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;
46 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
49 MPTCP_CMSG_TS = BIT(0),
50 MPTCP_CMSG_INQ = BIT(1),
53 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
55 static void __mptcp_destroy_sock(struct sock *sk);
56 static void mptcp_check_send_data_fin(struct sock *sk);
58 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
59 static struct net_device mptcp_napi_dev;
61 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
62 * completed yet or has failed, return the subflow socket.
63 * Otherwise return NULL.
65 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
67 if (!msk->subflow || READ_ONCE(msk->can_ack))
73 /* Returns end sequence number of the receiver's advertised window */
74 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
76 return READ_ONCE(msk->wnd_end);
79 static bool mptcp_is_tcpsk(struct sock *sk)
81 struct socket *sock = sk->sk_socket;
83 if (unlikely(sk->sk_prot == &tcp_prot)) {
84 /* we are being invoked after mptcp_accept() has
85 * accepted a non-mp-capable flow: sk is a tcp_sk,
88 * Hand the socket over to tcp so all further socket ops
91 sock->ops = &inet_stream_ops;
93 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
94 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
95 sock->ops = &inet6_stream_ops;
103 static int __mptcp_socket_create(struct mptcp_sock *msk)
105 struct mptcp_subflow_context *subflow;
106 struct sock *sk = (struct sock *)msk;
107 struct socket *ssock;
110 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
114 WRITE_ONCE(msk->first, ssock->sk);
115 WRITE_ONCE(msk->subflow, ssock);
116 subflow = mptcp_subflow_ctx(ssock->sk);
117 list_add(&subflow->node, &msk->conn_list);
118 sock_hold(ssock->sk);
119 subflow->request_mptcp = 1;
121 /* This is the first subflow, always with id 0 */
122 WRITE_ONCE(subflow->local_id, 0);
123 mptcp_sock_graft(msk->first, sk->sk_socket);
128 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
130 sk_drops_add(sk, skb);
134 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
136 WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
137 mptcp_sk(sk)->rmem_fwd_alloc + size);
140 static void mptcp_rmem_charge(struct sock *sk, int size)
142 mptcp_rmem_fwd_alloc_add(sk, -size);
145 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
146 struct sk_buff *from)
151 if (MPTCP_SKB_CB(from)->offset ||
152 !skb_try_coalesce(to, from, &fragstolen, &delta))
155 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
156 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
157 to->len, MPTCP_SKB_CB(from)->end_seq);
158 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
160 /* note the fwd memory can reach a negative value after accounting
161 * for the delta, but the later skb free will restore a non
164 atomic_add(delta, &sk->sk_rmem_alloc);
165 mptcp_rmem_charge(sk, delta);
166 kfree_skb_partial(from, fragstolen);
171 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
172 struct sk_buff *from)
174 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
177 return mptcp_try_coalesce((struct sock *)msk, to, from);
180 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
182 amount >>= PAGE_SHIFT;
183 mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
184 __sk_mem_reduce_allocated(sk, amount);
187 static void mptcp_rmem_uncharge(struct sock *sk, int size)
189 struct mptcp_sock *msk = mptcp_sk(sk);
192 mptcp_rmem_fwd_alloc_add(sk, size);
193 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
195 /* see sk_mem_uncharge() for the rationale behind the following schema */
196 if (unlikely(reclaimable >= PAGE_SIZE))
197 __mptcp_rmem_reclaim(sk, reclaimable);
200 static void mptcp_rfree(struct sk_buff *skb)
202 unsigned int len = skb->truesize;
203 struct sock *sk = skb->sk;
205 atomic_sub(len, &sk->sk_rmem_alloc);
206 mptcp_rmem_uncharge(sk, len);
209 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
213 skb->destructor = mptcp_rfree;
214 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
215 mptcp_rmem_charge(sk, skb->truesize);
218 /* "inspired" by tcp_data_queue_ofo(), main differences:
220 * - don't cope with sacks
222 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
224 struct sock *sk = (struct sock *)msk;
225 struct rb_node **p, *parent;
226 u64 seq, end_seq, max_seq;
227 struct sk_buff *skb1;
229 seq = MPTCP_SKB_CB(skb)->map_seq;
230 end_seq = MPTCP_SKB_CB(skb)->end_seq;
231 max_seq = atomic64_read(&msk->rcv_wnd_sent);
233 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
234 RB_EMPTY_ROOT(&msk->out_of_order_queue));
235 if (after64(end_seq, max_seq)) {
238 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
239 (unsigned long long)end_seq - (unsigned long)max_seq,
240 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
241 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
245 p = &msk->out_of_order_queue.rb_node;
246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
247 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
248 rb_link_node(&skb->rbnode, NULL, p);
249 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
250 msk->ooo_last_skb = skb;
254 /* with 2 subflows, adding at end of ooo queue is quite likely
255 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
257 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
258 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
259 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
263 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
264 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
265 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
266 parent = &msk->ooo_last_skb->rbnode;
267 p = &parent->rb_right;
271 /* Find place to insert this segment. Handle overlaps on the way. */
275 skb1 = rb_to_skb(parent);
276 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
277 p = &parent->rb_left;
280 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
281 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
282 /* All the bits are present. Drop. */
284 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
287 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
291 * continue traversing
294 /* skb's seq == skb1's seq and skb covers skb1.
295 * Replace skb1 with skb.
297 rb_replace_node(&skb1->rbnode, &skb->rbnode,
298 &msk->out_of_order_queue);
299 mptcp_drop(sk, skb1);
300 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
303 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
304 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
307 p = &parent->rb_right;
311 /* Insert segment into RB tree. */
312 rb_link_node(&skb->rbnode, parent, p);
313 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
316 /* Remove other segments covered by skb. */
317 while ((skb1 = skb_rb_next(skb)) != NULL) {
318 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
320 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
321 mptcp_drop(sk, skb1);
322 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
324 /* If there is no skb after us, we are the last_skb ! */
326 msk->ooo_last_skb = skb;
330 mptcp_set_owner_r(skb, sk);
333 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
335 struct mptcp_sock *msk = mptcp_sk(sk);
338 if (size <= msk->rmem_fwd_alloc)
341 size -= msk->rmem_fwd_alloc;
342 amt = sk_mem_pages(size);
343 amount = amt << PAGE_SHIFT;
344 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
347 mptcp_rmem_fwd_alloc_add(sk, amount);
351 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
352 struct sk_buff *skb, unsigned int offset,
355 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
356 struct sock *sk = (struct sock *)msk;
357 struct sk_buff *tail;
360 __skb_unlink(skb, &ssk->sk_receive_queue);
365 /* try to fetch required memory from subflow */
366 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
369 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
371 /* the skb map_seq accounts for the skb offset:
372 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
375 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
376 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
377 MPTCP_SKB_CB(skb)->offset = offset;
378 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
380 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
382 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
383 tail = skb_peek_tail(&sk->sk_receive_queue);
384 if (tail && mptcp_try_coalesce(sk, tail, skb))
387 mptcp_set_owner_r(skb, sk);
388 __skb_queue_tail(&sk->sk_receive_queue, skb);
390 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
391 mptcp_data_queue_ofo(msk, skb);
395 /* old data, keep it simple and drop the whole pkt, sender
396 * will retransmit as needed, if needed.
398 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
404 static void mptcp_stop_rtx_timer(struct sock *sk)
406 struct inet_connection_sock *icsk = inet_csk(sk);
408 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
409 mptcp_sk(sk)->timer_ival = 0;
412 static void mptcp_close_wake_up(struct sock *sk)
414 if (sock_flag(sk, SOCK_DEAD))
417 sk->sk_state_change(sk);
418 if (sk->sk_shutdown == SHUTDOWN_MASK ||
419 sk->sk_state == TCP_CLOSE)
420 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
422 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
425 static bool mptcp_pending_data_fin_ack(struct sock *sk)
427 struct mptcp_sock *msk = mptcp_sk(sk);
429 return ((1 << sk->sk_state) &
430 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
431 msk->write_seq == READ_ONCE(msk->snd_una);
434 static void mptcp_check_data_fin_ack(struct sock *sk)
436 struct mptcp_sock *msk = mptcp_sk(sk);
438 /* Look for an acknowledged DATA_FIN */
439 if (mptcp_pending_data_fin_ack(sk)) {
440 WRITE_ONCE(msk->snd_data_fin_enable, 0);
442 switch (sk->sk_state) {
444 inet_sk_state_store(sk, TCP_FIN_WAIT2);
448 inet_sk_state_store(sk, TCP_CLOSE);
452 mptcp_close_wake_up(sk);
456 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
458 struct mptcp_sock *msk = mptcp_sk(sk);
460 if (READ_ONCE(msk->rcv_data_fin) &&
461 ((1 << sk->sk_state) &
462 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
463 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
465 if (msk->ack_seq == rcv_data_fin_seq) {
467 *seq = rcv_data_fin_seq;
476 static void mptcp_set_datafin_timeout(const struct sock *sk)
478 struct inet_connection_sock *icsk = inet_csk(sk);
481 retransmits = min_t(u32, icsk->icsk_retransmits,
482 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
484 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
487 static void __mptcp_set_timeout(struct sock *sk, long tout)
489 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
492 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
494 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
496 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
497 inet_csk(ssk)->icsk_timeout - jiffies : 0;
500 static void mptcp_set_timeout(struct sock *sk)
502 struct mptcp_subflow_context *subflow;
505 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
506 tout = max(tout, mptcp_timeout_from_subflow(subflow));
507 __mptcp_set_timeout(sk, tout);
510 static inline bool tcp_can_send_ack(const struct sock *ssk)
512 return !((1 << inet_sk_state_load(ssk)) &
513 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
516 void __mptcp_subflow_send_ack(struct sock *ssk)
518 if (tcp_can_send_ack(ssk))
522 static void mptcp_subflow_send_ack(struct sock *ssk)
526 slow = lock_sock_fast(ssk);
527 __mptcp_subflow_send_ack(ssk);
528 unlock_sock_fast(ssk, slow);
531 static void mptcp_send_ack(struct mptcp_sock *msk)
533 struct mptcp_subflow_context *subflow;
535 mptcp_for_each_subflow(msk, subflow)
536 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
539 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
543 slow = lock_sock_fast(ssk);
544 if (tcp_can_send_ack(ssk))
545 tcp_cleanup_rbuf(ssk, 1);
546 unlock_sock_fast(ssk, slow);
549 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
551 const struct inet_connection_sock *icsk = inet_csk(ssk);
552 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
553 const struct tcp_sock *tp = tcp_sk(ssk);
555 return (ack_pending & ICSK_ACK_SCHED) &&
556 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
557 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
558 (rx_empty && ack_pending &
559 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
562 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
564 int old_space = READ_ONCE(msk->old_wspace);
565 struct mptcp_subflow_context *subflow;
566 struct sock *sk = (struct sock *)msk;
567 int space = __mptcp_space(sk);
568 bool cleanup, rx_empty;
570 cleanup = (space > 0) && (space >= (old_space << 1));
571 rx_empty = !__mptcp_rmem(sk);
573 mptcp_for_each_subflow(msk, subflow) {
574 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
576 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
577 mptcp_subflow_cleanup_rbuf(ssk);
581 static bool mptcp_check_data_fin(struct sock *sk)
583 struct mptcp_sock *msk = mptcp_sk(sk);
584 u64 rcv_data_fin_seq;
587 /* Need to ack a DATA_FIN received from a peer while this side
588 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
589 * msk->rcv_data_fin was set when parsing the incoming options
590 * at the subflow level and the msk lock was not held, so this
591 * is the first opportunity to act on the DATA_FIN and change
594 * If we are caught up to the sequence number of the incoming
595 * DATA_FIN, send the DATA_ACK now and do state transition. If
596 * not caught up, do nothing and let the recv code send DATA_ACK
600 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
601 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
602 WRITE_ONCE(msk->rcv_data_fin, 0);
604 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
605 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
607 switch (sk->sk_state) {
608 case TCP_ESTABLISHED:
609 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
612 inet_sk_state_store(sk, TCP_CLOSING);
615 inet_sk_state_store(sk, TCP_CLOSE);
618 /* Other states not expected */
624 if (!__mptcp_check_fallback(msk))
626 mptcp_close_wake_up(sk);
631 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
635 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
636 struct sock *sk = (struct sock *)msk;
637 unsigned int moved = 0;
638 bool more_data_avail;
643 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
645 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
646 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
648 if (unlikely(ssk_rbuf > sk_rbuf)) {
649 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
654 pr_debug("msk=%p ssk=%p", msk, ssk);
657 u32 map_remaining, offset;
658 u32 seq = tp->copied_seq;
662 /* try to move as much data as available */
663 map_remaining = subflow->map_data_len -
664 mptcp_subflow_get_map_offset(subflow);
666 skb = skb_peek(&ssk->sk_receive_queue);
668 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
669 * a different CPU can have already processed the pending
670 * data, stop here or we can enter an infinite loop
677 if (__mptcp_check_fallback(msk)) {
678 /* Under fallback skbs have no MPTCP extension and TCP could
679 * collapse them between the dummy map creation and the
680 * current dequeue. Be sure to adjust the map size.
682 map_remaining = skb->len;
683 subflow->map_data_len = skb->len;
686 offset = seq - TCP_SKB_CB(skb)->seq;
687 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
693 if (offset < skb->len) {
694 size_t len = skb->len - offset;
699 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
703 if (WARN_ON_ONCE(map_remaining < len))
707 sk_eat_skb(ssk, skb);
711 WRITE_ONCE(tp->copied_seq, seq);
712 more_data_avail = mptcp_subflow_data_available(ssk);
714 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
718 } while (more_data_avail);
724 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
726 struct sock *sk = (struct sock *)msk;
727 struct sk_buff *skb, *tail;
732 p = rb_first(&msk->out_of_order_queue);
733 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
736 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
740 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
742 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
745 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
749 end_seq = MPTCP_SKB_CB(skb)->end_seq;
750 tail = skb_peek_tail(&sk->sk_receive_queue);
751 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
752 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
754 /* skip overlapping data, if any */
755 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
756 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
758 MPTCP_SKB_CB(skb)->offset += delta;
759 MPTCP_SKB_CB(skb)->map_seq += delta;
760 __skb_queue_tail(&sk->sk_receive_queue, skb);
762 msk->ack_seq = end_seq;
768 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
770 int err = sock_error(ssk);
776 /* only propagate errors on fallen-back sockets or
779 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
782 /* We need to propagate only transition to CLOSE state.
783 * Orphaned socket will see such state change via
784 * subflow_sched_work_if_closed() and that path will properly
785 * destroy the msk as needed.
787 ssk_state = inet_sk_state_load(ssk);
788 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
789 inet_sk_state_store(sk, ssk_state);
790 WRITE_ONCE(sk->sk_err, -err);
792 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
798 void __mptcp_error_report(struct sock *sk)
800 struct mptcp_subflow_context *subflow;
801 struct mptcp_sock *msk = mptcp_sk(sk);
803 mptcp_for_each_subflow(msk, subflow)
804 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
808 /* In most cases we will be able to lock the mptcp socket. If its already
809 * owned, we need to defer to the work queue to avoid ABBA deadlock.
811 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
813 struct sock *sk = (struct sock *)msk;
814 unsigned int moved = 0;
816 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
817 __mptcp_ofo_queue(msk);
818 if (unlikely(ssk->sk_err)) {
819 if (!sock_owned_by_user(sk))
820 __mptcp_error_report(sk);
822 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
825 /* If the moves have caught up with the DATA_FIN sequence number
826 * it's time to ack the DATA_FIN and change socket state, but
827 * this is not a good place to change state. Let the workqueue
830 if (mptcp_pending_data_fin(sk, NULL))
831 mptcp_schedule_work(sk);
835 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
837 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
838 struct mptcp_sock *msk = mptcp_sk(sk);
839 int sk_rbuf, ssk_rbuf;
841 /* The peer can send data while we are shutting down this
842 * subflow at msk destruction time, but we must avoid enqueuing
843 * more data to the msk receive queue
845 if (unlikely(subflow->disposable))
848 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
849 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
850 if (unlikely(ssk_rbuf > sk_rbuf))
853 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
854 if (__mptcp_rmem(sk) > sk_rbuf) {
855 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
859 /* Wake-up the reader only for in-sequence data */
861 if (move_skbs_to_msk(msk, ssk))
862 sk->sk_data_ready(sk);
864 mptcp_data_unlock(sk);
867 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
869 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
870 WRITE_ONCE(msk->allow_infinite_fallback, false);
871 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
874 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
876 struct sock *sk = (struct sock *)msk;
878 if (sk->sk_state != TCP_ESTABLISHED)
881 /* attach to msk socket only after we are sure we will deal with it
884 if (sk->sk_socket && !ssk->sk_socket)
885 mptcp_sock_graft(ssk, sk->sk_socket);
887 mptcp_sockopt_sync_locked(msk, ssk);
888 mptcp_subflow_joined(msk, ssk);
889 mptcp_stop_tout_timer(sk);
893 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
895 struct mptcp_subflow_context *tmp, *subflow;
896 struct mptcp_sock *msk = mptcp_sk(sk);
898 list_for_each_entry_safe(subflow, tmp, join_list, node) {
899 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
900 bool slow = lock_sock_fast(ssk);
902 list_move_tail(&subflow->node, &msk->conn_list);
903 if (!__mptcp_finish_join(msk, ssk))
904 mptcp_subflow_reset(ssk);
905 unlock_sock_fast(ssk, slow);
909 static bool mptcp_rtx_timer_pending(struct sock *sk)
911 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
914 static void mptcp_reset_rtx_timer(struct sock *sk)
916 struct inet_connection_sock *icsk = inet_csk(sk);
919 /* prevent rescheduling on close */
920 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
923 tout = mptcp_sk(sk)->timer_ival;
924 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
927 bool mptcp_schedule_work(struct sock *sk)
929 if (inet_sk_state_load(sk) != TCP_CLOSE &&
930 schedule_work(&mptcp_sk(sk)->work)) {
931 /* each subflow already holds a reference to the sk, and the
932 * workqueue is invoked by a subflow, so sk can't go away here.
940 void mptcp_subflow_eof(struct sock *sk)
942 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
943 mptcp_schedule_work(sk);
946 static void mptcp_check_for_eof(struct mptcp_sock *msk)
948 struct mptcp_subflow_context *subflow;
949 struct sock *sk = (struct sock *)msk;
952 mptcp_for_each_subflow(msk, subflow)
953 receivers += !subflow->rx_eof;
957 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
958 /* hopefully temporary hack: propagate shutdown status
959 * to msk, when all subflows agree on it
961 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
963 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
964 sk->sk_data_ready(sk);
967 switch (sk->sk_state) {
968 case TCP_ESTABLISHED:
969 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
972 inet_sk_state_store(sk, TCP_CLOSING);
975 inet_sk_state_store(sk, TCP_CLOSE);
980 mptcp_close_wake_up(sk);
983 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
985 struct mptcp_subflow_context *subflow;
986 struct sock *sk = (struct sock *)msk;
988 sock_owned_by_me(sk);
990 mptcp_for_each_subflow(msk, subflow) {
991 if (READ_ONCE(subflow->data_avail))
992 return mptcp_subflow_tcp_sock(subflow);
998 static bool mptcp_skb_can_collapse_to(u64 write_seq,
999 const struct sk_buff *skb,
1000 const struct mptcp_ext *mpext)
1002 if (!tcp_skb_can_collapse_to(skb))
1005 /* can collapse only if MPTCP level sequence is in order and this
1006 * mapping has not been xmitted yet
1008 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
1012 /* we can append data to the given data frag if:
1013 * - there is space available in the backing page_frag
1014 * - the data frag tail matches the current page_frag free offset
1015 * - the data frag end sequence number matches the current write seq
1017 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
1018 const struct page_frag *pfrag,
1019 const struct mptcp_data_frag *df)
1021 return df && pfrag->page == df->page &&
1022 pfrag->size - pfrag->offset > 0 &&
1023 pfrag->offset == (df->offset + df->data_len) &&
1024 df->data_seq + df->data_len == msk->write_seq;
1027 static void dfrag_uncharge(struct sock *sk, int len)
1029 sk_mem_uncharge(sk, len);
1030 sk_wmem_queued_add(sk, -len);
1033 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1035 int len = dfrag->data_len + dfrag->overhead;
1037 list_del(&dfrag->list);
1038 dfrag_uncharge(sk, len);
1039 put_page(dfrag->page);
1042 static void __mptcp_clean_una(struct sock *sk)
1044 struct mptcp_sock *msk = mptcp_sk(sk);
1045 struct mptcp_data_frag *dtmp, *dfrag;
1048 /* on fallback we just need to ignore snd_una, as this is really
1051 if (__mptcp_check_fallback(msk))
1052 msk->snd_una = READ_ONCE(msk->snd_nxt);
1054 snd_una = msk->snd_una;
1055 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1056 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1059 if (unlikely(dfrag == msk->first_pending)) {
1060 /* in recovery mode can see ack after the current snd head */
1061 if (WARN_ON_ONCE(!msk->recovery))
1064 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1067 dfrag_clear(sk, dfrag);
1070 dfrag = mptcp_rtx_head(sk);
1071 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1072 u64 delta = snd_una - dfrag->data_seq;
1074 /* prevent wrap around in recovery mode */
1075 if (unlikely(delta > dfrag->already_sent)) {
1076 if (WARN_ON_ONCE(!msk->recovery))
1078 if (WARN_ON_ONCE(delta > dfrag->data_len))
1080 dfrag->already_sent += delta - dfrag->already_sent;
1083 dfrag->data_seq += delta;
1084 dfrag->offset += delta;
1085 dfrag->data_len -= delta;
1086 dfrag->already_sent -= delta;
1088 dfrag_uncharge(sk, delta);
1091 /* all retransmitted data acked, recovery completed */
1092 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1093 msk->recovery = false;
1096 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1097 snd_una == READ_ONCE(msk->write_seq)) {
1098 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1099 mptcp_stop_rtx_timer(sk);
1101 mptcp_reset_rtx_timer(sk);
1105 static void __mptcp_clean_una_wakeup(struct sock *sk)
1107 lockdep_assert_held_once(&sk->sk_lock.slock);
1109 __mptcp_clean_una(sk);
1110 mptcp_write_space(sk);
1113 static void mptcp_clean_una_wakeup(struct sock *sk)
1115 mptcp_data_lock(sk);
1116 __mptcp_clean_una_wakeup(sk);
1117 mptcp_data_unlock(sk);
1120 static void mptcp_enter_memory_pressure(struct sock *sk)
1122 struct mptcp_subflow_context *subflow;
1123 struct mptcp_sock *msk = mptcp_sk(sk);
1126 sk_stream_moderate_sndbuf(sk);
1127 mptcp_for_each_subflow(msk, subflow) {
1128 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1131 tcp_enter_memory_pressure(ssk);
1132 sk_stream_moderate_sndbuf(ssk);
1137 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1140 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1142 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1143 pfrag, sk->sk_allocation)))
1146 mptcp_enter_memory_pressure(sk);
1150 static struct mptcp_data_frag *
1151 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1154 int offset = ALIGN(orig_offset, sizeof(long));
1155 struct mptcp_data_frag *dfrag;
1157 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1158 dfrag->data_len = 0;
1159 dfrag->data_seq = msk->write_seq;
1160 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1161 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1162 dfrag->already_sent = 0;
1163 dfrag->page = pfrag->page;
1168 struct mptcp_sendmsg_info {
1174 bool data_lock_held;
1177 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1178 u64 data_seq, int avail_size)
1180 u64 window_end = mptcp_wnd_end(msk);
1183 if (__mptcp_check_fallback(msk))
1186 mptcp_snd_wnd = window_end - data_seq;
1187 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1189 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1190 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1191 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1197 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1199 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1203 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1207 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1209 struct sk_buff *skb;
1211 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1213 if (likely(__mptcp_add_ext(skb, gfp))) {
1214 skb_reserve(skb, MAX_TCP_HEADER);
1215 skb->ip_summed = CHECKSUM_PARTIAL;
1216 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1221 mptcp_enter_memory_pressure(sk);
1226 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1228 struct sk_buff *skb;
1230 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1234 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1235 tcp_skb_entail(ssk, skb);
1238 tcp_skb_tsorted_anchor_cleanup(skb);
1243 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1245 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1247 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1250 /* note: this always recompute the csum on the whole skb, even
1251 * if we just appended a single frag. More status info needed
1253 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1255 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1256 __wsum csum = ~csum_unfold(mpext->csum);
1257 int offset = skb->len - added;
1259 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1262 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1264 struct mptcp_ext *mpext)
1269 mpext->infinite_map = 1;
1270 mpext->data_len = 0;
1272 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1273 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1275 mptcp_do_fallback(ssk);
1278 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1280 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1281 struct mptcp_data_frag *dfrag,
1282 struct mptcp_sendmsg_info *info)
1284 u64 data_seq = dfrag->data_seq + info->sent;
1285 int offset = dfrag->offset + info->sent;
1286 struct mptcp_sock *msk = mptcp_sk(sk);
1287 bool zero_window_probe = false;
1288 struct mptcp_ext *mpext = NULL;
1289 bool can_coalesce = false;
1290 bool reuse_skb = true;
1291 struct sk_buff *skb;
1295 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1296 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1298 if (WARN_ON_ONCE(info->sent > info->limit ||
1299 info->limit > dfrag->data_len))
1302 if (unlikely(!__tcp_can_send(ssk)))
1305 /* compute send limit */
1306 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1307 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1308 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1309 copy = info->size_goal;
1311 skb = tcp_write_queue_tail(ssk);
1312 if (skb && copy > skb->len) {
1313 /* Limit the write to the size available in the
1314 * current skb, if any, so that we create at most a new skb.
1315 * Explicitly tells TCP internals to avoid collapsing on later
1316 * queue management operation, to avoid breaking the ext <->
1317 * SSN association set here
1319 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1320 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1321 TCP_SKB_CB(skb)->eor = 1;
1322 tcp_mark_push(tcp_sk(ssk), skb);
1326 i = skb_shinfo(skb)->nr_frags;
1327 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1328 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1329 tcp_mark_push(tcp_sk(ssk), skb);
1336 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1340 i = skb_shinfo(skb)->nr_frags;
1342 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1345 /* Zero window and all data acked? Probe. */
1346 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1348 u64 snd_una = READ_ONCE(msk->snd_una);
1350 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1351 tcp_remove_empty_skb(ssk);
1355 zero_window_probe = true;
1356 data_seq = snd_una - 1;
1360 copy = min_t(size_t, copy, info->limit - info->sent);
1361 if (!sk_wmem_schedule(ssk, copy)) {
1362 tcp_remove_empty_skb(ssk);
1367 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1369 get_page(dfrag->page);
1370 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1374 skb->data_len += copy;
1375 skb->truesize += copy;
1376 sk_wmem_queued_add(ssk, copy);
1377 sk_mem_charge(ssk, copy);
1378 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1379 TCP_SKB_CB(skb)->end_seq += copy;
1380 tcp_skb_pcount_set(skb, 0);
1382 /* on skb reuse we just need to update the DSS len */
1384 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1385 mpext->data_len += copy;
1389 memset(mpext, 0, sizeof(*mpext));
1390 mpext->data_seq = data_seq;
1391 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1392 mpext->data_len = copy;
1396 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1397 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1400 if (zero_window_probe) {
1401 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1403 if (READ_ONCE(msk->csum_enabled))
1404 mptcp_update_data_checksum(skb, copy);
1405 tcp_push_pending_frames(ssk);
1409 if (READ_ONCE(msk->csum_enabled))
1410 mptcp_update_data_checksum(skb, copy);
1411 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1412 mptcp_update_infinite_map(msk, ssk, mpext);
1413 trace_mptcp_sendmsg_frag(mpext);
1414 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1418 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1419 sizeof(struct tcphdr) - \
1420 MAX_TCP_OPTION_SPACE - \
1421 sizeof(struct ipv6hdr) - \
1422 sizeof(struct frag_hdr))
1424 struct subflow_send_info {
1429 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1431 if (!subflow->stale)
1435 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1438 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1440 if (unlikely(subflow->stale)) {
1441 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1443 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1446 mptcp_subflow_set_active(subflow);
1448 return __mptcp_subflow_active(subflow);
1451 #define SSK_MODE_ACTIVE 0
1452 #define SSK_MODE_BACKUP 1
1453 #define SSK_MODE_MAX 2
1455 /* implement the mptcp packet scheduler;
1456 * returns the subflow that will transmit the next DSS
1457 * additionally updates the rtx timeout
1459 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1461 struct subflow_send_info send_info[SSK_MODE_MAX];
1462 struct mptcp_subflow_context *subflow;
1463 struct sock *sk = (struct sock *)msk;
1464 u32 pace, burst, wmem;
1465 int i, nr_active = 0;
1470 sock_owned_by_me(sk);
1472 if (__mptcp_check_fallback(msk)) {
1475 return __tcp_can_send(msk->first) &&
1476 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1479 /* re-use last subflow, if the burst allow that */
1480 if (msk->last_snd && msk->snd_burst > 0 &&
1481 sk_stream_memory_free(msk->last_snd) &&
1482 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1483 mptcp_set_timeout(sk);
1484 return msk->last_snd;
1487 /* pick the subflow with the lower wmem/wspace ratio */
1488 for (i = 0; i < SSK_MODE_MAX; ++i) {
1489 send_info[i].ssk = NULL;
1490 send_info[i].linger_time = -1;
1493 mptcp_for_each_subflow(msk, subflow) {
1494 trace_mptcp_subflow_get_send(subflow);
1495 ssk = mptcp_subflow_tcp_sock(subflow);
1496 if (!mptcp_subflow_active(subflow))
1499 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1500 nr_active += !subflow->backup;
1501 pace = subflow->avg_pacing_rate;
1502 if (unlikely(!pace)) {
1503 /* init pacing rate from socket */
1504 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1505 pace = subflow->avg_pacing_rate;
1510 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1511 if (linger_time < send_info[subflow->backup].linger_time) {
1512 send_info[subflow->backup].ssk = ssk;
1513 send_info[subflow->backup].linger_time = linger_time;
1516 __mptcp_set_timeout(sk, tout);
1518 /* pick the best backup if no other subflow is active */
1520 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1522 /* According to the blest algorithm, to avoid HoL blocking for the
1523 * faster flow, we need to:
1524 * - estimate the faster flow linger time
1525 * - use the above to estimate the amount of byte transferred
1526 * by the faster flow
1527 * - check that the amount of queued data is greter than the above,
1528 * otherwise do not use the picked, slower, subflow
1529 * We select the subflow with the shorter estimated time to flush
1530 * the queued mem, which basically ensure the above. We just need
1531 * to check that subflow has a non empty cwin.
1533 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1534 if (!ssk || !sk_stream_memory_free(ssk))
1537 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1538 wmem = READ_ONCE(ssk->sk_wmem_queued);
1540 msk->last_snd = NULL;
1544 subflow = mptcp_subflow_ctx(ssk);
1545 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1546 READ_ONCE(ssk->sk_pacing_rate) * burst,
1548 msk->last_snd = ssk;
1549 msk->snd_burst = burst;
1553 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1555 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1559 static void mptcp_update_post_push(struct mptcp_sock *msk,
1560 struct mptcp_data_frag *dfrag,
1563 u64 snd_nxt_new = dfrag->data_seq;
1565 dfrag->already_sent += sent;
1567 msk->snd_burst -= sent;
1569 snd_nxt_new += dfrag->already_sent;
1571 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1572 * is recovering after a failover. In that event, this re-sends
1575 * Thus compute snd_nxt_new candidate based on
1576 * the dfrag->data_seq that was sent and the data
1577 * that has been handed to the subflow for transmission
1578 * and skip update in case it was old dfrag.
1580 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1581 msk->snd_nxt = snd_nxt_new;
1584 void mptcp_check_and_set_pending(struct sock *sk)
1586 if (mptcp_send_head(sk)) {
1587 mptcp_data_lock(sk);
1588 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1589 mptcp_data_unlock(sk);
1593 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1595 struct sock *prev_ssk = NULL, *ssk = NULL;
1596 struct mptcp_sock *msk = mptcp_sk(sk);
1597 struct mptcp_sendmsg_info info = {
1600 bool do_check_data_fin = false;
1601 struct mptcp_data_frag *dfrag;
1604 while ((dfrag = mptcp_send_head(sk))) {
1605 info.sent = dfrag->already_sent;
1606 info.limit = dfrag->data_len;
1607 len = dfrag->data_len - dfrag->already_sent;
1612 ssk = mptcp_subflow_get_send(msk);
1614 /* First check. If the ssk has changed since
1615 * the last round, release prev_ssk
1617 if (ssk != prev_ssk && prev_ssk)
1618 mptcp_push_release(prev_ssk, &info);
1622 /* Need to lock the new subflow only if different
1623 * from the previous one, otherwise we are still
1624 * helding the relevant lock
1626 if (ssk != prev_ssk)
1629 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1633 mptcp_push_release(ssk, &info);
1637 do_check_data_fin = true;
1641 mptcp_update_post_push(msk, dfrag, ret);
1643 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1646 /* at this point we held the socket lock for the last subflow we used */
1648 mptcp_push_release(ssk, &info);
1651 /* ensure the rtx timer is running */
1652 if (!mptcp_rtx_timer_pending(sk))
1653 mptcp_reset_rtx_timer(sk);
1654 if (do_check_data_fin)
1655 mptcp_check_send_data_fin(sk);
1658 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1660 struct mptcp_sock *msk = mptcp_sk(sk);
1661 struct mptcp_sendmsg_info info = {
1662 .data_lock_held = true,
1664 struct mptcp_data_frag *dfrag;
1665 struct sock *xmit_ssk;
1666 int len, copied = 0;
1670 while ((dfrag = mptcp_send_head(sk))) {
1671 info.sent = dfrag->already_sent;
1672 info.limit = dfrag->data_len;
1673 len = dfrag->data_len - dfrag->already_sent;
1677 /* the caller already invoked the packet scheduler,
1678 * check for a different subflow usage only after
1679 * spooling the first chunk of data
1681 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1684 if (xmit_ssk != ssk) {
1685 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1686 MPTCP_DELEGATE_SEND);
1690 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1699 mptcp_update_post_push(msk, dfrag, ret);
1701 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1705 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1706 * not going to flush it via release_sock()
1709 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1711 if (!mptcp_rtx_timer_pending(sk))
1712 mptcp_reset_rtx_timer(sk);
1714 if (msk->snd_data_fin_enable &&
1715 msk->snd_nxt + 1 == msk->write_seq)
1716 mptcp_schedule_work(sk);
1720 static void mptcp_set_nospace(struct sock *sk)
1722 /* enable autotune */
1723 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1725 /* will be cleared on avail space */
1726 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1729 static int mptcp_disconnect(struct sock *sk, int flags);
1731 static int mptcp_sendmsg_fastopen(struct sock *sk, struct sock *ssk, struct msghdr *msg,
1732 size_t len, int *copied_syn)
1734 unsigned int saved_flags = msg->msg_flags;
1735 struct mptcp_sock *msk = mptcp_sk(sk);
1739 msg->msg_flags |= MSG_DONTWAIT;
1740 msk->fastopening = 1;
1741 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1742 msk->fastopening = 0;
1743 msg->msg_flags = saved_flags;
1746 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1747 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1748 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1749 msg->msg_namelen, msg->msg_flags, 1);
1751 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1752 * case of any error, except timeout or signal
1754 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1756 } else if (ret && ret != -EINPROGRESS) {
1757 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1758 * __inet_stream_connect() can fail, due to looking check,
1759 * see mptcp_disconnect().
1760 * Attempt it again outside the problematic scope.
1762 if (!mptcp_disconnect(sk, 0))
1763 sk->sk_socket->state = SS_UNCONNECTED;
1769 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1771 struct mptcp_sock *msk = mptcp_sk(sk);
1772 struct page_frag *pfrag;
1773 struct socket *ssock;
1778 /* we don't support FASTOPEN yet */
1779 if (msg->msg_flags & MSG_FASTOPEN)
1782 /* silently ignore everything else */
1783 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1787 ssock = __mptcp_nmpc_socket(msk);
1788 if (unlikely(ssock && inet_sk(ssock->sk)->defer_connect)) {
1791 ret = mptcp_sendmsg_fastopen(sk, ssock->sk, msg, len, &copied_syn);
1792 copied += copied_syn;
1793 if (ret == -EINPROGRESS && copied_syn > 0)
1799 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1801 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1802 ret = sk_stream_wait_connect(sk, &timeo);
1808 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1811 pfrag = sk_page_frag(sk);
1813 while (msg_data_left(msg)) {
1814 int total_ts, frag_truesize = 0;
1815 struct mptcp_data_frag *dfrag;
1816 bool dfrag_collapsed;
1817 size_t psize, offset;
1819 /* reuse tail pfrag, if possible, or carve a new one from the
1822 dfrag = mptcp_pending_tail(sk);
1823 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1824 if (!dfrag_collapsed) {
1825 if (!sk_stream_memory_free(sk))
1826 goto wait_for_memory;
1828 if (!mptcp_page_frag_refill(sk, pfrag))
1829 goto wait_for_memory;
1831 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1832 frag_truesize = dfrag->overhead;
1835 /* we do not bound vs wspace, to allow a single packet.
1836 * memory accounting will prevent execessive memory usage
1839 offset = dfrag->offset + dfrag->data_len;
1840 psize = pfrag->size - offset;
1841 psize = min_t(size_t, psize, msg_data_left(msg));
1842 total_ts = psize + frag_truesize;
1844 if (!sk_wmem_schedule(sk, total_ts))
1845 goto wait_for_memory;
1847 if (copy_page_from_iter(dfrag->page, offset, psize,
1848 &msg->msg_iter) != psize) {
1853 /* data successfully copied into the write queue */
1854 sk_forward_alloc_add(sk, -total_ts);
1856 dfrag->data_len += psize;
1857 frag_truesize += psize;
1858 pfrag->offset += frag_truesize;
1859 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1861 /* charge data on mptcp pending queue to the msk socket
1862 * Note: we charge such data both to sk and ssk
1864 sk_wmem_queued_add(sk, frag_truesize);
1865 if (!dfrag_collapsed) {
1866 get_page(dfrag->page);
1867 list_add_tail(&dfrag->list, &msk->rtx_queue);
1868 if (!msk->first_pending)
1869 WRITE_ONCE(msk->first_pending, dfrag);
1871 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1872 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1878 mptcp_set_nospace(sk);
1879 __mptcp_push_pending(sk, msg->msg_flags);
1880 ret = sk_stream_wait_memory(sk, &timeo);
1886 __mptcp_push_pending(sk, msg->msg_flags);
1896 copied = sk_stream_error(sk, msg->msg_flags, ret);
1900 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1902 size_t len, int flags,
1903 struct scm_timestamping_internal *tss,
1906 struct sk_buff *skb, *tmp;
1909 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1910 u32 offset = MPTCP_SKB_CB(skb)->offset;
1911 u32 data_len = skb->len - offset;
1912 u32 count = min_t(size_t, len - copied, data_len);
1915 if (!(flags & MSG_TRUNC)) {
1916 err = skb_copy_datagram_msg(skb, offset, msg, count);
1917 if (unlikely(err < 0)) {
1924 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1925 tcp_update_recv_tstamps(skb, tss);
1926 *cmsg_flags |= MPTCP_CMSG_TS;
1931 if (count < data_len) {
1932 if (!(flags & MSG_PEEK)) {
1933 MPTCP_SKB_CB(skb)->offset += count;
1934 MPTCP_SKB_CB(skb)->map_seq += count;
1939 if (!(flags & MSG_PEEK)) {
1940 /* we will bulk release the skb memory later */
1941 skb->destructor = NULL;
1942 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1943 __skb_unlink(skb, &msk->receive_queue);
1954 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1956 * Only difference: Use highest rtt estimate of the subflows in use.
1958 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1960 struct mptcp_subflow_context *subflow;
1961 struct sock *sk = (struct sock *)msk;
1962 u32 time, advmss = 1;
1965 sock_owned_by_me(sk);
1970 msk->rcvq_space.copied += copied;
1972 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1973 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1975 rtt_us = msk->rcvq_space.rtt_us;
1976 if (rtt_us && time < (rtt_us >> 3))
1980 mptcp_for_each_subflow(msk, subflow) {
1981 const struct tcp_sock *tp;
1985 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1987 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1988 sf_advmss = READ_ONCE(tp->advmss);
1990 rtt_us = max(sf_rtt_us, rtt_us);
1991 advmss = max(sf_advmss, advmss);
1994 msk->rcvq_space.rtt_us = rtt_us;
1995 if (time < (rtt_us >> 3) || rtt_us == 0)
1998 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2001 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2002 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2006 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2008 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2010 do_div(grow, msk->rcvq_space.space);
2011 rcvwin += (grow << 1);
2013 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
2014 while (tcp_win_from_space(sk, rcvmem) < advmss)
2017 do_div(rcvwin, advmss);
2018 rcvbuf = min_t(u64, rcvwin * rcvmem,
2019 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2021 if (rcvbuf > sk->sk_rcvbuf) {
2024 window_clamp = tcp_win_from_space(sk, rcvbuf);
2025 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2027 /* Make subflows follow along. If we do not do this, we
2028 * get drops at subflow level if skbs can't be moved to
2029 * the mptcp rx queue fast enough (announced rcv_win can
2030 * exceed ssk->sk_rcvbuf).
2032 mptcp_for_each_subflow(msk, subflow) {
2036 ssk = mptcp_subflow_tcp_sock(subflow);
2037 slow = lock_sock_fast(ssk);
2038 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2039 tcp_sk(ssk)->window_clamp = window_clamp;
2040 tcp_cleanup_rbuf(ssk, 1);
2041 unlock_sock_fast(ssk, slow);
2046 msk->rcvq_space.space = msk->rcvq_space.copied;
2048 msk->rcvq_space.copied = 0;
2049 msk->rcvq_space.time = mstamp;
2052 static void __mptcp_update_rmem(struct sock *sk)
2054 struct mptcp_sock *msk = mptcp_sk(sk);
2056 if (!msk->rmem_released)
2059 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2060 mptcp_rmem_uncharge(sk, msk->rmem_released);
2061 WRITE_ONCE(msk->rmem_released, 0);
2064 static void __mptcp_splice_receive_queue(struct sock *sk)
2066 struct mptcp_sock *msk = mptcp_sk(sk);
2068 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2071 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2073 struct sock *sk = (struct sock *)msk;
2074 unsigned int moved = 0;
2078 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2081 /* we can have data pending in the subflows only if the msk
2082 * receive buffer was full at subflow_data_ready() time,
2083 * that is an unlikely slow path.
2088 slowpath = lock_sock_fast(ssk);
2089 mptcp_data_lock(sk);
2090 __mptcp_update_rmem(sk);
2091 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2092 mptcp_data_unlock(sk);
2094 if (unlikely(ssk->sk_err))
2095 __mptcp_error_report(sk);
2096 unlock_sock_fast(ssk, slowpath);
2099 /* acquire the data lock only if some input data is pending */
2101 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2102 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2103 mptcp_data_lock(sk);
2104 __mptcp_update_rmem(sk);
2105 ret |= __mptcp_ofo_queue(msk);
2106 __mptcp_splice_receive_queue(sk);
2107 mptcp_data_unlock(sk);
2110 mptcp_check_data_fin((struct sock *)msk);
2111 return !skb_queue_empty(&msk->receive_queue);
2114 static unsigned int mptcp_inq_hint(const struct sock *sk)
2116 const struct mptcp_sock *msk = mptcp_sk(sk);
2117 const struct sk_buff *skb;
2119 skb = skb_peek(&msk->receive_queue);
2121 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2123 if (hint_val >= INT_MAX)
2126 return (unsigned int)hint_val;
2129 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2135 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2136 int flags, int *addr_len)
2138 struct mptcp_sock *msk = mptcp_sk(sk);
2139 struct scm_timestamping_internal tss;
2140 int copied = 0, cmsg_flags = 0;
2144 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2145 if (unlikely(flags & MSG_ERRQUEUE))
2146 return inet_recv_error(sk, msg, len, addr_len);
2149 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2154 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2156 len = min_t(size_t, len, INT_MAX);
2157 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2159 if (unlikely(msk->recvmsg_inq))
2160 cmsg_flags = MPTCP_CMSG_INQ;
2162 while (copied < len) {
2165 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2166 if (unlikely(bytes_read < 0)) {
2168 copied = bytes_read;
2172 copied += bytes_read;
2174 /* be sure to advertise window change */
2175 mptcp_cleanup_rbuf(msk);
2177 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2180 /* only the master socket status is relevant here. The exit
2181 * conditions mirror closely tcp_recvmsg()
2183 if (copied >= target)
2188 sk->sk_state == TCP_CLOSE ||
2189 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2191 signal_pending(current))
2195 copied = sock_error(sk);
2199 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2200 mptcp_check_for_eof(msk);
2202 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2203 /* race breaker: the shutdown could be after the
2204 * previous receive queue check
2206 if (__mptcp_move_skbs(msk))
2211 if (sk->sk_state == TCP_CLOSE) {
2221 if (signal_pending(current)) {
2222 copied = sock_intr_errno(timeo);
2227 pr_debug("block timeout %ld", timeo);
2228 sk_wait_data(sk, &timeo, NULL);
2232 if (cmsg_flags && copied >= 0) {
2233 if (cmsg_flags & MPTCP_CMSG_TS)
2234 tcp_recv_timestamp(msg, sk, &tss);
2236 if (cmsg_flags & MPTCP_CMSG_INQ) {
2237 unsigned int inq = mptcp_inq_hint(sk);
2239 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2243 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2244 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2245 skb_queue_empty(&msk->receive_queue), copied);
2246 if (!(flags & MSG_PEEK))
2247 mptcp_rcv_space_adjust(msk, copied);
2253 static void mptcp_retransmit_timer(struct timer_list *t)
2255 struct inet_connection_sock *icsk = from_timer(icsk, t,
2256 icsk_retransmit_timer);
2257 struct sock *sk = &icsk->icsk_inet.sk;
2258 struct mptcp_sock *msk = mptcp_sk(sk);
2261 if (!sock_owned_by_user(sk)) {
2262 /* we need a process context to retransmit */
2263 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2264 mptcp_schedule_work(sk);
2266 /* delegate our work to tcp_release_cb() */
2267 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2273 static void mptcp_tout_timer(struct timer_list *t)
2275 struct sock *sk = from_timer(sk, t, sk_timer);
2277 mptcp_schedule_work(sk);
2281 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2284 * A backup subflow is returned only if that is the only kind available.
2286 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2288 struct sock *backup = NULL, *pick = NULL;
2289 struct mptcp_subflow_context *subflow;
2290 int min_stale_count = INT_MAX;
2292 sock_owned_by_me((const struct sock *)msk);
2294 if (__mptcp_check_fallback(msk))
2297 mptcp_for_each_subflow(msk, subflow) {
2298 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2300 if (!__mptcp_subflow_active(subflow))
2303 /* still data outstanding at TCP level? skip this */
2304 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2305 mptcp_pm_subflow_chk_stale(msk, ssk);
2306 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2310 if (subflow->backup) {
2323 /* use backup only if there are no progresses anywhere */
2324 return min_stale_count > 1 ? backup : NULL;
2327 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2330 iput(SOCK_INODE(msk->subflow));
2331 WRITE_ONCE(msk->subflow, NULL);
2335 bool __mptcp_retransmit_pending_data(struct sock *sk)
2337 struct mptcp_data_frag *cur, *rtx_head;
2338 struct mptcp_sock *msk = mptcp_sk(sk);
2340 if (__mptcp_check_fallback(mptcp_sk(sk)))
2343 /* the closing socket has some data untransmitted and/or unacked:
2344 * some data in the mptcp rtx queue has not really xmitted yet.
2345 * keep it simple and re-inject the whole mptcp level rtx queue
2347 mptcp_data_lock(sk);
2348 __mptcp_clean_una_wakeup(sk);
2349 rtx_head = mptcp_rtx_head(sk);
2351 mptcp_data_unlock(sk);
2355 msk->recovery_snd_nxt = msk->snd_nxt;
2356 msk->recovery = true;
2357 mptcp_data_unlock(sk);
2359 msk->first_pending = rtx_head;
2362 /* be sure to clear the "sent status" on all re-injected fragments */
2363 list_for_each_entry(cur, &msk->rtx_queue, list) {
2364 if (!cur->already_sent)
2366 cur->already_sent = 0;
2372 /* flags for __mptcp_close_ssk() */
2373 #define MPTCP_CF_PUSH BIT(1)
2374 #define MPTCP_CF_FASTCLOSE BIT(2)
2376 /* be sure to send a reset only if the caller asked for it, also
2377 * clean completely the subflow status when the subflow reaches
2380 static void __mptcp_subflow_disconnect(struct sock *ssk,
2381 struct mptcp_subflow_context *subflow,
2384 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2385 (flags & MPTCP_CF_FASTCLOSE)) {
2386 /* The MPTCP code never wait on the subflow sockets, TCP-level
2387 * disconnect should never fail
2389 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2390 mptcp_subflow_ctx_reset(subflow);
2392 tcp_shutdown(ssk, SEND_SHUTDOWN);
2396 /* subflow sockets can be either outgoing (connect) or incoming
2399 * Outgoing subflows use in-kernel sockets.
2400 * Incoming subflows do not have their own 'struct socket' allocated,
2401 * so we need to use tcp_close() after detaching them from the mptcp
2404 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2405 struct mptcp_subflow_context *subflow,
2408 struct mptcp_sock *msk = mptcp_sk(sk);
2409 bool dispose_it, need_push = false;
2411 /* If the first subflow moved to a close state before accept, e.g. due
2412 * to an incoming reset or listener shutdown, the subflow socket is
2413 * already deleted by inet_child_forget() and the mptcp socket can't
2416 if (msk->in_accept_queue && msk->first == ssk &&
2417 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2418 /* ensure later check in mptcp_worker() will dispose the msk */
2419 mptcp_set_close_tout(sk, tcp_jiffies32 - (TCP_TIMEWAIT_LEN + 1));
2420 sock_set_flag(sk, SOCK_DEAD);
2421 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2422 mptcp_subflow_drop_ctx(ssk);
2426 dispose_it = msk->free_first || ssk != msk->first;
2428 list_del(&subflow->node);
2430 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2432 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2433 /* be sure to force the tcp_close path
2434 * to generate the egress reset
2436 ssk->sk_lingertime = 0;
2437 sock_set_flag(ssk, SOCK_LINGER);
2438 subflow->send_fastclose = 1;
2441 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2443 __mptcp_subflow_disconnect(ssk, subflow, flags);
2444 if (msk->subflow && ssk == msk->subflow->sk)
2445 msk->subflow->state = SS_UNCONNECTED;
2451 subflow->disposable = 1;
2453 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2454 * the ssk has been already destroyed, we just need to release the
2455 * reference owned by msk;
2457 if (!inet_csk(ssk)->icsk_ulp_ops) {
2458 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2459 kfree_rcu(subflow, rcu);
2461 /* otherwise tcp will dispose of the ssk and subflow ctx */
2462 __tcp_close(ssk, 0);
2464 /* close acquired an extra ref */
2469 __mptcp_subflow_error_report(sk, ssk);
2474 if (ssk == msk->first)
2475 WRITE_ONCE(msk->first, NULL);
2478 if (ssk == msk->last_snd)
2479 msk->last_snd = NULL;
2482 __mptcp_push_pending(sk, 0);
2484 /* Catch every 'all subflows closed' scenario, including peers silently
2485 * closing them, e.g. due to timeout.
2486 * For established sockets, allow an additional timeout before closing,
2487 * as the protocol can still create more subflows.
2489 if (list_is_singular(&msk->conn_list) && msk->first &&
2490 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2491 if (sk->sk_state != TCP_ESTABLISHED ||
2492 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2493 inet_sk_state_store(sk, TCP_CLOSE);
2494 mptcp_close_wake_up(sk);
2496 mptcp_start_tout_timer(sk);
2501 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2502 struct mptcp_subflow_context *subflow)
2504 if (sk->sk_state == TCP_ESTABLISHED)
2505 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2507 /* subflow aborted before reaching the fully_established status
2508 * attempt the creation of the next subflow
2510 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2512 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2515 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2520 static void __mptcp_close_subflow(struct sock *sk)
2522 struct mptcp_subflow_context *subflow, *tmp;
2523 struct mptcp_sock *msk = mptcp_sk(sk);
2527 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2528 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2530 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2533 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2534 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2537 mptcp_close_ssk(sk, ssk, subflow);
2542 static bool mptcp_close_tout_expired(const struct sock *sk)
2544 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2545 sk->sk_state == TCP_CLOSE)
2548 return time_after32(tcp_jiffies32,
2549 inet_csk(sk)->icsk_mtup.probe_timestamp + TCP_TIMEWAIT_LEN);
2552 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2554 struct mptcp_subflow_context *subflow, *tmp;
2555 struct sock *sk = &msk->sk.icsk_inet.sk;
2557 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2560 mptcp_token_destroy(msk);
2562 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2563 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2566 slow = lock_sock_fast(tcp_sk);
2567 if (tcp_sk->sk_state != TCP_CLOSE) {
2568 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2569 tcp_set_state(tcp_sk, TCP_CLOSE);
2571 unlock_sock_fast(tcp_sk, slow);
2574 /* Mirror the tcp_reset() error propagation */
2575 switch (sk->sk_state) {
2577 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2579 case TCP_CLOSE_WAIT:
2580 WRITE_ONCE(sk->sk_err, EPIPE);
2585 WRITE_ONCE(sk->sk_err, ECONNRESET);
2588 inet_sk_state_store(sk, TCP_CLOSE);
2589 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2590 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2591 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2593 /* the calling mptcp_worker will properly destroy the socket */
2594 if (sock_flag(sk, SOCK_DEAD))
2597 sk->sk_state_change(sk);
2598 sk_error_report(sk);
2601 static void __mptcp_retrans(struct sock *sk)
2603 struct mptcp_sock *msk = mptcp_sk(sk);
2604 struct mptcp_sendmsg_info info = {};
2605 struct mptcp_data_frag *dfrag;
2610 mptcp_clean_una_wakeup(sk);
2612 /* first check ssk: need to kick "stale" logic */
2613 ssk = mptcp_subflow_get_retrans(msk);
2614 dfrag = mptcp_rtx_head(sk);
2616 if (mptcp_data_fin_enabled(msk)) {
2617 struct inet_connection_sock *icsk = inet_csk(sk);
2619 icsk->icsk_retransmits++;
2620 mptcp_set_datafin_timeout(sk);
2621 mptcp_send_ack(msk);
2626 if (!mptcp_send_head(sk))
2637 /* limit retransmission to the bytes already sent on some subflows */
2639 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2640 while (info.sent < info.limit) {
2641 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2645 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2650 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2651 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2653 WRITE_ONCE(msk->allow_infinite_fallback, false);
2659 mptcp_check_and_set_pending(sk);
2661 if (!mptcp_rtx_timer_pending(sk))
2662 mptcp_reset_rtx_timer(sk);
2665 /* schedule the timeout timer for the relevant event: either close timeout
2666 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2668 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2670 struct sock *sk = (struct sock *)msk;
2671 unsigned long timeout, close_timeout;
2673 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2676 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2679 /* the close timeout takes precedence on the fail one, and here at least one of
2682 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2684 sk_reset_timer(sk, &sk->sk_timer, timeout);
2687 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2689 struct sock *ssk = msk->first;
2695 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2697 slow = lock_sock_fast(ssk);
2698 mptcp_subflow_reset(ssk);
2699 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2700 unlock_sock_fast(ssk, slow);
2703 static void mptcp_do_fastclose(struct sock *sk)
2705 struct mptcp_subflow_context *subflow, *tmp;
2706 struct mptcp_sock *msk = mptcp_sk(sk);
2708 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2709 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2710 subflow, MPTCP_CF_FASTCLOSE);
2713 static void mptcp_worker(struct work_struct *work)
2715 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2716 struct sock *sk = &msk->sk.icsk_inet.sk;
2717 unsigned long fail_tout;
2721 state = sk->sk_state;
2722 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2725 mptcp_check_fastclose(msk);
2727 mptcp_pm_nl_work(msk);
2729 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2730 mptcp_check_for_eof(msk);
2732 mptcp_check_send_data_fin(sk);
2733 mptcp_check_data_fin_ack(sk);
2734 mptcp_check_data_fin(sk);
2736 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2737 __mptcp_close_subflow(sk);
2739 if (mptcp_close_tout_expired(sk)) {
2740 inet_sk_state_store(sk, TCP_CLOSE);
2741 mptcp_do_fastclose(sk);
2742 mptcp_close_wake_up(sk);
2745 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2746 __mptcp_destroy_sock(sk);
2750 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2751 __mptcp_retrans(sk);
2753 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2754 if (fail_tout && time_after(jiffies, fail_tout))
2755 mptcp_mp_fail_no_response(msk);
2762 static int __mptcp_init_sock(struct sock *sk)
2764 struct mptcp_sock *msk = mptcp_sk(sk);
2766 INIT_LIST_HEAD(&msk->conn_list);
2767 INIT_LIST_HEAD(&msk->join_list);
2768 INIT_LIST_HEAD(&msk->rtx_queue);
2769 INIT_WORK(&msk->work, mptcp_worker);
2770 __skb_queue_head_init(&msk->receive_queue);
2771 msk->out_of_order_queue = RB_ROOT;
2772 msk->first_pending = NULL;
2773 msk->rmem_fwd_alloc = 0;
2774 WRITE_ONCE(msk->rmem_released, 0);
2775 msk->timer_ival = TCP_RTO_MIN;
2777 WRITE_ONCE(msk->first, NULL);
2778 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2779 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2780 WRITE_ONCE(msk->allow_infinite_fallback, true);
2781 msk->recovery = false;
2783 mptcp_pm_data_init(msk);
2785 /* re-use the csk retrans timer for MPTCP-level retrans */
2786 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2787 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2792 static void mptcp_ca_reset(struct sock *sk)
2794 struct inet_connection_sock *icsk = inet_csk(sk);
2796 tcp_assign_congestion_control(sk);
2797 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2799 /* no need to keep a reference to the ops, the name will suffice */
2800 tcp_cleanup_congestion_control(sk);
2801 icsk->icsk_ca_ops = NULL;
2804 static int mptcp_init_sock(struct sock *sk)
2806 struct net *net = sock_net(sk);
2809 ret = __mptcp_init_sock(sk);
2813 if (!mptcp_is_enabled(net))
2814 return -ENOPROTOOPT;
2816 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2819 ret = __mptcp_socket_create(mptcp_sk(sk));
2823 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2824 * propagate the correct value
2828 sk_sockets_allocated_inc(sk);
2829 sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2830 sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2835 static void __mptcp_clear_xmit(struct sock *sk)
2837 struct mptcp_sock *msk = mptcp_sk(sk);
2838 struct mptcp_data_frag *dtmp, *dfrag;
2840 WRITE_ONCE(msk->first_pending, NULL);
2841 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2842 dfrag_clear(sk, dfrag);
2845 void mptcp_cancel_work(struct sock *sk)
2847 struct mptcp_sock *msk = mptcp_sk(sk);
2849 if (cancel_work_sync(&msk->work))
2853 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2857 switch (ssk->sk_state) {
2859 if (!(how & RCV_SHUTDOWN))
2863 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2866 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2867 pr_debug("Fallback");
2868 ssk->sk_shutdown |= how;
2869 tcp_shutdown(ssk, how);
2871 /* simulate the data_fin ack reception to let the state
2872 * machine move forward
2874 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2875 mptcp_schedule_work(sk);
2877 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2879 if (!mptcp_rtx_timer_pending(sk))
2880 mptcp_reset_rtx_timer(sk);
2888 static const unsigned char new_state[16] = {
2889 /* current state: new state: action: */
2890 [0 /* (Invalid) */] = TCP_CLOSE,
2891 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2892 [TCP_SYN_SENT] = TCP_CLOSE,
2893 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2894 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2895 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2896 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2897 [TCP_CLOSE] = TCP_CLOSE,
2898 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2899 [TCP_LAST_ACK] = TCP_LAST_ACK,
2900 [TCP_LISTEN] = TCP_CLOSE,
2901 [TCP_CLOSING] = TCP_CLOSING,
2902 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2905 static int mptcp_close_state(struct sock *sk)
2907 int next = (int)new_state[sk->sk_state];
2908 int ns = next & TCP_STATE_MASK;
2910 inet_sk_state_store(sk, ns);
2912 return next & TCP_ACTION_FIN;
2915 static void mptcp_check_send_data_fin(struct sock *sk)
2917 struct mptcp_subflow_context *subflow;
2918 struct mptcp_sock *msk = mptcp_sk(sk);
2920 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2921 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2922 msk->snd_nxt, msk->write_seq);
2924 /* we still need to enqueue subflows or not really shutting down,
2927 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2928 mptcp_send_head(sk))
2931 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2933 mptcp_for_each_subflow(msk, subflow) {
2934 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2936 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2940 static void __mptcp_wr_shutdown(struct sock *sk)
2942 struct mptcp_sock *msk = mptcp_sk(sk);
2944 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2945 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2946 !!mptcp_send_head(sk));
2948 /* will be ignored by fallback sockets */
2949 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2950 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2952 mptcp_check_send_data_fin(sk);
2955 static void __mptcp_destroy_sock(struct sock *sk)
2957 struct mptcp_sock *msk = mptcp_sk(sk);
2959 pr_debug("msk=%p", msk);
2963 mptcp_stop_rtx_timer(sk);
2964 sk_stop_timer(sk, &sk->sk_timer);
2967 sk->sk_prot->destroy(sk);
2969 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2970 WARN_ON_ONCE(msk->rmem_released);
2971 sk_stream_kill_queues(sk);
2972 xfrm_sk_free_policy(sk);
2974 sk_refcnt_debug_release(sk);
2978 void __mptcp_unaccepted_force_close(struct sock *sk)
2980 sock_set_flag(sk, SOCK_DEAD);
2981 inet_sk_state_store(sk, TCP_CLOSE);
2982 mptcp_do_fastclose(sk);
2983 __mptcp_destroy_sock(sk);
2986 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2988 /* Concurrent splices from sk_receive_queue into receive_queue will
2989 * always show at least one non-empty queue when checked in this order.
2991 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2992 skb_queue_empty_lockless(&msk->receive_queue))
2995 return EPOLLIN | EPOLLRDNORM;
2998 static void mptcp_check_listen_stop(struct sock *sk)
3002 if (inet_sk_state_load(sk) != TCP_LISTEN)
3005 ssk = mptcp_sk(sk)->first;
3006 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3009 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3010 tcp_set_state(ssk, TCP_CLOSE);
3011 mptcp_subflow_queue_clean(sk, ssk);
3012 inet_csk_listen_stop(ssk);
3016 bool __mptcp_close(struct sock *sk, long timeout)
3018 struct mptcp_subflow_context *subflow;
3019 struct mptcp_sock *msk = mptcp_sk(sk);
3020 bool do_cancel_work = false;
3021 int subflows_alive = 0;
3023 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3025 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3026 mptcp_check_listen_stop(sk);
3027 inet_sk_state_store(sk, TCP_CLOSE);
3031 if (mptcp_check_readable(msk)) {
3032 /* the msk has read data, do the MPTCP equivalent of TCP reset */
3033 inet_sk_state_store(sk, TCP_CLOSE);
3034 mptcp_do_fastclose(sk);
3035 } else if (mptcp_close_state(sk)) {
3036 __mptcp_wr_shutdown(sk);
3039 sk_stream_wait_close(sk, timeout);
3042 /* orphan all the subflows */
3043 mptcp_for_each_subflow(msk, subflow) {
3044 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3045 bool slow = lock_sock_fast_nested(ssk);
3047 subflows_alive += ssk->sk_state != TCP_CLOSE;
3049 /* since the close timeout takes precedence on the fail one,
3052 if (ssk == msk->first)
3053 subflow->fail_tout = 0;
3055 /* detach from the parent socket, but allow data_ready to
3056 * push incoming data into the mptcp stack, to properly ack it
3058 ssk->sk_socket = NULL;
3060 unlock_sock_fast(ssk, slow);
3064 /* all the subflows are closed, only timeout can change the msk
3065 * state, let's not keep resources busy for no reasons
3067 if (subflows_alive == 0)
3068 inet_sk_state_store(sk, TCP_CLOSE);
3071 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3072 if (mptcp_sk(sk)->token)
3073 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3075 if (sk->sk_state == TCP_CLOSE) {
3076 __mptcp_destroy_sock(sk);
3077 do_cancel_work = true;
3079 mptcp_start_tout_timer(sk);
3082 return do_cancel_work;
3085 static void mptcp_close(struct sock *sk, long timeout)
3087 bool do_cancel_work;
3091 do_cancel_work = __mptcp_close(sk, timeout);
3094 mptcp_cancel_work(sk);
3099 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3101 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3102 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3103 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3105 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3106 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3109 msk6->saddr = ssk6->saddr;
3110 msk6->flow_label = ssk6->flow_label;
3114 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3115 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3116 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3117 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3118 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3119 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3122 static int mptcp_disconnect(struct sock *sk, int flags)
3124 struct mptcp_sock *msk = mptcp_sk(sk);
3126 /* We are on the fastopen error path. We can't call straight into the
3127 * subflows cleanup code due to lock nesting (we are already under
3128 * msk->firstsocket lock).
3130 if (msk->fastopening)
3133 mptcp_check_listen_stop(sk);
3134 inet_sk_state_store(sk, TCP_CLOSE);
3136 mptcp_stop_rtx_timer(sk);
3137 mptcp_stop_tout_timer(sk);
3139 if (mptcp_sk(sk)->token)
3140 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
3142 /* msk->subflow is still intact, the following will not free the first
3145 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3146 msk->last_snd = NULL;
3147 WRITE_ONCE(msk->flags, 0);
3149 msk->recovery = false;
3150 msk->can_ack = false;
3151 msk->fully_established = false;
3152 msk->rcv_data_fin = false;
3153 msk->snd_data_fin_enable = false;
3154 msk->rcv_fastclose = false;
3155 msk->use_64bit_ack = false;
3156 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3157 mptcp_pm_data_reset(msk);
3160 WRITE_ONCE(sk->sk_shutdown, 0);
3161 sk_error_report(sk);
3165 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3166 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3168 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3170 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3173 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3175 const struct ipv6_pinfo *np = inet6_sk(sk);
3176 struct ipv6_txoptions *opt;
3177 struct ipv6_pinfo *newnp;
3179 newnp = inet6_sk(newsk);
3182 opt = rcu_dereference(np->opt);
3184 opt = ipv6_dup_options(newsk, opt);
3186 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3188 RCU_INIT_POINTER(newnp->opt, opt);
3193 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3195 struct ip_options_rcu *inet_opt, *newopt = NULL;
3196 const struct inet_sock *inet = inet_sk(sk);
3197 struct inet_sock *newinet;
3199 newinet = inet_sk(newsk);
3202 inet_opt = rcu_dereference(inet->inet_opt);
3204 newopt = sock_kmalloc(newsk, sizeof(*inet_opt) +
3205 inet_opt->opt.optlen, GFP_ATOMIC);
3207 memcpy(newopt, inet_opt, sizeof(*inet_opt) +
3208 inet_opt->opt.optlen);
3210 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3212 RCU_INIT_POINTER(newinet->inet_opt, newopt);
3216 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3217 const struct mptcp_options_received *mp_opt,
3219 struct request_sock *req)
3221 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3222 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3223 struct mptcp_sock *msk;
3229 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3230 if (nsk->sk_family == AF_INET6)
3231 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3234 __mptcp_init_sock(nsk);
3236 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3237 if (nsk->sk_family == AF_INET6)
3238 mptcp_copy_ip6_options(nsk, sk);
3241 mptcp_copy_ip_options(nsk, sk);
3243 msk = mptcp_sk(nsk);
3244 msk->local_key = subflow_req->local_key;
3245 msk->token = subflow_req->token;
3246 WRITE_ONCE(msk->subflow, NULL);
3247 msk->in_accept_queue = 1;
3248 WRITE_ONCE(msk->fully_established, false);
3249 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3250 WRITE_ONCE(msk->csum_enabled, true);
3252 msk->write_seq = subflow_req->idsn + 1;
3253 msk->snd_nxt = msk->write_seq;
3254 msk->snd_una = msk->write_seq;
3255 msk->wnd_end = msk->snd_nxt + tcp_sk(ssk)->snd_wnd;
3256 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3258 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
3259 msk->can_ack = true;
3260 msk->remote_key = mp_opt->sndr_key;
3261 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
3263 WRITE_ONCE(msk->ack_seq, ack_seq);
3264 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3267 sock_reset_flag(nsk, SOCK_RCU_FREE);
3268 security_inet_csk_clone(nsk, req);
3270 /* this can't race with mptcp_close(), as the msk is
3271 * not yet exposted to user-space
3273 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3275 /* The msk maintain a ref to each subflow in the connections list */
3276 WRITE_ONCE(msk->first, ssk);
3277 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3280 /* new mpc subflow takes ownership of the newly
3281 * created mptcp socket
3283 mptcp_token_accept(subflow_req, msk);
3285 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3286 * uses the correct data
3288 mptcp_copy_inaddrs(nsk, ssk);
3289 mptcp_propagate_sndbuf(nsk, ssk);
3291 mptcp_rcv_space_init(msk, ssk);
3292 bh_unlock_sock(nsk);
3294 /* note: the newly allocated socket refcount is 2 now */
3298 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3300 const struct tcp_sock *tp = tcp_sk(ssk);
3302 msk->rcvq_space.copied = 0;
3303 msk->rcvq_space.rtt_us = 0;
3305 msk->rcvq_space.time = tp->tcp_mstamp;
3307 /* initial rcv_space offering made to peer */
3308 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3309 TCP_INIT_CWND * tp->advmss);
3310 if (msk->rcvq_space.space == 0)
3311 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3313 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3316 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3319 struct mptcp_sock *msk = mptcp_sk(sk);
3320 struct socket *listener;
3323 listener = READ_ONCE(msk->subflow);
3324 if (WARN_ON_ONCE(!listener)) {
3329 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3330 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3334 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3335 if (sk_is_mptcp(newsk)) {
3336 struct mptcp_subflow_context *subflow;
3337 struct sock *new_mptcp_sock;
3339 subflow = mptcp_subflow_ctx(newsk);
3340 new_mptcp_sock = subflow->conn;
3342 /* is_mptcp should be false if subflow->conn is missing, see
3343 * subflow_syn_recv_sock()
3345 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3346 tcp_sk(newsk)->is_mptcp = 0;
3350 newsk = new_mptcp_sock;
3351 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3355 newsk->sk_kern_sock = kern;
3359 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3361 struct mptcp_subflow_context *subflow, *tmp;
3362 struct sock *sk = (struct sock *)msk;
3364 __mptcp_clear_xmit(sk);
3366 /* join list will be eventually flushed (with rst) at sock lock release time */
3367 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3368 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3370 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3371 mptcp_data_lock(sk);
3372 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3373 __skb_queue_purge(&sk->sk_receive_queue);
3374 skb_rbtree_purge(&msk->out_of_order_queue);
3375 mptcp_data_unlock(sk);
3377 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3378 * inet_sock_destruct() will dispose it
3380 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3381 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3382 mptcp_token_destroy(msk);
3383 mptcp_pm_free_anno_list(msk);
3384 mptcp_free_local_addr_list(msk);
3387 static void mptcp_destroy(struct sock *sk)
3389 struct mptcp_sock *msk = mptcp_sk(sk);
3391 mptcp_dispose_initial_subflow(msk);
3393 /* allow the following to close even the initial subflow */
3394 msk->free_first = 1;
3395 mptcp_destroy_common(msk, 0);
3396 sk_sockets_allocated_dec(sk);
3399 void __mptcp_data_acked(struct sock *sk)
3401 if (!sock_owned_by_user(sk))
3402 __mptcp_clean_una(sk);
3404 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3406 if (mptcp_pending_data_fin_ack(sk))
3407 mptcp_schedule_work(sk);
3410 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3412 if (!mptcp_send_head(sk))
3415 if (!sock_owned_by_user(sk)) {
3416 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3418 if (xmit_ssk == ssk)
3419 __mptcp_subflow_push_pending(sk, ssk);
3421 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3423 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3427 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3428 BIT(MPTCP_RETRANSMIT) | \
3429 BIT(MPTCP_FLUSH_JOIN_LIST))
3431 /* processes deferred events and flush wmem */
3432 static void mptcp_release_cb(struct sock *sk)
3433 __must_hold(&sk->sk_lock.slock)
3435 struct mptcp_sock *msk = mptcp_sk(sk);
3438 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3439 struct list_head join_list;
3444 INIT_LIST_HEAD(&join_list);
3445 list_splice_init(&msk->join_list, &join_list);
3447 /* the following actions acquire the subflow socket lock
3449 * 1) can't be invoked in atomic scope
3450 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3451 * datapath acquires the msk socket spinlock while helding
3452 * the subflow socket lock
3454 msk->cb_flags &= ~flags;
3455 spin_unlock_bh(&sk->sk_lock.slock);
3457 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3458 __mptcp_flush_join_list(sk, &join_list);
3459 if (flags & BIT(MPTCP_PUSH_PENDING))
3460 __mptcp_push_pending(sk, 0);
3461 if (flags & BIT(MPTCP_RETRANSMIT))
3462 __mptcp_retrans(sk);
3465 spin_lock_bh(&sk->sk_lock.slock);
3468 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3469 __mptcp_clean_una_wakeup(sk);
3470 if (unlikely(msk->cb_flags)) {
3471 /* be sure to set the current sk state before tacking actions
3472 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3474 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3475 __mptcp_set_connected(sk);
3476 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3477 __mptcp_error_report(sk);
3478 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3479 msk->last_snd = NULL;
3482 __mptcp_update_rmem(sk);
3485 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3486 * TCP can't schedule delack timer before the subflow is fully established.
3487 * MPTCP uses the delack timer to do 3rd ack retransmissions
3489 static void schedule_3rdack_retransmission(struct sock *ssk)
3491 struct inet_connection_sock *icsk = inet_csk(ssk);
3492 struct tcp_sock *tp = tcp_sk(ssk);
3493 unsigned long timeout;
3495 if (mptcp_subflow_ctx(ssk)->fully_established)
3498 /* reschedule with a timeout above RTT, as we must look only for drop */
3500 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3502 timeout = TCP_TIMEOUT_INIT;
3505 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3506 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3507 icsk->icsk_ack.timeout = timeout;
3508 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3511 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3513 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3514 struct sock *sk = subflow->conn;
3516 if (status & BIT(MPTCP_DELEGATE_SEND)) {
3517 mptcp_data_lock(sk);
3518 if (!sock_owned_by_user(sk))
3519 __mptcp_subflow_push_pending(sk, ssk);
3521 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3522 mptcp_data_unlock(sk);
3524 if (status & BIT(MPTCP_DELEGATE_ACK))
3525 schedule_3rdack_retransmission(ssk);
3528 static int mptcp_hash(struct sock *sk)
3530 /* should never be called,
3531 * we hash the TCP subflows not the master socket
3537 static void mptcp_unhash(struct sock *sk)
3539 /* called from sk_common_release(), but nothing to do here */
3542 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3544 struct mptcp_sock *msk = mptcp_sk(sk);
3545 struct socket *ssock;
3547 ssock = msk->subflow;
3548 pr_debug("msk=%p, subflow=%p", msk, ssock);
3549 if (WARN_ON_ONCE(!ssock))
3552 return inet_csk_get_port(ssock->sk, snum);
3555 void mptcp_finish_connect(struct sock *ssk)
3557 struct mptcp_subflow_context *subflow;
3558 struct mptcp_sock *msk;
3562 subflow = mptcp_subflow_ctx(ssk);
3566 pr_debug("msk=%p, token=%u", sk, subflow->token);
3568 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3570 subflow->map_seq = ack_seq;
3571 subflow->map_subflow_seq = 1;
3573 /* the socket is not connected yet, no msk/subflow ops can access/race
3574 * accessing the field below
3576 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3577 WRITE_ONCE(msk->local_key, subflow->local_key);
3578 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3579 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3580 WRITE_ONCE(msk->ack_seq, ack_seq);
3581 WRITE_ONCE(msk->can_ack, 1);
3582 WRITE_ONCE(msk->snd_una, msk->write_seq);
3583 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3585 mptcp_pm_new_connection(msk, ssk, 0);
3587 mptcp_rcv_space_init(msk, ssk);
3590 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3592 write_lock_bh(&sk->sk_callback_lock);
3593 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3594 sk_set_socket(sk, parent);
3595 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3596 write_unlock_bh(&sk->sk_callback_lock);
3599 bool mptcp_finish_join(struct sock *ssk)
3601 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3602 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3603 struct sock *parent = (void *)msk;
3606 pr_debug("msk=%p, subflow=%p", msk, subflow);
3608 /* mptcp socket already closing? */
3609 if (!mptcp_is_fully_established(parent)) {
3610 subflow->reset_reason = MPTCP_RST_EMPTCP;
3614 /* active subflow, already present inside the conn_list */
3615 if (!list_empty(&subflow->node)) {
3616 mptcp_subflow_joined(msk, ssk);
3620 if (!mptcp_pm_allow_new_subflow(msk))
3621 goto err_prohibited;
3623 /* If we can't acquire msk socket lock here, let the release callback
3626 mptcp_data_lock(parent);
3627 if (!sock_owned_by_user(parent)) {
3628 ret = __mptcp_finish_join(msk, ssk);
3631 list_add_tail(&subflow->node, &msk->conn_list);
3635 list_add_tail(&subflow->node, &msk->join_list);
3636 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3638 mptcp_data_unlock(parent);
3642 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3649 static void mptcp_shutdown(struct sock *sk, int how)
3651 pr_debug("sk=%p, how=%d", sk, how);
3653 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3654 __mptcp_wr_shutdown(sk);
3657 static int mptcp_forward_alloc_get(const struct sock *sk)
3659 return READ_ONCE(sk->sk_forward_alloc) +
3660 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3663 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3665 const struct sock *sk = (void *)msk;
3668 if (sk->sk_state == TCP_LISTEN)
3671 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3674 delta = msk->write_seq - v;
3675 if (__mptcp_check_fallback(msk) && msk->first) {
3676 struct tcp_sock *tp = tcp_sk(msk->first);
3678 /* the first subflow is disconnected after close - see
3679 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3680 * so ignore that status, too.
3682 if (!((1 << msk->first->sk_state) &
3683 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3684 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3686 if (delta > INT_MAX)
3692 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3694 struct mptcp_sock *msk = mptcp_sk(sk);
3700 if (sk->sk_state == TCP_LISTEN)
3704 __mptcp_move_skbs(msk);
3705 answ = mptcp_inq_hint(sk);
3709 slow = lock_sock_fast(sk);
3710 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3711 unlock_sock_fast(sk, slow);
3714 slow = lock_sock_fast(sk);
3715 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3716 unlock_sock_fast(sk, slow);
3719 return -ENOIOCTLCMD;
3722 return put_user(answ, (int __user *)arg);
3725 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3726 struct mptcp_subflow_context *subflow)
3728 subflow->request_mptcp = 0;
3729 __mptcp_do_fallback(msk);
3732 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3734 struct mptcp_subflow_context *subflow;
3735 struct mptcp_sock *msk = mptcp_sk(sk);
3736 struct socket *ssock;
3739 ssock = __mptcp_nmpc_socket(msk);
3743 mptcp_token_destroy(msk);
3744 inet_sk_state_store(sk, TCP_SYN_SENT);
3745 subflow = mptcp_subflow_ctx(ssock->sk);
3746 #ifdef CONFIG_TCP_MD5SIG
3747 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3750 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3751 mptcp_subflow_early_fallback(msk, subflow);
3753 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3754 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3755 mptcp_subflow_early_fallback(msk, subflow);
3757 if (likely(!__mptcp_check_fallback(msk)))
3758 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3760 /* if reaching here via the fastopen/sendmsg path, the caller already
3761 * acquired the subflow socket lock, too.
3763 if (msk->fastopening)
3764 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3766 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3767 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3769 /* on successful connect, the msk state will be moved to established by
3770 * subflow_finish_connect()
3772 if (unlikely(err && err != -EINPROGRESS)) {
3773 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3777 mptcp_copy_inaddrs(sk, ssock->sk);
3779 /* silence EINPROGRESS and let the caller inet_stream_connect
3780 * handle the connection in progress
3785 static struct proto mptcp_prot = {
3787 .owner = THIS_MODULE,
3788 .init = mptcp_init_sock,
3789 .connect = mptcp_connect,
3790 .disconnect = mptcp_disconnect,
3791 .close = mptcp_close,
3792 .accept = mptcp_accept,
3793 .setsockopt = mptcp_setsockopt,
3794 .getsockopt = mptcp_getsockopt,
3795 .shutdown = mptcp_shutdown,
3796 .destroy = mptcp_destroy,
3797 .sendmsg = mptcp_sendmsg,
3798 .ioctl = mptcp_ioctl,
3799 .recvmsg = mptcp_recvmsg,
3800 .release_cb = mptcp_release_cb,
3802 .unhash = mptcp_unhash,
3803 .get_port = mptcp_get_port,
3804 .forward_alloc_get = mptcp_forward_alloc_get,
3805 .sockets_allocated = &mptcp_sockets_allocated,
3807 .memory_allocated = &tcp_memory_allocated,
3808 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3810 .memory_pressure = &tcp_memory_pressure,
3811 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3812 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3813 .sysctl_mem = sysctl_tcp_mem,
3814 .obj_size = sizeof(struct mptcp_sock),
3815 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3816 .no_autobind = true,
3819 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3821 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3822 struct socket *ssock;
3825 lock_sock(sock->sk);
3826 ssock = __mptcp_nmpc_socket(msk);
3832 err = ssock->ops->bind(ssock, uaddr, addr_len);
3834 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3837 release_sock(sock->sk);
3841 static int mptcp_listen(struct socket *sock, int backlog)
3843 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3844 struct sock *sk = sock->sk;
3845 struct socket *ssock;
3848 pr_debug("msk=%p", msk);
3853 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3856 ssock = __mptcp_nmpc_socket(msk);
3862 mptcp_token_destroy(msk);
3863 inet_sk_state_store(sk, TCP_LISTEN);
3864 sock_set_flag(sk, SOCK_RCU_FREE);
3866 err = ssock->ops->listen(ssock, backlog);
3867 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3869 mptcp_copy_inaddrs(sk, ssock->sk);
3876 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3877 int flags, bool kern)
3879 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3880 struct socket *ssock;
3883 pr_debug("msk=%p", msk);
3885 /* Buggy applications can call accept on socket states other then LISTEN
3886 * but no need to allocate the first subflow just to error out.
3888 ssock = READ_ONCE(msk->subflow);
3892 err = ssock->ops->accept(sock, newsock, flags, kern);
3893 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3894 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3895 struct mptcp_subflow_context *subflow;
3896 struct sock *newsk = newsock->sk;
3898 msk->in_accept_queue = 0;
3902 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3903 * This is needed so NOSPACE flag can be set from tcp stack.
3905 mptcp_for_each_subflow(msk, subflow) {
3906 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3908 if (!ssk->sk_socket)
3909 mptcp_sock_graft(ssk, newsock);
3912 /* Do late cleanup for the first subflow as necessary. Also
3913 * deal with bad peers not doing a complete shutdown.
3916 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3917 __mptcp_close_ssk(newsk, msk->first,
3918 mptcp_subflow_ctx(msk->first), 0);
3919 if (unlikely(list_empty(&msk->conn_list)))
3920 inet_sk_state_store(newsk, TCP_CLOSE);
3923 release_sock(newsk);
3929 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3931 struct sock *sk = (struct sock *)msk;
3933 if (sk_stream_is_writeable(sk))
3934 return EPOLLOUT | EPOLLWRNORM;
3936 mptcp_set_nospace(sk);
3937 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3938 if (sk_stream_is_writeable(sk))
3939 return EPOLLOUT | EPOLLWRNORM;
3944 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3945 struct poll_table_struct *wait)
3947 struct sock *sk = sock->sk;
3948 struct mptcp_sock *msk;
3954 sock_poll_wait(file, sock, wait);
3956 state = inet_sk_state_load(sk);
3957 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3958 if (state == TCP_LISTEN) {
3959 struct socket *ssock = READ_ONCE(msk->subflow);
3961 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3964 return inet_csk_listen_poll(ssock->sk);
3967 shutdown = READ_ONCE(sk->sk_shutdown);
3968 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3970 if (shutdown & RCV_SHUTDOWN)
3971 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3973 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3974 mask |= mptcp_check_readable(msk);
3975 if (shutdown & SEND_SHUTDOWN)
3976 mask |= EPOLLOUT | EPOLLWRNORM;
3978 mask |= mptcp_check_writeable(msk);
3979 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3980 /* cf tcp_poll() note about TFO */
3981 mask |= EPOLLOUT | EPOLLWRNORM;
3984 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3986 if (READ_ONCE(sk->sk_err))
3992 static const struct proto_ops mptcp_stream_ops = {
3994 .owner = THIS_MODULE,
3995 .release = inet_release,
3997 .connect = inet_stream_connect,
3998 .socketpair = sock_no_socketpair,
3999 .accept = mptcp_stream_accept,
4000 .getname = inet_getname,
4002 .ioctl = inet_ioctl,
4003 .gettstamp = sock_gettstamp,
4004 .listen = mptcp_listen,
4005 .shutdown = inet_shutdown,
4006 .setsockopt = sock_common_setsockopt,
4007 .getsockopt = sock_common_getsockopt,
4008 .sendmsg = inet_sendmsg,
4009 .recvmsg = inet_recvmsg,
4010 .mmap = sock_no_mmap,
4011 .sendpage = inet_sendpage,
4014 static struct inet_protosw mptcp_protosw = {
4015 .type = SOCK_STREAM,
4016 .protocol = IPPROTO_MPTCP,
4017 .prot = &mptcp_prot,
4018 .ops = &mptcp_stream_ops,
4019 .flags = INET_PROTOSW_ICSK,
4022 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4024 struct mptcp_delegated_action *delegated;
4025 struct mptcp_subflow_context *subflow;
4028 delegated = container_of(napi, struct mptcp_delegated_action, napi);
4029 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4030 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4032 bh_lock_sock_nested(ssk);
4033 if (!sock_owned_by_user(ssk)) {
4034 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4036 /* tcp_release_cb_override already processed
4037 * the action or will do at next release_sock().
4038 * In both case must dequeue the subflow here - on the same
4039 * CPU that scheduled it.
4042 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4044 bh_unlock_sock(ssk);
4047 if (++work_done == budget)
4051 /* always provide a 0 'work_done' argument, so that napi_complete_done
4052 * will not try accessing the NULL napi->dev ptr
4054 napi_complete_done(napi, 0);
4058 void __init mptcp_proto_init(void)
4060 struct mptcp_delegated_action *delegated;
4063 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4065 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4066 panic("Failed to allocate MPTCP pcpu counter\n");
4068 init_dummy_netdev(&mptcp_napi_dev);
4069 for_each_possible_cpu(cpu) {
4070 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4071 INIT_LIST_HEAD(&delegated->head);
4072 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4074 napi_enable(&delegated->napi);
4077 mptcp_subflow_init();
4081 if (proto_register(&mptcp_prot, 1) != 0)
4082 panic("Failed to register MPTCP proto.\n");
4084 inet_register_protosw(&mptcp_protosw);
4086 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4089 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4090 static const struct proto_ops mptcp_v6_stream_ops = {
4092 .owner = THIS_MODULE,
4093 .release = inet6_release,
4095 .connect = inet_stream_connect,
4096 .socketpair = sock_no_socketpair,
4097 .accept = mptcp_stream_accept,
4098 .getname = inet6_getname,
4100 .ioctl = inet6_ioctl,
4101 .gettstamp = sock_gettstamp,
4102 .listen = mptcp_listen,
4103 .shutdown = inet_shutdown,
4104 .setsockopt = sock_common_setsockopt,
4105 .getsockopt = sock_common_getsockopt,
4106 .sendmsg = inet6_sendmsg,
4107 .recvmsg = inet6_recvmsg,
4108 .mmap = sock_no_mmap,
4109 .sendpage = inet_sendpage,
4110 #ifdef CONFIG_COMPAT
4111 .compat_ioctl = inet6_compat_ioctl,
4115 static struct proto mptcp_v6_prot;
4117 static struct inet_protosw mptcp_v6_protosw = {
4118 .type = SOCK_STREAM,
4119 .protocol = IPPROTO_MPTCP,
4120 .prot = &mptcp_v6_prot,
4121 .ops = &mptcp_v6_stream_ops,
4122 .flags = INET_PROTOSW_ICSK,
4125 int __init mptcp_proto_v6_init(void)
4129 mptcp_v6_prot = mptcp_prot;
4130 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4131 mptcp_v6_prot.slab = NULL;
4132 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4134 err = proto_register(&mptcp_v6_prot, 1);
4138 err = inet6_register_protosw(&mptcp_v6_protosw);
4140 proto_unregister(&mptcp_v6_prot);