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 void mptcp_rmem_charge(struct sock *sk, int size)
131 mptcp_sk(sk)->rmem_fwd_alloc -= size;
134 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
135 struct sk_buff *from)
140 if (MPTCP_SKB_CB(from)->offset ||
141 !skb_try_coalesce(to, from, &fragstolen, &delta))
144 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
145 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
146 to->len, MPTCP_SKB_CB(from)->end_seq);
147 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
148 kfree_skb_partial(from, fragstolen);
149 atomic_add(delta, &sk->sk_rmem_alloc);
150 mptcp_rmem_charge(sk, delta);
154 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
155 struct sk_buff *from)
157 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
160 return mptcp_try_coalesce((struct sock *)msk, to, from);
163 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
165 amount >>= SK_MEM_QUANTUM_SHIFT;
166 mptcp_sk(sk)->rmem_fwd_alloc -= amount << SK_MEM_QUANTUM_SHIFT;
167 __sk_mem_reduce_allocated(sk, amount);
170 static void mptcp_rmem_uncharge(struct sock *sk, int size)
172 struct mptcp_sock *msk = mptcp_sk(sk);
175 msk->rmem_fwd_alloc += size;
176 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
178 /* see sk_mem_uncharge() for the rationale behind the following schema */
179 if (unlikely(reclaimable >= SK_RECLAIM_THRESHOLD))
180 __mptcp_rmem_reclaim(sk, SK_RECLAIM_CHUNK);
183 static void mptcp_rfree(struct sk_buff *skb)
185 unsigned int len = skb->truesize;
186 struct sock *sk = skb->sk;
188 atomic_sub(len, &sk->sk_rmem_alloc);
189 mptcp_rmem_uncharge(sk, len);
192 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
196 skb->destructor = mptcp_rfree;
197 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
198 mptcp_rmem_charge(sk, skb->truesize);
201 /* "inspired" by tcp_data_queue_ofo(), main differences:
203 * - don't cope with sacks
205 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
207 struct sock *sk = (struct sock *)msk;
208 struct rb_node **p, *parent;
209 u64 seq, end_seq, max_seq;
210 struct sk_buff *skb1;
212 seq = MPTCP_SKB_CB(skb)->map_seq;
213 end_seq = MPTCP_SKB_CB(skb)->end_seq;
214 max_seq = READ_ONCE(msk->rcv_wnd_sent);
216 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
217 RB_EMPTY_ROOT(&msk->out_of_order_queue));
218 if (after64(end_seq, max_seq)) {
221 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
222 (unsigned long long)end_seq - (unsigned long)max_seq,
223 (unsigned long long)msk->rcv_wnd_sent);
224 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
228 p = &msk->out_of_order_queue.rb_node;
229 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
230 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
231 rb_link_node(&skb->rbnode, NULL, p);
232 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
233 msk->ooo_last_skb = skb;
237 /* with 2 subflows, adding at end of ooo queue is quite likely
238 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
240 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
241 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
242 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
246 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
247 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
248 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
249 parent = &msk->ooo_last_skb->rbnode;
250 p = &parent->rb_right;
254 /* Find place to insert this segment. Handle overlaps on the way. */
258 skb1 = rb_to_skb(parent);
259 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
260 p = &parent->rb_left;
263 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
264 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
265 /* All the bits are present. Drop. */
267 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
270 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
274 * continue traversing
277 /* skb's seq == skb1's seq and skb covers skb1.
278 * Replace skb1 with skb.
280 rb_replace_node(&skb1->rbnode, &skb->rbnode,
281 &msk->out_of_order_queue);
282 mptcp_drop(sk, skb1);
283 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
286 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
287 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
290 p = &parent->rb_right;
294 /* Insert segment into RB tree. */
295 rb_link_node(&skb->rbnode, parent, p);
296 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
299 /* Remove other segments covered by skb. */
300 while ((skb1 = skb_rb_next(skb)) != NULL) {
301 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
303 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
304 mptcp_drop(sk, skb1);
305 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
307 /* If there is no skb after us, we are the last_skb ! */
309 msk->ooo_last_skb = skb;
313 mptcp_set_owner_r(skb, sk);
316 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
318 struct mptcp_sock *msk = mptcp_sk(sk);
321 if (size < msk->rmem_fwd_alloc)
324 amt = sk_mem_pages(size);
325 amount = amt << SK_MEM_QUANTUM_SHIFT;
326 msk->rmem_fwd_alloc += amount;
327 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV)) {
328 if (ssk->sk_forward_alloc < amount) {
329 msk->rmem_fwd_alloc -= amount;
333 ssk->sk_forward_alloc -= amount;
338 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
339 struct sk_buff *skb, unsigned int offset,
342 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
343 struct sock *sk = (struct sock *)msk;
344 struct sk_buff *tail;
347 __skb_unlink(skb, &ssk->sk_receive_queue);
352 /* try to fetch required memory from subflow */
353 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
356 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
358 /* the skb map_seq accounts for the skb offset:
359 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
362 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
363 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
364 MPTCP_SKB_CB(skb)->offset = offset;
365 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
367 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
369 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
370 tail = skb_peek_tail(&sk->sk_receive_queue);
371 if (tail && mptcp_try_coalesce(sk, tail, skb))
374 mptcp_set_owner_r(skb, sk);
375 __skb_queue_tail(&sk->sk_receive_queue, skb);
377 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
378 mptcp_data_queue_ofo(msk, skb);
382 /* old data, keep it simple and drop the whole pkt, sender
383 * will retransmit as needed, if needed.
385 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
391 static void mptcp_stop_timer(struct sock *sk)
393 struct inet_connection_sock *icsk = inet_csk(sk);
395 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
396 mptcp_sk(sk)->timer_ival = 0;
399 static void mptcp_close_wake_up(struct sock *sk)
401 if (sock_flag(sk, SOCK_DEAD))
404 sk->sk_state_change(sk);
405 if (sk->sk_shutdown == SHUTDOWN_MASK ||
406 sk->sk_state == TCP_CLOSE)
407 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
409 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
412 static bool mptcp_pending_data_fin_ack(struct sock *sk)
414 struct mptcp_sock *msk = mptcp_sk(sk);
416 return !__mptcp_check_fallback(msk) &&
417 ((1 << sk->sk_state) &
418 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
419 msk->write_seq == READ_ONCE(msk->snd_una);
422 static void mptcp_check_data_fin_ack(struct sock *sk)
424 struct mptcp_sock *msk = mptcp_sk(sk);
426 /* Look for an acknowledged DATA_FIN */
427 if (mptcp_pending_data_fin_ack(sk)) {
428 WRITE_ONCE(msk->snd_data_fin_enable, 0);
430 switch (sk->sk_state) {
432 inet_sk_state_store(sk, TCP_FIN_WAIT2);
436 inet_sk_state_store(sk, TCP_CLOSE);
440 mptcp_close_wake_up(sk);
444 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
446 struct mptcp_sock *msk = mptcp_sk(sk);
448 if (READ_ONCE(msk->rcv_data_fin) &&
449 ((1 << sk->sk_state) &
450 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
451 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
453 if (msk->ack_seq == rcv_data_fin_seq) {
455 *seq = rcv_data_fin_seq;
464 static void mptcp_set_datafin_timeout(const struct sock *sk)
466 struct inet_connection_sock *icsk = inet_csk(sk);
468 mptcp_sk(sk)->timer_ival = min(TCP_RTO_MAX,
469 TCP_RTO_MIN << icsk->icsk_retransmits);
472 static void __mptcp_set_timeout(struct sock *sk, long tout)
474 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
477 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
479 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
481 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
482 inet_csk(ssk)->icsk_timeout - jiffies : 0;
485 static void mptcp_set_timeout(struct sock *sk)
487 struct mptcp_subflow_context *subflow;
490 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
491 tout = max(tout, mptcp_timeout_from_subflow(subflow));
492 __mptcp_set_timeout(sk, tout);
495 static bool tcp_can_send_ack(const struct sock *ssk)
497 return !((1 << inet_sk_state_load(ssk)) &
498 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
501 void mptcp_subflow_send_ack(struct sock *ssk)
505 slow = lock_sock_fast(ssk);
506 if (tcp_can_send_ack(ssk))
508 unlock_sock_fast(ssk, slow);
511 static void mptcp_send_ack(struct mptcp_sock *msk)
513 struct mptcp_subflow_context *subflow;
515 mptcp_for_each_subflow(msk, subflow)
516 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
519 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
523 slow = lock_sock_fast(ssk);
524 if (tcp_can_send_ack(ssk))
525 tcp_cleanup_rbuf(ssk, 1);
526 unlock_sock_fast(ssk, slow);
529 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
531 const struct inet_connection_sock *icsk = inet_csk(ssk);
532 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
533 const struct tcp_sock *tp = tcp_sk(ssk);
535 return (ack_pending & ICSK_ACK_SCHED) &&
536 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
537 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
538 (rx_empty && ack_pending &
539 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
542 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
544 int old_space = READ_ONCE(msk->old_wspace);
545 struct mptcp_subflow_context *subflow;
546 struct sock *sk = (struct sock *)msk;
547 int space = __mptcp_space(sk);
548 bool cleanup, rx_empty;
550 cleanup = (space > 0) && (space >= (old_space << 1));
551 rx_empty = !__mptcp_rmem(sk);
553 mptcp_for_each_subflow(msk, subflow) {
554 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
556 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
557 mptcp_subflow_cleanup_rbuf(ssk);
561 static bool mptcp_check_data_fin(struct sock *sk)
563 struct mptcp_sock *msk = mptcp_sk(sk);
564 u64 rcv_data_fin_seq;
567 if (__mptcp_check_fallback(msk))
570 /* Need to ack a DATA_FIN received from a peer while this side
571 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
572 * msk->rcv_data_fin was set when parsing the incoming options
573 * at the subflow level and the msk lock was not held, so this
574 * is the first opportunity to act on the DATA_FIN and change
577 * If we are caught up to the sequence number of the incoming
578 * DATA_FIN, send the DATA_ACK now and do state transition. If
579 * not caught up, do nothing and let the recv code send DATA_ACK
583 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
584 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
585 WRITE_ONCE(msk->rcv_data_fin, 0);
587 sk->sk_shutdown |= RCV_SHUTDOWN;
588 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
590 switch (sk->sk_state) {
591 case TCP_ESTABLISHED:
592 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
595 inet_sk_state_store(sk, TCP_CLOSING);
598 inet_sk_state_store(sk, TCP_CLOSE);
601 /* Other states not expected */
608 mptcp_close_wake_up(sk);
613 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
617 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
618 struct sock *sk = (struct sock *)msk;
619 unsigned int moved = 0;
620 bool more_data_avail;
625 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
627 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
628 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
630 if (unlikely(ssk_rbuf > sk_rbuf)) {
631 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
636 pr_debug("msk=%p ssk=%p", msk, ssk);
639 u32 map_remaining, offset;
640 u32 seq = tp->copied_seq;
644 /* try to move as much data as available */
645 map_remaining = subflow->map_data_len -
646 mptcp_subflow_get_map_offset(subflow);
648 skb = skb_peek(&ssk->sk_receive_queue);
650 /* if no data is found, a racing workqueue/recvmsg
651 * already processed the new data, stop here or we
652 * can enter an infinite loop
659 if (__mptcp_check_fallback(msk)) {
660 /* if we are running under the workqueue, TCP could have
661 * collapsed skbs between dummy map creation and now
662 * be sure to adjust the size
664 map_remaining = skb->len;
665 subflow->map_data_len = skb->len;
668 offset = seq - TCP_SKB_CB(skb)->seq;
669 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
675 if (offset < skb->len) {
676 size_t len = skb->len - offset;
681 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
685 if (WARN_ON_ONCE(map_remaining < len))
689 sk_eat_skb(ssk, skb);
693 WRITE_ONCE(tp->copied_seq, seq);
694 more_data_avail = mptcp_subflow_data_available(ssk);
696 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
700 } while (more_data_avail);
706 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
708 struct sock *sk = (struct sock *)msk;
709 struct sk_buff *skb, *tail;
714 p = rb_first(&msk->out_of_order_queue);
715 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
718 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
722 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
724 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
727 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
731 end_seq = MPTCP_SKB_CB(skb)->end_seq;
732 tail = skb_peek_tail(&sk->sk_receive_queue);
733 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
734 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
736 /* skip overlapping data, if any */
737 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
738 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
740 MPTCP_SKB_CB(skb)->offset += delta;
741 __skb_queue_tail(&sk->sk_receive_queue, skb);
743 msk->ack_seq = end_seq;
749 /* In most cases we will be able to lock the mptcp socket. If its already
750 * owned, we need to defer to the work queue to avoid ABBA deadlock.
752 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
754 struct sock *sk = (struct sock *)msk;
755 unsigned int moved = 0;
757 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
758 __mptcp_ofo_queue(msk);
759 if (unlikely(ssk->sk_err)) {
760 if (!sock_owned_by_user(sk))
761 __mptcp_error_report(sk);
763 set_bit(MPTCP_ERROR_REPORT, &msk->flags);
766 /* If the moves have caught up with the DATA_FIN sequence number
767 * it's time to ack the DATA_FIN and change socket state, but
768 * this is not a good place to change state. Let the workqueue
771 if (mptcp_pending_data_fin(sk, NULL))
772 mptcp_schedule_work(sk);
776 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
778 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
779 struct mptcp_sock *msk = mptcp_sk(sk);
780 int sk_rbuf, ssk_rbuf;
782 /* The peer can send data while we are shutting down this
783 * subflow at msk destruction time, but we must avoid enqueuing
784 * more data to the msk receive queue
786 if (unlikely(subflow->disposable))
789 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
790 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
791 if (unlikely(ssk_rbuf > sk_rbuf))
794 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
795 if (__mptcp_rmem(sk) > sk_rbuf) {
796 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
800 /* Wake-up the reader only for in-sequence data */
802 if (move_skbs_to_msk(msk, ssk))
803 sk->sk_data_ready(sk);
805 mptcp_data_unlock(sk);
808 static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
810 struct mptcp_subflow_context *subflow;
813 if (likely(list_empty(&msk->join_list)))
816 spin_lock_bh(&msk->join_list_lock);
817 list_for_each_entry(subflow, &msk->join_list, node) {
818 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
820 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
821 if (READ_ONCE(msk->setsockopt_seq) != sseq)
824 list_splice_tail_init(&msk->join_list, &msk->conn_list);
825 spin_unlock_bh(&msk->join_list_lock);
830 void __mptcp_flush_join_list(struct mptcp_sock *msk)
832 if (likely(!mptcp_do_flush_join_list(msk)))
835 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
836 mptcp_schedule_work((struct sock *)msk);
839 static void mptcp_flush_join_list(struct mptcp_sock *msk)
841 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
845 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
848 mptcp_sockopt_sync_all(msk);
851 static bool mptcp_timer_pending(struct sock *sk)
853 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
856 static void mptcp_reset_timer(struct sock *sk)
858 struct inet_connection_sock *icsk = inet_csk(sk);
861 /* prevent rescheduling on close */
862 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
865 tout = mptcp_sk(sk)->timer_ival;
866 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
869 bool mptcp_schedule_work(struct sock *sk)
871 if (inet_sk_state_load(sk) != TCP_CLOSE &&
872 schedule_work(&mptcp_sk(sk)->work)) {
873 /* each subflow already holds a reference to the sk, and the
874 * workqueue is invoked by a subflow, so sk can't go away here.
882 void mptcp_subflow_eof(struct sock *sk)
884 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
885 mptcp_schedule_work(sk);
888 static void mptcp_check_for_eof(struct mptcp_sock *msk)
890 struct mptcp_subflow_context *subflow;
891 struct sock *sk = (struct sock *)msk;
894 mptcp_for_each_subflow(msk, subflow)
895 receivers += !subflow->rx_eof;
899 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
900 /* hopefully temporary hack: propagate shutdown status
901 * to msk, when all subflows agree on it
903 sk->sk_shutdown |= RCV_SHUTDOWN;
905 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
906 sk->sk_data_ready(sk);
909 switch (sk->sk_state) {
910 case TCP_ESTABLISHED:
911 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
914 inet_sk_state_store(sk, TCP_CLOSING);
917 inet_sk_state_store(sk, TCP_CLOSE);
922 mptcp_close_wake_up(sk);
925 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
927 struct mptcp_subflow_context *subflow;
928 struct sock *sk = (struct sock *)msk;
930 sock_owned_by_me(sk);
932 mptcp_for_each_subflow(msk, subflow) {
933 if (READ_ONCE(subflow->data_avail))
934 return mptcp_subflow_tcp_sock(subflow);
940 static bool mptcp_skb_can_collapse_to(u64 write_seq,
941 const struct sk_buff *skb,
942 const struct mptcp_ext *mpext)
944 if (!tcp_skb_can_collapse_to(skb))
947 /* can collapse only if MPTCP level sequence is in order and this
948 * mapping has not been xmitted yet
950 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
954 /* we can append data to the given data frag if:
955 * - there is space available in the backing page_frag
956 * - the data frag tail matches the current page_frag free offset
957 * - the data frag end sequence number matches the current write seq
959 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
960 const struct page_frag *pfrag,
961 const struct mptcp_data_frag *df)
963 return df && pfrag->page == df->page &&
964 pfrag->size - pfrag->offset > 0 &&
965 pfrag->offset == (df->offset + df->data_len) &&
966 df->data_seq + df->data_len == msk->write_seq;
969 static void __mptcp_mem_reclaim_partial(struct sock *sk)
971 int reclaimable = mptcp_sk(sk)->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
973 lockdep_assert_held_once(&sk->sk_lock.slock);
975 __mptcp_rmem_reclaim(sk, reclaimable - 1);
976 sk_mem_reclaim_partial(sk);
979 static void mptcp_mem_reclaim_partial(struct sock *sk)
982 __mptcp_mem_reclaim_partial(sk);
983 mptcp_data_unlock(sk);
986 static void dfrag_uncharge(struct sock *sk, int len)
988 sk_mem_uncharge(sk, len);
989 sk_wmem_queued_add(sk, -len);
992 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
994 int len = dfrag->data_len + dfrag->overhead;
996 list_del(&dfrag->list);
997 dfrag_uncharge(sk, len);
998 put_page(dfrag->page);
1001 static void __mptcp_clean_una(struct sock *sk)
1003 struct mptcp_sock *msk = mptcp_sk(sk);
1004 struct mptcp_data_frag *dtmp, *dfrag;
1005 bool cleaned = false;
1008 /* on fallback we just need to ignore snd_una, as this is really
1011 if (__mptcp_check_fallback(msk))
1012 msk->snd_una = READ_ONCE(msk->snd_nxt);
1014 snd_una = msk->snd_una;
1015 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1016 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1019 if (unlikely(dfrag == msk->first_pending)) {
1020 /* in recovery mode can see ack after the current snd head */
1021 if (WARN_ON_ONCE(!msk->recovery))
1024 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1027 dfrag_clear(sk, dfrag);
1031 dfrag = mptcp_rtx_head(sk);
1032 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1033 u64 delta = snd_una - dfrag->data_seq;
1035 /* prevent wrap around in recovery mode */
1036 if (unlikely(delta > dfrag->already_sent)) {
1037 if (WARN_ON_ONCE(!msk->recovery))
1039 if (WARN_ON_ONCE(delta > dfrag->data_len))
1041 dfrag->already_sent += delta - dfrag->already_sent;
1044 dfrag->data_seq += delta;
1045 dfrag->offset += delta;
1046 dfrag->data_len -= delta;
1047 dfrag->already_sent -= delta;
1049 dfrag_uncharge(sk, delta);
1053 /* all retransmitted data acked, recovery completed */
1054 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1055 msk->recovery = false;
1058 if (cleaned && tcp_under_memory_pressure(sk))
1059 __mptcp_mem_reclaim_partial(sk);
1061 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1062 snd_una == READ_ONCE(msk->write_seq)) {
1063 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1064 mptcp_stop_timer(sk);
1066 mptcp_reset_timer(sk);
1070 static void __mptcp_clean_una_wakeup(struct sock *sk)
1072 lockdep_assert_held_once(&sk->sk_lock.slock);
1074 __mptcp_clean_una(sk);
1075 mptcp_write_space(sk);
1078 static void mptcp_clean_una_wakeup(struct sock *sk)
1080 mptcp_data_lock(sk);
1081 __mptcp_clean_una_wakeup(sk);
1082 mptcp_data_unlock(sk);
1085 static void mptcp_enter_memory_pressure(struct sock *sk)
1087 struct mptcp_subflow_context *subflow;
1088 struct mptcp_sock *msk = mptcp_sk(sk);
1091 sk_stream_moderate_sndbuf(sk);
1092 mptcp_for_each_subflow(msk, subflow) {
1093 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1096 tcp_enter_memory_pressure(ssk);
1097 sk_stream_moderate_sndbuf(ssk);
1102 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1105 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1107 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1108 pfrag, sk->sk_allocation)))
1111 mptcp_enter_memory_pressure(sk);
1115 static struct mptcp_data_frag *
1116 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1119 int offset = ALIGN(orig_offset, sizeof(long));
1120 struct mptcp_data_frag *dfrag;
1122 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1123 dfrag->data_len = 0;
1124 dfrag->data_seq = msk->write_seq;
1125 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1126 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1127 dfrag->already_sent = 0;
1128 dfrag->page = pfrag->page;
1133 struct mptcp_sendmsg_info {
1139 bool data_lock_held;
1142 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1145 u64 window_end = mptcp_wnd_end(msk);
1147 if (__mptcp_check_fallback(msk))
1150 if (!before64(data_seq + avail_size, window_end)) {
1151 u64 allowed_size = window_end - data_seq;
1153 return min_t(unsigned int, allowed_size, avail_size);
1159 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1161 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1165 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1169 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1171 struct sk_buff *skb;
1173 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1175 if (likely(__mptcp_add_ext(skb, gfp))) {
1176 skb_reserve(skb, MAX_TCP_HEADER);
1177 skb->ip_summed = CHECKSUM_PARTIAL;
1178 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1183 mptcp_enter_memory_pressure(sk);
1188 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1190 struct sk_buff *skb;
1192 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1196 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1197 tcp_skb_entail(ssk, skb);
1204 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1206 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1208 if (unlikely(tcp_under_memory_pressure(sk))) {
1210 __mptcp_mem_reclaim_partial(sk);
1212 mptcp_mem_reclaim_partial(sk);
1214 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1217 /* note: this always recompute the csum on the whole skb, even
1218 * if we just appended a single frag. More status info needed
1220 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1222 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1223 __wsum csum = ~csum_unfold(mpext->csum);
1224 int offset = skb->len - added;
1226 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1229 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1230 struct mptcp_data_frag *dfrag,
1231 struct mptcp_sendmsg_info *info)
1233 u64 data_seq = dfrag->data_seq + info->sent;
1234 int offset = dfrag->offset + info->sent;
1235 struct mptcp_sock *msk = mptcp_sk(sk);
1236 bool zero_window_probe = false;
1237 struct mptcp_ext *mpext = NULL;
1238 bool can_coalesce = false;
1239 bool reuse_skb = true;
1240 struct sk_buff *skb;
1244 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1245 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1247 if (WARN_ON_ONCE(info->sent > info->limit ||
1248 info->limit > dfrag->data_len))
1251 /* compute send limit */
1252 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1253 copy = info->size_goal;
1255 skb = tcp_write_queue_tail(ssk);
1256 if (skb && copy > skb->len) {
1257 /* Limit the write to the size available in the
1258 * current skb, if any, so that we create at most a new skb.
1259 * Explicitly tells TCP internals to avoid collapsing on later
1260 * queue management operation, to avoid breaking the ext <->
1261 * SSN association set here
1263 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1264 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1265 TCP_SKB_CB(skb)->eor = 1;
1269 i = skb_shinfo(skb)->nr_frags;
1270 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1271 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1272 tcp_mark_push(tcp_sk(ssk), skb);
1279 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1283 i = skb_shinfo(skb)->nr_frags;
1285 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1288 /* Zero window and all data acked? Probe. */
1289 copy = mptcp_check_allowed_size(msk, data_seq, copy);
1291 u64 snd_una = READ_ONCE(msk->snd_una);
1293 if (snd_una != msk->snd_nxt) {
1294 tcp_remove_empty_skb(ssk);
1298 zero_window_probe = true;
1299 data_seq = snd_una - 1;
1302 /* all mptcp-level data is acked, no skbs should be present into the
1305 WARN_ON_ONCE(reuse_skb);
1308 copy = min_t(size_t, copy, info->limit - info->sent);
1309 if (!sk_wmem_schedule(ssk, copy)) {
1310 tcp_remove_empty_skb(ssk);
1315 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1317 get_page(dfrag->page);
1318 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1322 skb->data_len += copy;
1323 skb->truesize += copy;
1324 sk_wmem_queued_add(ssk, copy);
1325 sk_mem_charge(ssk, copy);
1326 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1327 TCP_SKB_CB(skb)->end_seq += copy;
1328 tcp_skb_pcount_set(skb, 0);
1330 /* on skb reuse we just need to update the DSS len */
1332 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1333 mpext->data_len += copy;
1334 WARN_ON_ONCE(zero_window_probe);
1338 memset(mpext, 0, sizeof(*mpext));
1339 mpext->data_seq = data_seq;
1340 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1341 mpext->data_len = copy;
1345 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1346 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1349 if (zero_window_probe) {
1350 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1352 if (READ_ONCE(msk->csum_enabled))
1353 mptcp_update_data_checksum(skb, copy);
1354 tcp_push_pending_frames(ssk);
1358 if (READ_ONCE(msk->csum_enabled))
1359 mptcp_update_data_checksum(skb, copy);
1360 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1364 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1365 sizeof(struct tcphdr) - \
1366 MAX_TCP_OPTION_SPACE - \
1367 sizeof(struct ipv6hdr) - \
1368 sizeof(struct frag_hdr))
1370 struct subflow_send_info {
1375 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1377 if (!subflow->stale)
1381 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1384 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1386 if (unlikely(subflow->stale)) {
1387 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1389 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1392 mptcp_subflow_set_active(subflow);
1394 return __mptcp_subflow_active(subflow);
1397 /* implement the mptcp packet scheduler;
1398 * returns the subflow that will transmit the next DSS
1399 * additionally updates the rtx timeout
1401 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1403 struct subflow_send_info send_info[2];
1404 struct mptcp_subflow_context *subflow;
1405 struct sock *sk = (struct sock *)msk;
1406 int i, nr_active = 0;
1412 sock_owned_by_me(sk);
1414 if (__mptcp_check_fallback(msk)) {
1417 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1420 /* re-use last subflow, if the burst allow that */
1421 if (msk->last_snd && msk->snd_burst > 0 &&
1422 sk_stream_memory_free(msk->last_snd) &&
1423 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1424 mptcp_set_timeout(sk);
1425 return msk->last_snd;
1428 /* pick the subflow with the lower wmem/wspace ratio */
1429 for (i = 0; i < 2; ++i) {
1430 send_info[i].ssk = NULL;
1431 send_info[i].ratio = -1;
1433 mptcp_for_each_subflow(msk, subflow) {
1434 trace_mptcp_subflow_get_send(subflow);
1435 ssk = mptcp_subflow_tcp_sock(subflow);
1436 if (!mptcp_subflow_active(subflow))
1439 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1440 nr_active += !subflow->backup;
1441 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1444 pace = READ_ONCE(ssk->sk_pacing_rate);
1448 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1450 if (ratio < send_info[subflow->backup].ratio) {
1451 send_info[subflow->backup].ssk = ssk;
1452 send_info[subflow->backup].ratio = ratio;
1455 __mptcp_set_timeout(sk, tout);
1457 /* pick the best backup if no other subflow is active */
1459 send_info[0].ssk = send_info[1].ssk;
1461 if (send_info[0].ssk) {
1462 msk->last_snd = send_info[0].ssk;
1463 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1464 tcp_sk(msk->last_snd)->snd_wnd);
1465 return msk->last_snd;
1471 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1473 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1477 static void mptcp_update_post_push(struct mptcp_sock *msk,
1478 struct mptcp_data_frag *dfrag,
1481 u64 snd_nxt_new = dfrag->data_seq;
1483 dfrag->already_sent += sent;
1485 msk->snd_burst -= sent;
1487 snd_nxt_new += dfrag->already_sent;
1489 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1490 * is recovering after a failover. In that event, this re-sends
1493 * Thus compute snd_nxt_new candidate based on
1494 * the dfrag->data_seq that was sent and the data
1495 * that has been handed to the subflow for transmission
1496 * and skip update in case it was old dfrag.
1498 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1499 msk->snd_nxt = snd_nxt_new;
1502 static void mptcp_check_and_set_pending(struct sock *sk)
1504 if (mptcp_send_head(sk) &&
1505 !test_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
1506 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
1509 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1511 struct sock *prev_ssk = NULL, *ssk = NULL;
1512 struct mptcp_sock *msk = mptcp_sk(sk);
1513 struct mptcp_sendmsg_info info = {
1516 struct mptcp_data_frag *dfrag;
1517 int len, copied = 0;
1519 while ((dfrag = mptcp_send_head(sk))) {
1520 info.sent = dfrag->already_sent;
1521 info.limit = dfrag->data_len;
1522 len = dfrag->data_len - dfrag->already_sent;
1527 mptcp_flush_join_list(msk);
1528 ssk = mptcp_subflow_get_send(msk);
1530 /* First check. If the ssk has changed since
1531 * the last round, release prev_ssk
1533 if (ssk != prev_ssk && prev_ssk)
1534 mptcp_push_release(prev_ssk, &info);
1538 /* Need to lock the new subflow only if different
1539 * from the previous one, otherwise we are still
1540 * helding the relevant lock
1542 if (ssk != prev_ssk)
1545 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1547 mptcp_push_release(ssk, &info);
1555 mptcp_update_post_push(msk, dfrag, ret);
1557 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1560 /* at this point we held the socket lock for the last subflow we used */
1562 mptcp_push_release(ssk, &info);
1565 /* ensure the rtx timer is running */
1566 if (!mptcp_timer_pending(sk))
1567 mptcp_reset_timer(sk);
1569 __mptcp_check_send_data_fin(sk);
1572 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1574 struct mptcp_sock *msk = mptcp_sk(sk);
1575 struct mptcp_sendmsg_info info = {
1576 .data_lock_held = true,
1578 struct mptcp_data_frag *dfrag;
1579 struct sock *xmit_ssk;
1580 int len, copied = 0;
1584 while ((dfrag = mptcp_send_head(sk))) {
1585 info.sent = dfrag->already_sent;
1586 info.limit = dfrag->data_len;
1587 len = dfrag->data_len - dfrag->already_sent;
1591 /* the caller already invoked the packet scheduler,
1592 * check for a different subflow usage only after
1593 * spooling the first chunk of data
1595 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1598 if (xmit_ssk != ssk) {
1599 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1600 MPTCP_DELEGATE_SEND);
1604 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1613 mptcp_update_post_push(msk, dfrag, ret);
1615 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1619 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1620 * not going to flush it via release_sock()
1623 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1625 if (!mptcp_timer_pending(sk))
1626 mptcp_reset_timer(sk);
1628 if (msk->snd_data_fin_enable &&
1629 msk->snd_nxt + 1 == msk->write_seq)
1630 mptcp_schedule_work(sk);
1634 static void mptcp_set_nospace(struct sock *sk)
1636 /* enable autotune */
1637 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1639 /* will be cleared on avail space */
1640 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1643 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1645 struct mptcp_sock *msk = mptcp_sk(sk);
1646 struct page_frag *pfrag;
1651 /* we don't support FASTOPEN yet */
1652 if (msg->msg_flags & MSG_FASTOPEN)
1655 /* silently ignore everything else */
1656 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1660 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1662 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1663 ret = sk_stream_wait_connect(sk, &timeo);
1668 pfrag = sk_page_frag(sk);
1670 while (msg_data_left(msg)) {
1671 int total_ts, frag_truesize = 0;
1672 struct mptcp_data_frag *dfrag;
1673 bool dfrag_collapsed;
1674 size_t psize, offset;
1676 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1681 /* reuse tail pfrag, if possible, or carve a new one from the
1684 dfrag = mptcp_pending_tail(sk);
1685 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1686 if (!dfrag_collapsed) {
1687 if (!sk_stream_memory_free(sk))
1688 goto wait_for_memory;
1690 if (!mptcp_page_frag_refill(sk, pfrag))
1691 goto wait_for_memory;
1693 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1694 frag_truesize = dfrag->overhead;
1697 /* we do not bound vs wspace, to allow a single packet.
1698 * memory accounting will prevent execessive memory usage
1701 offset = dfrag->offset + dfrag->data_len;
1702 psize = pfrag->size - offset;
1703 psize = min_t(size_t, psize, msg_data_left(msg));
1704 total_ts = psize + frag_truesize;
1706 if (!sk_wmem_schedule(sk, total_ts))
1707 goto wait_for_memory;
1709 if (copy_page_from_iter(dfrag->page, offset, psize,
1710 &msg->msg_iter) != psize) {
1715 /* data successfully copied into the write queue */
1716 sk->sk_forward_alloc -= total_ts;
1718 dfrag->data_len += psize;
1719 frag_truesize += psize;
1720 pfrag->offset += frag_truesize;
1721 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1723 /* charge data on mptcp pending queue to the msk socket
1724 * Note: we charge such data both to sk and ssk
1726 sk_wmem_queued_add(sk, frag_truesize);
1727 if (!dfrag_collapsed) {
1728 get_page(dfrag->page);
1729 list_add_tail(&dfrag->list, &msk->rtx_queue);
1730 if (!msk->first_pending)
1731 WRITE_ONCE(msk->first_pending, dfrag);
1733 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1734 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1740 mptcp_set_nospace(sk);
1741 __mptcp_push_pending(sk, msg->msg_flags);
1742 ret = sk_stream_wait_memory(sk, &timeo);
1748 __mptcp_push_pending(sk, msg->msg_flags);
1752 return copied ? : ret;
1755 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1757 size_t len, int flags,
1758 struct scm_timestamping_internal *tss,
1761 struct sk_buff *skb, *tmp;
1764 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1765 u32 offset = MPTCP_SKB_CB(skb)->offset;
1766 u32 data_len = skb->len - offset;
1767 u32 count = min_t(size_t, len - copied, data_len);
1770 if (!(flags & MSG_TRUNC)) {
1771 err = skb_copy_datagram_msg(skb, offset, msg, count);
1772 if (unlikely(err < 0)) {
1779 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1780 tcp_update_recv_tstamps(skb, tss);
1781 *cmsg_flags |= MPTCP_CMSG_TS;
1786 if (count < data_len) {
1787 if (!(flags & MSG_PEEK))
1788 MPTCP_SKB_CB(skb)->offset += count;
1792 if (!(flags & MSG_PEEK)) {
1793 /* we will bulk release the skb memory later */
1794 skb->destructor = NULL;
1795 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1796 __skb_unlink(skb, &msk->receive_queue);
1807 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1809 * Only difference: Use highest rtt estimate of the subflows in use.
1811 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1813 struct mptcp_subflow_context *subflow;
1814 struct sock *sk = (struct sock *)msk;
1815 u32 time, advmss = 1;
1818 sock_owned_by_me(sk);
1823 msk->rcvq_space.copied += copied;
1825 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1826 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1828 rtt_us = msk->rcvq_space.rtt_us;
1829 if (rtt_us && time < (rtt_us >> 3))
1833 mptcp_for_each_subflow(msk, subflow) {
1834 const struct tcp_sock *tp;
1838 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1840 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1841 sf_advmss = READ_ONCE(tp->advmss);
1843 rtt_us = max(sf_rtt_us, rtt_us);
1844 advmss = max(sf_advmss, advmss);
1847 msk->rcvq_space.rtt_us = rtt_us;
1848 if (time < (rtt_us >> 3) || rtt_us == 0)
1851 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1854 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1855 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1859 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1861 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1863 do_div(grow, msk->rcvq_space.space);
1864 rcvwin += (grow << 1);
1866 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1867 while (tcp_win_from_space(sk, rcvmem) < advmss)
1870 do_div(rcvwin, advmss);
1871 rcvbuf = min_t(u64, rcvwin * rcvmem,
1872 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1874 if (rcvbuf > sk->sk_rcvbuf) {
1877 window_clamp = tcp_win_from_space(sk, rcvbuf);
1878 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1880 /* Make subflows follow along. If we do not do this, we
1881 * get drops at subflow level if skbs can't be moved to
1882 * the mptcp rx queue fast enough (announced rcv_win can
1883 * exceed ssk->sk_rcvbuf).
1885 mptcp_for_each_subflow(msk, subflow) {
1889 ssk = mptcp_subflow_tcp_sock(subflow);
1890 slow = lock_sock_fast(ssk);
1891 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1892 tcp_sk(ssk)->window_clamp = window_clamp;
1893 tcp_cleanup_rbuf(ssk, 1);
1894 unlock_sock_fast(ssk, slow);
1899 msk->rcvq_space.space = msk->rcvq_space.copied;
1901 msk->rcvq_space.copied = 0;
1902 msk->rcvq_space.time = mstamp;
1905 static void __mptcp_update_rmem(struct sock *sk)
1907 struct mptcp_sock *msk = mptcp_sk(sk);
1909 if (!msk->rmem_released)
1912 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1913 mptcp_rmem_uncharge(sk, msk->rmem_released);
1914 WRITE_ONCE(msk->rmem_released, 0);
1917 static void __mptcp_splice_receive_queue(struct sock *sk)
1919 struct mptcp_sock *msk = mptcp_sk(sk);
1921 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1924 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1926 struct sock *sk = (struct sock *)msk;
1927 unsigned int moved = 0;
1930 mptcp_flush_join_list(msk);
1932 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1935 /* we can have data pending in the subflows only if the msk
1936 * receive buffer was full at subflow_data_ready() time,
1937 * that is an unlikely slow path.
1942 slowpath = lock_sock_fast(ssk);
1943 mptcp_data_lock(sk);
1944 __mptcp_update_rmem(sk);
1945 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1946 mptcp_data_unlock(sk);
1948 if (unlikely(ssk->sk_err))
1949 __mptcp_error_report(sk);
1950 unlock_sock_fast(ssk, slowpath);
1953 /* acquire the data lock only if some input data is pending */
1955 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1956 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1957 mptcp_data_lock(sk);
1958 __mptcp_update_rmem(sk);
1959 ret |= __mptcp_ofo_queue(msk);
1960 __mptcp_splice_receive_queue(sk);
1961 mptcp_data_unlock(sk);
1964 mptcp_check_data_fin((struct sock *)msk);
1965 return !skb_queue_empty(&msk->receive_queue);
1968 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1969 int nonblock, int flags, int *addr_len)
1971 struct mptcp_sock *msk = mptcp_sk(sk);
1972 struct scm_timestamping_internal tss;
1973 int copied = 0, cmsg_flags = 0;
1977 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
1978 if (unlikely(flags & MSG_ERRQUEUE))
1979 return inet_recv_error(sk, msg, len, addr_len);
1982 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1987 timeo = sock_rcvtimeo(sk, nonblock);
1989 len = min_t(size_t, len, INT_MAX);
1990 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1992 while (copied < len) {
1995 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
1996 if (unlikely(bytes_read < 0)) {
1998 copied = bytes_read;
2002 copied += bytes_read;
2004 /* be sure to advertise window change */
2005 mptcp_cleanup_rbuf(msk);
2007 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2010 /* only the master socket status is relevant here. The exit
2011 * conditions mirror closely tcp_recvmsg()
2013 if (copied >= target)
2018 sk->sk_state == TCP_CLOSE ||
2019 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2021 signal_pending(current))
2025 copied = sock_error(sk);
2029 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2030 mptcp_check_for_eof(msk);
2032 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2033 /* race breaker: the shutdown could be after the
2034 * previous receive queue check
2036 if (__mptcp_move_skbs(msk))
2041 if (sk->sk_state == TCP_CLOSE) {
2051 if (signal_pending(current)) {
2052 copied = sock_intr_errno(timeo);
2057 pr_debug("block timeout %ld", timeo);
2058 sk_wait_data(sk, &timeo, NULL);
2062 if (cmsg_flags && copied >= 0) {
2063 if (cmsg_flags & MPTCP_CMSG_TS)
2064 tcp_recv_timestamp(msg, sk, &tss);
2067 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2068 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2069 skb_queue_empty(&msk->receive_queue), copied);
2070 if (!(flags & MSG_PEEK))
2071 mptcp_rcv_space_adjust(msk, copied);
2077 static void mptcp_retransmit_timer(struct timer_list *t)
2079 struct inet_connection_sock *icsk = from_timer(icsk, t,
2080 icsk_retransmit_timer);
2081 struct sock *sk = &icsk->icsk_inet.sk;
2082 struct mptcp_sock *msk = mptcp_sk(sk);
2085 if (!sock_owned_by_user(sk)) {
2086 /* we need a process context to retransmit */
2087 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2088 mptcp_schedule_work(sk);
2090 /* delegate our work to tcp_release_cb() */
2091 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2097 static void mptcp_timeout_timer(struct timer_list *t)
2099 struct sock *sk = from_timer(sk, t, sk_timer);
2101 mptcp_schedule_work(sk);
2105 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2108 * A backup subflow is returned only if that is the only kind available.
2110 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2112 struct sock *backup = NULL, *pick = NULL;
2113 struct mptcp_subflow_context *subflow;
2114 int min_stale_count = INT_MAX;
2116 sock_owned_by_me((const struct sock *)msk);
2118 if (__mptcp_check_fallback(msk))
2121 mptcp_for_each_subflow(msk, subflow) {
2122 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2124 if (!__mptcp_subflow_active(subflow))
2127 /* still data outstanding at TCP level? skip this */
2128 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2129 mptcp_pm_subflow_chk_stale(msk, ssk);
2130 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2134 if (subflow->backup) {
2147 /* use backup only if there are no progresses anywhere */
2148 return min_stale_count > 1 ? backup : NULL;
2151 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2154 iput(SOCK_INODE(msk->subflow));
2155 msk->subflow = NULL;
2159 bool __mptcp_retransmit_pending_data(struct sock *sk)
2161 struct mptcp_data_frag *cur, *rtx_head;
2162 struct mptcp_sock *msk = mptcp_sk(sk);
2164 if (__mptcp_check_fallback(mptcp_sk(sk)))
2167 if (tcp_rtx_and_write_queues_empty(sk))
2170 /* the closing socket has some data untransmitted and/or unacked:
2171 * some data in the mptcp rtx queue has not really xmitted yet.
2172 * keep it simple and re-inject the whole mptcp level rtx queue
2174 mptcp_data_lock(sk);
2175 __mptcp_clean_una_wakeup(sk);
2176 rtx_head = mptcp_rtx_head(sk);
2178 mptcp_data_unlock(sk);
2182 msk->recovery_snd_nxt = msk->snd_nxt;
2183 msk->recovery = true;
2184 mptcp_data_unlock(sk);
2186 msk->first_pending = rtx_head;
2189 /* be sure to clear the "sent status" on all re-injected fragments */
2190 list_for_each_entry(cur, &msk->rtx_queue, list) {
2191 if (!cur->already_sent)
2193 cur->already_sent = 0;
2199 /* subflow sockets can be either outgoing (connect) or incoming
2202 * Outgoing subflows use in-kernel sockets.
2203 * Incoming subflows do not have their own 'struct socket' allocated,
2204 * so we need to use tcp_close() after detaching them from the mptcp
2207 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2208 struct mptcp_subflow_context *subflow)
2210 struct mptcp_sock *msk = mptcp_sk(sk);
2213 list_del(&subflow->node);
2215 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2217 /* if we are invoked by the msk cleanup code, the subflow is
2223 need_push = __mptcp_retransmit_pending_data(sk);
2224 subflow->disposable = 1;
2226 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2227 * the ssk has been already destroyed, we just need to release the
2228 * reference owned by msk;
2230 if (!inet_csk(ssk)->icsk_ulp_ops) {
2231 kfree_rcu(subflow, rcu);
2233 /* otherwise tcp will dispose of the ssk and subflow ctx */
2234 __tcp_close(ssk, 0);
2236 /* close acquired an extra ref */
2243 if (ssk == msk->last_snd)
2244 msk->last_snd = NULL;
2246 if (ssk == msk->first)
2249 if (msk->subflow && ssk == msk->subflow->sk)
2250 mptcp_dispose_initial_subflow(msk);
2253 __mptcp_push_pending(sk, 0);
2256 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2257 struct mptcp_subflow_context *subflow)
2259 if (sk->sk_state == TCP_ESTABLISHED)
2260 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2261 __mptcp_close_ssk(sk, ssk, subflow);
2264 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2269 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2271 struct mptcp_subflow_context *subflow, *tmp;
2275 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2276 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2278 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2281 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2282 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2285 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2289 static bool mptcp_check_close_timeout(const struct sock *sk)
2291 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2292 struct mptcp_subflow_context *subflow;
2294 if (delta >= TCP_TIMEWAIT_LEN)
2297 /* if all subflows are in closed status don't bother with additional
2300 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2301 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2308 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2310 struct mptcp_subflow_context *subflow, *tmp;
2311 struct sock *sk = &msk->sk.icsk_inet.sk;
2313 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2316 mptcp_token_destroy(msk);
2318 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2319 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2322 slow = lock_sock_fast(tcp_sk);
2323 if (tcp_sk->sk_state != TCP_CLOSE) {
2324 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2325 tcp_set_state(tcp_sk, TCP_CLOSE);
2327 unlock_sock_fast(tcp_sk, slow);
2330 inet_sk_state_store(sk, TCP_CLOSE);
2331 sk->sk_shutdown = SHUTDOWN_MASK;
2332 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2333 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2335 mptcp_close_wake_up(sk);
2338 static void __mptcp_retrans(struct sock *sk)
2340 struct mptcp_sock *msk = mptcp_sk(sk);
2341 struct mptcp_sendmsg_info info = {};
2342 struct mptcp_data_frag *dfrag;
2347 mptcp_clean_una_wakeup(sk);
2349 /* first check ssk: need to kick "stale" logic */
2350 ssk = mptcp_subflow_get_retrans(msk);
2351 dfrag = mptcp_rtx_head(sk);
2353 if (mptcp_data_fin_enabled(msk)) {
2354 struct inet_connection_sock *icsk = inet_csk(sk);
2356 icsk->icsk_retransmits++;
2357 mptcp_set_datafin_timeout(sk);
2358 mptcp_send_ack(msk);
2363 if (!mptcp_send_head(sk))
2374 /* limit retransmission to the bytes already sent on some subflows */
2376 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2377 while (info.sent < info.limit) {
2378 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2382 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2387 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2388 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2395 mptcp_check_and_set_pending(sk);
2397 if (!mptcp_timer_pending(sk))
2398 mptcp_reset_timer(sk);
2401 static void mptcp_worker(struct work_struct *work)
2403 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2404 struct sock *sk = &msk->sk.icsk_inet.sk;
2408 state = sk->sk_state;
2409 if (unlikely(state == TCP_CLOSE))
2412 mptcp_check_data_fin_ack(sk);
2413 mptcp_flush_join_list(msk);
2415 mptcp_check_fastclose(msk);
2418 mptcp_pm_nl_work(msk);
2420 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2421 mptcp_check_for_eof(msk);
2423 __mptcp_check_send_data_fin(sk);
2424 mptcp_check_data_fin(sk);
2426 /* There is no point in keeping around an orphaned sk timedout or
2427 * closed, but we need the msk around to reply to incoming DATA_FIN,
2428 * even if it is orphaned and in FIN_WAIT2 state
2430 if (sock_flag(sk, SOCK_DEAD) &&
2431 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2432 inet_sk_state_store(sk, TCP_CLOSE);
2433 __mptcp_destroy_sock(sk);
2437 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2438 __mptcp_close_subflow(msk);
2440 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2441 __mptcp_retrans(sk);
2448 static int __mptcp_init_sock(struct sock *sk)
2450 struct mptcp_sock *msk = mptcp_sk(sk);
2452 spin_lock_init(&msk->join_list_lock);
2454 INIT_LIST_HEAD(&msk->conn_list);
2455 INIT_LIST_HEAD(&msk->join_list);
2456 INIT_LIST_HEAD(&msk->rtx_queue);
2457 INIT_WORK(&msk->work, mptcp_worker);
2458 __skb_queue_head_init(&msk->receive_queue);
2459 msk->out_of_order_queue = RB_ROOT;
2460 msk->first_pending = NULL;
2461 msk->rmem_fwd_alloc = 0;
2462 WRITE_ONCE(msk->rmem_released, 0);
2463 msk->timer_ival = TCP_RTO_MIN;
2466 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2467 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2468 msk->recovery = false;
2470 mptcp_pm_data_init(msk);
2472 /* re-use the csk retrans timer for MPTCP-level retrans */
2473 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2474 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2479 static int mptcp_init_sock(struct sock *sk)
2481 struct inet_connection_sock *icsk = inet_csk(sk);
2482 struct net *net = sock_net(sk);
2485 ret = __mptcp_init_sock(sk);
2489 if (!mptcp_is_enabled(net))
2490 return -ENOPROTOOPT;
2492 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2495 ret = __mptcp_socket_create(mptcp_sk(sk));
2499 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2500 * propagate the correct value
2502 tcp_assign_congestion_control(sk);
2503 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2505 /* no need to keep a reference to the ops, the name will suffice */
2506 tcp_cleanup_congestion_control(sk);
2507 icsk->icsk_ca_ops = NULL;
2509 sk_sockets_allocated_inc(sk);
2510 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2511 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2516 static void __mptcp_clear_xmit(struct sock *sk)
2518 struct mptcp_sock *msk = mptcp_sk(sk);
2519 struct mptcp_data_frag *dtmp, *dfrag;
2521 WRITE_ONCE(msk->first_pending, NULL);
2522 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2523 dfrag_clear(sk, dfrag);
2526 static void mptcp_cancel_work(struct sock *sk)
2528 struct mptcp_sock *msk = mptcp_sk(sk);
2530 if (cancel_work_sync(&msk->work))
2534 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2538 switch (ssk->sk_state) {
2540 if (!(how & RCV_SHUTDOWN))
2544 tcp_disconnect(ssk, O_NONBLOCK);
2547 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2548 pr_debug("Fallback");
2549 ssk->sk_shutdown |= how;
2550 tcp_shutdown(ssk, how);
2552 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2554 if (!mptcp_timer_pending(sk))
2555 mptcp_reset_timer(sk);
2563 static const unsigned char new_state[16] = {
2564 /* current state: new state: action: */
2565 [0 /* (Invalid) */] = TCP_CLOSE,
2566 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2567 [TCP_SYN_SENT] = TCP_CLOSE,
2568 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2569 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2570 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2571 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2572 [TCP_CLOSE] = TCP_CLOSE,
2573 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2574 [TCP_LAST_ACK] = TCP_LAST_ACK,
2575 [TCP_LISTEN] = TCP_CLOSE,
2576 [TCP_CLOSING] = TCP_CLOSING,
2577 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2580 static int mptcp_close_state(struct sock *sk)
2582 int next = (int)new_state[sk->sk_state];
2583 int ns = next & TCP_STATE_MASK;
2585 inet_sk_state_store(sk, ns);
2587 return next & TCP_ACTION_FIN;
2590 static void __mptcp_check_send_data_fin(struct sock *sk)
2592 struct mptcp_subflow_context *subflow;
2593 struct mptcp_sock *msk = mptcp_sk(sk);
2595 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2596 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2597 msk->snd_nxt, msk->write_seq);
2599 /* we still need to enqueue subflows or not really shutting down,
2602 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2603 mptcp_send_head(sk))
2606 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2608 /* fallback socket will not get data_fin/ack, can move to the next
2611 if (__mptcp_check_fallback(msk)) {
2612 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2613 inet_sk_state_store(sk, TCP_CLOSE);
2614 mptcp_close_wake_up(sk);
2615 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2616 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2620 mptcp_flush_join_list(msk);
2621 mptcp_for_each_subflow(msk, subflow) {
2622 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2624 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2628 static void __mptcp_wr_shutdown(struct sock *sk)
2630 struct mptcp_sock *msk = mptcp_sk(sk);
2632 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2633 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2634 !!mptcp_send_head(sk));
2636 /* will be ignored by fallback sockets */
2637 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2638 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2640 __mptcp_check_send_data_fin(sk);
2643 static void __mptcp_destroy_sock(struct sock *sk)
2645 struct mptcp_subflow_context *subflow, *tmp;
2646 struct mptcp_sock *msk = mptcp_sk(sk);
2647 LIST_HEAD(conn_list);
2649 pr_debug("msk=%p", msk);
2653 /* be sure to always acquire the join list lock, to sync vs
2654 * mptcp_finish_join().
2656 spin_lock_bh(&msk->join_list_lock);
2657 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2658 spin_unlock_bh(&msk->join_list_lock);
2659 list_splice_init(&msk->conn_list, &conn_list);
2661 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2662 sk_stop_timer(sk, &sk->sk_timer);
2665 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2666 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2667 __mptcp_close_ssk(sk, ssk, subflow);
2670 sk->sk_prot->destroy(sk);
2672 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2673 WARN_ON_ONCE(msk->rmem_released);
2674 sk_stream_kill_queues(sk);
2675 xfrm_sk_free_policy(sk);
2677 sk_refcnt_debug_release(sk);
2678 mptcp_dispose_initial_subflow(msk);
2682 static void mptcp_close(struct sock *sk, long timeout)
2684 struct mptcp_subflow_context *subflow;
2685 bool do_cancel_work = false;
2688 sk->sk_shutdown = SHUTDOWN_MASK;
2690 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2691 inet_sk_state_store(sk, TCP_CLOSE);
2695 if (mptcp_close_state(sk))
2696 __mptcp_wr_shutdown(sk);
2698 sk_stream_wait_close(sk, timeout);
2701 /* orphan all the subflows */
2702 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2703 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2704 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2705 bool slow = lock_sock_fast_nested(ssk);
2708 unlock_sock_fast(ssk, slow);
2713 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2714 if (sk->sk_state == TCP_CLOSE) {
2715 __mptcp_destroy_sock(sk);
2716 do_cancel_work = true;
2718 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2722 mptcp_cancel_work(sk);
2724 if (mptcp_sk(sk)->token)
2725 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2730 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2732 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2733 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2734 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2736 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2737 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2740 msk6->saddr = ssk6->saddr;
2741 msk6->flow_label = ssk6->flow_label;
2745 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2746 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2747 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2748 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2749 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2750 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2753 static int mptcp_disconnect(struct sock *sk, int flags)
2755 struct mptcp_subflow_context *subflow;
2756 struct mptcp_sock *msk = mptcp_sk(sk);
2758 mptcp_do_flush_join_list(msk);
2760 mptcp_for_each_subflow(msk, subflow) {
2761 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2764 tcp_disconnect(ssk, flags);
2770 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2771 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2773 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2775 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2779 struct sock *mptcp_sk_clone(const struct sock *sk,
2780 const struct mptcp_options_received *mp_opt,
2781 struct request_sock *req)
2783 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2784 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2785 struct mptcp_sock *msk;
2791 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2792 if (nsk->sk_family == AF_INET6)
2793 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2796 __mptcp_init_sock(nsk);
2798 msk = mptcp_sk(nsk);
2799 msk->local_key = subflow_req->local_key;
2800 msk->token = subflow_req->token;
2801 msk->subflow = NULL;
2802 WRITE_ONCE(msk->fully_established, false);
2803 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
2804 WRITE_ONCE(msk->csum_enabled, true);
2806 msk->write_seq = subflow_req->idsn + 1;
2807 msk->snd_nxt = msk->write_seq;
2808 msk->snd_una = msk->write_seq;
2809 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2810 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2812 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
2813 msk->can_ack = true;
2814 msk->remote_key = mp_opt->sndr_key;
2815 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2817 WRITE_ONCE(msk->ack_seq, ack_seq);
2818 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2821 sock_reset_flag(nsk, SOCK_RCU_FREE);
2822 /* will be fully established after successful MPC subflow creation */
2823 inet_sk_state_store(nsk, TCP_SYN_RECV);
2825 security_inet_csk_clone(nsk, req);
2826 bh_unlock_sock(nsk);
2828 /* keep a single reference */
2833 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2835 const struct tcp_sock *tp = tcp_sk(ssk);
2837 msk->rcvq_space.copied = 0;
2838 msk->rcvq_space.rtt_us = 0;
2840 msk->rcvq_space.time = tp->tcp_mstamp;
2842 /* initial rcv_space offering made to peer */
2843 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2844 TCP_INIT_CWND * tp->advmss);
2845 if (msk->rcvq_space.space == 0)
2846 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2848 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2851 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2854 struct mptcp_sock *msk = mptcp_sk(sk);
2855 struct socket *listener;
2858 listener = __mptcp_nmpc_socket(msk);
2859 if (WARN_ON_ONCE(!listener)) {
2864 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2865 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2869 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2870 if (sk_is_mptcp(newsk)) {
2871 struct mptcp_subflow_context *subflow;
2872 struct sock *new_mptcp_sock;
2874 subflow = mptcp_subflow_ctx(newsk);
2875 new_mptcp_sock = subflow->conn;
2877 /* is_mptcp should be false if subflow->conn is missing, see
2878 * subflow_syn_recv_sock()
2880 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2881 tcp_sk(newsk)->is_mptcp = 0;
2885 /* acquire the 2nd reference for the owning socket */
2886 sock_hold(new_mptcp_sock);
2887 newsk = new_mptcp_sock;
2888 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2890 MPTCP_INC_STATS(sock_net(sk),
2891 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2897 void mptcp_destroy_common(struct mptcp_sock *msk)
2899 struct sock *sk = (struct sock *)msk;
2901 __mptcp_clear_xmit(sk);
2903 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2904 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2905 __skb_queue_purge(&sk->sk_receive_queue);
2906 skb_rbtree_purge(&msk->out_of_order_queue);
2908 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
2909 * inet_sock_destruct() will dispose it
2911 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
2912 msk->rmem_fwd_alloc = 0;
2913 mptcp_token_destroy(msk);
2914 mptcp_pm_free_anno_list(msk);
2917 static void mptcp_destroy(struct sock *sk)
2919 struct mptcp_sock *msk = mptcp_sk(sk);
2921 mptcp_destroy_common(msk);
2922 sk_sockets_allocated_dec(sk);
2925 void __mptcp_data_acked(struct sock *sk)
2927 if (!sock_owned_by_user(sk))
2928 __mptcp_clean_una(sk);
2930 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2932 if (mptcp_pending_data_fin_ack(sk))
2933 mptcp_schedule_work(sk);
2936 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2938 if (!mptcp_send_head(sk))
2941 if (!sock_owned_by_user(sk)) {
2942 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2944 if (xmit_ssk == ssk)
2945 __mptcp_subflow_push_pending(sk, ssk);
2947 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
2949 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2953 /* processes deferred events and flush wmem */
2954 static void mptcp_release_cb(struct sock *sk)
2957 unsigned long flags = 0;
2959 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2960 flags |= BIT(MPTCP_PUSH_PENDING);
2961 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2962 flags |= BIT(MPTCP_RETRANSMIT);
2966 /* the following actions acquire the subflow socket lock
2968 * 1) can't be invoked in atomic scope
2969 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2970 * datapath acquires the msk socket spinlock while helding
2971 * the subflow socket lock
2974 spin_unlock_bh(&sk->sk_lock.slock);
2975 if (flags & BIT(MPTCP_PUSH_PENDING))
2976 __mptcp_push_pending(sk, 0);
2977 if (flags & BIT(MPTCP_RETRANSMIT))
2978 __mptcp_retrans(sk);
2981 spin_lock_bh(&sk->sk_lock.slock);
2984 /* be sure to set the current sk state before tacking actions
2985 * depending on sk_state
2987 if (test_and_clear_bit(MPTCP_CONNECTED, &mptcp_sk(sk)->flags))
2988 __mptcp_set_connected(sk);
2989 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2990 __mptcp_clean_una_wakeup(sk);
2991 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2992 __mptcp_error_report(sk);
2994 __mptcp_update_rmem(sk);
2997 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
2998 * TCP can't schedule delack timer before the subflow is fully established.
2999 * MPTCP uses the delack timer to do 3rd ack retransmissions
3001 static void schedule_3rdack_retransmission(struct sock *ssk)
3003 struct inet_connection_sock *icsk = inet_csk(ssk);
3004 struct tcp_sock *tp = tcp_sk(ssk);
3005 unsigned long timeout;
3007 if (mptcp_subflow_ctx(ssk)->fully_established)
3010 /* reschedule with a timeout above RTT, as we must look only for drop */
3012 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3014 timeout = TCP_TIMEOUT_INIT;
3017 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3018 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3019 icsk->icsk_ack.timeout = timeout;
3020 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3023 void mptcp_subflow_process_delegated(struct sock *ssk)
3025 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3026 struct sock *sk = subflow->conn;
3028 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3029 mptcp_data_lock(sk);
3030 if (!sock_owned_by_user(sk))
3031 __mptcp_subflow_push_pending(sk, ssk);
3033 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
3034 mptcp_data_unlock(sk);
3035 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3037 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3038 schedule_3rdack_retransmission(ssk);
3039 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3043 static int mptcp_hash(struct sock *sk)
3045 /* should never be called,
3046 * we hash the TCP subflows not the master socket
3052 static void mptcp_unhash(struct sock *sk)
3054 /* called from sk_common_release(), but nothing to do here */
3057 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3059 struct mptcp_sock *msk = mptcp_sk(sk);
3060 struct socket *ssock;
3062 ssock = __mptcp_nmpc_socket(msk);
3063 pr_debug("msk=%p, subflow=%p", msk, ssock);
3064 if (WARN_ON_ONCE(!ssock))
3067 return inet_csk_get_port(ssock->sk, snum);
3070 void mptcp_finish_connect(struct sock *ssk)
3072 struct mptcp_subflow_context *subflow;
3073 struct mptcp_sock *msk;
3077 subflow = mptcp_subflow_ctx(ssk);
3081 pr_debug("msk=%p, token=%u", sk, subflow->token);
3083 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3085 subflow->map_seq = ack_seq;
3086 subflow->map_subflow_seq = 1;
3088 /* the socket is not connected yet, no msk/subflow ops can access/race
3089 * accessing the field below
3091 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3092 WRITE_ONCE(msk->local_key, subflow->local_key);
3093 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3094 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3095 WRITE_ONCE(msk->ack_seq, ack_seq);
3096 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3097 WRITE_ONCE(msk->can_ack, 1);
3098 WRITE_ONCE(msk->snd_una, msk->write_seq);
3100 mptcp_pm_new_connection(msk, ssk, 0);
3102 mptcp_rcv_space_init(msk, ssk);
3105 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3107 write_lock_bh(&sk->sk_callback_lock);
3108 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3109 sk_set_socket(sk, parent);
3110 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3111 write_unlock_bh(&sk->sk_callback_lock);
3114 bool mptcp_finish_join(struct sock *ssk)
3116 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3117 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3118 struct sock *parent = (void *)msk;
3119 struct socket *parent_sock;
3122 pr_debug("msk=%p, subflow=%p", msk, subflow);
3124 /* mptcp socket already closing? */
3125 if (!mptcp_is_fully_established(parent)) {
3126 subflow->reset_reason = MPTCP_RST_EMPTCP;
3130 if (!msk->pm.server_side)
3133 if (!mptcp_pm_allow_new_subflow(msk)) {
3134 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3138 /* active connections are already on conn_list, and we can't acquire
3140 * use the join list lock as synchronization point and double-check
3141 * msk status to avoid racing with __mptcp_destroy_sock()
3143 spin_lock_bh(&msk->join_list_lock);
3144 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3145 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3146 list_add_tail(&subflow->node, &msk->join_list);
3149 spin_unlock_bh(&msk->join_list_lock);
3151 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3155 /* attach to msk socket only after we are sure he will deal with us
3158 parent_sock = READ_ONCE(parent->sk_socket);
3159 if (parent_sock && !ssk->sk_socket)
3160 mptcp_sock_graft(ssk, parent_sock);
3161 subflow->map_seq = READ_ONCE(msk->ack_seq);
3163 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3167 static void mptcp_shutdown(struct sock *sk, int how)
3169 pr_debug("sk=%p, how=%d", sk, how);
3171 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3172 __mptcp_wr_shutdown(sk);
3175 static int mptcp_forward_alloc_get(const struct sock *sk)
3177 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3180 static struct proto mptcp_prot = {
3182 .owner = THIS_MODULE,
3183 .init = mptcp_init_sock,
3184 .disconnect = mptcp_disconnect,
3185 .close = mptcp_close,
3186 .accept = mptcp_accept,
3187 .setsockopt = mptcp_setsockopt,
3188 .getsockopt = mptcp_getsockopt,
3189 .shutdown = mptcp_shutdown,
3190 .destroy = mptcp_destroy,
3191 .sendmsg = mptcp_sendmsg,
3192 .recvmsg = mptcp_recvmsg,
3193 .release_cb = mptcp_release_cb,
3195 .unhash = mptcp_unhash,
3196 .get_port = mptcp_get_port,
3197 .forward_alloc_get = mptcp_forward_alloc_get,
3198 .sockets_allocated = &mptcp_sockets_allocated,
3199 .memory_allocated = &tcp_memory_allocated,
3200 .memory_pressure = &tcp_memory_pressure,
3201 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3202 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3203 .sysctl_mem = sysctl_tcp_mem,
3204 .obj_size = sizeof(struct mptcp_sock),
3205 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3206 .no_autobind = true,
3209 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3211 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3212 struct socket *ssock;
3215 lock_sock(sock->sk);
3216 ssock = __mptcp_nmpc_socket(msk);
3222 err = ssock->ops->bind(ssock, uaddr, addr_len);
3224 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3227 release_sock(sock->sk);
3231 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3232 struct mptcp_subflow_context *subflow)
3234 subflow->request_mptcp = 0;
3235 __mptcp_do_fallback(msk);
3238 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3239 int addr_len, int flags)
3241 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3242 struct mptcp_subflow_context *subflow;
3243 struct socket *ssock;
3246 lock_sock(sock->sk);
3247 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3248 /* pending connection or invalid state, let existing subflow
3251 ssock = msk->subflow;
3255 ssock = __mptcp_nmpc_socket(msk);
3261 mptcp_token_destroy(msk);
3262 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3263 subflow = mptcp_subflow_ctx(ssock->sk);
3264 #ifdef CONFIG_TCP_MD5SIG
3265 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3268 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3269 mptcp_subflow_early_fallback(msk, subflow);
3271 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3272 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3273 mptcp_subflow_early_fallback(msk, subflow);
3275 if (likely(!__mptcp_check_fallback(msk)))
3276 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3279 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3280 sock->state = ssock->state;
3282 /* on successful connect, the msk state will be moved to established by
3283 * subflow_finish_connect()
3285 if (!err || err == -EINPROGRESS)
3286 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3288 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3291 release_sock(sock->sk);
3295 static int mptcp_listen(struct socket *sock, int backlog)
3297 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3298 struct socket *ssock;
3301 pr_debug("msk=%p", msk);
3303 lock_sock(sock->sk);
3304 ssock = __mptcp_nmpc_socket(msk);
3310 mptcp_token_destroy(msk);
3311 inet_sk_state_store(sock->sk, TCP_LISTEN);
3312 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3314 err = ssock->ops->listen(ssock, backlog);
3315 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3317 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3320 release_sock(sock->sk);
3324 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3325 int flags, bool kern)
3327 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3328 struct socket *ssock;
3331 pr_debug("msk=%p", msk);
3333 lock_sock(sock->sk);
3334 if (sock->sk->sk_state != TCP_LISTEN)
3337 ssock = __mptcp_nmpc_socket(msk);
3341 clear_bit(MPTCP_DATA_READY, &msk->flags);
3342 sock_hold(ssock->sk);
3343 release_sock(sock->sk);
3345 err = ssock->ops->accept(sock, newsock, flags, kern);
3346 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3347 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3348 struct mptcp_subflow_context *subflow;
3349 struct sock *newsk = newsock->sk;
3353 /* PM/worker can now acquire the first subflow socket
3354 * lock without racing with listener queue cleanup,
3355 * we can notify it, if needed.
3357 * Even if remote has reset the initial subflow by now
3358 * the refcnt is still at least one.
3360 subflow = mptcp_subflow_ctx(msk->first);
3361 list_add(&subflow->node, &msk->conn_list);
3362 sock_hold(msk->first);
3363 if (mptcp_is_fully_established(newsk))
3364 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3366 mptcp_copy_inaddrs(newsk, msk->first);
3367 mptcp_rcv_space_init(msk, msk->first);
3368 mptcp_propagate_sndbuf(newsk, msk->first);
3370 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3371 * This is needed so NOSPACE flag can be set from tcp stack.
3373 mptcp_flush_join_list(msk);
3374 mptcp_for_each_subflow(msk, subflow) {
3375 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3377 if (!ssk->sk_socket)
3378 mptcp_sock_graft(ssk, newsock);
3380 release_sock(newsk);
3383 if (inet_csk_listen_poll(ssock->sk))
3384 set_bit(MPTCP_DATA_READY, &msk->flags);
3385 sock_put(ssock->sk);
3389 release_sock(sock->sk);
3393 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3395 /* Concurrent splices from sk_receive_queue into receive_queue will
3396 * always show at least one non-empty queue when checked in this order.
3398 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
3399 skb_queue_empty_lockless(&msk->receive_queue))
3402 return EPOLLIN | EPOLLRDNORM;
3405 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3407 struct sock *sk = (struct sock *)msk;
3409 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3410 return EPOLLOUT | EPOLLWRNORM;
3412 if (sk_stream_is_writeable(sk))
3413 return EPOLLOUT | EPOLLWRNORM;
3415 mptcp_set_nospace(sk);
3416 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3417 if (sk_stream_is_writeable(sk))
3418 return EPOLLOUT | EPOLLWRNORM;
3423 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3424 struct poll_table_struct *wait)
3426 struct sock *sk = sock->sk;
3427 struct mptcp_sock *msk;
3432 sock_poll_wait(file, sock, wait);
3434 state = inet_sk_state_load(sk);
3435 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3436 if (state == TCP_LISTEN)
3437 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM : 0;
3439 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3440 mask |= mptcp_check_readable(msk);
3441 mask |= mptcp_check_writeable(msk);
3443 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3445 if (sk->sk_shutdown & RCV_SHUTDOWN)
3446 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3448 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3456 static const struct proto_ops mptcp_stream_ops = {
3458 .owner = THIS_MODULE,
3459 .release = inet_release,
3461 .connect = mptcp_stream_connect,
3462 .socketpair = sock_no_socketpair,
3463 .accept = mptcp_stream_accept,
3464 .getname = inet_getname,
3466 .ioctl = inet_ioctl,
3467 .gettstamp = sock_gettstamp,
3468 .listen = mptcp_listen,
3469 .shutdown = inet_shutdown,
3470 .setsockopt = sock_common_setsockopt,
3471 .getsockopt = sock_common_getsockopt,
3472 .sendmsg = inet_sendmsg,
3473 .recvmsg = inet_recvmsg,
3474 .mmap = sock_no_mmap,
3475 .sendpage = inet_sendpage,
3478 static struct inet_protosw mptcp_protosw = {
3479 .type = SOCK_STREAM,
3480 .protocol = IPPROTO_MPTCP,
3481 .prot = &mptcp_prot,
3482 .ops = &mptcp_stream_ops,
3483 .flags = INET_PROTOSW_ICSK,
3486 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3488 struct mptcp_delegated_action *delegated;
3489 struct mptcp_subflow_context *subflow;
3492 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3493 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3494 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3496 bh_lock_sock_nested(ssk);
3497 if (!sock_owned_by_user(ssk) &&
3498 mptcp_subflow_has_delegated_action(subflow))
3499 mptcp_subflow_process_delegated(ssk);
3500 /* ... elsewhere tcp_release_cb_override already processed
3501 * the action or will do at next release_sock().
3502 * In both case must dequeue the subflow here - on the same
3503 * CPU that scheduled it.
3505 bh_unlock_sock(ssk);
3508 if (++work_done == budget)
3512 /* always provide a 0 'work_done' argument, so that napi_complete_done
3513 * will not try accessing the NULL napi->dev ptr
3515 napi_complete_done(napi, 0);
3519 void __init mptcp_proto_init(void)
3521 struct mptcp_delegated_action *delegated;
3524 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3526 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3527 panic("Failed to allocate MPTCP pcpu counter\n");
3529 init_dummy_netdev(&mptcp_napi_dev);
3530 for_each_possible_cpu(cpu) {
3531 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3532 INIT_LIST_HEAD(&delegated->head);
3533 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3535 napi_enable(&delegated->napi);
3538 mptcp_subflow_init();
3542 if (proto_register(&mptcp_prot, 1) != 0)
3543 panic("Failed to register MPTCP proto.\n");
3545 inet_register_protosw(&mptcp_protosw);
3547 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3550 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3551 static const struct proto_ops mptcp_v6_stream_ops = {
3553 .owner = THIS_MODULE,
3554 .release = inet6_release,
3556 .connect = mptcp_stream_connect,
3557 .socketpair = sock_no_socketpair,
3558 .accept = mptcp_stream_accept,
3559 .getname = inet6_getname,
3561 .ioctl = inet6_ioctl,
3562 .gettstamp = sock_gettstamp,
3563 .listen = mptcp_listen,
3564 .shutdown = inet_shutdown,
3565 .setsockopt = sock_common_setsockopt,
3566 .getsockopt = sock_common_getsockopt,
3567 .sendmsg = inet6_sendmsg,
3568 .recvmsg = inet6_recvmsg,
3569 .mmap = sock_no_mmap,
3570 .sendpage = inet_sendpage,
3571 #ifdef CONFIG_COMPAT
3572 .compat_ioctl = inet6_compat_ioctl,
3576 static struct proto mptcp_v6_prot;
3578 static void mptcp_v6_destroy(struct sock *sk)
3581 inet6_destroy_sock(sk);
3584 static struct inet_protosw mptcp_v6_protosw = {
3585 .type = SOCK_STREAM,
3586 .protocol = IPPROTO_MPTCP,
3587 .prot = &mptcp_v6_prot,
3588 .ops = &mptcp_v6_stream_ops,
3589 .flags = INET_PROTOSW_ICSK,
3592 int __init mptcp_proto_v6_init(void)
3596 mptcp_v6_prot = mptcp_prot;
3597 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3598 mptcp_v6_prot.slab = NULL;
3599 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3600 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3602 err = proto_register(&mptcp_v6_prot, 1);
3606 err = inet6_register_protosw(&mptcp_v6_protosw);
3608 proto_unregister(&mptcp_v6_prot);