1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
46 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
49 MPTCP_CMSG_TS = BIT(0),
50 MPTCP_CMSG_INQ = BIT(1),
53 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
55 static void __mptcp_destroy_sock(struct sock *sk);
56 static void __mptcp_check_send_data_fin(struct sock *sk);
58 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
59 static struct net_device mptcp_napi_dev;
61 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
62 * completed yet or has failed, return the subflow socket.
63 * Otherwise return NULL.
65 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
67 if (!msk->subflow || READ_ONCE(msk->can_ack))
73 /* Returns end sequence number of the receiver's advertised window */
74 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
76 return READ_ONCE(msk->wnd_end);
79 static bool mptcp_is_tcpsk(struct sock *sk)
81 struct socket *sock = sk->sk_socket;
83 if (unlikely(sk->sk_prot == &tcp_prot)) {
84 /* we are being invoked after mptcp_accept() has
85 * accepted a non-mp-capable flow: sk is a tcp_sk,
88 * Hand the socket over to tcp so all further socket ops
91 sock->ops = &inet_stream_ops;
93 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
94 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
95 sock->ops = &inet6_stream_ops;
103 static int __mptcp_socket_create(struct mptcp_sock *msk)
105 struct mptcp_subflow_context *subflow;
106 struct sock *sk = (struct sock *)msk;
107 struct socket *ssock;
110 err = mptcp_subflow_create_socket(sk, &ssock);
114 msk->first = ssock->sk;
115 msk->subflow = ssock;
116 subflow = mptcp_subflow_ctx(ssock->sk);
117 list_add(&subflow->node, &msk->conn_list);
118 sock_hold(ssock->sk);
119 subflow->request_mptcp = 1;
120 mptcp_sock_graft(msk->first, sk->sk_socket);
125 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
127 sk_drops_add(sk, skb);
131 static void mptcp_rmem_charge(struct sock *sk, int size)
133 mptcp_sk(sk)->rmem_fwd_alloc -= size;
136 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
137 struct sk_buff *from)
142 if (MPTCP_SKB_CB(from)->offset ||
143 !skb_try_coalesce(to, from, &fragstolen, &delta))
146 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
147 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
148 to->len, MPTCP_SKB_CB(from)->end_seq);
149 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
150 kfree_skb_partial(from, fragstolen);
151 atomic_add(delta, &sk->sk_rmem_alloc);
152 mptcp_rmem_charge(sk, delta);
156 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
157 struct sk_buff *from)
159 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
162 return mptcp_try_coalesce((struct sock *)msk, to, from);
165 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
167 amount >>= SK_MEM_QUANTUM_SHIFT;
168 mptcp_sk(sk)->rmem_fwd_alloc -= amount << SK_MEM_QUANTUM_SHIFT;
169 __sk_mem_reduce_allocated(sk, amount);
172 static void mptcp_rmem_uncharge(struct sock *sk, int size)
174 struct mptcp_sock *msk = mptcp_sk(sk);
177 msk->rmem_fwd_alloc += size;
178 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
180 /* see sk_mem_uncharge() for the rationale behind the following schema */
181 if (unlikely(reclaimable >= SK_RECLAIM_THRESHOLD))
182 __mptcp_rmem_reclaim(sk, SK_RECLAIM_CHUNK);
185 static void mptcp_rfree(struct sk_buff *skb)
187 unsigned int len = skb->truesize;
188 struct sock *sk = skb->sk;
190 atomic_sub(len, &sk->sk_rmem_alloc);
191 mptcp_rmem_uncharge(sk, len);
194 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
198 skb->destructor = mptcp_rfree;
199 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
200 mptcp_rmem_charge(sk, skb->truesize);
203 /* "inspired" by tcp_data_queue_ofo(), main differences:
205 * - don't cope with sacks
207 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
209 struct sock *sk = (struct sock *)msk;
210 struct rb_node **p, *parent;
211 u64 seq, end_seq, max_seq;
212 struct sk_buff *skb1;
214 seq = MPTCP_SKB_CB(skb)->map_seq;
215 end_seq = MPTCP_SKB_CB(skb)->end_seq;
216 max_seq = READ_ONCE(msk->rcv_wnd_sent);
218 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
219 RB_EMPTY_ROOT(&msk->out_of_order_queue));
220 if (after64(end_seq, max_seq)) {
223 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
224 (unsigned long long)end_seq - (unsigned long)max_seq,
225 (unsigned long long)msk->rcv_wnd_sent);
226 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
230 p = &msk->out_of_order_queue.rb_node;
231 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
232 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
233 rb_link_node(&skb->rbnode, NULL, p);
234 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
235 msk->ooo_last_skb = skb;
239 /* with 2 subflows, adding at end of ooo queue is quite likely
240 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
242 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
248 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
249 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
250 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
251 parent = &msk->ooo_last_skb->rbnode;
252 p = &parent->rb_right;
256 /* Find place to insert this segment. Handle overlaps on the way. */
260 skb1 = rb_to_skb(parent);
261 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
262 p = &parent->rb_left;
265 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
266 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
267 /* All the bits are present. Drop. */
269 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
272 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
276 * continue traversing
279 /* skb's seq == skb1's seq and skb covers skb1.
280 * Replace skb1 with skb.
282 rb_replace_node(&skb1->rbnode, &skb->rbnode,
283 &msk->out_of_order_queue);
284 mptcp_drop(sk, skb1);
285 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
288 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
289 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
292 p = &parent->rb_right;
296 /* Insert segment into RB tree. */
297 rb_link_node(&skb->rbnode, parent, p);
298 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
301 /* Remove other segments covered by skb. */
302 while ((skb1 = skb_rb_next(skb)) != NULL) {
303 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
305 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
306 mptcp_drop(sk, skb1);
307 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
309 /* If there is no skb after us, we are the last_skb ! */
311 msk->ooo_last_skb = skb;
315 mptcp_set_owner_r(skb, sk);
318 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
320 struct mptcp_sock *msk = mptcp_sk(sk);
323 if (size < msk->rmem_fwd_alloc)
326 amt = sk_mem_pages(size);
327 amount = amt << SK_MEM_QUANTUM_SHIFT;
328 msk->rmem_fwd_alloc += amount;
329 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV)) {
330 if (ssk->sk_forward_alloc < amount) {
331 msk->rmem_fwd_alloc -= amount;
335 ssk->sk_forward_alloc -= amount;
340 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
341 struct sk_buff *skb, unsigned int offset,
344 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
345 struct sock *sk = (struct sock *)msk;
346 struct sk_buff *tail;
349 __skb_unlink(skb, &ssk->sk_receive_queue);
354 /* try to fetch required memory from subflow */
355 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
358 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
360 /* the skb map_seq accounts for the skb offset:
361 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
364 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
365 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
366 MPTCP_SKB_CB(skb)->offset = offset;
367 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
369 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
371 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
372 tail = skb_peek_tail(&sk->sk_receive_queue);
373 if (tail && mptcp_try_coalesce(sk, tail, skb))
376 mptcp_set_owner_r(skb, sk);
377 __skb_queue_tail(&sk->sk_receive_queue, skb);
379 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
380 mptcp_data_queue_ofo(msk, skb);
384 /* old data, keep it simple and drop the whole pkt, sender
385 * will retransmit as needed, if needed.
387 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
393 static void mptcp_stop_timer(struct sock *sk)
395 struct inet_connection_sock *icsk = inet_csk(sk);
397 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
398 mptcp_sk(sk)->timer_ival = 0;
401 static void mptcp_close_wake_up(struct sock *sk)
403 if (sock_flag(sk, SOCK_DEAD))
406 sk->sk_state_change(sk);
407 if (sk->sk_shutdown == SHUTDOWN_MASK ||
408 sk->sk_state == TCP_CLOSE)
409 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
411 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
414 static bool mptcp_pending_data_fin_ack(struct sock *sk)
416 struct mptcp_sock *msk = mptcp_sk(sk);
418 return !__mptcp_check_fallback(msk) &&
419 ((1 << sk->sk_state) &
420 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
421 msk->write_seq == READ_ONCE(msk->snd_una);
424 static void mptcp_check_data_fin_ack(struct sock *sk)
426 struct mptcp_sock *msk = mptcp_sk(sk);
428 /* Look for an acknowledged DATA_FIN */
429 if (mptcp_pending_data_fin_ack(sk)) {
430 WRITE_ONCE(msk->snd_data_fin_enable, 0);
432 switch (sk->sk_state) {
434 inet_sk_state_store(sk, TCP_FIN_WAIT2);
438 inet_sk_state_store(sk, TCP_CLOSE);
442 mptcp_close_wake_up(sk);
446 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
448 struct mptcp_sock *msk = mptcp_sk(sk);
450 if (READ_ONCE(msk->rcv_data_fin) &&
451 ((1 << sk->sk_state) &
452 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
453 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
455 if (msk->ack_seq == rcv_data_fin_seq) {
457 *seq = rcv_data_fin_seq;
466 static void mptcp_set_datafin_timeout(const struct sock *sk)
468 struct inet_connection_sock *icsk = inet_csk(sk);
471 retransmits = min_t(u32, icsk->icsk_retransmits,
472 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
474 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
477 static void __mptcp_set_timeout(struct sock *sk, long tout)
479 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
482 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
484 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
486 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
487 inet_csk(ssk)->icsk_timeout - jiffies : 0;
490 static void mptcp_set_timeout(struct sock *sk)
492 struct mptcp_subflow_context *subflow;
495 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
496 tout = max(tout, mptcp_timeout_from_subflow(subflow));
497 __mptcp_set_timeout(sk, tout);
500 static bool tcp_can_send_ack(const struct sock *ssk)
502 return !((1 << inet_sk_state_load(ssk)) &
503 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
506 void mptcp_subflow_send_ack(struct sock *ssk)
510 slow = lock_sock_fast(ssk);
511 if (tcp_can_send_ack(ssk))
513 unlock_sock_fast(ssk, slow);
516 static void mptcp_send_ack(struct mptcp_sock *msk)
518 struct mptcp_subflow_context *subflow;
520 mptcp_for_each_subflow(msk, subflow)
521 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
524 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
528 slow = lock_sock_fast(ssk);
529 if (tcp_can_send_ack(ssk))
530 tcp_cleanup_rbuf(ssk, 1);
531 unlock_sock_fast(ssk, slow);
534 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
536 const struct inet_connection_sock *icsk = inet_csk(ssk);
537 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
538 const struct tcp_sock *tp = tcp_sk(ssk);
540 return (ack_pending & ICSK_ACK_SCHED) &&
541 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
542 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
543 (rx_empty && ack_pending &
544 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
547 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
549 int old_space = READ_ONCE(msk->old_wspace);
550 struct mptcp_subflow_context *subflow;
551 struct sock *sk = (struct sock *)msk;
552 int space = __mptcp_space(sk);
553 bool cleanup, rx_empty;
555 cleanup = (space > 0) && (space >= (old_space << 1));
556 rx_empty = !__mptcp_rmem(sk);
558 mptcp_for_each_subflow(msk, subflow) {
559 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
561 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
562 mptcp_subflow_cleanup_rbuf(ssk);
566 static bool mptcp_check_data_fin(struct sock *sk)
568 struct mptcp_sock *msk = mptcp_sk(sk);
569 u64 rcv_data_fin_seq;
572 if (__mptcp_check_fallback(msk))
575 /* Need to ack a DATA_FIN received from a peer while this side
576 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
577 * msk->rcv_data_fin was set when parsing the incoming options
578 * at the subflow level and the msk lock was not held, so this
579 * is the first opportunity to act on the DATA_FIN and change
582 * If we are caught up to the sequence number of the incoming
583 * DATA_FIN, send the DATA_ACK now and do state transition. If
584 * not caught up, do nothing and let the recv code send DATA_ACK
588 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
589 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
590 WRITE_ONCE(msk->rcv_data_fin, 0);
592 sk->sk_shutdown |= RCV_SHUTDOWN;
593 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
595 switch (sk->sk_state) {
596 case TCP_ESTABLISHED:
597 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
600 inet_sk_state_store(sk, TCP_CLOSING);
603 inet_sk_state_store(sk, TCP_CLOSE);
606 /* Other states not expected */
613 mptcp_close_wake_up(sk);
618 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
622 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
623 struct sock *sk = (struct sock *)msk;
624 unsigned int moved = 0;
625 bool more_data_avail;
630 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
632 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
633 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
635 if (unlikely(ssk_rbuf > sk_rbuf)) {
636 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
641 pr_debug("msk=%p ssk=%p", msk, ssk);
644 u32 map_remaining, offset;
645 u32 seq = tp->copied_seq;
649 /* try to move as much data as available */
650 map_remaining = subflow->map_data_len -
651 mptcp_subflow_get_map_offset(subflow);
653 skb = skb_peek(&ssk->sk_receive_queue);
655 /* if no data is found, a racing workqueue/recvmsg
656 * already processed the new data, stop here or we
657 * can enter an infinite loop
664 if (__mptcp_check_fallback(msk)) {
665 /* if we are running under the workqueue, TCP could have
666 * collapsed skbs between dummy map creation and now
667 * be sure to adjust the size
669 map_remaining = skb->len;
670 subflow->map_data_len = skb->len;
673 offset = seq - TCP_SKB_CB(skb)->seq;
674 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
680 if (offset < skb->len) {
681 size_t len = skb->len - offset;
686 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
690 if (WARN_ON_ONCE(map_remaining < len))
694 sk_eat_skb(ssk, skb);
698 WRITE_ONCE(tp->copied_seq, seq);
699 more_data_avail = mptcp_subflow_data_available(ssk);
701 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
705 } while (more_data_avail);
711 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
713 struct sock *sk = (struct sock *)msk;
714 struct sk_buff *skb, *tail;
719 p = rb_first(&msk->out_of_order_queue);
720 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
723 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
727 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
729 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
732 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
736 end_seq = MPTCP_SKB_CB(skb)->end_seq;
737 tail = skb_peek_tail(&sk->sk_receive_queue);
738 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
739 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
741 /* skip overlapping data, if any */
742 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
743 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
745 MPTCP_SKB_CB(skb)->offset += delta;
746 MPTCP_SKB_CB(skb)->map_seq += delta;
747 __skb_queue_tail(&sk->sk_receive_queue, skb);
749 msk->ack_seq = end_seq;
755 /* In most cases we will be able to lock the mptcp socket. If its already
756 * owned, we need to defer to the work queue to avoid ABBA deadlock.
758 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
760 struct sock *sk = (struct sock *)msk;
761 unsigned int moved = 0;
763 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
764 __mptcp_ofo_queue(msk);
765 if (unlikely(ssk->sk_err)) {
766 if (!sock_owned_by_user(sk))
767 __mptcp_error_report(sk);
769 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
772 /* If the moves have caught up with the DATA_FIN sequence number
773 * it's time to ack the DATA_FIN and change socket state, but
774 * this is not a good place to change state. Let the workqueue
777 if (mptcp_pending_data_fin(sk, NULL))
778 mptcp_schedule_work(sk);
782 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
784 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
785 struct mptcp_sock *msk = mptcp_sk(sk);
786 int sk_rbuf, ssk_rbuf;
788 /* The peer can send data while we are shutting down this
789 * subflow at msk destruction time, but we must avoid enqueuing
790 * more data to the msk receive queue
792 if (unlikely(subflow->disposable))
795 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
796 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
797 if (unlikely(ssk_rbuf > sk_rbuf))
800 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
801 if (__mptcp_rmem(sk) > sk_rbuf) {
802 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
806 /* Wake-up the reader only for in-sequence data */
808 if (move_skbs_to_msk(msk, ssk))
809 sk->sk_data_ready(sk);
811 mptcp_data_unlock(sk);
814 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
816 struct sock *sk = (struct sock *)msk;
818 if (sk->sk_state != TCP_ESTABLISHED)
821 /* attach to msk socket only after we are sure we will deal with it
824 if (sk->sk_socket && !ssk->sk_socket)
825 mptcp_sock_graft(ssk, sk->sk_socket);
827 mptcp_propagate_sndbuf((struct sock *)msk, ssk);
828 mptcp_sockopt_sync_locked(msk, ssk);
832 static void __mptcp_flush_join_list(struct sock *sk)
834 struct mptcp_subflow_context *tmp, *subflow;
835 struct mptcp_sock *msk = mptcp_sk(sk);
837 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
838 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
839 bool slow = lock_sock_fast(ssk);
841 list_move_tail(&subflow->node, &msk->conn_list);
842 if (!__mptcp_finish_join(msk, ssk))
843 mptcp_subflow_reset(ssk);
844 unlock_sock_fast(ssk, slow);
848 static bool mptcp_timer_pending(struct sock *sk)
850 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
853 static void mptcp_reset_timer(struct sock *sk)
855 struct inet_connection_sock *icsk = inet_csk(sk);
858 /* prevent rescheduling on close */
859 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
862 tout = mptcp_sk(sk)->timer_ival;
863 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
866 bool mptcp_schedule_work(struct sock *sk)
868 if (inet_sk_state_load(sk) != TCP_CLOSE &&
869 schedule_work(&mptcp_sk(sk)->work)) {
870 /* each subflow already holds a reference to the sk, and the
871 * workqueue is invoked by a subflow, so sk can't go away here.
879 void mptcp_subflow_eof(struct sock *sk)
881 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
882 mptcp_schedule_work(sk);
885 static void mptcp_check_for_eof(struct mptcp_sock *msk)
887 struct mptcp_subflow_context *subflow;
888 struct sock *sk = (struct sock *)msk;
891 mptcp_for_each_subflow(msk, subflow)
892 receivers += !subflow->rx_eof;
896 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
897 /* hopefully temporary hack: propagate shutdown status
898 * to msk, when all subflows agree on it
900 sk->sk_shutdown |= RCV_SHUTDOWN;
902 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
903 sk->sk_data_ready(sk);
906 switch (sk->sk_state) {
907 case TCP_ESTABLISHED:
908 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
911 inet_sk_state_store(sk, TCP_CLOSING);
914 inet_sk_state_store(sk, TCP_CLOSE);
919 mptcp_close_wake_up(sk);
922 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
924 struct mptcp_subflow_context *subflow;
925 struct sock *sk = (struct sock *)msk;
927 sock_owned_by_me(sk);
929 mptcp_for_each_subflow(msk, subflow) {
930 if (READ_ONCE(subflow->data_avail))
931 return mptcp_subflow_tcp_sock(subflow);
937 static bool mptcp_skb_can_collapse_to(u64 write_seq,
938 const struct sk_buff *skb,
939 const struct mptcp_ext *mpext)
941 if (!tcp_skb_can_collapse_to(skb))
944 /* can collapse only if MPTCP level sequence is in order and this
945 * mapping has not been xmitted yet
947 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
951 /* we can append data to the given data frag if:
952 * - there is space available in the backing page_frag
953 * - the data frag tail matches the current page_frag free offset
954 * - the data frag end sequence number matches the current write seq
956 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
957 const struct page_frag *pfrag,
958 const struct mptcp_data_frag *df)
960 return df && pfrag->page == df->page &&
961 pfrag->size - pfrag->offset > 0 &&
962 pfrag->offset == (df->offset + df->data_len) &&
963 df->data_seq + df->data_len == msk->write_seq;
966 static void __mptcp_mem_reclaim_partial(struct sock *sk)
968 int reclaimable = mptcp_sk(sk)->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
970 lockdep_assert_held_once(&sk->sk_lock.slock);
972 if (reclaimable > SK_MEM_QUANTUM)
973 __mptcp_rmem_reclaim(sk, reclaimable - 1);
975 sk_mem_reclaim_partial(sk);
978 static void mptcp_mem_reclaim_partial(struct sock *sk)
981 __mptcp_mem_reclaim_partial(sk);
982 mptcp_data_unlock(sk);
985 static void dfrag_uncharge(struct sock *sk, int len)
987 sk_mem_uncharge(sk, len);
988 sk_wmem_queued_add(sk, -len);
991 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
993 int len = dfrag->data_len + dfrag->overhead;
995 list_del(&dfrag->list);
996 dfrag_uncharge(sk, len);
997 put_page(dfrag->page);
1000 static void __mptcp_clean_una(struct sock *sk)
1002 struct mptcp_sock *msk = mptcp_sk(sk);
1003 struct mptcp_data_frag *dtmp, *dfrag;
1004 bool cleaned = false;
1007 /* on fallback we just need to ignore snd_una, as this is really
1010 if (__mptcp_check_fallback(msk))
1011 msk->snd_una = READ_ONCE(msk->snd_nxt);
1013 snd_una = msk->snd_una;
1014 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1015 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1018 if (unlikely(dfrag == msk->first_pending)) {
1019 /* in recovery mode can see ack after the current snd head */
1020 if (WARN_ON_ONCE(!msk->recovery))
1023 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1026 dfrag_clear(sk, dfrag);
1030 dfrag = mptcp_rtx_head(sk);
1031 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1032 u64 delta = snd_una - dfrag->data_seq;
1034 /* prevent wrap around in recovery mode */
1035 if (unlikely(delta > dfrag->already_sent)) {
1036 if (WARN_ON_ONCE(!msk->recovery))
1038 if (WARN_ON_ONCE(delta > dfrag->data_len))
1040 dfrag->already_sent += delta - dfrag->already_sent;
1043 dfrag->data_seq += delta;
1044 dfrag->offset += delta;
1045 dfrag->data_len -= delta;
1046 dfrag->already_sent -= delta;
1048 dfrag_uncharge(sk, delta);
1052 /* all retransmitted data acked, recovery completed */
1053 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1054 msk->recovery = false;
1057 if (cleaned && tcp_under_memory_pressure(sk))
1058 __mptcp_mem_reclaim_partial(sk);
1060 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1061 snd_una == READ_ONCE(msk->write_seq)) {
1062 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1063 mptcp_stop_timer(sk);
1065 mptcp_reset_timer(sk);
1069 static void __mptcp_clean_una_wakeup(struct sock *sk)
1071 lockdep_assert_held_once(&sk->sk_lock.slock);
1073 __mptcp_clean_una(sk);
1074 mptcp_write_space(sk);
1077 static void mptcp_clean_una_wakeup(struct sock *sk)
1079 mptcp_data_lock(sk);
1080 __mptcp_clean_una_wakeup(sk);
1081 mptcp_data_unlock(sk);
1084 static void mptcp_enter_memory_pressure(struct sock *sk)
1086 struct mptcp_subflow_context *subflow;
1087 struct mptcp_sock *msk = mptcp_sk(sk);
1090 sk_stream_moderate_sndbuf(sk);
1091 mptcp_for_each_subflow(msk, subflow) {
1092 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1095 tcp_enter_memory_pressure(ssk);
1096 sk_stream_moderate_sndbuf(ssk);
1101 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1104 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1106 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1107 pfrag, sk->sk_allocation)))
1110 mptcp_enter_memory_pressure(sk);
1114 static struct mptcp_data_frag *
1115 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1118 int offset = ALIGN(orig_offset, sizeof(long));
1119 struct mptcp_data_frag *dfrag;
1121 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1122 dfrag->data_len = 0;
1123 dfrag->data_seq = msk->write_seq;
1124 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1125 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1126 dfrag->already_sent = 0;
1127 dfrag->page = pfrag->page;
1132 struct mptcp_sendmsg_info {
1138 bool data_lock_held;
1141 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1144 u64 window_end = mptcp_wnd_end(msk);
1146 if (__mptcp_check_fallback(msk))
1149 if (!before64(data_seq + avail_size, window_end)) {
1150 u64 allowed_size = window_end - data_seq;
1152 return min_t(unsigned int, allowed_size, avail_size);
1158 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1160 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1164 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1168 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1170 struct sk_buff *skb;
1172 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1174 if (likely(__mptcp_add_ext(skb, gfp))) {
1175 skb_reserve(skb, MAX_TCP_HEADER);
1176 skb->ip_summed = CHECKSUM_PARTIAL;
1177 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1182 mptcp_enter_memory_pressure(sk);
1187 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1189 struct sk_buff *skb;
1191 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1195 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1196 tcp_skb_entail(ssk, skb);
1203 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1205 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1207 if (unlikely(tcp_under_memory_pressure(sk))) {
1209 __mptcp_mem_reclaim_partial(sk);
1211 mptcp_mem_reclaim_partial(sk);
1213 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1216 /* note: this always recompute the csum on the whole skb, even
1217 * if we just appended a single frag. More status info needed
1219 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1221 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1222 __wsum csum = ~csum_unfold(mpext->csum);
1223 int offset = skb->len - added;
1225 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1228 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1229 struct mptcp_data_frag *dfrag,
1230 struct mptcp_sendmsg_info *info)
1232 u64 data_seq = dfrag->data_seq + info->sent;
1233 int offset = dfrag->offset + info->sent;
1234 struct mptcp_sock *msk = mptcp_sk(sk);
1235 bool zero_window_probe = false;
1236 struct mptcp_ext *mpext = NULL;
1237 bool can_coalesce = false;
1238 bool reuse_skb = true;
1239 struct sk_buff *skb;
1243 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1244 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1246 if (WARN_ON_ONCE(info->sent > info->limit ||
1247 info->limit > dfrag->data_len))
1250 /* compute send limit */
1251 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1252 copy = info->size_goal;
1254 skb = tcp_write_queue_tail(ssk);
1255 if (skb && copy > skb->len) {
1256 /* Limit the write to the size available in the
1257 * current skb, if any, so that we create at most a new skb.
1258 * Explicitly tells TCP internals to avoid collapsing on later
1259 * queue management operation, to avoid breaking the ext <->
1260 * SSN association set here
1262 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1263 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1264 TCP_SKB_CB(skb)->eor = 1;
1268 i = skb_shinfo(skb)->nr_frags;
1269 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1270 if (!can_coalesce && i >= sysctl_max_skb_frags) {
1271 tcp_mark_push(tcp_sk(ssk), skb);
1278 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1282 i = skb_shinfo(skb)->nr_frags;
1284 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1287 /* Zero window and all data acked? Probe. */
1288 copy = mptcp_check_allowed_size(msk, data_seq, copy);
1290 u64 snd_una = READ_ONCE(msk->snd_una);
1292 if (snd_una != msk->snd_nxt) {
1293 tcp_remove_empty_skb(ssk);
1297 zero_window_probe = true;
1298 data_seq = snd_una - 1;
1301 /* all mptcp-level data is acked, no skbs should be present into the
1304 WARN_ON_ONCE(reuse_skb);
1307 copy = min_t(size_t, copy, info->limit - info->sent);
1308 if (!sk_wmem_schedule(ssk, copy)) {
1309 tcp_remove_empty_skb(ssk);
1314 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1316 get_page(dfrag->page);
1317 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1321 skb->data_len += copy;
1322 skb->truesize += copy;
1323 sk_wmem_queued_add(ssk, copy);
1324 sk_mem_charge(ssk, copy);
1325 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1326 TCP_SKB_CB(skb)->end_seq += copy;
1327 tcp_skb_pcount_set(skb, 0);
1329 /* on skb reuse we just need to update the DSS len */
1331 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1332 mpext->data_len += copy;
1333 WARN_ON_ONCE(zero_window_probe);
1337 memset(mpext, 0, sizeof(*mpext));
1338 mpext->data_seq = data_seq;
1339 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1340 mpext->data_len = copy;
1344 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1345 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1348 if (zero_window_probe) {
1349 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1351 if (READ_ONCE(msk->csum_enabled))
1352 mptcp_update_data_checksum(skb, copy);
1353 tcp_push_pending_frames(ssk);
1357 if (READ_ONCE(msk->csum_enabled))
1358 mptcp_update_data_checksum(skb, copy);
1359 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1363 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1364 sizeof(struct tcphdr) - \
1365 MAX_TCP_OPTION_SPACE - \
1366 sizeof(struct ipv6hdr) - \
1367 sizeof(struct frag_hdr))
1369 struct subflow_send_info {
1374 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1376 if (!subflow->stale)
1380 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1383 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1385 if (unlikely(subflow->stale)) {
1386 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1388 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1391 mptcp_subflow_set_active(subflow);
1393 return __mptcp_subflow_active(subflow);
1396 #define SSK_MODE_ACTIVE 0
1397 #define SSK_MODE_BACKUP 1
1398 #define SSK_MODE_MAX 2
1400 /* implement the mptcp packet scheduler;
1401 * returns the subflow that will transmit the next DSS
1402 * additionally updates the rtx timeout
1404 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1406 struct subflow_send_info send_info[SSK_MODE_MAX];
1407 struct mptcp_subflow_context *subflow;
1408 struct sock *sk = (struct sock *)msk;
1409 u32 pace, burst, wmem;
1410 int i, nr_active = 0;
1415 sock_owned_by_me(sk);
1417 if (__mptcp_check_fallback(msk)) {
1420 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1423 /* re-use last subflow, if the burst allow that */
1424 if (msk->last_snd && msk->snd_burst > 0 &&
1425 sk_stream_memory_free(msk->last_snd) &&
1426 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1427 mptcp_set_timeout(sk);
1428 return msk->last_snd;
1431 /* pick the subflow with the lower wmem/wspace ratio */
1432 for (i = 0; i < SSK_MODE_MAX; ++i) {
1433 send_info[i].ssk = NULL;
1434 send_info[i].linger_time = -1;
1437 mptcp_for_each_subflow(msk, subflow) {
1438 trace_mptcp_subflow_get_send(subflow);
1439 ssk = mptcp_subflow_tcp_sock(subflow);
1440 if (!mptcp_subflow_active(subflow))
1443 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1444 nr_active += !subflow->backup;
1445 pace = subflow->avg_pacing_rate;
1446 if (unlikely(!pace)) {
1447 /* init pacing rate from socket */
1448 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1449 pace = subflow->avg_pacing_rate;
1454 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1455 if (linger_time < send_info[subflow->backup].linger_time) {
1456 send_info[subflow->backup].ssk = ssk;
1457 send_info[subflow->backup].linger_time = linger_time;
1460 __mptcp_set_timeout(sk, tout);
1462 /* pick the best backup if no other subflow is active */
1464 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1466 /* According to the blest algorithm, to avoid HoL blocking for the
1467 * faster flow, we need to:
1468 * - estimate the faster flow linger time
1469 * - use the above to estimate the amount of byte transferred
1470 * by the faster flow
1471 * - check that the amount of queued data is greter than the above,
1472 * otherwise do not use the picked, slower, subflow
1473 * We select the subflow with the shorter estimated time to flush
1474 * the queued mem, which basically ensure the above. We just need
1475 * to check that subflow has a non empty cwin.
1477 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1478 if (!ssk || !sk_stream_memory_free(ssk) || !tcp_sk(ssk)->snd_wnd)
1481 burst = min_t(int, MPTCP_SEND_BURST_SIZE, tcp_sk(ssk)->snd_wnd);
1482 wmem = READ_ONCE(ssk->sk_wmem_queued);
1483 subflow = mptcp_subflow_ctx(ssk);
1484 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1485 READ_ONCE(ssk->sk_pacing_rate) * burst,
1487 msk->last_snd = ssk;
1488 msk->snd_burst = burst;
1492 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1494 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1498 static void mptcp_update_post_push(struct mptcp_sock *msk,
1499 struct mptcp_data_frag *dfrag,
1502 u64 snd_nxt_new = dfrag->data_seq;
1504 dfrag->already_sent += sent;
1506 msk->snd_burst -= sent;
1508 snd_nxt_new += dfrag->already_sent;
1510 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1511 * is recovering after a failover. In that event, this re-sends
1514 * Thus compute snd_nxt_new candidate based on
1515 * the dfrag->data_seq that was sent and the data
1516 * that has been handed to the subflow for transmission
1517 * and skip update in case it was old dfrag.
1519 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1520 msk->snd_nxt = snd_nxt_new;
1523 void mptcp_check_and_set_pending(struct sock *sk)
1525 if (mptcp_send_head(sk))
1526 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1529 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1531 struct sock *prev_ssk = NULL, *ssk = NULL;
1532 struct mptcp_sock *msk = mptcp_sk(sk);
1533 struct mptcp_sendmsg_info info = {
1536 struct mptcp_data_frag *dfrag;
1537 int len, copied = 0;
1539 while ((dfrag = mptcp_send_head(sk))) {
1540 info.sent = dfrag->already_sent;
1541 info.limit = dfrag->data_len;
1542 len = dfrag->data_len - dfrag->already_sent;
1547 ssk = mptcp_subflow_get_send(msk);
1549 /* First check. If the ssk has changed since
1550 * the last round, release prev_ssk
1552 if (ssk != prev_ssk && prev_ssk)
1553 mptcp_push_release(prev_ssk, &info);
1557 /* Need to lock the new subflow only if different
1558 * from the previous one, otherwise we are still
1559 * helding the relevant lock
1561 if (ssk != prev_ssk)
1564 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1566 mptcp_push_release(ssk, &info);
1574 mptcp_update_post_push(msk, dfrag, ret);
1576 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1579 /* at this point we held the socket lock for the last subflow we used */
1581 mptcp_push_release(ssk, &info);
1584 /* ensure the rtx timer is running */
1585 if (!mptcp_timer_pending(sk))
1586 mptcp_reset_timer(sk);
1588 __mptcp_check_send_data_fin(sk);
1591 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1593 struct mptcp_sock *msk = mptcp_sk(sk);
1594 struct mptcp_sendmsg_info info = {
1595 .data_lock_held = true,
1597 struct mptcp_data_frag *dfrag;
1598 struct sock *xmit_ssk;
1599 int len, copied = 0;
1603 while ((dfrag = mptcp_send_head(sk))) {
1604 info.sent = dfrag->already_sent;
1605 info.limit = dfrag->data_len;
1606 len = dfrag->data_len - dfrag->already_sent;
1610 /* the caller already invoked the packet scheduler,
1611 * check for a different subflow usage only after
1612 * spooling the first chunk of data
1614 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1617 if (xmit_ssk != ssk) {
1618 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1619 MPTCP_DELEGATE_SEND);
1623 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1632 mptcp_update_post_push(msk, dfrag, ret);
1634 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1638 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1639 * not going to flush it via release_sock()
1642 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1644 if (!mptcp_timer_pending(sk))
1645 mptcp_reset_timer(sk);
1647 if (msk->snd_data_fin_enable &&
1648 msk->snd_nxt + 1 == msk->write_seq)
1649 mptcp_schedule_work(sk);
1653 static void mptcp_set_nospace(struct sock *sk)
1655 /* enable autotune */
1656 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1658 /* will be cleared on avail space */
1659 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1662 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1664 struct mptcp_sock *msk = mptcp_sk(sk);
1665 struct page_frag *pfrag;
1670 /* we don't support FASTOPEN yet */
1671 if (msg->msg_flags & MSG_FASTOPEN)
1674 /* silently ignore everything else */
1675 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1679 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1681 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1682 ret = sk_stream_wait_connect(sk, &timeo);
1687 pfrag = sk_page_frag(sk);
1689 while (msg_data_left(msg)) {
1690 int total_ts, frag_truesize = 0;
1691 struct mptcp_data_frag *dfrag;
1692 bool dfrag_collapsed;
1693 size_t psize, offset;
1695 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1700 /* reuse tail pfrag, if possible, or carve a new one from the
1703 dfrag = mptcp_pending_tail(sk);
1704 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1705 if (!dfrag_collapsed) {
1706 if (!sk_stream_memory_free(sk))
1707 goto wait_for_memory;
1709 if (!mptcp_page_frag_refill(sk, pfrag))
1710 goto wait_for_memory;
1712 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1713 frag_truesize = dfrag->overhead;
1716 /* we do not bound vs wspace, to allow a single packet.
1717 * memory accounting will prevent execessive memory usage
1720 offset = dfrag->offset + dfrag->data_len;
1721 psize = pfrag->size - offset;
1722 psize = min_t(size_t, psize, msg_data_left(msg));
1723 total_ts = psize + frag_truesize;
1725 if (!sk_wmem_schedule(sk, total_ts))
1726 goto wait_for_memory;
1728 if (copy_page_from_iter(dfrag->page, offset, psize,
1729 &msg->msg_iter) != psize) {
1734 /* data successfully copied into the write queue */
1735 sk->sk_forward_alloc -= total_ts;
1737 dfrag->data_len += psize;
1738 frag_truesize += psize;
1739 pfrag->offset += frag_truesize;
1740 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1742 /* charge data on mptcp pending queue to the msk socket
1743 * Note: we charge such data both to sk and ssk
1745 sk_wmem_queued_add(sk, frag_truesize);
1746 if (!dfrag_collapsed) {
1747 get_page(dfrag->page);
1748 list_add_tail(&dfrag->list, &msk->rtx_queue);
1749 if (!msk->first_pending)
1750 WRITE_ONCE(msk->first_pending, dfrag);
1752 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1753 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1759 mptcp_set_nospace(sk);
1760 __mptcp_push_pending(sk, msg->msg_flags);
1761 ret = sk_stream_wait_memory(sk, &timeo);
1767 __mptcp_push_pending(sk, msg->msg_flags);
1771 return copied ? : ret;
1774 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1776 size_t len, int flags,
1777 struct scm_timestamping_internal *tss,
1780 struct sk_buff *skb, *tmp;
1783 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1784 u32 offset = MPTCP_SKB_CB(skb)->offset;
1785 u32 data_len = skb->len - offset;
1786 u32 count = min_t(size_t, len - copied, data_len);
1789 if (!(flags & MSG_TRUNC)) {
1790 err = skb_copy_datagram_msg(skb, offset, msg, count);
1791 if (unlikely(err < 0)) {
1798 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1799 tcp_update_recv_tstamps(skb, tss);
1800 *cmsg_flags |= MPTCP_CMSG_TS;
1805 if (count < data_len) {
1806 if (!(flags & MSG_PEEK)) {
1807 MPTCP_SKB_CB(skb)->offset += count;
1808 MPTCP_SKB_CB(skb)->map_seq += count;
1813 if (!(flags & MSG_PEEK)) {
1814 /* we will bulk release the skb memory later */
1815 skb->destructor = NULL;
1816 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1817 __skb_unlink(skb, &msk->receive_queue);
1828 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1830 * Only difference: Use highest rtt estimate of the subflows in use.
1832 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1834 struct mptcp_subflow_context *subflow;
1835 struct sock *sk = (struct sock *)msk;
1836 u32 time, advmss = 1;
1839 sock_owned_by_me(sk);
1844 msk->rcvq_space.copied += copied;
1846 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1847 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1849 rtt_us = msk->rcvq_space.rtt_us;
1850 if (rtt_us && time < (rtt_us >> 3))
1854 mptcp_for_each_subflow(msk, subflow) {
1855 const struct tcp_sock *tp;
1859 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1861 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1862 sf_advmss = READ_ONCE(tp->advmss);
1864 rtt_us = max(sf_rtt_us, rtt_us);
1865 advmss = max(sf_advmss, advmss);
1868 msk->rcvq_space.rtt_us = rtt_us;
1869 if (time < (rtt_us >> 3) || rtt_us == 0)
1872 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1875 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1876 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1880 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1882 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1884 do_div(grow, msk->rcvq_space.space);
1885 rcvwin += (grow << 1);
1887 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1888 while (tcp_win_from_space(sk, rcvmem) < advmss)
1891 do_div(rcvwin, advmss);
1892 rcvbuf = min_t(u64, rcvwin * rcvmem,
1893 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1895 if (rcvbuf > sk->sk_rcvbuf) {
1898 window_clamp = tcp_win_from_space(sk, rcvbuf);
1899 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1901 /* Make subflows follow along. If we do not do this, we
1902 * get drops at subflow level if skbs can't be moved to
1903 * the mptcp rx queue fast enough (announced rcv_win can
1904 * exceed ssk->sk_rcvbuf).
1906 mptcp_for_each_subflow(msk, subflow) {
1910 ssk = mptcp_subflow_tcp_sock(subflow);
1911 slow = lock_sock_fast(ssk);
1912 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1913 tcp_sk(ssk)->window_clamp = window_clamp;
1914 tcp_cleanup_rbuf(ssk, 1);
1915 unlock_sock_fast(ssk, slow);
1920 msk->rcvq_space.space = msk->rcvq_space.copied;
1922 msk->rcvq_space.copied = 0;
1923 msk->rcvq_space.time = mstamp;
1926 static void __mptcp_update_rmem(struct sock *sk)
1928 struct mptcp_sock *msk = mptcp_sk(sk);
1930 if (!msk->rmem_released)
1933 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1934 mptcp_rmem_uncharge(sk, msk->rmem_released);
1935 WRITE_ONCE(msk->rmem_released, 0);
1938 static void __mptcp_splice_receive_queue(struct sock *sk)
1940 struct mptcp_sock *msk = mptcp_sk(sk);
1942 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1945 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1947 struct sock *sk = (struct sock *)msk;
1948 unsigned int moved = 0;
1952 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1955 /* we can have data pending in the subflows only if the msk
1956 * receive buffer was full at subflow_data_ready() time,
1957 * that is an unlikely slow path.
1962 slowpath = lock_sock_fast(ssk);
1963 mptcp_data_lock(sk);
1964 __mptcp_update_rmem(sk);
1965 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1966 mptcp_data_unlock(sk);
1968 if (unlikely(ssk->sk_err))
1969 __mptcp_error_report(sk);
1970 unlock_sock_fast(ssk, slowpath);
1973 /* acquire the data lock only if some input data is pending */
1975 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1976 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1977 mptcp_data_lock(sk);
1978 __mptcp_update_rmem(sk);
1979 ret |= __mptcp_ofo_queue(msk);
1980 __mptcp_splice_receive_queue(sk);
1981 mptcp_data_unlock(sk);
1984 mptcp_check_data_fin((struct sock *)msk);
1985 return !skb_queue_empty(&msk->receive_queue);
1988 static unsigned int mptcp_inq_hint(const struct sock *sk)
1990 const struct mptcp_sock *msk = mptcp_sk(sk);
1991 const struct sk_buff *skb;
1993 skb = skb_peek(&msk->receive_queue);
1995 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
1997 if (hint_val >= INT_MAX)
2000 return (unsigned int)hint_val;
2003 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2009 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2010 int nonblock, int flags, int *addr_len)
2012 struct mptcp_sock *msk = mptcp_sk(sk);
2013 struct scm_timestamping_internal tss;
2014 int copied = 0, cmsg_flags = 0;
2018 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2019 if (unlikely(flags & MSG_ERRQUEUE))
2020 return inet_recv_error(sk, msg, len, addr_len);
2023 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2028 timeo = sock_rcvtimeo(sk, nonblock);
2030 len = min_t(size_t, len, INT_MAX);
2031 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2033 if (unlikely(msk->recvmsg_inq))
2034 cmsg_flags = MPTCP_CMSG_INQ;
2036 while (copied < len) {
2039 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2040 if (unlikely(bytes_read < 0)) {
2042 copied = bytes_read;
2046 copied += bytes_read;
2048 /* be sure to advertise window change */
2049 mptcp_cleanup_rbuf(msk);
2051 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2054 /* only the master socket status is relevant here. The exit
2055 * conditions mirror closely tcp_recvmsg()
2057 if (copied >= target)
2062 sk->sk_state == TCP_CLOSE ||
2063 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2065 signal_pending(current))
2069 copied = sock_error(sk);
2073 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2074 mptcp_check_for_eof(msk);
2076 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2077 /* race breaker: the shutdown could be after the
2078 * previous receive queue check
2080 if (__mptcp_move_skbs(msk))
2085 if (sk->sk_state == TCP_CLOSE) {
2095 if (signal_pending(current)) {
2096 copied = sock_intr_errno(timeo);
2101 pr_debug("block timeout %ld", timeo);
2102 sk_wait_data(sk, &timeo, NULL);
2106 if (cmsg_flags && copied >= 0) {
2107 if (cmsg_flags & MPTCP_CMSG_TS)
2108 tcp_recv_timestamp(msg, sk, &tss);
2110 if (cmsg_flags & MPTCP_CMSG_INQ) {
2111 unsigned int inq = mptcp_inq_hint(sk);
2113 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2117 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2118 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2119 skb_queue_empty(&msk->receive_queue), copied);
2120 if (!(flags & MSG_PEEK))
2121 mptcp_rcv_space_adjust(msk, copied);
2127 static void mptcp_retransmit_timer(struct timer_list *t)
2129 struct inet_connection_sock *icsk = from_timer(icsk, t,
2130 icsk_retransmit_timer);
2131 struct sock *sk = &icsk->icsk_inet.sk;
2132 struct mptcp_sock *msk = mptcp_sk(sk);
2135 if (!sock_owned_by_user(sk)) {
2136 /* we need a process context to retransmit */
2137 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2138 mptcp_schedule_work(sk);
2140 /* delegate our work to tcp_release_cb() */
2141 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2147 static void mptcp_timeout_timer(struct timer_list *t)
2149 struct sock *sk = from_timer(sk, t, sk_timer);
2151 mptcp_schedule_work(sk);
2155 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2158 * A backup subflow is returned only if that is the only kind available.
2160 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2162 struct sock *backup = NULL, *pick = NULL;
2163 struct mptcp_subflow_context *subflow;
2164 int min_stale_count = INT_MAX;
2166 sock_owned_by_me((const struct sock *)msk);
2168 if (__mptcp_check_fallback(msk))
2171 mptcp_for_each_subflow(msk, subflow) {
2172 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2174 if (!__mptcp_subflow_active(subflow))
2177 /* still data outstanding at TCP level? skip this */
2178 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2179 mptcp_pm_subflow_chk_stale(msk, ssk);
2180 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2184 if (subflow->backup) {
2197 /* use backup only if there are no progresses anywhere */
2198 return min_stale_count > 1 ? backup : NULL;
2201 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2204 iput(SOCK_INODE(msk->subflow));
2205 msk->subflow = NULL;
2209 bool __mptcp_retransmit_pending_data(struct sock *sk)
2211 struct mptcp_data_frag *cur, *rtx_head;
2212 struct mptcp_sock *msk = mptcp_sk(sk);
2214 if (__mptcp_check_fallback(mptcp_sk(sk)))
2217 if (tcp_rtx_and_write_queues_empty(sk))
2220 /* the closing socket has some data untransmitted and/or unacked:
2221 * some data in the mptcp rtx queue has not really xmitted yet.
2222 * keep it simple and re-inject the whole mptcp level rtx queue
2224 mptcp_data_lock(sk);
2225 __mptcp_clean_una_wakeup(sk);
2226 rtx_head = mptcp_rtx_head(sk);
2228 mptcp_data_unlock(sk);
2232 msk->recovery_snd_nxt = msk->snd_nxt;
2233 msk->recovery = true;
2234 mptcp_data_unlock(sk);
2236 msk->first_pending = rtx_head;
2239 /* be sure to clear the "sent status" on all re-injected fragments */
2240 list_for_each_entry(cur, &msk->rtx_queue, list) {
2241 if (!cur->already_sent)
2243 cur->already_sent = 0;
2249 /* flags for __mptcp_close_ssk() */
2250 #define MPTCP_CF_PUSH BIT(1)
2251 #define MPTCP_CF_FASTCLOSE BIT(2)
2253 /* subflow sockets can be either outgoing (connect) or incoming
2256 * Outgoing subflows use in-kernel sockets.
2257 * Incoming subflows do not have their own 'struct socket' allocated,
2258 * so we need to use tcp_close() after detaching them from the mptcp
2261 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2262 struct mptcp_subflow_context *subflow,
2265 struct mptcp_sock *msk = mptcp_sk(sk);
2266 bool need_push, dispose_it;
2268 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2270 list_del(&subflow->node);
2272 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2274 if (flags & MPTCP_CF_FASTCLOSE)
2275 subflow->send_fastclose = 1;
2277 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2279 tcp_disconnect(ssk, 0);
2280 msk->subflow->state = SS_UNCONNECTED;
2281 mptcp_subflow_ctx_reset(subflow);
2287 /* if we are invoked by the msk cleanup code, the subflow is
2293 subflow->disposable = 1;
2295 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2296 * the ssk has been already destroyed, we just need to release the
2297 * reference owned by msk;
2299 if (!inet_csk(ssk)->icsk_ulp_ops) {
2300 kfree_rcu(subflow, rcu);
2302 /* otherwise tcp will dispose of the ssk and subflow ctx */
2303 __tcp_close(ssk, 0);
2305 /* close acquired an extra ref */
2312 if (ssk == msk->first)
2316 if (ssk == msk->last_snd)
2317 msk->last_snd = NULL;
2320 __mptcp_push_pending(sk, 0);
2323 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2324 struct mptcp_subflow_context *subflow)
2326 if (sk->sk_state == TCP_ESTABLISHED)
2327 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2329 /* subflow aborted before reaching the fully_established status
2330 * attempt the creation of the next subflow
2332 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2334 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2337 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2342 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2344 struct mptcp_subflow_context *subflow, *tmp;
2348 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2349 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2351 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2354 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2355 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2358 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2362 static bool mptcp_check_close_timeout(const struct sock *sk)
2364 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2365 struct mptcp_subflow_context *subflow;
2367 if (delta >= TCP_TIMEWAIT_LEN)
2370 /* if all subflows are in closed status don't bother with additional
2373 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2374 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2381 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2383 struct mptcp_subflow_context *subflow, *tmp;
2384 struct sock *sk = &msk->sk.icsk_inet.sk;
2386 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2389 mptcp_token_destroy(msk);
2391 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2392 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2395 slow = lock_sock_fast(tcp_sk);
2396 if (tcp_sk->sk_state != TCP_CLOSE) {
2397 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2398 tcp_set_state(tcp_sk, TCP_CLOSE);
2400 unlock_sock_fast(tcp_sk, slow);
2403 inet_sk_state_store(sk, TCP_CLOSE);
2404 sk->sk_shutdown = SHUTDOWN_MASK;
2405 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2406 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2408 mptcp_close_wake_up(sk);
2411 static void __mptcp_retrans(struct sock *sk)
2413 struct mptcp_sock *msk = mptcp_sk(sk);
2414 struct mptcp_sendmsg_info info = {};
2415 struct mptcp_data_frag *dfrag;
2420 mptcp_clean_una_wakeup(sk);
2422 /* first check ssk: need to kick "stale" logic */
2423 ssk = mptcp_subflow_get_retrans(msk);
2424 dfrag = mptcp_rtx_head(sk);
2426 if (mptcp_data_fin_enabled(msk)) {
2427 struct inet_connection_sock *icsk = inet_csk(sk);
2429 icsk->icsk_retransmits++;
2430 mptcp_set_datafin_timeout(sk);
2431 mptcp_send_ack(msk);
2436 if (!mptcp_send_head(sk))
2447 /* limit retransmission to the bytes already sent on some subflows */
2449 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2450 while (info.sent < info.limit) {
2451 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2455 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2460 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2461 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2468 mptcp_check_and_set_pending(sk);
2470 if (!mptcp_timer_pending(sk))
2471 mptcp_reset_timer(sk);
2474 static void mptcp_worker(struct work_struct *work)
2476 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2477 struct sock *sk = &msk->sk.icsk_inet.sk;
2481 state = sk->sk_state;
2482 if (unlikely(state == TCP_CLOSE))
2485 mptcp_check_data_fin_ack(sk);
2487 mptcp_check_fastclose(msk);
2489 mptcp_pm_nl_work(msk);
2491 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2492 mptcp_check_for_eof(msk);
2494 __mptcp_check_send_data_fin(sk);
2495 mptcp_check_data_fin(sk);
2497 /* There is no point in keeping around an orphaned sk timedout or
2498 * closed, but we need the msk around to reply to incoming DATA_FIN,
2499 * even if it is orphaned and in FIN_WAIT2 state
2501 if (sock_flag(sk, SOCK_DEAD) &&
2502 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2503 inet_sk_state_store(sk, TCP_CLOSE);
2504 __mptcp_destroy_sock(sk);
2508 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2509 __mptcp_close_subflow(msk);
2511 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2512 __mptcp_retrans(sk);
2519 static int __mptcp_init_sock(struct sock *sk)
2521 struct mptcp_sock *msk = mptcp_sk(sk);
2523 INIT_LIST_HEAD(&msk->conn_list);
2524 INIT_LIST_HEAD(&msk->join_list);
2525 INIT_LIST_HEAD(&msk->rtx_queue);
2526 INIT_WORK(&msk->work, mptcp_worker);
2527 __skb_queue_head_init(&msk->receive_queue);
2528 msk->out_of_order_queue = RB_ROOT;
2529 msk->first_pending = NULL;
2530 msk->rmem_fwd_alloc = 0;
2531 WRITE_ONCE(msk->rmem_released, 0);
2532 msk->timer_ival = TCP_RTO_MIN;
2535 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2536 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2537 msk->recovery = false;
2539 mptcp_pm_data_init(msk);
2541 /* re-use the csk retrans timer for MPTCP-level retrans */
2542 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2543 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2548 static void mptcp_ca_reset(struct sock *sk)
2550 struct inet_connection_sock *icsk = inet_csk(sk);
2552 tcp_assign_congestion_control(sk);
2553 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2555 /* no need to keep a reference to the ops, the name will suffice */
2556 tcp_cleanup_congestion_control(sk);
2557 icsk->icsk_ca_ops = NULL;
2560 static int mptcp_init_sock(struct sock *sk)
2562 struct net *net = sock_net(sk);
2565 ret = __mptcp_init_sock(sk);
2569 if (!mptcp_is_enabled(net))
2570 return -ENOPROTOOPT;
2572 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2575 ret = __mptcp_socket_create(mptcp_sk(sk));
2579 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2580 * propagate the correct value
2584 sk_sockets_allocated_inc(sk);
2585 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2586 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2591 static void __mptcp_clear_xmit(struct sock *sk)
2593 struct mptcp_sock *msk = mptcp_sk(sk);
2594 struct mptcp_data_frag *dtmp, *dfrag;
2596 WRITE_ONCE(msk->first_pending, NULL);
2597 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2598 dfrag_clear(sk, dfrag);
2601 static void mptcp_cancel_work(struct sock *sk)
2603 struct mptcp_sock *msk = mptcp_sk(sk);
2605 if (cancel_work_sync(&msk->work))
2609 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2613 switch (ssk->sk_state) {
2615 if (!(how & RCV_SHUTDOWN))
2619 tcp_disconnect(ssk, O_NONBLOCK);
2622 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2623 pr_debug("Fallback");
2624 ssk->sk_shutdown |= how;
2625 tcp_shutdown(ssk, how);
2627 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2629 if (!mptcp_timer_pending(sk))
2630 mptcp_reset_timer(sk);
2638 static const unsigned char new_state[16] = {
2639 /* current state: new state: action: */
2640 [0 /* (Invalid) */] = TCP_CLOSE,
2641 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2642 [TCP_SYN_SENT] = TCP_CLOSE,
2643 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2644 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2645 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2646 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2647 [TCP_CLOSE] = TCP_CLOSE,
2648 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2649 [TCP_LAST_ACK] = TCP_LAST_ACK,
2650 [TCP_LISTEN] = TCP_CLOSE,
2651 [TCP_CLOSING] = TCP_CLOSING,
2652 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2655 static int mptcp_close_state(struct sock *sk)
2657 int next = (int)new_state[sk->sk_state];
2658 int ns = next & TCP_STATE_MASK;
2660 inet_sk_state_store(sk, ns);
2662 return next & TCP_ACTION_FIN;
2665 static void __mptcp_check_send_data_fin(struct sock *sk)
2667 struct mptcp_subflow_context *subflow;
2668 struct mptcp_sock *msk = mptcp_sk(sk);
2670 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2671 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2672 msk->snd_nxt, msk->write_seq);
2674 /* we still need to enqueue subflows or not really shutting down,
2677 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2678 mptcp_send_head(sk))
2681 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2683 /* fallback socket will not get data_fin/ack, can move to the next
2686 if (__mptcp_check_fallback(msk)) {
2687 WRITE_ONCE(msk->snd_una, msk->write_seq);
2688 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2689 inet_sk_state_store(sk, TCP_CLOSE);
2690 mptcp_close_wake_up(sk);
2691 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2692 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2696 mptcp_for_each_subflow(msk, subflow) {
2697 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2699 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2703 static void __mptcp_wr_shutdown(struct sock *sk)
2705 struct mptcp_sock *msk = mptcp_sk(sk);
2707 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2708 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2709 !!mptcp_send_head(sk));
2711 /* will be ignored by fallback sockets */
2712 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2713 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2715 __mptcp_check_send_data_fin(sk);
2718 static void __mptcp_destroy_sock(struct sock *sk)
2720 struct mptcp_subflow_context *subflow, *tmp;
2721 struct mptcp_sock *msk = mptcp_sk(sk);
2722 LIST_HEAD(conn_list);
2724 pr_debug("msk=%p", msk);
2728 /* join list will be eventually flushed (with rst) at sock lock release time*/
2729 list_splice_init(&msk->conn_list, &conn_list);
2731 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2732 sk_stop_timer(sk, &sk->sk_timer);
2735 /* clears msk->subflow, allowing the following loop to close
2736 * even the initial subflow
2738 mptcp_dispose_initial_subflow(msk);
2739 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2740 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2741 __mptcp_close_ssk(sk, ssk, subflow, 0);
2744 sk->sk_prot->destroy(sk);
2746 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2747 WARN_ON_ONCE(msk->rmem_released);
2748 sk_stream_kill_queues(sk);
2749 xfrm_sk_free_policy(sk);
2751 sk_refcnt_debug_release(sk);
2755 static void mptcp_close(struct sock *sk, long timeout)
2757 struct mptcp_subflow_context *subflow;
2758 bool do_cancel_work = false;
2761 sk->sk_shutdown = SHUTDOWN_MASK;
2763 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2764 inet_sk_state_store(sk, TCP_CLOSE);
2768 if (mptcp_close_state(sk))
2769 __mptcp_wr_shutdown(sk);
2771 sk_stream_wait_close(sk, timeout);
2774 /* orphan all the subflows */
2775 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2776 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2777 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2778 bool slow = lock_sock_fast_nested(ssk);
2781 unlock_sock_fast(ssk, slow);
2786 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2787 if (mptcp_sk(sk)->token)
2788 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2790 if (sk->sk_state == TCP_CLOSE) {
2791 __mptcp_destroy_sock(sk);
2792 do_cancel_work = true;
2794 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2798 mptcp_cancel_work(sk);
2803 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2805 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2806 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2807 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2809 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2810 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2813 msk6->saddr = ssk6->saddr;
2814 msk6->flow_label = ssk6->flow_label;
2818 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2819 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2820 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2821 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2822 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2823 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2826 static int mptcp_disconnect(struct sock *sk, int flags)
2828 struct mptcp_subflow_context *subflow;
2829 struct mptcp_sock *msk = mptcp_sk(sk);
2831 inet_sk_state_store(sk, TCP_CLOSE);
2833 mptcp_for_each_subflow(msk, subflow) {
2834 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2836 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_FASTCLOSE);
2839 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2840 sk_stop_timer(sk, &sk->sk_timer);
2842 if (mptcp_sk(sk)->token)
2843 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2845 mptcp_destroy_common(msk);
2846 msk->last_snd = NULL;
2847 WRITE_ONCE(msk->flags, 0);
2849 msk->push_pending = 0;
2850 msk->recovery = false;
2851 msk->can_ack = false;
2852 msk->fully_established = false;
2853 msk->rcv_data_fin = false;
2854 msk->snd_data_fin_enable = false;
2855 msk->rcv_fastclose = false;
2856 msk->use_64bit_ack = false;
2857 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2858 mptcp_pm_data_reset(msk);
2861 sk->sk_shutdown = 0;
2862 sk_error_report(sk);
2866 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2867 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2869 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2871 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2875 struct sock *mptcp_sk_clone(const struct sock *sk,
2876 const struct mptcp_options_received *mp_opt,
2877 struct request_sock *req)
2879 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2880 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2881 struct mptcp_sock *msk;
2887 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2888 if (nsk->sk_family == AF_INET6)
2889 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2892 __mptcp_init_sock(nsk);
2894 msk = mptcp_sk(nsk);
2895 msk->local_key = subflow_req->local_key;
2896 msk->token = subflow_req->token;
2897 msk->subflow = NULL;
2898 WRITE_ONCE(msk->fully_established, false);
2899 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
2900 WRITE_ONCE(msk->csum_enabled, true);
2902 msk->write_seq = subflow_req->idsn + 1;
2903 msk->snd_nxt = msk->write_seq;
2904 msk->snd_una = msk->write_seq;
2905 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2906 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2908 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
2909 msk->can_ack = true;
2910 msk->remote_key = mp_opt->sndr_key;
2911 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2913 WRITE_ONCE(msk->ack_seq, ack_seq);
2914 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2917 sock_reset_flag(nsk, SOCK_RCU_FREE);
2918 /* will be fully established after successful MPC subflow creation */
2919 inet_sk_state_store(nsk, TCP_SYN_RECV);
2921 security_inet_csk_clone(nsk, req);
2922 bh_unlock_sock(nsk);
2924 /* keep a single reference */
2929 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2931 const struct tcp_sock *tp = tcp_sk(ssk);
2933 msk->rcvq_space.copied = 0;
2934 msk->rcvq_space.rtt_us = 0;
2936 msk->rcvq_space.time = tp->tcp_mstamp;
2938 /* initial rcv_space offering made to peer */
2939 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2940 TCP_INIT_CWND * tp->advmss);
2941 if (msk->rcvq_space.space == 0)
2942 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2944 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2947 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2950 struct mptcp_sock *msk = mptcp_sk(sk);
2951 struct socket *listener;
2954 listener = __mptcp_nmpc_socket(msk);
2955 if (WARN_ON_ONCE(!listener)) {
2960 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2961 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2965 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2966 if (sk_is_mptcp(newsk)) {
2967 struct mptcp_subflow_context *subflow;
2968 struct sock *new_mptcp_sock;
2970 subflow = mptcp_subflow_ctx(newsk);
2971 new_mptcp_sock = subflow->conn;
2973 /* is_mptcp should be false if subflow->conn is missing, see
2974 * subflow_syn_recv_sock()
2976 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2977 tcp_sk(newsk)->is_mptcp = 0;
2981 /* acquire the 2nd reference for the owning socket */
2982 sock_hold(new_mptcp_sock);
2983 newsk = new_mptcp_sock;
2984 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2986 MPTCP_INC_STATS(sock_net(sk),
2987 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2991 newsk->sk_kern_sock = kern;
2995 void mptcp_destroy_common(struct mptcp_sock *msk)
2997 struct sock *sk = (struct sock *)msk;
2999 __mptcp_clear_xmit(sk);
3001 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3002 mptcp_data_lock(sk);
3003 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3004 __skb_queue_purge(&sk->sk_receive_queue);
3005 skb_rbtree_purge(&msk->out_of_order_queue);
3006 mptcp_data_unlock(sk);
3008 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3009 * inet_sock_destruct() will dispose it
3011 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3012 msk->rmem_fwd_alloc = 0;
3013 mptcp_token_destroy(msk);
3014 mptcp_pm_free_anno_list(msk);
3017 static void mptcp_destroy(struct sock *sk)
3019 struct mptcp_sock *msk = mptcp_sk(sk);
3021 mptcp_destroy_common(msk);
3022 sk_sockets_allocated_dec(sk);
3025 void __mptcp_data_acked(struct sock *sk)
3027 if (!sock_owned_by_user(sk))
3028 __mptcp_clean_una(sk);
3030 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3032 if (mptcp_pending_data_fin_ack(sk))
3033 mptcp_schedule_work(sk);
3036 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3038 if (!mptcp_send_head(sk))
3041 if (!sock_owned_by_user(sk)) {
3042 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3044 if (xmit_ssk == ssk)
3045 __mptcp_subflow_push_pending(sk, ssk);
3047 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3049 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3053 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3054 BIT(MPTCP_RETRANSMIT) | \
3055 BIT(MPTCP_FLUSH_JOIN_LIST))
3057 /* processes deferred events and flush wmem */
3058 static void mptcp_release_cb(struct sock *sk)
3059 __must_hold(&sk->sk_lock.slock)
3061 struct mptcp_sock *msk = mptcp_sk(sk);
3064 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3069 /* the following actions acquire the subflow socket lock
3071 * 1) can't be invoked in atomic scope
3072 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3073 * datapath acquires the msk socket spinlock while helding
3074 * the subflow socket lock
3076 msk->push_pending = 0;
3077 msk->cb_flags &= ~flags;
3078 spin_unlock_bh(&sk->sk_lock.slock);
3079 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3080 __mptcp_flush_join_list(sk);
3081 if (flags & BIT(MPTCP_PUSH_PENDING))
3082 __mptcp_push_pending(sk, 0);
3083 if (flags & BIT(MPTCP_RETRANSMIT))
3084 __mptcp_retrans(sk);
3087 spin_lock_bh(&sk->sk_lock.slock);
3090 /* be sure to set the current sk state before tacking actions
3091 * depending on sk_state
3093 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3094 __mptcp_set_connected(sk);
3095 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3096 __mptcp_clean_una_wakeup(sk);
3097 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3098 __mptcp_error_report(sk);
3100 __mptcp_update_rmem(sk);
3103 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3104 * TCP can't schedule delack timer before the subflow is fully established.
3105 * MPTCP uses the delack timer to do 3rd ack retransmissions
3107 static void schedule_3rdack_retransmission(struct sock *ssk)
3109 struct inet_connection_sock *icsk = inet_csk(ssk);
3110 struct tcp_sock *tp = tcp_sk(ssk);
3111 unsigned long timeout;
3113 if (mptcp_subflow_ctx(ssk)->fully_established)
3116 /* reschedule with a timeout above RTT, as we must look only for drop */
3118 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3120 timeout = TCP_TIMEOUT_INIT;
3123 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3124 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3125 icsk->icsk_ack.timeout = timeout;
3126 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3129 void mptcp_subflow_process_delegated(struct sock *ssk)
3131 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3132 struct sock *sk = subflow->conn;
3134 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3135 mptcp_data_lock(sk);
3136 if (!sock_owned_by_user(sk))
3137 __mptcp_subflow_push_pending(sk, ssk);
3139 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3140 mptcp_data_unlock(sk);
3141 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3143 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3144 schedule_3rdack_retransmission(ssk);
3145 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3149 static int mptcp_hash(struct sock *sk)
3151 /* should never be called,
3152 * we hash the TCP subflows not the master socket
3158 static void mptcp_unhash(struct sock *sk)
3160 /* called from sk_common_release(), but nothing to do here */
3163 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3165 struct mptcp_sock *msk = mptcp_sk(sk);
3166 struct socket *ssock;
3168 ssock = __mptcp_nmpc_socket(msk);
3169 pr_debug("msk=%p, subflow=%p", msk, ssock);
3170 if (WARN_ON_ONCE(!ssock))
3173 return inet_csk_get_port(ssock->sk, snum);
3176 void mptcp_finish_connect(struct sock *ssk)
3178 struct mptcp_subflow_context *subflow;
3179 struct mptcp_sock *msk;
3183 subflow = mptcp_subflow_ctx(ssk);
3187 pr_debug("msk=%p, token=%u", sk, subflow->token);
3189 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3191 subflow->map_seq = ack_seq;
3192 subflow->map_subflow_seq = 1;
3194 /* the socket is not connected yet, no msk/subflow ops can access/race
3195 * accessing the field below
3197 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3198 WRITE_ONCE(msk->local_key, subflow->local_key);
3199 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3200 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3201 WRITE_ONCE(msk->ack_seq, ack_seq);
3202 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3203 WRITE_ONCE(msk->can_ack, 1);
3204 WRITE_ONCE(msk->snd_una, msk->write_seq);
3206 mptcp_pm_new_connection(msk, ssk, 0);
3208 mptcp_rcv_space_init(msk, ssk);
3211 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3213 write_lock_bh(&sk->sk_callback_lock);
3214 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3215 sk_set_socket(sk, parent);
3216 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3217 write_unlock_bh(&sk->sk_callback_lock);
3220 bool mptcp_finish_join(struct sock *ssk)
3222 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3223 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3224 struct sock *parent = (void *)msk;
3227 pr_debug("msk=%p, subflow=%p", msk, subflow);
3229 /* mptcp socket already closing? */
3230 if (!mptcp_is_fully_established(parent)) {
3231 subflow->reset_reason = MPTCP_RST_EMPTCP;
3235 if (!msk->pm.server_side)
3238 if (!mptcp_pm_allow_new_subflow(msk))
3239 goto err_prohibited;
3241 if (WARN_ON_ONCE(!list_empty(&subflow->node)))
3242 goto err_prohibited;
3244 /* active connections are already on conn_list.
3245 * If we can't acquire msk socket lock here, let the release callback
3248 mptcp_data_lock(parent);
3249 if (!sock_owned_by_user(parent)) {
3250 ret = __mptcp_finish_join(msk, ssk);
3253 list_add_tail(&subflow->node, &msk->conn_list);
3257 list_add_tail(&subflow->node, &msk->join_list);
3258 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3260 mptcp_data_unlock(parent);
3264 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3268 subflow->map_seq = READ_ONCE(msk->ack_seq);
3271 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3275 static void mptcp_shutdown(struct sock *sk, int how)
3277 pr_debug("sk=%p, how=%d", sk, how);
3279 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3280 __mptcp_wr_shutdown(sk);
3283 static int mptcp_forward_alloc_get(const struct sock *sk)
3285 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3288 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3290 const struct sock *sk = (void *)msk;
3293 if (sk->sk_state == TCP_LISTEN)
3296 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3299 delta = msk->write_seq - v;
3300 if (__mptcp_check_fallback(msk) && msk->first) {
3301 struct tcp_sock *tp = tcp_sk(msk->first);
3303 /* the first subflow is disconnected after close - see
3304 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3305 * so ignore that status, too.
3307 if (!((1 << msk->first->sk_state) &
3308 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3309 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3311 if (delta > INT_MAX)
3317 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3319 struct mptcp_sock *msk = mptcp_sk(sk);
3325 if (sk->sk_state == TCP_LISTEN)
3329 __mptcp_move_skbs(msk);
3330 answ = mptcp_inq_hint(sk);
3334 slow = lock_sock_fast(sk);
3335 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3336 unlock_sock_fast(sk, slow);
3339 slow = lock_sock_fast(sk);
3340 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3341 unlock_sock_fast(sk, slow);
3344 return -ENOIOCTLCMD;
3347 return put_user(answ, (int __user *)arg);
3350 static struct proto mptcp_prot = {
3352 .owner = THIS_MODULE,
3353 .init = mptcp_init_sock,
3354 .disconnect = mptcp_disconnect,
3355 .close = mptcp_close,
3356 .accept = mptcp_accept,
3357 .setsockopt = mptcp_setsockopt,
3358 .getsockopt = mptcp_getsockopt,
3359 .shutdown = mptcp_shutdown,
3360 .destroy = mptcp_destroy,
3361 .sendmsg = mptcp_sendmsg,
3362 .ioctl = mptcp_ioctl,
3363 .recvmsg = mptcp_recvmsg,
3364 .release_cb = mptcp_release_cb,
3366 .unhash = mptcp_unhash,
3367 .get_port = mptcp_get_port,
3368 .forward_alloc_get = mptcp_forward_alloc_get,
3369 .sockets_allocated = &mptcp_sockets_allocated,
3370 .memory_allocated = &tcp_memory_allocated,
3371 .memory_pressure = &tcp_memory_pressure,
3372 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3373 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3374 .sysctl_mem = sysctl_tcp_mem,
3375 .obj_size = sizeof(struct mptcp_sock),
3376 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3377 .no_autobind = true,
3380 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3382 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3383 struct socket *ssock;
3386 lock_sock(sock->sk);
3387 ssock = __mptcp_nmpc_socket(msk);
3393 err = ssock->ops->bind(ssock, uaddr, addr_len);
3395 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3398 release_sock(sock->sk);
3402 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3403 struct mptcp_subflow_context *subflow)
3405 subflow->request_mptcp = 0;
3406 __mptcp_do_fallback(msk);
3409 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3410 int addr_len, int flags)
3412 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3413 struct mptcp_subflow_context *subflow;
3414 struct socket *ssock;
3417 lock_sock(sock->sk);
3419 if (addr_len < sizeof(uaddr->sa_family))
3422 if (uaddr->sa_family == AF_UNSPEC) {
3423 err = mptcp_disconnect(sock->sk, flags);
3424 sock->state = err ? SS_DISCONNECTING : SS_UNCONNECTED;
3429 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3430 /* pending connection or invalid state, let existing subflow
3433 ssock = msk->subflow;
3437 ssock = __mptcp_nmpc_socket(msk);
3441 mptcp_token_destroy(msk);
3442 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3443 subflow = mptcp_subflow_ctx(ssock->sk);
3444 #ifdef CONFIG_TCP_MD5SIG
3445 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3448 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3449 mptcp_subflow_early_fallback(msk, subflow);
3451 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3452 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3453 mptcp_subflow_early_fallback(msk, subflow);
3455 if (likely(!__mptcp_check_fallback(msk)))
3456 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3459 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3460 sock->state = ssock->state;
3462 /* on successful connect, the msk state will be moved to established by
3463 * subflow_finish_connect()
3465 if (!err || err == -EINPROGRESS)
3466 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3468 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3471 release_sock(sock->sk);
3475 static int mptcp_listen(struct socket *sock, int backlog)
3477 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3478 struct socket *ssock;
3481 pr_debug("msk=%p", msk);
3483 lock_sock(sock->sk);
3484 ssock = __mptcp_nmpc_socket(msk);
3490 mptcp_token_destroy(msk);
3491 inet_sk_state_store(sock->sk, TCP_LISTEN);
3492 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3494 err = ssock->ops->listen(ssock, backlog);
3495 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3497 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3500 release_sock(sock->sk);
3504 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3505 int flags, bool kern)
3507 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3508 struct socket *ssock;
3511 pr_debug("msk=%p", msk);
3513 ssock = __mptcp_nmpc_socket(msk);
3517 err = ssock->ops->accept(sock, newsock, flags, kern);
3518 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3519 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3520 struct mptcp_subflow_context *subflow;
3521 struct sock *newsk = newsock->sk;
3525 /* PM/worker can now acquire the first subflow socket
3526 * lock without racing with listener queue cleanup,
3527 * we can notify it, if needed.
3529 * Even if remote has reset the initial subflow by now
3530 * the refcnt is still at least one.
3532 subflow = mptcp_subflow_ctx(msk->first);
3533 list_add(&subflow->node, &msk->conn_list);
3534 sock_hold(msk->first);
3535 if (mptcp_is_fully_established(newsk))
3536 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3538 mptcp_copy_inaddrs(newsk, msk->first);
3539 mptcp_rcv_space_init(msk, msk->first);
3540 mptcp_propagate_sndbuf(newsk, msk->first);
3542 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3543 * This is needed so NOSPACE flag can be set from tcp stack.
3545 mptcp_for_each_subflow(msk, subflow) {
3546 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3548 if (!ssk->sk_socket)
3549 mptcp_sock_graft(ssk, newsock);
3551 release_sock(newsk);
3557 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3559 /* Concurrent splices from sk_receive_queue into receive_queue will
3560 * always show at least one non-empty queue when checked in this order.
3562 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
3563 skb_queue_empty_lockless(&msk->receive_queue))
3566 return EPOLLIN | EPOLLRDNORM;
3569 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3571 struct sock *sk = (struct sock *)msk;
3573 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3574 return EPOLLOUT | EPOLLWRNORM;
3576 if (sk_stream_is_writeable(sk))
3577 return EPOLLOUT | EPOLLWRNORM;
3579 mptcp_set_nospace(sk);
3580 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3581 if (sk_stream_is_writeable(sk))
3582 return EPOLLOUT | EPOLLWRNORM;
3587 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3588 struct poll_table_struct *wait)
3590 struct sock *sk = sock->sk;
3591 struct mptcp_sock *msk;
3596 sock_poll_wait(file, sock, wait);
3598 state = inet_sk_state_load(sk);
3599 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3600 if (state == TCP_LISTEN) {
3601 if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk))
3604 return inet_csk_listen_poll(msk->subflow->sk);
3607 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3608 mask |= mptcp_check_readable(msk);
3609 mask |= mptcp_check_writeable(msk);
3611 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3613 if (sk->sk_shutdown & RCV_SHUTDOWN)
3614 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3616 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3624 static const struct proto_ops mptcp_stream_ops = {
3626 .owner = THIS_MODULE,
3627 .release = inet_release,
3629 .connect = mptcp_stream_connect,
3630 .socketpair = sock_no_socketpair,
3631 .accept = mptcp_stream_accept,
3632 .getname = inet_getname,
3634 .ioctl = inet_ioctl,
3635 .gettstamp = sock_gettstamp,
3636 .listen = mptcp_listen,
3637 .shutdown = inet_shutdown,
3638 .setsockopt = sock_common_setsockopt,
3639 .getsockopt = sock_common_getsockopt,
3640 .sendmsg = inet_sendmsg,
3641 .recvmsg = inet_recvmsg,
3642 .mmap = sock_no_mmap,
3643 .sendpage = inet_sendpage,
3646 static struct inet_protosw mptcp_protosw = {
3647 .type = SOCK_STREAM,
3648 .protocol = IPPROTO_MPTCP,
3649 .prot = &mptcp_prot,
3650 .ops = &mptcp_stream_ops,
3651 .flags = INET_PROTOSW_ICSK,
3654 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3656 struct mptcp_delegated_action *delegated;
3657 struct mptcp_subflow_context *subflow;
3660 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3661 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3662 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3664 bh_lock_sock_nested(ssk);
3665 if (!sock_owned_by_user(ssk) &&
3666 mptcp_subflow_has_delegated_action(subflow))
3667 mptcp_subflow_process_delegated(ssk);
3668 /* ... elsewhere tcp_release_cb_override already processed
3669 * the action or will do at next release_sock().
3670 * In both case must dequeue the subflow here - on the same
3671 * CPU that scheduled it.
3673 bh_unlock_sock(ssk);
3676 if (++work_done == budget)
3680 /* always provide a 0 'work_done' argument, so that napi_complete_done
3681 * will not try accessing the NULL napi->dev ptr
3683 napi_complete_done(napi, 0);
3687 void __init mptcp_proto_init(void)
3689 struct mptcp_delegated_action *delegated;
3692 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3694 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3695 panic("Failed to allocate MPTCP pcpu counter\n");
3697 init_dummy_netdev(&mptcp_napi_dev);
3698 for_each_possible_cpu(cpu) {
3699 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3700 INIT_LIST_HEAD(&delegated->head);
3701 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3703 napi_enable(&delegated->napi);
3706 mptcp_subflow_init();
3710 if (proto_register(&mptcp_prot, 1) != 0)
3711 panic("Failed to register MPTCP proto.\n");
3713 inet_register_protosw(&mptcp_protosw);
3715 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3718 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3719 static const struct proto_ops mptcp_v6_stream_ops = {
3721 .owner = THIS_MODULE,
3722 .release = inet6_release,
3724 .connect = mptcp_stream_connect,
3725 .socketpair = sock_no_socketpair,
3726 .accept = mptcp_stream_accept,
3727 .getname = inet6_getname,
3729 .ioctl = inet6_ioctl,
3730 .gettstamp = sock_gettstamp,
3731 .listen = mptcp_listen,
3732 .shutdown = inet_shutdown,
3733 .setsockopt = sock_common_setsockopt,
3734 .getsockopt = sock_common_getsockopt,
3735 .sendmsg = inet6_sendmsg,
3736 .recvmsg = inet6_recvmsg,
3737 .mmap = sock_no_mmap,
3738 .sendpage = inet_sendpage,
3739 #ifdef CONFIG_COMPAT
3740 .compat_ioctl = inet6_compat_ioctl,
3744 static struct proto mptcp_v6_prot;
3746 static void mptcp_v6_destroy(struct sock *sk)
3749 inet6_destroy_sock(sk);
3752 static struct inet_protosw mptcp_v6_protosw = {
3753 .type = SOCK_STREAM,
3754 .protocol = IPPROTO_MPTCP,
3755 .prot = &mptcp_v6_prot,
3756 .ops = &mptcp_v6_stream_ops,
3757 .flags = INET_PROTOSW_ICSK,
3760 int __init mptcp_proto_v6_init(void)
3764 mptcp_v6_prot = mptcp_prot;
3765 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3766 mptcp_v6_prot.slab = NULL;
3767 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3768 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3770 err = proto_register(&mptcp_v6_prot, 1);
3774 err = inet6_register_protosw(&mptcp_v6_protosw);
3776 proto_unregister(&mptcp_v6_prot);