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_rtx_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 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
775 int err = sock_error(ssk);
781 /* only propagate errors on fallen-back sockets or
784 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
787 /* We need to propagate only transition to CLOSE state.
788 * Orphaned socket will see such state change via
789 * subflow_sched_work_if_closed() and that path will properly
790 * destroy the msk as needed.
792 ssk_state = inet_sk_state_load(ssk);
793 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
794 inet_sk_state_store(sk, ssk_state);
795 WRITE_ONCE(sk->sk_err, -err);
797 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
803 void __mptcp_error_report(struct sock *sk)
805 struct mptcp_subflow_context *subflow;
806 struct mptcp_sock *msk = mptcp_sk(sk);
808 mptcp_for_each_subflow(msk, subflow)
809 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
813 /* In most cases we will be able to lock the mptcp socket. If its already
814 * owned, we need to defer to the work queue to avoid ABBA deadlock.
816 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
818 struct sock *sk = (struct sock *)msk;
819 unsigned int moved = 0;
821 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
822 __mptcp_ofo_queue(msk);
823 if (unlikely(ssk->sk_err)) {
824 if (!sock_owned_by_user(sk))
825 __mptcp_error_report(sk);
827 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
830 /* If the moves have caught up with the DATA_FIN sequence number
831 * it's time to ack the DATA_FIN and change socket state, but
832 * this is not a good place to change state. Let the workqueue
835 if (mptcp_pending_data_fin(sk, NULL))
836 mptcp_schedule_work(sk);
840 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
842 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
843 struct mptcp_sock *msk = mptcp_sk(sk);
844 int sk_rbuf, ssk_rbuf;
846 /* The peer can send data while we are shutting down this
847 * subflow at msk destruction time, but we must avoid enqueuing
848 * more data to the msk receive queue
850 if (unlikely(subflow->disposable))
853 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
854 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
855 if (unlikely(ssk_rbuf > sk_rbuf))
858 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
859 if (__mptcp_rmem(sk) > sk_rbuf) {
860 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
864 /* Wake-up the reader only for in-sequence data */
866 if (move_skbs_to_msk(msk, ssk))
867 sk->sk_data_ready(sk);
869 mptcp_data_unlock(sk);
872 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
874 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
875 WRITE_ONCE(msk->allow_infinite_fallback, false);
876 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
879 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
881 struct sock *sk = (struct sock *)msk;
883 if (sk->sk_state != TCP_ESTABLISHED)
886 /* attach to msk socket only after we are sure we will deal with it
889 if (sk->sk_socket && !ssk->sk_socket)
890 mptcp_sock_graft(ssk, sk->sk_socket);
892 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
893 mptcp_sockopt_sync_locked(msk, ssk);
894 mptcp_subflow_joined(msk, ssk);
895 mptcp_stop_tout_timer(sk);
899 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
901 struct mptcp_subflow_context *tmp, *subflow;
902 struct mptcp_sock *msk = mptcp_sk(sk);
904 list_for_each_entry_safe(subflow, tmp, join_list, node) {
905 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
906 bool slow = lock_sock_fast(ssk);
908 list_move_tail(&subflow->node, &msk->conn_list);
909 if (!__mptcp_finish_join(msk, ssk))
910 mptcp_subflow_reset(ssk);
911 unlock_sock_fast(ssk, slow);
915 static bool mptcp_rtx_timer_pending(struct sock *sk)
917 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
920 static void mptcp_reset_rtx_timer(struct sock *sk)
922 struct inet_connection_sock *icsk = inet_csk(sk);
925 /* prevent rescheduling on close */
926 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
929 tout = mptcp_sk(sk)->timer_ival;
930 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
933 bool mptcp_schedule_work(struct sock *sk)
935 if (inet_sk_state_load(sk) != TCP_CLOSE &&
936 schedule_work(&mptcp_sk(sk)->work)) {
937 /* each subflow already holds a reference to the sk, and the
938 * workqueue is invoked by a subflow, so sk can't go away here.
946 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
948 struct mptcp_subflow_context *subflow;
950 msk_owned_by_me(msk);
952 mptcp_for_each_subflow(msk, subflow) {
953 if (READ_ONCE(subflow->data_avail))
954 return mptcp_subflow_tcp_sock(subflow);
960 static bool mptcp_skb_can_collapse_to(u64 write_seq,
961 const struct sk_buff *skb,
962 const struct mptcp_ext *mpext)
964 if (!tcp_skb_can_collapse_to(skb))
967 /* can collapse only if MPTCP level sequence is in order and this
968 * mapping has not been xmitted yet
970 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
974 /* we can append data to the given data frag if:
975 * - there is space available in the backing page_frag
976 * - the data frag tail matches the current page_frag free offset
977 * - the data frag end sequence number matches the current write seq
979 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
980 const struct page_frag *pfrag,
981 const struct mptcp_data_frag *df)
983 return df && pfrag->page == df->page &&
984 pfrag->size - pfrag->offset > 0 &&
985 pfrag->offset == (df->offset + df->data_len) &&
986 df->data_seq + df->data_len == msk->write_seq;
989 static void dfrag_uncharge(struct sock *sk, int len)
991 sk_mem_uncharge(sk, len);
992 sk_wmem_queued_add(sk, -len);
995 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
997 int len = dfrag->data_len + dfrag->overhead;
999 list_del(&dfrag->list);
1000 dfrag_uncharge(sk, len);
1001 put_page(dfrag->page);
1004 static void __mptcp_clean_una(struct sock *sk)
1006 struct mptcp_sock *msk = mptcp_sk(sk);
1007 struct mptcp_data_frag *dtmp, *dfrag;
1010 snd_una = msk->snd_una;
1011 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1012 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1015 if (unlikely(dfrag == msk->first_pending)) {
1016 /* in recovery mode can see ack after the current snd head */
1017 if (WARN_ON_ONCE(!msk->recovery))
1020 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1023 dfrag_clear(sk, dfrag);
1026 dfrag = mptcp_rtx_head(sk);
1027 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1028 u64 delta = snd_una - dfrag->data_seq;
1030 /* prevent wrap around in recovery mode */
1031 if (unlikely(delta > dfrag->already_sent)) {
1032 if (WARN_ON_ONCE(!msk->recovery))
1034 if (WARN_ON_ONCE(delta > dfrag->data_len))
1036 dfrag->already_sent += delta - dfrag->already_sent;
1039 dfrag->data_seq += delta;
1040 dfrag->offset += delta;
1041 dfrag->data_len -= delta;
1042 dfrag->already_sent -= delta;
1044 dfrag_uncharge(sk, delta);
1047 /* all retransmitted data acked, recovery completed */
1048 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1049 msk->recovery = false;
1052 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1053 snd_una == READ_ONCE(msk->write_seq)) {
1054 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1055 mptcp_stop_rtx_timer(sk);
1057 mptcp_reset_rtx_timer(sk);
1061 static void __mptcp_clean_una_wakeup(struct sock *sk)
1063 lockdep_assert_held_once(&sk->sk_lock.slock);
1065 __mptcp_clean_una(sk);
1066 mptcp_write_space(sk);
1069 static void mptcp_clean_una_wakeup(struct sock *sk)
1071 mptcp_data_lock(sk);
1072 __mptcp_clean_una_wakeup(sk);
1073 mptcp_data_unlock(sk);
1076 static void mptcp_enter_memory_pressure(struct sock *sk)
1078 struct mptcp_subflow_context *subflow;
1079 struct mptcp_sock *msk = mptcp_sk(sk);
1082 sk_stream_moderate_sndbuf(sk);
1083 mptcp_for_each_subflow(msk, subflow) {
1084 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1087 tcp_enter_memory_pressure(ssk);
1088 sk_stream_moderate_sndbuf(ssk);
1093 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1096 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1098 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1099 pfrag, sk->sk_allocation)))
1102 mptcp_enter_memory_pressure(sk);
1106 static struct mptcp_data_frag *
1107 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1110 int offset = ALIGN(orig_offset, sizeof(long));
1111 struct mptcp_data_frag *dfrag;
1113 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1114 dfrag->data_len = 0;
1115 dfrag->data_seq = msk->write_seq;
1116 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1117 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1118 dfrag->already_sent = 0;
1119 dfrag->page = pfrag->page;
1124 struct mptcp_sendmsg_info {
1130 bool data_lock_held;
1133 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1134 u64 data_seq, int avail_size)
1136 u64 window_end = mptcp_wnd_end(msk);
1139 if (__mptcp_check_fallback(msk))
1142 mptcp_snd_wnd = window_end - data_seq;
1143 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1145 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1146 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1147 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1153 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1155 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1159 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1163 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1165 struct sk_buff *skb;
1167 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1169 if (likely(__mptcp_add_ext(skb, gfp))) {
1170 skb_reserve(skb, MAX_TCP_HEADER);
1171 skb->ip_summed = CHECKSUM_PARTIAL;
1172 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1177 mptcp_enter_memory_pressure(sk);
1182 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1184 struct sk_buff *skb;
1186 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1190 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1191 tcp_skb_entail(ssk, skb);
1194 tcp_skb_tsorted_anchor_cleanup(skb);
1199 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1201 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1203 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1206 /* note: this always recompute the csum on the whole skb, even
1207 * if we just appended a single frag. More status info needed
1209 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1211 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1212 __wsum csum = ~csum_unfold(mpext->csum);
1213 int offset = skb->len - added;
1215 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1218 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1220 struct mptcp_ext *mpext)
1225 mpext->infinite_map = 1;
1226 mpext->data_len = 0;
1228 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1229 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1231 mptcp_do_fallback(ssk);
1234 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1235 struct mptcp_data_frag *dfrag,
1236 struct mptcp_sendmsg_info *info)
1238 u64 data_seq = dfrag->data_seq + info->sent;
1239 int offset = dfrag->offset + info->sent;
1240 struct mptcp_sock *msk = mptcp_sk(sk);
1241 bool zero_window_probe = false;
1242 struct mptcp_ext *mpext = NULL;
1243 bool can_coalesce = false;
1244 bool reuse_skb = true;
1245 struct sk_buff *skb;
1249 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1250 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1252 if (WARN_ON_ONCE(info->sent > info->limit ||
1253 info->limit > dfrag->data_len))
1256 if (unlikely(!__tcp_can_send(ssk)))
1259 /* compute send limit */
1260 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1261 copy = info->size_goal;
1263 skb = tcp_write_queue_tail(ssk);
1264 if (skb && copy > skb->len) {
1265 /* Limit the write to the size available in the
1266 * current skb, if any, so that we create at most a new skb.
1267 * Explicitly tells TCP internals to avoid collapsing on later
1268 * queue management operation, to avoid breaking the ext <->
1269 * SSN association set here
1271 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1272 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1273 TCP_SKB_CB(skb)->eor = 1;
1277 i = skb_shinfo(skb)->nr_frags;
1278 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1279 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1280 tcp_mark_push(tcp_sk(ssk), skb);
1287 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1291 i = skb_shinfo(skb)->nr_frags;
1293 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1296 /* Zero window and all data acked? Probe. */
1297 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1299 u64 snd_una = READ_ONCE(msk->snd_una);
1301 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1302 tcp_remove_empty_skb(ssk);
1306 zero_window_probe = true;
1307 data_seq = snd_una - 1;
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;
1340 memset(mpext, 0, sizeof(*mpext));
1341 mpext->data_seq = data_seq;
1342 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1343 mpext->data_len = copy;
1347 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1348 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1351 if (zero_window_probe) {
1352 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1354 if (READ_ONCE(msk->csum_enabled))
1355 mptcp_update_data_checksum(skb, copy);
1356 tcp_push_pending_frames(ssk);
1360 if (READ_ONCE(msk->csum_enabled))
1361 mptcp_update_data_checksum(skb, copy);
1362 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1363 mptcp_update_infinite_map(msk, ssk, mpext);
1364 trace_mptcp_sendmsg_frag(mpext);
1365 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1369 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1370 sizeof(struct tcphdr) - \
1371 MAX_TCP_OPTION_SPACE - \
1372 sizeof(struct ipv6hdr) - \
1373 sizeof(struct frag_hdr))
1375 struct subflow_send_info {
1380 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1382 if (!subflow->stale)
1386 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1389 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1391 if (unlikely(subflow->stale)) {
1392 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1394 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1397 mptcp_subflow_set_active(subflow);
1399 return __mptcp_subflow_active(subflow);
1402 #define SSK_MODE_ACTIVE 0
1403 #define SSK_MODE_BACKUP 1
1404 #define SSK_MODE_MAX 2
1406 /* implement the mptcp packet scheduler;
1407 * returns the subflow that will transmit the next DSS
1408 * additionally updates the rtx timeout
1410 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1412 struct subflow_send_info send_info[SSK_MODE_MAX];
1413 struct mptcp_subflow_context *subflow;
1414 struct sock *sk = (struct sock *)msk;
1415 u32 pace, burst, wmem;
1416 int i, nr_active = 0;
1421 /* pick the subflow with the lower wmem/wspace ratio */
1422 for (i = 0; i < SSK_MODE_MAX; ++i) {
1423 send_info[i].ssk = NULL;
1424 send_info[i].linger_time = -1;
1427 mptcp_for_each_subflow(msk, subflow) {
1428 trace_mptcp_subflow_get_send(subflow);
1429 ssk = mptcp_subflow_tcp_sock(subflow);
1430 if (!mptcp_subflow_active(subflow))
1433 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1434 nr_active += !subflow->backup;
1435 pace = subflow->avg_pacing_rate;
1436 if (unlikely(!pace)) {
1437 /* init pacing rate from socket */
1438 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1439 pace = subflow->avg_pacing_rate;
1444 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1445 if (linger_time < send_info[subflow->backup].linger_time) {
1446 send_info[subflow->backup].ssk = ssk;
1447 send_info[subflow->backup].linger_time = linger_time;
1450 __mptcp_set_timeout(sk, tout);
1452 /* pick the best backup if no other subflow is active */
1454 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1456 /* According to the blest algorithm, to avoid HoL blocking for the
1457 * faster flow, we need to:
1458 * - estimate the faster flow linger time
1459 * - use the above to estimate the amount of byte transferred
1460 * by the faster flow
1461 * - check that the amount of queued data is greter than the above,
1462 * otherwise do not use the picked, slower, subflow
1463 * We select the subflow with the shorter estimated time to flush
1464 * the queued mem, which basically ensure the above. We just need
1465 * to check that subflow has a non empty cwin.
1467 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1468 if (!ssk || !sk_stream_memory_free(ssk))
1471 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1472 wmem = READ_ONCE(ssk->sk_wmem_queued);
1476 subflow = mptcp_subflow_ctx(ssk);
1477 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1478 READ_ONCE(ssk->sk_pacing_rate) * burst,
1480 msk->snd_burst = burst;
1484 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1486 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1490 static void mptcp_update_post_push(struct mptcp_sock *msk,
1491 struct mptcp_data_frag *dfrag,
1494 u64 snd_nxt_new = dfrag->data_seq;
1496 dfrag->already_sent += sent;
1498 msk->snd_burst -= sent;
1500 snd_nxt_new += dfrag->already_sent;
1502 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1503 * is recovering after a failover. In that event, this re-sends
1506 * Thus compute snd_nxt_new candidate based on
1507 * the dfrag->data_seq that was sent and the data
1508 * that has been handed to the subflow for transmission
1509 * and skip update in case it was old dfrag.
1511 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1512 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1513 msk->snd_nxt = snd_nxt_new;
1517 void mptcp_check_and_set_pending(struct sock *sk)
1519 if (mptcp_send_head(sk))
1520 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1523 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1524 struct mptcp_sendmsg_info *info)
1526 struct mptcp_sock *msk = mptcp_sk(sk);
1527 struct mptcp_data_frag *dfrag;
1528 int len, copied = 0, err = 0;
1530 while ((dfrag = mptcp_send_head(sk))) {
1531 info->sent = dfrag->already_sent;
1532 info->limit = dfrag->data_len;
1533 len = dfrag->data_len - dfrag->already_sent;
1537 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1539 err = copied ? : ret;
1547 mptcp_update_post_push(msk, dfrag, ret);
1549 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1551 if (msk->snd_burst <= 0 ||
1552 !sk_stream_memory_free(ssk) ||
1553 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1557 mptcp_set_timeout(sk);
1565 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1567 struct sock *prev_ssk = NULL, *ssk = NULL;
1568 struct mptcp_sock *msk = mptcp_sk(sk);
1569 struct mptcp_sendmsg_info info = {
1572 bool do_check_data_fin = false;
1575 while (mptcp_send_head(sk) && (push_count > 0)) {
1576 struct mptcp_subflow_context *subflow;
1579 if (mptcp_sched_get_send(msk))
1584 mptcp_for_each_subflow(msk, subflow) {
1585 if (READ_ONCE(subflow->scheduled)) {
1586 mptcp_subflow_set_scheduled(subflow, false);
1589 ssk = mptcp_subflow_tcp_sock(subflow);
1590 if (ssk != prev_ssk) {
1591 /* First check. If the ssk has changed since
1592 * the last round, release prev_ssk
1595 mptcp_push_release(prev_ssk, &info);
1597 /* Need to lock the new subflow only if different
1598 * from the previous one, otherwise we are still
1599 * helding the relevant lock
1606 ret = __subflow_push_pending(sk, ssk, &info);
1608 if (ret != -EAGAIN ||
1609 (1 << ssk->sk_state) &
1610 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1614 do_check_data_fin = true;
1619 /* at this point we held the socket lock for the last subflow we used */
1621 mptcp_push_release(ssk, &info);
1623 /* ensure the rtx timer is running */
1624 if (!mptcp_rtx_timer_pending(sk))
1625 mptcp_reset_rtx_timer(sk);
1626 if (do_check_data_fin)
1627 mptcp_check_send_data_fin(sk);
1630 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1632 struct mptcp_sock *msk = mptcp_sk(sk);
1633 struct mptcp_sendmsg_info info = {
1634 .data_lock_held = true,
1636 bool keep_pushing = true;
1637 struct sock *xmit_ssk;
1641 while (mptcp_send_head(sk) && keep_pushing) {
1642 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1645 /* check for a different subflow usage only after
1646 * spooling the first chunk of data
1649 mptcp_subflow_set_scheduled(subflow, false);
1650 ret = __subflow_push_pending(sk, ssk, &info);
1658 if (mptcp_sched_get_send(msk))
1661 if (READ_ONCE(subflow->scheduled)) {
1662 mptcp_subflow_set_scheduled(subflow, false);
1663 ret = __subflow_push_pending(sk, ssk, &info);
1665 keep_pushing = false;
1669 mptcp_for_each_subflow(msk, subflow) {
1670 if (READ_ONCE(subflow->scheduled)) {
1671 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1672 if (xmit_ssk != ssk) {
1673 mptcp_subflow_delegate(subflow,
1674 MPTCP_DELEGATE_SEND);
1675 keep_pushing = false;
1682 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1683 * not going to flush it via release_sock()
1686 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1688 if (!mptcp_rtx_timer_pending(sk))
1689 mptcp_reset_rtx_timer(sk);
1691 if (msk->snd_data_fin_enable &&
1692 msk->snd_nxt + 1 == msk->write_seq)
1693 mptcp_schedule_work(sk);
1697 static void mptcp_set_nospace(struct sock *sk)
1699 /* enable autotune */
1700 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1702 /* will be cleared on avail space */
1703 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1706 static int mptcp_disconnect(struct sock *sk, int flags);
1708 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1709 size_t len, int *copied_syn)
1711 unsigned int saved_flags = msg->msg_flags;
1712 struct mptcp_sock *msk = mptcp_sk(sk);
1716 /* on flags based fastopen the mptcp is supposed to create the
1717 * first subflow right now. Otherwise we are in the defer_connect
1718 * path, and the first subflow must be already present.
1719 * Since the defer_connect flag is cleared after the first succsful
1720 * fastopen attempt, no need to check for additional subflow status.
1722 if (msg->msg_flags & MSG_FASTOPEN) {
1723 ssk = __mptcp_nmpc_sk(msk);
1725 return PTR_ERR(ssk);
1733 msg->msg_flags |= MSG_DONTWAIT;
1734 msk->fastopening = 1;
1735 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1736 msk->fastopening = 0;
1737 msg->msg_flags = saved_flags;
1740 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1741 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1742 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1743 msg->msg_namelen, msg->msg_flags, 1);
1745 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1746 * case of any error, except timeout or signal
1748 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1750 } else if (ret && ret != -EINPROGRESS) {
1751 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1752 * __inet_stream_connect() can fail, due to looking check,
1753 * see mptcp_disconnect().
1754 * Attempt it again outside the problematic scope.
1756 if (!mptcp_disconnect(sk, 0))
1757 sk->sk_socket->state = SS_UNCONNECTED;
1759 inet_clear_bit(DEFER_CONNECT, sk);
1764 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1766 struct mptcp_sock *msk = mptcp_sk(sk);
1767 struct page_frag *pfrag;
1772 /* silently ignore everything else */
1773 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1777 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1778 msg->msg_flags & MSG_FASTOPEN)) {
1781 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1782 copied += copied_syn;
1783 if (ret == -EINPROGRESS && copied_syn > 0)
1789 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1791 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1792 ret = sk_stream_wait_connect(sk, &timeo);
1798 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1801 pfrag = sk_page_frag(sk);
1803 while (msg_data_left(msg)) {
1804 int total_ts, frag_truesize = 0;
1805 struct mptcp_data_frag *dfrag;
1806 bool dfrag_collapsed;
1807 size_t psize, offset;
1809 /* reuse tail pfrag, if possible, or carve a new one from the
1812 dfrag = mptcp_pending_tail(sk);
1813 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1814 if (!dfrag_collapsed) {
1815 if (!sk_stream_memory_free(sk))
1816 goto wait_for_memory;
1818 if (!mptcp_page_frag_refill(sk, pfrag))
1819 goto wait_for_memory;
1821 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1822 frag_truesize = dfrag->overhead;
1825 /* we do not bound vs wspace, to allow a single packet.
1826 * memory accounting will prevent execessive memory usage
1829 offset = dfrag->offset + dfrag->data_len;
1830 psize = pfrag->size - offset;
1831 psize = min_t(size_t, psize, msg_data_left(msg));
1832 total_ts = psize + frag_truesize;
1834 if (!sk_wmem_schedule(sk, total_ts))
1835 goto wait_for_memory;
1837 if (copy_page_from_iter(dfrag->page, offset, psize,
1838 &msg->msg_iter) != psize) {
1843 /* data successfully copied into the write queue */
1844 sk_forward_alloc_add(sk, -total_ts);
1846 dfrag->data_len += psize;
1847 frag_truesize += psize;
1848 pfrag->offset += frag_truesize;
1849 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1851 /* charge data on mptcp pending queue to the msk socket
1852 * Note: we charge such data both to sk and ssk
1854 sk_wmem_queued_add(sk, frag_truesize);
1855 if (!dfrag_collapsed) {
1856 get_page(dfrag->page);
1857 list_add_tail(&dfrag->list, &msk->rtx_queue);
1858 if (!msk->first_pending)
1859 WRITE_ONCE(msk->first_pending, dfrag);
1861 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1862 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1868 mptcp_set_nospace(sk);
1869 __mptcp_push_pending(sk, msg->msg_flags);
1870 ret = sk_stream_wait_memory(sk, &timeo);
1876 __mptcp_push_pending(sk, msg->msg_flags);
1886 copied = sk_stream_error(sk, msg->msg_flags, ret);
1890 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1892 size_t len, int flags,
1893 struct scm_timestamping_internal *tss,
1896 struct sk_buff *skb, *tmp;
1899 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1900 u32 offset = MPTCP_SKB_CB(skb)->offset;
1901 u32 data_len = skb->len - offset;
1902 u32 count = min_t(size_t, len - copied, data_len);
1905 if (!(flags & MSG_TRUNC)) {
1906 err = skb_copy_datagram_msg(skb, offset, msg, count);
1907 if (unlikely(err < 0)) {
1914 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1915 tcp_update_recv_tstamps(skb, tss);
1916 *cmsg_flags |= MPTCP_CMSG_TS;
1921 if (count < data_len) {
1922 if (!(flags & MSG_PEEK)) {
1923 MPTCP_SKB_CB(skb)->offset += count;
1924 MPTCP_SKB_CB(skb)->map_seq += count;
1929 if (!(flags & MSG_PEEK)) {
1930 /* we will bulk release the skb memory later */
1931 skb->destructor = NULL;
1932 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1933 __skb_unlink(skb, &msk->receive_queue);
1944 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1946 * Only difference: Use highest rtt estimate of the subflows in use.
1948 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1950 struct mptcp_subflow_context *subflow;
1951 struct sock *sk = (struct sock *)msk;
1952 u8 scaling_ratio = U8_MAX;
1953 u32 time, advmss = 1;
1956 msk_owned_by_me(msk);
1961 msk->rcvq_space.copied += copied;
1963 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1964 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1966 rtt_us = msk->rcvq_space.rtt_us;
1967 if (rtt_us && time < (rtt_us >> 3))
1971 mptcp_for_each_subflow(msk, subflow) {
1972 const struct tcp_sock *tp;
1976 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1978 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1979 sf_advmss = READ_ONCE(tp->advmss);
1981 rtt_us = max(sf_rtt_us, rtt_us);
1982 advmss = max(sf_advmss, advmss);
1983 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1986 msk->rcvq_space.rtt_us = rtt_us;
1987 msk->scaling_ratio = scaling_ratio;
1988 if (time < (rtt_us >> 3) || rtt_us == 0)
1991 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1994 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1995 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1999 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2001 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2003 do_div(grow, msk->rcvq_space.space);
2004 rcvwin += (grow << 1);
2006 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2007 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2009 if (rcvbuf > sk->sk_rcvbuf) {
2012 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2013 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2015 /* Make subflows follow along. If we do not do this, we
2016 * get drops at subflow level if skbs can't be moved to
2017 * the mptcp rx queue fast enough (announced rcv_win can
2018 * exceed ssk->sk_rcvbuf).
2020 mptcp_for_each_subflow(msk, subflow) {
2024 ssk = mptcp_subflow_tcp_sock(subflow);
2025 slow = lock_sock_fast(ssk);
2026 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2027 tcp_sk(ssk)->window_clamp = window_clamp;
2028 tcp_cleanup_rbuf(ssk, 1);
2029 unlock_sock_fast(ssk, slow);
2034 msk->rcvq_space.space = msk->rcvq_space.copied;
2036 msk->rcvq_space.copied = 0;
2037 msk->rcvq_space.time = mstamp;
2040 static void __mptcp_update_rmem(struct sock *sk)
2042 struct mptcp_sock *msk = mptcp_sk(sk);
2044 if (!msk->rmem_released)
2047 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2048 mptcp_rmem_uncharge(sk, msk->rmem_released);
2049 WRITE_ONCE(msk->rmem_released, 0);
2052 static void __mptcp_splice_receive_queue(struct sock *sk)
2054 struct mptcp_sock *msk = mptcp_sk(sk);
2056 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2059 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2061 struct sock *sk = (struct sock *)msk;
2062 unsigned int moved = 0;
2066 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2069 /* we can have data pending in the subflows only if the msk
2070 * receive buffer was full at subflow_data_ready() time,
2071 * that is an unlikely slow path.
2076 slowpath = lock_sock_fast(ssk);
2077 mptcp_data_lock(sk);
2078 __mptcp_update_rmem(sk);
2079 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2080 mptcp_data_unlock(sk);
2082 if (unlikely(ssk->sk_err))
2083 __mptcp_error_report(sk);
2084 unlock_sock_fast(ssk, slowpath);
2087 /* acquire the data lock only if some input data is pending */
2089 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2090 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2091 mptcp_data_lock(sk);
2092 __mptcp_update_rmem(sk);
2093 ret |= __mptcp_ofo_queue(msk);
2094 __mptcp_splice_receive_queue(sk);
2095 mptcp_data_unlock(sk);
2098 mptcp_check_data_fin((struct sock *)msk);
2099 return !skb_queue_empty(&msk->receive_queue);
2102 static unsigned int mptcp_inq_hint(const struct sock *sk)
2104 const struct mptcp_sock *msk = mptcp_sk(sk);
2105 const struct sk_buff *skb;
2107 skb = skb_peek(&msk->receive_queue);
2109 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2111 if (hint_val >= INT_MAX)
2114 return (unsigned int)hint_val;
2117 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2123 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2124 int flags, int *addr_len)
2126 struct mptcp_sock *msk = mptcp_sk(sk);
2127 struct scm_timestamping_internal tss;
2128 int copied = 0, cmsg_flags = 0;
2132 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2133 if (unlikely(flags & MSG_ERRQUEUE))
2134 return inet_recv_error(sk, msg, len, addr_len);
2137 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2142 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2144 len = min_t(size_t, len, INT_MAX);
2145 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2147 if (unlikely(msk->recvmsg_inq))
2148 cmsg_flags = MPTCP_CMSG_INQ;
2150 while (copied < len) {
2153 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2154 if (unlikely(bytes_read < 0)) {
2156 copied = bytes_read;
2160 copied += bytes_read;
2162 /* be sure to advertise window change */
2163 mptcp_cleanup_rbuf(msk);
2165 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2168 /* only the master socket status is relevant here. The exit
2169 * conditions mirror closely tcp_recvmsg()
2171 if (copied >= target)
2176 sk->sk_state == TCP_CLOSE ||
2177 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2179 signal_pending(current))
2183 copied = sock_error(sk);
2187 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2188 /* race breaker: the shutdown could be after the
2189 * previous receive queue check
2191 if (__mptcp_move_skbs(msk))
2196 if (sk->sk_state == TCP_CLOSE) {
2206 if (signal_pending(current)) {
2207 copied = sock_intr_errno(timeo);
2212 pr_debug("block timeout %ld", timeo);
2213 sk_wait_data(sk, &timeo, NULL);
2217 if (cmsg_flags && copied >= 0) {
2218 if (cmsg_flags & MPTCP_CMSG_TS)
2219 tcp_recv_timestamp(msg, sk, &tss);
2221 if (cmsg_flags & MPTCP_CMSG_INQ) {
2222 unsigned int inq = mptcp_inq_hint(sk);
2224 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2228 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2229 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2230 skb_queue_empty(&msk->receive_queue), copied);
2231 if (!(flags & MSG_PEEK))
2232 mptcp_rcv_space_adjust(msk, copied);
2238 static void mptcp_retransmit_timer(struct timer_list *t)
2240 struct inet_connection_sock *icsk = from_timer(icsk, t,
2241 icsk_retransmit_timer);
2242 struct sock *sk = &icsk->icsk_inet.sk;
2243 struct mptcp_sock *msk = mptcp_sk(sk);
2246 if (!sock_owned_by_user(sk)) {
2247 /* we need a process context to retransmit */
2248 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2249 mptcp_schedule_work(sk);
2251 /* delegate our work to tcp_release_cb() */
2252 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2258 static void mptcp_tout_timer(struct timer_list *t)
2260 struct sock *sk = from_timer(sk, t, sk_timer);
2262 mptcp_schedule_work(sk);
2266 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2269 * A backup subflow is returned only if that is the only kind available.
2271 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2273 struct sock *backup = NULL, *pick = NULL;
2274 struct mptcp_subflow_context *subflow;
2275 int min_stale_count = INT_MAX;
2277 mptcp_for_each_subflow(msk, subflow) {
2278 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2280 if (!__mptcp_subflow_active(subflow))
2283 /* still data outstanding at TCP level? skip this */
2284 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2285 mptcp_pm_subflow_chk_stale(msk, ssk);
2286 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2290 if (subflow->backup) {
2303 /* use backup only if there are no progresses anywhere */
2304 return min_stale_count > 1 ? backup : NULL;
2307 bool __mptcp_retransmit_pending_data(struct sock *sk)
2309 struct mptcp_data_frag *cur, *rtx_head;
2310 struct mptcp_sock *msk = mptcp_sk(sk);
2312 if (__mptcp_check_fallback(msk))
2315 if (tcp_rtx_and_write_queues_empty(sk))
2318 /* the closing socket has some data untransmitted and/or unacked:
2319 * some data in the mptcp rtx queue has not really xmitted yet.
2320 * keep it simple and re-inject the whole mptcp level rtx queue
2322 mptcp_data_lock(sk);
2323 __mptcp_clean_una_wakeup(sk);
2324 rtx_head = mptcp_rtx_head(sk);
2326 mptcp_data_unlock(sk);
2330 msk->recovery_snd_nxt = msk->snd_nxt;
2331 msk->recovery = true;
2332 mptcp_data_unlock(sk);
2334 msk->first_pending = rtx_head;
2337 /* be sure to clear the "sent status" on all re-injected fragments */
2338 list_for_each_entry(cur, &msk->rtx_queue, list) {
2339 if (!cur->already_sent)
2341 cur->already_sent = 0;
2347 /* flags for __mptcp_close_ssk() */
2348 #define MPTCP_CF_PUSH BIT(1)
2349 #define MPTCP_CF_FASTCLOSE BIT(2)
2351 /* subflow sockets can be either outgoing (connect) or incoming
2354 * Outgoing subflows use in-kernel sockets.
2355 * Incoming subflows do not have their own 'struct socket' allocated,
2356 * so we need to use tcp_close() after detaching them from the mptcp
2359 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2360 struct mptcp_subflow_context *subflow,
2363 struct mptcp_sock *msk = mptcp_sk(sk);
2364 bool dispose_it, need_push = false;
2366 /* If the first subflow moved to a close state before accept, e.g. due
2367 * to an incoming reset or listener shutdown, the subflow socket is
2368 * already deleted by inet_child_forget() and the mptcp socket can't
2371 if (msk->in_accept_queue && msk->first == ssk &&
2372 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2373 /* ensure later check in mptcp_worker() will dispose the msk */
2374 mptcp_set_close_tout(sk, tcp_jiffies32 - (TCP_TIMEWAIT_LEN + 1));
2375 sock_set_flag(sk, SOCK_DEAD);
2376 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2377 mptcp_subflow_drop_ctx(ssk);
2381 dispose_it = msk->free_first || ssk != msk->first;
2383 list_del(&subflow->node);
2385 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2387 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2388 /* be sure to force the tcp_disconnect() path,
2389 * to generate the egress reset
2391 ssk->sk_lingertime = 0;
2392 sock_set_flag(ssk, SOCK_LINGER);
2393 subflow->send_fastclose = 1;
2396 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2398 /* The MPTCP code never wait on the subflow sockets, TCP-level
2399 * disconnect should never fail
2401 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2402 mptcp_subflow_ctx_reset(subflow);
2408 subflow->disposable = 1;
2410 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2411 * the ssk has been already destroyed, we just need to release the
2412 * reference owned by msk;
2414 if (!inet_csk(ssk)->icsk_ulp_ops) {
2415 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2416 kfree_rcu(subflow, rcu);
2418 /* otherwise tcp will dispose of the ssk and subflow ctx */
2419 __tcp_close(ssk, 0);
2421 /* close acquired an extra ref */
2426 __mptcp_subflow_error_report(sk, ssk);
2431 if (ssk == msk->first)
2432 WRITE_ONCE(msk->first, NULL);
2436 __mptcp_push_pending(sk, 0);
2438 /* Catch every 'all subflows closed' scenario, including peers silently
2439 * closing them, e.g. due to timeout.
2440 * For established sockets, allow an additional timeout before closing,
2441 * as the protocol can still create more subflows.
2443 if (list_is_singular(&msk->conn_list) && msk->first &&
2444 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2445 if (sk->sk_state != TCP_ESTABLISHED ||
2446 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2447 inet_sk_state_store(sk, TCP_CLOSE);
2448 mptcp_close_wake_up(sk);
2450 mptcp_start_tout_timer(sk);
2455 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2456 struct mptcp_subflow_context *subflow)
2458 if (sk->sk_state == TCP_ESTABLISHED)
2459 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2461 /* subflow aborted before reaching the fully_established status
2462 * attempt the creation of the next subflow
2464 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2466 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2469 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2474 static void __mptcp_close_subflow(struct sock *sk)
2476 struct mptcp_subflow_context *subflow, *tmp;
2477 struct mptcp_sock *msk = mptcp_sk(sk);
2481 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2482 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2484 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2487 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2488 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2491 mptcp_close_ssk(sk, ssk, subflow);
2496 static bool mptcp_close_tout_expired(const struct sock *sk)
2498 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2499 sk->sk_state == TCP_CLOSE)
2502 return time_after32(tcp_jiffies32,
2503 inet_csk(sk)->icsk_mtup.probe_timestamp + TCP_TIMEWAIT_LEN);
2506 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2508 struct mptcp_subflow_context *subflow, *tmp;
2509 struct sock *sk = (struct sock *)msk;
2511 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2514 mptcp_token_destroy(msk);
2516 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2517 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2520 slow = lock_sock_fast(tcp_sk);
2521 if (tcp_sk->sk_state != TCP_CLOSE) {
2522 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2523 tcp_set_state(tcp_sk, TCP_CLOSE);
2525 unlock_sock_fast(tcp_sk, slow);
2528 /* Mirror the tcp_reset() error propagation */
2529 switch (sk->sk_state) {
2531 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2533 case TCP_CLOSE_WAIT:
2534 WRITE_ONCE(sk->sk_err, EPIPE);
2539 WRITE_ONCE(sk->sk_err, ECONNRESET);
2542 inet_sk_state_store(sk, TCP_CLOSE);
2543 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2544 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2545 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2547 /* the calling mptcp_worker will properly destroy the socket */
2548 if (sock_flag(sk, SOCK_DEAD))
2551 sk->sk_state_change(sk);
2552 sk_error_report(sk);
2555 static void __mptcp_retrans(struct sock *sk)
2557 struct mptcp_sock *msk = mptcp_sk(sk);
2558 struct mptcp_subflow_context *subflow;
2559 struct mptcp_sendmsg_info info = {};
2560 struct mptcp_data_frag *dfrag;
2565 mptcp_clean_una_wakeup(sk);
2567 /* first check ssk: need to kick "stale" logic */
2568 err = mptcp_sched_get_retrans(msk);
2569 dfrag = mptcp_rtx_head(sk);
2571 if (mptcp_data_fin_enabled(msk)) {
2572 struct inet_connection_sock *icsk = inet_csk(sk);
2574 icsk->icsk_retransmits++;
2575 mptcp_set_datafin_timeout(sk);
2576 mptcp_send_ack(msk);
2581 if (!mptcp_send_head(sk))
2590 mptcp_for_each_subflow(msk, subflow) {
2591 if (READ_ONCE(subflow->scheduled)) {
2594 mptcp_subflow_set_scheduled(subflow, false);
2596 ssk = mptcp_subflow_tcp_sock(subflow);
2600 /* limit retransmission to the bytes already sent on some subflows */
2602 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2603 dfrag->already_sent;
2604 while (info.sent < info.limit) {
2605 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2609 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2614 len = max(copied, len);
2615 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2617 WRITE_ONCE(msk->allow_infinite_fallback, false);
2624 msk->bytes_retrans += len;
2625 dfrag->already_sent = max(dfrag->already_sent, len);
2628 mptcp_check_and_set_pending(sk);
2630 if (!mptcp_rtx_timer_pending(sk))
2631 mptcp_reset_rtx_timer(sk);
2634 /* schedule the timeout timer for the relevant event: either close timeout
2635 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2637 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2639 struct sock *sk = (struct sock *)msk;
2640 unsigned long timeout, close_timeout;
2642 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2645 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2648 /* the close timeout takes precedence on the fail one, and here at least one of
2651 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2653 sk_reset_timer(sk, &sk->sk_timer, timeout);
2656 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2658 struct sock *ssk = msk->first;
2664 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2666 slow = lock_sock_fast(ssk);
2667 mptcp_subflow_reset(ssk);
2668 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2669 unlock_sock_fast(ssk, slow);
2672 static void mptcp_do_fastclose(struct sock *sk)
2674 struct mptcp_subflow_context *subflow, *tmp;
2675 struct mptcp_sock *msk = mptcp_sk(sk);
2677 inet_sk_state_store(sk, TCP_CLOSE);
2678 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2679 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2680 subflow, MPTCP_CF_FASTCLOSE);
2683 static void mptcp_worker(struct work_struct *work)
2685 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2686 struct sock *sk = (struct sock *)msk;
2687 unsigned long fail_tout;
2691 state = sk->sk_state;
2692 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2695 mptcp_check_fastclose(msk);
2697 mptcp_pm_nl_work(msk);
2699 mptcp_check_send_data_fin(sk);
2700 mptcp_check_data_fin_ack(sk);
2701 mptcp_check_data_fin(sk);
2703 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2704 __mptcp_close_subflow(sk);
2706 if (mptcp_close_tout_expired(sk)) {
2707 mptcp_do_fastclose(sk);
2708 mptcp_close_wake_up(sk);
2711 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2712 __mptcp_destroy_sock(sk);
2716 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2717 __mptcp_retrans(sk);
2719 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2720 if (fail_tout && time_after(jiffies, fail_tout))
2721 mptcp_mp_fail_no_response(msk);
2728 static void __mptcp_init_sock(struct sock *sk)
2730 struct mptcp_sock *msk = mptcp_sk(sk);
2732 INIT_LIST_HEAD(&msk->conn_list);
2733 INIT_LIST_HEAD(&msk->join_list);
2734 INIT_LIST_HEAD(&msk->rtx_queue);
2735 INIT_WORK(&msk->work, mptcp_worker);
2736 __skb_queue_head_init(&msk->receive_queue);
2737 msk->out_of_order_queue = RB_ROOT;
2738 msk->first_pending = NULL;
2739 msk->rmem_fwd_alloc = 0;
2740 WRITE_ONCE(msk->rmem_released, 0);
2741 msk->timer_ival = TCP_RTO_MIN;
2743 WRITE_ONCE(msk->first, NULL);
2744 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2745 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2746 WRITE_ONCE(msk->allow_infinite_fallback, true);
2747 msk->recovery = false;
2748 msk->subflow_id = 1;
2750 mptcp_pm_data_init(msk);
2752 /* re-use the csk retrans timer for MPTCP-level retrans */
2753 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2754 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2757 static void mptcp_ca_reset(struct sock *sk)
2759 struct inet_connection_sock *icsk = inet_csk(sk);
2761 tcp_assign_congestion_control(sk);
2762 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2764 /* no need to keep a reference to the ops, the name will suffice */
2765 tcp_cleanup_congestion_control(sk);
2766 icsk->icsk_ca_ops = NULL;
2769 static int mptcp_init_sock(struct sock *sk)
2771 struct net *net = sock_net(sk);
2774 __mptcp_init_sock(sk);
2776 if (!mptcp_is_enabled(net))
2777 return -ENOPROTOOPT;
2779 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2782 ret = mptcp_init_sched(mptcp_sk(sk),
2783 mptcp_sched_find(mptcp_get_scheduler(net)));
2787 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2789 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2790 * propagate the correct value
2794 sk_sockets_allocated_inc(sk);
2795 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2796 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2801 static void __mptcp_clear_xmit(struct sock *sk)
2803 struct mptcp_sock *msk = mptcp_sk(sk);
2804 struct mptcp_data_frag *dtmp, *dfrag;
2806 WRITE_ONCE(msk->first_pending, NULL);
2807 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2808 dfrag_clear(sk, dfrag);
2811 void mptcp_cancel_work(struct sock *sk)
2813 struct mptcp_sock *msk = mptcp_sk(sk);
2815 if (cancel_work_sync(&msk->work))
2819 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2823 switch (ssk->sk_state) {
2825 if (!(how & RCV_SHUTDOWN))
2829 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2832 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2833 pr_debug("Fallback");
2834 ssk->sk_shutdown |= how;
2835 tcp_shutdown(ssk, how);
2837 /* simulate the data_fin ack reception to let the state
2838 * machine move forward
2840 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2841 mptcp_schedule_work(sk);
2843 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2845 if (!mptcp_rtx_timer_pending(sk))
2846 mptcp_reset_rtx_timer(sk);
2854 static const unsigned char new_state[16] = {
2855 /* current state: new state: action: */
2856 [0 /* (Invalid) */] = TCP_CLOSE,
2857 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2858 [TCP_SYN_SENT] = TCP_CLOSE,
2859 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2860 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2861 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2862 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2863 [TCP_CLOSE] = TCP_CLOSE,
2864 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2865 [TCP_LAST_ACK] = TCP_LAST_ACK,
2866 [TCP_LISTEN] = TCP_CLOSE,
2867 [TCP_CLOSING] = TCP_CLOSING,
2868 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2871 static int mptcp_close_state(struct sock *sk)
2873 int next = (int)new_state[sk->sk_state];
2874 int ns = next & TCP_STATE_MASK;
2876 inet_sk_state_store(sk, ns);
2878 return next & TCP_ACTION_FIN;
2881 static void mptcp_check_send_data_fin(struct sock *sk)
2883 struct mptcp_subflow_context *subflow;
2884 struct mptcp_sock *msk = mptcp_sk(sk);
2886 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2887 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2888 msk->snd_nxt, msk->write_seq);
2890 /* we still need to enqueue subflows or not really shutting down,
2893 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2894 mptcp_send_head(sk))
2897 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2899 mptcp_for_each_subflow(msk, subflow) {
2900 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2902 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2906 static void __mptcp_wr_shutdown(struct sock *sk)
2908 struct mptcp_sock *msk = mptcp_sk(sk);
2910 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2911 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2912 !!mptcp_send_head(sk));
2914 /* will be ignored by fallback sockets */
2915 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2916 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2918 mptcp_check_send_data_fin(sk);
2921 static void __mptcp_destroy_sock(struct sock *sk)
2923 struct mptcp_sock *msk = mptcp_sk(sk);
2925 pr_debug("msk=%p", msk);
2929 mptcp_stop_rtx_timer(sk);
2930 sk_stop_timer(sk, &sk->sk_timer);
2932 mptcp_release_sched(msk);
2934 sk->sk_prot->destroy(sk);
2936 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2937 WARN_ON_ONCE(msk->rmem_released);
2938 sk_stream_kill_queues(sk);
2939 xfrm_sk_free_policy(sk);
2944 void __mptcp_unaccepted_force_close(struct sock *sk)
2946 sock_set_flag(sk, SOCK_DEAD);
2947 mptcp_do_fastclose(sk);
2948 __mptcp_destroy_sock(sk);
2951 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2953 /* Concurrent splices from sk_receive_queue into receive_queue will
2954 * always show at least one non-empty queue when checked in this order.
2956 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2957 skb_queue_empty_lockless(&msk->receive_queue))
2960 return EPOLLIN | EPOLLRDNORM;
2963 static void mptcp_check_listen_stop(struct sock *sk)
2967 if (inet_sk_state_load(sk) != TCP_LISTEN)
2970 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2971 ssk = mptcp_sk(sk)->first;
2972 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2975 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2976 tcp_set_state(ssk, TCP_CLOSE);
2977 mptcp_subflow_queue_clean(sk, ssk);
2978 inet_csk_listen_stop(ssk);
2979 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2983 bool __mptcp_close(struct sock *sk, long timeout)
2985 struct mptcp_subflow_context *subflow;
2986 struct mptcp_sock *msk = mptcp_sk(sk);
2987 bool do_cancel_work = false;
2988 int subflows_alive = 0;
2990 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2992 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2993 mptcp_check_listen_stop(sk);
2994 inet_sk_state_store(sk, TCP_CLOSE);
2998 if (mptcp_check_readable(msk) || timeout < 0) {
2999 /* If the msk has read data, or the caller explicitly ask it,
3000 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3002 mptcp_do_fastclose(sk);
3004 } else if (mptcp_close_state(sk)) {
3005 __mptcp_wr_shutdown(sk);
3008 sk_stream_wait_close(sk, timeout);
3011 /* orphan all the subflows */
3012 mptcp_for_each_subflow(msk, subflow) {
3013 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3014 bool slow = lock_sock_fast_nested(ssk);
3016 subflows_alive += ssk->sk_state != TCP_CLOSE;
3018 /* since the close timeout takes precedence on the fail one,
3021 if (ssk == msk->first)
3022 subflow->fail_tout = 0;
3024 /* detach from the parent socket, but allow data_ready to
3025 * push incoming data into the mptcp stack, to properly ack it
3027 ssk->sk_socket = NULL;
3029 unlock_sock_fast(ssk, slow);
3033 /* all the subflows are closed, only timeout can change the msk
3034 * state, let's not keep resources busy for no reasons
3036 if (subflows_alive == 0)
3037 inet_sk_state_store(sk, TCP_CLOSE);
3040 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3042 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3044 if (sk->sk_state == TCP_CLOSE) {
3045 __mptcp_destroy_sock(sk);
3046 do_cancel_work = true;
3048 mptcp_start_tout_timer(sk);
3051 return do_cancel_work;
3054 static void mptcp_close(struct sock *sk, long timeout)
3056 bool do_cancel_work;
3060 do_cancel_work = __mptcp_close(sk, timeout);
3063 mptcp_cancel_work(sk);
3068 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3070 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3071 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3072 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3074 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3075 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3078 msk6->saddr = ssk6->saddr;
3079 msk6->flow_label = ssk6->flow_label;
3083 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3084 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3085 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3086 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3087 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3088 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3091 static int mptcp_disconnect(struct sock *sk, int flags)
3093 struct mptcp_sock *msk = mptcp_sk(sk);
3095 /* We are on the fastopen error path. We can't call straight into the
3096 * subflows cleanup code due to lock nesting (we are already under
3097 * msk->firstsocket lock).
3099 if (msk->fastopening)
3102 mptcp_check_listen_stop(sk);
3103 inet_sk_state_store(sk, TCP_CLOSE);
3105 mptcp_stop_rtx_timer(sk);
3106 mptcp_stop_tout_timer(sk);
3109 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3111 /* msk->subflow is still intact, the following will not free the first
3114 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3115 WRITE_ONCE(msk->flags, 0);
3117 msk->push_pending = 0;
3118 msk->recovery = false;
3119 msk->can_ack = false;
3120 msk->fully_established = false;
3121 msk->rcv_data_fin = false;
3122 msk->snd_data_fin_enable = false;
3123 msk->rcv_fastclose = false;
3124 msk->use_64bit_ack = false;
3125 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3126 mptcp_pm_data_reset(msk);
3128 msk->bytes_acked = 0;
3129 msk->bytes_received = 0;
3130 msk->bytes_sent = 0;
3131 msk->bytes_retrans = 0;
3133 WRITE_ONCE(sk->sk_shutdown, 0);
3134 sk_error_report(sk);
3138 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3139 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3141 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3143 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3147 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3148 const struct mptcp_options_received *mp_opt,
3150 struct request_sock *req)
3152 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3153 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3154 struct mptcp_sock *msk;
3159 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3160 if (nsk->sk_family == AF_INET6)
3161 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3164 __mptcp_init_sock(nsk);
3166 msk = mptcp_sk(nsk);
3167 msk->local_key = subflow_req->local_key;
3168 msk->token = subflow_req->token;
3169 msk->in_accept_queue = 1;
3170 WRITE_ONCE(msk->fully_established, false);
3171 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3172 WRITE_ONCE(msk->csum_enabled, true);
3174 msk->write_seq = subflow_req->idsn + 1;
3175 msk->snd_nxt = msk->write_seq;
3176 msk->snd_una = msk->write_seq;
3177 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3178 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3179 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3181 /* passive msk is created after the first/MPC subflow */
3182 msk->subflow_id = 2;
3184 sock_reset_flag(nsk, SOCK_RCU_FREE);
3185 security_inet_csk_clone(nsk, req);
3187 /* this can't race with mptcp_close(), as the msk is
3188 * not yet exposted to user-space
3190 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3192 /* The msk maintain a ref to each subflow in the connections list */
3193 WRITE_ONCE(msk->first, ssk);
3194 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3197 /* new mpc subflow takes ownership of the newly
3198 * created mptcp socket
3200 mptcp_token_accept(subflow_req, msk);
3202 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3203 * uses the correct data
3205 mptcp_copy_inaddrs(nsk, ssk);
3206 mptcp_propagate_sndbuf(nsk, ssk);
3208 mptcp_rcv_space_init(msk, ssk);
3209 bh_unlock_sock(nsk);
3211 /* note: the newly allocated socket refcount is 2 now */
3215 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3217 const struct tcp_sock *tp = tcp_sk(ssk);
3219 msk->rcvq_space.copied = 0;
3220 msk->rcvq_space.rtt_us = 0;
3222 msk->rcvq_space.time = tp->tcp_mstamp;
3224 /* initial rcv_space offering made to peer */
3225 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3226 TCP_INIT_CWND * tp->advmss);
3227 if (msk->rcvq_space.space == 0)
3228 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3230 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3233 static struct sock *mptcp_accept(struct sock *ssk, int flags, int *err,
3238 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3239 newsk = inet_csk_accept(ssk, flags, err, kern);
3243 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3244 if (sk_is_mptcp(newsk)) {
3245 struct mptcp_subflow_context *subflow;
3246 struct sock *new_mptcp_sock;
3248 subflow = mptcp_subflow_ctx(newsk);
3249 new_mptcp_sock = subflow->conn;
3251 /* is_mptcp should be false if subflow->conn is missing, see
3252 * subflow_syn_recv_sock()
3254 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3255 tcp_sk(newsk)->is_mptcp = 0;
3259 newsk = new_mptcp_sock;
3260 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3262 MPTCP_INC_STATS(sock_net(ssk),
3263 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3267 newsk->sk_kern_sock = kern;
3271 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3273 struct mptcp_subflow_context *subflow, *tmp;
3274 struct sock *sk = (struct sock *)msk;
3276 __mptcp_clear_xmit(sk);
3278 /* join list will be eventually flushed (with rst) at sock lock release time */
3279 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3280 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3282 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3283 mptcp_data_lock(sk);
3284 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3285 __skb_queue_purge(&sk->sk_receive_queue);
3286 skb_rbtree_purge(&msk->out_of_order_queue);
3287 mptcp_data_unlock(sk);
3289 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3290 * inet_sock_destruct() will dispose it
3292 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3293 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3294 mptcp_token_destroy(msk);
3295 mptcp_pm_free_anno_list(msk);
3296 mptcp_free_local_addr_list(msk);
3299 static void mptcp_destroy(struct sock *sk)
3301 struct mptcp_sock *msk = mptcp_sk(sk);
3303 /* allow the following to close even the initial subflow */
3304 msk->free_first = 1;
3305 mptcp_destroy_common(msk, 0);
3306 sk_sockets_allocated_dec(sk);
3309 void __mptcp_data_acked(struct sock *sk)
3311 if (!sock_owned_by_user(sk))
3312 __mptcp_clean_una(sk);
3314 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3316 if (mptcp_pending_data_fin_ack(sk))
3317 mptcp_schedule_work(sk);
3320 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3322 if (!mptcp_send_head(sk))
3325 if (!sock_owned_by_user(sk))
3326 __mptcp_subflow_push_pending(sk, ssk, false);
3328 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3331 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3332 BIT(MPTCP_RETRANSMIT) | \
3333 BIT(MPTCP_FLUSH_JOIN_LIST))
3335 /* processes deferred events and flush wmem */
3336 static void mptcp_release_cb(struct sock *sk)
3337 __must_hold(&sk->sk_lock.slock)
3339 struct mptcp_sock *msk = mptcp_sk(sk);
3342 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3344 struct list_head join_list;
3349 INIT_LIST_HEAD(&join_list);
3350 list_splice_init(&msk->join_list, &join_list);
3352 /* the following actions acquire the subflow socket lock
3354 * 1) can't be invoked in atomic scope
3355 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3356 * datapath acquires the msk socket spinlock while helding
3357 * the subflow socket lock
3359 msk->push_pending = 0;
3360 msk->cb_flags &= ~flags;
3361 spin_unlock_bh(&sk->sk_lock.slock);
3363 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3364 __mptcp_flush_join_list(sk, &join_list);
3365 if (flags & BIT(MPTCP_PUSH_PENDING))
3366 __mptcp_push_pending(sk, 0);
3367 if (flags & BIT(MPTCP_RETRANSMIT))
3368 __mptcp_retrans(sk);
3371 spin_lock_bh(&sk->sk_lock.slock);
3374 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3375 __mptcp_clean_una_wakeup(sk);
3376 if (unlikely(msk->cb_flags)) {
3377 /* be sure to set the current sk state before tacking actions
3378 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3380 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3381 __mptcp_set_connected(sk);
3382 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3383 __mptcp_error_report(sk);
3386 __mptcp_update_rmem(sk);
3389 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3390 * TCP can't schedule delack timer before the subflow is fully established.
3391 * MPTCP uses the delack timer to do 3rd ack retransmissions
3393 static void schedule_3rdack_retransmission(struct sock *ssk)
3395 struct inet_connection_sock *icsk = inet_csk(ssk);
3396 struct tcp_sock *tp = tcp_sk(ssk);
3397 unsigned long timeout;
3399 if (mptcp_subflow_ctx(ssk)->fully_established)
3402 /* reschedule with a timeout above RTT, as we must look only for drop */
3404 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3406 timeout = TCP_TIMEOUT_INIT;
3409 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3410 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3411 icsk->icsk_ack.timeout = timeout;
3412 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3415 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3417 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3418 struct sock *sk = subflow->conn;
3420 if (status & BIT(MPTCP_DELEGATE_SEND)) {
3421 mptcp_data_lock(sk);
3422 if (!sock_owned_by_user(sk))
3423 __mptcp_subflow_push_pending(sk, ssk, true);
3425 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3426 mptcp_data_unlock(sk);
3428 if (status & BIT(MPTCP_DELEGATE_ACK))
3429 schedule_3rdack_retransmission(ssk);
3432 static int mptcp_hash(struct sock *sk)
3434 /* should never be called,
3435 * we hash the TCP subflows not the master socket
3441 static void mptcp_unhash(struct sock *sk)
3443 /* called from sk_common_release(), but nothing to do here */
3446 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3448 struct mptcp_sock *msk = mptcp_sk(sk);
3450 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3451 if (WARN_ON_ONCE(!msk->first))
3454 return inet_csk_get_port(msk->first, snum);
3457 void mptcp_finish_connect(struct sock *ssk)
3459 struct mptcp_subflow_context *subflow;
3460 struct mptcp_sock *msk;
3463 subflow = mptcp_subflow_ctx(ssk);
3467 pr_debug("msk=%p, token=%u", sk, subflow->token);
3469 subflow->map_seq = subflow->iasn;
3470 subflow->map_subflow_seq = 1;
3472 /* the socket is not connected yet, no msk/subflow ops can access/race
3473 * accessing the field below
3475 WRITE_ONCE(msk->local_key, subflow->local_key);
3476 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3477 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3478 WRITE_ONCE(msk->snd_una, msk->write_seq);
3480 mptcp_pm_new_connection(msk, ssk, 0);
3482 mptcp_rcv_space_init(msk, ssk);
3485 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3487 write_lock_bh(&sk->sk_callback_lock);
3488 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3489 sk_set_socket(sk, parent);
3490 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3491 write_unlock_bh(&sk->sk_callback_lock);
3494 bool mptcp_finish_join(struct sock *ssk)
3496 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3497 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3498 struct sock *parent = (void *)msk;
3501 pr_debug("msk=%p, subflow=%p", msk, subflow);
3503 /* mptcp socket already closing? */
3504 if (!mptcp_is_fully_established(parent)) {
3505 subflow->reset_reason = MPTCP_RST_EMPTCP;
3509 /* active subflow, already present inside the conn_list */
3510 if (!list_empty(&subflow->node)) {
3511 mptcp_subflow_joined(msk, ssk);
3515 if (!mptcp_pm_allow_new_subflow(msk))
3516 goto err_prohibited;
3518 /* If we can't acquire msk socket lock here, let the release callback
3521 mptcp_data_lock(parent);
3522 if (!sock_owned_by_user(parent)) {
3523 ret = __mptcp_finish_join(msk, ssk);
3526 list_add_tail(&subflow->node, &msk->conn_list);
3530 list_add_tail(&subflow->node, &msk->join_list);
3531 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3533 mptcp_data_unlock(parent);
3537 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3544 static void mptcp_shutdown(struct sock *sk, int how)
3546 pr_debug("sk=%p, how=%d", sk, how);
3548 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3549 __mptcp_wr_shutdown(sk);
3552 static int mptcp_forward_alloc_get(const struct sock *sk)
3554 return READ_ONCE(sk->sk_forward_alloc) +
3555 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3558 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3560 const struct sock *sk = (void *)msk;
3563 if (sk->sk_state == TCP_LISTEN)
3566 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3569 delta = msk->write_seq - v;
3570 if (__mptcp_check_fallback(msk) && msk->first) {
3571 struct tcp_sock *tp = tcp_sk(msk->first);
3573 /* the first subflow is disconnected after close - see
3574 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3575 * so ignore that status, too.
3577 if (!((1 << msk->first->sk_state) &
3578 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3579 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3581 if (delta > INT_MAX)
3587 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3589 struct mptcp_sock *msk = mptcp_sk(sk);
3594 if (sk->sk_state == TCP_LISTEN)
3598 __mptcp_move_skbs(msk);
3599 *karg = mptcp_inq_hint(sk);
3603 slow = lock_sock_fast(sk);
3604 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3605 unlock_sock_fast(sk, slow);
3608 slow = lock_sock_fast(sk);
3609 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3610 unlock_sock_fast(sk, slow);
3613 return -ENOIOCTLCMD;
3619 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3620 struct mptcp_subflow_context *subflow)
3622 subflow->request_mptcp = 0;
3623 __mptcp_do_fallback(msk);
3626 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3628 struct mptcp_subflow_context *subflow;
3629 struct mptcp_sock *msk = mptcp_sk(sk);
3633 ssk = __mptcp_nmpc_sk(msk);
3635 return PTR_ERR(ssk);
3637 inet_sk_state_store(sk, TCP_SYN_SENT);
3638 subflow = mptcp_subflow_ctx(ssk);
3639 #ifdef CONFIG_TCP_MD5SIG
3640 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3643 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3644 mptcp_subflow_early_fallback(msk, subflow);
3646 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3647 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3648 mptcp_subflow_early_fallback(msk, subflow);
3650 if (likely(!__mptcp_check_fallback(msk)))
3651 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3653 /* if reaching here via the fastopen/sendmsg path, the caller already
3654 * acquired the subflow socket lock, too.
3656 if (!msk->fastopening)
3659 /* the following mirrors closely a very small chunk of code from
3660 * __inet_stream_connect()
3662 if (ssk->sk_state != TCP_CLOSE)
3665 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3666 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3671 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3675 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3678 if (!msk->fastopening)
3681 /* on successful connect, the msk state will be moved to established by
3682 * subflow_finish_connect()
3684 if (unlikely(err)) {
3685 /* avoid leaving a dangling token in an unconnected socket */
3686 mptcp_token_destroy(msk);
3687 inet_sk_state_store(sk, TCP_CLOSE);
3691 mptcp_copy_inaddrs(sk, ssk);
3695 static struct proto mptcp_prot = {
3697 .owner = THIS_MODULE,
3698 .init = mptcp_init_sock,
3699 .connect = mptcp_connect,
3700 .disconnect = mptcp_disconnect,
3701 .close = mptcp_close,
3702 .accept = mptcp_accept,
3703 .setsockopt = mptcp_setsockopt,
3704 .getsockopt = mptcp_getsockopt,
3705 .shutdown = mptcp_shutdown,
3706 .destroy = mptcp_destroy,
3707 .sendmsg = mptcp_sendmsg,
3708 .ioctl = mptcp_ioctl,
3709 .recvmsg = mptcp_recvmsg,
3710 .release_cb = mptcp_release_cb,
3712 .unhash = mptcp_unhash,
3713 .get_port = mptcp_get_port,
3714 .forward_alloc_get = mptcp_forward_alloc_get,
3715 .sockets_allocated = &mptcp_sockets_allocated,
3717 .memory_allocated = &tcp_memory_allocated,
3718 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3720 .memory_pressure = &tcp_memory_pressure,
3721 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3722 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3723 .sysctl_mem = sysctl_tcp_mem,
3724 .obj_size = sizeof(struct mptcp_sock),
3725 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3726 .no_autobind = true,
3729 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3731 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3732 struct sock *ssk, *sk = sock->sk;
3736 ssk = __mptcp_nmpc_sk(msk);
3742 if (sk->sk_family == AF_INET)
3743 err = inet_bind_sk(ssk, uaddr, addr_len);
3744 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3745 else if (sk->sk_family == AF_INET6)
3746 err = inet6_bind_sk(ssk, uaddr, addr_len);
3749 mptcp_copy_inaddrs(sk, ssk);
3756 static int mptcp_listen(struct socket *sock, int backlog)
3758 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3759 struct sock *sk = sock->sk;
3763 pr_debug("msk=%p", msk);
3768 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3771 ssk = __mptcp_nmpc_sk(msk);
3777 inet_sk_state_store(sk, TCP_LISTEN);
3778 sock_set_flag(sk, SOCK_RCU_FREE);
3781 err = __inet_listen_sk(ssk, backlog);
3783 inet_sk_state_store(sk, inet_sk_state_load(ssk));
3786 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3787 mptcp_copy_inaddrs(sk, ssk);
3788 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3796 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3797 int flags, bool kern)
3799 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3800 struct sock *ssk, *newsk;
3803 pr_debug("msk=%p", msk);
3805 /* Buggy applications can call accept on socket states other then LISTEN
3806 * but no need to allocate the first subflow just to error out.
3808 ssk = READ_ONCE(msk->first);
3812 newsk = mptcp_accept(ssk, flags, &err, kern);
3818 __inet_accept(sock, newsock, newsk);
3819 if (!mptcp_is_tcpsk(newsock->sk)) {
3820 struct mptcp_sock *msk = mptcp_sk(newsk);
3821 struct mptcp_subflow_context *subflow;
3823 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3824 msk->in_accept_queue = 0;
3826 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3827 * This is needed so NOSPACE flag can be set from tcp stack.
3829 mptcp_for_each_subflow(msk, subflow) {
3830 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3832 if (!ssk->sk_socket)
3833 mptcp_sock_graft(ssk, newsock);
3836 /* Do late cleanup for the first subflow as necessary. Also
3837 * deal with bad peers not doing a complete shutdown.
3839 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3840 __mptcp_close_ssk(newsk, msk->first,
3841 mptcp_subflow_ctx(msk->first), 0);
3842 if (unlikely(list_is_singular(&msk->conn_list)))
3843 inet_sk_state_store(newsk, TCP_CLOSE);
3846 release_sock(newsk);
3851 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3853 struct sock *sk = (struct sock *)msk;
3855 if (sk_stream_is_writeable(sk))
3856 return EPOLLOUT | EPOLLWRNORM;
3858 mptcp_set_nospace(sk);
3859 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3860 if (sk_stream_is_writeable(sk))
3861 return EPOLLOUT | EPOLLWRNORM;
3866 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3867 struct poll_table_struct *wait)
3869 struct sock *sk = sock->sk;
3870 struct mptcp_sock *msk;
3876 sock_poll_wait(file, sock, wait);
3878 state = inet_sk_state_load(sk);
3879 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3880 if (state == TCP_LISTEN) {
3881 struct sock *ssk = READ_ONCE(msk->first);
3883 if (WARN_ON_ONCE(!ssk))
3886 return inet_csk_listen_poll(ssk);
3889 shutdown = READ_ONCE(sk->sk_shutdown);
3890 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3892 if (shutdown & RCV_SHUTDOWN)
3893 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3895 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3896 mask |= mptcp_check_readable(msk);
3897 if (shutdown & SEND_SHUTDOWN)
3898 mask |= EPOLLOUT | EPOLLWRNORM;
3900 mask |= mptcp_check_writeable(msk);
3901 } else if (state == TCP_SYN_SENT &&
3902 inet_test_bit(DEFER_CONNECT, sk)) {
3903 /* cf tcp_poll() note about TFO */
3904 mask |= EPOLLOUT | EPOLLWRNORM;
3907 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3909 if (READ_ONCE(sk->sk_err))
3915 static const struct proto_ops mptcp_stream_ops = {
3917 .owner = THIS_MODULE,
3918 .release = inet_release,
3920 .connect = inet_stream_connect,
3921 .socketpair = sock_no_socketpair,
3922 .accept = mptcp_stream_accept,
3923 .getname = inet_getname,
3925 .ioctl = inet_ioctl,
3926 .gettstamp = sock_gettstamp,
3927 .listen = mptcp_listen,
3928 .shutdown = inet_shutdown,
3929 .setsockopt = sock_common_setsockopt,
3930 .getsockopt = sock_common_getsockopt,
3931 .sendmsg = inet_sendmsg,
3932 .recvmsg = inet_recvmsg,
3933 .mmap = sock_no_mmap,
3936 static struct inet_protosw mptcp_protosw = {
3937 .type = SOCK_STREAM,
3938 .protocol = IPPROTO_MPTCP,
3939 .prot = &mptcp_prot,
3940 .ops = &mptcp_stream_ops,
3941 .flags = INET_PROTOSW_ICSK,
3944 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3946 struct mptcp_delegated_action *delegated;
3947 struct mptcp_subflow_context *subflow;
3950 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3951 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3952 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3954 bh_lock_sock_nested(ssk);
3955 if (!sock_owned_by_user(ssk)) {
3956 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
3958 /* tcp_release_cb_override already processed
3959 * the action or will do at next release_sock().
3960 * In both case must dequeue the subflow here - on the same
3961 * CPU that scheduled it.
3964 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
3966 bh_unlock_sock(ssk);
3969 if (++work_done == budget)
3973 /* always provide a 0 'work_done' argument, so that napi_complete_done
3974 * will not try accessing the NULL napi->dev ptr
3976 napi_complete_done(napi, 0);
3980 void __init mptcp_proto_init(void)
3982 struct mptcp_delegated_action *delegated;
3985 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3987 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3988 panic("Failed to allocate MPTCP pcpu counter\n");
3990 init_dummy_netdev(&mptcp_napi_dev);
3991 for_each_possible_cpu(cpu) {
3992 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3993 INIT_LIST_HEAD(&delegated->head);
3994 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3996 napi_enable(&delegated->napi);
3999 mptcp_subflow_init();
4004 if (proto_register(&mptcp_prot, 1) != 0)
4005 panic("Failed to register MPTCP proto.\n");
4007 inet_register_protosw(&mptcp_protosw);
4009 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4012 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4013 static const struct proto_ops mptcp_v6_stream_ops = {
4015 .owner = THIS_MODULE,
4016 .release = inet6_release,
4018 .connect = inet_stream_connect,
4019 .socketpair = sock_no_socketpair,
4020 .accept = mptcp_stream_accept,
4021 .getname = inet6_getname,
4023 .ioctl = inet6_ioctl,
4024 .gettstamp = sock_gettstamp,
4025 .listen = mptcp_listen,
4026 .shutdown = inet_shutdown,
4027 .setsockopt = sock_common_setsockopt,
4028 .getsockopt = sock_common_getsockopt,
4029 .sendmsg = inet6_sendmsg,
4030 .recvmsg = inet6_recvmsg,
4031 .mmap = sock_no_mmap,
4032 #ifdef CONFIG_COMPAT
4033 .compat_ioctl = inet6_compat_ioctl,
4037 static struct proto mptcp_v6_prot;
4039 static struct inet_protosw mptcp_v6_protosw = {
4040 .type = SOCK_STREAM,
4041 .protocol = IPPROTO_MPTCP,
4042 .prot = &mptcp_v6_prot,
4043 .ops = &mptcp_v6_stream_ops,
4044 .flags = INET_PROTOSW_ICSK,
4047 int __init mptcp_proto_v6_init(void)
4051 mptcp_v6_prot = mptcp_prot;
4052 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4053 mptcp_v6_prot.slab = NULL;
4054 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4055 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4057 err = proto_register(&mptcp_v6_prot, 1);
4061 err = inet6_register_protosw(&mptcp_v6_protosw);
4063 proto_unregister(&mptcp_v6_prot);