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 sock->ops = &inet_stream_ops;
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 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 WRITE_ONCE(msk->first, ssock->sk);
94 WRITE_ONCE(msk->subflow, ssock);
95 subflow = mptcp_subflow_ctx(ssock->sk);
96 list_add(&subflow->node, &msk->conn_list);
98 subflow->request_mptcp = 1;
100 /* This is the first subflow, always with id 0 */
101 subflow->local_id_valid = 1;
102 mptcp_sock_graft(msk->first, sk->sk_socket);
107 /* If the MPC handshake is not started, returns the first subflow,
108 * eventually allocating it.
110 struct socket *__mptcp_nmpc_socket(struct mptcp_sock *msk)
112 struct sock *sk = (struct sock *)msk;
115 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
116 return ERR_PTR(-EINVAL);
120 return ERR_PTR(-EINVAL);
122 ret = __mptcp_socket_create(msk);
126 mptcp_sockopt_sync(msk, msk->first);
132 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
134 sk_drops_add(sk, skb);
138 static void mptcp_rmem_charge(struct sock *sk, int size)
140 mptcp_sk(sk)->rmem_fwd_alloc -= size;
143 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
144 struct sk_buff *from)
149 if (MPTCP_SKB_CB(from)->offset ||
150 !skb_try_coalesce(to, from, &fragstolen, &delta))
153 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
154 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
155 to->len, MPTCP_SKB_CB(from)->end_seq);
156 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
158 /* note the fwd memory can reach a negative value after accounting
159 * for the delta, but the later skb free will restore a non
162 atomic_add(delta, &sk->sk_rmem_alloc);
163 mptcp_rmem_charge(sk, delta);
164 kfree_skb_partial(from, fragstolen);
169 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
170 struct sk_buff *from)
172 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
175 return mptcp_try_coalesce((struct sock *)msk, to, from);
178 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
180 amount >>= PAGE_SHIFT;
181 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
182 __sk_mem_reduce_allocated(sk, amount);
185 static void mptcp_rmem_uncharge(struct sock *sk, int size)
187 struct mptcp_sock *msk = mptcp_sk(sk);
190 msk->rmem_fwd_alloc += size;
191 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
193 /* see sk_mem_uncharge() for the rationale behind the following schema */
194 if (unlikely(reclaimable >= PAGE_SIZE))
195 __mptcp_rmem_reclaim(sk, reclaimable);
198 static void mptcp_rfree(struct sk_buff *skb)
200 unsigned int len = skb->truesize;
201 struct sock *sk = skb->sk;
203 atomic_sub(len, &sk->sk_rmem_alloc);
204 mptcp_rmem_uncharge(sk, len);
207 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
211 skb->destructor = mptcp_rfree;
212 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
213 mptcp_rmem_charge(sk, skb->truesize);
216 /* "inspired" by tcp_data_queue_ofo(), main differences:
218 * - don't cope with sacks
220 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
222 struct sock *sk = (struct sock *)msk;
223 struct rb_node **p, *parent;
224 u64 seq, end_seq, max_seq;
225 struct sk_buff *skb1;
227 seq = MPTCP_SKB_CB(skb)->map_seq;
228 end_seq = MPTCP_SKB_CB(skb)->end_seq;
229 max_seq = atomic64_read(&msk->rcv_wnd_sent);
231 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
232 RB_EMPTY_ROOT(&msk->out_of_order_queue));
233 if (after64(end_seq, max_seq)) {
236 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
237 (unsigned long long)end_seq - (unsigned long)max_seq,
238 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
239 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
243 p = &msk->out_of_order_queue.rb_node;
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
245 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
246 rb_link_node(&skb->rbnode, NULL, p);
247 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
248 msk->ooo_last_skb = skb;
252 /* with 2 subflows, adding at end of ooo queue is quite likely
253 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
255 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
256 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
257 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
261 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
262 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
263 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
264 parent = &msk->ooo_last_skb->rbnode;
265 p = &parent->rb_right;
269 /* Find place to insert this segment. Handle overlaps on the way. */
273 skb1 = rb_to_skb(parent);
274 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
275 p = &parent->rb_left;
278 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
279 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
280 /* All the bits are present. Drop. */
282 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
285 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
289 * continue traversing
292 /* skb's seq == skb1's seq and skb covers skb1.
293 * Replace skb1 with skb.
295 rb_replace_node(&skb1->rbnode, &skb->rbnode,
296 &msk->out_of_order_queue);
297 mptcp_drop(sk, skb1);
298 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
301 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
302 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
305 p = &parent->rb_right;
309 /* Insert segment into RB tree. */
310 rb_link_node(&skb->rbnode, parent, p);
311 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
314 /* Remove other segments covered by skb. */
315 while ((skb1 = skb_rb_next(skb)) != NULL) {
316 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
318 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
319 mptcp_drop(sk, skb1);
320 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
322 /* If there is no skb after us, we are the last_skb ! */
324 msk->ooo_last_skb = skb;
328 mptcp_set_owner_r(skb, sk);
331 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
333 struct mptcp_sock *msk = mptcp_sk(sk);
336 if (size <= msk->rmem_fwd_alloc)
339 size -= msk->rmem_fwd_alloc;
340 amt = sk_mem_pages(size);
341 amount = amt << PAGE_SHIFT;
342 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
345 msk->rmem_fwd_alloc += amount;
349 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
350 struct sk_buff *skb, unsigned int offset,
353 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
354 struct sock *sk = (struct sock *)msk;
355 struct sk_buff *tail;
358 __skb_unlink(skb, &ssk->sk_receive_queue);
363 /* try to fetch required memory from subflow */
364 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
367 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
369 /* the skb map_seq accounts for the skb offset:
370 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
373 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
374 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
375 MPTCP_SKB_CB(skb)->offset = offset;
376 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
378 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
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 !__mptcp_check_fallback(msk) &&
428 ((1 << sk->sk_state) &
429 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
430 msk->write_seq == READ_ONCE(msk->snd_una);
433 static void mptcp_check_data_fin_ack(struct sock *sk)
435 struct mptcp_sock *msk = mptcp_sk(sk);
437 /* Look for an acknowledged DATA_FIN */
438 if (mptcp_pending_data_fin_ack(sk)) {
439 WRITE_ONCE(msk->snd_data_fin_enable, 0);
441 switch (sk->sk_state) {
443 inet_sk_state_store(sk, TCP_FIN_WAIT2);
447 inet_sk_state_store(sk, TCP_CLOSE);
451 mptcp_close_wake_up(sk);
455 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
457 struct mptcp_sock *msk = mptcp_sk(sk);
459 if (READ_ONCE(msk->rcv_data_fin) &&
460 ((1 << sk->sk_state) &
461 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
462 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
464 if (msk->ack_seq == rcv_data_fin_seq) {
466 *seq = rcv_data_fin_seq;
475 static void mptcp_set_datafin_timeout(struct sock *sk)
477 struct inet_connection_sock *icsk = inet_csk(sk);
480 retransmits = min_t(u32, icsk->icsk_retransmits,
481 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
483 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
486 static void __mptcp_set_timeout(struct sock *sk, long tout)
488 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
491 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
493 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
495 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
496 inet_csk(ssk)->icsk_timeout - jiffies : 0;
499 static void mptcp_set_timeout(struct sock *sk)
501 struct mptcp_subflow_context *subflow;
504 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
505 tout = max(tout, mptcp_timeout_from_subflow(subflow));
506 __mptcp_set_timeout(sk, tout);
509 static inline bool tcp_can_send_ack(const struct sock *ssk)
511 return !((1 << inet_sk_state_load(ssk)) &
512 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
515 void __mptcp_subflow_send_ack(struct sock *ssk)
517 if (tcp_can_send_ack(ssk))
521 static void mptcp_subflow_send_ack(struct sock *ssk)
525 slow = lock_sock_fast(ssk);
526 __mptcp_subflow_send_ack(ssk);
527 unlock_sock_fast(ssk, slow);
530 static void mptcp_send_ack(struct mptcp_sock *msk)
532 struct mptcp_subflow_context *subflow;
534 mptcp_for_each_subflow(msk, subflow)
535 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
538 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
542 slow = lock_sock_fast(ssk);
543 if (tcp_can_send_ack(ssk))
544 tcp_cleanup_rbuf(ssk, 1);
545 unlock_sock_fast(ssk, slow);
548 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
550 const struct inet_connection_sock *icsk = inet_csk(ssk);
551 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
552 const struct tcp_sock *tp = tcp_sk(ssk);
554 return (ack_pending & ICSK_ACK_SCHED) &&
555 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
556 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
557 (rx_empty && ack_pending &
558 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
561 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
563 int old_space = READ_ONCE(msk->old_wspace);
564 struct mptcp_subflow_context *subflow;
565 struct sock *sk = (struct sock *)msk;
566 int space = __mptcp_space(sk);
567 bool cleanup, rx_empty;
569 cleanup = (space > 0) && (space >= (old_space << 1));
570 rx_empty = !__mptcp_rmem(sk);
572 mptcp_for_each_subflow(msk, subflow) {
573 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
575 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
576 mptcp_subflow_cleanup_rbuf(ssk);
580 static bool mptcp_check_data_fin(struct sock *sk)
582 struct mptcp_sock *msk = mptcp_sk(sk);
583 u64 rcv_data_fin_seq;
586 if (__mptcp_check_fallback(msk))
589 /* Need to ack a DATA_FIN received from a peer while this side
590 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
591 * msk->rcv_data_fin was set when parsing the incoming options
592 * at the subflow level and the msk lock was not held, so this
593 * is the first opportunity to act on the DATA_FIN and change
596 * If we are caught up to the sequence number of the incoming
597 * DATA_FIN, send the DATA_ACK now and do state transition. If
598 * not caught up, do nothing and let the recv code send DATA_ACK
602 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
603 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
604 WRITE_ONCE(msk->rcv_data_fin, 0);
606 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
607 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
609 switch (sk->sk_state) {
610 case TCP_ESTABLISHED:
611 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
614 inet_sk_state_store(sk, TCP_CLOSING);
617 inet_sk_state_store(sk, TCP_CLOSE);
620 /* Other states not expected */
627 mptcp_close_wake_up(sk);
632 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
636 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
637 struct sock *sk = (struct sock *)msk;
638 unsigned int moved = 0;
639 bool more_data_avail;
644 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
646 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
647 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
649 if (unlikely(ssk_rbuf > sk_rbuf)) {
650 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
655 pr_debug("msk=%p ssk=%p", msk, ssk);
658 u32 map_remaining, offset;
659 u32 seq = tp->copied_seq;
663 /* try to move as much data as available */
664 map_remaining = subflow->map_data_len -
665 mptcp_subflow_get_map_offset(subflow);
667 skb = skb_peek(&ssk->sk_receive_queue);
669 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
670 * a different CPU can have already processed the pending
671 * data, stop here or we can enter an infinite loop
678 if (__mptcp_check_fallback(msk)) {
679 /* Under fallback skbs have no MPTCP extension and TCP could
680 * collapse them between the dummy map creation and the
681 * current dequeue. Be sure to adjust the map size.
683 map_remaining = skb->len;
684 subflow->map_data_len = skb->len;
687 offset = seq - TCP_SKB_CB(skb)->seq;
688 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
694 if (offset < skb->len) {
695 size_t len = skb->len - offset;
700 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
704 if (WARN_ON_ONCE(map_remaining < len))
708 sk_eat_skb(ssk, skb);
712 WRITE_ONCE(tp->copied_seq, seq);
713 more_data_avail = mptcp_subflow_data_available(ssk);
715 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
719 } while (more_data_avail);
725 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
727 struct sock *sk = (struct sock *)msk;
728 struct sk_buff *skb, *tail;
733 p = rb_first(&msk->out_of_order_queue);
734 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
737 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
741 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
743 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
746 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
750 end_seq = MPTCP_SKB_CB(skb)->end_seq;
751 tail = skb_peek_tail(&sk->sk_receive_queue);
752 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
753 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
755 /* skip overlapping data, if any */
756 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
757 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
759 MPTCP_SKB_CB(skb)->offset += delta;
760 MPTCP_SKB_CB(skb)->map_seq += delta;
761 __skb_queue_tail(&sk->sk_receive_queue, skb);
763 msk->ack_seq = end_seq;
769 /* In most cases we will be able to lock the mptcp socket. If its already
770 * owned, we need to defer to the work queue to avoid ABBA deadlock.
772 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
774 struct sock *sk = (struct sock *)msk;
775 unsigned int moved = 0;
777 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
778 __mptcp_ofo_queue(msk);
779 if (unlikely(ssk->sk_err)) {
780 if (!sock_owned_by_user(sk))
781 __mptcp_error_report(sk);
783 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
786 /* If the moves have caught up with the DATA_FIN sequence number
787 * it's time to ack the DATA_FIN and change socket state, but
788 * this is not a good place to change state. Let the workqueue
791 if (mptcp_pending_data_fin(sk, NULL))
792 mptcp_schedule_work(sk);
796 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
798 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
799 struct mptcp_sock *msk = mptcp_sk(sk);
800 int sk_rbuf, ssk_rbuf;
802 /* The peer can send data while we are shutting down this
803 * subflow at msk destruction time, but we must avoid enqueuing
804 * more data to the msk receive queue
806 if (unlikely(subflow->disposable))
809 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
810 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
811 if (unlikely(ssk_rbuf > sk_rbuf))
814 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
815 if (__mptcp_rmem(sk) > sk_rbuf) {
816 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
820 /* Wake-up the reader only for in-sequence data */
822 if (move_skbs_to_msk(msk, ssk))
823 sk->sk_data_ready(sk);
825 mptcp_data_unlock(sk);
828 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
830 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
831 WRITE_ONCE(msk->allow_infinite_fallback, false);
832 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
835 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
837 struct sock *sk = (struct sock *)msk;
839 if (sk->sk_state != TCP_ESTABLISHED)
842 /* attach to msk socket only after we are sure we will deal with it
845 if (sk->sk_socket && !ssk->sk_socket)
846 mptcp_sock_graft(ssk, sk->sk_socket);
848 mptcp_sockopt_sync_locked(msk, ssk);
849 mptcp_subflow_joined(msk, ssk);
853 static void __mptcp_flush_join_list(struct sock *sk)
855 struct mptcp_subflow_context *tmp, *subflow;
856 struct mptcp_sock *msk = mptcp_sk(sk);
858 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
859 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
860 bool slow = lock_sock_fast(ssk);
862 list_move_tail(&subflow->node, &msk->conn_list);
863 if (!__mptcp_finish_join(msk, ssk))
864 mptcp_subflow_reset(ssk);
865 unlock_sock_fast(ssk, slow);
869 static bool mptcp_timer_pending(struct sock *sk)
871 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
874 static void mptcp_reset_timer(struct sock *sk)
876 struct inet_connection_sock *icsk = inet_csk(sk);
879 /* prevent rescheduling on close */
880 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
883 tout = mptcp_sk(sk)->timer_ival;
884 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
887 bool mptcp_schedule_work(struct sock *sk)
889 if (inet_sk_state_load(sk) != TCP_CLOSE &&
890 schedule_work(&mptcp_sk(sk)->work)) {
891 /* each subflow already holds a reference to the sk, and the
892 * workqueue is invoked by a subflow, so sk can't go away here.
900 void mptcp_subflow_eof(struct sock *sk)
902 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
903 mptcp_schedule_work(sk);
906 static void mptcp_check_for_eof(struct mptcp_sock *msk)
908 struct mptcp_subflow_context *subflow;
909 struct sock *sk = (struct sock *)msk;
912 mptcp_for_each_subflow(msk, subflow)
913 receivers += !subflow->rx_eof;
917 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
918 /* hopefully temporary hack: propagate shutdown status
919 * to msk, when all subflows agree on it
921 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
923 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
924 sk->sk_data_ready(sk);
927 switch (sk->sk_state) {
928 case TCP_ESTABLISHED:
929 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
932 inet_sk_state_store(sk, TCP_CLOSING);
935 inet_sk_state_store(sk, TCP_CLOSE);
940 mptcp_close_wake_up(sk);
943 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
945 struct mptcp_subflow_context *subflow;
947 msk_owned_by_me(msk);
949 mptcp_for_each_subflow(msk, subflow) {
950 if (READ_ONCE(subflow->data_avail))
951 return mptcp_subflow_tcp_sock(subflow);
957 static bool mptcp_skb_can_collapse_to(u64 write_seq,
958 const struct sk_buff *skb,
959 const struct mptcp_ext *mpext)
961 if (!tcp_skb_can_collapse_to(skb))
964 /* can collapse only if MPTCP level sequence is in order and this
965 * mapping has not been xmitted yet
967 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
971 /* we can append data to the given data frag if:
972 * - there is space available in the backing page_frag
973 * - the data frag tail matches the current page_frag free offset
974 * - the data frag end sequence number matches the current write seq
976 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
977 const struct page_frag *pfrag,
978 const struct mptcp_data_frag *df)
980 return df && pfrag->page == df->page &&
981 pfrag->size - pfrag->offset > 0 &&
982 pfrag->offset == (df->offset + df->data_len) &&
983 df->data_seq + df->data_len == msk->write_seq;
986 static void dfrag_uncharge(struct sock *sk, int len)
988 sk_mem_uncharge(sk, len);
989 sk_wmem_queued_add(sk, -len);
992 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
994 int len = dfrag->data_len + dfrag->overhead;
996 list_del(&dfrag->list);
997 dfrag_uncharge(sk, len);
998 put_page(dfrag->page);
1001 static void __mptcp_clean_una(struct sock *sk)
1003 struct mptcp_sock *msk = mptcp_sk(sk);
1004 struct mptcp_data_frag *dtmp, *dfrag;
1007 /* on fallback we just need to ignore snd_una, as this is really
1010 if (__mptcp_check_fallback(msk))
1011 msk->snd_una = READ_ONCE(msk->snd_nxt);
1013 snd_una = msk->snd_una;
1014 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1015 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1018 if (unlikely(dfrag == msk->first_pending)) {
1019 /* in recovery mode can see ack after the current snd head */
1020 if (WARN_ON_ONCE(!msk->recovery))
1023 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1026 dfrag_clear(sk, dfrag);
1029 dfrag = mptcp_rtx_head(sk);
1030 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1031 u64 delta = snd_una - dfrag->data_seq;
1033 /* prevent wrap around in recovery mode */
1034 if (unlikely(delta > dfrag->already_sent)) {
1035 if (WARN_ON_ONCE(!msk->recovery))
1037 if (WARN_ON_ONCE(delta > dfrag->data_len))
1039 dfrag->already_sent += delta - dfrag->already_sent;
1042 dfrag->data_seq += delta;
1043 dfrag->offset += delta;
1044 dfrag->data_len -= delta;
1045 dfrag->already_sent -= delta;
1047 dfrag_uncharge(sk, delta);
1050 /* all retransmitted data acked, recovery completed */
1051 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1052 msk->recovery = false;
1055 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1056 snd_una == READ_ONCE(msk->write_seq)) {
1057 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1058 mptcp_stop_timer(sk);
1060 mptcp_reset_timer(sk);
1064 static void __mptcp_clean_una_wakeup(struct sock *sk)
1066 lockdep_assert_held_once(&sk->sk_lock.slock);
1068 __mptcp_clean_una(sk);
1069 mptcp_write_space(sk);
1072 static void mptcp_clean_una_wakeup(struct sock *sk)
1074 mptcp_data_lock(sk);
1075 __mptcp_clean_una_wakeup(sk);
1076 mptcp_data_unlock(sk);
1079 static void mptcp_enter_memory_pressure(struct sock *sk)
1081 struct mptcp_subflow_context *subflow;
1082 struct mptcp_sock *msk = mptcp_sk(sk);
1085 sk_stream_moderate_sndbuf(sk);
1086 mptcp_for_each_subflow(msk, subflow) {
1087 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1090 tcp_enter_memory_pressure(ssk);
1091 sk_stream_moderate_sndbuf(ssk);
1096 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1099 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1101 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1102 pfrag, sk->sk_allocation)))
1105 mptcp_enter_memory_pressure(sk);
1109 static struct mptcp_data_frag *
1110 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1113 int offset = ALIGN(orig_offset, sizeof(long));
1114 struct mptcp_data_frag *dfrag;
1116 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1117 dfrag->data_len = 0;
1118 dfrag->data_seq = msk->write_seq;
1119 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1120 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1121 dfrag->already_sent = 0;
1122 dfrag->page = pfrag->page;
1127 struct mptcp_sendmsg_info {
1133 bool data_lock_held;
1136 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1137 u64 data_seq, int avail_size)
1139 u64 window_end = mptcp_wnd_end(msk);
1142 if (__mptcp_check_fallback(msk))
1145 mptcp_snd_wnd = window_end - data_seq;
1146 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1148 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1149 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1150 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1156 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1158 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1162 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1166 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1168 struct sk_buff *skb;
1170 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1172 if (likely(__mptcp_add_ext(skb, gfp))) {
1173 skb_reserve(skb, MAX_TCP_HEADER);
1174 skb->ip_summed = CHECKSUM_PARTIAL;
1175 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1180 mptcp_enter_memory_pressure(sk);
1185 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1187 struct sk_buff *skb;
1189 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1193 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1194 tcp_skb_entail(ssk, skb);
1197 tcp_skb_tsorted_anchor_cleanup(skb);
1202 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1204 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1206 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1209 /* note: this always recompute the csum on the whole skb, even
1210 * if we just appended a single frag. More status info needed
1212 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1214 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1215 __wsum csum = ~csum_unfold(mpext->csum);
1216 int offset = skb->len - added;
1218 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1221 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1223 struct mptcp_ext *mpext)
1228 mpext->infinite_map = 1;
1229 mpext->data_len = 0;
1231 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1232 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1234 mptcp_do_fallback(ssk);
1237 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1238 struct mptcp_data_frag *dfrag,
1239 struct mptcp_sendmsg_info *info)
1241 u64 data_seq = dfrag->data_seq + info->sent;
1242 int offset = dfrag->offset + info->sent;
1243 struct mptcp_sock *msk = mptcp_sk(sk);
1244 bool zero_window_probe = false;
1245 struct mptcp_ext *mpext = NULL;
1246 bool can_coalesce = false;
1247 bool reuse_skb = true;
1248 struct sk_buff *skb;
1252 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1253 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1255 if (WARN_ON_ONCE(info->sent > info->limit ||
1256 info->limit > dfrag->data_len))
1259 if (unlikely(!__tcp_can_send(ssk)))
1262 /* compute send limit */
1263 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1264 copy = info->size_goal;
1266 skb = tcp_write_queue_tail(ssk);
1267 if (skb && copy > skb->len) {
1268 /* Limit the write to the size available in the
1269 * current skb, if any, so that we create at most a new skb.
1270 * Explicitly tells TCP internals to avoid collapsing on later
1271 * queue management operation, to avoid breaking the ext <->
1272 * SSN association set here
1274 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1275 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1276 TCP_SKB_CB(skb)->eor = 1;
1280 i = skb_shinfo(skb)->nr_frags;
1281 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1282 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1283 tcp_mark_push(tcp_sk(ssk), skb);
1290 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1294 i = skb_shinfo(skb)->nr_frags;
1296 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1299 /* Zero window and all data acked? Probe. */
1300 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1302 u64 snd_una = READ_ONCE(msk->snd_una);
1304 if (snd_una != msk->snd_nxt) {
1305 tcp_remove_empty_skb(ssk);
1309 zero_window_probe = true;
1310 data_seq = snd_una - 1;
1313 /* all mptcp-level data is acked, no skbs should be present into the
1316 WARN_ON_ONCE(reuse_skb);
1319 copy = min_t(size_t, copy, info->limit - info->sent);
1320 if (!sk_wmem_schedule(ssk, copy)) {
1321 tcp_remove_empty_skb(ssk);
1326 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1328 get_page(dfrag->page);
1329 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1333 skb->data_len += copy;
1334 skb->truesize += copy;
1335 sk_wmem_queued_add(ssk, copy);
1336 sk_mem_charge(ssk, copy);
1337 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1338 TCP_SKB_CB(skb)->end_seq += copy;
1339 tcp_skb_pcount_set(skb, 0);
1341 /* on skb reuse we just need to update the DSS len */
1343 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1344 mpext->data_len += copy;
1345 WARN_ON_ONCE(zero_window_probe);
1349 memset(mpext, 0, sizeof(*mpext));
1350 mpext->data_seq = data_seq;
1351 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1352 mpext->data_len = copy;
1356 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1357 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1360 if (zero_window_probe) {
1361 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1363 if (READ_ONCE(msk->csum_enabled))
1364 mptcp_update_data_checksum(skb, copy);
1365 tcp_push_pending_frames(ssk);
1369 if (READ_ONCE(msk->csum_enabled))
1370 mptcp_update_data_checksum(skb, copy);
1371 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1372 mptcp_update_infinite_map(msk, ssk, mpext);
1373 trace_mptcp_sendmsg_frag(mpext);
1374 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1378 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1379 sizeof(struct tcphdr) - \
1380 MAX_TCP_OPTION_SPACE - \
1381 sizeof(struct ipv6hdr) - \
1382 sizeof(struct frag_hdr))
1384 struct subflow_send_info {
1389 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1391 if (!subflow->stale)
1395 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1398 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1400 if (unlikely(subflow->stale)) {
1401 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1403 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1406 mptcp_subflow_set_active(subflow);
1408 return __mptcp_subflow_active(subflow);
1411 #define SSK_MODE_ACTIVE 0
1412 #define SSK_MODE_BACKUP 1
1413 #define SSK_MODE_MAX 2
1415 /* implement the mptcp packet scheduler;
1416 * returns the subflow that will transmit the next DSS
1417 * additionally updates the rtx timeout
1419 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1421 struct subflow_send_info send_info[SSK_MODE_MAX];
1422 struct mptcp_subflow_context *subflow;
1423 struct sock *sk = (struct sock *)msk;
1424 u32 pace, burst, wmem;
1425 int i, nr_active = 0;
1430 msk_owned_by_me(msk);
1432 if (__mptcp_check_fallback(msk)) {
1435 return __tcp_can_send(msk->first) &&
1436 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1439 /* re-use last subflow, if the burst allow that */
1440 if (msk->last_snd && msk->snd_burst > 0 &&
1441 sk_stream_memory_free(msk->last_snd) &&
1442 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1443 mptcp_set_timeout(sk);
1444 return msk->last_snd;
1447 /* pick the subflow with the lower wmem/wspace ratio */
1448 for (i = 0; i < SSK_MODE_MAX; ++i) {
1449 send_info[i].ssk = NULL;
1450 send_info[i].linger_time = -1;
1453 mptcp_for_each_subflow(msk, subflow) {
1454 trace_mptcp_subflow_get_send(subflow);
1455 ssk = mptcp_subflow_tcp_sock(subflow);
1456 if (!mptcp_subflow_active(subflow))
1459 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1460 nr_active += !subflow->backup;
1461 pace = subflow->avg_pacing_rate;
1462 if (unlikely(!pace)) {
1463 /* init pacing rate from socket */
1464 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1465 pace = subflow->avg_pacing_rate;
1470 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1471 if (linger_time < send_info[subflow->backup].linger_time) {
1472 send_info[subflow->backup].ssk = ssk;
1473 send_info[subflow->backup].linger_time = linger_time;
1476 __mptcp_set_timeout(sk, tout);
1478 /* pick the best backup if no other subflow is active */
1480 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1482 /* According to the blest algorithm, to avoid HoL blocking for the
1483 * faster flow, we need to:
1484 * - estimate the faster flow linger time
1485 * - use the above to estimate the amount of byte transferred
1486 * by the faster flow
1487 * - check that the amount of queued data is greter than the above,
1488 * otherwise do not use the picked, slower, subflow
1489 * We select the subflow with the shorter estimated time to flush
1490 * the queued mem, which basically ensure the above. We just need
1491 * to check that subflow has a non empty cwin.
1493 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1494 if (!ssk || !sk_stream_memory_free(ssk))
1497 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1498 wmem = READ_ONCE(ssk->sk_wmem_queued);
1500 msk->last_snd = NULL;
1504 subflow = mptcp_subflow_ctx(ssk);
1505 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1506 READ_ONCE(ssk->sk_pacing_rate) * burst,
1508 msk->last_snd = ssk;
1509 msk->snd_burst = burst;
1513 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1515 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1519 static void mptcp_update_post_push(struct mptcp_sock *msk,
1520 struct mptcp_data_frag *dfrag,
1523 u64 snd_nxt_new = dfrag->data_seq;
1525 dfrag->already_sent += sent;
1527 msk->snd_burst -= sent;
1529 snd_nxt_new += dfrag->already_sent;
1531 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1532 * is recovering after a failover. In that event, this re-sends
1535 * Thus compute snd_nxt_new candidate based on
1536 * the dfrag->data_seq that was sent and the data
1537 * that has been handed to the subflow for transmission
1538 * and skip update in case it was old dfrag.
1540 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1541 msk->snd_nxt = snd_nxt_new;
1544 void mptcp_check_and_set_pending(struct sock *sk)
1546 if (mptcp_send_head(sk))
1547 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1550 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1552 struct sock *prev_ssk = NULL, *ssk = NULL;
1553 struct mptcp_sock *msk = mptcp_sk(sk);
1554 struct mptcp_sendmsg_info info = {
1557 bool do_check_data_fin = false;
1558 struct mptcp_data_frag *dfrag;
1561 while ((dfrag = mptcp_send_head(sk))) {
1562 info.sent = dfrag->already_sent;
1563 info.limit = dfrag->data_len;
1564 len = dfrag->data_len - dfrag->already_sent;
1569 ssk = mptcp_subflow_get_send(msk);
1571 /* First check. If the ssk has changed since
1572 * the last round, release prev_ssk
1574 if (ssk != prev_ssk && prev_ssk)
1575 mptcp_push_release(prev_ssk, &info);
1579 /* Need to lock the new subflow only if different
1580 * from the previous one, otherwise we are still
1581 * helding the relevant lock
1583 if (ssk != prev_ssk)
1586 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1590 mptcp_push_release(ssk, &info);
1594 do_check_data_fin = true;
1598 mptcp_update_post_push(msk, dfrag, ret);
1600 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1603 /* at this point we held the socket lock for the last subflow we used */
1605 mptcp_push_release(ssk, &info);
1608 /* ensure the rtx timer is running */
1609 if (!mptcp_timer_pending(sk))
1610 mptcp_reset_timer(sk);
1611 if (do_check_data_fin)
1612 __mptcp_check_send_data_fin(sk);
1615 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1617 struct mptcp_sock *msk = mptcp_sk(sk);
1618 struct mptcp_sendmsg_info info = {
1619 .data_lock_held = true,
1621 struct mptcp_data_frag *dfrag;
1622 struct sock *xmit_ssk;
1623 int len, copied = 0;
1626 while ((dfrag = mptcp_send_head(sk))) {
1627 info.sent = dfrag->already_sent;
1628 info.limit = dfrag->data_len;
1629 len = dfrag->data_len - dfrag->already_sent;
1633 /* check for a different subflow usage only after
1634 * spooling the first chunk of data
1636 xmit_ssk = first ? ssk : mptcp_subflow_get_send(msk);
1639 if (xmit_ssk != ssk) {
1640 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1641 MPTCP_DELEGATE_SEND);
1645 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1654 mptcp_update_post_push(msk, dfrag, ret);
1656 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1660 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1661 * not going to flush it via release_sock()
1664 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1666 if (!mptcp_timer_pending(sk))
1667 mptcp_reset_timer(sk);
1669 if (msk->snd_data_fin_enable &&
1670 msk->snd_nxt + 1 == msk->write_seq)
1671 mptcp_schedule_work(sk);
1675 static void mptcp_set_nospace(struct sock *sk)
1677 /* enable autotune */
1678 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1680 /* will be cleared on avail space */
1681 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1684 static int mptcp_disconnect(struct sock *sk, int flags);
1686 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1687 size_t len, int *copied_syn)
1689 unsigned int saved_flags = msg->msg_flags;
1690 struct mptcp_sock *msk = mptcp_sk(sk);
1691 struct socket *ssock;
1695 /* on flags based fastopen the mptcp is supposed to create the
1696 * first subflow right now. Otherwise we are in the defer_connect
1697 * path, and the first subflow must be already present.
1698 * Since the defer_connect flag is cleared after the first succsful
1699 * fastopen attempt, no need to check for additional subflow status.
1701 if (msg->msg_flags & MSG_FASTOPEN) {
1702 ssock = __mptcp_nmpc_socket(msk);
1704 return PTR_ERR(ssock);
1712 msg->msg_flags |= MSG_DONTWAIT;
1713 msk->fastopening = 1;
1714 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1715 msk->fastopening = 0;
1716 msg->msg_flags = saved_flags;
1719 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1720 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1721 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1722 msg->msg_namelen, msg->msg_flags, 1);
1724 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1725 * case of any error, except timeout or signal
1727 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1729 } else if (ret && ret != -EINPROGRESS) {
1730 mptcp_disconnect(sk, 0);
1732 inet_sk(sk)->defer_connect = 0;
1737 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1739 struct mptcp_sock *msk = mptcp_sk(sk);
1740 struct page_frag *pfrag;
1745 /* silently ignore everything else */
1746 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1750 if (unlikely(inet_sk(sk)->defer_connect || msg->msg_flags & MSG_FASTOPEN)) {
1753 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1754 copied += copied_syn;
1755 if (ret == -EINPROGRESS && copied_syn > 0)
1761 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1763 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1764 ret = sk_stream_wait_connect(sk, &timeo);
1770 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1773 pfrag = sk_page_frag(sk);
1775 while (msg_data_left(msg)) {
1776 int total_ts, frag_truesize = 0;
1777 struct mptcp_data_frag *dfrag;
1778 bool dfrag_collapsed;
1779 size_t psize, offset;
1781 /* reuse tail pfrag, if possible, or carve a new one from the
1784 dfrag = mptcp_pending_tail(sk);
1785 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1786 if (!dfrag_collapsed) {
1787 if (!sk_stream_memory_free(sk))
1788 goto wait_for_memory;
1790 if (!mptcp_page_frag_refill(sk, pfrag))
1791 goto wait_for_memory;
1793 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1794 frag_truesize = dfrag->overhead;
1797 /* we do not bound vs wspace, to allow a single packet.
1798 * memory accounting will prevent execessive memory usage
1801 offset = dfrag->offset + dfrag->data_len;
1802 psize = pfrag->size - offset;
1803 psize = min_t(size_t, psize, msg_data_left(msg));
1804 total_ts = psize + frag_truesize;
1806 if (!sk_wmem_schedule(sk, total_ts))
1807 goto wait_for_memory;
1809 if (copy_page_from_iter(dfrag->page, offset, psize,
1810 &msg->msg_iter) != psize) {
1815 /* data successfully copied into the write queue */
1816 sk->sk_forward_alloc -= total_ts;
1818 dfrag->data_len += psize;
1819 frag_truesize += psize;
1820 pfrag->offset += frag_truesize;
1821 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1823 /* charge data on mptcp pending queue to the msk socket
1824 * Note: we charge such data both to sk and ssk
1826 sk_wmem_queued_add(sk, frag_truesize);
1827 if (!dfrag_collapsed) {
1828 get_page(dfrag->page);
1829 list_add_tail(&dfrag->list, &msk->rtx_queue);
1830 if (!msk->first_pending)
1831 WRITE_ONCE(msk->first_pending, dfrag);
1833 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1834 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1840 mptcp_set_nospace(sk);
1841 __mptcp_push_pending(sk, msg->msg_flags);
1842 ret = sk_stream_wait_memory(sk, &timeo);
1848 __mptcp_push_pending(sk, msg->msg_flags);
1858 copied = sk_stream_error(sk, msg->msg_flags, ret);
1862 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1864 size_t len, int flags,
1865 struct scm_timestamping_internal *tss,
1868 struct sk_buff *skb, *tmp;
1871 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1872 u32 offset = MPTCP_SKB_CB(skb)->offset;
1873 u32 data_len = skb->len - offset;
1874 u32 count = min_t(size_t, len - copied, data_len);
1877 if (!(flags & MSG_TRUNC)) {
1878 err = skb_copy_datagram_msg(skb, offset, msg, count);
1879 if (unlikely(err < 0)) {
1886 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1887 tcp_update_recv_tstamps(skb, tss);
1888 *cmsg_flags |= MPTCP_CMSG_TS;
1893 if (count < data_len) {
1894 if (!(flags & MSG_PEEK)) {
1895 MPTCP_SKB_CB(skb)->offset += count;
1896 MPTCP_SKB_CB(skb)->map_seq += count;
1901 if (!(flags & MSG_PEEK)) {
1902 /* we will bulk release the skb memory later */
1903 skb->destructor = NULL;
1904 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1905 __skb_unlink(skb, &msk->receive_queue);
1916 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1918 * Only difference: Use highest rtt estimate of the subflows in use.
1920 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1922 struct mptcp_subflow_context *subflow;
1923 struct sock *sk = (struct sock *)msk;
1924 u32 time, advmss = 1;
1927 msk_owned_by_me(msk);
1932 msk->rcvq_space.copied += copied;
1934 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1935 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1937 rtt_us = msk->rcvq_space.rtt_us;
1938 if (rtt_us && time < (rtt_us >> 3))
1942 mptcp_for_each_subflow(msk, subflow) {
1943 const struct tcp_sock *tp;
1947 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1949 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1950 sf_advmss = READ_ONCE(tp->advmss);
1952 rtt_us = max(sf_rtt_us, rtt_us);
1953 advmss = max(sf_advmss, advmss);
1956 msk->rcvq_space.rtt_us = rtt_us;
1957 if (time < (rtt_us >> 3) || rtt_us == 0)
1960 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1963 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1964 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1968 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1970 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1972 do_div(grow, msk->rcvq_space.space);
1973 rcvwin += (grow << 1);
1975 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1976 while (tcp_win_from_space(sk, rcvmem) < advmss)
1979 do_div(rcvwin, advmss);
1980 rcvbuf = min_t(u64, rcvwin * rcvmem,
1981 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1983 if (rcvbuf > sk->sk_rcvbuf) {
1986 window_clamp = tcp_win_from_space(sk, rcvbuf);
1987 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1989 /* Make subflows follow along. If we do not do this, we
1990 * get drops at subflow level if skbs can't be moved to
1991 * the mptcp rx queue fast enough (announced rcv_win can
1992 * exceed ssk->sk_rcvbuf).
1994 mptcp_for_each_subflow(msk, subflow) {
1998 ssk = mptcp_subflow_tcp_sock(subflow);
1999 slow = lock_sock_fast(ssk);
2000 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2001 tcp_sk(ssk)->window_clamp = window_clamp;
2002 tcp_cleanup_rbuf(ssk, 1);
2003 unlock_sock_fast(ssk, slow);
2008 msk->rcvq_space.space = msk->rcvq_space.copied;
2010 msk->rcvq_space.copied = 0;
2011 msk->rcvq_space.time = mstamp;
2014 static void __mptcp_update_rmem(struct sock *sk)
2016 struct mptcp_sock *msk = mptcp_sk(sk);
2018 if (!msk->rmem_released)
2021 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2022 mptcp_rmem_uncharge(sk, msk->rmem_released);
2023 WRITE_ONCE(msk->rmem_released, 0);
2026 static void __mptcp_splice_receive_queue(struct sock *sk)
2028 struct mptcp_sock *msk = mptcp_sk(sk);
2030 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2033 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2035 struct sock *sk = (struct sock *)msk;
2036 unsigned int moved = 0;
2040 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2043 /* we can have data pending in the subflows only if the msk
2044 * receive buffer was full at subflow_data_ready() time,
2045 * that is an unlikely slow path.
2050 slowpath = lock_sock_fast(ssk);
2051 mptcp_data_lock(sk);
2052 __mptcp_update_rmem(sk);
2053 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2054 mptcp_data_unlock(sk);
2056 if (unlikely(ssk->sk_err))
2057 __mptcp_error_report(sk);
2058 unlock_sock_fast(ssk, slowpath);
2061 /* acquire the data lock only if some input data is pending */
2063 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2064 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2065 mptcp_data_lock(sk);
2066 __mptcp_update_rmem(sk);
2067 ret |= __mptcp_ofo_queue(msk);
2068 __mptcp_splice_receive_queue(sk);
2069 mptcp_data_unlock(sk);
2072 mptcp_check_data_fin((struct sock *)msk);
2073 return !skb_queue_empty(&msk->receive_queue);
2076 static unsigned int mptcp_inq_hint(const struct sock *sk)
2078 const struct mptcp_sock *msk = mptcp_sk(sk);
2079 const struct sk_buff *skb;
2081 skb = skb_peek(&msk->receive_queue);
2083 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2085 if (hint_val >= INT_MAX)
2088 return (unsigned int)hint_val;
2091 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2097 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2098 int flags, int *addr_len)
2100 struct mptcp_sock *msk = mptcp_sk(sk);
2101 struct scm_timestamping_internal tss;
2102 int copied = 0, cmsg_flags = 0;
2106 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2107 if (unlikely(flags & MSG_ERRQUEUE))
2108 return inet_recv_error(sk, msg, len, addr_len);
2111 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2116 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2118 len = min_t(size_t, len, INT_MAX);
2119 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2121 if (unlikely(msk->recvmsg_inq))
2122 cmsg_flags = MPTCP_CMSG_INQ;
2124 while (copied < len) {
2127 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2128 if (unlikely(bytes_read < 0)) {
2130 copied = bytes_read;
2134 copied += bytes_read;
2136 /* be sure to advertise window change */
2137 mptcp_cleanup_rbuf(msk);
2139 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2142 /* only the master socket status is relevant here. The exit
2143 * conditions mirror closely tcp_recvmsg()
2145 if (copied >= target)
2150 sk->sk_state == TCP_CLOSE ||
2151 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2153 signal_pending(current))
2157 copied = sock_error(sk);
2161 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2162 mptcp_check_for_eof(msk);
2164 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2165 /* race breaker: the shutdown could be after the
2166 * previous receive queue check
2168 if (__mptcp_move_skbs(msk))
2173 if (sk->sk_state == TCP_CLOSE) {
2183 if (signal_pending(current)) {
2184 copied = sock_intr_errno(timeo);
2189 pr_debug("block timeout %ld", timeo);
2190 sk_wait_data(sk, &timeo, NULL);
2194 if (cmsg_flags && copied >= 0) {
2195 if (cmsg_flags & MPTCP_CMSG_TS)
2196 tcp_recv_timestamp(msg, sk, &tss);
2198 if (cmsg_flags & MPTCP_CMSG_INQ) {
2199 unsigned int inq = mptcp_inq_hint(sk);
2201 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2205 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2206 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2207 skb_queue_empty(&msk->receive_queue), copied);
2208 if (!(flags & MSG_PEEK))
2209 mptcp_rcv_space_adjust(msk, copied);
2215 static void mptcp_retransmit_timer(struct timer_list *t)
2217 struct inet_connection_sock *icsk = from_timer(icsk, t,
2218 icsk_retransmit_timer);
2219 struct sock *sk = &icsk->icsk_inet.sk;
2220 struct mptcp_sock *msk = mptcp_sk(sk);
2223 if (!sock_owned_by_user(sk)) {
2224 /* we need a process context to retransmit */
2225 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2226 mptcp_schedule_work(sk);
2228 /* delegate our work to tcp_release_cb() */
2229 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2235 static void mptcp_timeout_timer(struct timer_list *t)
2237 struct sock *sk = from_timer(sk, t, sk_timer);
2239 mptcp_schedule_work(sk);
2243 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2246 * A backup subflow is returned only if that is the only kind available.
2248 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2250 struct sock *backup = NULL, *pick = NULL;
2251 struct mptcp_subflow_context *subflow;
2252 int min_stale_count = INT_MAX;
2254 msk_owned_by_me(msk);
2256 if (__mptcp_check_fallback(msk))
2259 mptcp_for_each_subflow(msk, subflow) {
2260 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2262 if (!__mptcp_subflow_active(subflow))
2265 /* still data outstanding at TCP level? skip this */
2266 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2267 mptcp_pm_subflow_chk_stale(msk, ssk);
2268 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2272 if (subflow->backup) {
2285 /* use backup only if there are no progresses anywhere */
2286 return min_stale_count > 1 ? backup : NULL;
2289 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2292 iput(SOCK_INODE(msk->subflow));
2293 WRITE_ONCE(msk->subflow, NULL);
2297 bool __mptcp_retransmit_pending_data(struct sock *sk)
2299 struct mptcp_data_frag *cur, *rtx_head;
2300 struct mptcp_sock *msk = mptcp_sk(sk);
2302 if (__mptcp_check_fallback(msk))
2305 if (tcp_rtx_and_write_queues_empty(sk))
2308 /* the closing socket has some data untransmitted and/or unacked:
2309 * some data in the mptcp rtx queue has not really xmitted yet.
2310 * keep it simple and re-inject the whole mptcp level rtx queue
2312 mptcp_data_lock(sk);
2313 __mptcp_clean_una_wakeup(sk);
2314 rtx_head = mptcp_rtx_head(sk);
2316 mptcp_data_unlock(sk);
2320 msk->recovery_snd_nxt = msk->snd_nxt;
2321 msk->recovery = true;
2322 mptcp_data_unlock(sk);
2324 msk->first_pending = rtx_head;
2327 /* be sure to clear the "sent status" on all re-injected fragments */
2328 list_for_each_entry(cur, &msk->rtx_queue, list) {
2329 if (!cur->already_sent)
2331 cur->already_sent = 0;
2337 /* flags for __mptcp_close_ssk() */
2338 #define MPTCP_CF_PUSH BIT(1)
2339 #define MPTCP_CF_FASTCLOSE BIT(2)
2341 /* subflow sockets can be either outgoing (connect) or incoming
2344 * Outgoing subflows use in-kernel sockets.
2345 * Incoming subflows do not have their own 'struct socket' allocated,
2346 * so we need to use tcp_close() after detaching them from the mptcp
2349 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2350 struct mptcp_subflow_context *subflow,
2353 struct mptcp_sock *msk = mptcp_sk(sk);
2354 bool dispose_it, need_push = false;
2356 /* If the first subflow moved to a close state before accept, e.g. due
2357 * to an incoming reset, mptcp either:
2358 * - if either the subflow or the msk are dead, destroy the context
2359 * (the subflow socket is deleted by inet_child_forget) and the msk
2360 * - otherwise do nothing at the moment and take action at accept and/or
2361 * listener shutdown - user-space must be able to accept() the closed
2364 if (msk->in_accept_queue && msk->first == ssk) {
2365 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2368 /* ensure later check in mptcp_worker() will dispose the msk */
2369 sock_set_flag(sk, SOCK_DEAD);
2370 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2371 mptcp_subflow_drop_ctx(ssk);
2375 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2377 list_del(&subflow->node);
2379 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2381 if (flags & MPTCP_CF_FASTCLOSE) {
2382 /* be sure to force the tcp_disconnect() path,
2383 * to generate the egress reset
2385 ssk->sk_lingertime = 0;
2386 sock_set_flag(ssk, SOCK_LINGER);
2387 subflow->send_fastclose = 1;
2390 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2392 tcp_disconnect(ssk, 0);
2393 msk->subflow->state = SS_UNCONNECTED;
2394 mptcp_subflow_ctx_reset(subflow);
2400 subflow->disposable = 1;
2402 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2403 * the ssk has been already destroyed, we just need to release the
2404 * reference owned by msk;
2406 if (!inet_csk(ssk)->icsk_ulp_ops) {
2407 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2408 kfree_rcu(subflow, rcu);
2410 /* otherwise tcp will dispose of the ssk and subflow ctx */
2411 if (ssk->sk_state == TCP_LISTEN) {
2412 tcp_set_state(ssk, TCP_CLOSE);
2413 mptcp_subflow_queue_clean(sk, ssk);
2414 inet_csk_listen_stop(ssk);
2415 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2418 __tcp_close(ssk, 0);
2420 /* close acquired an extra ref */
2429 if (ssk == msk->first)
2430 WRITE_ONCE(msk->first, NULL);
2433 if (ssk == msk->last_snd)
2434 msk->last_snd = NULL;
2437 __mptcp_push_pending(sk, 0);
2440 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2441 struct mptcp_subflow_context *subflow)
2443 if (sk->sk_state == TCP_ESTABLISHED)
2444 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2446 /* subflow aborted before reaching the fully_established status
2447 * attempt the creation of the next subflow
2449 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2451 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2454 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2459 static void __mptcp_close_subflow(struct sock *sk)
2461 struct mptcp_subflow_context *subflow, *tmp;
2462 struct mptcp_sock *msk = mptcp_sk(sk);
2466 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2467 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2469 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2472 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2473 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2476 mptcp_close_ssk(sk, ssk, subflow);
2481 static bool mptcp_should_close(const struct sock *sk)
2483 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2484 struct mptcp_subflow_context *subflow;
2486 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2489 /* if all subflows are in closed status don't bother with additional
2492 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2493 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2500 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2502 struct mptcp_subflow_context *subflow, *tmp;
2503 struct sock *sk = (struct sock *)msk;
2505 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2508 mptcp_token_destroy(msk);
2510 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2511 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2514 slow = lock_sock_fast(tcp_sk);
2515 if (tcp_sk->sk_state != TCP_CLOSE) {
2516 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2517 tcp_set_state(tcp_sk, TCP_CLOSE);
2519 unlock_sock_fast(tcp_sk, slow);
2522 /* Mirror the tcp_reset() error propagation */
2523 switch (sk->sk_state) {
2525 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2527 case TCP_CLOSE_WAIT:
2528 WRITE_ONCE(sk->sk_err, EPIPE);
2533 WRITE_ONCE(sk->sk_err, ECONNRESET);
2536 inet_sk_state_store(sk, TCP_CLOSE);
2537 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2538 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2539 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2541 /* the calling mptcp_worker will properly destroy the socket */
2542 if (sock_flag(sk, SOCK_DEAD))
2545 sk->sk_state_change(sk);
2546 sk_error_report(sk);
2549 static void __mptcp_retrans(struct sock *sk)
2551 struct mptcp_sock *msk = mptcp_sk(sk);
2552 struct mptcp_sendmsg_info info = {};
2553 struct mptcp_data_frag *dfrag;
2558 mptcp_clean_una_wakeup(sk);
2560 /* first check ssk: need to kick "stale" logic */
2561 ssk = mptcp_subflow_get_retrans(msk);
2562 dfrag = mptcp_rtx_head(sk);
2564 if (mptcp_data_fin_enabled(msk)) {
2565 struct inet_connection_sock *icsk = inet_csk(sk);
2567 icsk->icsk_retransmits++;
2568 mptcp_set_datafin_timeout(sk);
2569 mptcp_send_ack(msk);
2574 if (!mptcp_send_head(sk))
2585 /* limit retransmission to the bytes already sent on some subflows */
2587 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2588 while (info.sent < info.limit) {
2589 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2593 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2598 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2599 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2601 WRITE_ONCE(msk->allow_infinite_fallback, false);
2607 mptcp_check_and_set_pending(sk);
2609 if (!mptcp_timer_pending(sk))
2610 mptcp_reset_timer(sk);
2613 /* schedule the timeout timer for the relevant event: either close timeout
2614 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2616 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2618 struct sock *sk = (struct sock *)msk;
2619 unsigned long timeout, close_timeout;
2621 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2624 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2626 /* the close timeout takes precedence on the fail one, and here at least one of
2629 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2631 sk_reset_timer(sk, &sk->sk_timer, timeout);
2634 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2636 struct sock *ssk = msk->first;
2642 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2644 slow = lock_sock_fast(ssk);
2645 mptcp_subflow_reset(ssk);
2646 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2647 unlock_sock_fast(ssk, slow);
2649 mptcp_reset_timeout(msk, 0);
2652 static void mptcp_do_fastclose(struct sock *sk)
2654 struct mptcp_subflow_context *subflow, *tmp;
2655 struct mptcp_sock *msk = mptcp_sk(sk);
2657 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2658 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2659 subflow, MPTCP_CF_FASTCLOSE);
2662 static void mptcp_worker(struct work_struct *work)
2664 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2665 struct sock *sk = (struct sock *)msk;
2666 unsigned long fail_tout;
2670 state = sk->sk_state;
2671 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2674 mptcp_check_data_fin_ack(sk);
2676 mptcp_check_fastclose(msk);
2678 mptcp_pm_nl_work(msk);
2680 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2681 mptcp_check_for_eof(msk);
2683 __mptcp_check_send_data_fin(sk);
2684 mptcp_check_data_fin(sk);
2686 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2687 __mptcp_close_subflow(sk);
2689 /* There is no point in keeping around an orphaned sk timedout or
2690 * closed, but we need the msk around to reply to incoming DATA_FIN,
2691 * even if it is orphaned and in FIN_WAIT2 state
2693 if (sock_flag(sk, SOCK_DEAD)) {
2694 if (mptcp_should_close(sk)) {
2695 inet_sk_state_store(sk, TCP_CLOSE);
2696 mptcp_do_fastclose(sk);
2698 if (sk->sk_state == TCP_CLOSE) {
2699 __mptcp_destroy_sock(sk);
2704 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2705 __mptcp_retrans(sk);
2707 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2708 if (fail_tout && time_after(jiffies, fail_tout))
2709 mptcp_mp_fail_no_response(msk);
2716 static int __mptcp_init_sock(struct sock *sk)
2718 struct mptcp_sock *msk = mptcp_sk(sk);
2720 INIT_LIST_HEAD(&msk->conn_list);
2721 INIT_LIST_HEAD(&msk->join_list);
2722 INIT_LIST_HEAD(&msk->rtx_queue);
2723 INIT_WORK(&msk->work, mptcp_worker);
2724 __skb_queue_head_init(&msk->receive_queue);
2725 msk->out_of_order_queue = RB_ROOT;
2726 msk->first_pending = NULL;
2727 msk->rmem_fwd_alloc = 0;
2728 WRITE_ONCE(msk->rmem_released, 0);
2729 msk->timer_ival = TCP_RTO_MIN;
2731 WRITE_ONCE(msk->first, NULL);
2732 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2733 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2734 WRITE_ONCE(msk->allow_infinite_fallback, true);
2735 msk->recovery = false;
2737 mptcp_pm_data_init(msk);
2739 /* re-use the csk retrans timer for MPTCP-level retrans */
2740 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2741 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2746 static void mptcp_ca_reset(struct sock *sk)
2748 struct inet_connection_sock *icsk = inet_csk(sk);
2750 tcp_assign_congestion_control(sk);
2751 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2753 /* no need to keep a reference to the ops, the name will suffice */
2754 tcp_cleanup_congestion_control(sk);
2755 icsk->icsk_ca_ops = NULL;
2758 static int mptcp_init_sock(struct sock *sk)
2760 struct net *net = sock_net(sk);
2763 ret = __mptcp_init_sock(sk);
2767 if (!mptcp_is_enabled(net))
2768 return -ENOPROTOOPT;
2770 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2773 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2775 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2776 * propagate the correct value
2780 sk_sockets_allocated_inc(sk);
2781 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2782 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2787 static void __mptcp_clear_xmit(struct sock *sk)
2789 struct mptcp_sock *msk = mptcp_sk(sk);
2790 struct mptcp_data_frag *dtmp, *dfrag;
2792 WRITE_ONCE(msk->first_pending, NULL);
2793 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2794 dfrag_clear(sk, dfrag);
2797 void mptcp_cancel_work(struct sock *sk)
2799 struct mptcp_sock *msk = mptcp_sk(sk);
2801 if (cancel_work_sync(&msk->work))
2805 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2809 switch (ssk->sk_state) {
2811 if (!(how & RCV_SHUTDOWN))
2815 tcp_disconnect(ssk, O_NONBLOCK);
2818 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2819 pr_debug("Fallback");
2820 ssk->sk_shutdown |= how;
2821 tcp_shutdown(ssk, how);
2823 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2825 if (!mptcp_timer_pending(sk))
2826 mptcp_reset_timer(sk);
2834 static const unsigned char new_state[16] = {
2835 /* current state: new state: action: */
2836 [0 /* (Invalid) */] = TCP_CLOSE,
2837 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2838 [TCP_SYN_SENT] = TCP_CLOSE,
2839 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2840 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2841 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2842 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2843 [TCP_CLOSE] = TCP_CLOSE,
2844 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2845 [TCP_LAST_ACK] = TCP_LAST_ACK,
2846 [TCP_LISTEN] = TCP_CLOSE,
2847 [TCP_CLOSING] = TCP_CLOSING,
2848 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2851 static int mptcp_close_state(struct sock *sk)
2853 int next = (int)new_state[sk->sk_state];
2854 int ns = next & TCP_STATE_MASK;
2856 inet_sk_state_store(sk, ns);
2858 return next & TCP_ACTION_FIN;
2861 static void __mptcp_check_send_data_fin(struct sock *sk)
2863 struct mptcp_subflow_context *subflow;
2864 struct mptcp_sock *msk = mptcp_sk(sk);
2866 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2867 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2868 msk->snd_nxt, msk->write_seq);
2870 /* we still need to enqueue subflows or not really shutting down,
2873 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2874 mptcp_send_head(sk))
2877 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2879 /* fallback socket will not get data_fin/ack, can move to the next
2882 if (__mptcp_check_fallback(msk)) {
2883 WRITE_ONCE(msk->snd_una, msk->write_seq);
2884 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2885 inet_sk_state_store(sk, TCP_CLOSE);
2886 mptcp_close_wake_up(sk);
2887 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2888 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2892 mptcp_for_each_subflow(msk, subflow) {
2893 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2895 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2899 static void __mptcp_wr_shutdown(struct sock *sk)
2901 struct mptcp_sock *msk = mptcp_sk(sk);
2903 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2904 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2905 !!mptcp_send_head(sk));
2907 /* will be ignored by fallback sockets */
2908 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2909 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2911 __mptcp_check_send_data_fin(sk);
2914 static void __mptcp_destroy_sock(struct sock *sk)
2916 struct mptcp_sock *msk = mptcp_sk(sk);
2918 pr_debug("msk=%p", msk);
2922 mptcp_stop_timer(sk);
2923 sk_stop_timer(sk, &sk->sk_timer);
2926 sk->sk_prot->destroy(sk);
2928 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2929 WARN_ON_ONCE(msk->rmem_released);
2930 sk_stream_kill_queues(sk);
2931 xfrm_sk_free_policy(sk);
2936 void __mptcp_unaccepted_force_close(struct sock *sk)
2938 sock_set_flag(sk, SOCK_DEAD);
2939 inet_sk_state_store(sk, TCP_CLOSE);
2940 mptcp_do_fastclose(sk);
2941 __mptcp_destroy_sock(sk);
2944 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2946 /* Concurrent splices from sk_receive_queue into receive_queue will
2947 * always show at least one non-empty queue when checked in this order.
2949 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2950 skb_queue_empty_lockless(&msk->receive_queue))
2953 return EPOLLIN | EPOLLRDNORM;
2956 static void mptcp_listen_inuse_dec(struct sock *sk)
2958 if (inet_sk_state_load(sk) == TCP_LISTEN)
2959 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2962 bool __mptcp_close(struct sock *sk, long timeout)
2964 struct mptcp_subflow_context *subflow;
2965 struct mptcp_sock *msk = mptcp_sk(sk);
2966 bool do_cancel_work = false;
2967 int subflows_alive = 0;
2969 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2971 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2972 mptcp_listen_inuse_dec(sk);
2973 inet_sk_state_store(sk, TCP_CLOSE);
2977 if (mptcp_check_readable(msk) || timeout < 0) {
2978 /* If the msk has read data, or the caller explicitly ask it,
2979 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
2981 inet_sk_state_store(sk, TCP_CLOSE);
2982 mptcp_do_fastclose(sk);
2984 } else if (mptcp_close_state(sk)) {
2985 __mptcp_wr_shutdown(sk);
2988 sk_stream_wait_close(sk, timeout);
2991 /* orphan all the subflows */
2992 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2993 mptcp_for_each_subflow(msk, subflow) {
2994 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2995 bool slow = lock_sock_fast_nested(ssk);
2997 subflows_alive += ssk->sk_state != TCP_CLOSE;
2999 /* since the close timeout takes precedence on the fail one,
3002 if (ssk == msk->first)
3003 subflow->fail_tout = 0;
3005 /* detach from the parent socket, but allow data_ready to
3006 * push incoming data into the mptcp stack, to properly ack it
3008 ssk->sk_socket = NULL;
3010 unlock_sock_fast(ssk, slow);
3014 /* all the subflows are closed, only timeout can change the msk
3015 * state, let's not keep resources busy for no reasons
3017 if (subflows_alive == 0)
3018 inet_sk_state_store(sk, TCP_CLOSE);
3021 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3023 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3025 if (sk->sk_state == TCP_CLOSE) {
3026 __mptcp_destroy_sock(sk);
3027 do_cancel_work = true;
3029 mptcp_reset_timeout(msk, 0);
3032 return do_cancel_work;
3035 static void mptcp_close(struct sock *sk, long timeout)
3037 bool do_cancel_work;
3041 do_cancel_work = __mptcp_close(sk, timeout);
3044 mptcp_cancel_work(sk);
3049 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3051 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3052 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3053 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3055 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3056 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3059 msk6->saddr = ssk6->saddr;
3060 msk6->flow_label = ssk6->flow_label;
3064 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3065 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3066 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3067 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3068 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3069 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3072 static int mptcp_disconnect(struct sock *sk, int flags)
3074 struct mptcp_sock *msk = mptcp_sk(sk);
3076 /* We are on the fastopen error path. We can't call straight into the
3077 * subflows cleanup code due to lock nesting (we are already under
3078 * msk->firstsocket lock). Do nothing and leave the cleanup to the
3081 if (msk->fastopening)
3084 mptcp_listen_inuse_dec(sk);
3085 inet_sk_state_store(sk, TCP_CLOSE);
3087 mptcp_stop_timer(sk);
3088 sk_stop_timer(sk, &sk->sk_timer);
3091 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3093 /* msk->subflow is still intact, the following will not free the first
3096 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3097 msk->last_snd = NULL;
3098 WRITE_ONCE(msk->flags, 0);
3100 msk->push_pending = 0;
3101 msk->recovery = false;
3102 msk->can_ack = false;
3103 msk->fully_established = false;
3104 msk->rcv_data_fin = false;
3105 msk->snd_data_fin_enable = false;
3106 msk->rcv_fastclose = false;
3107 msk->use_64bit_ack = false;
3108 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3109 mptcp_pm_data_reset(msk);
3112 WRITE_ONCE(sk->sk_shutdown, 0);
3113 sk_error_report(sk);
3117 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3118 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3120 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3122 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3126 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3127 const struct mptcp_options_received *mp_opt,
3129 struct request_sock *req)
3131 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3132 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3133 struct mptcp_sock *msk;
3138 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3139 if (nsk->sk_family == AF_INET6)
3140 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3143 __mptcp_init_sock(nsk);
3145 msk = mptcp_sk(nsk);
3146 msk->local_key = subflow_req->local_key;
3147 msk->token = subflow_req->token;
3148 WRITE_ONCE(msk->subflow, NULL);
3149 msk->in_accept_queue = 1;
3150 WRITE_ONCE(msk->fully_established, false);
3151 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3152 WRITE_ONCE(msk->csum_enabled, true);
3154 msk->write_seq = subflow_req->idsn + 1;
3155 msk->snd_nxt = msk->write_seq;
3156 msk->snd_una = msk->write_seq;
3157 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3158 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3160 sock_reset_flag(nsk, SOCK_RCU_FREE);
3161 security_inet_csk_clone(nsk, req);
3163 /* this can't race with mptcp_close(), as the msk is
3164 * not yet exposted to user-space
3166 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3168 /* The msk maintain a ref to each subflow in the connections list */
3169 WRITE_ONCE(msk->first, ssk);
3170 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3173 /* new mpc subflow takes ownership of the newly
3174 * created mptcp socket
3176 mptcp_token_accept(subflow_req, msk);
3178 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3179 * uses the correct data
3181 mptcp_copy_inaddrs(nsk, ssk);
3182 mptcp_propagate_sndbuf(nsk, ssk);
3184 mptcp_rcv_space_init(msk, ssk);
3185 bh_unlock_sock(nsk);
3187 /* note: the newly allocated socket refcount is 2 now */
3191 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3193 const struct tcp_sock *tp = tcp_sk(ssk);
3195 msk->rcvq_space.copied = 0;
3196 msk->rcvq_space.rtt_us = 0;
3198 msk->rcvq_space.time = tp->tcp_mstamp;
3200 /* initial rcv_space offering made to peer */
3201 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3202 TCP_INIT_CWND * tp->advmss);
3203 if (msk->rcvq_space.space == 0)
3204 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3206 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3209 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3212 struct mptcp_sock *msk = mptcp_sk(sk);
3213 struct socket *listener;
3216 listener = READ_ONCE(msk->subflow);
3217 if (WARN_ON_ONCE(!listener)) {
3222 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3223 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3227 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3228 if (sk_is_mptcp(newsk)) {
3229 struct mptcp_subflow_context *subflow;
3230 struct sock *new_mptcp_sock;
3232 subflow = mptcp_subflow_ctx(newsk);
3233 new_mptcp_sock = subflow->conn;
3235 /* is_mptcp should be false if subflow->conn is missing, see
3236 * subflow_syn_recv_sock()
3238 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3239 tcp_sk(newsk)->is_mptcp = 0;
3243 newsk = new_mptcp_sock;
3244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3246 MPTCP_INC_STATS(sock_net(sk),
3247 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3251 newsk->sk_kern_sock = kern;
3255 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3257 struct mptcp_subflow_context *subflow, *tmp;
3258 struct sock *sk = (struct sock *)msk;
3260 __mptcp_clear_xmit(sk);
3262 /* join list will be eventually flushed (with rst) at sock lock release time */
3263 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3264 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3266 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3267 mptcp_data_lock(sk);
3268 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3269 __skb_queue_purge(&sk->sk_receive_queue);
3270 skb_rbtree_purge(&msk->out_of_order_queue);
3271 mptcp_data_unlock(sk);
3273 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3274 * inet_sock_destruct() will dispose it
3276 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3277 msk->rmem_fwd_alloc = 0;
3278 mptcp_token_destroy(msk);
3279 mptcp_pm_free_anno_list(msk);
3280 mptcp_free_local_addr_list(msk);
3283 static void mptcp_destroy(struct sock *sk)
3285 struct mptcp_sock *msk = mptcp_sk(sk);
3287 /* clears msk->subflow, allowing the following to close
3288 * even the initial subflow
3290 mptcp_dispose_initial_subflow(msk);
3291 mptcp_destroy_common(msk, 0);
3292 sk_sockets_allocated_dec(sk);
3295 void __mptcp_data_acked(struct sock *sk)
3297 if (!sock_owned_by_user(sk))
3298 __mptcp_clean_una(sk);
3300 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3302 if (mptcp_pending_data_fin_ack(sk))
3303 mptcp_schedule_work(sk);
3306 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3308 if (!mptcp_send_head(sk))
3311 if (!sock_owned_by_user(sk))
3312 __mptcp_subflow_push_pending(sk, ssk, false);
3314 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3317 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3318 BIT(MPTCP_RETRANSMIT) | \
3319 BIT(MPTCP_FLUSH_JOIN_LIST))
3321 /* processes deferred events and flush wmem */
3322 static void mptcp_release_cb(struct sock *sk)
3323 __must_hold(&sk->sk_lock.slock)
3325 struct mptcp_sock *msk = mptcp_sk(sk);
3328 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3333 /* the following actions acquire the subflow socket lock
3335 * 1) can't be invoked in atomic scope
3336 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3337 * datapath acquires the msk socket spinlock while helding
3338 * the subflow socket lock
3340 msk->push_pending = 0;
3341 msk->cb_flags &= ~flags;
3342 spin_unlock_bh(&sk->sk_lock.slock);
3343 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3344 __mptcp_flush_join_list(sk);
3345 if (flags & BIT(MPTCP_PUSH_PENDING))
3346 __mptcp_push_pending(sk, 0);
3347 if (flags & BIT(MPTCP_RETRANSMIT))
3348 __mptcp_retrans(sk);
3351 spin_lock_bh(&sk->sk_lock.slock);
3354 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3355 __mptcp_clean_una_wakeup(sk);
3356 if (unlikely(&msk->cb_flags)) {
3357 /* be sure to set the current sk state before tacking actions
3358 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3360 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3361 __mptcp_set_connected(sk);
3362 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3363 __mptcp_error_report(sk);
3364 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3365 msk->last_snd = NULL;
3368 __mptcp_update_rmem(sk);
3371 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3372 * TCP can't schedule delack timer before the subflow is fully established.
3373 * MPTCP uses the delack timer to do 3rd ack retransmissions
3375 static void schedule_3rdack_retransmission(struct sock *ssk)
3377 struct inet_connection_sock *icsk = inet_csk(ssk);
3378 struct tcp_sock *tp = tcp_sk(ssk);
3379 unsigned long timeout;
3381 if (mptcp_subflow_ctx(ssk)->fully_established)
3384 /* reschedule with a timeout above RTT, as we must look only for drop */
3386 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3388 timeout = TCP_TIMEOUT_INIT;
3391 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3392 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3393 icsk->icsk_ack.timeout = timeout;
3394 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3397 void mptcp_subflow_process_delegated(struct sock *ssk)
3399 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3400 struct sock *sk = subflow->conn;
3402 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3403 mptcp_data_lock(sk);
3404 if (!sock_owned_by_user(sk))
3405 __mptcp_subflow_push_pending(sk, ssk, true);
3407 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3408 mptcp_data_unlock(sk);
3409 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3411 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3412 schedule_3rdack_retransmission(ssk);
3413 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3417 static int mptcp_hash(struct sock *sk)
3419 /* should never be called,
3420 * we hash the TCP subflows not the master socket
3426 static void mptcp_unhash(struct sock *sk)
3428 /* called from sk_common_release(), but nothing to do here */
3431 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3433 struct mptcp_sock *msk = mptcp_sk(sk);
3434 struct socket *ssock;
3436 ssock = msk->subflow;
3437 pr_debug("msk=%p, subflow=%p", msk, ssock);
3438 if (WARN_ON_ONCE(!ssock))
3441 return inet_csk_get_port(ssock->sk, snum);
3444 void mptcp_finish_connect(struct sock *ssk)
3446 struct mptcp_subflow_context *subflow;
3447 struct mptcp_sock *msk;
3450 subflow = mptcp_subflow_ctx(ssk);
3454 pr_debug("msk=%p, token=%u", sk, subflow->token);
3456 subflow->map_seq = subflow->iasn;
3457 subflow->map_subflow_seq = 1;
3459 /* the socket is not connected yet, no msk/subflow ops can access/race
3460 * accessing the field below
3462 WRITE_ONCE(msk->local_key, subflow->local_key);
3463 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3464 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3465 WRITE_ONCE(msk->snd_una, msk->write_seq);
3467 mptcp_pm_new_connection(msk, ssk, 0);
3469 mptcp_rcv_space_init(msk, ssk);
3472 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3474 write_lock_bh(&sk->sk_callback_lock);
3475 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3476 sk_set_socket(sk, parent);
3477 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3478 write_unlock_bh(&sk->sk_callback_lock);
3481 bool mptcp_finish_join(struct sock *ssk)
3483 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3484 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3485 struct sock *parent = (void *)msk;
3488 pr_debug("msk=%p, subflow=%p", msk, subflow);
3490 /* mptcp socket already closing? */
3491 if (!mptcp_is_fully_established(parent)) {
3492 subflow->reset_reason = MPTCP_RST_EMPTCP;
3496 /* active subflow, already present inside the conn_list */
3497 if (!list_empty(&subflow->node)) {
3498 mptcp_subflow_joined(msk, ssk);
3502 if (!mptcp_pm_allow_new_subflow(msk))
3503 goto err_prohibited;
3505 /* If we can't acquire msk socket lock here, let the release callback
3508 mptcp_data_lock(parent);
3509 if (!sock_owned_by_user(parent)) {
3510 ret = __mptcp_finish_join(msk, ssk);
3513 list_add_tail(&subflow->node, &msk->conn_list);
3517 list_add_tail(&subflow->node, &msk->join_list);
3518 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3520 mptcp_data_unlock(parent);
3524 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3531 static void mptcp_shutdown(struct sock *sk, int how)
3533 pr_debug("sk=%p, how=%d", sk, how);
3535 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3536 __mptcp_wr_shutdown(sk);
3539 static int mptcp_forward_alloc_get(const struct sock *sk)
3541 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3544 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3546 const struct sock *sk = (void *)msk;
3549 if (sk->sk_state == TCP_LISTEN)
3552 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3555 delta = msk->write_seq - v;
3556 if (__mptcp_check_fallback(msk) && msk->first) {
3557 struct tcp_sock *tp = tcp_sk(msk->first);
3559 /* the first subflow is disconnected after close - see
3560 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3561 * so ignore that status, too.
3563 if (!((1 << msk->first->sk_state) &
3564 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3565 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3567 if (delta > INT_MAX)
3573 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3575 struct mptcp_sock *msk = mptcp_sk(sk);
3581 if (sk->sk_state == TCP_LISTEN)
3585 __mptcp_move_skbs(msk);
3586 answ = mptcp_inq_hint(sk);
3590 slow = lock_sock_fast(sk);
3591 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3592 unlock_sock_fast(sk, slow);
3595 slow = lock_sock_fast(sk);
3596 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3597 unlock_sock_fast(sk, slow);
3600 return -ENOIOCTLCMD;
3603 return put_user(answ, (int __user *)arg);
3606 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3607 struct mptcp_subflow_context *subflow)
3609 subflow->request_mptcp = 0;
3610 __mptcp_do_fallback(msk);
3613 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3615 struct mptcp_subflow_context *subflow;
3616 struct mptcp_sock *msk = mptcp_sk(sk);
3617 struct socket *ssock;
3620 ssock = __mptcp_nmpc_socket(msk);
3622 return PTR_ERR(ssock);
3624 mptcp_token_destroy(msk);
3625 inet_sk_state_store(sk, TCP_SYN_SENT);
3626 subflow = mptcp_subflow_ctx(ssock->sk);
3627 #ifdef CONFIG_TCP_MD5SIG
3628 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3631 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3632 mptcp_subflow_early_fallback(msk, subflow);
3634 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3635 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3636 mptcp_subflow_early_fallback(msk, subflow);
3638 if (likely(!__mptcp_check_fallback(msk)))
3639 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3641 /* if reaching here via the fastopen/sendmsg path, the caller already
3642 * acquired the subflow socket lock, too.
3644 if (msk->fastopening)
3645 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3647 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3648 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3650 /* on successful connect, the msk state will be moved to established by
3651 * subflow_finish_connect()
3653 if (unlikely(err && err != -EINPROGRESS)) {
3654 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3658 mptcp_copy_inaddrs(sk, ssock->sk);
3660 /* silence EINPROGRESS and let the caller inet_stream_connect
3661 * handle the connection in progress
3666 static struct proto mptcp_prot = {
3668 .owner = THIS_MODULE,
3669 .init = mptcp_init_sock,
3670 .connect = mptcp_connect,
3671 .disconnect = mptcp_disconnect,
3672 .close = mptcp_close,
3673 .accept = mptcp_accept,
3674 .setsockopt = mptcp_setsockopt,
3675 .getsockopt = mptcp_getsockopt,
3676 .shutdown = mptcp_shutdown,
3677 .destroy = mptcp_destroy,
3678 .sendmsg = mptcp_sendmsg,
3679 .ioctl = mptcp_ioctl,
3680 .recvmsg = mptcp_recvmsg,
3681 .release_cb = mptcp_release_cb,
3683 .unhash = mptcp_unhash,
3684 .get_port = mptcp_get_port,
3685 .forward_alloc_get = mptcp_forward_alloc_get,
3686 .sockets_allocated = &mptcp_sockets_allocated,
3688 .memory_allocated = &tcp_memory_allocated,
3689 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3691 .memory_pressure = &tcp_memory_pressure,
3692 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3693 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3694 .sysctl_mem = sysctl_tcp_mem,
3695 .obj_size = sizeof(struct mptcp_sock),
3696 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3697 .no_autobind = true,
3700 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3702 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3703 struct socket *ssock;
3706 lock_sock(sock->sk);
3707 ssock = __mptcp_nmpc_socket(msk);
3708 if (IS_ERR(ssock)) {
3709 err = PTR_ERR(ssock);
3713 err = ssock->ops->bind(ssock, uaddr, addr_len);
3715 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3718 release_sock(sock->sk);
3722 static int mptcp_listen(struct socket *sock, int backlog)
3724 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3725 struct sock *sk = sock->sk;
3726 struct socket *ssock;
3729 pr_debug("msk=%p", msk);
3732 ssock = __mptcp_nmpc_socket(msk);
3733 if (IS_ERR(ssock)) {
3734 err = PTR_ERR(ssock);
3738 mptcp_token_destroy(msk);
3739 inet_sk_state_store(sk, TCP_LISTEN);
3740 sock_set_flag(sk, SOCK_RCU_FREE);
3742 err = ssock->ops->listen(ssock, backlog);
3743 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3745 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3746 mptcp_copy_inaddrs(sk, ssock->sk);
3749 mptcp_event_pm_listener(ssock->sk, MPTCP_EVENT_LISTENER_CREATED);
3756 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3757 int flags, bool kern)
3759 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3760 struct socket *ssock;
3763 pr_debug("msk=%p", msk);
3765 /* Buggy applications can call accept on socket states other then LISTEN
3766 * but no need to allocate the first subflow just to error out.
3768 ssock = READ_ONCE(msk->subflow);
3772 err = ssock->ops->accept(sock, newsock, flags, kern);
3773 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3774 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3775 struct mptcp_subflow_context *subflow;
3776 struct sock *newsk = newsock->sk;
3778 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3779 msk->in_accept_queue = 0;
3783 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3784 * This is needed so NOSPACE flag can be set from tcp stack.
3786 mptcp_for_each_subflow(msk, subflow) {
3787 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3789 if (!ssk->sk_socket)
3790 mptcp_sock_graft(ssk, newsock);
3793 /* Do late cleanup for the first subflow as necessary. Also
3794 * deal with bad peers not doing a complete shutdown.
3797 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3798 __mptcp_close_ssk(newsk, msk->first,
3799 mptcp_subflow_ctx(msk->first), 0);
3800 if (unlikely(list_empty(&msk->conn_list)))
3801 inet_sk_state_store(newsk, TCP_CLOSE);
3804 release_sock(newsk);
3810 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3812 struct sock *sk = (struct sock *)msk;
3814 if (sk_stream_is_writeable(sk))
3815 return EPOLLOUT | EPOLLWRNORM;
3817 mptcp_set_nospace(sk);
3818 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3819 if (sk_stream_is_writeable(sk))
3820 return EPOLLOUT | EPOLLWRNORM;
3825 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3826 struct poll_table_struct *wait)
3828 struct sock *sk = sock->sk;
3829 struct mptcp_sock *msk;
3835 sock_poll_wait(file, sock, wait);
3837 state = inet_sk_state_load(sk);
3838 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3839 if (state == TCP_LISTEN) {
3840 struct socket *ssock = READ_ONCE(msk->subflow);
3842 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3845 return inet_csk_listen_poll(ssock->sk);
3848 shutdown = READ_ONCE(sk->sk_shutdown);
3849 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3851 if (shutdown & RCV_SHUTDOWN)
3852 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3854 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3855 mask |= mptcp_check_readable(msk);
3856 if (shutdown & SEND_SHUTDOWN)
3857 mask |= EPOLLOUT | EPOLLWRNORM;
3859 mask |= mptcp_check_writeable(msk);
3860 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3861 /* cf tcp_poll() note about TFO */
3862 mask |= EPOLLOUT | EPOLLWRNORM;
3865 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3867 if (READ_ONCE(sk->sk_err))
3873 static const struct proto_ops mptcp_stream_ops = {
3875 .owner = THIS_MODULE,
3876 .release = inet_release,
3878 .connect = inet_stream_connect,
3879 .socketpair = sock_no_socketpair,
3880 .accept = mptcp_stream_accept,
3881 .getname = inet_getname,
3883 .ioctl = inet_ioctl,
3884 .gettstamp = sock_gettstamp,
3885 .listen = mptcp_listen,
3886 .shutdown = inet_shutdown,
3887 .setsockopt = sock_common_setsockopt,
3888 .getsockopt = sock_common_getsockopt,
3889 .sendmsg = inet_sendmsg,
3890 .recvmsg = inet_recvmsg,
3891 .mmap = sock_no_mmap,
3892 .sendpage = inet_sendpage,
3895 static struct inet_protosw mptcp_protosw = {
3896 .type = SOCK_STREAM,
3897 .protocol = IPPROTO_MPTCP,
3898 .prot = &mptcp_prot,
3899 .ops = &mptcp_stream_ops,
3900 .flags = INET_PROTOSW_ICSK,
3903 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3905 struct mptcp_delegated_action *delegated;
3906 struct mptcp_subflow_context *subflow;
3909 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3910 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3911 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3913 bh_lock_sock_nested(ssk);
3914 if (!sock_owned_by_user(ssk) &&
3915 mptcp_subflow_has_delegated_action(subflow))
3916 mptcp_subflow_process_delegated(ssk);
3917 /* ... elsewhere tcp_release_cb_override already processed
3918 * the action or will do at next release_sock().
3919 * In both case must dequeue the subflow here - on the same
3920 * CPU that scheduled it.
3922 bh_unlock_sock(ssk);
3925 if (++work_done == budget)
3929 /* always provide a 0 'work_done' argument, so that napi_complete_done
3930 * will not try accessing the NULL napi->dev ptr
3932 napi_complete_done(napi, 0);
3936 void __init mptcp_proto_init(void)
3938 struct mptcp_delegated_action *delegated;
3941 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3943 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3944 panic("Failed to allocate MPTCP pcpu counter\n");
3946 init_dummy_netdev(&mptcp_napi_dev);
3947 for_each_possible_cpu(cpu) {
3948 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3949 INIT_LIST_HEAD(&delegated->head);
3950 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3952 napi_enable(&delegated->napi);
3955 mptcp_subflow_init();
3959 if (proto_register(&mptcp_prot, 1) != 0)
3960 panic("Failed to register MPTCP proto.\n");
3962 inet_register_protosw(&mptcp_protosw);
3964 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3967 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3968 static const struct proto_ops mptcp_v6_stream_ops = {
3970 .owner = THIS_MODULE,
3971 .release = inet6_release,
3973 .connect = inet_stream_connect,
3974 .socketpair = sock_no_socketpair,
3975 .accept = mptcp_stream_accept,
3976 .getname = inet6_getname,
3978 .ioctl = inet6_ioctl,
3979 .gettstamp = sock_gettstamp,
3980 .listen = mptcp_listen,
3981 .shutdown = inet_shutdown,
3982 .setsockopt = sock_common_setsockopt,
3983 .getsockopt = sock_common_getsockopt,
3984 .sendmsg = inet6_sendmsg,
3985 .recvmsg = inet6_recvmsg,
3986 .mmap = sock_no_mmap,
3987 .sendpage = inet_sendpage,
3988 #ifdef CONFIG_COMPAT
3989 .compat_ioctl = inet6_compat_ioctl,
3993 static struct proto mptcp_v6_prot;
3995 static struct inet_protosw mptcp_v6_protosw = {
3996 .type = SOCK_STREAM,
3997 .protocol = IPPROTO_MPTCP,
3998 .prot = &mptcp_v6_prot,
3999 .ops = &mptcp_v6_stream_ops,
4000 .flags = INET_PROTOSW_ICSK,
4003 int __init mptcp_proto_v6_init(void)
4007 mptcp_v6_prot = mptcp_prot;
4008 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4009 mptcp_v6_prot.slab = NULL;
4010 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4012 err = proto_register(&mptcp_v6_prot, 1);
4016 err = inet6_register_protosw(&mptcp_v6_protosw);
4018 proto_unregister(&mptcp_v6_prot);