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 WRITE_ONCE(msk->subflow, ssock);
96 subflow = mptcp_subflow_ctx(ssock->sk);
97 list_add(&subflow->node, &msk->conn_list);
99 subflow->request_mptcp = 1;
100 subflow->subflow_id = msk->subflow_id++;
102 /* This is the first subflow, always with id 0 */
103 subflow->local_id_valid = 1;
104 mptcp_sock_graft(msk->first, sk->sk_socket);
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_charge(struct sock *sk, int size)
139 mptcp_sk(sk)->rmem_fwd_alloc -= size;
142 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
143 struct sk_buff *from)
148 if (MPTCP_SKB_CB(from)->offset ||
149 !skb_try_coalesce(to, from, &fragstolen, &delta))
152 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
153 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
154 to->len, MPTCP_SKB_CB(from)->end_seq);
155 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
157 /* note the fwd memory can reach a negative value after accounting
158 * for the delta, but the later skb free will restore a non
161 atomic_add(delta, &sk->sk_rmem_alloc);
162 mptcp_rmem_charge(sk, delta);
163 kfree_skb_partial(from, fragstolen);
168 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
169 struct sk_buff *from)
171 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
174 return mptcp_try_coalesce((struct sock *)msk, to, from);
177 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
179 amount >>= PAGE_SHIFT;
180 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
181 __sk_mem_reduce_allocated(sk, amount);
184 static void mptcp_rmem_uncharge(struct sock *sk, int size)
186 struct mptcp_sock *msk = mptcp_sk(sk);
189 msk->rmem_fwd_alloc += size;
190 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
192 /* see sk_mem_uncharge() for the rationale behind the following schema */
193 if (unlikely(reclaimable >= PAGE_SIZE))
194 __mptcp_rmem_reclaim(sk, reclaimable);
197 static void mptcp_rfree(struct sk_buff *skb)
199 unsigned int len = skb->truesize;
200 struct sock *sk = skb->sk;
202 atomic_sub(len, &sk->sk_rmem_alloc);
203 mptcp_rmem_uncharge(sk, len);
206 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
210 skb->destructor = mptcp_rfree;
211 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
212 mptcp_rmem_charge(sk, skb->truesize);
215 /* "inspired" by tcp_data_queue_ofo(), main differences:
217 * - don't cope with sacks
219 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
221 struct sock *sk = (struct sock *)msk;
222 struct rb_node **p, *parent;
223 u64 seq, end_seq, max_seq;
224 struct sk_buff *skb1;
226 seq = MPTCP_SKB_CB(skb)->map_seq;
227 end_seq = MPTCP_SKB_CB(skb)->end_seq;
228 max_seq = atomic64_read(&msk->rcv_wnd_sent);
230 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
231 RB_EMPTY_ROOT(&msk->out_of_order_queue));
232 if (after64(end_seq, max_seq)) {
235 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
236 (unsigned long long)end_seq - (unsigned long)max_seq,
237 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
238 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
242 p = &msk->out_of_order_queue.rb_node;
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
244 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
245 rb_link_node(&skb->rbnode, NULL, p);
246 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
247 msk->ooo_last_skb = skb;
251 /* with 2 subflows, adding at end of ooo queue is quite likely
252 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
254 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
255 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
256 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
260 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
261 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
263 parent = &msk->ooo_last_skb->rbnode;
264 p = &parent->rb_right;
268 /* Find place to insert this segment. Handle overlaps on the way. */
272 skb1 = rb_to_skb(parent);
273 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
274 p = &parent->rb_left;
277 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
278 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
279 /* All the bits are present. Drop. */
281 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
284 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
288 * continue traversing
291 /* skb's seq == skb1's seq and skb covers skb1.
292 * Replace skb1 with skb.
294 rb_replace_node(&skb1->rbnode, &skb->rbnode,
295 &msk->out_of_order_queue);
296 mptcp_drop(sk, skb1);
297 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
300 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
301 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
304 p = &parent->rb_right;
308 /* Insert segment into RB tree. */
309 rb_link_node(&skb->rbnode, parent, p);
310 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
313 /* Remove other segments covered by skb. */
314 while ((skb1 = skb_rb_next(skb)) != NULL) {
315 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
317 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
318 mptcp_drop(sk, skb1);
319 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
321 /* If there is no skb after us, we are the last_skb ! */
323 msk->ooo_last_skb = skb;
327 mptcp_set_owner_r(skb, sk);
330 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
332 struct mptcp_sock *msk = mptcp_sk(sk);
335 if (size <= msk->rmem_fwd_alloc)
338 size -= msk->rmem_fwd_alloc;
339 amt = sk_mem_pages(size);
340 amount = amt << PAGE_SHIFT;
341 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
344 msk->rmem_fwd_alloc += amount;
348 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
349 struct sk_buff *skb, unsigned int offset,
352 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
353 struct sock *sk = (struct sock *)msk;
354 struct sk_buff *tail;
357 __skb_unlink(skb, &ssk->sk_receive_queue);
362 /* try to fetch required memory from subflow */
363 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
366 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
368 /* the skb map_seq accounts for the skb offset:
369 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
372 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
373 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
374 MPTCP_SKB_CB(skb)->offset = offset;
375 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
377 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
379 msk->bytes_received += copy_len;
380 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
381 tail = skb_peek_tail(&sk->sk_receive_queue);
382 if (tail && mptcp_try_coalesce(sk, tail, skb))
385 mptcp_set_owner_r(skb, sk);
386 __skb_queue_tail(&sk->sk_receive_queue, skb);
388 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
389 mptcp_data_queue_ofo(msk, skb);
393 /* old data, keep it simple and drop the whole pkt, sender
394 * will retransmit as needed, if needed.
396 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
402 static void mptcp_stop_timer(struct sock *sk)
404 struct inet_connection_sock *icsk = inet_csk(sk);
406 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
407 mptcp_sk(sk)->timer_ival = 0;
410 static void mptcp_close_wake_up(struct sock *sk)
412 if (sock_flag(sk, SOCK_DEAD))
415 sk->sk_state_change(sk);
416 if (sk->sk_shutdown == SHUTDOWN_MASK ||
417 sk->sk_state == TCP_CLOSE)
418 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
420 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
423 static bool mptcp_pending_data_fin_ack(struct sock *sk)
425 struct mptcp_sock *msk = mptcp_sk(sk);
427 return ((1 << sk->sk_state) &
428 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
429 msk->write_seq == READ_ONCE(msk->snd_una);
432 static void mptcp_check_data_fin_ack(struct sock *sk)
434 struct mptcp_sock *msk = mptcp_sk(sk);
436 /* Look for an acknowledged DATA_FIN */
437 if (mptcp_pending_data_fin_ack(sk)) {
438 WRITE_ONCE(msk->snd_data_fin_enable, 0);
440 switch (sk->sk_state) {
442 inet_sk_state_store(sk, TCP_FIN_WAIT2);
446 inet_sk_state_store(sk, TCP_CLOSE);
450 mptcp_close_wake_up(sk);
454 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
456 struct mptcp_sock *msk = mptcp_sk(sk);
458 if (READ_ONCE(msk->rcv_data_fin) &&
459 ((1 << sk->sk_state) &
460 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
461 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
463 if (msk->ack_seq == rcv_data_fin_seq) {
465 *seq = rcv_data_fin_seq;
474 static void mptcp_set_datafin_timeout(struct sock *sk)
476 struct inet_connection_sock *icsk = inet_csk(sk);
479 retransmits = min_t(u32, icsk->icsk_retransmits,
480 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
482 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
485 static void __mptcp_set_timeout(struct sock *sk, long tout)
487 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
490 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
492 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
494 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
495 inet_csk(ssk)->icsk_timeout - jiffies : 0;
498 static void mptcp_set_timeout(struct sock *sk)
500 struct mptcp_subflow_context *subflow;
503 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
504 tout = max(tout, mptcp_timeout_from_subflow(subflow));
505 __mptcp_set_timeout(sk, tout);
508 static inline bool tcp_can_send_ack(const struct sock *ssk)
510 return !((1 << inet_sk_state_load(ssk)) &
511 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
514 void __mptcp_subflow_send_ack(struct sock *ssk)
516 if (tcp_can_send_ack(ssk))
520 static void mptcp_subflow_send_ack(struct sock *ssk)
524 slow = lock_sock_fast(ssk);
525 __mptcp_subflow_send_ack(ssk);
526 unlock_sock_fast(ssk, slow);
529 static void mptcp_send_ack(struct mptcp_sock *msk)
531 struct mptcp_subflow_context *subflow;
533 mptcp_for_each_subflow(msk, subflow)
534 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
537 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
541 slow = lock_sock_fast(ssk);
542 if (tcp_can_send_ack(ssk))
543 tcp_cleanup_rbuf(ssk, 1);
544 unlock_sock_fast(ssk, slow);
547 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
549 const struct inet_connection_sock *icsk = inet_csk(ssk);
550 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
551 const struct tcp_sock *tp = tcp_sk(ssk);
553 return (ack_pending & ICSK_ACK_SCHED) &&
554 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
555 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
556 (rx_empty && ack_pending &
557 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
560 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
562 int old_space = READ_ONCE(msk->old_wspace);
563 struct mptcp_subflow_context *subflow;
564 struct sock *sk = (struct sock *)msk;
565 int space = __mptcp_space(sk);
566 bool cleanup, rx_empty;
568 cleanup = (space > 0) && (space >= (old_space << 1));
569 rx_empty = !__mptcp_rmem(sk);
571 mptcp_for_each_subflow(msk, subflow) {
572 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
574 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
575 mptcp_subflow_cleanup_rbuf(ssk);
579 static bool mptcp_check_data_fin(struct sock *sk)
581 struct mptcp_sock *msk = mptcp_sk(sk);
582 u64 rcv_data_fin_seq;
585 /* Need to ack a DATA_FIN received from a peer while this side
586 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
587 * msk->rcv_data_fin was set when parsing the incoming options
588 * at the subflow level and the msk lock was not held, so this
589 * is the first opportunity to act on the DATA_FIN and change
592 * If we are caught up to the sequence number of the incoming
593 * DATA_FIN, send the DATA_ACK now and do state transition. If
594 * not caught up, do nothing and let the recv code send DATA_ACK
598 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
599 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
600 WRITE_ONCE(msk->rcv_data_fin, 0);
602 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
603 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
605 switch (sk->sk_state) {
606 case TCP_ESTABLISHED:
607 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
610 inet_sk_state_store(sk, TCP_CLOSING);
613 inet_sk_state_store(sk, TCP_CLOSE);
616 /* Other states not expected */
622 if (!__mptcp_check_fallback(msk))
624 mptcp_close_wake_up(sk);
629 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
633 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
634 struct sock *sk = (struct sock *)msk;
635 unsigned int moved = 0;
636 bool more_data_avail;
641 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
643 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
644 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
646 if (unlikely(ssk_rbuf > sk_rbuf)) {
647 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
652 pr_debug("msk=%p ssk=%p", msk, ssk);
655 u32 map_remaining, offset;
656 u32 seq = tp->copied_seq;
660 /* try to move as much data as available */
661 map_remaining = subflow->map_data_len -
662 mptcp_subflow_get_map_offset(subflow);
664 skb = skb_peek(&ssk->sk_receive_queue);
666 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
667 * a different CPU can have already processed the pending
668 * data, stop here or we can enter an infinite loop
675 if (__mptcp_check_fallback(msk)) {
676 /* Under fallback skbs have no MPTCP extension and TCP could
677 * collapse them between the dummy map creation and the
678 * current dequeue. Be sure to adjust the map size.
680 map_remaining = skb->len;
681 subflow->map_data_len = skb->len;
684 offset = seq - TCP_SKB_CB(skb)->seq;
685 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
691 if (offset < skb->len) {
692 size_t len = skb->len - offset;
697 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
701 if (WARN_ON_ONCE(map_remaining < len))
705 sk_eat_skb(ssk, skb);
709 WRITE_ONCE(tp->copied_seq, seq);
710 more_data_avail = mptcp_subflow_data_available(ssk);
712 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
716 } while (more_data_avail);
722 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
724 struct sock *sk = (struct sock *)msk;
725 struct sk_buff *skb, *tail;
730 p = rb_first(&msk->out_of_order_queue);
731 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
734 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
738 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
740 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
743 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
747 end_seq = MPTCP_SKB_CB(skb)->end_seq;
748 tail = skb_peek_tail(&sk->sk_receive_queue);
749 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
750 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
752 /* skip overlapping data, if any */
753 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
754 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
756 MPTCP_SKB_CB(skb)->offset += delta;
757 MPTCP_SKB_CB(skb)->map_seq += delta;
758 __skb_queue_tail(&sk->sk_receive_queue, skb);
760 msk->bytes_received += end_seq - msk->ack_seq;
761 msk->ack_seq = end_seq;
767 /* In most cases we will be able to lock the mptcp socket. If its already
768 * owned, we need to defer to the work queue to avoid ABBA deadlock.
770 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
772 struct sock *sk = (struct sock *)msk;
773 unsigned int moved = 0;
775 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
776 __mptcp_ofo_queue(msk);
777 if (unlikely(ssk->sk_err)) {
778 if (!sock_owned_by_user(sk))
779 __mptcp_error_report(sk);
781 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
784 /* If the moves have caught up with the DATA_FIN sequence number
785 * it's time to ack the DATA_FIN and change socket state, but
786 * this is not a good place to change state. Let the workqueue
789 if (mptcp_pending_data_fin(sk, NULL))
790 mptcp_schedule_work(sk);
794 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
796 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
797 struct mptcp_sock *msk = mptcp_sk(sk);
798 int sk_rbuf, ssk_rbuf;
800 /* The peer can send data while we are shutting down this
801 * subflow at msk destruction time, but we must avoid enqueuing
802 * more data to the msk receive queue
804 if (unlikely(subflow->disposable))
807 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
808 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
809 if (unlikely(ssk_rbuf > sk_rbuf))
812 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
813 if (__mptcp_rmem(sk) > sk_rbuf) {
814 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
818 /* Wake-up the reader only for in-sequence data */
820 if (move_skbs_to_msk(msk, ssk))
821 sk->sk_data_ready(sk);
823 mptcp_data_unlock(sk);
826 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
828 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
829 WRITE_ONCE(msk->allow_infinite_fallback, false);
830 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
833 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
835 struct sock *sk = (struct sock *)msk;
837 if (sk->sk_state != TCP_ESTABLISHED)
840 /* attach to msk socket only after we are sure we will deal with it
843 if (sk->sk_socket && !ssk->sk_socket)
844 mptcp_sock_graft(ssk, sk->sk_socket);
846 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
847 mptcp_sockopt_sync_locked(msk, ssk);
848 mptcp_subflow_joined(msk, ssk);
852 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
854 struct mptcp_subflow_context *tmp, *subflow;
855 struct mptcp_sock *msk = mptcp_sk(sk);
857 list_for_each_entry_safe(subflow, tmp, join_list, node) {
858 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
859 bool slow = lock_sock_fast(ssk);
861 list_move_tail(&subflow->node, &msk->conn_list);
862 if (!__mptcp_finish_join(msk, ssk))
863 mptcp_subflow_reset(ssk);
864 unlock_sock_fast(ssk, slow);
868 static bool mptcp_timer_pending(struct sock *sk)
870 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
873 static void mptcp_reset_timer(struct sock *sk)
875 struct inet_connection_sock *icsk = inet_csk(sk);
878 /* prevent rescheduling on close */
879 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
882 tout = mptcp_sk(sk)->timer_ival;
883 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
886 bool mptcp_schedule_work(struct sock *sk)
888 if (inet_sk_state_load(sk) != TCP_CLOSE &&
889 schedule_work(&mptcp_sk(sk)->work)) {
890 /* each subflow already holds a reference to the sk, and the
891 * workqueue is invoked by a subflow, so sk can't go away here.
899 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
901 struct mptcp_subflow_context *subflow;
903 msk_owned_by_me(msk);
905 mptcp_for_each_subflow(msk, subflow) {
906 if (READ_ONCE(subflow->data_avail))
907 return mptcp_subflow_tcp_sock(subflow);
913 static bool mptcp_skb_can_collapse_to(u64 write_seq,
914 const struct sk_buff *skb,
915 const struct mptcp_ext *mpext)
917 if (!tcp_skb_can_collapse_to(skb))
920 /* can collapse only if MPTCP level sequence is in order and this
921 * mapping has not been xmitted yet
923 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
927 /* we can append data to the given data frag if:
928 * - there is space available in the backing page_frag
929 * - the data frag tail matches the current page_frag free offset
930 * - the data frag end sequence number matches the current write seq
932 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
933 const struct page_frag *pfrag,
934 const struct mptcp_data_frag *df)
936 return df && pfrag->page == df->page &&
937 pfrag->size - pfrag->offset > 0 &&
938 pfrag->offset == (df->offset + df->data_len) &&
939 df->data_seq + df->data_len == msk->write_seq;
942 static void dfrag_uncharge(struct sock *sk, int len)
944 sk_mem_uncharge(sk, len);
945 sk_wmem_queued_add(sk, -len);
948 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
950 int len = dfrag->data_len + dfrag->overhead;
952 list_del(&dfrag->list);
953 dfrag_uncharge(sk, len);
954 put_page(dfrag->page);
957 static void __mptcp_clean_una(struct sock *sk)
959 struct mptcp_sock *msk = mptcp_sk(sk);
960 struct mptcp_data_frag *dtmp, *dfrag;
963 snd_una = msk->snd_una;
964 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
965 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
968 if (unlikely(dfrag == msk->first_pending)) {
969 /* in recovery mode can see ack after the current snd head */
970 if (WARN_ON_ONCE(!msk->recovery))
973 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
976 dfrag_clear(sk, dfrag);
979 dfrag = mptcp_rtx_head(sk);
980 if (dfrag && after64(snd_una, dfrag->data_seq)) {
981 u64 delta = snd_una - dfrag->data_seq;
983 /* prevent wrap around in recovery mode */
984 if (unlikely(delta > dfrag->already_sent)) {
985 if (WARN_ON_ONCE(!msk->recovery))
987 if (WARN_ON_ONCE(delta > dfrag->data_len))
989 dfrag->already_sent += delta - dfrag->already_sent;
992 dfrag->data_seq += delta;
993 dfrag->offset += delta;
994 dfrag->data_len -= delta;
995 dfrag->already_sent -= delta;
997 dfrag_uncharge(sk, delta);
1000 /* all retransmitted data acked, recovery completed */
1001 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1002 msk->recovery = false;
1005 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1006 snd_una == READ_ONCE(msk->write_seq)) {
1007 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1008 mptcp_stop_timer(sk);
1010 mptcp_reset_timer(sk);
1014 static void __mptcp_clean_una_wakeup(struct sock *sk)
1016 lockdep_assert_held_once(&sk->sk_lock.slock);
1018 __mptcp_clean_una(sk);
1019 mptcp_write_space(sk);
1022 static void mptcp_clean_una_wakeup(struct sock *sk)
1024 mptcp_data_lock(sk);
1025 __mptcp_clean_una_wakeup(sk);
1026 mptcp_data_unlock(sk);
1029 static void mptcp_enter_memory_pressure(struct sock *sk)
1031 struct mptcp_subflow_context *subflow;
1032 struct mptcp_sock *msk = mptcp_sk(sk);
1035 sk_stream_moderate_sndbuf(sk);
1036 mptcp_for_each_subflow(msk, subflow) {
1037 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1040 tcp_enter_memory_pressure(ssk);
1041 sk_stream_moderate_sndbuf(ssk);
1046 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1049 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1051 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1052 pfrag, sk->sk_allocation)))
1055 mptcp_enter_memory_pressure(sk);
1059 static struct mptcp_data_frag *
1060 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1063 int offset = ALIGN(orig_offset, sizeof(long));
1064 struct mptcp_data_frag *dfrag;
1066 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1067 dfrag->data_len = 0;
1068 dfrag->data_seq = msk->write_seq;
1069 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1070 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1071 dfrag->already_sent = 0;
1072 dfrag->page = pfrag->page;
1077 struct mptcp_sendmsg_info {
1083 bool data_lock_held;
1086 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1087 u64 data_seq, int avail_size)
1089 u64 window_end = mptcp_wnd_end(msk);
1092 if (__mptcp_check_fallback(msk))
1095 mptcp_snd_wnd = window_end - data_seq;
1096 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1098 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1099 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1100 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1106 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1108 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1112 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1116 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1118 struct sk_buff *skb;
1120 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1122 if (likely(__mptcp_add_ext(skb, gfp))) {
1123 skb_reserve(skb, MAX_TCP_HEADER);
1124 skb->ip_summed = CHECKSUM_PARTIAL;
1125 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1130 mptcp_enter_memory_pressure(sk);
1135 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1137 struct sk_buff *skb;
1139 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1143 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1144 tcp_skb_entail(ssk, skb);
1147 tcp_skb_tsorted_anchor_cleanup(skb);
1152 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1154 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1156 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1159 /* note: this always recompute the csum on the whole skb, even
1160 * if we just appended a single frag. More status info needed
1162 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1164 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1165 __wsum csum = ~csum_unfold(mpext->csum);
1166 int offset = skb->len - added;
1168 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1171 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1173 struct mptcp_ext *mpext)
1178 mpext->infinite_map = 1;
1179 mpext->data_len = 0;
1181 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1182 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1184 mptcp_do_fallback(ssk);
1187 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1188 struct mptcp_data_frag *dfrag,
1189 struct mptcp_sendmsg_info *info)
1191 u64 data_seq = dfrag->data_seq + info->sent;
1192 int offset = dfrag->offset + info->sent;
1193 struct mptcp_sock *msk = mptcp_sk(sk);
1194 bool zero_window_probe = false;
1195 struct mptcp_ext *mpext = NULL;
1196 bool can_coalesce = false;
1197 bool reuse_skb = true;
1198 struct sk_buff *skb;
1202 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1203 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1205 if (WARN_ON_ONCE(info->sent > info->limit ||
1206 info->limit > dfrag->data_len))
1209 if (unlikely(!__tcp_can_send(ssk)))
1212 /* compute send limit */
1213 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1214 copy = info->size_goal;
1216 skb = tcp_write_queue_tail(ssk);
1217 if (skb && copy > skb->len) {
1218 /* Limit the write to the size available in the
1219 * current skb, if any, so that we create at most a new skb.
1220 * Explicitly tells TCP internals to avoid collapsing on later
1221 * queue management operation, to avoid breaking the ext <->
1222 * SSN association set here
1224 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1225 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1226 TCP_SKB_CB(skb)->eor = 1;
1230 i = skb_shinfo(skb)->nr_frags;
1231 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1232 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1233 tcp_mark_push(tcp_sk(ssk), skb);
1240 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1244 i = skb_shinfo(skb)->nr_frags;
1246 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1249 /* Zero window and all data acked? Probe. */
1250 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1252 u64 snd_una = READ_ONCE(msk->snd_una);
1254 if (snd_una != msk->snd_nxt) {
1255 tcp_remove_empty_skb(ssk);
1259 zero_window_probe = true;
1260 data_seq = snd_una - 1;
1263 /* all mptcp-level data is acked, no skbs should be present into the
1266 WARN_ON_ONCE(reuse_skb);
1269 copy = min_t(size_t, copy, info->limit - info->sent);
1270 if (!sk_wmem_schedule(ssk, copy)) {
1271 tcp_remove_empty_skb(ssk);
1276 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1278 get_page(dfrag->page);
1279 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1283 skb->data_len += copy;
1284 skb->truesize += copy;
1285 sk_wmem_queued_add(ssk, copy);
1286 sk_mem_charge(ssk, copy);
1287 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1288 TCP_SKB_CB(skb)->end_seq += copy;
1289 tcp_skb_pcount_set(skb, 0);
1291 /* on skb reuse we just need to update the DSS len */
1293 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1294 mpext->data_len += copy;
1295 WARN_ON_ONCE(zero_window_probe);
1299 memset(mpext, 0, sizeof(*mpext));
1300 mpext->data_seq = data_seq;
1301 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1302 mpext->data_len = copy;
1306 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1307 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1310 if (zero_window_probe) {
1311 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1313 if (READ_ONCE(msk->csum_enabled))
1314 mptcp_update_data_checksum(skb, copy);
1315 tcp_push_pending_frames(ssk);
1319 if (READ_ONCE(msk->csum_enabled))
1320 mptcp_update_data_checksum(skb, copy);
1321 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1322 mptcp_update_infinite_map(msk, ssk, mpext);
1323 trace_mptcp_sendmsg_frag(mpext);
1324 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1328 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1329 sizeof(struct tcphdr) - \
1330 MAX_TCP_OPTION_SPACE - \
1331 sizeof(struct ipv6hdr) - \
1332 sizeof(struct frag_hdr))
1334 struct subflow_send_info {
1339 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1341 if (!subflow->stale)
1345 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1348 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1350 if (unlikely(subflow->stale)) {
1351 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1353 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1356 mptcp_subflow_set_active(subflow);
1358 return __mptcp_subflow_active(subflow);
1361 #define SSK_MODE_ACTIVE 0
1362 #define SSK_MODE_BACKUP 1
1363 #define SSK_MODE_MAX 2
1365 /* implement the mptcp packet scheduler;
1366 * returns the subflow that will transmit the next DSS
1367 * additionally updates the rtx timeout
1369 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1371 struct subflow_send_info send_info[SSK_MODE_MAX];
1372 struct mptcp_subflow_context *subflow;
1373 struct sock *sk = (struct sock *)msk;
1374 u32 pace, burst, wmem;
1375 int i, nr_active = 0;
1380 msk_owned_by_me(msk);
1382 if (__mptcp_check_fallback(msk)) {
1385 return __tcp_can_send(msk->first) &&
1386 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1389 /* re-use last subflow, if the burst allow that */
1390 if (msk->last_snd && msk->snd_burst > 0 &&
1391 sk_stream_memory_free(msk->last_snd) &&
1392 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1393 mptcp_set_timeout(sk);
1394 return msk->last_snd;
1397 /* pick the subflow with the lower wmem/wspace ratio */
1398 for (i = 0; i < SSK_MODE_MAX; ++i) {
1399 send_info[i].ssk = NULL;
1400 send_info[i].linger_time = -1;
1403 mptcp_for_each_subflow(msk, subflow) {
1404 trace_mptcp_subflow_get_send(subflow);
1405 ssk = mptcp_subflow_tcp_sock(subflow);
1406 if (!mptcp_subflow_active(subflow))
1409 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1410 nr_active += !subflow->backup;
1411 pace = subflow->avg_pacing_rate;
1412 if (unlikely(!pace)) {
1413 /* init pacing rate from socket */
1414 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1415 pace = subflow->avg_pacing_rate;
1420 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1421 if (linger_time < send_info[subflow->backup].linger_time) {
1422 send_info[subflow->backup].ssk = ssk;
1423 send_info[subflow->backup].linger_time = linger_time;
1426 __mptcp_set_timeout(sk, tout);
1428 /* pick the best backup if no other subflow is active */
1430 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1432 /* According to the blest algorithm, to avoid HoL blocking for the
1433 * faster flow, we need to:
1434 * - estimate the faster flow linger time
1435 * - use the above to estimate the amount of byte transferred
1436 * by the faster flow
1437 * - check that the amount of queued data is greter than the above,
1438 * otherwise do not use the picked, slower, subflow
1439 * We select the subflow with the shorter estimated time to flush
1440 * the queued mem, which basically ensure the above. We just need
1441 * to check that subflow has a non empty cwin.
1443 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1444 if (!ssk || !sk_stream_memory_free(ssk))
1447 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1448 wmem = READ_ONCE(ssk->sk_wmem_queued);
1450 msk->last_snd = NULL;
1454 subflow = mptcp_subflow_ctx(ssk);
1455 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1456 READ_ONCE(ssk->sk_pacing_rate) * burst,
1458 msk->last_snd = ssk;
1459 msk->snd_burst = burst;
1463 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1465 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1469 static void mptcp_update_post_push(struct mptcp_sock *msk,
1470 struct mptcp_data_frag *dfrag,
1473 u64 snd_nxt_new = dfrag->data_seq;
1475 dfrag->already_sent += sent;
1477 msk->snd_burst -= sent;
1479 snd_nxt_new += dfrag->already_sent;
1481 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1482 * is recovering after a failover. In that event, this re-sends
1485 * Thus compute snd_nxt_new candidate based on
1486 * the dfrag->data_seq that was sent and the data
1487 * that has been handed to the subflow for transmission
1488 * and skip update in case it was old dfrag.
1490 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1491 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1492 msk->snd_nxt = snd_nxt_new;
1496 void mptcp_check_and_set_pending(struct sock *sk)
1498 if (mptcp_send_head(sk))
1499 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1502 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1504 struct sock *prev_ssk = NULL, *ssk = NULL;
1505 struct mptcp_sock *msk = mptcp_sk(sk);
1506 struct mptcp_sendmsg_info info = {
1509 bool do_check_data_fin = false;
1510 struct mptcp_data_frag *dfrag;
1513 while ((dfrag = mptcp_send_head(sk))) {
1514 info.sent = dfrag->already_sent;
1515 info.limit = dfrag->data_len;
1516 len = dfrag->data_len - dfrag->already_sent;
1521 ssk = mptcp_subflow_get_send(msk);
1523 /* First check. If the ssk has changed since
1524 * the last round, release prev_ssk
1526 if (ssk != prev_ssk && prev_ssk)
1527 mptcp_push_release(prev_ssk, &info);
1531 /* Need to lock the new subflow only if different
1532 * from the previous one, otherwise we are still
1533 * helding the relevant lock
1535 if (ssk != prev_ssk)
1538 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1542 mptcp_push_release(ssk, &info);
1546 do_check_data_fin = true;
1550 mptcp_update_post_push(msk, dfrag, ret);
1552 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1555 /* at this point we held the socket lock for the last subflow we used */
1557 mptcp_push_release(ssk, &info);
1560 /* ensure the rtx timer is running */
1561 if (!mptcp_timer_pending(sk))
1562 mptcp_reset_timer(sk);
1563 if (do_check_data_fin)
1564 mptcp_check_send_data_fin(sk);
1567 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1569 struct mptcp_sock *msk = mptcp_sk(sk);
1570 struct mptcp_sendmsg_info info = {
1571 .data_lock_held = true,
1573 struct mptcp_data_frag *dfrag;
1574 struct sock *xmit_ssk;
1575 int len, copied = 0;
1578 while ((dfrag = mptcp_send_head(sk))) {
1579 info.sent = dfrag->already_sent;
1580 info.limit = dfrag->data_len;
1581 len = dfrag->data_len - dfrag->already_sent;
1585 /* check for a different subflow usage only after
1586 * spooling the first chunk of data
1588 xmit_ssk = first ? ssk : mptcp_subflow_get_send(msk);
1591 if (xmit_ssk != ssk) {
1592 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1593 MPTCP_DELEGATE_SEND);
1597 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1606 mptcp_update_post_push(msk, dfrag, ret);
1608 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1612 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1613 * not going to flush it via release_sock()
1616 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1618 if (!mptcp_timer_pending(sk))
1619 mptcp_reset_timer(sk);
1621 if (msk->snd_data_fin_enable &&
1622 msk->snd_nxt + 1 == msk->write_seq)
1623 mptcp_schedule_work(sk);
1627 static void mptcp_set_nospace(struct sock *sk)
1629 /* enable autotune */
1630 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1632 /* will be cleared on avail space */
1633 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1636 static int mptcp_disconnect(struct sock *sk, int flags);
1638 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1639 size_t len, int *copied_syn)
1641 unsigned int saved_flags = msg->msg_flags;
1642 struct mptcp_sock *msk = mptcp_sk(sk);
1646 /* on flags based fastopen the mptcp is supposed to create the
1647 * first subflow right now. Otherwise we are in the defer_connect
1648 * path, and the first subflow must be already present.
1649 * Since the defer_connect flag is cleared after the first succsful
1650 * fastopen attempt, no need to check for additional subflow status.
1652 if (msg->msg_flags & MSG_FASTOPEN) {
1653 ssk = __mptcp_nmpc_sk(msk);
1655 return PTR_ERR(ssk);
1663 msg->msg_flags |= MSG_DONTWAIT;
1664 msk->fastopening = 1;
1665 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1666 msk->fastopening = 0;
1667 msg->msg_flags = saved_flags;
1670 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1671 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1672 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1673 msg->msg_namelen, msg->msg_flags, 1);
1675 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1676 * case of any error, except timeout or signal
1678 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1680 } else if (ret && ret != -EINPROGRESS) {
1681 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1682 * __inet_stream_connect() can fail, due to looking check,
1683 * see mptcp_disconnect().
1684 * Attempt it again outside the problematic scope.
1686 if (!mptcp_disconnect(sk, 0))
1687 sk->sk_socket->state = SS_UNCONNECTED;
1689 inet_sk(sk)->defer_connect = 0;
1694 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1696 struct mptcp_sock *msk = mptcp_sk(sk);
1697 struct page_frag *pfrag;
1702 /* silently ignore everything else */
1703 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1707 if (unlikely(inet_sk(sk)->defer_connect || msg->msg_flags & MSG_FASTOPEN)) {
1710 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1711 copied += copied_syn;
1712 if (ret == -EINPROGRESS && copied_syn > 0)
1718 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1720 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1721 ret = sk_stream_wait_connect(sk, &timeo);
1727 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1730 pfrag = sk_page_frag(sk);
1732 while (msg_data_left(msg)) {
1733 int total_ts, frag_truesize = 0;
1734 struct mptcp_data_frag *dfrag;
1735 bool dfrag_collapsed;
1736 size_t psize, offset;
1738 /* reuse tail pfrag, if possible, or carve a new one from the
1741 dfrag = mptcp_pending_tail(sk);
1742 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1743 if (!dfrag_collapsed) {
1744 if (!sk_stream_memory_free(sk))
1745 goto wait_for_memory;
1747 if (!mptcp_page_frag_refill(sk, pfrag))
1748 goto wait_for_memory;
1750 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1751 frag_truesize = dfrag->overhead;
1754 /* we do not bound vs wspace, to allow a single packet.
1755 * memory accounting will prevent execessive memory usage
1758 offset = dfrag->offset + dfrag->data_len;
1759 psize = pfrag->size - offset;
1760 psize = min_t(size_t, psize, msg_data_left(msg));
1761 total_ts = psize + frag_truesize;
1763 if (!sk_wmem_schedule(sk, total_ts))
1764 goto wait_for_memory;
1766 if (copy_page_from_iter(dfrag->page, offset, psize,
1767 &msg->msg_iter) != psize) {
1772 /* data successfully copied into the write queue */
1773 sk->sk_forward_alloc -= total_ts;
1775 dfrag->data_len += psize;
1776 frag_truesize += psize;
1777 pfrag->offset += frag_truesize;
1778 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1780 /* charge data on mptcp pending queue to the msk socket
1781 * Note: we charge such data both to sk and ssk
1783 sk_wmem_queued_add(sk, frag_truesize);
1784 if (!dfrag_collapsed) {
1785 get_page(dfrag->page);
1786 list_add_tail(&dfrag->list, &msk->rtx_queue);
1787 if (!msk->first_pending)
1788 WRITE_ONCE(msk->first_pending, dfrag);
1790 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1791 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1797 mptcp_set_nospace(sk);
1798 __mptcp_push_pending(sk, msg->msg_flags);
1799 ret = sk_stream_wait_memory(sk, &timeo);
1805 __mptcp_push_pending(sk, msg->msg_flags);
1815 copied = sk_stream_error(sk, msg->msg_flags, ret);
1819 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1821 size_t len, int flags,
1822 struct scm_timestamping_internal *tss,
1825 struct sk_buff *skb, *tmp;
1828 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1829 u32 offset = MPTCP_SKB_CB(skb)->offset;
1830 u32 data_len = skb->len - offset;
1831 u32 count = min_t(size_t, len - copied, data_len);
1834 if (!(flags & MSG_TRUNC)) {
1835 err = skb_copy_datagram_msg(skb, offset, msg, count);
1836 if (unlikely(err < 0)) {
1843 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1844 tcp_update_recv_tstamps(skb, tss);
1845 *cmsg_flags |= MPTCP_CMSG_TS;
1850 if (count < data_len) {
1851 if (!(flags & MSG_PEEK)) {
1852 MPTCP_SKB_CB(skb)->offset += count;
1853 MPTCP_SKB_CB(skb)->map_seq += count;
1858 if (!(flags & MSG_PEEK)) {
1859 /* we will bulk release the skb memory later */
1860 skb->destructor = NULL;
1861 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1862 __skb_unlink(skb, &msk->receive_queue);
1873 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1875 * Only difference: Use highest rtt estimate of the subflows in use.
1877 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1879 struct mptcp_subflow_context *subflow;
1880 struct sock *sk = (struct sock *)msk;
1881 u8 scaling_ratio = U8_MAX;
1882 u32 time, advmss = 1;
1885 msk_owned_by_me(msk);
1890 msk->rcvq_space.copied += copied;
1892 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1893 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1895 rtt_us = msk->rcvq_space.rtt_us;
1896 if (rtt_us && time < (rtt_us >> 3))
1900 mptcp_for_each_subflow(msk, subflow) {
1901 const struct tcp_sock *tp;
1905 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1907 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1908 sf_advmss = READ_ONCE(tp->advmss);
1910 rtt_us = max(sf_rtt_us, rtt_us);
1911 advmss = max(sf_advmss, advmss);
1912 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1915 msk->rcvq_space.rtt_us = rtt_us;
1916 msk->scaling_ratio = scaling_ratio;
1917 if (time < (rtt_us >> 3) || rtt_us == 0)
1920 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1923 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1924 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1928 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1930 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1932 do_div(grow, msk->rcvq_space.space);
1933 rcvwin += (grow << 1);
1935 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
1936 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1938 if (rcvbuf > sk->sk_rcvbuf) {
1941 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
1942 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1944 /* Make subflows follow along. If we do not do this, we
1945 * get drops at subflow level if skbs can't be moved to
1946 * the mptcp rx queue fast enough (announced rcv_win can
1947 * exceed ssk->sk_rcvbuf).
1949 mptcp_for_each_subflow(msk, subflow) {
1953 ssk = mptcp_subflow_tcp_sock(subflow);
1954 slow = lock_sock_fast(ssk);
1955 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1956 tcp_sk(ssk)->window_clamp = window_clamp;
1957 tcp_cleanup_rbuf(ssk, 1);
1958 unlock_sock_fast(ssk, slow);
1963 msk->rcvq_space.space = msk->rcvq_space.copied;
1965 msk->rcvq_space.copied = 0;
1966 msk->rcvq_space.time = mstamp;
1969 static void __mptcp_update_rmem(struct sock *sk)
1971 struct mptcp_sock *msk = mptcp_sk(sk);
1973 if (!msk->rmem_released)
1976 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1977 mptcp_rmem_uncharge(sk, msk->rmem_released);
1978 WRITE_ONCE(msk->rmem_released, 0);
1981 static void __mptcp_splice_receive_queue(struct sock *sk)
1983 struct mptcp_sock *msk = mptcp_sk(sk);
1985 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1988 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1990 struct sock *sk = (struct sock *)msk;
1991 unsigned int moved = 0;
1995 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1998 /* we can have data pending in the subflows only if the msk
1999 * receive buffer was full at subflow_data_ready() time,
2000 * that is an unlikely slow path.
2005 slowpath = lock_sock_fast(ssk);
2006 mptcp_data_lock(sk);
2007 __mptcp_update_rmem(sk);
2008 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2009 mptcp_data_unlock(sk);
2011 if (unlikely(ssk->sk_err))
2012 __mptcp_error_report(sk);
2013 unlock_sock_fast(ssk, slowpath);
2016 /* acquire the data lock only if some input data is pending */
2018 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2019 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2020 mptcp_data_lock(sk);
2021 __mptcp_update_rmem(sk);
2022 ret |= __mptcp_ofo_queue(msk);
2023 __mptcp_splice_receive_queue(sk);
2024 mptcp_data_unlock(sk);
2027 mptcp_check_data_fin((struct sock *)msk);
2028 return !skb_queue_empty(&msk->receive_queue);
2031 static unsigned int mptcp_inq_hint(const struct sock *sk)
2033 const struct mptcp_sock *msk = mptcp_sk(sk);
2034 const struct sk_buff *skb;
2036 skb = skb_peek(&msk->receive_queue);
2038 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2040 if (hint_val >= INT_MAX)
2043 return (unsigned int)hint_val;
2046 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2052 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2053 int flags, int *addr_len)
2055 struct mptcp_sock *msk = mptcp_sk(sk);
2056 struct scm_timestamping_internal tss;
2057 int copied = 0, cmsg_flags = 0;
2061 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2062 if (unlikely(flags & MSG_ERRQUEUE))
2063 return inet_recv_error(sk, msg, len, addr_len);
2066 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2071 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2073 len = min_t(size_t, len, INT_MAX);
2074 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2076 if (unlikely(msk->recvmsg_inq))
2077 cmsg_flags = MPTCP_CMSG_INQ;
2079 while (copied < len) {
2082 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2083 if (unlikely(bytes_read < 0)) {
2085 copied = bytes_read;
2089 copied += bytes_read;
2091 /* be sure to advertise window change */
2092 mptcp_cleanup_rbuf(msk);
2094 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2097 /* only the master socket status is relevant here. The exit
2098 * conditions mirror closely tcp_recvmsg()
2100 if (copied >= target)
2105 sk->sk_state == TCP_CLOSE ||
2106 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2108 signal_pending(current))
2112 copied = sock_error(sk);
2116 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2117 /* race breaker: the shutdown could be after the
2118 * previous receive queue check
2120 if (__mptcp_move_skbs(msk))
2125 if (sk->sk_state == TCP_CLOSE) {
2135 if (signal_pending(current)) {
2136 copied = sock_intr_errno(timeo);
2141 pr_debug("block timeout %ld", timeo);
2142 sk_wait_data(sk, &timeo, NULL);
2146 if (cmsg_flags && copied >= 0) {
2147 if (cmsg_flags & MPTCP_CMSG_TS)
2148 tcp_recv_timestamp(msg, sk, &tss);
2150 if (cmsg_flags & MPTCP_CMSG_INQ) {
2151 unsigned int inq = mptcp_inq_hint(sk);
2153 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2157 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2158 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2159 skb_queue_empty(&msk->receive_queue), copied);
2160 if (!(flags & MSG_PEEK))
2161 mptcp_rcv_space_adjust(msk, copied);
2167 static void mptcp_retransmit_timer(struct timer_list *t)
2169 struct inet_connection_sock *icsk = from_timer(icsk, t,
2170 icsk_retransmit_timer);
2171 struct sock *sk = &icsk->icsk_inet.sk;
2172 struct mptcp_sock *msk = mptcp_sk(sk);
2175 if (!sock_owned_by_user(sk)) {
2176 /* we need a process context to retransmit */
2177 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2178 mptcp_schedule_work(sk);
2180 /* delegate our work to tcp_release_cb() */
2181 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2187 static void mptcp_timeout_timer(struct timer_list *t)
2189 struct sock *sk = from_timer(sk, t, sk_timer);
2191 mptcp_schedule_work(sk);
2195 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2198 * A backup subflow is returned only if that is the only kind available.
2200 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2202 struct sock *backup = NULL, *pick = NULL;
2203 struct mptcp_subflow_context *subflow;
2204 int min_stale_count = INT_MAX;
2206 msk_owned_by_me(msk);
2208 if (__mptcp_check_fallback(msk))
2211 mptcp_for_each_subflow(msk, subflow) {
2212 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2214 if (!__mptcp_subflow_active(subflow))
2217 /* still data outstanding at TCP level? skip this */
2218 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2219 mptcp_pm_subflow_chk_stale(msk, ssk);
2220 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2224 if (subflow->backup) {
2237 /* use backup only if there are no progresses anywhere */
2238 return min_stale_count > 1 ? backup : NULL;
2241 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2244 iput(SOCK_INODE(msk->subflow));
2245 WRITE_ONCE(msk->subflow, NULL);
2249 bool __mptcp_retransmit_pending_data(struct sock *sk)
2251 struct mptcp_data_frag *cur, *rtx_head;
2252 struct mptcp_sock *msk = mptcp_sk(sk);
2254 if (__mptcp_check_fallback(msk))
2257 if (tcp_rtx_and_write_queues_empty(sk))
2260 /* the closing socket has some data untransmitted and/or unacked:
2261 * some data in the mptcp rtx queue has not really xmitted yet.
2262 * keep it simple and re-inject the whole mptcp level rtx queue
2264 mptcp_data_lock(sk);
2265 __mptcp_clean_una_wakeup(sk);
2266 rtx_head = mptcp_rtx_head(sk);
2268 mptcp_data_unlock(sk);
2272 msk->recovery_snd_nxt = msk->snd_nxt;
2273 msk->recovery = true;
2274 mptcp_data_unlock(sk);
2276 msk->first_pending = rtx_head;
2279 /* be sure to clear the "sent status" on all re-injected fragments */
2280 list_for_each_entry(cur, &msk->rtx_queue, list) {
2281 if (!cur->already_sent)
2283 cur->already_sent = 0;
2289 /* flags for __mptcp_close_ssk() */
2290 #define MPTCP_CF_PUSH BIT(1)
2291 #define MPTCP_CF_FASTCLOSE BIT(2)
2293 /* subflow sockets can be either outgoing (connect) or incoming
2296 * Outgoing subflows use in-kernel sockets.
2297 * Incoming subflows do not have their own 'struct socket' allocated,
2298 * so we need to use tcp_close() after detaching them from the mptcp
2301 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2302 struct mptcp_subflow_context *subflow,
2305 struct mptcp_sock *msk = mptcp_sk(sk);
2306 bool dispose_it, need_push = false;
2308 /* If the first subflow moved to a close state before accept, e.g. due
2309 * to an incoming reset, mptcp either:
2310 * - if either the subflow or the msk are dead, destroy the context
2311 * (the subflow socket is deleted by inet_child_forget) and the msk
2312 * - otherwise do nothing at the moment and take action at accept and/or
2313 * listener shutdown - user-space must be able to accept() the closed
2316 if (msk->in_accept_queue && msk->first == ssk) {
2317 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2320 /* ensure later check in mptcp_worker() will dispose the msk */
2321 sock_set_flag(sk, SOCK_DEAD);
2322 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2323 mptcp_subflow_drop_ctx(ssk);
2327 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2329 list_del(&subflow->node);
2331 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2333 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2334 /* be sure to force the tcp_disconnect() path,
2335 * to generate the egress reset
2337 ssk->sk_lingertime = 0;
2338 sock_set_flag(ssk, SOCK_LINGER);
2339 subflow->send_fastclose = 1;
2342 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2344 /* The MPTCP code never wait on the subflow sockets, TCP-level
2345 * disconnect should never fail
2347 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2348 msk->subflow->state = SS_UNCONNECTED;
2349 mptcp_subflow_ctx_reset(subflow);
2355 subflow->disposable = 1;
2357 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2358 * the ssk has been already destroyed, we just need to release the
2359 * reference owned by msk;
2361 if (!inet_csk(ssk)->icsk_ulp_ops) {
2362 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2363 kfree_rcu(subflow, rcu);
2365 /* otherwise tcp will dispose of the ssk and subflow ctx */
2366 __tcp_close(ssk, 0);
2368 /* close acquired an extra ref */
2377 if (ssk == msk->first)
2378 WRITE_ONCE(msk->first, NULL);
2381 if (ssk == msk->last_snd)
2382 msk->last_snd = NULL;
2385 __mptcp_push_pending(sk, 0);
2388 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2389 struct mptcp_subflow_context *subflow)
2391 if (sk->sk_state == TCP_ESTABLISHED)
2392 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2394 /* subflow aborted before reaching the fully_established status
2395 * attempt the creation of the next subflow
2397 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2399 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2402 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2407 static void __mptcp_close_subflow(struct sock *sk)
2409 struct mptcp_subflow_context *subflow, *tmp;
2410 struct mptcp_sock *msk = mptcp_sk(sk);
2414 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2415 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2417 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2420 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2421 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2424 mptcp_close_ssk(sk, ssk, subflow);
2429 static bool mptcp_should_close(const struct sock *sk)
2431 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2432 struct mptcp_subflow_context *subflow;
2434 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2437 /* if all subflows are in closed status don't bother with additional
2440 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2441 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2448 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2450 struct mptcp_subflow_context *subflow, *tmp;
2451 struct sock *sk = (struct sock *)msk;
2453 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2456 mptcp_token_destroy(msk);
2458 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2459 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2462 slow = lock_sock_fast(tcp_sk);
2463 if (tcp_sk->sk_state != TCP_CLOSE) {
2464 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2465 tcp_set_state(tcp_sk, TCP_CLOSE);
2467 unlock_sock_fast(tcp_sk, slow);
2470 /* Mirror the tcp_reset() error propagation */
2471 switch (sk->sk_state) {
2473 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2475 case TCP_CLOSE_WAIT:
2476 WRITE_ONCE(sk->sk_err, EPIPE);
2481 WRITE_ONCE(sk->sk_err, ECONNRESET);
2484 inet_sk_state_store(sk, TCP_CLOSE);
2485 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2486 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2487 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2489 /* the calling mptcp_worker will properly destroy the socket */
2490 if (sock_flag(sk, SOCK_DEAD))
2493 sk->sk_state_change(sk);
2494 sk_error_report(sk);
2497 static void __mptcp_retrans(struct sock *sk)
2499 struct mptcp_sock *msk = mptcp_sk(sk);
2500 struct mptcp_sendmsg_info info = {};
2501 struct mptcp_data_frag *dfrag;
2506 mptcp_clean_una_wakeup(sk);
2508 /* first check ssk: need to kick "stale" logic */
2509 ssk = mptcp_subflow_get_retrans(msk);
2510 dfrag = mptcp_rtx_head(sk);
2512 if (mptcp_data_fin_enabled(msk)) {
2513 struct inet_connection_sock *icsk = inet_csk(sk);
2515 icsk->icsk_retransmits++;
2516 mptcp_set_datafin_timeout(sk);
2517 mptcp_send_ack(msk);
2522 if (!mptcp_send_head(sk))
2533 /* limit retransmission to the bytes already sent on some subflows */
2535 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2536 while (info.sent < info.limit) {
2537 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2541 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2546 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2547 msk->bytes_retrans += copied;
2548 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2550 WRITE_ONCE(msk->allow_infinite_fallback, false);
2556 mptcp_check_and_set_pending(sk);
2558 if (!mptcp_timer_pending(sk))
2559 mptcp_reset_timer(sk);
2562 /* schedule the timeout timer for the relevant event: either close timeout
2563 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2565 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2567 struct sock *sk = (struct sock *)msk;
2568 unsigned long timeout, close_timeout;
2570 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2573 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2575 /* the close timeout takes precedence on the fail one, and here at least one of
2578 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2580 sk_reset_timer(sk, &sk->sk_timer, timeout);
2583 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2585 struct sock *ssk = msk->first;
2591 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2593 slow = lock_sock_fast(ssk);
2594 mptcp_subflow_reset(ssk);
2595 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2596 unlock_sock_fast(ssk, slow);
2598 mptcp_reset_timeout(msk, 0);
2601 static void mptcp_do_fastclose(struct sock *sk)
2603 struct mptcp_subflow_context *subflow, *tmp;
2604 struct mptcp_sock *msk = mptcp_sk(sk);
2606 inet_sk_state_store(sk, TCP_CLOSE);
2607 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2608 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2609 subflow, MPTCP_CF_FASTCLOSE);
2612 static void mptcp_worker(struct work_struct *work)
2614 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2615 struct sock *sk = (struct sock *)msk;
2616 unsigned long fail_tout;
2620 state = sk->sk_state;
2621 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2624 mptcp_check_fastclose(msk);
2626 mptcp_pm_nl_work(msk);
2628 mptcp_check_send_data_fin(sk);
2629 mptcp_check_data_fin_ack(sk);
2630 mptcp_check_data_fin(sk);
2632 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2633 __mptcp_close_subflow(sk);
2635 /* There is no point in keeping around an orphaned sk timedout or
2636 * closed, but we need the msk around to reply to incoming DATA_FIN,
2637 * even if it is orphaned and in FIN_WAIT2 state
2639 if (sock_flag(sk, SOCK_DEAD)) {
2640 if (mptcp_should_close(sk))
2641 mptcp_do_fastclose(sk);
2643 if (sk->sk_state == TCP_CLOSE) {
2644 __mptcp_destroy_sock(sk);
2649 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2650 __mptcp_retrans(sk);
2652 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2653 if (fail_tout && time_after(jiffies, fail_tout))
2654 mptcp_mp_fail_no_response(msk);
2661 static int __mptcp_init_sock(struct sock *sk)
2663 struct mptcp_sock *msk = mptcp_sk(sk);
2665 INIT_LIST_HEAD(&msk->conn_list);
2666 INIT_LIST_HEAD(&msk->join_list);
2667 INIT_LIST_HEAD(&msk->rtx_queue);
2668 INIT_WORK(&msk->work, mptcp_worker);
2669 __skb_queue_head_init(&msk->receive_queue);
2670 msk->out_of_order_queue = RB_ROOT;
2671 msk->first_pending = NULL;
2672 msk->rmem_fwd_alloc = 0;
2673 WRITE_ONCE(msk->rmem_released, 0);
2674 msk->timer_ival = TCP_RTO_MIN;
2676 WRITE_ONCE(msk->first, NULL);
2677 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2678 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2679 WRITE_ONCE(msk->allow_infinite_fallback, true);
2680 msk->recovery = false;
2681 msk->subflow_id = 1;
2683 mptcp_pm_data_init(msk);
2685 /* re-use the csk retrans timer for MPTCP-level retrans */
2686 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2687 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2692 static void mptcp_ca_reset(struct sock *sk)
2694 struct inet_connection_sock *icsk = inet_csk(sk);
2696 tcp_assign_congestion_control(sk);
2697 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2699 /* no need to keep a reference to the ops, the name will suffice */
2700 tcp_cleanup_congestion_control(sk);
2701 icsk->icsk_ca_ops = NULL;
2704 static int mptcp_init_sock(struct sock *sk)
2706 struct net *net = sock_net(sk);
2709 ret = __mptcp_init_sock(sk);
2713 if (!mptcp_is_enabled(net))
2714 return -ENOPROTOOPT;
2716 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2719 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2721 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2722 * propagate the correct value
2726 sk_sockets_allocated_inc(sk);
2727 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2728 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2733 static void __mptcp_clear_xmit(struct sock *sk)
2735 struct mptcp_sock *msk = mptcp_sk(sk);
2736 struct mptcp_data_frag *dtmp, *dfrag;
2738 WRITE_ONCE(msk->first_pending, NULL);
2739 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2740 dfrag_clear(sk, dfrag);
2743 void mptcp_cancel_work(struct sock *sk)
2745 struct mptcp_sock *msk = mptcp_sk(sk);
2747 if (cancel_work_sync(&msk->work))
2751 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2755 switch (ssk->sk_state) {
2757 if (!(how & RCV_SHUTDOWN))
2761 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2764 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2765 pr_debug("Fallback");
2766 ssk->sk_shutdown |= how;
2767 tcp_shutdown(ssk, how);
2769 /* simulate the data_fin ack reception to let the state
2770 * machine move forward
2772 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2773 mptcp_schedule_work(sk);
2775 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2777 if (!mptcp_timer_pending(sk))
2778 mptcp_reset_timer(sk);
2786 static const unsigned char new_state[16] = {
2787 /* current state: new state: action: */
2788 [0 /* (Invalid) */] = TCP_CLOSE,
2789 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2790 [TCP_SYN_SENT] = TCP_CLOSE,
2791 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2792 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2793 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2794 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2795 [TCP_CLOSE] = TCP_CLOSE,
2796 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2797 [TCP_LAST_ACK] = TCP_LAST_ACK,
2798 [TCP_LISTEN] = TCP_CLOSE,
2799 [TCP_CLOSING] = TCP_CLOSING,
2800 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2803 static int mptcp_close_state(struct sock *sk)
2805 int next = (int)new_state[sk->sk_state];
2806 int ns = next & TCP_STATE_MASK;
2808 inet_sk_state_store(sk, ns);
2810 return next & TCP_ACTION_FIN;
2813 static void mptcp_check_send_data_fin(struct sock *sk)
2815 struct mptcp_subflow_context *subflow;
2816 struct mptcp_sock *msk = mptcp_sk(sk);
2818 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2819 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2820 msk->snd_nxt, msk->write_seq);
2822 /* we still need to enqueue subflows or not really shutting down,
2825 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2826 mptcp_send_head(sk))
2829 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2831 mptcp_for_each_subflow(msk, subflow) {
2832 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2834 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2838 static void __mptcp_wr_shutdown(struct sock *sk)
2840 struct mptcp_sock *msk = mptcp_sk(sk);
2842 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2843 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2844 !!mptcp_send_head(sk));
2846 /* will be ignored by fallback sockets */
2847 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2848 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2850 mptcp_check_send_data_fin(sk);
2853 static void __mptcp_destroy_sock(struct sock *sk)
2855 struct mptcp_sock *msk = mptcp_sk(sk);
2857 pr_debug("msk=%p", msk);
2861 mptcp_stop_timer(sk);
2862 sk_stop_timer(sk, &sk->sk_timer);
2865 sk->sk_prot->destroy(sk);
2867 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2868 WARN_ON_ONCE(msk->rmem_released);
2869 sk_stream_kill_queues(sk);
2870 xfrm_sk_free_policy(sk);
2875 void __mptcp_unaccepted_force_close(struct sock *sk)
2877 sock_set_flag(sk, SOCK_DEAD);
2878 mptcp_do_fastclose(sk);
2879 __mptcp_destroy_sock(sk);
2882 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2884 /* Concurrent splices from sk_receive_queue into receive_queue will
2885 * always show at least one non-empty queue when checked in this order.
2887 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2888 skb_queue_empty_lockless(&msk->receive_queue))
2891 return EPOLLIN | EPOLLRDNORM;
2894 static void mptcp_check_listen_stop(struct sock *sk)
2898 if (inet_sk_state_load(sk) != TCP_LISTEN)
2901 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2902 ssk = mptcp_sk(sk)->first;
2903 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2906 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2907 tcp_set_state(ssk, TCP_CLOSE);
2908 mptcp_subflow_queue_clean(sk, ssk);
2909 inet_csk_listen_stop(ssk);
2910 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2914 bool __mptcp_close(struct sock *sk, long timeout)
2916 struct mptcp_subflow_context *subflow;
2917 struct mptcp_sock *msk = mptcp_sk(sk);
2918 bool do_cancel_work = false;
2919 int subflows_alive = 0;
2921 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2923 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2924 mptcp_check_listen_stop(sk);
2925 inet_sk_state_store(sk, TCP_CLOSE);
2929 if (mptcp_check_readable(msk) || timeout < 0) {
2930 /* If the msk has read data, or the caller explicitly ask it,
2931 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
2933 mptcp_do_fastclose(sk);
2935 } else if (mptcp_close_state(sk)) {
2936 __mptcp_wr_shutdown(sk);
2939 sk_stream_wait_close(sk, timeout);
2942 /* orphan all the subflows */
2943 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2944 mptcp_for_each_subflow(msk, subflow) {
2945 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2946 bool slow = lock_sock_fast_nested(ssk);
2948 subflows_alive += ssk->sk_state != TCP_CLOSE;
2950 /* since the close timeout takes precedence on the fail one,
2953 if (ssk == msk->first)
2954 subflow->fail_tout = 0;
2956 /* detach from the parent socket, but allow data_ready to
2957 * push incoming data into the mptcp stack, to properly ack it
2959 ssk->sk_socket = NULL;
2961 unlock_sock_fast(ssk, slow);
2965 /* all the subflows are closed, only timeout can change the msk
2966 * state, let's not keep resources busy for no reasons
2968 if (subflows_alive == 0)
2969 inet_sk_state_store(sk, TCP_CLOSE);
2972 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2974 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2976 if (sk->sk_state == TCP_CLOSE) {
2977 __mptcp_destroy_sock(sk);
2978 do_cancel_work = true;
2980 mptcp_reset_timeout(msk, 0);
2983 return do_cancel_work;
2986 static void mptcp_close(struct sock *sk, long timeout)
2988 bool do_cancel_work;
2992 do_cancel_work = __mptcp_close(sk, timeout);
2995 mptcp_cancel_work(sk);
3000 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3002 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3003 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3004 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3006 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3007 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3010 msk6->saddr = ssk6->saddr;
3011 msk6->flow_label = ssk6->flow_label;
3015 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3016 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3017 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3018 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3019 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3020 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3023 static int mptcp_disconnect(struct sock *sk, int flags)
3025 struct mptcp_sock *msk = mptcp_sk(sk);
3027 /* Deny disconnect if other threads are blocked in sk_wait_event()
3028 * or inet_wait_for_connect().
3030 if (sk->sk_wait_pending)
3033 /* We are on the fastopen error path. We can't call straight into the
3034 * subflows cleanup code due to lock nesting (we are already under
3035 * msk->firstsocket lock).
3037 if (msk->fastopening)
3040 mptcp_check_listen_stop(sk);
3041 inet_sk_state_store(sk, TCP_CLOSE);
3043 mptcp_stop_timer(sk);
3044 sk_stop_timer(sk, &sk->sk_timer);
3047 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3049 /* msk->subflow is still intact, the following will not free the first
3052 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3053 msk->last_snd = NULL;
3054 WRITE_ONCE(msk->flags, 0);
3056 msk->push_pending = 0;
3057 msk->recovery = false;
3058 msk->can_ack = false;
3059 msk->fully_established = false;
3060 msk->rcv_data_fin = false;
3061 msk->snd_data_fin_enable = false;
3062 msk->rcv_fastclose = false;
3063 msk->use_64bit_ack = false;
3064 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3065 mptcp_pm_data_reset(msk);
3067 msk->bytes_acked = 0;
3068 msk->bytes_received = 0;
3069 msk->bytes_sent = 0;
3070 msk->bytes_retrans = 0;
3072 WRITE_ONCE(sk->sk_shutdown, 0);
3073 sk_error_report(sk);
3077 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3078 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3080 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3082 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3086 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3087 const struct mptcp_options_received *mp_opt,
3089 struct request_sock *req)
3091 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3092 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3093 struct mptcp_sock *msk;
3098 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3099 if (nsk->sk_family == AF_INET6)
3100 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3103 nsk->sk_wait_pending = 0;
3104 __mptcp_init_sock(nsk);
3106 msk = mptcp_sk(nsk);
3107 msk->local_key = subflow_req->local_key;
3108 msk->token = subflow_req->token;
3109 WRITE_ONCE(msk->subflow, NULL);
3110 msk->in_accept_queue = 1;
3111 WRITE_ONCE(msk->fully_established, false);
3112 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3113 WRITE_ONCE(msk->csum_enabled, true);
3115 msk->write_seq = subflow_req->idsn + 1;
3116 msk->snd_nxt = msk->write_seq;
3117 msk->snd_una = msk->write_seq;
3118 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3119 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3121 /* passive msk is created after the first/MPC subflow */
3122 msk->subflow_id = 2;
3124 sock_reset_flag(nsk, SOCK_RCU_FREE);
3125 security_inet_csk_clone(nsk, req);
3127 /* this can't race with mptcp_close(), as the msk is
3128 * not yet exposted to user-space
3130 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3132 /* The msk maintain a ref to each subflow in the connections list */
3133 WRITE_ONCE(msk->first, ssk);
3134 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3137 /* new mpc subflow takes ownership of the newly
3138 * created mptcp socket
3140 mptcp_token_accept(subflow_req, msk);
3142 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3143 * uses the correct data
3145 mptcp_copy_inaddrs(nsk, ssk);
3146 mptcp_propagate_sndbuf(nsk, ssk);
3148 mptcp_rcv_space_init(msk, ssk);
3149 bh_unlock_sock(nsk);
3151 /* note: the newly allocated socket refcount is 2 now */
3155 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3157 const struct tcp_sock *tp = tcp_sk(ssk);
3159 msk->rcvq_space.copied = 0;
3160 msk->rcvq_space.rtt_us = 0;
3162 msk->rcvq_space.time = tp->tcp_mstamp;
3164 /* initial rcv_space offering made to peer */
3165 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3166 TCP_INIT_CWND * tp->advmss);
3167 if (msk->rcvq_space.space == 0)
3168 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3170 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3173 static struct sock *mptcp_accept(struct sock *ssk, int flags, int *err,
3178 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3179 newsk = inet_csk_accept(ssk, flags, err, kern);
3183 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3184 if (sk_is_mptcp(newsk)) {
3185 struct mptcp_subflow_context *subflow;
3186 struct sock *new_mptcp_sock;
3188 subflow = mptcp_subflow_ctx(newsk);
3189 new_mptcp_sock = subflow->conn;
3191 /* is_mptcp should be false if subflow->conn is missing, see
3192 * subflow_syn_recv_sock()
3194 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3195 tcp_sk(newsk)->is_mptcp = 0;
3199 newsk = new_mptcp_sock;
3200 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3202 MPTCP_INC_STATS(sock_net(ssk),
3203 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3207 newsk->sk_kern_sock = kern;
3211 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3213 struct mptcp_subflow_context *subflow, *tmp;
3214 struct sock *sk = (struct sock *)msk;
3216 __mptcp_clear_xmit(sk);
3218 /* join list will be eventually flushed (with rst) at sock lock release time */
3219 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3220 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3222 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3223 mptcp_data_lock(sk);
3224 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3225 __skb_queue_purge(&sk->sk_receive_queue);
3226 skb_rbtree_purge(&msk->out_of_order_queue);
3227 mptcp_data_unlock(sk);
3229 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3230 * inet_sock_destruct() will dispose it
3232 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3233 msk->rmem_fwd_alloc = 0;
3234 mptcp_token_destroy(msk);
3235 mptcp_pm_free_anno_list(msk);
3236 mptcp_free_local_addr_list(msk);
3239 static void mptcp_destroy(struct sock *sk)
3241 struct mptcp_sock *msk = mptcp_sk(sk);
3243 /* clears msk->subflow, allowing the following to close
3244 * even the initial subflow
3246 mptcp_dispose_initial_subflow(msk);
3247 mptcp_destroy_common(msk, 0);
3248 sk_sockets_allocated_dec(sk);
3251 void __mptcp_data_acked(struct sock *sk)
3253 if (!sock_owned_by_user(sk))
3254 __mptcp_clean_una(sk);
3256 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3258 if (mptcp_pending_data_fin_ack(sk))
3259 mptcp_schedule_work(sk);
3262 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3264 if (!mptcp_send_head(sk))
3267 if (!sock_owned_by_user(sk))
3268 __mptcp_subflow_push_pending(sk, ssk, false);
3270 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3273 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3274 BIT(MPTCP_RETRANSMIT) | \
3275 BIT(MPTCP_FLUSH_JOIN_LIST))
3277 /* processes deferred events and flush wmem */
3278 static void mptcp_release_cb(struct sock *sk)
3279 __must_hold(&sk->sk_lock.slock)
3281 struct mptcp_sock *msk = mptcp_sk(sk);
3284 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3286 struct list_head join_list;
3291 INIT_LIST_HEAD(&join_list);
3292 list_splice_init(&msk->join_list, &join_list);
3294 /* the following actions acquire the subflow socket lock
3296 * 1) can't be invoked in atomic scope
3297 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3298 * datapath acquires the msk socket spinlock while helding
3299 * the subflow socket lock
3301 msk->push_pending = 0;
3302 msk->cb_flags &= ~flags;
3303 spin_unlock_bh(&sk->sk_lock.slock);
3305 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3306 __mptcp_flush_join_list(sk, &join_list);
3307 if (flags & BIT(MPTCP_PUSH_PENDING))
3308 __mptcp_push_pending(sk, 0);
3309 if (flags & BIT(MPTCP_RETRANSMIT))
3310 __mptcp_retrans(sk);
3313 spin_lock_bh(&sk->sk_lock.slock);
3316 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3317 __mptcp_clean_una_wakeup(sk);
3318 if (unlikely(msk->cb_flags)) {
3319 /* be sure to set the current sk state before tacking actions
3320 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3322 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3323 __mptcp_set_connected(sk);
3324 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3325 __mptcp_error_report(sk);
3326 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3327 msk->last_snd = NULL;
3330 __mptcp_update_rmem(sk);
3333 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3334 * TCP can't schedule delack timer before the subflow is fully established.
3335 * MPTCP uses the delack timer to do 3rd ack retransmissions
3337 static void schedule_3rdack_retransmission(struct sock *ssk)
3339 struct inet_connection_sock *icsk = inet_csk(ssk);
3340 struct tcp_sock *tp = tcp_sk(ssk);
3341 unsigned long timeout;
3343 if (mptcp_subflow_ctx(ssk)->fully_established)
3346 /* reschedule with a timeout above RTT, as we must look only for drop */
3348 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3350 timeout = TCP_TIMEOUT_INIT;
3353 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3354 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3355 icsk->icsk_ack.timeout = timeout;
3356 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3359 void mptcp_subflow_process_delegated(struct sock *ssk)
3361 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3362 struct sock *sk = subflow->conn;
3364 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3365 mptcp_data_lock(sk);
3366 if (!sock_owned_by_user(sk))
3367 __mptcp_subflow_push_pending(sk, ssk, true);
3369 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3370 mptcp_data_unlock(sk);
3371 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3373 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3374 schedule_3rdack_retransmission(ssk);
3375 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3379 static int mptcp_hash(struct sock *sk)
3381 /* should never be called,
3382 * we hash the TCP subflows not the master socket
3388 static void mptcp_unhash(struct sock *sk)
3390 /* called from sk_common_release(), but nothing to do here */
3393 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3395 struct mptcp_sock *msk = mptcp_sk(sk);
3397 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3398 if (WARN_ON_ONCE(!msk->first))
3401 return inet_csk_get_port(msk->first, snum);
3404 void mptcp_finish_connect(struct sock *ssk)
3406 struct mptcp_subflow_context *subflow;
3407 struct mptcp_sock *msk;
3410 subflow = mptcp_subflow_ctx(ssk);
3414 pr_debug("msk=%p, token=%u", sk, subflow->token);
3416 subflow->map_seq = subflow->iasn;
3417 subflow->map_subflow_seq = 1;
3419 /* the socket is not connected yet, no msk/subflow ops can access/race
3420 * accessing the field below
3422 WRITE_ONCE(msk->local_key, subflow->local_key);
3423 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3424 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3425 WRITE_ONCE(msk->snd_una, msk->write_seq);
3427 mptcp_pm_new_connection(msk, ssk, 0);
3429 mptcp_rcv_space_init(msk, ssk);
3432 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3434 write_lock_bh(&sk->sk_callback_lock);
3435 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3436 sk_set_socket(sk, parent);
3437 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3438 write_unlock_bh(&sk->sk_callback_lock);
3441 bool mptcp_finish_join(struct sock *ssk)
3443 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3444 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3445 struct sock *parent = (void *)msk;
3448 pr_debug("msk=%p, subflow=%p", msk, subflow);
3450 /* mptcp socket already closing? */
3451 if (!mptcp_is_fully_established(parent)) {
3452 subflow->reset_reason = MPTCP_RST_EMPTCP;
3456 /* active subflow, already present inside the conn_list */
3457 if (!list_empty(&subflow->node)) {
3458 mptcp_subflow_joined(msk, ssk);
3462 if (!mptcp_pm_allow_new_subflow(msk))
3463 goto err_prohibited;
3465 /* If we can't acquire msk socket lock here, let the release callback
3468 mptcp_data_lock(parent);
3469 if (!sock_owned_by_user(parent)) {
3470 ret = __mptcp_finish_join(msk, ssk);
3473 list_add_tail(&subflow->node, &msk->conn_list);
3477 list_add_tail(&subflow->node, &msk->join_list);
3478 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3480 mptcp_data_unlock(parent);
3484 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3491 static void mptcp_shutdown(struct sock *sk, int how)
3493 pr_debug("sk=%p, how=%d", sk, how);
3495 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3496 __mptcp_wr_shutdown(sk);
3499 static int mptcp_forward_alloc_get(const struct sock *sk)
3501 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3504 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3506 const struct sock *sk = (void *)msk;
3509 if (sk->sk_state == TCP_LISTEN)
3512 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3515 delta = msk->write_seq - v;
3516 if (__mptcp_check_fallback(msk) && msk->first) {
3517 struct tcp_sock *tp = tcp_sk(msk->first);
3519 /* the first subflow is disconnected after close - see
3520 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3521 * so ignore that status, too.
3523 if (!((1 << msk->first->sk_state) &
3524 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3525 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3527 if (delta > INT_MAX)
3533 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3535 struct mptcp_sock *msk = mptcp_sk(sk);
3540 if (sk->sk_state == TCP_LISTEN)
3544 __mptcp_move_skbs(msk);
3545 *karg = mptcp_inq_hint(sk);
3549 slow = lock_sock_fast(sk);
3550 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3551 unlock_sock_fast(sk, slow);
3554 slow = lock_sock_fast(sk);
3555 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3556 unlock_sock_fast(sk, slow);
3559 return -ENOIOCTLCMD;
3565 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3566 struct mptcp_subflow_context *subflow)
3568 subflow->request_mptcp = 0;
3569 __mptcp_do_fallback(msk);
3572 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3574 struct mptcp_subflow_context *subflow;
3575 struct mptcp_sock *msk = mptcp_sk(sk);
3579 ssk = __mptcp_nmpc_sk(msk);
3581 return PTR_ERR(ssk);
3583 inet_sk_state_store(sk, TCP_SYN_SENT);
3584 subflow = mptcp_subflow_ctx(ssk);
3585 #ifdef CONFIG_TCP_MD5SIG
3586 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3589 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3590 mptcp_subflow_early_fallback(msk, subflow);
3592 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3593 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3594 mptcp_subflow_early_fallback(msk, subflow);
3596 if (likely(!__mptcp_check_fallback(msk)))
3597 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3599 /* if reaching here via the fastopen/sendmsg path, the caller already
3600 * acquired the subflow socket lock, too.
3602 if (!msk->fastopening)
3605 /* the following mirrors closely a very small chunk of code from
3606 * __inet_stream_connect()
3608 if (ssk->sk_state != TCP_CLOSE)
3611 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3612 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3617 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3621 inet_sk(sk)->defer_connect = inet_sk(ssk)->defer_connect;
3624 if (!msk->fastopening)
3627 /* on successful connect, the msk state will be moved to established by
3628 * subflow_finish_connect()
3630 if (unlikely(err)) {
3631 /* avoid leaving a dangling token in an unconnected socket */
3632 mptcp_token_destroy(msk);
3633 inet_sk_state_store(sk, TCP_CLOSE);
3637 mptcp_copy_inaddrs(sk, ssk);
3641 static struct proto mptcp_prot = {
3643 .owner = THIS_MODULE,
3644 .init = mptcp_init_sock,
3645 .connect = mptcp_connect,
3646 .disconnect = mptcp_disconnect,
3647 .close = mptcp_close,
3648 .accept = mptcp_accept,
3649 .setsockopt = mptcp_setsockopt,
3650 .getsockopt = mptcp_getsockopt,
3651 .shutdown = mptcp_shutdown,
3652 .destroy = mptcp_destroy,
3653 .sendmsg = mptcp_sendmsg,
3654 .ioctl = mptcp_ioctl,
3655 .recvmsg = mptcp_recvmsg,
3656 .release_cb = mptcp_release_cb,
3658 .unhash = mptcp_unhash,
3659 .get_port = mptcp_get_port,
3660 .forward_alloc_get = mptcp_forward_alloc_get,
3661 .sockets_allocated = &mptcp_sockets_allocated,
3663 .memory_allocated = &tcp_memory_allocated,
3664 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3666 .memory_pressure = &tcp_memory_pressure,
3667 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3668 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3669 .sysctl_mem = sysctl_tcp_mem,
3670 .obj_size = sizeof(struct mptcp_sock),
3671 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3672 .no_autobind = true,
3675 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3677 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3678 struct sock *ssk, *sk = sock->sk;
3682 ssk = __mptcp_nmpc_sk(msk);
3688 if (sk->sk_family == AF_INET)
3689 err = inet_bind_sk(ssk, uaddr, addr_len);
3690 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3691 else if (sk->sk_family == AF_INET6)
3692 err = inet6_bind_sk(ssk, uaddr, addr_len);
3695 mptcp_copy_inaddrs(sk, ssk);
3702 static int mptcp_listen(struct socket *sock, int backlog)
3704 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3705 struct sock *sk = sock->sk;
3709 pr_debug("msk=%p", msk);
3714 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3717 ssk = __mptcp_nmpc_sk(msk);
3723 inet_sk_state_store(sk, TCP_LISTEN);
3724 sock_set_flag(sk, SOCK_RCU_FREE);
3727 err = __inet_listen_sk(ssk, backlog);
3729 inet_sk_state_store(sk, inet_sk_state_load(ssk));
3732 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3733 mptcp_copy_inaddrs(sk, ssk);
3734 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3742 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3743 int flags, bool kern)
3745 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3746 struct sock *ssk, *newsk;
3749 pr_debug("msk=%p", msk);
3751 /* Buggy applications can call accept on socket states other then LISTEN
3752 * but no need to allocate the first subflow just to error out.
3754 ssk = READ_ONCE(msk->first);
3758 newsk = mptcp_accept(ssk, flags, &err, kern);
3764 __inet_accept(sock, newsock, newsk);
3765 if (!mptcp_is_tcpsk(newsock->sk)) {
3766 struct mptcp_sock *msk = mptcp_sk(newsk);
3767 struct mptcp_subflow_context *subflow;
3769 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3770 msk->in_accept_queue = 0;
3772 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3773 * This is needed so NOSPACE flag can be set from tcp stack.
3775 mptcp_for_each_subflow(msk, subflow) {
3776 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3778 if (!ssk->sk_socket)
3779 mptcp_sock_graft(ssk, newsock);
3782 /* Do late cleanup for the first subflow as necessary. Also
3783 * deal with bad peers not doing a complete shutdown.
3786 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3787 __mptcp_close_ssk(newsk, msk->first,
3788 mptcp_subflow_ctx(msk->first), 0);
3789 if (unlikely(list_empty(&msk->conn_list)))
3790 inet_sk_state_store(newsk, TCP_CLOSE);
3793 release_sock(newsk);
3798 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3800 struct sock *sk = (struct sock *)msk;
3802 if (sk_stream_is_writeable(sk))
3803 return EPOLLOUT | EPOLLWRNORM;
3805 mptcp_set_nospace(sk);
3806 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3807 if (sk_stream_is_writeable(sk))
3808 return EPOLLOUT | EPOLLWRNORM;
3813 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3814 struct poll_table_struct *wait)
3816 struct sock *sk = sock->sk;
3817 struct mptcp_sock *msk;
3823 sock_poll_wait(file, sock, wait);
3825 state = inet_sk_state_load(sk);
3826 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3827 if (state == TCP_LISTEN) {
3828 struct sock *ssk = READ_ONCE(msk->first);
3830 if (WARN_ON_ONCE(!ssk))
3833 return inet_csk_listen_poll(ssk);
3836 shutdown = READ_ONCE(sk->sk_shutdown);
3837 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3839 if (shutdown & RCV_SHUTDOWN)
3840 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3842 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3843 mask |= mptcp_check_readable(msk);
3844 if (shutdown & SEND_SHUTDOWN)
3845 mask |= EPOLLOUT | EPOLLWRNORM;
3847 mask |= mptcp_check_writeable(msk);
3848 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3849 /* cf tcp_poll() note about TFO */
3850 mask |= EPOLLOUT | EPOLLWRNORM;
3853 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3855 if (READ_ONCE(sk->sk_err))
3861 static const struct proto_ops mptcp_stream_ops = {
3863 .owner = THIS_MODULE,
3864 .release = inet_release,
3866 .connect = inet_stream_connect,
3867 .socketpair = sock_no_socketpair,
3868 .accept = mptcp_stream_accept,
3869 .getname = inet_getname,
3871 .ioctl = inet_ioctl,
3872 .gettstamp = sock_gettstamp,
3873 .listen = mptcp_listen,
3874 .shutdown = inet_shutdown,
3875 .setsockopt = sock_common_setsockopt,
3876 .getsockopt = sock_common_getsockopt,
3877 .sendmsg = inet_sendmsg,
3878 .recvmsg = inet_recvmsg,
3879 .mmap = sock_no_mmap,
3882 static struct inet_protosw mptcp_protosw = {
3883 .type = SOCK_STREAM,
3884 .protocol = IPPROTO_MPTCP,
3885 .prot = &mptcp_prot,
3886 .ops = &mptcp_stream_ops,
3887 .flags = INET_PROTOSW_ICSK,
3890 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3892 struct mptcp_delegated_action *delegated;
3893 struct mptcp_subflow_context *subflow;
3896 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3897 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3898 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3900 bh_lock_sock_nested(ssk);
3901 if (!sock_owned_by_user(ssk) &&
3902 mptcp_subflow_has_delegated_action(subflow))
3903 mptcp_subflow_process_delegated(ssk);
3904 /* ... elsewhere tcp_release_cb_override already processed
3905 * the action or will do at next release_sock().
3906 * In both case must dequeue the subflow here - on the same
3907 * CPU that scheduled it.
3909 bh_unlock_sock(ssk);
3912 if (++work_done == budget)
3916 /* always provide a 0 'work_done' argument, so that napi_complete_done
3917 * will not try accessing the NULL napi->dev ptr
3919 napi_complete_done(napi, 0);
3923 void __init mptcp_proto_init(void)
3925 struct mptcp_delegated_action *delegated;
3928 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3930 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3931 panic("Failed to allocate MPTCP pcpu counter\n");
3933 init_dummy_netdev(&mptcp_napi_dev);
3934 for_each_possible_cpu(cpu) {
3935 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3936 INIT_LIST_HEAD(&delegated->head);
3937 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3939 napi_enable(&delegated->napi);
3942 mptcp_subflow_init();
3946 if (proto_register(&mptcp_prot, 1) != 0)
3947 panic("Failed to register MPTCP proto.\n");
3949 inet_register_protosw(&mptcp_protosw);
3951 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3954 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3955 static const struct proto_ops mptcp_v6_stream_ops = {
3957 .owner = THIS_MODULE,
3958 .release = inet6_release,
3960 .connect = inet_stream_connect,
3961 .socketpair = sock_no_socketpair,
3962 .accept = mptcp_stream_accept,
3963 .getname = inet6_getname,
3965 .ioctl = inet6_ioctl,
3966 .gettstamp = sock_gettstamp,
3967 .listen = mptcp_listen,
3968 .shutdown = inet_shutdown,
3969 .setsockopt = sock_common_setsockopt,
3970 .getsockopt = sock_common_getsockopt,
3971 .sendmsg = inet6_sendmsg,
3972 .recvmsg = inet6_recvmsg,
3973 .mmap = sock_no_mmap,
3974 #ifdef CONFIG_COMPAT
3975 .compat_ioctl = inet6_compat_ioctl,
3979 static struct proto mptcp_v6_prot;
3981 static struct inet_protosw mptcp_v6_protosw = {
3982 .type = SOCK_STREAM,
3983 .protocol = IPPROTO_MPTCP,
3984 .prot = &mptcp_v6_prot,
3985 .ops = &mptcp_v6_stream_ops,
3986 .flags = INET_PROTOSW_ICSK,
3989 int __init mptcp_proto_v6_init(void)
3993 mptcp_v6_prot = mptcp_prot;
3994 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3995 mptcp_v6_prot.slab = NULL;
3996 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3997 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
3999 err = proto_register(&mptcp_v6_prot, 1);
4003 err = inet6_register_protosw(&mptcp_v6_protosw);
4005 proto_unregister(&mptcp_v6_prot);