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 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
53 * completed yet or has failed, return the subflow socket.
54 * Otherwise return NULL.
56 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
58 if (!msk->subflow || READ_ONCE(msk->can_ack))
64 /* Returns end sequence number of the receiver's advertised window */
65 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
67 return READ_ONCE(msk->wnd_end);
70 static bool mptcp_is_tcpsk(struct sock *sk)
72 struct socket *sock = sk->sk_socket;
74 if (unlikely(sk->sk_prot == &tcp_prot)) {
75 /* we are being invoked after mptcp_accept() has
76 * accepted a non-mp-capable flow: sk is a tcp_sk,
79 * Hand the socket over to tcp so all further socket ops
82 sock->ops = &inet_stream_ops;
84 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
85 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
86 sock->ops = &inet6_stream_ops;
94 static int __mptcp_socket_create(struct mptcp_sock *msk)
96 struct mptcp_subflow_context *subflow;
97 struct sock *sk = (struct sock *)msk;
101 err = mptcp_subflow_create_socket(sk, &ssock);
105 msk->first = ssock->sk;
106 msk->subflow = ssock;
107 subflow = mptcp_subflow_ctx(ssock->sk);
108 list_add(&subflow->node, &msk->conn_list);
109 sock_hold(ssock->sk);
110 subflow->request_mptcp = 1;
112 /* This is the first subflow, always with id 0 */
113 subflow->local_id_valid = 1;
114 mptcp_sock_graft(msk->first, sk->sk_socket);
119 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
121 sk_drops_add(sk, skb);
125 static void mptcp_rmem_charge(struct sock *sk, int size)
127 mptcp_sk(sk)->rmem_fwd_alloc -= size;
130 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
131 struct sk_buff *from)
136 if (MPTCP_SKB_CB(from)->offset ||
137 !skb_try_coalesce(to, from, &fragstolen, &delta))
140 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
141 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
142 to->len, MPTCP_SKB_CB(from)->end_seq);
143 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
145 /* note the fwd memory can reach a negative value after accounting
146 * for the delta, but the later skb free will restore a non
149 atomic_add(delta, &sk->sk_rmem_alloc);
150 mptcp_rmem_charge(sk, delta);
151 kfree_skb_partial(from, fragstolen);
156 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
157 struct sk_buff *from)
159 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
162 return mptcp_try_coalesce((struct sock *)msk, to, from);
165 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
167 amount >>= PAGE_SHIFT;
168 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
169 __sk_mem_reduce_allocated(sk, amount);
172 static void mptcp_rmem_uncharge(struct sock *sk, int size)
174 struct mptcp_sock *msk = mptcp_sk(sk);
177 msk->rmem_fwd_alloc += size;
178 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
180 /* see sk_mem_uncharge() for the rationale behind the following schema */
181 if (unlikely(reclaimable >= PAGE_SIZE))
182 __mptcp_rmem_reclaim(sk, reclaimable);
185 static void mptcp_rfree(struct sk_buff *skb)
187 unsigned int len = skb->truesize;
188 struct sock *sk = skb->sk;
190 atomic_sub(len, &sk->sk_rmem_alloc);
191 mptcp_rmem_uncharge(sk, len);
194 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
198 skb->destructor = mptcp_rfree;
199 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
200 mptcp_rmem_charge(sk, skb->truesize);
203 /* "inspired" by tcp_data_queue_ofo(), main differences:
205 * - don't cope with sacks
207 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
209 struct sock *sk = (struct sock *)msk;
210 struct rb_node **p, *parent;
211 u64 seq, end_seq, max_seq;
212 struct sk_buff *skb1;
214 seq = MPTCP_SKB_CB(skb)->map_seq;
215 end_seq = MPTCP_SKB_CB(skb)->end_seq;
216 max_seq = atomic64_read(&msk->rcv_wnd_sent);
218 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
219 RB_EMPTY_ROOT(&msk->out_of_order_queue));
220 if (after64(end_seq, max_seq)) {
223 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
224 (unsigned long long)end_seq - (unsigned long)max_seq,
225 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
226 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
230 p = &msk->out_of_order_queue.rb_node;
231 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
232 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
233 rb_link_node(&skb->rbnode, NULL, p);
234 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
235 msk->ooo_last_skb = skb;
239 /* with 2 subflows, adding at end of ooo queue is quite likely
240 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
242 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
243 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
248 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
249 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
250 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
251 parent = &msk->ooo_last_skb->rbnode;
252 p = &parent->rb_right;
256 /* Find place to insert this segment. Handle overlaps on the way. */
260 skb1 = rb_to_skb(parent);
261 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
262 p = &parent->rb_left;
265 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
266 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
267 /* All the bits are present. Drop. */
269 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
272 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
276 * continue traversing
279 /* skb's seq == skb1's seq and skb covers skb1.
280 * Replace skb1 with skb.
282 rb_replace_node(&skb1->rbnode, &skb->rbnode,
283 &msk->out_of_order_queue);
284 mptcp_drop(sk, skb1);
285 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
288 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
289 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
292 p = &parent->rb_right;
296 /* Insert segment into RB tree. */
297 rb_link_node(&skb->rbnode, parent, p);
298 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
301 /* Remove other segments covered by skb. */
302 while ((skb1 = skb_rb_next(skb)) != NULL) {
303 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
305 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
306 mptcp_drop(sk, skb1);
307 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
309 /* If there is no skb after us, we are the last_skb ! */
311 msk->ooo_last_skb = skb;
315 mptcp_set_owner_r(skb, sk);
318 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
320 struct mptcp_sock *msk = mptcp_sk(sk);
323 if (size <= msk->rmem_fwd_alloc)
326 size -= msk->rmem_fwd_alloc;
327 amt = sk_mem_pages(size);
328 amount = amt << PAGE_SHIFT;
329 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
332 msk->rmem_fwd_alloc += amount;
336 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
337 struct sk_buff *skb, unsigned int offset,
340 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
341 struct sock *sk = (struct sock *)msk;
342 struct sk_buff *tail;
345 __skb_unlink(skb, &ssk->sk_receive_queue);
350 /* try to fetch required memory from subflow */
351 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
354 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
356 /* the skb map_seq accounts for the skb offset:
357 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
360 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
361 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
362 MPTCP_SKB_CB(skb)->offset = offset;
363 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
365 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
367 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
368 tail = skb_peek_tail(&sk->sk_receive_queue);
369 if (tail && mptcp_try_coalesce(sk, tail, skb))
372 mptcp_set_owner_r(skb, sk);
373 __skb_queue_tail(&sk->sk_receive_queue, skb);
375 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
376 mptcp_data_queue_ofo(msk, skb);
380 /* old data, keep it simple and drop the whole pkt, sender
381 * will retransmit as needed, if needed.
383 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
389 static void mptcp_stop_timer(struct sock *sk)
391 struct inet_connection_sock *icsk = inet_csk(sk);
393 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
394 mptcp_sk(sk)->timer_ival = 0;
397 static void mptcp_close_wake_up(struct sock *sk)
399 if (sock_flag(sk, SOCK_DEAD))
402 sk->sk_state_change(sk);
403 if (sk->sk_shutdown == SHUTDOWN_MASK ||
404 sk->sk_state == TCP_CLOSE)
405 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
407 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
410 static bool mptcp_pending_data_fin_ack(struct sock *sk)
412 struct mptcp_sock *msk = mptcp_sk(sk);
414 return !__mptcp_check_fallback(msk) &&
415 ((1 << sk->sk_state) &
416 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
417 msk->write_seq == READ_ONCE(msk->snd_una);
420 static void mptcp_check_data_fin_ack(struct sock *sk)
422 struct mptcp_sock *msk = mptcp_sk(sk);
424 /* Look for an acknowledged DATA_FIN */
425 if (mptcp_pending_data_fin_ack(sk)) {
426 WRITE_ONCE(msk->snd_data_fin_enable, 0);
428 switch (sk->sk_state) {
430 inet_sk_state_store(sk, TCP_FIN_WAIT2);
434 inet_sk_state_store(sk, TCP_CLOSE);
438 mptcp_close_wake_up(sk);
442 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
444 struct mptcp_sock *msk = mptcp_sk(sk);
446 if (READ_ONCE(msk->rcv_data_fin) &&
447 ((1 << sk->sk_state) &
448 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
449 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
451 if (msk->ack_seq == rcv_data_fin_seq) {
453 *seq = rcv_data_fin_seq;
462 static void mptcp_set_datafin_timeout(const struct sock *sk)
464 struct inet_connection_sock *icsk = inet_csk(sk);
467 retransmits = min_t(u32, icsk->icsk_retransmits,
468 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
470 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
473 static void __mptcp_set_timeout(struct sock *sk, long tout)
475 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
478 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
480 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
482 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
483 inet_csk(ssk)->icsk_timeout - jiffies : 0;
486 static void mptcp_set_timeout(struct sock *sk)
488 struct mptcp_subflow_context *subflow;
491 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
492 tout = max(tout, mptcp_timeout_from_subflow(subflow));
493 __mptcp_set_timeout(sk, tout);
496 static inline bool tcp_can_send_ack(const struct sock *ssk)
498 return !((1 << inet_sk_state_load(ssk)) &
499 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
502 void __mptcp_subflow_send_ack(struct sock *ssk)
504 if (tcp_can_send_ack(ssk))
508 static void mptcp_subflow_send_ack(struct sock *ssk)
512 slow = lock_sock_fast(ssk);
513 __mptcp_subflow_send_ack(ssk);
514 unlock_sock_fast(ssk, slow);
517 static void mptcp_send_ack(struct mptcp_sock *msk)
519 struct mptcp_subflow_context *subflow;
521 mptcp_for_each_subflow(msk, subflow)
522 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
525 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
529 slow = lock_sock_fast(ssk);
530 if (tcp_can_send_ack(ssk))
531 tcp_cleanup_rbuf(ssk, 1);
532 unlock_sock_fast(ssk, slow);
535 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
537 const struct inet_connection_sock *icsk = inet_csk(ssk);
538 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
539 const struct tcp_sock *tp = tcp_sk(ssk);
541 return (ack_pending & ICSK_ACK_SCHED) &&
542 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
543 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
544 (rx_empty && ack_pending &
545 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
548 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
550 int old_space = READ_ONCE(msk->old_wspace);
551 struct mptcp_subflow_context *subflow;
552 struct sock *sk = (struct sock *)msk;
553 int space = __mptcp_space(sk);
554 bool cleanup, rx_empty;
556 cleanup = (space > 0) && (space >= (old_space << 1));
557 rx_empty = !__mptcp_rmem(sk);
559 mptcp_for_each_subflow(msk, subflow) {
560 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
562 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
563 mptcp_subflow_cleanup_rbuf(ssk);
567 static bool mptcp_check_data_fin(struct sock *sk)
569 struct mptcp_sock *msk = mptcp_sk(sk);
570 u64 rcv_data_fin_seq;
573 if (__mptcp_check_fallback(msk))
576 /* Need to ack a DATA_FIN received from a peer while this side
577 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
578 * msk->rcv_data_fin was set when parsing the incoming options
579 * at the subflow level and the msk lock was not held, so this
580 * is the first opportunity to act on the DATA_FIN and change
583 * If we are caught up to the sequence number of the incoming
584 * DATA_FIN, send the DATA_ACK now and do state transition. If
585 * not caught up, do nothing and let the recv code send DATA_ACK
589 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
590 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
591 WRITE_ONCE(msk->rcv_data_fin, 0);
593 sk->sk_shutdown |= RCV_SHUTDOWN;
594 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
596 switch (sk->sk_state) {
597 case TCP_ESTABLISHED:
598 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
601 inet_sk_state_store(sk, TCP_CLOSING);
604 inet_sk_state_store(sk, TCP_CLOSE);
607 /* Other states not expected */
614 mptcp_close_wake_up(sk);
619 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
623 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
624 struct sock *sk = (struct sock *)msk;
625 unsigned int moved = 0;
626 bool more_data_avail;
631 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
633 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
634 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
636 if (unlikely(ssk_rbuf > sk_rbuf)) {
637 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
642 pr_debug("msk=%p ssk=%p", msk, ssk);
645 u32 map_remaining, offset;
646 u32 seq = tp->copied_seq;
650 /* try to move as much data as available */
651 map_remaining = subflow->map_data_len -
652 mptcp_subflow_get_map_offset(subflow);
654 skb = skb_peek(&ssk->sk_receive_queue);
656 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
657 * a different CPU can have already processed the pending
658 * data, stop here or we can enter an infinite loop
665 if (__mptcp_check_fallback(msk)) {
666 /* Under fallback skbs have no MPTCP extension and TCP could
667 * collapse them between the dummy map creation and the
668 * current dequeue. Be sure to adjust the map size.
670 map_remaining = skb->len;
671 subflow->map_data_len = skb->len;
674 offset = seq - TCP_SKB_CB(skb)->seq;
675 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
681 if (offset < skb->len) {
682 size_t len = skb->len - offset;
687 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
691 if (WARN_ON_ONCE(map_remaining < len))
695 sk_eat_skb(ssk, skb);
699 WRITE_ONCE(tp->copied_seq, seq);
700 more_data_avail = mptcp_subflow_data_available(ssk);
702 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
706 } while (more_data_avail);
712 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
714 struct sock *sk = (struct sock *)msk;
715 struct sk_buff *skb, *tail;
720 p = rb_first(&msk->out_of_order_queue);
721 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
724 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
728 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
730 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
733 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
737 end_seq = MPTCP_SKB_CB(skb)->end_seq;
738 tail = skb_peek_tail(&sk->sk_receive_queue);
739 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
740 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
742 /* skip overlapping data, if any */
743 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
744 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
746 MPTCP_SKB_CB(skb)->offset += delta;
747 MPTCP_SKB_CB(skb)->map_seq += delta;
748 __skb_queue_tail(&sk->sk_receive_queue, skb);
750 msk->ack_seq = end_seq;
756 /* In most cases we will be able to lock the mptcp socket. If its already
757 * owned, we need to defer to the work queue to avoid ABBA deadlock.
759 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
761 struct sock *sk = (struct sock *)msk;
762 unsigned int moved = 0;
764 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
765 __mptcp_ofo_queue(msk);
766 if (unlikely(ssk->sk_err)) {
767 if (!sock_owned_by_user(sk))
768 __mptcp_error_report(sk);
770 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
773 /* If the moves have caught up with the DATA_FIN sequence number
774 * it's time to ack the DATA_FIN and change socket state, but
775 * this is not a good place to change state. Let the workqueue
778 if (mptcp_pending_data_fin(sk, NULL))
779 mptcp_schedule_work(sk);
783 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
785 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
786 struct mptcp_sock *msk = mptcp_sk(sk);
787 int sk_rbuf, ssk_rbuf;
789 /* The peer can send data while we are shutting down this
790 * subflow at msk destruction time, but we must avoid enqueuing
791 * more data to the msk receive queue
793 if (unlikely(subflow->disposable))
796 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
797 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
798 if (unlikely(ssk_rbuf > sk_rbuf))
801 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
802 if (__mptcp_rmem(sk) > sk_rbuf) {
803 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
807 /* Wake-up the reader only for in-sequence data */
809 if (move_skbs_to_msk(msk, ssk))
810 sk->sk_data_ready(sk);
812 mptcp_data_unlock(sk);
815 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
817 struct sock *sk = (struct sock *)msk;
819 if (sk->sk_state != TCP_ESTABLISHED)
822 /* attach to msk socket only after we are sure we will deal with it
825 if (sk->sk_socket && !ssk->sk_socket)
826 mptcp_sock_graft(ssk, sk->sk_socket);
828 mptcp_propagate_sndbuf((struct sock *)msk, ssk);
829 mptcp_sockopt_sync_locked(msk, ssk);
833 static void __mptcp_flush_join_list(struct sock *sk)
835 struct mptcp_subflow_context *tmp, *subflow;
836 struct mptcp_sock *msk = mptcp_sk(sk);
838 list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
839 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
840 bool slow = lock_sock_fast(ssk);
842 list_move_tail(&subflow->node, &msk->conn_list);
843 if (!__mptcp_finish_join(msk, ssk))
844 mptcp_subflow_reset(ssk);
845 unlock_sock_fast(ssk, slow);
849 static bool mptcp_timer_pending(struct sock *sk)
851 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
854 static void mptcp_reset_timer(struct sock *sk)
856 struct inet_connection_sock *icsk = inet_csk(sk);
859 /* prevent rescheduling on close */
860 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
863 tout = mptcp_sk(sk)->timer_ival;
864 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
867 bool mptcp_schedule_work(struct sock *sk)
869 if (inet_sk_state_load(sk) != TCP_CLOSE &&
870 schedule_work(&mptcp_sk(sk)->work)) {
871 /* each subflow already holds a reference to the sk, and the
872 * workqueue is invoked by a subflow, so sk can't go away here.
880 void mptcp_subflow_eof(struct sock *sk)
882 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
883 mptcp_schedule_work(sk);
886 static void mptcp_check_for_eof(struct mptcp_sock *msk)
888 struct mptcp_subflow_context *subflow;
889 struct sock *sk = (struct sock *)msk;
892 mptcp_for_each_subflow(msk, subflow)
893 receivers += !subflow->rx_eof;
897 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
898 /* hopefully temporary hack: propagate shutdown status
899 * to msk, when all subflows agree on it
901 sk->sk_shutdown |= RCV_SHUTDOWN;
903 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
904 sk->sk_data_ready(sk);
907 switch (sk->sk_state) {
908 case TCP_ESTABLISHED:
909 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
912 inet_sk_state_store(sk, TCP_CLOSING);
915 inet_sk_state_store(sk, TCP_CLOSE);
920 mptcp_close_wake_up(sk);
923 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
925 struct mptcp_subflow_context *subflow;
926 struct sock *sk = (struct sock *)msk;
928 sock_owned_by_me(sk);
930 mptcp_for_each_subflow(msk, subflow) {
931 if (READ_ONCE(subflow->data_avail))
932 return mptcp_subflow_tcp_sock(subflow);
938 static bool mptcp_skb_can_collapse_to(u64 write_seq,
939 const struct sk_buff *skb,
940 const struct mptcp_ext *mpext)
942 if (!tcp_skb_can_collapse_to(skb))
945 /* can collapse only if MPTCP level sequence is in order and this
946 * mapping has not been xmitted yet
948 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
952 /* we can append data to the given data frag if:
953 * - there is space available in the backing page_frag
954 * - the data frag tail matches the current page_frag free offset
955 * - the data frag end sequence number matches the current write seq
957 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
958 const struct page_frag *pfrag,
959 const struct mptcp_data_frag *df)
961 return df && pfrag->page == df->page &&
962 pfrag->size - pfrag->offset > 0 &&
963 pfrag->offset == (df->offset + df->data_len) &&
964 df->data_seq + df->data_len == msk->write_seq;
967 static void dfrag_uncharge(struct sock *sk, int len)
969 sk_mem_uncharge(sk, len);
970 sk_wmem_queued_add(sk, -len);
973 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
975 int len = dfrag->data_len + dfrag->overhead;
977 list_del(&dfrag->list);
978 dfrag_uncharge(sk, len);
979 put_page(dfrag->page);
982 static void __mptcp_clean_una(struct sock *sk)
984 struct mptcp_sock *msk = mptcp_sk(sk);
985 struct mptcp_data_frag *dtmp, *dfrag;
988 /* on fallback we just need to ignore snd_una, as this is really
991 if (__mptcp_check_fallback(msk))
992 msk->snd_una = READ_ONCE(msk->snd_nxt);
994 snd_una = msk->snd_una;
995 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
996 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
999 if (unlikely(dfrag == msk->first_pending)) {
1000 /* in recovery mode can see ack after the current snd head */
1001 if (WARN_ON_ONCE(!msk->recovery))
1004 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1007 dfrag_clear(sk, dfrag);
1010 dfrag = mptcp_rtx_head(sk);
1011 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1012 u64 delta = snd_una - dfrag->data_seq;
1014 /* prevent wrap around in recovery mode */
1015 if (unlikely(delta > dfrag->already_sent)) {
1016 if (WARN_ON_ONCE(!msk->recovery))
1018 if (WARN_ON_ONCE(delta > dfrag->data_len))
1020 dfrag->already_sent += delta - dfrag->already_sent;
1023 dfrag->data_seq += delta;
1024 dfrag->offset += delta;
1025 dfrag->data_len -= delta;
1026 dfrag->already_sent -= delta;
1028 dfrag_uncharge(sk, delta);
1031 /* all retransmitted data acked, recovery completed */
1032 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1033 msk->recovery = false;
1036 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1037 snd_una == READ_ONCE(msk->write_seq)) {
1038 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1039 mptcp_stop_timer(sk);
1041 mptcp_reset_timer(sk);
1045 static void __mptcp_clean_una_wakeup(struct sock *sk)
1047 lockdep_assert_held_once(&sk->sk_lock.slock);
1049 __mptcp_clean_una(sk);
1050 mptcp_write_space(sk);
1053 static void mptcp_clean_una_wakeup(struct sock *sk)
1055 mptcp_data_lock(sk);
1056 __mptcp_clean_una_wakeup(sk);
1057 mptcp_data_unlock(sk);
1060 static void mptcp_enter_memory_pressure(struct sock *sk)
1062 struct mptcp_subflow_context *subflow;
1063 struct mptcp_sock *msk = mptcp_sk(sk);
1066 sk_stream_moderate_sndbuf(sk);
1067 mptcp_for_each_subflow(msk, subflow) {
1068 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1071 tcp_enter_memory_pressure(ssk);
1072 sk_stream_moderate_sndbuf(ssk);
1077 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1080 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1082 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1083 pfrag, sk->sk_allocation)))
1086 mptcp_enter_memory_pressure(sk);
1090 static struct mptcp_data_frag *
1091 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1094 int offset = ALIGN(orig_offset, sizeof(long));
1095 struct mptcp_data_frag *dfrag;
1097 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1098 dfrag->data_len = 0;
1099 dfrag->data_seq = msk->write_seq;
1100 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1101 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1102 dfrag->already_sent = 0;
1103 dfrag->page = pfrag->page;
1108 struct mptcp_sendmsg_info {
1114 bool data_lock_held;
1117 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1118 u64 data_seq, int avail_size)
1120 u64 window_end = mptcp_wnd_end(msk);
1123 if (__mptcp_check_fallback(msk))
1126 mptcp_snd_wnd = window_end - data_seq;
1127 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1129 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1130 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1131 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1137 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1139 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1143 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1147 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1149 struct sk_buff *skb;
1151 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1153 if (likely(__mptcp_add_ext(skb, gfp))) {
1154 skb_reserve(skb, MAX_TCP_HEADER);
1155 skb->ip_summed = CHECKSUM_PARTIAL;
1156 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1161 mptcp_enter_memory_pressure(sk);
1166 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1168 struct sk_buff *skb;
1170 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1174 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1175 tcp_skb_entail(ssk, skb);
1178 tcp_skb_tsorted_anchor_cleanup(skb);
1183 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1185 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1187 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1190 /* note: this always recompute the csum on the whole skb, even
1191 * if we just appended a single frag. More status info needed
1193 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1195 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1196 __wsum csum = ~csum_unfold(mpext->csum);
1197 int offset = skb->len - added;
1199 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1202 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1204 struct mptcp_ext *mpext)
1209 mpext->infinite_map = 1;
1210 mpext->data_len = 0;
1212 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1213 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1215 mptcp_do_fallback(ssk);
1218 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1219 struct mptcp_data_frag *dfrag,
1220 struct mptcp_sendmsg_info *info)
1222 u64 data_seq = dfrag->data_seq + info->sent;
1223 int offset = dfrag->offset + info->sent;
1224 struct mptcp_sock *msk = mptcp_sk(sk);
1225 bool zero_window_probe = false;
1226 struct mptcp_ext *mpext = NULL;
1227 bool can_coalesce = false;
1228 bool reuse_skb = true;
1229 struct sk_buff *skb;
1233 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1234 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1236 if (WARN_ON_ONCE(info->sent > info->limit ||
1237 info->limit > dfrag->data_len))
1240 if (unlikely(!__tcp_can_send(ssk)))
1243 /* compute send limit */
1244 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1245 copy = info->size_goal;
1247 skb = tcp_write_queue_tail(ssk);
1248 if (skb && copy > skb->len) {
1249 /* Limit the write to the size available in the
1250 * current skb, if any, so that we create at most a new skb.
1251 * Explicitly tells TCP internals to avoid collapsing on later
1252 * queue management operation, to avoid breaking the ext <->
1253 * SSN association set here
1255 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1256 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1257 TCP_SKB_CB(skb)->eor = 1;
1261 i = skb_shinfo(skb)->nr_frags;
1262 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1263 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1264 tcp_mark_push(tcp_sk(ssk), skb);
1271 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1275 i = skb_shinfo(skb)->nr_frags;
1277 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1280 /* Zero window and all data acked? Probe. */
1281 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1283 u64 snd_una = READ_ONCE(msk->snd_una);
1285 if (snd_una != msk->snd_nxt) {
1286 tcp_remove_empty_skb(ssk);
1290 zero_window_probe = true;
1291 data_seq = snd_una - 1;
1294 /* all mptcp-level data is acked, no skbs should be present into the
1297 WARN_ON_ONCE(reuse_skb);
1300 copy = min_t(size_t, copy, info->limit - info->sent);
1301 if (!sk_wmem_schedule(ssk, copy)) {
1302 tcp_remove_empty_skb(ssk);
1307 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1309 get_page(dfrag->page);
1310 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1314 skb->data_len += copy;
1315 skb->truesize += copy;
1316 sk_wmem_queued_add(ssk, copy);
1317 sk_mem_charge(ssk, copy);
1318 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1319 TCP_SKB_CB(skb)->end_seq += copy;
1320 tcp_skb_pcount_set(skb, 0);
1322 /* on skb reuse we just need to update the DSS len */
1324 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1325 mpext->data_len += copy;
1326 WARN_ON_ONCE(zero_window_probe);
1330 memset(mpext, 0, sizeof(*mpext));
1331 mpext->data_seq = data_seq;
1332 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1333 mpext->data_len = copy;
1337 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1338 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1341 if (zero_window_probe) {
1342 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1344 if (READ_ONCE(msk->csum_enabled))
1345 mptcp_update_data_checksum(skb, copy);
1346 tcp_push_pending_frames(ssk);
1350 if (READ_ONCE(msk->csum_enabled))
1351 mptcp_update_data_checksum(skb, copy);
1352 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1353 mptcp_update_infinite_map(msk, ssk, mpext);
1354 trace_mptcp_sendmsg_frag(mpext);
1355 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1359 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1360 sizeof(struct tcphdr) - \
1361 MAX_TCP_OPTION_SPACE - \
1362 sizeof(struct ipv6hdr) - \
1363 sizeof(struct frag_hdr))
1365 struct subflow_send_info {
1370 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1372 if (!subflow->stale)
1376 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1379 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1381 if (unlikely(subflow->stale)) {
1382 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1384 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1387 mptcp_subflow_set_active(subflow);
1389 return __mptcp_subflow_active(subflow);
1392 #define SSK_MODE_ACTIVE 0
1393 #define SSK_MODE_BACKUP 1
1394 #define SSK_MODE_MAX 2
1396 /* implement the mptcp packet scheduler;
1397 * returns the subflow that will transmit the next DSS
1398 * additionally updates the rtx timeout
1400 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1402 struct subflow_send_info send_info[SSK_MODE_MAX];
1403 struct mptcp_subflow_context *subflow;
1404 struct sock *sk = (struct sock *)msk;
1405 u32 pace, burst, wmem;
1406 int i, nr_active = 0;
1411 sock_owned_by_me(sk);
1413 if (__mptcp_check_fallback(msk)) {
1416 return __tcp_can_send(msk->first) &&
1417 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1420 /* re-use last subflow, if the burst allow that */
1421 if (msk->last_snd && msk->snd_burst > 0 &&
1422 sk_stream_memory_free(msk->last_snd) &&
1423 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1424 mptcp_set_timeout(sk);
1425 return msk->last_snd;
1428 /* pick the subflow with the lower wmem/wspace ratio */
1429 for (i = 0; i < SSK_MODE_MAX; ++i) {
1430 send_info[i].ssk = NULL;
1431 send_info[i].linger_time = -1;
1434 mptcp_for_each_subflow(msk, subflow) {
1435 trace_mptcp_subflow_get_send(subflow);
1436 ssk = mptcp_subflow_tcp_sock(subflow);
1437 if (!mptcp_subflow_active(subflow))
1440 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1441 nr_active += !subflow->backup;
1442 pace = subflow->avg_pacing_rate;
1443 if (unlikely(!pace)) {
1444 /* init pacing rate from socket */
1445 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1446 pace = subflow->avg_pacing_rate;
1451 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1452 if (linger_time < send_info[subflow->backup].linger_time) {
1453 send_info[subflow->backup].ssk = ssk;
1454 send_info[subflow->backup].linger_time = linger_time;
1457 __mptcp_set_timeout(sk, tout);
1459 /* pick the best backup if no other subflow is active */
1461 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1463 /* According to the blest algorithm, to avoid HoL blocking for the
1464 * faster flow, we need to:
1465 * - estimate the faster flow linger time
1466 * - use the above to estimate the amount of byte transferred
1467 * by the faster flow
1468 * - check that the amount of queued data is greter than the above,
1469 * otherwise do not use the picked, slower, subflow
1470 * We select the subflow with the shorter estimated time to flush
1471 * the queued mem, which basically ensure the above. We just need
1472 * to check that subflow has a non empty cwin.
1474 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1475 if (!ssk || !sk_stream_memory_free(ssk))
1478 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1479 wmem = READ_ONCE(ssk->sk_wmem_queued);
1481 msk->last_snd = NULL;
1485 subflow = mptcp_subflow_ctx(ssk);
1486 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1487 READ_ONCE(ssk->sk_pacing_rate) * burst,
1489 msk->last_snd = ssk;
1490 msk->snd_burst = burst;
1494 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1496 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1500 static void mptcp_update_post_push(struct mptcp_sock *msk,
1501 struct mptcp_data_frag *dfrag,
1504 u64 snd_nxt_new = dfrag->data_seq;
1506 dfrag->already_sent += sent;
1508 msk->snd_burst -= sent;
1510 snd_nxt_new += dfrag->already_sent;
1512 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1513 * is recovering after a failover. In that event, this re-sends
1516 * Thus compute snd_nxt_new candidate based on
1517 * the dfrag->data_seq that was sent and the data
1518 * that has been handed to the subflow for transmission
1519 * and skip update in case it was old dfrag.
1521 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1522 msk->snd_nxt = snd_nxt_new;
1525 void mptcp_check_and_set_pending(struct sock *sk)
1527 if (mptcp_send_head(sk))
1528 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1531 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1533 struct sock *prev_ssk = NULL, *ssk = NULL;
1534 struct mptcp_sock *msk = mptcp_sk(sk);
1535 struct mptcp_sendmsg_info info = {
1538 bool do_check_data_fin = false;
1539 struct mptcp_data_frag *dfrag;
1542 while ((dfrag = mptcp_send_head(sk))) {
1543 info.sent = dfrag->already_sent;
1544 info.limit = dfrag->data_len;
1545 len = dfrag->data_len - dfrag->already_sent;
1550 ssk = mptcp_subflow_get_send(msk);
1552 /* First check. If the ssk has changed since
1553 * the last round, release prev_ssk
1555 if (ssk != prev_ssk && prev_ssk)
1556 mptcp_push_release(prev_ssk, &info);
1560 /* Need to lock the new subflow only if different
1561 * from the previous one, otherwise we are still
1562 * helding the relevant lock
1564 if (ssk != prev_ssk)
1567 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1571 mptcp_push_release(ssk, &info);
1575 do_check_data_fin = true;
1579 mptcp_update_post_push(msk, dfrag, ret);
1581 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1584 /* at this point we held the socket lock for the last subflow we used */
1586 mptcp_push_release(ssk, &info);
1589 /* ensure the rtx timer is running */
1590 if (!mptcp_timer_pending(sk))
1591 mptcp_reset_timer(sk);
1592 if (do_check_data_fin)
1593 __mptcp_check_send_data_fin(sk);
1596 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1598 struct mptcp_sock *msk = mptcp_sk(sk);
1599 struct mptcp_sendmsg_info info = {
1600 .data_lock_held = true,
1602 struct mptcp_data_frag *dfrag;
1603 struct sock *xmit_ssk;
1604 int len, copied = 0;
1607 while ((dfrag = mptcp_send_head(sk))) {
1608 info.sent = dfrag->already_sent;
1609 info.limit = dfrag->data_len;
1610 len = dfrag->data_len - dfrag->already_sent;
1614 /* check for a different subflow usage only after
1615 * spooling the first chunk of data
1617 xmit_ssk = first ? ssk : mptcp_subflow_get_send(msk);
1620 if (xmit_ssk != ssk) {
1621 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1622 MPTCP_DELEGATE_SEND);
1626 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1635 mptcp_update_post_push(msk, dfrag, ret);
1637 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1641 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1642 * not going to flush it via release_sock()
1645 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1647 if (!mptcp_timer_pending(sk))
1648 mptcp_reset_timer(sk);
1650 if (msk->snd_data_fin_enable &&
1651 msk->snd_nxt + 1 == msk->write_seq)
1652 mptcp_schedule_work(sk);
1656 static void mptcp_set_nospace(struct sock *sk)
1658 /* enable autotune */
1659 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1661 /* will be cleared on avail space */
1662 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1665 static int mptcp_sendmsg_fastopen(struct sock *sk, struct sock *ssk, struct msghdr *msg,
1666 size_t len, int *copied_syn)
1668 unsigned int saved_flags = msg->msg_flags;
1669 struct mptcp_sock *msk = mptcp_sk(sk);
1673 msg->msg_flags |= MSG_DONTWAIT;
1674 msk->connect_flags = O_NONBLOCK;
1675 msk->is_sendmsg = 1;
1676 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1677 msk->is_sendmsg = 0;
1678 msg->msg_flags = saved_flags;
1681 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1682 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1683 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1684 msg->msg_namelen, msg->msg_flags, 1);
1686 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1687 * case of any error, except timeout or signal
1689 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1696 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1698 struct mptcp_sock *msk = mptcp_sk(sk);
1699 struct page_frag *pfrag;
1700 struct socket *ssock;
1705 /* silently ignore everything else */
1706 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1710 ssock = __mptcp_nmpc_socket(msk);
1711 if (unlikely(ssock && (inet_sk(ssock->sk)->defer_connect ||
1712 msg->msg_flags & MSG_FASTOPEN))) {
1715 ret = mptcp_sendmsg_fastopen(sk, ssock->sk, msg, len, &copied_syn);
1716 copied += copied_syn;
1717 if (ret == -EINPROGRESS && copied_syn > 0)
1723 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1725 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1726 ret = sk_stream_wait_connect(sk, &timeo);
1732 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1735 pfrag = sk_page_frag(sk);
1737 while (msg_data_left(msg)) {
1738 int total_ts, frag_truesize = 0;
1739 struct mptcp_data_frag *dfrag;
1740 bool dfrag_collapsed;
1741 size_t psize, offset;
1743 /* reuse tail pfrag, if possible, or carve a new one from the
1746 dfrag = mptcp_pending_tail(sk);
1747 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1748 if (!dfrag_collapsed) {
1749 if (!sk_stream_memory_free(sk))
1750 goto wait_for_memory;
1752 if (!mptcp_page_frag_refill(sk, pfrag))
1753 goto wait_for_memory;
1755 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1756 frag_truesize = dfrag->overhead;
1759 /* we do not bound vs wspace, to allow a single packet.
1760 * memory accounting will prevent execessive memory usage
1763 offset = dfrag->offset + dfrag->data_len;
1764 psize = pfrag->size - offset;
1765 psize = min_t(size_t, psize, msg_data_left(msg));
1766 total_ts = psize + frag_truesize;
1768 if (!sk_wmem_schedule(sk, total_ts))
1769 goto wait_for_memory;
1771 if (copy_page_from_iter(dfrag->page, offset, psize,
1772 &msg->msg_iter) != psize) {
1777 /* data successfully copied into the write queue */
1778 sk->sk_forward_alloc -= total_ts;
1780 dfrag->data_len += psize;
1781 frag_truesize += psize;
1782 pfrag->offset += frag_truesize;
1783 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1785 /* charge data on mptcp pending queue to the msk socket
1786 * Note: we charge such data both to sk and ssk
1788 sk_wmem_queued_add(sk, frag_truesize);
1789 if (!dfrag_collapsed) {
1790 get_page(dfrag->page);
1791 list_add_tail(&dfrag->list, &msk->rtx_queue);
1792 if (!msk->first_pending)
1793 WRITE_ONCE(msk->first_pending, dfrag);
1795 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1796 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1802 mptcp_set_nospace(sk);
1803 __mptcp_push_pending(sk, msg->msg_flags);
1804 ret = sk_stream_wait_memory(sk, &timeo);
1810 __mptcp_push_pending(sk, msg->msg_flags);
1820 copied = sk_stream_error(sk, msg->msg_flags, ret);
1824 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1826 size_t len, int flags,
1827 struct scm_timestamping_internal *tss,
1830 struct sk_buff *skb, *tmp;
1833 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1834 u32 offset = MPTCP_SKB_CB(skb)->offset;
1835 u32 data_len = skb->len - offset;
1836 u32 count = min_t(size_t, len - copied, data_len);
1839 if (!(flags & MSG_TRUNC)) {
1840 err = skb_copy_datagram_msg(skb, offset, msg, count);
1841 if (unlikely(err < 0)) {
1848 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1849 tcp_update_recv_tstamps(skb, tss);
1850 *cmsg_flags |= MPTCP_CMSG_TS;
1855 if (count < data_len) {
1856 if (!(flags & MSG_PEEK)) {
1857 MPTCP_SKB_CB(skb)->offset += count;
1858 MPTCP_SKB_CB(skb)->map_seq += count;
1863 if (!(flags & MSG_PEEK)) {
1864 /* we will bulk release the skb memory later */
1865 skb->destructor = NULL;
1866 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1867 __skb_unlink(skb, &msk->receive_queue);
1878 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1880 * Only difference: Use highest rtt estimate of the subflows in use.
1882 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1884 struct mptcp_subflow_context *subflow;
1885 struct sock *sk = (struct sock *)msk;
1886 u32 time, advmss = 1;
1889 sock_owned_by_me(sk);
1894 msk->rcvq_space.copied += copied;
1896 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1897 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1899 rtt_us = msk->rcvq_space.rtt_us;
1900 if (rtt_us && time < (rtt_us >> 3))
1904 mptcp_for_each_subflow(msk, subflow) {
1905 const struct tcp_sock *tp;
1909 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1911 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1912 sf_advmss = READ_ONCE(tp->advmss);
1914 rtt_us = max(sf_rtt_us, rtt_us);
1915 advmss = max(sf_advmss, advmss);
1918 msk->rcvq_space.rtt_us = rtt_us;
1919 if (time < (rtt_us >> 3) || rtt_us == 0)
1922 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1925 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1926 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1930 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1932 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1934 do_div(grow, msk->rcvq_space.space);
1935 rcvwin += (grow << 1);
1937 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1938 while (tcp_win_from_space(sk, rcvmem) < advmss)
1941 do_div(rcvwin, advmss);
1942 rcvbuf = min_t(u64, rcvwin * rcvmem,
1943 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1945 if (rcvbuf > sk->sk_rcvbuf) {
1948 window_clamp = tcp_win_from_space(sk, rcvbuf);
1949 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1951 /* Make subflows follow along. If we do not do this, we
1952 * get drops at subflow level if skbs can't be moved to
1953 * the mptcp rx queue fast enough (announced rcv_win can
1954 * exceed ssk->sk_rcvbuf).
1956 mptcp_for_each_subflow(msk, subflow) {
1960 ssk = mptcp_subflow_tcp_sock(subflow);
1961 slow = lock_sock_fast(ssk);
1962 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1963 tcp_sk(ssk)->window_clamp = window_clamp;
1964 tcp_cleanup_rbuf(ssk, 1);
1965 unlock_sock_fast(ssk, slow);
1970 msk->rcvq_space.space = msk->rcvq_space.copied;
1972 msk->rcvq_space.copied = 0;
1973 msk->rcvq_space.time = mstamp;
1976 static void __mptcp_update_rmem(struct sock *sk)
1978 struct mptcp_sock *msk = mptcp_sk(sk);
1980 if (!msk->rmem_released)
1983 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1984 mptcp_rmem_uncharge(sk, msk->rmem_released);
1985 WRITE_ONCE(msk->rmem_released, 0);
1988 static void __mptcp_splice_receive_queue(struct sock *sk)
1990 struct mptcp_sock *msk = mptcp_sk(sk);
1992 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1995 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1997 struct sock *sk = (struct sock *)msk;
1998 unsigned int moved = 0;
2002 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2005 /* we can have data pending in the subflows only if the msk
2006 * receive buffer was full at subflow_data_ready() time,
2007 * that is an unlikely slow path.
2012 slowpath = lock_sock_fast(ssk);
2013 mptcp_data_lock(sk);
2014 __mptcp_update_rmem(sk);
2015 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2016 mptcp_data_unlock(sk);
2018 if (unlikely(ssk->sk_err))
2019 __mptcp_error_report(sk);
2020 unlock_sock_fast(ssk, slowpath);
2023 /* acquire the data lock only if some input data is pending */
2025 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2026 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2027 mptcp_data_lock(sk);
2028 __mptcp_update_rmem(sk);
2029 ret |= __mptcp_ofo_queue(msk);
2030 __mptcp_splice_receive_queue(sk);
2031 mptcp_data_unlock(sk);
2034 mptcp_check_data_fin((struct sock *)msk);
2035 return !skb_queue_empty(&msk->receive_queue);
2038 static unsigned int mptcp_inq_hint(const struct sock *sk)
2040 const struct mptcp_sock *msk = mptcp_sk(sk);
2041 const struct sk_buff *skb;
2043 skb = skb_peek(&msk->receive_queue);
2045 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2047 if (hint_val >= INT_MAX)
2050 return (unsigned int)hint_val;
2053 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2059 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2060 int flags, int *addr_len)
2062 struct mptcp_sock *msk = mptcp_sk(sk);
2063 struct scm_timestamping_internal tss;
2064 int copied = 0, cmsg_flags = 0;
2068 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2069 if (unlikely(flags & MSG_ERRQUEUE))
2070 return inet_recv_error(sk, msg, len, addr_len);
2073 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2078 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2080 len = min_t(size_t, len, INT_MAX);
2081 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2083 if (unlikely(msk->recvmsg_inq))
2084 cmsg_flags = MPTCP_CMSG_INQ;
2086 while (copied < len) {
2089 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2090 if (unlikely(bytes_read < 0)) {
2092 copied = bytes_read;
2096 copied += bytes_read;
2098 /* be sure to advertise window change */
2099 mptcp_cleanup_rbuf(msk);
2101 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2104 /* only the master socket status is relevant here. The exit
2105 * conditions mirror closely tcp_recvmsg()
2107 if (copied >= target)
2112 sk->sk_state == TCP_CLOSE ||
2113 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2115 signal_pending(current))
2119 copied = sock_error(sk);
2123 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2124 mptcp_check_for_eof(msk);
2126 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2127 /* race breaker: the shutdown could be after the
2128 * previous receive queue check
2130 if (__mptcp_move_skbs(msk))
2135 if (sk->sk_state == TCP_CLOSE) {
2145 if (signal_pending(current)) {
2146 copied = sock_intr_errno(timeo);
2151 pr_debug("block timeout %ld", timeo);
2152 sk_wait_data(sk, &timeo, NULL);
2156 if (cmsg_flags && copied >= 0) {
2157 if (cmsg_flags & MPTCP_CMSG_TS)
2158 tcp_recv_timestamp(msg, sk, &tss);
2160 if (cmsg_flags & MPTCP_CMSG_INQ) {
2161 unsigned int inq = mptcp_inq_hint(sk);
2163 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2167 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2168 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2169 skb_queue_empty(&msk->receive_queue), copied);
2170 if (!(flags & MSG_PEEK))
2171 mptcp_rcv_space_adjust(msk, copied);
2177 static void mptcp_retransmit_timer(struct timer_list *t)
2179 struct inet_connection_sock *icsk = from_timer(icsk, t,
2180 icsk_retransmit_timer);
2181 struct sock *sk = &icsk->icsk_inet.sk;
2182 struct mptcp_sock *msk = mptcp_sk(sk);
2185 if (!sock_owned_by_user(sk)) {
2186 /* we need a process context to retransmit */
2187 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2188 mptcp_schedule_work(sk);
2190 /* delegate our work to tcp_release_cb() */
2191 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2197 static void mptcp_timeout_timer(struct timer_list *t)
2199 struct sock *sk = from_timer(sk, t, sk_timer);
2201 mptcp_schedule_work(sk);
2205 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2208 * A backup subflow is returned only if that is the only kind available.
2210 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2212 struct sock *backup = NULL, *pick = NULL;
2213 struct mptcp_subflow_context *subflow;
2214 int min_stale_count = INT_MAX;
2216 sock_owned_by_me((const struct sock *)msk);
2218 if (__mptcp_check_fallback(msk))
2221 mptcp_for_each_subflow(msk, subflow) {
2222 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2224 if (!__mptcp_subflow_active(subflow))
2227 /* still data outstanding at TCP level? skip this */
2228 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2229 mptcp_pm_subflow_chk_stale(msk, ssk);
2230 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2234 if (subflow->backup) {
2247 /* use backup only if there are no progresses anywhere */
2248 return min_stale_count > 1 ? backup : NULL;
2251 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2254 iput(SOCK_INODE(msk->subflow));
2255 msk->subflow = NULL;
2259 bool __mptcp_retransmit_pending_data(struct sock *sk)
2261 struct mptcp_data_frag *cur, *rtx_head;
2262 struct mptcp_sock *msk = mptcp_sk(sk);
2264 if (__mptcp_check_fallback(msk))
2267 if (tcp_rtx_and_write_queues_empty(sk))
2270 /* the closing socket has some data untransmitted and/or unacked:
2271 * some data in the mptcp rtx queue has not really xmitted yet.
2272 * keep it simple and re-inject the whole mptcp level rtx queue
2274 mptcp_data_lock(sk);
2275 __mptcp_clean_una_wakeup(sk);
2276 rtx_head = mptcp_rtx_head(sk);
2278 mptcp_data_unlock(sk);
2282 msk->recovery_snd_nxt = msk->snd_nxt;
2283 msk->recovery = true;
2284 mptcp_data_unlock(sk);
2286 msk->first_pending = rtx_head;
2289 /* be sure to clear the "sent status" on all re-injected fragments */
2290 list_for_each_entry(cur, &msk->rtx_queue, list) {
2291 if (!cur->already_sent)
2293 cur->already_sent = 0;
2299 /* flags for __mptcp_close_ssk() */
2300 #define MPTCP_CF_PUSH BIT(1)
2301 #define MPTCP_CF_FASTCLOSE BIT(2)
2303 /* subflow sockets can be either outgoing (connect) or incoming
2306 * Outgoing subflows use in-kernel sockets.
2307 * Incoming subflows do not have their own 'struct socket' allocated,
2308 * so we need to use tcp_close() after detaching them from the mptcp
2311 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2312 struct mptcp_subflow_context *subflow,
2315 struct mptcp_sock *msk = mptcp_sk(sk);
2316 bool need_push, dispose_it;
2318 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2320 list_del(&subflow->node);
2322 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2324 if (flags & MPTCP_CF_FASTCLOSE) {
2325 /* be sure to force the tcp_disconnect() path,
2326 * to generate the egress reset
2328 ssk->sk_lingertime = 0;
2329 sock_set_flag(ssk, SOCK_LINGER);
2330 subflow->send_fastclose = 1;
2333 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2335 tcp_disconnect(ssk, 0);
2336 msk->subflow->state = SS_UNCONNECTED;
2337 mptcp_subflow_ctx_reset(subflow);
2344 subflow->disposable = 1;
2346 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2347 * the ssk has been already destroyed, we just need to release the
2348 * reference owned by msk;
2350 if (!inet_csk(ssk)->icsk_ulp_ops) {
2351 kfree_rcu(subflow, rcu);
2353 /* otherwise tcp will dispose of the ssk and subflow ctx */
2354 if (ssk->sk_state == TCP_LISTEN) {
2355 tcp_set_state(ssk, TCP_CLOSE);
2356 mptcp_subflow_queue_clean(ssk);
2357 inet_csk_listen_stop(ssk);
2358 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2360 __tcp_close(ssk, 0);
2362 /* close acquired an extra ref */
2369 if (ssk == msk->first)
2373 if (ssk == msk->last_snd)
2374 msk->last_snd = NULL;
2377 __mptcp_push_pending(sk, 0);
2380 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2381 struct mptcp_subflow_context *subflow)
2383 if (sk->sk_state == TCP_ESTABLISHED)
2384 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2386 /* subflow aborted before reaching the fully_established status
2387 * attempt the creation of the next subflow
2389 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2391 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2394 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2399 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2401 struct mptcp_subflow_context *subflow, *tmp;
2405 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2406 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2408 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2411 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2412 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2415 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2419 static bool mptcp_check_close_timeout(const struct sock *sk)
2421 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2422 struct mptcp_subflow_context *subflow;
2424 if (delta >= TCP_TIMEWAIT_LEN)
2427 /* if all subflows are in closed status don't bother with additional
2430 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2431 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2438 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2440 struct mptcp_subflow_context *subflow, *tmp;
2441 struct sock *sk = (struct sock *)msk;
2443 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2446 mptcp_token_destroy(msk);
2448 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2449 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2452 slow = lock_sock_fast(tcp_sk);
2453 if (tcp_sk->sk_state != TCP_CLOSE) {
2454 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2455 tcp_set_state(tcp_sk, TCP_CLOSE);
2457 unlock_sock_fast(tcp_sk, slow);
2460 /* Mirror the tcp_reset() error propagation */
2461 switch (sk->sk_state) {
2463 sk->sk_err = ECONNREFUSED;
2465 case TCP_CLOSE_WAIT:
2471 sk->sk_err = ECONNRESET;
2474 inet_sk_state_store(sk, TCP_CLOSE);
2475 sk->sk_shutdown = SHUTDOWN_MASK;
2476 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2477 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2479 /* the calling mptcp_worker will properly destroy the socket */
2480 if (sock_flag(sk, SOCK_DEAD))
2483 sk->sk_state_change(sk);
2484 sk_error_report(sk);
2487 static void __mptcp_retrans(struct sock *sk)
2489 struct mptcp_sock *msk = mptcp_sk(sk);
2490 struct mptcp_sendmsg_info info = {};
2491 struct mptcp_data_frag *dfrag;
2496 mptcp_clean_una_wakeup(sk);
2498 /* first check ssk: need to kick "stale" logic */
2499 ssk = mptcp_subflow_get_retrans(msk);
2500 dfrag = mptcp_rtx_head(sk);
2502 if (mptcp_data_fin_enabled(msk)) {
2503 struct inet_connection_sock *icsk = inet_csk(sk);
2505 icsk->icsk_retransmits++;
2506 mptcp_set_datafin_timeout(sk);
2507 mptcp_send_ack(msk);
2512 if (!mptcp_send_head(sk))
2523 /* limit retransmission to the bytes already sent on some subflows */
2525 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2526 while (info.sent < info.limit) {
2527 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2531 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2536 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2537 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2539 WRITE_ONCE(msk->allow_infinite_fallback, false);
2545 mptcp_check_and_set_pending(sk);
2547 if (!mptcp_timer_pending(sk))
2548 mptcp_reset_timer(sk);
2551 /* schedule the timeout timer for the relevant event: either close timeout
2552 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2554 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2556 struct sock *sk = (struct sock *)msk;
2557 unsigned long timeout, close_timeout;
2559 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2562 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2564 /* the close timeout takes precedence on the fail one, and here at least one of
2567 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2569 sk_reset_timer(sk, &sk->sk_timer, timeout);
2572 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2574 struct sock *ssk = msk->first;
2580 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2582 slow = lock_sock_fast(ssk);
2583 mptcp_subflow_reset(ssk);
2584 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2585 unlock_sock_fast(ssk, slow);
2587 mptcp_reset_timeout(msk, 0);
2590 static void mptcp_do_fastclose(struct sock *sk)
2592 struct mptcp_subflow_context *subflow, *tmp;
2593 struct mptcp_sock *msk = mptcp_sk(sk);
2595 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2596 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2597 subflow, MPTCP_CF_FASTCLOSE);
2600 static void mptcp_worker(struct work_struct *work)
2602 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2603 struct sock *sk = (struct sock *)msk;
2604 unsigned long fail_tout;
2608 state = sk->sk_state;
2609 if (unlikely(state == TCP_CLOSE))
2612 mptcp_check_data_fin_ack(sk);
2614 mptcp_check_fastclose(msk);
2616 mptcp_pm_nl_work(msk);
2618 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2619 mptcp_check_for_eof(msk);
2621 __mptcp_check_send_data_fin(sk);
2622 mptcp_check_data_fin(sk);
2624 /* There is no point in keeping around an orphaned sk timedout or
2625 * closed, but we need the msk around to reply to incoming DATA_FIN,
2626 * even if it is orphaned and in FIN_WAIT2 state
2628 if (sock_flag(sk, SOCK_DEAD)) {
2629 if (mptcp_check_close_timeout(sk)) {
2630 inet_sk_state_store(sk, TCP_CLOSE);
2631 mptcp_do_fastclose(sk);
2633 if (sk->sk_state == TCP_CLOSE) {
2634 __mptcp_destroy_sock(sk);
2639 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2640 __mptcp_close_subflow(msk);
2642 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2643 __mptcp_retrans(sk);
2645 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2646 if (fail_tout && time_after(jiffies, fail_tout))
2647 mptcp_mp_fail_no_response(msk);
2654 static int __mptcp_init_sock(struct sock *sk)
2656 struct mptcp_sock *msk = mptcp_sk(sk);
2658 INIT_LIST_HEAD(&msk->conn_list);
2659 INIT_LIST_HEAD(&msk->join_list);
2660 INIT_LIST_HEAD(&msk->rtx_queue);
2661 INIT_WORK(&msk->work, mptcp_worker);
2662 __skb_queue_head_init(&msk->receive_queue);
2663 msk->out_of_order_queue = RB_ROOT;
2664 msk->first_pending = NULL;
2665 msk->rmem_fwd_alloc = 0;
2666 WRITE_ONCE(msk->rmem_released, 0);
2667 msk->timer_ival = TCP_RTO_MIN;
2670 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2671 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2672 WRITE_ONCE(msk->allow_infinite_fallback, true);
2673 msk->recovery = false;
2675 mptcp_pm_data_init(msk);
2677 /* re-use the csk retrans timer for MPTCP-level retrans */
2678 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2679 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2684 static void mptcp_ca_reset(struct sock *sk)
2686 struct inet_connection_sock *icsk = inet_csk(sk);
2688 tcp_assign_congestion_control(sk);
2689 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2691 /* no need to keep a reference to the ops, the name will suffice */
2692 tcp_cleanup_congestion_control(sk);
2693 icsk->icsk_ca_ops = NULL;
2696 static int mptcp_init_sock(struct sock *sk)
2698 struct net *net = sock_net(sk);
2701 ret = __mptcp_init_sock(sk);
2705 if (!mptcp_is_enabled(net))
2706 return -ENOPROTOOPT;
2708 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2711 ret = __mptcp_socket_create(mptcp_sk(sk));
2715 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2717 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2718 * propagate the correct value
2722 sk_sockets_allocated_inc(sk);
2723 sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2724 sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2729 static void __mptcp_clear_xmit(struct sock *sk)
2731 struct mptcp_sock *msk = mptcp_sk(sk);
2732 struct mptcp_data_frag *dtmp, *dfrag;
2734 WRITE_ONCE(msk->first_pending, NULL);
2735 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2736 dfrag_clear(sk, dfrag);
2739 void mptcp_cancel_work(struct sock *sk)
2741 struct mptcp_sock *msk = mptcp_sk(sk);
2743 if (cancel_work_sync(&msk->work))
2747 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2751 switch (ssk->sk_state) {
2753 if (!(how & RCV_SHUTDOWN))
2757 tcp_disconnect(ssk, O_NONBLOCK);
2760 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2761 pr_debug("Fallback");
2762 ssk->sk_shutdown |= how;
2763 tcp_shutdown(ssk, how);
2765 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2767 if (!mptcp_timer_pending(sk))
2768 mptcp_reset_timer(sk);
2776 static const unsigned char new_state[16] = {
2777 /* current state: new state: action: */
2778 [0 /* (Invalid) */] = TCP_CLOSE,
2779 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2780 [TCP_SYN_SENT] = TCP_CLOSE,
2781 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2782 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2783 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2784 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2785 [TCP_CLOSE] = TCP_CLOSE,
2786 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2787 [TCP_LAST_ACK] = TCP_LAST_ACK,
2788 [TCP_LISTEN] = TCP_CLOSE,
2789 [TCP_CLOSING] = TCP_CLOSING,
2790 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2793 static int mptcp_close_state(struct sock *sk)
2795 int next = (int)new_state[sk->sk_state];
2796 int ns = next & TCP_STATE_MASK;
2798 inet_sk_state_store(sk, ns);
2800 return next & TCP_ACTION_FIN;
2803 static void __mptcp_check_send_data_fin(struct sock *sk)
2805 struct mptcp_subflow_context *subflow;
2806 struct mptcp_sock *msk = mptcp_sk(sk);
2808 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2809 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2810 msk->snd_nxt, msk->write_seq);
2812 /* we still need to enqueue subflows or not really shutting down,
2815 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2816 mptcp_send_head(sk))
2819 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2821 /* fallback socket will not get data_fin/ack, can move to the next
2824 if (__mptcp_check_fallback(msk)) {
2825 WRITE_ONCE(msk->snd_una, msk->write_seq);
2826 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2827 inet_sk_state_store(sk, TCP_CLOSE);
2828 mptcp_close_wake_up(sk);
2829 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2830 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2834 mptcp_for_each_subflow(msk, subflow) {
2835 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2837 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2841 static void __mptcp_wr_shutdown(struct sock *sk)
2843 struct mptcp_sock *msk = mptcp_sk(sk);
2845 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2846 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2847 !!mptcp_send_head(sk));
2849 /* will be ignored by fallback sockets */
2850 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2851 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2853 __mptcp_check_send_data_fin(sk);
2856 static void __mptcp_destroy_sock(struct sock *sk)
2858 struct mptcp_sock *msk = mptcp_sk(sk);
2860 pr_debug("msk=%p", msk);
2864 mptcp_stop_timer(sk);
2865 sk_stop_timer(sk, &sk->sk_timer);
2868 sk->sk_prot->destroy(sk);
2870 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2871 WARN_ON_ONCE(msk->rmem_released);
2872 sk_stream_kill_queues(sk);
2873 xfrm_sk_free_policy(sk);
2875 sk_refcnt_debug_release(sk);
2879 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2881 /* Concurrent splices from sk_receive_queue into receive_queue will
2882 * always show at least one non-empty queue when checked in this order.
2884 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2885 skb_queue_empty_lockless(&msk->receive_queue))
2888 return EPOLLIN | EPOLLRDNORM;
2891 bool __mptcp_close(struct sock *sk, long timeout)
2893 struct mptcp_subflow_context *subflow;
2894 struct mptcp_sock *msk = mptcp_sk(sk);
2895 bool do_cancel_work = false;
2897 sk->sk_shutdown = SHUTDOWN_MASK;
2899 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2900 inet_sk_state_store(sk, TCP_CLOSE);
2904 if (mptcp_check_readable(msk)) {
2905 /* the msk has read data, do the MPTCP equivalent of TCP reset */
2906 inet_sk_state_store(sk, TCP_CLOSE);
2907 mptcp_do_fastclose(sk);
2908 } else if (mptcp_close_state(sk)) {
2909 __mptcp_wr_shutdown(sk);
2912 sk_stream_wait_close(sk, timeout);
2915 /* orphan all the subflows */
2916 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2917 mptcp_for_each_subflow(msk, subflow) {
2918 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2919 bool slow = lock_sock_fast_nested(ssk);
2921 /* since the close timeout takes precedence on the fail one,
2924 if (ssk == msk->first)
2925 subflow->fail_tout = 0;
2927 /* detach from the parent socket, but allow data_ready to
2928 * push incoming data into the mptcp stack, to properly ack it
2930 ssk->sk_socket = NULL;
2932 unlock_sock_fast(ssk, slow);
2937 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2939 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2941 if (sk->sk_state == TCP_CLOSE) {
2942 __mptcp_destroy_sock(sk);
2943 do_cancel_work = true;
2945 mptcp_reset_timeout(msk, 0);
2948 return do_cancel_work;
2951 static void mptcp_close(struct sock *sk, long timeout)
2953 bool do_cancel_work;
2957 do_cancel_work = __mptcp_close(sk, timeout);
2960 mptcp_cancel_work(sk);
2965 void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2967 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2968 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2969 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2971 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2972 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2975 msk6->saddr = ssk6->saddr;
2976 msk6->flow_label = ssk6->flow_label;
2980 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2981 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2982 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2983 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2984 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2985 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2988 static int mptcp_disconnect(struct sock *sk, int flags)
2990 struct mptcp_sock *msk = mptcp_sk(sk);
2992 inet_sk_state_store(sk, TCP_CLOSE);
2994 mptcp_stop_timer(sk);
2995 sk_stop_timer(sk, &sk->sk_timer);
2998 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3000 /* msk->subflow is still intact, the following will not free the first
3003 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3004 msk->last_snd = NULL;
3005 WRITE_ONCE(msk->flags, 0);
3007 msk->push_pending = 0;
3008 msk->recovery = false;
3009 msk->can_ack = false;
3010 msk->fully_established = false;
3011 msk->rcv_data_fin = false;
3012 msk->snd_data_fin_enable = false;
3013 msk->rcv_fastclose = false;
3014 msk->use_64bit_ack = false;
3015 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3016 mptcp_pm_data_reset(msk);
3019 sk->sk_shutdown = 0;
3020 sk_error_report(sk);
3024 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3025 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3027 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3029 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3033 struct sock *mptcp_sk_clone(const struct sock *sk,
3034 const struct mptcp_options_received *mp_opt,
3035 struct request_sock *req)
3037 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3038 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3039 struct mptcp_sock *msk;
3044 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3045 if (nsk->sk_family == AF_INET6)
3046 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3049 __mptcp_init_sock(nsk);
3051 msk = mptcp_sk(nsk);
3052 msk->local_key = subflow_req->local_key;
3053 msk->token = subflow_req->token;
3054 msk->subflow = NULL;
3055 WRITE_ONCE(msk->fully_established, false);
3056 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3057 WRITE_ONCE(msk->csum_enabled, true);
3059 msk->write_seq = subflow_req->idsn + 1;
3060 msk->snd_nxt = msk->write_seq;
3061 msk->snd_una = msk->write_seq;
3062 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3063 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3065 sock_reset_flag(nsk, SOCK_RCU_FREE);
3066 /* will be fully established after successful MPC subflow creation */
3067 inet_sk_state_store(nsk, TCP_SYN_RECV);
3069 security_inet_csk_clone(nsk, req);
3070 bh_unlock_sock(nsk);
3072 /* keep a single reference */
3077 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3079 const struct tcp_sock *tp = tcp_sk(ssk);
3081 msk->rcvq_space.copied = 0;
3082 msk->rcvq_space.rtt_us = 0;
3084 msk->rcvq_space.time = tp->tcp_mstamp;
3086 /* initial rcv_space offering made to peer */
3087 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3088 TCP_INIT_CWND * tp->advmss);
3089 if (msk->rcvq_space.space == 0)
3090 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3092 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3095 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3098 struct mptcp_sock *msk = mptcp_sk(sk);
3099 struct socket *listener;
3102 listener = __mptcp_nmpc_socket(msk);
3103 if (WARN_ON_ONCE(!listener)) {
3108 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3109 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3113 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3114 if (sk_is_mptcp(newsk)) {
3115 struct mptcp_subflow_context *subflow;
3116 struct sock *new_mptcp_sock;
3118 subflow = mptcp_subflow_ctx(newsk);
3119 new_mptcp_sock = subflow->conn;
3121 /* is_mptcp should be false if subflow->conn is missing, see
3122 * subflow_syn_recv_sock()
3124 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3125 tcp_sk(newsk)->is_mptcp = 0;
3129 /* acquire the 2nd reference for the owning socket */
3130 sock_hold(new_mptcp_sock);
3131 newsk = new_mptcp_sock;
3132 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3134 MPTCP_INC_STATS(sock_net(sk),
3135 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3139 newsk->sk_kern_sock = kern;
3143 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3145 struct mptcp_subflow_context *subflow, *tmp;
3146 struct sock *sk = (struct sock *)msk;
3148 __mptcp_clear_xmit(sk);
3150 /* join list will be eventually flushed (with rst) at sock lock release time */
3151 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3152 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3154 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3155 mptcp_data_lock(sk);
3156 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3157 __skb_queue_purge(&sk->sk_receive_queue);
3158 skb_rbtree_purge(&msk->out_of_order_queue);
3159 mptcp_data_unlock(sk);
3161 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3162 * inet_sock_destruct() will dispose it
3164 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3165 msk->rmem_fwd_alloc = 0;
3166 mptcp_token_destroy(msk);
3167 mptcp_pm_free_anno_list(msk);
3168 mptcp_free_local_addr_list(msk);
3171 static void mptcp_destroy(struct sock *sk)
3173 struct mptcp_sock *msk = mptcp_sk(sk);
3175 /* clears msk->subflow, allowing the following to close
3176 * even the initial subflow
3178 mptcp_dispose_initial_subflow(msk);
3179 mptcp_destroy_common(msk, 0);
3180 sk_sockets_allocated_dec(sk);
3183 void __mptcp_data_acked(struct sock *sk)
3185 if (!sock_owned_by_user(sk))
3186 __mptcp_clean_una(sk);
3188 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3190 if (mptcp_pending_data_fin_ack(sk))
3191 mptcp_schedule_work(sk);
3194 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3196 if (!mptcp_send_head(sk))
3199 if (!sock_owned_by_user(sk))
3200 __mptcp_subflow_push_pending(sk, ssk, false);
3202 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3205 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3206 BIT(MPTCP_RETRANSMIT) | \
3207 BIT(MPTCP_FLUSH_JOIN_LIST))
3209 /* processes deferred events and flush wmem */
3210 static void mptcp_release_cb(struct sock *sk)
3211 __must_hold(&sk->sk_lock.slock)
3213 struct mptcp_sock *msk = mptcp_sk(sk);
3216 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3221 /* the following actions acquire the subflow socket lock
3223 * 1) can't be invoked in atomic scope
3224 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3225 * datapath acquires the msk socket spinlock while helding
3226 * the subflow socket lock
3228 msk->push_pending = 0;
3229 msk->cb_flags &= ~flags;
3230 spin_unlock_bh(&sk->sk_lock.slock);
3231 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3232 __mptcp_flush_join_list(sk);
3233 if (flags & BIT(MPTCP_PUSH_PENDING))
3234 __mptcp_push_pending(sk, 0);
3235 if (flags & BIT(MPTCP_RETRANSMIT))
3236 __mptcp_retrans(sk);
3239 spin_lock_bh(&sk->sk_lock.slock);
3242 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3243 __mptcp_clean_una_wakeup(sk);
3244 if (unlikely(&msk->cb_flags)) {
3245 /* be sure to set the current sk state before tacking actions
3246 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3248 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3249 __mptcp_set_connected(sk);
3250 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3251 __mptcp_error_report(sk);
3252 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3253 msk->last_snd = NULL;
3256 __mptcp_update_rmem(sk);
3259 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3260 * TCP can't schedule delack timer before the subflow is fully established.
3261 * MPTCP uses the delack timer to do 3rd ack retransmissions
3263 static void schedule_3rdack_retransmission(struct sock *ssk)
3265 struct inet_connection_sock *icsk = inet_csk(ssk);
3266 struct tcp_sock *tp = tcp_sk(ssk);
3267 unsigned long timeout;
3269 if (mptcp_subflow_ctx(ssk)->fully_established)
3272 /* reschedule with a timeout above RTT, as we must look only for drop */
3274 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3276 timeout = TCP_TIMEOUT_INIT;
3279 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3280 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3281 icsk->icsk_ack.timeout = timeout;
3282 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3285 void mptcp_subflow_process_delegated(struct sock *ssk)
3287 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3288 struct sock *sk = subflow->conn;
3290 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3291 mptcp_data_lock(sk);
3292 if (!sock_owned_by_user(sk))
3293 __mptcp_subflow_push_pending(sk, ssk, true);
3295 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3296 mptcp_data_unlock(sk);
3297 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3299 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3300 schedule_3rdack_retransmission(ssk);
3301 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3305 static int mptcp_hash(struct sock *sk)
3307 /* should never be called,
3308 * we hash the TCP subflows not the master socket
3314 static void mptcp_unhash(struct sock *sk)
3316 /* called from sk_common_release(), but nothing to do here */
3319 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3321 struct mptcp_sock *msk = mptcp_sk(sk);
3322 struct socket *ssock;
3324 ssock = __mptcp_nmpc_socket(msk);
3325 pr_debug("msk=%p, subflow=%p", msk, ssock);
3326 if (WARN_ON_ONCE(!ssock))
3329 return inet_csk_get_port(ssock->sk, snum);
3332 void mptcp_finish_connect(struct sock *ssk)
3334 struct mptcp_subflow_context *subflow;
3335 struct mptcp_sock *msk;
3338 subflow = mptcp_subflow_ctx(ssk);
3342 pr_debug("msk=%p, token=%u", sk, subflow->token);
3344 subflow->map_seq = subflow->iasn;
3345 subflow->map_subflow_seq = 1;
3347 /* the socket is not connected yet, no msk/subflow ops can access/race
3348 * accessing the field below
3350 WRITE_ONCE(msk->local_key, subflow->local_key);
3351 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3352 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3353 WRITE_ONCE(msk->snd_una, msk->write_seq);
3355 mptcp_pm_new_connection(msk, ssk, 0);
3357 mptcp_rcv_space_init(msk, ssk);
3360 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3362 write_lock_bh(&sk->sk_callback_lock);
3363 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3364 sk_set_socket(sk, parent);
3365 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3366 write_unlock_bh(&sk->sk_callback_lock);
3369 bool mptcp_finish_join(struct sock *ssk)
3371 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3372 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3373 struct sock *parent = (void *)msk;
3376 pr_debug("msk=%p, subflow=%p", msk, subflow);
3378 /* mptcp socket already closing? */
3379 if (!mptcp_is_fully_established(parent)) {
3380 subflow->reset_reason = MPTCP_RST_EMPTCP;
3384 if (!list_empty(&subflow->node))
3387 if (!mptcp_pm_allow_new_subflow(msk))
3388 goto err_prohibited;
3390 /* active connections are already on conn_list.
3391 * If we can't acquire msk socket lock here, let the release callback
3394 mptcp_data_lock(parent);
3395 if (!sock_owned_by_user(parent)) {
3396 ret = __mptcp_finish_join(msk, ssk);
3399 list_add_tail(&subflow->node, &msk->conn_list);
3403 list_add_tail(&subflow->node, &msk->join_list);
3404 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3406 mptcp_data_unlock(parent);
3410 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3414 subflow->map_seq = READ_ONCE(msk->ack_seq);
3415 WRITE_ONCE(msk->allow_infinite_fallback, false);
3418 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3422 static void mptcp_shutdown(struct sock *sk, int how)
3424 pr_debug("sk=%p, how=%d", sk, how);
3426 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3427 __mptcp_wr_shutdown(sk);
3430 static int mptcp_forward_alloc_get(const struct sock *sk)
3432 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3435 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3437 const struct sock *sk = (void *)msk;
3440 if (sk->sk_state == TCP_LISTEN)
3443 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3446 delta = msk->write_seq - v;
3447 if (__mptcp_check_fallback(msk) && msk->first) {
3448 struct tcp_sock *tp = tcp_sk(msk->first);
3450 /* the first subflow is disconnected after close - see
3451 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3452 * so ignore that status, too.
3454 if (!((1 << msk->first->sk_state) &
3455 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3456 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3458 if (delta > INT_MAX)
3464 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3466 struct mptcp_sock *msk = mptcp_sk(sk);
3472 if (sk->sk_state == TCP_LISTEN)
3476 __mptcp_move_skbs(msk);
3477 answ = mptcp_inq_hint(sk);
3481 slow = lock_sock_fast(sk);
3482 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3483 unlock_sock_fast(sk, slow);
3486 slow = lock_sock_fast(sk);
3487 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3488 unlock_sock_fast(sk, slow);
3491 return -ENOIOCTLCMD;
3494 return put_user(answ, (int __user *)arg);
3497 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3498 struct mptcp_subflow_context *subflow)
3500 subflow->request_mptcp = 0;
3501 __mptcp_do_fallback(msk);
3504 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3506 struct mptcp_subflow_context *subflow;
3507 struct mptcp_sock *msk = mptcp_sk(sk);
3508 struct socket *ssock;
3511 ssock = __mptcp_nmpc_socket(msk);
3515 mptcp_token_destroy(msk);
3516 inet_sk_state_store(sk, TCP_SYN_SENT);
3517 subflow = mptcp_subflow_ctx(ssock->sk);
3518 #ifdef CONFIG_TCP_MD5SIG
3519 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3522 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3523 mptcp_subflow_early_fallback(msk, subflow);
3525 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3526 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3527 mptcp_subflow_early_fallback(msk, subflow);
3529 if (likely(!__mptcp_check_fallback(msk)))
3530 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3532 /* if reaching here via the fastopen/sendmsg path, the caller already
3533 * acquired the subflow socket lock, too.
3535 if (msk->is_sendmsg)
3536 err = __inet_stream_connect(ssock, uaddr, addr_len, msk->connect_flags, 1);
3538 err = inet_stream_connect(ssock, uaddr, addr_len, msk->connect_flags);
3539 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3541 /* on successful connect, the msk state will be moved to established by
3542 * subflow_finish_connect()
3544 if (unlikely(err && err != -EINPROGRESS)) {
3545 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3549 mptcp_copy_inaddrs(sk, ssock->sk);
3551 /* unblocking connect, mptcp-level inet_stream_connect will error out
3552 * without changing the socket state, update it here.
3554 if (err == -EINPROGRESS)
3555 sk->sk_socket->state = ssock->state;
3559 static struct proto mptcp_prot = {
3561 .owner = THIS_MODULE,
3562 .init = mptcp_init_sock,
3563 .connect = mptcp_connect,
3564 .disconnect = mptcp_disconnect,
3565 .close = mptcp_close,
3566 .accept = mptcp_accept,
3567 .setsockopt = mptcp_setsockopt,
3568 .getsockopt = mptcp_getsockopt,
3569 .shutdown = mptcp_shutdown,
3570 .destroy = mptcp_destroy,
3571 .sendmsg = mptcp_sendmsg,
3572 .ioctl = mptcp_ioctl,
3573 .recvmsg = mptcp_recvmsg,
3574 .release_cb = mptcp_release_cb,
3576 .unhash = mptcp_unhash,
3577 .get_port = mptcp_get_port,
3578 .forward_alloc_get = mptcp_forward_alloc_get,
3579 .sockets_allocated = &mptcp_sockets_allocated,
3581 .memory_allocated = &tcp_memory_allocated,
3582 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3584 .memory_pressure = &tcp_memory_pressure,
3585 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3586 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3587 .sysctl_mem = sysctl_tcp_mem,
3588 .obj_size = sizeof(struct mptcp_sock),
3589 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3590 .no_autobind = true,
3593 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3595 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3596 struct socket *ssock;
3599 lock_sock(sock->sk);
3600 ssock = __mptcp_nmpc_socket(msk);
3606 err = ssock->ops->bind(ssock, uaddr, addr_len);
3608 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3611 release_sock(sock->sk);
3615 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3616 int addr_len, int flags)
3620 lock_sock(sock->sk);
3621 mptcp_sk(sock->sk)->connect_flags = flags;
3622 ret = __inet_stream_connect(sock, uaddr, addr_len, flags, 0);
3623 release_sock(sock->sk);
3627 static int mptcp_listen(struct socket *sock, int backlog)
3629 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3630 struct socket *ssock;
3633 pr_debug("msk=%p", msk);
3635 lock_sock(sock->sk);
3636 ssock = __mptcp_nmpc_socket(msk);
3642 mptcp_token_destroy(msk);
3643 inet_sk_state_store(sock->sk, TCP_LISTEN);
3644 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3646 err = ssock->ops->listen(ssock, backlog);
3647 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3649 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3651 mptcp_event_pm_listener(ssock->sk, MPTCP_EVENT_LISTENER_CREATED);
3654 release_sock(sock->sk);
3658 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3659 int flags, bool kern)
3661 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3662 struct socket *ssock;
3665 pr_debug("msk=%p", msk);
3667 ssock = __mptcp_nmpc_socket(msk);
3671 err = ssock->ops->accept(sock, newsock, flags, kern);
3672 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3673 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3674 struct mptcp_subflow_context *subflow;
3675 struct sock *newsk = newsock->sk;
3677 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3681 /* PM/worker can now acquire the first subflow socket
3682 * lock without racing with listener queue cleanup,
3683 * we can notify it, if needed.
3685 * Even if remote has reset the initial subflow by now
3686 * the refcnt is still at least one.
3688 subflow = mptcp_subflow_ctx(msk->first);
3689 list_add(&subflow->node, &msk->conn_list);
3690 sock_hold(msk->first);
3691 if (mptcp_is_fully_established(newsk))
3692 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3694 mptcp_rcv_space_init(msk, msk->first);
3695 mptcp_propagate_sndbuf(newsk, msk->first);
3697 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3698 * This is needed so NOSPACE flag can be set from tcp stack.
3700 mptcp_for_each_subflow(msk, subflow) {
3701 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3703 if (!ssk->sk_socket)
3704 mptcp_sock_graft(ssk, newsock);
3706 release_sock(newsk);
3712 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3714 struct sock *sk = (struct sock *)msk;
3716 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3717 return EPOLLOUT | EPOLLWRNORM;
3719 if (sk_stream_is_writeable(sk))
3720 return EPOLLOUT | EPOLLWRNORM;
3722 mptcp_set_nospace(sk);
3723 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3724 if (sk_stream_is_writeable(sk))
3725 return EPOLLOUT | EPOLLWRNORM;
3730 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3731 struct poll_table_struct *wait)
3733 struct sock *sk = sock->sk;
3734 struct mptcp_sock *msk;
3739 sock_poll_wait(file, sock, wait);
3741 state = inet_sk_state_load(sk);
3742 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3743 if (state == TCP_LISTEN) {
3744 if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk))
3747 return inet_csk_listen_poll(msk->subflow->sk);
3750 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3751 mask |= mptcp_check_readable(msk);
3752 mask |= mptcp_check_writeable(msk);
3753 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3754 /* cf tcp_poll() note about TFO */
3755 mask |= EPOLLOUT | EPOLLWRNORM;
3757 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3759 if (sk->sk_shutdown & RCV_SHUTDOWN)
3760 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3762 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3770 static const struct proto_ops mptcp_stream_ops = {
3772 .owner = THIS_MODULE,
3773 .release = inet_release,
3775 .connect = mptcp_stream_connect,
3776 .socketpair = sock_no_socketpair,
3777 .accept = mptcp_stream_accept,
3778 .getname = inet_getname,
3780 .ioctl = inet_ioctl,
3781 .gettstamp = sock_gettstamp,
3782 .listen = mptcp_listen,
3783 .shutdown = inet_shutdown,
3784 .setsockopt = sock_common_setsockopt,
3785 .getsockopt = sock_common_getsockopt,
3786 .sendmsg = inet_sendmsg,
3787 .recvmsg = inet_recvmsg,
3788 .mmap = sock_no_mmap,
3789 .sendpage = inet_sendpage,
3792 static struct inet_protosw mptcp_protosw = {
3793 .type = SOCK_STREAM,
3794 .protocol = IPPROTO_MPTCP,
3795 .prot = &mptcp_prot,
3796 .ops = &mptcp_stream_ops,
3797 .flags = INET_PROTOSW_ICSK,
3800 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3802 struct mptcp_delegated_action *delegated;
3803 struct mptcp_subflow_context *subflow;
3806 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3807 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3808 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3810 bh_lock_sock_nested(ssk);
3811 if (!sock_owned_by_user(ssk) &&
3812 mptcp_subflow_has_delegated_action(subflow))
3813 mptcp_subflow_process_delegated(ssk);
3814 /* ... elsewhere tcp_release_cb_override already processed
3815 * the action or will do at next release_sock().
3816 * In both case must dequeue the subflow here - on the same
3817 * CPU that scheduled it.
3819 bh_unlock_sock(ssk);
3822 if (++work_done == budget)
3826 /* always provide a 0 'work_done' argument, so that napi_complete_done
3827 * will not try accessing the NULL napi->dev ptr
3829 napi_complete_done(napi, 0);
3833 void __init mptcp_proto_init(void)
3835 struct mptcp_delegated_action *delegated;
3838 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3840 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3841 panic("Failed to allocate MPTCP pcpu counter\n");
3843 init_dummy_netdev(&mptcp_napi_dev);
3844 for_each_possible_cpu(cpu) {
3845 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3846 INIT_LIST_HEAD(&delegated->head);
3847 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3849 napi_enable(&delegated->napi);
3852 mptcp_subflow_init();
3856 if (proto_register(&mptcp_prot, 1) != 0)
3857 panic("Failed to register MPTCP proto.\n");
3859 inet_register_protosw(&mptcp_protosw);
3861 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3864 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3865 static const struct proto_ops mptcp_v6_stream_ops = {
3867 .owner = THIS_MODULE,
3868 .release = inet6_release,
3870 .connect = mptcp_stream_connect,
3871 .socketpair = sock_no_socketpair,
3872 .accept = mptcp_stream_accept,
3873 .getname = inet6_getname,
3875 .ioctl = inet6_ioctl,
3876 .gettstamp = sock_gettstamp,
3877 .listen = mptcp_listen,
3878 .shutdown = inet_shutdown,
3879 .setsockopt = sock_common_setsockopt,
3880 .getsockopt = sock_common_getsockopt,
3881 .sendmsg = inet6_sendmsg,
3882 .recvmsg = inet6_recvmsg,
3883 .mmap = sock_no_mmap,
3884 .sendpage = inet_sendpage,
3885 #ifdef CONFIG_COMPAT
3886 .compat_ioctl = inet6_compat_ioctl,
3890 static struct proto mptcp_v6_prot;
3892 static struct inet_protosw mptcp_v6_protosw = {
3893 .type = SOCK_STREAM,
3894 .protocol = IPPROTO_MPTCP,
3895 .prot = &mptcp_v6_prot,
3896 .ops = &mptcp_v6_stream_ops,
3897 .flags = INET_PROTOSW_ICSK,
3900 int __init mptcp_proto_v6_init(void)
3904 mptcp_v6_prot = mptcp_prot;
3905 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3906 mptcp_v6_prot.slab = NULL;
3907 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3909 err = proto_register(&mptcp_v6_prot, 1);
3913 err = inet6_register_protosw(&mptcp_v6_protosw);
3915 proto_unregister(&mptcp_v6_prot);