct sock *newsk = newsock->sk;
3369 /* PM/worker can now acquire the first subflow socket
3370 * lock without racing with listener queue cleanup,
3371 * we can notify it, if needed.
3373 * Even if remote has reset the initial subflow by now
3374 * the refcnt is still at least one.
3376 subflow = mptcp_subflow_ctx(msk->first);
3377 list_add(&subflow->node, &msk->conn_list);
3378 sock_hold(msk->first);
3379 if (mptcp_is_fully_established(newsk))
3380 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3382 mptcp_copy_inaddrs(newsk, msk->first);
3383 mptcp_rcv_space_init(msk, msk->first);
3384 mptcp_propagate_sndbuf(newsk, msk->first);
3386 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3387 * This is needed so NOSPACE flag can be set from tcp stack.
3389 mptcp_flush_join_list(msk);
3390 mptcp_for_each_subflow(msk, subflow) {
3391 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3393 if (!ssk->sk_socket)
3394 mptcp_sock_graft(ssk, newsock);
3396 release_sock(newsk);
3399 if (inet_csk_listen_poll(ssock->sk))
3400 set_bit(MPTCP_DATA_READY, &msk->flags);
3401 sock_put(ssock->sk);
3405 release_sock(sock->sk);
3409 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3411 /* Concurrent splices from sk_receive_queue into receive_queue will
3412 * always show at least one non-empty queue when checked in this order.
3414 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
3415 skb_queue_empty_lockless(&msk->receive_queue))
3418 return EPOLLIN | EPOLLRDNORM;
3421 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3423 struct sock *sk = (struct sock *)msk;
3425 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3426 return EPOLLOUT | EPOLLWRNORM;
3428 if (sk_stream_is_writeable(sk))
3429 return EPOLLOUT | EPOLLWRNORM;
3431 mptcp_set_nospace(sk);
3432 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3433 if (sk_stream_is_writeable(sk))
3434 return EPOLLOUT | EPOLLWRNORM;
3439 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3440 struct poll_table_struct *wait)
3442 struct sock *sk = sock->sk;
3443 struct mptcp_sock *msk;
3448 sock_poll_wait(file, sock, wait);
3450 state = inet_sk_state_load(sk);
3451 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3452 if (state == TCP_LISTEN)
3453 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM : 0;
3455 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3456 mask |= mptcp_check_readable(msk);
3457 mask |= mptcp_check_writeable(msk);
3459 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3461 if (sk->sk_shutdown & RCV_SHUTDOWN)
3462 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3464 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3472 static const struct proto_ops mptcp_stream_ops = {
3474 .owner = THIS_MODULE,
3475 .release = inet_release,
3477 .connect = mptcp_stream_connect,
3478 .socketpair = sock_no_socketpair,
3479 .accept = mptcp_stream_accept,
3480 .getname = inet_getname,
3482 .ioctl = inet_ioctl,
3483 .gettstamp = sock_gettstamp,
3484 .listen = mptcp_listen,
3485 .shutdown = inet_shutdown,
3486 .setsockopt = sock_common_setsockopt,
3487 .getsockopt = sock_common_getsockopt,
3488 .sendmsg = inet_sendmsg,
3489 .recvmsg = inet_recvmsg,
3490 .mmap = sock_no_mmap,
3491 .sendpage = inet_sendpage,
3494 static struct inet_protosw mptcp_protosw = {
3495 .type = SOCK_STREAM,
3496 .protocol = IPPROTO_MPTCP,
3497 .prot = &mptcp_prot,
3498 .ops = &mptcp_stream_ops,
3499 .flags = INET_PROTOSW_ICSK,
3502 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3504 struct mptcp_delegated_action *delegated;
3505 struct mptcp_subflow_context *subflow;
3508 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3509 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3510 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3512 bh_lock_sock_nested(ssk);
3513 if (!sock_owned_by_user(ssk) &&
3514 mptcp_subflow_has_delegated_action(subflow))
3515 mptcp_subflow_process_delegated(ssk);
3516 /* ... elsewhere tcp_release_cb_override already processed
3517 * the action or will do at next release_sock().
3518 * In both case must dequeue the subflow here - on the same
3519 * CPU that scheduled it.
3521 bh_unlock_sock(ssk);
3524 if (++work_done == budget)
3528 /* always provide a 0 'work_done' argument, so that napi_complete_done
3529 * will not try accessing the NULL napi->dev ptr
3531 napi_complete_done(napi, 0);
3535 void __init mptcp_proto_init(void)
3537 struct mptcp_delegated_action *delegated;
3540 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3542 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3543 panic("Failed to allocate MPTCP pcpu counter\n");
3545 init_dummy_netdev(&mptcp_napi_dev);
3546 for_each_possible_cpu(cpu) {
3547 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3548 INIT_LIST_HEAD(&delegated->head);
3549 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3551 napi_enable(&delegated->napi);
3554 mptcp_subflow_init();
3558 if (proto_register(&mptcp_prot, 1) != 0)
3559 panic("Failed to register MPTCP proto.\n");
3561 inet_register_protosw(&mptcp_protosw);
3563 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3566 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3567 static const struct proto_ops mptcp_v6_stream_ops = {
3569 .owner = THIS_MODULE,
3570 .release = inet6_release,
3572 .connect = mptcp_stream_connect,
3573 .socketpair = sock_no_socketpair,
3574 .accept = mptcp_stream_accept,
3575 .getname = inet6_getname,
3577 .ioctl = inet6_ioctl,
3578 .gettstamp = sock_gettstamp,
3579 .listen = mptcp_listen,
3580 .shutdown = inet_shutdown,
3581 .setsockopt = sock_common_setsockopt,
3582 .getsockopt = sock_common_getsockopt,
3583 .sendmsg = inet6_sendmsg,
3584 .recvmsg = inet6_recvmsg,
3585 .mmap = sock_no_mmap,
3586 .sendpage = inet_sendpage,
3587 #ifdef CONFIG_COMPAT
3588 .compat_ioctl = inet6_compat_ioctl,
3592 static struct proto mptcp_v6_prot;
3594 static void mptcp_v6_destroy(struct sock *sk)
3597 inet6_destroy_sock(sk);
3600 static struct inet_protosw mptcp_v6_protosw = {
3601 .type = SOCK_STREAM,
3602 .protocol = IPPROTO_MPTCP,
3603 .prot = &mptcp_v6_prot,
3604 .ops = &mptcp_v6_stream_ops,
3605 .flags = INET_PROTOSW_ICSK,
3608 int __init mptcp_proto_v6_init(void)
3612 mptcp_v6_prot = mptcp_prot;
3613 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3614 mptcp_v6_prot.slab = NULL;
3615 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3616 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3618 err = proto_register(&mptcp_v6_prot, 1);
3622 err = inet6_register_protosw(&mptcp_v6_protosw);
3624 proto_unregister(&mptcp_v6_prot);