1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
40 MPTCP_CMSG_TS = BIT(0),
41 MPTCP_CMSG_INQ = BIT(1),
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void mptcp_check_send_data_fin(struct sock *sk);
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
52 /* Returns end sequence number of the receiver's advertised window */
53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
55 return READ_ONCE(msk->wnd_end);
58 static bool mptcp_is_tcpsk(struct sock *sk)
60 struct socket *sock = sk->sk_socket;
62 if (unlikely(sk->sk_prot == &tcp_prot)) {
63 /* we are being invoked after mptcp_accept() has
64 * accepted a non-mp-capable flow: sk is a tcp_sk,
67 * Hand the socket over to tcp so all further socket ops
70 WRITE_ONCE(sock->ops, &inet_stream_ops);
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 WRITE_ONCE(sock->ops, &inet6_stream_ops);
82 static int __mptcp_socket_create(struct mptcp_sock *msk)
84 struct mptcp_subflow_context *subflow;
85 struct sock *sk = (struct sock *)msk;
89 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
93 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
94 WRITE_ONCE(msk->first, ssock->sk);
95 subflow = mptcp_subflow_ctx(ssock->sk);
96 list_add(&subflow->node, &msk->conn_list);
98 subflow->request_mptcp = 1;
99 subflow->subflow_id = msk->subflow_id++;
101 /* This is the first subflow, always with id 0 */
102 subflow->local_id_valid = 1;
103 mptcp_sock_graft(msk->first, sk->sk_socket);
104 iput(SOCK_INODE(ssock));
109 /* If the MPC handshake is not started, returns the first subflow,
110 * eventually allocating it.
112 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
114 struct sock *sk = (struct sock *)msk;
117 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
118 return ERR_PTR(-EINVAL);
121 ret = __mptcp_socket_create(msk);
125 mptcp_sockopt_sync(msk, msk->first);
131 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
133 sk_drops_add(sk, skb);
137 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
139 WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
140 mptcp_sk(sk)->rmem_fwd_alloc + size);
143 static void mptcp_rmem_charge(struct sock *sk, int size)
145 mptcp_rmem_fwd_alloc_add(sk, -size);
148 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
149 struct sk_buff *from)
154 if (MPTCP_SKB_CB(from)->offset ||
155 !skb_try_coalesce(to, from, &fragstolen, &delta))
158 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
159 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
160 to->len, MPTCP_SKB_CB(from)->end_seq);
161 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
163 /* note the fwd memory can reach a negative value after accounting
164 * for the delta, but the later skb free will restore a non
167 atomic_add(delta, &sk->sk_rmem_alloc);
168 mptcp_rmem_charge(sk, delta);
169 kfree_skb_partial(from, fragstolen);
174 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
175 struct sk_buff *from)
177 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
180 return mptcp_try_coalesce((struct sock *)msk, to, from);
183 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
185 amount >>= PAGE_SHIFT;
186 mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
187 __sk_mem_reduce_allocated(sk, amount);
190 static void mptcp_rmem_uncharge(struct sock *sk, int size)
192 struct mptcp_sock *msk = mptcp_sk(sk);
195 mptcp_rmem_fwd_alloc_add(sk, size);
196 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
198 /* see sk_mem_uncharge() for the rationale behind the following schema */
199 if (unlikely(reclaimable >= PAGE_SIZE))
200 __mptcp_rmem_reclaim(sk, reclaimable);
203 static void mptcp_rfree(struct sk_buff *skb)
205 unsigned int len = skb->truesize;
206 struct sock *sk = skb->sk;
208 atomic_sub(len, &sk->sk_rmem_alloc);
209 mptcp_rmem_uncharge(sk, len);
212 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
216 skb->destructor = mptcp_rfree;
217 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
218 mptcp_rmem_charge(sk, skb->truesize);
221 /* "inspired" by tcp_data_queue_ofo(), main differences:
223 * - don't cope with sacks
225 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
227 struct sock *sk = (struct sock *)msk;
228 struct rb_node **p, *parent;
229 u64 seq, end_seq, max_seq;
230 struct sk_buff *skb1;
232 seq = MPTCP_SKB_CB(skb)->map_seq;
233 end_seq = MPTCP_SKB_CB(skb)->end_seq;
234 max_seq = atomic64_read(&msk->rcv_wnd_sent);
236 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
237 RB_EMPTY_ROOT(&msk->out_of_order_queue));
238 if (after64(end_seq, max_seq)) {
241 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
242 (unsigned long long)end_seq - (unsigned long)max_seq,
243 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
248 p = &msk->out_of_order_queue.rb_node;
249 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
250 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
251 rb_link_node(&skb->rbnode, NULL, p);
252 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
253 msk->ooo_last_skb = skb;
257 /* with 2 subflows, adding at end of ooo queue is quite likely
258 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
260 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
261 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
266 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
267 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
268 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
269 parent = &msk->ooo_last_skb->rbnode;
270 p = &parent->rb_right;
274 /* Find place to insert this segment. Handle overlaps on the way. */
278 skb1 = rb_to_skb(parent);
279 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
280 p = &parent->rb_left;
283 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
284 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
285 /* All the bits are present. Drop. */
287 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
290 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
294 * continue traversing
297 /* skb's seq == skb1's seq and skb covers skb1.
298 * Replace skb1 with skb.
300 rb_replace_node(&skb1->rbnode, &skb->rbnode,
301 &msk->out_of_order_queue);
302 mptcp_drop(sk, skb1);
303 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
306 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
307 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
310 p = &parent->rb_right;
314 /* Insert segment into RB tree. */
315 rb_link_node(&skb->rbnode, parent, p);
316 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
319 /* Remove other segments covered by skb. */
320 while ((skb1 = skb_rb_next(skb)) != NULL) {
321 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
323 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
324 mptcp_drop(sk, skb1);
325 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
327 /* If there is no skb after us, we are the last_skb ! */
329 msk->ooo_last_skb = skb;
333 mptcp_set_owner_r(skb, sk);
336 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
338 struct mptcp_sock *msk = mptcp_sk(sk);
341 if (size <= msk->rmem_fwd_alloc)
344 size -= msk->rmem_fwd_alloc;
345 amt = sk_mem_pages(size);
346 amount = amt << PAGE_SHIFT;
347 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
350 mptcp_rmem_fwd_alloc_add(sk, amount);
354 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
355 struct sk_buff *skb, unsigned int offset,
358 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
359 struct sock *sk = (struct sock *)msk;
360 struct sk_buff *tail;
363 __skb_unlink(skb, &ssk->sk_receive_queue);
368 /* try to fetch required memory from subflow */
369 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
372 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
374 /* the skb map_seq accounts for the skb offset:
375 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
378 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
379 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
380 MPTCP_SKB_CB(skb)->offset = offset;
381 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
383 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
385 msk->bytes_received += copy_len;
386 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
387 tail = skb_peek_tail(&sk->sk_receive_queue);
388 if (tail && mptcp_try_coalesce(sk, tail, skb))
391 mptcp_set_owner_r(skb, sk);
392 __skb_queue_tail(&sk->sk_receive_queue, skb);
394 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
395 mptcp_data_queue_ofo(msk, skb);
399 /* old data, keep it simple and drop the whole pkt, sender
400 * will retransmit as needed, if needed.
402 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
408 static void mptcp_stop_timer(struct sock *sk)
410 struct inet_connection_sock *icsk = inet_csk(sk);
412 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
413 mptcp_sk(sk)->timer_ival = 0;
416 static void mptcp_close_wake_up(struct sock *sk)
418 if (sock_flag(sk, SOCK_DEAD))
421 sk->sk_state_change(sk);
422 if (sk->sk_shutdown == SHUTDOWN_MASK ||
423 sk->sk_state == TCP_CLOSE)
424 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
426 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
429 static bool mptcp_pending_data_fin_ack(struct sock *sk)
431 struct mptcp_sock *msk = mptcp_sk(sk);
433 return ((1 << sk->sk_state) &
434 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
435 msk->write_seq == READ_ONCE(msk->snd_una);
438 static void mptcp_check_data_fin_ack(struct sock *sk)
440 struct mptcp_sock *msk = mptcp_sk(sk);
442 /* Look for an acknowledged DATA_FIN */
443 if (mptcp_pending_data_fin_ack(sk)) {
444 WRITE_ONCE(msk->snd_data_fin_enable, 0);
446 switch (sk->sk_state) {
448 inet_sk_state_store(sk, TCP_FIN_WAIT2);
452 inet_sk_state_store(sk, TCP_CLOSE);
456 mptcp_close_wake_up(sk);
460 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
462 struct mptcp_sock *msk = mptcp_sk(sk);
464 if (READ_ONCE(msk->rcv_data_fin) &&
465 ((1 << sk->sk_state) &
466 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
467 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
469 if (msk->ack_seq == rcv_data_fin_seq) {
471 *seq = rcv_data_fin_seq;
480 static void mptcp_set_datafin_timeout(struct sock *sk)
482 struct inet_connection_sock *icsk = inet_csk(sk);
485 retransmits = min_t(u32, icsk->icsk_retransmits,
486 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
488 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
491 static void __mptcp_set_timeout(struct sock *sk, long tout)
493 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
496 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
498 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
500 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
501 inet_csk(ssk)->icsk_timeout - jiffies : 0;
504 static void mptcp_set_timeout(struct sock *sk)
506 struct mptcp_subflow_context *subflow;
509 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
510 tout = max(tout, mptcp_timeout_from_subflow(subflow));
511 __mptcp_set_timeout(sk, tout);
514 static inline bool tcp_can_send_ack(const struct sock *ssk)
516 return !((1 << inet_sk_state_load(ssk)) &
517 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
520 void __mptcp_subflow_send_ack(struct sock *ssk)
522 if (tcp_can_send_ack(ssk))
526 static void mptcp_subflow_send_ack(struct sock *ssk)
530 slow = lock_sock_fast(ssk);
531 __mptcp_subflow_send_ack(ssk);
532 unlock_sock_fast(ssk, slow);
535 static void mptcp_send_ack(struct mptcp_sock *msk)
537 struct mptcp_subflow_context *subflow;
539 mptcp_for_each_subflow(msk, subflow)
540 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
543 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
547 slow = lock_sock_fast(ssk);
548 if (tcp_can_send_ack(ssk))
549 tcp_cleanup_rbuf(ssk, 1);
550 unlock_sock_fast(ssk, slow);
553 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
555 const struct inet_connection_sock *icsk = inet_csk(ssk);
556 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
557 const struct tcp_sock *tp = tcp_sk(ssk);
559 return (ack_pending & ICSK_ACK_SCHED) &&
560 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
561 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
562 (rx_empty && ack_pending &
563 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
566 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
568 int old_space = READ_ONCE(msk->old_wspace);
569 struct mptcp_subflow_context *subflow;
570 struct sock *sk = (struct sock *)msk;
571 int space = __mptcp_space(sk);
572 bool cleanup, rx_empty;
574 cleanup = (space > 0) && (space >= (old_space << 1));
575 rx_empty = !__mptcp_rmem(sk);
577 mptcp_for_each_subflow(msk, subflow) {
578 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
580 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
581 mptcp_subflow_cleanup_rbuf(ssk);
585 static bool mptcp_check_data_fin(struct sock *sk)
587 struct mptcp_sock *msk = mptcp_sk(sk);
588 u64 rcv_data_fin_seq;
591 /* Need to ack a DATA_FIN received from a peer while this side
592 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
593 * msk->rcv_data_fin was set when parsing the incoming options
594 * at the subflow level and the msk lock was not held, so this
595 * is the first opportunity to act on the DATA_FIN and change
598 * If we are caught up to the sequence number of the incoming
599 * DATA_FIN, send the DATA_ACK now and do state transition. If
600 * not caught up, do nothing and let the recv code send DATA_ACK
604 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
605 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
606 WRITE_ONCE(msk->rcv_data_fin, 0);
608 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
609 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
611 switch (sk->sk_state) {
612 case TCP_ESTABLISHED:
613 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
616 inet_sk_state_store(sk, TCP_CLOSING);
619 inet_sk_state_store(sk, TCP_CLOSE);
622 /* Other states not expected */
628 if (!__mptcp_check_fallback(msk))
630 mptcp_close_wake_up(sk);
635 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
639 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
640 struct sock *sk = (struct sock *)msk;
641 unsigned int moved = 0;
642 bool more_data_avail;
647 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
649 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
650 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
652 if (unlikely(ssk_rbuf > sk_rbuf)) {
653 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
658 pr_debug("msk=%p ssk=%p", msk, ssk);
661 u32 map_remaining, offset;
662 u32 seq = tp->copied_seq;
666 /* try to move as much data as available */
667 map_remaining = subflow->map_data_len -
668 mptcp_subflow_get_map_offset(subflow);
670 skb = skb_peek(&ssk->sk_receive_queue);
672 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
673 * a different CPU can have already processed the pending
674 * data, stop here or we can enter an infinite loop
681 if (__mptcp_check_fallback(msk)) {
682 /* Under fallback skbs have no MPTCP extension and TCP could
683 * collapse them between the dummy map creation and the
684 * current dequeue. Be sure to adjust the map size.
686 map_remaining = skb->len;
687 subflow->map_data_len = skb->len;
690 offset = seq - TCP_SKB_CB(skb)->seq;
691 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
697 if (offset < skb->len) {
698 size_t len = skb->len - offset;
703 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
707 if (WARN_ON_ONCE(map_remaining < len))
711 sk_eat_skb(ssk, skb);
715 WRITE_ONCE(tp->copied_seq, seq);
716 more_data_avail = mptcp_subflow_data_available(ssk);
718 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
722 } while (more_data_avail);
728 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
730 struct sock *sk = (struct sock *)msk;
731 struct sk_buff *skb, *tail;
736 p = rb_first(&msk->out_of_order_queue);
737 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
740 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
744 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
746 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
749 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
753 end_seq = MPTCP_SKB_CB(skb)->end_seq;
754 tail = skb_peek_tail(&sk->sk_receive_queue);
755 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
756 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
758 /* skip overlapping data, if any */
759 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
760 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
762 MPTCP_SKB_CB(skb)->offset += delta;
763 MPTCP_SKB_CB(skb)->map_seq += delta;
764 __skb_queue_tail(&sk->sk_receive_queue, skb);
766 msk->bytes_received += end_seq - msk->ack_seq;
767 msk->ack_seq = end_seq;
773 void __mptcp_error_report(struct sock *sk)
775 struct mptcp_subflow_context *subflow;
776 struct mptcp_sock *msk = mptcp_sk(sk);
778 mptcp_for_each_subflow(msk, subflow) {
779 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
780 int err = sock_error(ssk);
786 /* only propagate errors on fallen-back sockets or
789 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(msk))
792 /* We need to propagate only transition to CLOSE state.
793 * Orphaned socket will see such state change via
794 * subflow_sched_work_if_closed() and that path will properly
795 * destroy the msk as needed.
797 ssk_state = inet_sk_state_load(ssk);
798 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
799 inet_sk_state_store(sk, ssk_state);
800 WRITE_ONCE(sk->sk_err, -err);
802 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
809 /* In most cases we will be able to lock the mptcp socket. If its already
810 * owned, we need to defer to the work queue to avoid ABBA deadlock.
812 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
814 struct sock *sk = (struct sock *)msk;
815 unsigned int moved = 0;
817 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
818 __mptcp_ofo_queue(msk);
819 if (unlikely(ssk->sk_err)) {
820 if (!sock_owned_by_user(sk))
821 __mptcp_error_report(sk);
823 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
826 /* If the moves have caught up with the DATA_FIN sequence number
827 * it's time to ack the DATA_FIN and change socket state, but
828 * this is not a good place to change state. Let the workqueue
831 if (mptcp_pending_data_fin(sk, NULL))
832 mptcp_schedule_work(sk);
836 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
838 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
839 struct mptcp_sock *msk = mptcp_sk(sk);
840 int sk_rbuf, ssk_rbuf;
842 /* The peer can send data while we are shutting down this
843 * subflow at msk destruction time, but we must avoid enqueuing
844 * more data to the msk receive queue
846 if (unlikely(subflow->disposable))
849 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
850 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
851 if (unlikely(ssk_rbuf > sk_rbuf))
854 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
855 if (__mptcp_rmem(sk) > sk_rbuf) {
856 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
860 /* Wake-up the reader only for in-sequence data */
862 if (move_skbs_to_msk(msk, ssk))
863 sk->sk_data_ready(sk);
865 mptcp_data_unlock(sk);
868 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
870 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
871 WRITE_ONCE(msk->allow_infinite_fallback, false);
872 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
875 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
877 struct sock *sk = (struct sock *)msk;
879 if (sk->sk_state != TCP_ESTABLISHED)
882 /* attach to msk socket only after we are sure we will deal with it
885 if (sk->sk_socket && !ssk->sk_socket)
886 mptcp_sock_graft(ssk, sk->sk_socket);
888 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
889 mptcp_sockopt_sync_locked(msk, ssk);
890 mptcp_subflow_joined(msk, ssk);
894 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
896 struct mptcp_subflow_context *tmp, *subflow;
897 struct mptcp_sock *msk = mptcp_sk(sk);
899 list_for_each_entry_safe(subflow, tmp, join_list, node) {
900 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
901 bool slow = lock_sock_fast(ssk);
903 list_move_tail(&subflow->node, &msk->conn_list);
904 if (!__mptcp_finish_join(msk, ssk))
905 mptcp_subflow_reset(ssk);
906 unlock_sock_fast(ssk, slow);
910 static bool mptcp_timer_pending(struct sock *sk)
912 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
915 static void mptcp_reset_timer(struct sock *sk)
917 struct inet_connection_sock *icsk = inet_csk(sk);
920 /* prevent rescheduling on close */
921 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
924 tout = mptcp_sk(sk)->timer_ival;
925 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
928 bool mptcp_schedule_work(struct sock *sk)
930 if (inet_sk_state_load(sk) != TCP_CLOSE &&
931 schedule_work(&mptcp_sk(sk)->work)) {
932 /* each subflow already holds a reference to the sk, and the
933 * workqueue is invoked by a subflow, so sk can't go away here.
941 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
943 struct mptcp_subflow_context *subflow;
945 msk_owned_by_me(msk);
947 mptcp_for_each_subflow(msk, subflow) {
948 if (READ_ONCE(subflow->data_avail))
949 return mptcp_subflow_tcp_sock(subflow);
955 static bool mptcp_skb_can_collapse_to(u64 write_seq,
956 const struct sk_buff *skb,
957 const struct mptcp_ext *mpext)
959 if (!tcp_skb_can_collapse_to(skb))
962 /* can collapse only if MPTCP level sequence is in order and this
963 * mapping has not been xmitted yet
965 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
969 /* we can append data to the given data frag if:
970 * - there is space available in the backing page_frag
971 * - the data frag tail matches the current page_frag free offset
972 * - the data frag end sequence number matches the current write seq
974 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
975 const struct page_frag *pfrag,
976 const struct mptcp_data_frag *df)
978 return df && pfrag->page == df->page &&
979 pfrag->size - pfrag->offset > 0 &&
980 pfrag->offset == (df->offset + df->data_len) &&
981 df->data_seq + df->data_len == msk->write_seq;
984 static void dfrag_uncharge(struct sock *sk, int len)
986 sk_mem_uncharge(sk, len);
987 sk_wmem_queued_add(sk, -len);
990 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
992 int len = dfrag->data_len + dfrag->overhead;
994 list_del(&dfrag->list);
995 dfrag_uncharge(sk, len);
996 put_page(dfrag->page);
999 static void __mptcp_clean_una(struct sock *sk)
1001 struct mptcp_sock *msk = mptcp_sk(sk);
1002 struct mptcp_data_frag *dtmp, *dfrag;
1005 snd_una = msk->snd_una;
1006 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1007 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1010 if (unlikely(dfrag == msk->first_pending)) {
1011 /* in recovery mode can see ack after the current snd head */
1012 if (WARN_ON_ONCE(!msk->recovery))
1015 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1018 dfrag_clear(sk, dfrag);
1021 dfrag = mptcp_rtx_head(sk);
1022 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1023 u64 delta = snd_una - dfrag->data_seq;
1025 /* prevent wrap around in recovery mode */
1026 if (unlikely(delta > dfrag->already_sent)) {
1027 if (WARN_ON_ONCE(!msk->recovery))
1029 if (WARN_ON_ONCE(delta > dfrag->data_len))
1031 dfrag->already_sent += delta - dfrag->already_sent;
1034 dfrag->data_seq += delta;
1035 dfrag->offset += delta;
1036 dfrag->data_len -= delta;
1037 dfrag->already_sent -= delta;
1039 dfrag_uncharge(sk, delta);
1042 /* all retransmitted data acked, recovery completed */
1043 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1044 msk->recovery = false;
1047 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1048 snd_una == READ_ONCE(msk->write_seq)) {
1049 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1050 mptcp_stop_timer(sk);
1052 mptcp_reset_timer(sk);
1056 static void __mptcp_clean_una_wakeup(struct sock *sk)
1058 lockdep_assert_held_once(&sk->sk_lock.slock);
1060 __mptcp_clean_una(sk);
1061 mptcp_write_space(sk);
1064 static void mptcp_clean_una_wakeup(struct sock *sk)
1066 mptcp_data_lock(sk);
1067 __mptcp_clean_una_wakeup(sk);
1068 mptcp_data_unlock(sk);
1071 static void mptcp_enter_memory_pressure(struct sock *sk)
1073 struct mptcp_subflow_context *subflow;
1074 struct mptcp_sock *msk = mptcp_sk(sk);
1077 sk_stream_moderate_sndbuf(sk);
1078 mptcp_for_each_subflow(msk, subflow) {
1079 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1082 tcp_enter_memory_pressure(ssk);
1083 sk_stream_moderate_sndbuf(ssk);
1088 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1091 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1093 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1094 pfrag, sk->sk_allocation)))
1097 mptcp_enter_memory_pressure(sk);
1101 static struct mptcp_data_frag *
1102 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1105 int offset = ALIGN(orig_offset, sizeof(long));
1106 struct mptcp_data_frag *dfrag;
1108 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1109 dfrag->data_len = 0;
1110 dfrag->data_seq = msk->write_seq;
1111 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1112 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1113 dfrag->already_sent = 0;
1114 dfrag->page = pfrag->page;
1119 struct mptcp_sendmsg_info {
1125 bool data_lock_held;
1128 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1129 u64 data_seq, int avail_size)
1131 u64 window_end = mptcp_wnd_end(msk);
1134 if (__mptcp_check_fallback(msk))
1137 mptcp_snd_wnd = window_end - data_seq;
1138 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1140 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1141 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1142 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1148 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1150 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1154 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1158 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1160 struct sk_buff *skb;
1162 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1164 if (likely(__mptcp_add_ext(skb, gfp))) {
1165 skb_reserve(skb, MAX_TCP_HEADER);
1166 skb->ip_summed = CHECKSUM_PARTIAL;
1167 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1172 mptcp_enter_memory_pressure(sk);
1177 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1179 struct sk_buff *skb;
1181 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1185 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1186 tcp_skb_entail(ssk, skb);
1189 tcp_skb_tsorted_anchor_cleanup(skb);
1194 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1196 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1198 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1201 /* note: this always recompute the csum on the whole skb, even
1202 * if we just appended a single frag. More status info needed
1204 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1206 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1207 __wsum csum = ~csum_unfold(mpext->csum);
1208 int offset = skb->len - added;
1210 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1213 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1215 struct mptcp_ext *mpext)
1220 mpext->infinite_map = 1;
1221 mpext->data_len = 0;
1223 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1224 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1226 mptcp_do_fallback(ssk);
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 if (unlikely(!__tcp_can_send(ssk)))
1254 /* compute send limit */
1255 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1256 copy = info->size_goal;
1258 skb = tcp_write_queue_tail(ssk);
1259 if (skb && copy > skb->len) {
1260 /* Limit the write to the size available in the
1261 * current skb, if any, so that we create at most a new skb.
1262 * Explicitly tells TCP internals to avoid collapsing on later
1263 * queue management operation, to avoid breaking the ext <->
1264 * SSN association set here
1266 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1267 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1268 TCP_SKB_CB(skb)->eor = 1;
1272 i = skb_shinfo(skb)->nr_frags;
1273 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1274 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1275 tcp_mark_push(tcp_sk(ssk), skb);
1282 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1286 i = skb_shinfo(skb)->nr_frags;
1288 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1291 /* Zero window and all data acked? Probe. */
1292 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1294 u64 snd_una = READ_ONCE(msk->snd_una);
1296 if (snd_una != msk->snd_nxt) {
1297 tcp_remove_empty_skb(ssk);
1301 zero_window_probe = true;
1302 data_seq = snd_una - 1;
1305 /* all mptcp-level data is acked, no skbs should be present into the
1308 WARN_ON_ONCE(reuse_skb);
1311 copy = min_t(size_t, copy, info->limit - info->sent);
1312 if (!sk_wmem_schedule(ssk, copy)) {
1313 tcp_remove_empty_skb(ssk);
1318 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1320 get_page(dfrag->page);
1321 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1325 skb->data_len += copy;
1326 skb->truesize += copy;
1327 sk_wmem_queued_add(ssk, copy);
1328 sk_mem_charge(ssk, copy);
1329 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1330 TCP_SKB_CB(skb)->end_seq += copy;
1331 tcp_skb_pcount_set(skb, 0);
1333 /* on skb reuse we just need to update the DSS len */
1335 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1336 mpext->data_len += copy;
1337 WARN_ON_ONCE(zero_window_probe);
1341 memset(mpext, 0, sizeof(*mpext));
1342 mpext->data_seq = data_seq;
1343 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1344 mpext->data_len = copy;
1348 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1349 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1352 if (zero_window_probe) {
1353 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1355 if (READ_ONCE(msk->csum_enabled))
1356 mptcp_update_data_checksum(skb, copy);
1357 tcp_push_pending_frames(ssk);
1361 if (READ_ONCE(msk->csum_enabled))
1362 mptcp_update_data_checksum(skb, copy);
1363 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1364 mptcp_update_infinite_map(msk, ssk, mpext);
1365 trace_mptcp_sendmsg_frag(mpext);
1366 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1370 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1371 sizeof(struct tcphdr) - \
1372 MAX_TCP_OPTION_SPACE - \
1373 sizeof(struct ipv6hdr) - \
1374 sizeof(struct frag_hdr))
1376 struct subflow_send_info {
1381 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1383 if (!subflow->stale)
1387 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1390 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1392 if (unlikely(subflow->stale)) {
1393 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1395 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1398 mptcp_subflow_set_active(subflow);
1400 return __mptcp_subflow_active(subflow);
1403 #define SSK_MODE_ACTIVE 0
1404 #define SSK_MODE_BACKUP 1
1405 #define SSK_MODE_MAX 2
1407 /* implement the mptcp packet scheduler;
1408 * returns the subflow that will transmit the next DSS
1409 * additionally updates the rtx timeout
1411 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1413 struct subflow_send_info send_info[SSK_MODE_MAX];
1414 struct mptcp_subflow_context *subflow;
1415 struct sock *sk = (struct sock *)msk;
1416 u32 pace, burst, wmem;
1417 int i, nr_active = 0;
1422 /* pick the subflow with the lower wmem/wspace ratio */
1423 for (i = 0; i < SSK_MODE_MAX; ++i) {
1424 send_info[i].ssk = NULL;
1425 send_info[i].linger_time = -1;
1428 mptcp_for_each_subflow(msk, subflow) {
1429 trace_mptcp_subflow_get_send(subflow);
1430 ssk = mptcp_subflow_tcp_sock(subflow);
1431 if (!mptcp_subflow_active(subflow))
1434 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1435 nr_active += !subflow->backup;
1436 pace = subflow->avg_pacing_rate;
1437 if (unlikely(!pace)) {
1438 /* init pacing rate from socket */
1439 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1440 pace = subflow->avg_pacing_rate;
1445 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1446 if (linger_time < send_info[subflow->backup].linger_time) {
1447 send_info[subflow->backup].ssk = ssk;
1448 send_info[subflow->backup].linger_time = linger_time;
1451 __mptcp_set_timeout(sk, tout);
1453 /* pick the best backup if no other subflow is active */
1455 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1457 /* According to the blest algorithm, to avoid HoL blocking for the
1458 * faster flow, we need to:
1459 * - estimate the faster flow linger time
1460 * - use the above to estimate the amount of byte transferred
1461 * by the faster flow
1462 * - check that the amount of queued data is greter than the above,
1463 * otherwise do not use the picked, slower, subflow
1464 * We select the subflow with the shorter estimated time to flush
1465 * the queued mem, which basically ensure the above. We just need
1466 * to check that subflow has a non empty cwin.
1468 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1469 if (!ssk || !sk_stream_memory_free(ssk))
1472 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1473 wmem = READ_ONCE(ssk->sk_wmem_queued);
1477 subflow = mptcp_subflow_ctx(ssk);
1478 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1479 READ_ONCE(ssk->sk_pacing_rate) * burst,
1481 msk->snd_burst = burst;
1485 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1487 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1491 static void mptcp_update_post_push(struct mptcp_sock *msk,
1492 struct mptcp_data_frag *dfrag,
1495 u64 snd_nxt_new = dfrag->data_seq;
1497 dfrag->already_sent += sent;
1499 msk->snd_burst -= sent;
1501 snd_nxt_new += dfrag->already_sent;
1503 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1504 * is recovering after a failover. In that event, this re-sends
1507 * Thus compute snd_nxt_new candidate based on
1508 * the dfrag->data_seq that was sent and the data
1509 * that has been handed to the subflow for transmission
1510 * and skip update in case it was old dfrag.
1512 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1513 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1514 msk->snd_nxt = snd_nxt_new;
1518 void mptcp_check_and_set_pending(struct sock *sk)
1520 if (mptcp_send_head(sk))
1521 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1524 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1525 struct mptcp_sendmsg_info *info)
1527 struct mptcp_sock *msk = mptcp_sk(sk);
1528 struct mptcp_data_frag *dfrag;
1529 int len, copied = 0, err = 0;
1531 while ((dfrag = mptcp_send_head(sk))) {
1532 info->sent = dfrag->already_sent;
1533 info->limit = dfrag->data_len;
1534 len = dfrag->data_len - dfrag->already_sent;
1538 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1540 err = copied ? : ret;
1548 mptcp_update_post_push(msk, dfrag, ret);
1550 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1552 if (msk->snd_burst <= 0 ||
1553 !sk_stream_memory_free(ssk) ||
1554 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1558 mptcp_set_timeout(sk);
1566 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1568 struct sock *prev_ssk = NULL, *ssk = NULL;
1569 struct mptcp_sock *msk = mptcp_sk(sk);
1570 struct mptcp_sendmsg_info info = {
1573 bool do_check_data_fin = false;
1576 while (mptcp_send_head(sk) && (push_count > 0)) {
1577 struct mptcp_subflow_context *subflow;
1580 if (mptcp_sched_get_send(msk))
1585 mptcp_for_each_subflow(msk, subflow) {
1586 if (READ_ONCE(subflow->scheduled)) {
1587 mptcp_subflow_set_scheduled(subflow, false);
1590 ssk = mptcp_subflow_tcp_sock(subflow);
1591 if (ssk != prev_ssk) {
1592 /* First check. If the ssk has changed since
1593 * the last round, release prev_ssk
1596 mptcp_push_release(prev_ssk, &info);
1598 /* Need to lock the new subflow only if different
1599 * from the previous one, otherwise we are still
1600 * helding the relevant lock
1607 ret = __subflow_push_pending(sk, ssk, &info);
1609 if (ret != -EAGAIN ||
1610 (1 << ssk->sk_state) &
1611 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1615 do_check_data_fin = true;
1620 /* at this point we held the socket lock for the last subflow we used */
1622 mptcp_push_release(ssk, &info);
1624 /* ensure the rtx timer is running */
1625 if (!mptcp_timer_pending(sk))
1626 mptcp_reset_timer(sk);
1627 if (do_check_data_fin)
1628 mptcp_check_send_data_fin(sk);
1631 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1633 struct mptcp_sock *msk = mptcp_sk(sk);
1634 struct mptcp_sendmsg_info info = {
1635 .data_lock_held = true,
1637 bool keep_pushing = true;
1638 struct sock *xmit_ssk;
1642 while (mptcp_send_head(sk) && keep_pushing) {
1643 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1646 /* check for a different subflow usage only after
1647 * spooling the first chunk of data
1650 mptcp_subflow_set_scheduled(subflow, false);
1651 ret = __subflow_push_pending(sk, ssk, &info);
1659 if (mptcp_sched_get_send(msk))
1662 if (READ_ONCE(subflow->scheduled)) {
1663 mptcp_subflow_set_scheduled(subflow, false);
1664 ret = __subflow_push_pending(sk, ssk, &info);
1666 keep_pushing = false;
1670 mptcp_for_each_subflow(msk, subflow) {
1671 if (READ_ONCE(subflow->scheduled)) {
1672 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1673 if (xmit_ssk != ssk) {
1674 mptcp_subflow_delegate(subflow,
1675 MPTCP_DELEGATE_SEND);
1676 keep_pushing = false;
1683 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1684 * not going to flush it via release_sock()
1687 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1689 if (!mptcp_timer_pending(sk))
1690 mptcp_reset_timer(sk);
1692 if (msk->snd_data_fin_enable &&
1693 msk->snd_nxt + 1 == msk->write_seq)
1694 mptcp_schedule_work(sk);
1698 static void mptcp_set_nospace(struct sock *sk)
1700 /* enable autotune */
1701 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1703 /* will be cleared on avail space */
1704 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1707 static int mptcp_disconnect(struct sock *sk, int flags);
1709 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1710 size_t len, int *copied_syn)
1712 unsigned int saved_flags = msg->msg_flags;
1713 struct mptcp_sock *msk = mptcp_sk(sk);
1717 /* on flags based fastopen the mptcp is supposed to create the
1718 * first subflow right now. Otherwise we are in the defer_connect
1719 * path, and the first subflow must be already present.
1720 * Since the defer_connect flag is cleared after the first succsful
1721 * fastopen attempt, no need to check for additional subflow status.
1723 if (msg->msg_flags & MSG_FASTOPEN) {
1724 ssk = __mptcp_nmpc_sk(msk);
1726 return PTR_ERR(ssk);
1734 msg->msg_flags |= MSG_DONTWAIT;
1735 msk->fastopening = 1;
1736 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1737 msk->fastopening = 0;
1738 msg->msg_flags = saved_flags;
1741 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1742 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1743 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1744 msg->msg_namelen, msg->msg_flags, 1);
1746 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1747 * case of any error, except timeout or signal
1749 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1751 } else if (ret && ret != -EINPROGRESS) {
1752 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1753 * __inet_stream_connect() can fail, due to looking check,
1754 * see mptcp_disconnect().
1755 * Attempt it again outside the problematic scope.
1757 if (!mptcp_disconnect(sk, 0))
1758 sk->sk_socket->state = SS_UNCONNECTED;
1760 inet_clear_bit(DEFER_CONNECT, sk);
1765 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1767 struct mptcp_sock *msk = mptcp_sk(sk);
1768 struct page_frag *pfrag;
1773 /* silently ignore everything else */
1774 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1778 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1779 msg->msg_flags & MSG_FASTOPEN)) {
1782 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1783 copied += copied_syn;
1784 if (ret == -EINPROGRESS && copied_syn > 0)
1790 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1792 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1793 ret = sk_stream_wait_connect(sk, &timeo);
1799 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1802 pfrag = sk_page_frag(sk);
1804 while (msg_data_left(msg)) {
1805 int total_ts, frag_truesize = 0;
1806 struct mptcp_data_frag *dfrag;
1807 bool dfrag_collapsed;
1808 size_t psize, offset;
1810 /* reuse tail pfrag, if possible, or carve a new one from the
1813 dfrag = mptcp_pending_tail(sk);
1814 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1815 if (!dfrag_collapsed) {
1816 if (!sk_stream_memory_free(sk))
1817 goto wait_for_memory;
1819 if (!mptcp_page_frag_refill(sk, pfrag))
1820 goto wait_for_memory;
1822 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1823 frag_truesize = dfrag->overhead;
1826 /* we do not bound vs wspace, to allow a single packet.
1827 * memory accounting will prevent execessive memory usage
1830 offset = dfrag->offset + dfrag->data_len;
1831 psize = pfrag->size - offset;
1832 psize = min_t(size_t, psize, msg_data_left(msg));
1833 total_ts = psize + frag_truesize;
1835 if (!sk_wmem_schedule(sk, total_ts))
1836 goto wait_for_memory;
1838 if (copy_page_from_iter(dfrag->page, offset, psize,
1839 &msg->msg_iter) != psize) {
1844 /* data successfully copied into the write queue */
1845 sk_forward_alloc_add(sk, -total_ts);
1847 dfrag->data_len += psize;
1848 frag_truesize += psize;
1849 pfrag->offset += frag_truesize;
1850 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1852 /* charge data on mptcp pending queue to the msk socket
1853 * Note: we charge such data both to sk and ssk
1855 sk_wmem_queued_add(sk, frag_truesize);
1856 if (!dfrag_collapsed) {
1857 get_page(dfrag->page);
1858 list_add_tail(&dfrag->list, &msk->rtx_queue);
1859 if (!msk->first_pending)
1860 WRITE_ONCE(msk->first_pending, dfrag);
1862 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1863 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1869 mptcp_set_nospace(sk);
1870 __mptcp_push_pending(sk, msg->msg_flags);
1871 ret = sk_stream_wait_memory(sk, &timeo);
1877 __mptcp_push_pending(sk, msg->msg_flags);
1887 copied = sk_stream_error(sk, msg->msg_flags, ret);
1891 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1893 size_t len, int flags,
1894 struct scm_timestamping_internal *tss,
1897 struct sk_buff *skb, *tmp;
1900 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1901 u32 offset = MPTCP_SKB_CB(skb)->offset;
1902 u32 data_len = skb->len - offset;
1903 u32 count = min_t(size_t, len - copied, data_len);
1906 if (!(flags & MSG_TRUNC)) {
1907 err = skb_copy_datagram_msg(skb, offset, msg, count);
1908 if (unlikely(err < 0)) {
1915 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1916 tcp_update_recv_tstamps(skb, tss);
1917 *cmsg_flags |= MPTCP_CMSG_TS;
1922 if (count < data_len) {
1923 if (!(flags & MSG_PEEK)) {
1924 MPTCP_SKB_CB(skb)->offset += count;
1925 MPTCP_SKB_CB(skb)->map_seq += count;
1930 if (!(flags & MSG_PEEK)) {
1931 /* we will bulk release the skb memory later */
1932 skb->destructor = NULL;
1933 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1934 __skb_unlink(skb, &msk->receive_queue);
1945 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1947 * Only difference: Use highest rtt estimate of the subflows in use.
1949 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1951 struct mptcp_subflow_context *subflow;
1952 struct sock *sk = (struct sock *)msk;
1953 u8 scaling_ratio = U8_MAX;
1954 u32 time, advmss = 1;
1957 msk_owned_by_me(msk);
1962 msk->rcvq_space.copied += copied;
1964 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1965 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1967 rtt_us = msk->rcvq_space.rtt_us;
1968 if (rtt_us && time < (rtt_us >> 3))
1972 mptcp_for_each_subflow(msk, subflow) {
1973 const struct tcp_sock *tp;
1977 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1979 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1980 sf_advmss = READ_ONCE(tp->advmss);
1982 rtt_us = max(sf_rtt_us, rtt_us);
1983 advmss = max(sf_advmss, advmss);
1984 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1987 msk->rcvq_space.rtt_us = rtt_us;
1988 msk->scaling_ratio = scaling_ratio;
1989 if (time < (rtt_us >> 3) || rtt_us == 0)
1992 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1995 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1996 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2000 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2002 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2004 do_div(grow, msk->rcvq_space.space);
2005 rcvwin += (grow << 1);
2007 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2008 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2010 if (rcvbuf > sk->sk_rcvbuf) {
2013 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2014 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2016 /* Make subflows follow along. If we do not do this, we
2017 * get drops at subflow level if skbs can't be moved to
2018 * the mptcp rx queue fast enough (announced rcv_win can
2019 * exceed ssk->sk_rcvbuf).
2021 mptcp_for_each_subflow(msk, subflow) {
2025 ssk = mptcp_subflow_tcp_sock(subflow);
2026 slow = lock_sock_fast(ssk);
2027 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2028 tcp_sk(ssk)->window_clamp = window_clamp;
2029 tcp_cleanup_rbuf(ssk, 1);
2030 unlock_sock_fast(ssk, slow);
2035 msk->rcvq_space.space = msk->rcvq_space.copied;
2037 msk->rcvq_space.copied = 0;
2038 msk->rcvq_space.time = mstamp;
2041 static void __mptcp_update_rmem(struct sock *sk)
2043 struct mptcp_sock *msk = mptcp_sk(sk);
2045 if (!msk->rmem_released)
2048 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2049 mptcp_rmem_uncharge(sk, msk->rmem_released);
2050 WRITE_ONCE(msk->rmem_released, 0);
2053 static void __mptcp_splice_receive_queue(struct sock *sk)
2055 struct mptcp_sock *msk = mptcp_sk(sk);
2057 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2060 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2062 struct sock *sk = (struct sock *)msk;
2063 unsigned int moved = 0;
2067 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2070 /* we can have data pending in the subflows only if the msk
2071 * receive buffer was full at subflow_data_ready() time,
2072 * that is an unlikely slow path.
2077 slowpath = lock_sock_fast(ssk);
2078 mptcp_data_lock(sk);
2079 __mptcp_update_rmem(sk);
2080 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2081 mptcp_data_unlock(sk);
2083 if (unlikely(ssk->sk_err))
2084 __mptcp_error_report(sk);
2085 unlock_sock_fast(ssk, slowpath);
2088 /* acquire the data lock only if some input data is pending */
2090 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2091 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2092 mptcp_data_lock(sk);
2093 __mptcp_update_rmem(sk);
2094 ret |= __mptcp_ofo_queue(msk);
2095 __mptcp_splice_receive_queue(sk);
2096 mptcp_data_unlock(sk);
2099 mptcp_check_data_fin((struct sock *)msk);
2100 return !skb_queue_empty(&msk->receive_queue);
2103 static unsigned int mptcp_inq_hint(const struct sock *sk)
2105 const struct mptcp_sock *msk = mptcp_sk(sk);
2106 const struct sk_buff *skb;
2108 skb = skb_peek(&msk->receive_queue);
2110 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2112 if (hint_val >= INT_MAX)
2115 return (unsigned int)hint_val;
2118 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2124 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2125 int flags, int *addr_len)
2127 struct mptcp_sock *msk = mptcp_sk(sk);
2128 struct scm_timestamping_internal tss;
2129 int copied = 0, cmsg_flags = 0;
2133 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2134 if (unlikely(flags & MSG_ERRQUEUE))
2135 return inet_recv_error(sk, msg, len, addr_len);
2138 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2143 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2145 len = min_t(size_t, len, INT_MAX);
2146 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2148 if (unlikely(msk->recvmsg_inq))
2149 cmsg_flags = MPTCP_CMSG_INQ;
2151 while (copied < len) {
2154 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2155 if (unlikely(bytes_read < 0)) {
2157 copied = bytes_read;
2161 copied += bytes_read;
2163 /* be sure to advertise window change */
2164 mptcp_cleanup_rbuf(msk);
2166 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2169 /* only the master socket status is relevant here. The exit
2170 * conditions mirror closely tcp_recvmsg()
2172 if (copied >= target)
2177 sk->sk_state == TCP_CLOSE ||
2178 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2180 signal_pending(current))
2184 copied = sock_error(sk);
2188 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2189 /* race breaker: the shutdown could be after the
2190 * previous receive queue check
2192 if (__mptcp_move_skbs(msk))
2197 if (sk->sk_state == TCP_CLOSE) {
2207 if (signal_pending(current)) {
2208 copied = sock_intr_errno(timeo);
2213 pr_debug("block timeout %ld", timeo);
2214 sk_wait_data(sk, &timeo, NULL);
2218 if (cmsg_flags && copied >= 0) {
2219 if (cmsg_flags & MPTCP_CMSG_TS)
2220 tcp_recv_timestamp(msg, sk, &tss);
2222 if (cmsg_flags & MPTCP_CMSG_INQ) {
2223 unsigned int inq = mptcp_inq_hint(sk);
2225 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2229 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2230 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2231 skb_queue_empty(&msk->receive_queue), copied);
2232 if (!(flags & MSG_PEEK))
2233 mptcp_rcv_space_adjust(msk, copied);
2239 static void mptcp_retransmit_timer(struct timer_list *t)
2241 struct inet_connection_sock *icsk = from_timer(icsk, t,
2242 icsk_retransmit_timer);
2243 struct sock *sk = &icsk->icsk_inet.sk;
2244 struct mptcp_sock *msk = mptcp_sk(sk);
2247 if (!sock_owned_by_user(sk)) {
2248 /* we need a process context to retransmit */
2249 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2250 mptcp_schedule_work(sk);
2252 /* delegate our work to tcp_release_cb() */
2253 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2259 static void mptcp_timeout_timer(struct timer_list *t)
2261 struct sock *sk = from_timer(sk, t, sk_timer);
2263 mptcp_schedule_work(sk);
2267 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2270 * A backup subflow is returned only if that is the only kind available.
2272 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2274 struct sock *backup = NULL, *pick = NULL;
2275 struct mptcp_subflow_context *subflow;
2276 int min_stale_count = INT_MAX;
2278 mptcp_for_each_subflow(msk, subflow) {
2279 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2281 if (!__mptcp_subflow_active(subflow))
2284 /* still data outstanding at TCP level? skip this */
2285 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2286 mptcp_pm_subflow_chk_stale(msk, ssk);
2287 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2291 if (subflow->backup) {
2304 /* use backup only if there are no progresses anywhere */
2305 return min_stale_count > 1 ? backup : NULL;
2308 bool __mptcp_retransmit_pending_data(struct sock *sk)
2310 struct mptcp_data_frag *cur, *rtx_head;
2311 struct mptcp_sock *msk = mptcp_sk(sk);
2313 if (__mptcp_check_fallback(msk))
2316 if (tcp_rtx_and_write_queues_empty(sk))
2319 /* the closing socket has some data untransmitted and/or unacked:
2320 * some data in the mptcp rtx queue has not really xmitted yet.
2321 * keep it simple and re-inject the whole mptcp level rtx queue
2323 mptcp_data_lock(sk);
2324 __mptcp_clean_una_wakeup(sk);
2325 rtx_head = mptcp_rtx_head(sk);
2327 mptcp_data_unlock(sk);
2331 msk->recovery_snd_nxt = msk->snd_nxt;
2332 msk->recovery = true;
2333 mptcp_data_unlock(sk);
2335 msk->first_pending = rtx_head;
2338 /* be sure to clear the "sent status" on all re-injected fragments */
2339 list_for_each_entry(cur, &msk->rtx_queue, list) {
2340 if (!cur->already_sent)
2342 cur->already_sent = 0;
2348 /* flags for __mptcp_close_ssk() */
2349 #define MPTCP_CF_PUSH BIT(1)
2350 #define MPTCP_CF_FASTCLOSE BIT(2)
2352 /* subflow sockets can be either outgoing (connect) or incoming
2355 * Outgoing subflows use in-kernel sockets.
2356 * Incoming subflows do not have their own 'struct socket' allocated,
2357 * so we need to use tcp_close() after detaching them from the mptcp
2360 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2361 struct mptcp_subflow_context *subflow,
2364 struct mptcp_sock *msk = mptcp_sk(sk);
2365 bool dispose_it, need_push = false;
2367 /* If the first subflow moved to a close state before accept, e.g. due
2368 * to an incoming reset, mptcp either:
2369 * - if either the subflow or the msk are dead, destroy the context
2370 * (the subflow socket is deleted by inet_child_forget) and the msk
2371 * - otherwise do nothing at the moment and take action at accept and/or
2372 * listener shutdown - user-space must be able to accept() the closed
2375 if (msk->in_accept_queue && msk->first == ssk) {
2376 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2379 /* ensure later check in mptcp_worker() will dispose the msk */
2380 sock_set_flag(sk, SOCK_DEAD);
2381 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2382 mptcp_subflow_drop_ctx(ssk);
2386 dispose_it = msk->free_first || ssk != msk->first;
2388 list_del(&subflow->node);
2390 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2392 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2393 /* be sure to force the tcp_disconnect() path,
2394 * to generate the egress reset
2396 ssk->sk_lingertime = 0;
2397 sock_set_flag(ssk, SOCK_LINGER);
2398 subflow->send_fastclose = 1;
2401 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2403 /* The MPTCP code never wait on the subflow sockets, TCP-level
2404 * disconnect should never fail
2406 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2407 mptcp_subflow_ctx_reset(subflow);
2413 subflow->disposable = 1;
2415 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2416 * the ssk has been already destroyed, we just need to release the
2417 * reference owned by msk;
2419 if (!inet_csk(ssk)->icsk_ulp_ops) {
2420 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2421 kfree_rcu(subflow, rcu);
2423 /* otherwise tcp will dispose of the ssk and subflow ctx */
2424 __tcp_close(ssk, 0);
2426 /* close acquired an extra ref */
2435 if (ssk == msk->first)
2436 WRITE_ONCE(msk->first, NULL);
2440 __mptcp_push_pending(sk, 0);
2443 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2444 struct mptcp_subflow_context *subflow)
2446 if (sk->sk_state == TCP_ESTABLISHED)
2447 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2449 /* subflow aborted before reaching the fully_established status
2450 * attempt the creation of the next subflow
2452 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2454 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2457 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2462 static void __mptcp_close_subflow(struct sock *sk)
2464 struct mptcp_subflow_context *subflow, *tmp;
2465 struct mptcp_sock *msk = mptcp_sk(sk);
2469 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2470 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2472 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2475 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2476 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2479 mptcp_close_ssk(sk, ssk, subflow);
2484 static bool mptcp_should_close(const struct sock *sk)
2486 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2487 struct mptcp_subflow_context *subflow;
2489 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2492 /* if all subflows are in closed status don't bother with additional
2495 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2496 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2503 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2505 struct mptcp_subflow_context *subflow, *tmp;
2506 struct sock *sk = (struct sock *)msk;
2508 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2511 mptcp_token_destroy(msk);
2513 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2514 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2517 slow = lock_sock_fast(tcp_sk);
2518 if (tcp_sk->sk_state != TCP_CLOSE) {
2519 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2520 tcp_set_state(tcp_sk, TCP_CLOSE);
2522 unlock_sock_fast(tcp_sk, slow);
2525 /* Mirror the tcp_reset() error propagation */
2526 switch (sk->sk_state) {
2528 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2530 case TCP_CLOSE_WAIT:
2531 WRITE_ONCE(sk->sk_err, EPIPE);
2536 WRITE_ONCE(sk->sk_err, ECONNRESET);
2539 inet_sk_state_store(sk, TCP_CLOSE);
2540 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2541 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2542 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2544 /* the calling mptcp_worker will properly destroy the socket */
2545 if (sock_flag(sk, SOCK_DEAD))
2548 sk->sk_state_change(sk);
2549 sk_error_report(sk);
2552 static void __mptcp_retrans(struct sock *sk)
2554 struct mptcp_sock *msk = mptcp_sk(sk);
2555 struct mptcp_subflow_context *subflow;
2556 struct mptcp_sendmsg_info info = {};
2557 struct mptcp_data_frag *dfrag;
2562 mptcp_clean_una_wakeup(sk);
2564 /* first check ssk: need to kick "stale" logic */
2565 err = mptcp_sched_get_retrans(msk);
2566 dfrag = mptcp_rtx_head(sk);
2568 if (mptcp_data_fin_enabled(msk)) {
2569 struct inet_connection_sock *icsk = inet_csk(sk);
2571 icsk->icsk_retransmits++;
2572 mptcp_set_datafin_timeout(sk);
2573 mptcp_send_ack(msk);
2578 if (!mptcp_send_head(sk))
2587 mptcp_for_each_subflow(msk, subflow) {
2588 if (READ_ONCE(subflow->scheduled)) {
2591 mptcp_subflow_set_scheduled(subflow, false);
2593 ssk = mptcp_subflow_tcp_sock(subflow);
2597 /* limit retransmission to the bytes already sent on some subflows */
2599 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2600 dfrag->already_sent;
2601 while (info.sent < info.limit) {
2602 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2606 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2611 len = max(copied, len);
2612 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2614 WRITE_ONCE(msk->allow_infinite_fallback, false);
2621 msk->bytes_retrans += len;
2622 dfrag->already_sent = max(dfrag->already_sent, len);
2625 mptcp_check_and_set_pending(sk);
2627 if (!mptcp_timer_pending(sk))
2628 mptcp_reset_timer(sk);
2631 /* schedule the timeout timer for the relevant event: either close timeout
2632 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2634 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2636 struct sock *sk = (struct sock *)msk;
2637 unsigned long timeout, close_timeout;
2639 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2642 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2644 /* the close timeout takes precedence on the fail one, and here at least one of
2647 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2649 sk_reset_timer(sk, &sk->sk_timer, timeout);
2652 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2654 struct sock *ssk = msk->first;
2660 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2662 slow = lock_sock_fast(ssk);
2663 mptcp_subflow_reset(ssk);
2664 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2665 unlock_sock_fast(ssk, slow);
2667 mptcp_reset_timeout(msk, 0);
2670 static void mptcp_do_fastclose(struct sock *sk)
2672 struct mptcp_subflow_context *subflow, *tmp;
2673 struct mptcp_sock *msk = mptcp_sk(sk);
2675 inet_sk_state_store(sk, TCP_CLOSE);
2676 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2677 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2678 subflow, MPTCP_CF_FASTCLOSE);
2681 static void mptcp_worker(struct work_struct *work)
2683 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2684 struct sock *sk = (struct sock *)msk;
2685 unsigned long fail_tout;
2689 state = sk->sk_state;
2690 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2693 mptcp_check_fastclose(msk);
2695 mptcp_pm_nl_work(msk);
2697 mptcp_check_send_data_fin(sk);
2698 mptcp_check_data_fin_ack(sk);
2699 mptcp_check_data_fin(sk);
2701 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2702 __mptcp_close_subflow(sk);
2704 /* There is no point in keeping around an orphaned sk timedout or
2705 * closed, but we need the msk around to reply to incoming DATA_FIN,
2706 * even if it is orphaned and in FIN_WAIT2 state
2708 if (sock_flag(sk, SOCK_DEAD)) {
2709 if (mptcp_should_close(sk))
2710 mptcp_do_fastclose(sk);
2712 if (sk->sk_state == TCP_CLOSE) {
2713 __mptcp_destroy_sock(sk);
2718 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2719 __mptcp_retrans(sk);
2721 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2722 if (fail_tout && time_after(jiffies, fail_tout))
2723 mptcp_mp_fail_no_response(msk);
2730 static void __mptcp_init_sock(struct sock *sk)
2732 struct mptcp_sock *msk = mptcp_sk(sk);
2734 INIT_LIST_HEAD(&msk->conn_list);
2735 INIT_LIST_HEAD(&msk->join_list);
2736 INIT_LIST_HEAD(&msk->rtx_queue);
2737 INIT_WORK(&msk->work, mptcp_worker);
2738 __skb_queue_head_init(&msk->receive_queue);
2739 msk->out_of_order_queue = RB_ROOT;
2740 msk->first_pending = NULL;
2741 msk->rmem_fwd_alloc = 0;
2742 WRITE_ONCE(msk->rmem_released, 0);
2743 msk->timer_ival = TCP_RTO_MIN;
2745 WRITE_ONCE(msk->first, NULL);
2746 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2747 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2748 WRITE_ONCE(msk->allow_infinite_fallback, true);
2749 msk->recovery = false;
2750 msk->subflow_id = 1;
2752 mptcp_pm_data_init(msk);
2754 /* re-use the csk retrans timer for MPTCP-level retrans */
2755 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2756 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2759 static void mptcp_ca_reset(struct sock *sk)
2761 struct inet_connection_sock *icsk = inet_csk(sk);
2763 tcp_assign_congestion_control(sk);
2764 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2766 /* no need to keep a reference to the ops, the name will suffice */
2767 tcp_cleanup_congestion_control(sk);
2768 icsk->icsk_ca_ops = NULL;
2771 static int mptcp_init_sock(struct sock *sk)
2773 struct net *net = sock_net(sk);
2776 __mptcp_init_sock(sk);
2778 if (!mptcp_is_enabled(net))
2779 return -ENOPROTOOPT;
2781 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2784 ret = mptcp_init_sched(mptcp_sk(sk),
2785 mptcp_sched_find(mptcp_get_scheduler(net)));
2789 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2791 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2792 * propagate the correct value
2796 sk_sockets_allocated_inc(sk);
2797 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2798 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2803 static void __mptcp_clear_xmit(struct sock *sk)
2805 struct mptcp_sock *msk = mptcp_sk(sk);
2806 struct mptcp_data_frag *dtmp, *dfrag;
2808 WRITE_ONCE(msk->first_pending, NULL);
2809 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2810 dfrag_clear(sk, dfrag);
2813 void mptcp_cancel_work(struct sock *sk)
2815 struct mptcp_sock *msk = mptcp_sk(sk);
2817 if (cancel_work_sync(&msk->work))
2821 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2825 switch (ssk->sk_state) {
2827 if (!(how & RCV_SHUTDOWN))
2831 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2834 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2835 pr_debug("Fallback");
2836 ssk->sk_shutdown |= how;
2837 tcp_shutdown(ssk, how);
2839 /* simulate the data_fin ack reception to let the state
2840 * machine move forward
2842 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2843 mptcp_schedule_work(sk);
2845 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2847 if (!mptcp_timer_pending(sk))
2848 mptcp_reset_timer(sk);
2856 static const unsigned char new_state[16] = {
2857 /* current state: new state: action: */
2858 [0 /* (Invalid) */] = TCP_CLOSE,
2859 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2860 [TCP_SYN_SENT] = TCP_CLOSE,
2861 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2862 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2863 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2864 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2865 [TCP_CLOSE] = TCP_CLOSE,
2866 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2867 [TCP_LAST_ACK] = TCP_LAST_ACK,
2868 [TCP_LISTEN] = TCP_CLOSE,
2869 [TCP_CLOSING] = TCP_CLOSING,
2870 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2873 static int mptcp_close_state(struct sock *sk)
2875 int next = (int)new_state[sk->sk_state];
2876 int ns = next & TCP_STATE_MASK;
2878 inet_sk_state_store(sk, ns);
2880 return next & TCP_ACTION_FIN;
2883 static void mptcp_check_send_data_fin(struct sock *sk)
2885 struct mptcp_subflow_context *subflow;
2886 struct mptcp_sock *msk = mptcp_sk(sk);
2888 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2889 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2890 msk->snd_nxt, msk->write_seq);
2892 /* we still need to enqueue subflows or not really shutting down,
2895 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2896 mptcp_send_head(sk))
2899 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2901 mptcp_for_each_subflow(msk, subflow) {
2902 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2904 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2908 static void __mptcp_wr_shutdown(struct sock *sk)
2910 struct mptcp_sock *msk = mptcp_sk(sk);
2912 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2913 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2914 !!mptcp_send_head(sk));
2916 /* will be ignored by fallback sockets */
2917 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2918 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2920 mptcp_check_send_data_fin(sk);
2923 static void __mptcp_destroy_sock(struct sock *sk)
2925 struct mptcp_sock *msk = mptcp_sk(sk);
2927 pr_debug("msk=%p", msk);
2931 mptcp_stop_timer(sk);
2932 sk_stop_timer(sk, &sk->sk_timer);
2934 mptcp_release_sched(msk);
2936 sk->sk_prot->destroy(sk);
2938 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2939 WARN_ON_ONCE(msk->rmem_released);
2940 sk_stream_kill_queues(sk);
2941 xfrm_sk_free_policy(sk);
2946 void __mptcp_unaccepted_force_close(struct sock *sk)
2948 sock_set_flag(sk, SOCK_DEAD);
2949 mptcp_do_fastclose(sk);
2950 __mptcp_destroy_sock(sk);
2953 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2955 /* Concurrent splices from sk_receive_queue into receive_queue will
2956 * always show at least one non-empty queue when checked in this order.
2958 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2959 skb_queue_empty_lockless(&msk->receive_queue))
2962 return EPOLLIN | EPOLLRDNORM;
2965 static void mptcp_check_listen_stop(struct sock *sk)
2969 if (inet_sk_state_load(sk) != TCP_LISTEN)
2972 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2973 ssk = mptcp_sk(sk)->first;
2974 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2977 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2978 tcp_set_state(ssk, TCP_CLOSE);
2979 mptcp_subflow_queue_clean(sk, ssk);
2980 inet_csk_listen_stop(ssk);
2981 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2985 bool __mptcp_close(struct sock *sk, long timeout)
2987 struct mptcp_subflow_context *subflow;
2988 struct mptcp_sock *msk = mptcp_sk(sk);
2989 bool do_cancel_work = false;
2990 int subflows_alive = 0;
2992 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2994 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2995 mptcp_check_listen_stop(sk);
2996 inet_sk_state_store(sk, TCP_CLOSE);
3000 if (mptcp_check_readable(msk) || timeout < 0) {
3001 /* If the msk has read data, or the caller explicitly ask it,
3002 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3004 mptcp_do_fastclose(sk);
3006 } else if (mptcp_close_state(sk)) {
3007 __mptcp_wr_shutdown(sk);
3010 sk_stream_wait_close(sk, timeout);
3013 /* orphan all the subflows */
3014 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
3015 mptcp_for_each_subflow(msk, subflow) {
3016 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3017 bool slow = lock_sock_fast_nested(ssk);
3019 subflows_alive += ssk->sk_state != TCP_CLOSE;
3021 /* since the close timeout takes precedence on the fail one,
3024 if (ssk == msk->first)
3025 subflow->fail_tout = 0;
3027 /* detach from the parent socket, but allow data_ready to
3028 * push incoming data into the mptcp stack, to properly ack it
3030 ssk->sk_socket = NULL;
3032 unlock_sock_fast(ssk, slow);
3036 /* all the subflows are closed, only timeout can change the msk
3037 * state, let's not keep resources busy for no reasons
3039 if (subflows_alive == 0)
3040 inet_sk_state_store(sk, TCP_CLOSE);
3043 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3045 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3047 if (sk->sk_state == TCP_CLOSE) {
3048 __mptcp_destroy_sock(sk);
3049 do_cancel_work = true;
3051 mptcp_reset_timeout(msk, 0);
3054 return do_cancel_work;
3057 static void mptcp_close(struct sock *sk, long timeout)
3059 bool do_cancel_work;
3063 do_cancel_work = __mptcp_close(sk, timeout);
3066 mptcp_cancel_work(sk);
3071 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3073 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3074 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3075 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3077 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3078 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3081 msk6->saddr = ssk6->saddr;
3082 msk6->flow_label = ssk6->flow_label;
3086 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3087 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3088 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3089 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3090 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3091 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3094 static int mptcp_disconnect(struct sock *sk, int flags)
3096 struct mptcp_sock *msk = mptcp_sk(sk);
3098 /* Deny disconnect if other threads are blocked in sk_wait_event()
3099 * or inet_wait_for_connect().
3101 if (sk->sk_wait_pending)
3104 /* We are on the fastopen error path. We can't call straight into the
3105 * subflows cleanup code due to lock nesting (we are already under
3106 * msk->firstsocket lock).
3108 if (msk->fastopening)
3111 mptcp_check_listen_stop(sk);
3112 inet_sk_state_store(sk, TCP_CLOSE);
3114 mptcp_stop_timer(sk);
3115 sk_stop_timer(sk, &sk->sk_timer);
3118 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3120 /* msk->subflow is still intact, the following will not free the first
3123 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3124 WRITE_ONCE(msk->flags, 0);
3126 msk->push_pending = 0;
3127 msk->recovery = false;
3128 msk->can_ack = false;
3129 msk->fully_established = false;
3130 msk->rcv_data_fin = false;
3131 msk->snd_data_fin_enable = false;
3132 msk->rcv_fastclose = false;
3133 msk->use_64bit_ack = false;
3134 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3135 mptcp_pm_data_reset(msk);
3137 msk->bytes_acked = 0;
3138 msk->bytes_received = 0;
3139 msk->bytes_sent = 0;
3140 msk->bytes_retrans = 0;
3142 WRITE_ONCE(sk->sk_shutdown, 0);
3143 sk_error_report(sk);
3147 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3148 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3150 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3152 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3156 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3157 const struct mptcp_options_received *mp_opt,
3159 struct request_sock *req)
3161 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3162 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3163 struct mptcp_sock *msk;
3168 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3169 if (nsk->sk_family == AF_INET6)
3170 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3173 nsk->sk_wait_pending = 0;
3174 __mptcp_init_sock(nsk);
3176 msk = mptcp_sk(nsk);
3177 msk->local_key = subflow_req->local_key;
3178 msk->token = subflow_req->token;
3179 msk->in_accept_queue = 1;
3180 WRITE_ONCE(msk->fully_established, false);
3181 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3182 WRITE_ONCE(msk->csum_enabled, true);
3184 msk->write_seq = subflow_req->idsn + 1;
3185 msk->snd_nxt = msk->write_seq;
3186 msk->snd_una = msk->write_seq;
3187 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3188 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3189 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3191 /* passive msk is created after the first/MPC subflow */
3192 msk->subflow_id = 2;
3194 sock_reset_flag(nsk, SOCK_RCU_FREE);
3195 security_inet_csk_clone(nsk, req);
3197 /* this can't race with mptcp_close(), as the msk is
3198 * not yet exposted to user-space
3200 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3202 /* The msk maintain a ref to each subflow in the connections list */
3203 WRITE_ONCE(msk->first, ssk);
3204 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3207 /* new mpc subflow takes ownership of the newly
3208 * created mptcp socket
3210 mptcp_token_accept(subflow_req, msk);
3212 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3213 * uses the correct data
3215 mptcp_copy_inaddrs(nsk, ssk);
3216 mptcp_propagate_sndbuf(nsk, ssk);
3218 mptcp_rcv_space_init(msk, ssk);
3219 bh_unlock_sock(nsk);
3221 /* note: the newly allocated socket refcount is 2 now */
3225 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3227 const struct tcp_sock *tp = tcp_sk(ssk);
3229 msk->rcvq_space.copied = 0;
3230 msk->rcvq_space.rtt_us = 0;
3232 msk->rcvq_space.time = tp->tcp_mstamp;
3234 /* initial rcv_space offering made to peer */
3235 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3236 TCP_INIT_CWND * tp->advmss);
3237 if (msk->rcvq_space.space == 0)
3238 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3240 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3243 static struct sock *mptcp_accept(struct sock *ssk, int flags, int *err,
3248 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3249 newsk = inet_csk_accept(ssk, flags, err, kern);
3253 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3254 if (sk_is_mptcp(newsk)) {
3255 struct mptcp_subflow_context *subflow;
3256 struct sock *new_mptcp_sock;
3258 subflow = mptcp_subflow_ctx(newsk);
3259 new_mptcp_sock = subflow->conn;
3261 /* is_mptcp should be false if subflow->conn is missing, see
3262 * subflow_syn_recv_sock()
3264 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3265 tcp_sk(newsk)->is_mptcp = 0;
3269 newsk = new_mptcp_sock;
3270 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3272 MPTCP_INC_STATS(sock_net(ssk),
3273 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3277 newsk->sk_kern_sock = kern;
3281 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3283 struct mptcp_subflow_context *subflow, *tmp;
3284 struct sock *sk = (struct sock *)msk;
3286 __mptcp_clear_xmit(sk);
3288 /* join list will be eventually flushed (with rst) at sock lock release time */
3289 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3290 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3292 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3293 mptcp_data_lock(sk);
3294 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3295 __skb_queue_purge(&sk->sk_receive_queue);
3296 skb_rbtree_purge(&msk->out_of_order_queue);
3297 mptcp_data_unlock(sk);
3299 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3300 * inet_sock_destruct() will dispose it
3302 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3303 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3304 mptcp_token_destroy(msk);
3305 mptcp_pm_free_anno_list(msk);
3306 mptcp_free_local_addr_list(msk);
3309 static void mptcp_destroy(struct sock *sk)
3311 struct mptcp_sock *msk = mptcp_sk(sk);
3313 /* allow the following to close even the initial subflow */
3314 msk->free_first = 1;
3315 mptcp_destroy_common(msk, 0);
3316 sk_sockets_allocated_dec(sk);
3319 void __mptcp_data_acked(struct sock *sk)
3321 if (!sock_owned_by_user(sk))
3322 __mptcp_clean_una(sk);
3324 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3326 if (mptcp_pending_data_fin_ack(sk))
3327 mptcp_schedule_work(sk);
3330 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3332 if (!mptcp_send_head(sk))
3335 if (!sock_owned_by_user(sk))
3336 __mptcp_subflow_push_pending(sk, ssk, false);
3338 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3341 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3342 BIT(MPTCP_RETRANSMIT) | \
3343 BIT(MPTCP_FLUSH_JOIN_LIST))
3345 /* processes deferred events and flush wmem */
3346 static void mptcp_release_cb(struct sock *sk)
3347 __must_hold(&sk->sk_lock.slock)
3349 struct mptcp_sock *msk = mptcp_sk(sk);
3352 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3354 struct list_head join_list;
3359 INIT_LIST_HEAD(&join_list);
3360 list_splice_init(&msk->join_list, &join_list);
3362 /* the following actions acquire the subflow socket lock
3364 * 1) can't be invoked in atomic scope
3365 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3366 * datapath acquires the msk socket spinlock while helding
3367 * the subflow socket lock
3369 msk->push_pending = 0;
3370 msk->cb_flags &= ~flags;
3371 spin_unlock_bh(&sk->sk_lock.slock);
3373 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3374 __mptcp_flush_join_list(sk, &join_list);
3375 if (flags & BIT(MPTCP_PUSH_PENDING))
3376 __mptcp_push_pending(sk, 0);
3377 if (flags & BIT(MPTCP_RETRANSMIT))
3378 __mptcp_retrans(sk);
3381 spin_lock_bh(&sk->sk_lock.slock);
3384 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3385 __mptcp_clean_una_wakeup(sk);
3386 if (unlikely(msk->cb_flags)) {
3387 /* be sure to set the current sk state before tacking actions
3388 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3390 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3391 __mptcp_set_connected(sk);
3392 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3393 __mptcp_error_report(sk);
3396 __mptcp_update_rmem(sk);
3399 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3400 * TCP can't schedule delack timer before the subflow is fully established.
3401 * MPTCP uses the delack timer to do 3rd ack retransmissions
3403 static void schedule_3rdack_retransmission(struct sock *ssk)
3405 struct inet_connection_sock *icsk = inet_csk(ssk);
3406 struct tcp_sock *tp = tcp_sk(ssk);
3407 unsigned long timeout;
3409 if (mptcp_subflow_ctx(ssk)->fully_established)
3412 /* reschedule with a timeout above RTT, as we must look only for drop */
3414 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3416 timeout = TCP_TIMEOUT_INIT;
3419 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3420 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3421 icsk->icsk_ack.timeout = timeout;
3422 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3425 void mptcp_subflow_process_delegated(struct sock *ssk)
3427 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3428 struct sock *sk = subflow->conn;
3430 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3431 mptcp_data_lock(sk);
3432 if (!sock_owned_by_user(sk))
3433 __mptcp_subflow_push_pending(sk, ssk, true);
3435 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3436 mptcp_data_unlock(sk);
3437 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3439 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3440 schedule_3rdack_retransmission(ssk);
3441 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3445 static int mptcp_hash(struct sock *sk)
3447 /* should never be called,
3448 * we hash the TCP subflows not the master socket
3454 static void mptcp_unhash(struct sock *sk)
3456 /* called from sk_common_release(), but nothing to do here */
3459 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3461 struct mptcp_sock *msk = mptcp_sk(sk);
3463 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3464 if (WARN_ON_ONCE(!msk->first))
3467 return inet_csk_get_port(msk->first, snum);
3470 void mptcp_finish_connect(struct sock *ssk)
3472 struct mptcp_subflow_context *subflow;
3473 struct mptcp_sock *msk;
3476 subflow = mptcp_subflow_ctx(ssk);
3480 pr_debug("msk=%p, token=%u", sk, subflow->token);
3482 subflow->map_seq = subflow->iasn;
3483 subflow->map_subflow_seq = 1;
3485 /* the socket is not connected yet, no msk/subflow ops can access/race
3486 * accessing the field below
3488 WRITE_ONCE(msk->local_key, subflow->local_key);
3489 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3490 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3491 WRITE_ONCE(msk->snd_una, msk->write_seq);
3493 mptcp_pm_new_connection(msk, ssk, 0);
3495 mptcp_rcv_space_init(msk, ssk);
3498 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3500 write_lock_bh(&sk->sk_callback_lock);
3501 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3502 sk_set_socket(sk, parent);
3503 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3504 write_unlock_bh(&sk->sk_callback_lock);
3507 bool mptcp_finish_join(struct sock *ssk)
3509 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3510 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3511 struct sock *parent = (void *)msk;
3514 pr_debug("msk=%p, subflow=%p", msk, subflow);
3516 /* mptcp socket already closing? */
3517 if (!mptcp_is_fully_established(parent)) {
3518 subflow->reset_reason = MPTCP_RST_EMPTCP;
3522 /* active subflow, already present inside the conn_list */
3523 if (!list_empty(&subflow->node)) {
3524 mptcp_subflow_joined(msk, ssk);
3528 if (!mptcp_pm_allow_new_subflow(msk))
3529 goto err_prohibited;
3531 /* If we can't acquire msk socket lock here, let the release callback
3534 mptcp_data_lock(parent);
3535 if (!sock_owned_by_user(parent)) {
3536 ret = __mptcp_finish_join(msk, ssk);
3539 list_add_tail(&subflow->node, &msk->conn_list);
3543 list_add_tail(&subflow->node, &msk->join_list);
3544 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3546 mptcp_data_unlock(parent);
3550 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3557 static void mptcp_shutdown(struct sock *sk, int how)
3559 pr_debug("sk=%p, how=%d", sk, how);
3561 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3562 __mptcp_wr_shutdown(sk);
3565 static int mptcp_forward_alloc_get(const struct sock *sk)
3567 return READ_ONCE(sk->sk_forward_alloc) +
3568 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3571 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3573 const struct sock *sk = (void *)msk;
3576 if (sk->sk_state == TCP_LISTEN)
3579 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3582 delta = msk->write_seq - v;
3583 if (__mptcp_check_fallback(msk) && msk->first) {
3584 struct tcp_sock *tp = tcp_sk(msk->first);
3586 /* the first subflow is disconnected after close - see
3587 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3588 * so ignore that status, too.
3590 if (!((1 << msk->first->sk_state) &
3591 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3592 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3594 if (delta > INT_MAX)
3600 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3602 struct mptcp_sock *msk = mptcp_sk(sk);
3607 if (sk->sk_state == TCP_LISTEN)
3611 __mptcp_move_skbs(msk);
3612 *karg = mptcp_inq_hint(sk);
3616 slow = lock_sock_fast(sk);
3617 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3618 unlock_sock_fast(sk, slow);
3621 slow = lock_sock_fast(sk);
3622 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3623 unlock_sock_fast(sk, slow);
3626 return -ENOIOCTLCMD;
3632 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3633 struct mptcp_subflow_context *subflow)
3635 subflow->request_mptcp = 0;
3636 __mptcp_do_fallback(msk);
3639 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3641 struct mptcp_subflow_context *subflow;
3642 struct mptcp_sock *msk = mptcp_sk(sk);
3646 ssk = __mptcp_nmpc_sk(msk);
3648 return PTR_ERR(ssk);
3650 inet_sk_state_store(sk, TCP_SYN_SENT);
3651 subflow = mptcp_subflow_ctx(ssk);
3652 #ifdef CONFIG_TCP_MD5SIG
3653 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3656 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3657 mptcp_subflow_early_fallback(msk, subflow);
3659 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3660 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3661 mptcp_subflow_early_fallback(msk, subflow);
3663 if (likely(!__mptcp_check_fallback(msk)))
3664 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3666 /* if reaching here via the fastopen/sendmsg path, the caller already
3667 * acquired the subflow socket lock, too.
3669 if (!msk->fastopening)
3672 /* the following mirrors closely a very small chunk of code from
3673 * __inet_stream_connect()
3675 if (ssk->sk_state != TCP_CLOSE)
3678 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3679 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3684 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3688 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3691 if (!msk->fastopening)
3694 /* on successful connect, the msk state will be moved to established by
3695 * subflow_finish_connect()
3697 if (unlikely(err)) {
3698 /* avoid leaving a dangling token in an unconnected socket */
3699 mptcp_token_destroy(msk);
3700 inet_sk_state_store(sk, TCP_CLOSE);
3704 mptcp_copy_inaddrs(sk, ssk);
3708 static struct proto mptcp_prot = {
3710 .owner = THIS_MODULE,
3711 .init = mptcp_init_sock,
3712 .connect = mptcp_connect,
3713 .disconnect = mptcp_disconnect,
3714 .close = mptcp_close,
3715 .accept = mptcp_accept,
3716 .setsockopt = mptcp_setsockopt,
3717 .getsockopt = mptcp_getsockopt,
3718 .shutdown = mptcp_shutdown,
3719 .destroy = mptcp_destroy,
3720 .sendmsg = mptcp_sendmsg,
3721 .ioctl = mptcp_ioctl,
3722 .recvmsg = mptcp_recvmsg,
3723 .release_cb = mptcp_release_cb,
3725 .unhash = mptcp_unhash,
3726 .get_port = mptcp_get_port,
3727 .forward_alloc_get = mptcp_forward_alloc_get,
3728 .sockets_allocated = &mptcp_sockets_allocated,
3730 .memory_allocated = &tcp_memory_allocated,
3731 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3733 .memory_pressure = &tcp_memory_pressure,
3734 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3735 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3736 .sysctl_mem = sysctl_tcp_mem,
3737 .obj_size = sizeof(struct mptcp_sock),
3738 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3739 .no_autobind = true,
3742 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3744 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3745 struct sock *ssk, *sk = sock->sk;
3749 ssk = __mptcp_nmpc_sk(msk);
3755 if (sk->sk_family == AF_INET)
3756 err = inet_bind_sk(ssk, uaddr, addr_len);
3757 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3758 else if (sk->sk_family == AF_INET6)
3759 err = inet6_bind_sk(ssk, uaddr, addr_len);
3762 mptcp_copy_inaddrs(sk, ssk);
3769 static int mptcp_listen(struct socket *sock, int backlog)
3771 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3772 struct sock *sk = sock->sk;
3776 pr_debug("msk=%p", msk);
3781 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3784 ssk = __mptcp_nmpc_sk(msk);
3790 inet_sk_state_store(sk, TCP_LISTEN);
3791 sock_set_flag(sk, SOCK_RCU_FREE);
3794 err = __inet_listen_sk(ssk, backlog);
3796 inet_sk_state_store(sk, inet_sk_state_load(ssk));
3799 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3800 mptcp_copy_inaddrs(sk, ssk);
3801 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3809 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3810 int flags, bool kern)
3812 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3813 struct sock *ssk, *newsk;
3816 pr_debug("msk=%p", msk);
3818 /* Buggy applications can call accept on socket states other then LISTEN
3819 * but no need to allocate the first subflow just to error out.
3821 ssk = READ_ONCE(msk->first);
3825 newsk = mptcp_accept(ssk, flags, &err, kern);
3831 __inet_accept(sock, newsock, newsk);
3832 if (!mptcp_is_tcpsk(newsock->sk)) {
3833 struct mptcp_sock *msk = mptcp_sk(newsk);
3834 struct mptcp_subflow_context *subflow;
3836 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3837 msk->in_accept_queue = 0;
3839 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3840 * This is needed so NOSPACE flag can be set from tcp stack.
3842 mptcp_for_each_subflow(msk, subflow) {
3843 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3845 if (!ssk->sk_socket)
3846 mptcp_sock_graft(ssk, newsock);
3849 /* Do late cleanup for the first subflow as necessary. Also
3850 * deal with bad peers not doing a complete shutdown.
3852 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3853 __mptcp_close_ssk(newsk, msk->first,
3854 mptcp_subflow_ctx(msk->first), 0);
3855 if (unlikely(list_is_singular(&msk->conn_list)))
3856 inet_sk_state_store(newsk, TCP_CLOSE);
3859 release_sock(newsk);
3864 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3866 struct sock *sk = (struct sock *)msk;
3868 if (sk_stream_is_writeable(sk))
3869 return EPOLLOUT | EPOLLWRNORM;
3871 mptcp_set_nospace(sk);
3872 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3873 if (sk_stream_is_writeable(sk))
3874 return EPOLLOUT | EPOLLWRNORM;
3879 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3880 struct poll_table_struct *wait)
3882 struct sock *sk = sock->sk;
3883 struct mptcp_sock *msk;
3889 sock_poll_wait(file, sock, wait);
3891 state = inet_sk_state_load(sk);
3892 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3893 if (state == TCP_LISTEN) {
3894 struct sock *ssk = READ_ONCE(msk->first);
3896 if (WARN_ON_ONCE(!ssk))
3899 return inet_csk_listen_poll(ssk);
3902 shutdown = READ_ONCE(sk->sk_shutdown);
3903 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3905 if (shutdown & RCV_SHUTDOWN)
3906 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3908 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3909 mask |= mptcp_check_readable(msk);
3910 if (shutdown & SEND_SHUTDOWN)
3911 mask |= EPOLLOUT | EPOLLWRNORM;
3913 mask |= mptcp_check_writeable(msk);
3914 } else if (state == TCP_SYN_SENT &&
3915 inet_test_bit(DEFER_CONNECT, sk)) {
3916 /* cf tcp_poll() note about TFO */
3917 mask |= EPOLLOUT | EPOLLWRNORM;
3920 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3922 if (READ_ONCE(sk->sk_err))
3928 static const struct proto_ops mptcp_stream_ops = {
3930 .owner = THIS_MODULE,
3931 .release = inet_release,
3933 .connect = inet_stream_connect,
3934 .socketpair = sock_no_socketpair,
3935 .accept = mptcp_stream_accept,
3936 .getname = inet_getname,
3938 .ioctl = inet_ioctl,
3939 .gettstamp = sock_gettstamp,
3940 .listen = mptcp_listen,
3941 .shutdown = inet_shutdown,
3942 .setsockopt = sock_common_setsockopt,
3943 .getsockopt = sock_common_getsockopt,
3944 .sendmsg = inet_sendmsg,
3945 .recvmsg = inet_recvmsg,
3946 .mmap = sock_no_mmap,
3949 static struct inet_protosw mptcp_protosw = {
3950 .type = SOCK_STREAM,
3951 .protocol = IPPROTO_MPTCP,
3952 .prot = &mptcp_prot,
3953 .ops = &mptcp_stream_ops,
3954 .flags = INET_PROTOSW_ICSK,
3957 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3959 struct mptcp_delegated_action *delegated;
3960 struct mptcp_subflow_context *subflow;
3963 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3964 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3965 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3967 bh_lock_sock_nested(ssk);
3968 if (!sock_owned_by_user(ssk) &&
3969 mptcp_subflow_has_delegated_action(subflow))
3970 mptcp_subflow_process_delegated(ssk);
3971 /* ... elsewhere tcp_release_cb_override already processed
3972 * the action or will do at next release_sock().
3973 * In both case must dequeue the subflow here - on the same
3974 * CPU that scheduled it.
3976 bh_unlock_sock(ssk);
3979 if (++work_done == budget)
3983 /* always provide a 0 'work_done' argument, so that napi_complete_done
3984 * will not try accessing the NULL napi->dev ptr
3986 napi_complete_done(napi, 0);
3990 void __init mptcp_proto_init(void)
3992 struct mptcp_delegated_action *delegated;
3995 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3997 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3998 panic("Failed to allocate MPTCP pcpu counter\n");
4000 init_dummy_netdev(&mptcp_napi_dev);
4001 for_each_possible_cpu(cpu) {
4002 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4003 INIT_LIST_HEAD(&delegated->head);
4004 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4006 napi_enable(&delegated->napi);
4009 mptcp_subflow_init();
4014 if (proto_register(&mptcp_prot, 1) != 0)
4015 panic("Failed to register MPTCP proto.\n");
4017 inet_register_protosw(&mptcp_protosw);
4019 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4022 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4023 static const struct proto_ops mptcp_v6_stream_ops = {
4025 .owner = THIS_MODULE,
4026 .release = inet6_release,
4028 .connect = inet_stream_connect,
4029 .socketpair = sock_no_socketpair,
4030 .accept = mptcp_stream_accept,
4031 .getname = inet6_getname,
4033 .ioctl = inet6_ioctl,
4034 .gettstamp = sock_gettstamp,
4035 .listen = mptcp_listen,
4036 .shutdown = inet_shutdown,
4037 .setsockopt = sock_common_setsockopt,
4038 .getsockopt = sock_common_getsockopt,
4039 .sendmsg = inet6_sendmsg,
4040 .recvmsg = inet6_recvmsg,
4041 .mmap = sock_no_mmap,
4042 #ifdef CONFIG_COMPAT
4043 .compat_ioctl = inet6_compat_ioctl,
4047 static struct proto mptcp_v6_prot;
4049 static struct inet_protosw mptcp_v6_protosw = {
4050 .type = SOCK_STREAM,
4051 .protocol = IPPROTO_MPTCP,
4052 .prot = &mptcp_v6_prot,
4053 .ops = &mptcp_v6_stream_ops,
4054 .flags = INET_PROTOSW_ICSK,
4057 int __init mptcp_proto_v6_init(void)
4061 mptcp_v6_prot = mptcp_prot;
4062 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4063 mptcp_v6_prot.slab = NULL;
4064 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4065 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4067 err = proto_register(&mptcp_v6_prot, 1);
4071 err = inet6_register_protosw(&mptcp_v6_protosw);
4073 proto_unregister(&mptcp_v6_prot);