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 subflow->local_id_valid = 1;
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 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1279 struct mptcp_data_frag *dfrag,
1280 struct mptcp_sendmsg_info *info)
1282 u64 data_seq = dfrag->data_seq + info->sent;
1283 int offset = dfrag->offset + info->sent;
1284 struct mptcp_sock *msk = mptcp_sk(sk);
1285 bool zero_window_probe = false;
1286 struct mptcp_ext *mpext = NULL;
1287 bool can_coalesce = false;
1288 bool reuse_skb = true;
1289 struct sk_buff *skb;
1293 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1294 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1296 if (WARN_ON_ONCE(info->sent > info->limit ||
1297 info->limit > dfrag->data_len))
1300 if (unlikely(!__tcp_can_send(ssk)))
1303 /* compute send limit */
1304 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1305 copy = info->size_goal;
1307 skb = tcp_write_queue_tail(ssk);
1308 if (skb && copy > skb->len) {
1309 /* Limit the write to the size available in the
1310 * current skb, if any, so that we create at most a new skb.
1311 * Explicitly tells TCP internals to avoid collapsing on later
1312 * queue management operation, to avoid breaking the ext <->
1313 * SSN association set here
1315 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1316 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1317 TCP_SKB_CB(skb)->eor = 1;
1321 i = skb_shinfo(skb)->nr_frags;
1322 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1323 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1324 tcp_mark_push(tcp_sk(ssk), skb);
1331 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1335 i = skb_shinfo(skb)->nr_frags;
1337 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1340 /* Zero window and all data acked? Probe. */
1341 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1343 u64 snd_una = READ_ONCE(msk->snd_una);
1345 if (snd_una != msk->snd_nxt) {
1346 tcp_remove_empty_skb(ssk);
1350 zero_window_probe = true;
1351 data_seq = snd_una - 1;
1354 /* all mptcp-level data is acked, no skbs should be present into the
1357 WARN_ON_ONCE(reuse_skb);
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;
1386 WARN_ON_ONCE(zero_window_probe);
1390 memset(mpext, 0, sizeof(*mpext));
1391 mpext->data_seq = data_seq;
1392 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1393 mpext->data_len = copy;
1397 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1398 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1401 if (zero_window_probe) {
1402 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1404 if (READ_ONCE(msk->csum_enabled))
1405 mptcp_update_data_checksum(skb, copy);
1406 tcp_push_pending_frames(ssk);
1410 if (READ_ONCE(msk->csum_enabled))
1411 mptcp_update_data_checksum(skb, copy);
1412 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1413 mptcp_update_infinite_map(msk, ssk, mpext);
1414 trace_mptcp_sendmsg_frag(mpext);
1415 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1419 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1420 sizeof(struct tcphdr) - \
1421 MAX_TCP_OPTION_SPACE - \
1422 sizeof(struct ipv6hdr) - \
1423 sizeof(struct frag_hdr))
1425 struct subflow_send_info {
1430 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1432 if (!subflow->stale)
1436 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1439 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1441 if (unlikely(subflow->stale)) {
1442 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1444 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1447 mptcp_subflow_set_active(subflow);
1449 return __mptcp_subflow_active(subflow);
1452 #define SSK_MODE_ACTIVE 0
1453 #define SSK_MODE_BACKUP 1
1454 #define SSK_MODE_MAX 2
1456 /* implement the mptcp packet scheduler;
1457 * returns the subflow that will transmit the next DSS
1458 * additionally updates the rtx timeout
1460 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1462 struct subflow_send_info send_info[SSK_MODE_MAX];
1463 struct mptcp_subflow_context *subflow;
1464 struct sock *sk = (struct sock *)msk;
1465 u32 pace, burst, wmem;
1466 int i, nr_active = 0;
1471 sock_owned_by_me(sk);
1473 if (__mptcp_check_fallback(msk)) {
1476 return __tcp_can_send(msk->first) &&
1477 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1480 /* re-use last subflow, if the burst allow that */
1481 if (msk->last_snd && msk->snd_burst > 0 &&
1482 sk_stream_memory_free(msk->last_snd) &&
1483 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1484 mptcp_set_timeout(sk);
1485 return msk->last_snd;
1488 /* pick the subflow with the lower wmem/wspace ratio */
1489 for (i = 0; i < SSK_MODE_MAX; ++i) {
1490 send_info[i].ssk = NULL;
1491 send_info[i].linger_time = -1;
1494 mptcp_for_each_subflow(msk, subflow) {
1495 trace_mptcp_subflow_get_send(subflow);
1496 ssk = mptcp_subflow_tcp_sock(subflow);
1497 if (!mptcp_subflow_active(subflow))
1500 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1501 nr_active += !subflow->backup;
1502 pace = subflow->avg_pacing_rate;
1503 if (unlikely(!pace)) {
1504 /* init pacing rate from socket */
1505 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1506 pace = subflow->avg_pacing_rate;
1511 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1512 if (linger_time < send_info[subflow->backup].linger_time) {
1513 send_info[subflow->backup].ssk = ssk;
1514 send_info[subflow->backup].linger_time = linger_time;
1517 __mptcp_set_timeout(sk, tout);
1519 /* pick the best backup if no other subflow is active */
1521 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1523 /* According to the blest algorithm, to avoid HoL blocking for the
1524 * faster flow, we need to:
1525 * - estimate the faster flow linger time
1526 * - use the above to estimate the amount of byte transferred
1527 * by the faster flow
1528 * - check that the amount of queued data is greter than the above,
1529 * otherwise do not use the picked, slower, subflow
1530 * We select the subflow with the shorter estimated time to flush
1531 * the queued mem, which basically ensure the above. We just need
1532 * to check that subflow has a non empty cwin.
1534 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1535 if (!ssk || !sk_stream_memory_free(ssk))
1538 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1539 wmem = READ_ONCE(ssk->sk_wmem_queued);
1541 msk->last_snd = NULL;
1545 subflow = mptcp_subflow_ctx(ssk);
1546 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1547 READ_ONCE(ssk->sk_pacing_rate) * burst,
1549 msk->last_snd = ssk;
1550 msk->snd_burst = burst;
1554 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1556 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1560 static void mptcp_update_post_push(struct mptcp_sock *msk,
1561 struct mptcp_data_frag *dfrag,
1564 u64 snd_nxt_new = dfrag->data_seq;
1566 dfrag->already_sent += sent;
1568 msk->snd_burst -= sent;
1570 snd_nxt_new += dfrag->already_sent;
1572 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1573 * is recovering after a failover. In that event, this re-sends
1576 * Thus compute snd_nxt_new candidate based on
1577 * the dfrag->data_seq that was sent and the data
1578 * that has been handed to the subflow for transmission
1579 * and skip update in case it was old dfrag.
1581 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1582 msk->snd_nxt = snd_nxt_new;
1585 void mptcp_check_and_set_pending(struct sock *sk)
1587 if (mptcp_send_head(sk))
1588 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1591 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1593 struct sock *prev_ssk = NULL, *ssk = NULL;
1594 struct mptcp_sock *msk = mptcp_sk(sk);
1595 struct mptcp_sendmsg_info info = {
1598 bool do_check_data_fin = false;
1599 struct mptcp_data_frag *dfrag;
1602 while ((dfrag = mptcp_send_head(sk))) {
1603 info.sent = dfrag->already_sent;
1604 info.limit = dfrag->data_len;
1605 len = dfrag->data_len - dfrag->already_sent;
1610 ssk = mptcp_subflow_get_send(msk);
1612 /* First check. If the ssk has changed since
1613 * the last round, release prev_ssk
1615 if (ssk != prev_ssk && prev_ssk)
1616 mptcp_push_release(prev_ssk, &info);
1620 /* Need to lock the new subflow only if different
1621 * from the previous one, otherwise we are still
1622 * helding the relevant lock
1624 if (ssk != prev_ssk)
1627 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1631 mptcp_push_release(ssk, &info);
1635 do_check_data_fin = true;
1639 mptcp_update_post_push(msk, dfrag, ret);
1641 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1644 /* at this point we held the socket lock for the last subflow we used */
1646 mptcp_push_release(ssk, &info);
1649 /* ensure the rtx timer is running */
1650 if (!mptcp_rtx_timer_pending(sk))
1651 mptcp_reset_rtx_timer(sk);
1652 if (do_check_data_fin)
1653 mptcp_check_send_data_fin(sk);
1656 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1658 struct mptcp_sock *msk = mptcp_sk(sk);
1659 struct mptcp_sendmsg_info info = {
1660 .data_lock_held = true,
1662 struct mptcp_data_frag *dfrag;
1663 struct sock *xmit_ssk;
1664 int len, copied = 0;
1668 while ((dfrag = mptcp_send_head(sk))) {
1669 info.sent = dfrag->already_sent;
1670 info.limit = dfrag->data_len;
1671 len = dfrag->data_len - dfrag->already_sent;
1675 /* the caller already invoked the packet scheduler,
1676 * check for a different subflow usage only after
1677 * spooling the first chunk of data
1679 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1682 if (xmit_ssk != ssk) {
1683 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1684 MPTCP_DELEGATE_SEND);
1688 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1697 mptcp_update_post_push(msk, dfrag, ret);
1699 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1703 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1704 * not going to flush it via release_sock()
1707 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1709 if (!mptcp_rtx_timer_pending(sk))
1710 mptcp_reset_rtx_timer(sk);
1712 if (msk->snd_data_fin_enable &&
1713 msk->snd_nxt + 1 == msk->write_seq)
1714 mptcp_schedule_work(sk);
1718 static void mptcp_set_nospace(struct sock *sk)
1720 /* enable autotune */
1721 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1723 /* will be cleared on avail space */
1724 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1727 static int mptcp_disconnect(struct sock *sk, int flags);
1729 static int mptcp_sendmsg_fastopen(struct sock *sk, struct sock *ssk, struct msghdr *msg,
1730 size_t len, int *copied_syn)
1732 unsigned int saved_flags = msg->msg_flags;
1733 struct mptcp_sock *msk = mptcp_sk(sk);
1737 msg->msg_flags |= MSG_DONTWAIT;
1738 msk->fastopening = 1;
1739 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1740 msk->fastopening = 0;
1741 msg->msg_flags = saved_flags;
1744 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1745 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1746 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1747 msg->msg_namelen, msg->msg_flags, 1);
1749 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1750 * case of any error, except timeout or signal
1752 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1754 } else if (ret && ret != -EINPROGRESS) {
1755 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1756 * __inet_stream_connect() can fail, due to looking check,
1757 * see mptcp_disconnect().
1758 * Attempt it again outside the problematic scope.
1760 if (!mptcp_disconnect(sk, 0))
1761 sk->sk_socket->state = SS_UNCONNECTED;
1767 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1769 struct mptcp_sock *msk = mptcp_sk(sk);
1770 struct page_frag *pfrag;
1771 struct socket *ssock;
1776 /* we don't support FASTOPEN yet */
1777 if (msg->msg_flags & MSG_FASTOPEN)
1780 /* silently ignore everything else */
1781 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1785 ssock = __mptcp_nmpc_socket(msk);
1786 if (unlikely(ssock && inet_sk(ssock->sk)->defer_connect)) {
1789 ret = mptcp_sendmsg_fastopen(sk, ssock->sk, msg, len, &copied_syn);
1790 copied += copied_syn;
1791 if (ret == -EINPROGRESS && copied_syn > 0)
1797 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1799 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1800 ret = sk_stream_wait_connect(sk, &timeo);
1806 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1809 pfrag = sk_page_frag(sk);
1811 while (msg_data_left(msg)) {
1812 int total_ts, frag_truesize = 0;
1813 struct mptcp_data_frag *dfrag;
1814 bool dfrag_collapsed;
1815 size_t psize, offset;
1817 /* reuse tail pfrag, if possible, or carve a new one from the
1820 dfrag = mptcp_pending_tail(sk);
1821 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1822 if (!dfrag_collapsed) {
1823 if (!sk_stream_memory_free(sk))
1824 goto wait_for_memory;
1826 if (!mptcp_page_frag_refill(sk, pfrag))
1827 goto wait_for_memory;
1829 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1830 frag_truesize = dfrag->overhead;
1833 /* we do not bound vs wspace, to allow a single packet.
1834 * memory accounting will prevent execessive memory usage
1837 offset = dfrag->offset + dfrag->data_len;
1838 psize = pfrag->size - offset;
1839 psize = min_t(size_t, psize, msg_data_left(msg));
1840 total_ts = psize + frag_truesize;
1842 if (!sk_wmem_schedule(sk, total_ts))
1843 goto wait_for_memory;
1845 if (copy_page_from_iter(dfrag->page, offset, psize,
1846 &msg->msg_iter) != psize) {
1851 /* data successfully copied into the write queue */
1852 sk_forward_alloc_add(sk, -total_ts);
1854 dfrag->data_len += psize;
1855 frag_truesize += psize;
1856 pfrag->offset += frag_truesize;
1857 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1859 /* charge data on mptcp pending queue to the msk socket
1860 * Note: we charge such data both to sk and ssk
1862 sk_wmem_queued_add(sk, frag_truesize);
1863 if (!dfrag_collapsed) {
1864 get_page(dfrag->page);
1865 list_add_tail(&dfrag->list, &msk->rtx_queue);
1866 if (!msk->first_pending)
1867 WRITE_ONCE(msk->first_pending, dfrag);
1869 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1870 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1876 mptcp_set_nospace(sk);
1877 __mptcp_push_pending(sk, msg->msg_flags);
1878 ret = sk_stream_wait_memory(sk, &timeo);
1884 __mptcp_push_pending(sk, msg->msg_flags);
1894 copied = sk_stream_error(sk, msg->msg_flags, ret);
1898 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1900 size_t len, int flags,
1901 struct scm_timestamping_internal *tss,
1904 struct sk_buff *skb, *tmp;
1907 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1908 u32 offset = MPTCP_SKB_CB(skb)->offset;
1909 u32 data_len = skb->len - offset;
1910 u32 count = min_t(size_t, len - copied, data_len);
1913 if (!(flags & MSG_TRUNC)) {
1914 err = skb_copy_datagram_msg(skb, offset, msg, count);
1915 if (unlikely(err < 0)) {
1922 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1923 tcp_update_recv_tstamps(skb, tss);
1924 *cmsg_flags |= MPTCP_CMSG_TS;
1929 if (count < data_len) {
1930 if (!(flags & MSG_PEEK)) {
1931 MPTCP_SKB_CB(skb)->offset += count;
1932 MPTCP_SKB_CB(skb)->map_seq += count;
1937 if (!(flags & MSG_PEEK)) {
1938 /* we will bulk release the skb memory later */
1939 skb->destructor = NULL;
1940 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1941 __skb_unlink(skb, &msk->receive_queue);
1952 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1954 * Only difference: Use highest rtt estimate of the subflows in use.
1956 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1958 struct mptcp_subflow_context *subflow;
1959 struct sock *sk = (struct sock *)msk;
1960 u32 time, advmss = 1;
1963 sock_owned_by_me(sk);
1968 msk->rcvq_space.copied += copied;
1970 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1971 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1973 rtt_us = msk->rcvq_space.rtt_us;
1974 if (rtt_us && time < (rtt_us >> 3))
1978 mptcp_for_each_subflow(msk, subflow) {
1979 const struct tcp_sock *tp;
1983 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1985 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1986 sf_advmss = READ_ONCE(tp->advmss);
1988 rtt_us = max(sf_rtt_us, rtt_us);
1989 advmss = max(sf_advmss, advmss);
1992 msk->rcvq_space.rtt_us = rtt_us;
1993 if (time < (rtt_us >> 3) || rtt_us == 0)
1996 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1999 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2000 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2004 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2006 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2008 do_div(grow, msk->rcvq_space.space);
2009 rcvwin += (grow << 1);
2011 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
2012 while (tcp_win_from_space(sk, rcvmem) < advmss)
2015 do_div(rcvwin, advmss);
2016 rcvbuf = min_t(u64, rcvwin * rcvmem,
2017 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2019 if (rcvbuf > sk->sk_rcvbuf) {
2022 window_clamp = tcp_win_from_space(sk, rcvbuf);
2023 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2025 /* Make subflows follow along. If we do not do this, we
2026 * get drops at subflow level if skbs can't be moved to
2027 * the mptcp rx queue fast enough (announced rcv_win can
2028 * exceed ssk->sk_rcvbuf).
2030 mptcp_for_each_subflow(msk, subflow) {
2034 ssk = mptcp_subflow_tcp_sock(subflow);
2035 slow = lock_sock_fast(ssk);
2036 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2037 tcp_sk(ssk)->window_clamp = window_clamp;
2038 tcp_cleanup_rbuf(ssk, 1);
2039 unlock_sock_fast(ssk, slow);
2044 msk->rcvq_space.space = msk->rcvq_space.copied;
2046 msk->rcvq_space.copied = 0;
2047 msk->rcvq_space.time = mstamp;
2050 static void __mptcp_update_rmem(struct sock *sk)
2052 struct mptcp_sock *msk = mptcp_sk(sk);
2054 if (!msk->rmem_released)
2057 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2058 mptcp_rmem_uncharge(sk, msk->rmem_released);
2059 WRITE_ONCE(msk->rmem_released, 0);
2062 static void __mptcp_splice_receive_queue(struct sock *sk)
2064 struct mptcp_sock *msk = mptcp_sk(sk);
2066 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2069 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2071 struct sock *sk = (struct sock *)msk;
2072 unsigned int moved = 0;
2076 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2079 /* we can have data pending in the subflows only if the msk
2080 * receive buffer was full at subflow_data_ready() time,
2081 * that is an unlikely slow path.
2086 slowpath = lock_sock_fast(ssk);
2087 mptcp_data_lock(sk);
2088 __mptcp_update_rmem(sk);
2089 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2090 mptcp_data_unlock(sk);
2092 if (unlikely(ssk->sk_err))
2093 __mptcp_error_report(sk);
2094 unlock_sock_fast(ssk, slowpath);
2097 /* acquire the data lock only if some input data is pending */
2099 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2100 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2101 mptcp_data_lock(sk);
2102 __mptcp_update_rmem(sk);
2103 ret |= __mptcp_ofo_queue(msk);
2104 __mptcp_splice_receive_queue(sk);
2105 mptcp_data_unlock(sk);
2108 mptcp_check_data_fin((struct sock *)msk);
2109 return !skb_queue_empty(&msk->receive_queue);
2112 static unsigned int mptcp_inq_hint(const struct sock *sk)
2114 const struct mptcp_sock *msk = mptcp_sk(sk);
2115 const struct sk_buff *skb;
2117 skb = skb_peek(&msk->receive_queue);
2119 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2121 if (hint_val >= INT_MAX)
2124 return (unsigned int)hint_val;
2127 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2133 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2134 int flags, int *addr_len)
2136 struct mptcp_sock *msk = mptcp_sk(sk);
2137 struct scm_timestamping_internal tss;
2138 int copied = 0, cmsg_flags = 0;
2142 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2143 if (unlikely(flags & MSG_ERRQUEUE))
2144 return inet_recv_error(sk, msg, len, addr_len);
2147 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2152 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2154 len = min_t(size_t, len, INT_MAX);
2155 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2157 if (unlikely(msk->recvmsg_inq))
2158 cmsg_flags = MPTCP_CMSG_INQ;
2160 while (copied < len) {
2163 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2164 if (unlikely(bytes_read < 0)) {
2166 copied = bytes_read;
2170 copied += bytes_read;
2172 /* be sure to advertise window change */
2173 mptcp_cleanup_rbuf(msk);
2175 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2178 /* only the master socket status is relevant here. The exit
2179 * conditions mirror closely tcp_recvmsg()
2181 if (copied >= target)
2186 sk->sk_state == TCP_CLOSE ||
2187 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2189 signal_pending(current))
2193 copied = sock_error(sk);
2197 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2198 mptcp_check_for_eof(msk);
2200 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2201 /* race breaker: the shutdown could be after the
2202 * previous receive queue check
2204 if (__mptcp_move_skbs(msk))
2209 if (sk->sk_state == TCP_CLOSE) {
2219 if (signal_pending(current)) {
2220 copied = sock_intr_errno(timeo);
2225 pr_debug("block timeout %ld", timeo);
2226 sk_wait_data(sk, &timeo, NULL);
2230 if (cmsg_flags && copied >= 0) {
2231 if (cmsg_flags & MPTCP_CMSG_TS)
2232 tcp_recv_timestamp(msg, sk, &tss);
2234 if (cmsg_flags & MPTCP_CMSG_INQ) {
2235 unsigned int inq = mptcp_inq_hint(sk);
2237 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2241 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2242 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2243 skb_queue_empty(&msk->receive_queue), copied);
2244 if (!(flags & MSG_PEEK))
2245 mptcp_rcv_space_adjust(msk, copied);
2251 static void mptcp_retransmit_timer(struct timer_list *t)
2253 struct inet_connection_sock *icsk = from_timer(icsk, t,
2254 icsk_retransmit_timer);
2255 struct sock *sk = &icsk->icsk_inet.sk;
2256 struct mptcp_sock *msk = mptcp_sk(sk);
2259 if (!sock_owned_by_user(sk)) {
2260 /* we need a process context to retransmit */
2261 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2262 mptcp_schedule_work(sk);
2264 /* delegate our work to tcp_release_cb() */
2265 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2271 static void mptcp_tout_timer(struct timer_list *t)
2273 struct sock *sk = from_timer(sk, t, sk_timer);
2275 mptcp_schedule_work(sk);
2279 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2282 * A backup subflow is returned only if that is the only kind available.
2284 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2286 struct sock *backup = NULL, *pick = NULL;
2287 struct mptcp_subflow_context *subflow;
2288 int min_stale_count = INT_MAX;
2290 sock_owned_by_me((const struct sock *)msk);
2292 if (__mptcp_check_fallback(msk))
2295 mptcp_for_each_subflow(msk, subflow) {
2296 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2298 if (!__mptcp_subflow_active(subflow))
2301 /* still data outstanding at TCP level? skip this */
2302 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2303 mptcp_pm_subflow_chk_stale(msk, ssk);
2304 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2308 if (subflow->backup) {
2321 /* use backup only if there are no progresses anywhere */
2322 return min_stale_count > 1 ? backup : NULL;
2325 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2328 iput(SOCK_INODE(msk->subflow));
2329 WRITE_ONCE(msk->subflow, NULL);
2333 bool __mptcp_retransmit_pending_data(struct sock *sk)
2335 struct mptcp_data_frag *cur, *rtx_head;
2336 struct mptcp_sock *msk = mptcp_sk(sk);
2338 if (__mptcp_check_fallback(mptcp_sk(sk)))
2341 if (tcp_rtx_and_write_queues_empty(sk))
2344 /* the closing socket has some data untransmitted and/or unacked:
2345 * some data in the mptcp rtx queue has not really xmitted yet.
2346 * keep it simple and re-inject the whole mptcp level rtx queue
2348 mptcp_data_lock(sk);
2349 __mptcp_clean_una_wakeup(sk);
2350 rtx_head = mptcp_rtx_head(sk);
2352 mptcp_data_unlock(sk);
2356 msk->recovery_snd_nxt = msk->snd_nxt;
2357 msk->recovery = true;
2358 mptcp_data_unlock(sk);
2360 msk->first_pending = rtx_head;
2363 /* be sure to clear the "sent status" on all re-injected fragments */
2364 list_for_each_entry(cur, &msk->rtx_queue, list) {
2365 if (!cur->already_sent)
2367 cur->already_sent = 0;
2373 /* flags for __mptcp_close_ssk() */
2374 #define MPTCP_CF_PUSH BIT(1)
2375 #define MPTCP_CF_FASTCLOSE BIT(2)
2377 /* subflow sockets can be either outgoing (connect) or incoming
2380 * Outgoing subflows use in-kernel sockets.
2381 * Incoming subflows do not have their own 'struct socket' allocated,
2382 * so we need to use tcp_close() after detaching them from the mptcp
2385 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2386 struct mptcp_subflow_context *subflow,
2389 struct mptcp_sock *msk = mptcp_sk(sk);
2390 bool dispose_it, need_push = false;
2392 /* If the first subflow moved to a close state before accept, e.g. due
2393 * to an incoming reset or listener shutdown, the subflow socket is
2394 * already deleted by inet_child_forget() and the mptcp socket can't
2397 if (msk->in_accept_queue && msk->first == ssk &&
2398 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2399 /* ensure later check in mptcp_worker() will dispose the msk */
2400 mptcp_set_close_tout(sk, tcp_jiffies32 - (TCP_TIMEWAIT_LEN + 1));
2401 sock_set_flag(sk, SOCK_DEAD);
2402 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2403 mptcp_subflow_drop_ctx(ssk);
2407 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2409 list_del(&subflow->node);
2411 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2413 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2414 /* be sure to force the tcp_disconnect() path,
2415 * to generate the egress reset
2417 ssk->sk_lingertime = 0;
2418 sock_set_flag(ssk, SOCK_LINGER);
2419 subflow->send_fastclose = 1;
2422 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2424 /* The MPTCP code never wait on the subflow sockets, TCP-level
2425 * disconnect should never fail
2427 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2428 msk->subflow->state = SS_UNCONNECTED;
2429 mptcp_subflow_ctx_reset(subflow);
2435 subflow->disposable = 1;
2437 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2438 * the ssk has been already destroyed, we just need to release the
2439 * reference owned by msk;
2441 if (!inet_csk(ssk)->icsk_ulp_ops) {
2442 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2443 kfree_rcu(subflow, rcu);
2445 /* otherwise tcp will dispose of the ssk and subflow ctx */
2446 __tcp_close(ssk, 0);
2448 /* close acquired an extra ref */
2453 __mptcp_subflow_error_report(sk, ssk);
2458 if (ssk == msk->first)
2459 WRITE_ONCE(msk->first, NULL);
2462 if (ssk == msk->last_snd)
2463 msk->last_snd = NULL;
2466 __mptcp_push_pending(sk, 0);
2468 /* Catch every 'all subflows closed' scenario, including peers silently
2469 * closing them, e.g. due to timeout.
2470 * For established sockets, allow an additional timeout before closing,
2471 * as the protocol can still create more subflows.
2473 if (list_is_singular(&msk->conn_list) && msk->first &&
2474 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2475 if (sk->sk_state != TCP_ESTABLISHED ||
2476 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2477 inet_sk_state_store(sk, TCP_CLOSE);
2478 mptcp_close_wake_up(sk);
2480 mptcp_start_tout_timer(sk);
2485 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2486 struct mptcp_subflow_context *subflow)
2488 if (sk->sk_state == TCP_ESTABLISHED)
2489 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2491 /* subflow aborted before reaching the fully_established status
2492 * attempt the creation of the next subflow
2494 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2496 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2499 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2504 static void __mptcp_close_subflow(struct sock *sk)
2506 struct mptcp_subflow_context *subflow, *tmp;
2507 struct mptcp_sock *msk = mptcp_sk(sk);
2511 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2512 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2514 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2517 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2518 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2521 mptcp_close_ssk(sk, ssk, subflow);
2526 static bool mptcp_close_tout_expired(const struct sock *sk)
2528 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2529 sk->sk_state == TCP_CLOSE)
2532 return time_after32(tcp_jiffies32,
2533 inet_csk(sk)->icsk_mtup.probe_timestamp + TCP_TIMEWAIT_LEN);
2536 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2538 struct mptcp_subflow_context *subflow, *tmp;
2539 struct sock *sk = &msk->sk.icsk_inet.sk;
2541 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2544 mptcp_token_destroy(msk);
2546 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2547 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2550 slow = lock_sock_fast(tcp_sk);
2551 if (tcp_sk->sk_state != TCP_CLOSE) {
2552 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2553 tcp_set_state(tcp_sk, TCP_CLOSE);
2555 unlock_sock_fast(tcp_sk, slow);
2558 /* Mirror the tcp_reset() error propagation */
2559 switch (sk->sk_state) {
2561 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2563 case TCP_CLOSE_WAIT:
2564 WRITE_ONCE(sk->sk_err, EPIPE);
2569 WRITE_ONCE(sk->sk_err, ECONNRESET);
2572 inet_sk_state_store(sk, TCP_CLOSE);
2573 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2574 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2575 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2577 /* the calling mptcp_worker will properly destroy the socket */
2578 if (sock_flag(sk, SOCK_DEAD))
2581 sk->sk_state_change(sk);
2582 sk_error_report(sk);
2585 static void __mptcp_retrans(struct sock *sk)
2587 struct mptcp_sock *msk = mptcp_sk(sk);
2588 struct mptcp_sendmsg_info info = {};
2589 struct mptcp_data_frag *dfrag;
2594 mptcp_clean_una_wakeup(sk);
2596 /* first check ssk: need to kick "stale" logic */
2597 ssk = mptcp_subflow_get_retrans(msk);
2598 dfrag = mptcp_rtx_head(sk);
2600 if (mptcp_data_fin_enabled(msk)) {
2601 struct inet_connection_sock *icsk = inet_csk(sk);
2603 icsk->icsk_retransmits++;
2604 mptcp_set_datafin_timeout(sk);
2605 mptcp_send_ack(msk);
2610 if (!mptcp_send_head(sk))
2621 /* limit retransmission to the bytes already sent on some subflows */
2623 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2624 while (info.sent < info.limit) {
2625 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2629 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2634 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2635 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2637 WRITE_ONCE(msk->allow_infinite_fallback, false);
2643 mptcp_check_and_set_pending(sk);
2645 if (!mptcp_rtx_timer_pending(sk))
2646 mptcp_reset_rtx_timer(sk);
2649 /* schedule the timeout timer for the relevant event: either close timeout
2650 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2652 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2654 struct sock *sk = (struct sock *)msk;
2655 unsigned long timeout, close_timeout;
2657 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2660 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2663 /* the close timeout takes precedence on the fail one, and here at least one of
2666 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2668 sk_reset_timer(sk, &sk->sk_timer, timeout);
2671 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2673 struct sock *ssk = msk->first;
2679 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2681 slow = lock_sock_fast(ssk);
2682 mptcp_subflow_reset(ssk);
2683 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2684 unlock_sock_fast(ssk, slow);
2687 static void mptcp_do_fastclose(struct sock *sk)
2689 struct mptcp_subflow_context *subflow, *tmp;
2690 struct mptcp_sock *msk = mptcp_sk(sk);
2692 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2693 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2694 subflow, MPTCP_CF_FASTCLOSE);
2697 static void mptcp_worker(struct work_struct *work)
2699 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2700 struct sock *sk = &msk->sk.icsk_inet.sk;
2701 unsigned long fail_tout;
2705 state = sk->sk_state;
2706 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2709 mptcp_check_fastclose(msk);
2711 mptcp_pm_nl_work(msk);
2713 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2714 mptcp_check_for_eof(msk);
2716 mptcp_check_send_data_fin(sk);
2717 mptcp_check_data_fin_ack(sk);
2718 mptcp_check_data_fin(sk);
2720 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2721 __mptcp_close_subflow(sk);
2723 if (mptcp_close_tout_expired(sk)) {
2724 inet_sk_state_store(sk, TCP_CLOSE);
2725 mptcp_do_fastclose(sk);
2726 mptcp_close_wake_up(sk);
2729 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2730 __mptcp_destroy_sock(sk);
2734 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2735 __mptcp_retrans(sk);
2737 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2738 if (fail_tout && time_after(jiffies, fail_tout))
2739 mptcp_mp_fail_no_response(msk);
2746 static int __mptcp_init_sock(struct sock *sk)
2748 struct mptcp_sock *msk = mptcp_sk(sk);
2750 INIT_LIST_HEAD(&msk->conn_list);
2751 INIT_LIST_HEAD(&msk->join_list);
2752 INIT_LIST_HEAD(&msk->rtx_queue);
2753 INIT_WORK(&msk->work, mptcp_worker);
2754 __skb_queue_head_init(&msk->receive_queue);
2755 msk->out_of_order_queue = RB_ROOT;
2756 msk->first_pending = NULL;
2757 msk->rmem_fwd_alloc = 0;
2758 WRITE_ONCE(msk->rmem_released, 0);
2759 msk->timer_ival = TCP_RTO_MIN;
2761 WRITE_ONCE(msk->first, NULL);
2762 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2763 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2764 WRITE_ONCE(msk->allow_infinite_fallback, true);
2765 msk->recovery = false;
2767 mptcp_pm_data_init(msk);
2769 /* re-use the csk retrans timer for MPTCP-level retrans */
2770 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2771 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2776 static void mptcp_ca_reset(struct sock *sk)
2778 struct inet_connection_sock *icsk = inet_csk(sk);
2780 tcp_assign_congestion_control(sk);
2781 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2783 /* no need to keep a reference to the ops, the name will suffice */
2784 tcp_cleanup_congestion_control(sk);
2785 icsk->icsk_ca_ops = NULL;
2788 static int mptcp_init_sock(struct sock *sk)
2790 struct net *net = sock_net(sk);
2793 ret = __mptcp_init_sock(sk);
2797 if (!mptcp_is_enabled(net))
2798 return -ENOPROTOOPT;
2800 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2803 ret = __mptcp_socket_create(mptcp_sk(sk));
2807 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2808 * propagate the correct value
2812 sk_sockets_allocated_inc(sk);
2813 sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2814 sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2819 static void __mptcp_clear_xmit(struct sock *sk)
2821 struct mptcp_sock *msk = mptcp_sk(sk);
2822 struct mptcp_data_frag *dtmp, *dfrag;
2824 WRITE_ONCE(msk->first_pending, NULL);
2825 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2826 dfrag_clear(sk, dfrag);
2829 void mptcp_cancel_work(struct sock *sk)
2831 struct mptcp_sock *msk = mptcp_sk(sk);
2833 if (cancel_work_sync(&msk->work))
2837 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2841 switch (ssk->sk_state) {
2843 if (!(how & RCV_SHUTDOWN))
2847 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2850 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2851 pr_debug("Fallback");
2852 ssk->sk_shutdown |= how;
2853 tcp_shutdown(ssk, how);
2855 /* simulate the data_fin ack reception to let the state
2856 * machine move forward
2858 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2859 mptcp_schedule_work(sk);
2861 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2863 if (!mptcp_rtx_timer_pending(sk))
2864 mptcp_reset_rtx_timer(sk);
2872 static const unsigned char new_state[16] = {
2873 /* current state: new state: action: */
2874 [0 /* (Invalid) */] = TCP_CLOSE,
2875 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2876 [TCP_SYN_SENT] = TCP_CLOSE,
2877 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2878 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2879 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2880 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2881 [TCP_CLOSE] = TCP_CLOSE,
2882 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2883 [TCP_LAST_ACK] = TCP_LAST_ACK,
2884 [TCP_LISTEN] = TCP_CLOSE,
2885 [TCP_CLOSING] = TCP_CLOSING,
2886 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2889 static int mptcp_close_state(struct sock *sk)
2891 int next = (int)new_state[sk->sk_state];
2892 int ns = next & TCP_STATE_MASK;
2894 inet_sk_state_store(sk, ns);
2896 return next & TCP_ACTION_FIN;
2899 static void mptcp_check_send_data_fin(struct sock *sk)
2901 struct mptcp_subflow_context *subflow;
2902 struct mptcp_sock *msk = mptcp_sk(sk);
2904 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2905 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2906 msk->snd_nxt, msk->write_seq);
2908 /* we still need to enqueue subflows or not really shutting down,
2911 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2912 mptcp_send_head(sk))
2915 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2917 mptcp_for_each_subflow(msk, subflow) {
2918 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2920 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2924 static void __mptcp_wr_shutdown(struct sock *sk)
2926 struct mptcp_sock *msk = mptcp_sk(sk);
2928 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2929 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2930 !!mptcp_send_head(sk));
2932 /* will be ignored by fallback sockets */
2933 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2934 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2936 mptcp_check_send_data_fin(sk);
2939 static void __mptcp_destroy_sock(struct sock *sk)
2941 struct mptcp_sock *msk = mptcp_sk(sk);
2943 pr_debug("msk=%p", msk);
2947 mptcp_stop_rtx_timer(sk);
2948 sk_stop_timer(sk, &sk->sk_timer);
2951 sk->sk_prot->destroy(sk);
2953 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2954 WARN_ON_ONCE(msk->rmem_released);
2955 sk_stream_kill_queues(sk);
2956 xfrm_sk_free_policy(sk);
2958 sk_refcnt_debug_release(sk);
2962 void __mptcp_unaccepted_force_close(struct sock *sk)
2964 sock_set_flag(sk, SOCK_DEAD);
2965 inet_sk_state_store(sk, TCP_CLOSE);
2966 mptcp_do_fastclose(sk);
2967 __mptcp_destroy_sock(sk);
2970 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2972 /* Concurrent splices from sk_receive_queue into receive_queue will
2973 * always show at least one non-empty queue when checked in this order.
2975 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2976 skb_queue_empty_lockless(&msk->receive_queue))
2979 return EPOLLIN | EPOLLRDNORM;
2982 static void mptcp_check_listen_stop(struct sock *sk)
2986 if (inet_sk_state_load(sk) != TCP_LISTEN)
2989 ssk = mptcp_sk(sk)->first;
2990 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2993 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2994 tcp_set_state(ssk, TCP_CLOSE);
2995 mptcp_subflow_queue_clean(sk, ssk);
2996 inet_csk_listen_stop(ssk);
3000 bool __mptcp_close(struct sock *sk, long timeout)
3002 struct mptcp_subflow_context *subflow;
3003 struct mptcp_sock *msk = mptcp_sk(sk);
3004 bool do_cancel_work = false;
3005 int subflows_alive = 0;
3007 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3009 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3010 mptcp_check_listen_stop(sk);
3011 inet_sk_state_store(sk, TCP_CLOSE);
3015 if (mptcp_check_readable(msk)) {
3016 /* the msk has read data, do the MPTCP equivalent of TCP reset */
3017 inet_sk_state_store(sk, TCP_CLOSE);
3018 mptcp_do_fastclose(sk);
3019 } else if (mptcp_close_state(sk)) {
3020 __mptcp_wr_shutdown(sk);
3023 sk_stream_wait_close(sk, timeout);
3026 /* orphan all the subflows */
3027 mptcp_for_each_subflow(msk, subflow) {
3028 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3029 bool slow = lock_sock_fast_nested(ssk);
3031 subflows_alive += ssk->sk_state != TCP_CLOSE;
3033 /* since the close timeout takes precedence on the fail one,
3036 if (ssk == msk->first)
3037 subflow->fail_tout = 0;
3039 /* detach from the parent socket, but allow data_ready to
3040 * push incoming data into the mptcp stack, to properly ack it
3042 ssk->sk_socket = NULL;
3044 unlock_sock_fast(ssk, slow);
3048 /* all the subflows are closed, only timeout can change the msk
3049 * state, let's not keep resources busy for no reasons
3051 if (subflows_alive == 0)
3052 inet_sk_state_store(sk, TCP_CLOSE);
3055 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3056 if (mptcp_sk(sk)->token)
3057 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3059 if (sk->sk_state == TCP_CLOSE) {
3060 __mptcp_destroy_sock(sk);
3061 do_cancel_work = true;
3063 mptcp_start_tout_timer(sk);
3066 return do_cancel_work;
3069 static void mptcp_close(struct sock *sk, long timeout)
3071 bool do_cancel_work;
3075 do_cancel_work = __mptcp_close(sk, timeout);
3078 mptcp_cancel_work(sk);
3083 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3085 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3086 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3087 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3089 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3090 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3093 msk6->saddr = ssk6->saddr;
3094 msk6->flow_label = ssk6->flow_label;
3098 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3099 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3100 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3101 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3102 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3103 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3106 static int mptcp_disconnect(struct sock *sk, int flags)
3108 struct mptcp_sock *msk = mptcp_sk(sk);
3110 /* Deny disconnect if other threads are blocked in sk_wait_event()
3111 * or inet_wait_for_connect().
3113 if (sk->sk_wait_pending)
3116 /* We are on the fastopen error path. We can't call straight into the
3117 * subflows cleanup code due to lock nesting (we are already under
3118 * msk->firstsocket lock).
3120 if (msk->fastopening)
3123 mptcp_check_listen_stop(sk);
3124 inet_sk_state_store(sk, TCP_CLOSE);
3126 mptcp_stop_rtx_timer(sk);
3127 mptcp_stop_tout_timer(sk);
3129 if (mptcp_sk(sk)->token)
3130 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
3132 /* msk->subflow is still intact, the following will not free the first
3135 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3136 msk->last_snd = NULL;
3137 WRITE_ONCE(msk->flags, 0);
3139 msk->push_pending = 0;
3140 msk->recovery = false;
3141 msk->can_ack = false;
3142 msk->fully_established = false;
3143 msk->rcv_data_fin = false;
3144 msk->snd_data_fin_enable = false;
3145 msk->rcv_fastclose = false;
3146 msk->use_64bit_ack = false;
3147 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3148 mptcp_pm_data_reset(msk);
3151 WRITE_ONCE(sk->sk_shutdown, 0);
3152 sk_error_report(sk);
3156 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3157 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3159 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3161 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3165 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3166 const struct mptcp_options_received *mp_opt,
3168 struct request_sock *req)
3170 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3171 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3172 struct mptcp_sock *msk;
3178 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3179 if (nsk->sk_family == AF_INET6)
3180 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3183 nsk->sk_wait_pending = 0;
3184 __mptcp_init_sock(nsk);
3186 msk = mptcp_sk(nsk);
3187 msk->local_key = subflow_req->local_key;
3188 msk->token = subflow_req->token;
3189 WRITE_ONCE(msk->subflow, NULL);
3190 msk->in_accept_queue = 1;
3191 WRITE_ONCE(msk->fully_established, false);
3192 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3193 WRITE_ONCE(msk->csum_enabled, true);
3195 msk->write_seq = subflow_req->idsn + 1;
3196 msk->snd_nxt = msk->write_seq;
3197 msk->snd_una = msk->write_seq;
3198 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3199 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3201 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
3202 msk->can_ack = true;
3203 msk->remote_key = mp_opt->sndr_key;
3204 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
3206 WRITE_ONCE(msk->ack_seq, ack_seq);
3207 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3210 sock_reset_flag(nsk, SOCK_RCU_FREE);
3211 security_inet_csk_clone(nsk, req);
3213 /* this can't race with mptcp_close(), as the msk is
3214 * not yet exposted to user-space
3216 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3218 /* The msk maintain a ref to each subflow in the connections list */
3219 WRITE_ONCE(msk->first, ssk);
3220 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3223 /* new mpc subflow takes ownership of the newly
3224 * created mptcp socket
3226 mptcp_token_accept(subflow_req, msk);
3228 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3229 * uses the correct data
3231 mptcp_copy_inaddrs(nsk, ssk);
3232 mptcp_propagate_sndbuf(nsk, ssk);
3234 mptcp_rcv_space_init(msk, ssk);
3235 bh_unlock_sock(nsk);
3237 /* note: the newly allocated socket refcount is 2 now */
3241 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3243 const struct tcp_sock *tp = tcp_sk(ssk);
3245 msk->rcvq_space.copied = 0;
3246 msk->rcvq_space.rtt_us = 0;
3248 msk->rcvq_space.time = tp->tcp_mstamp;
3250 /* initial rcv_space offering made to peer */
3251 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3252 TCP_INIT_CWND * tp->advmss);
3253 if (msk->rcvq_space.space == 0)
3254 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3256 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3259 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3262 struct mptcp_sock *msk = mptcp_sk(sk);
3263 struct socket *listener;
3266 listener = READ_ONCE(msk->subflow);
3267 if (WARN_ON_ONCE(!listener)) {
3272 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3273 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3277 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3278 if (sk_is_mptcp(newsk)) {
3279 struct mptcp_subflow_context *subflow;
3280 struct sock *new_mptcp_sock;
3282 subflow = mptcp_subflow_ctx(newsk);
3283 new_mptcp_sock = subflow->conn;
3285 /* is_mptcp should be false if subflow->conn is missing, see
3286 * subflow_syn_recv_sock()
3288 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3289 tcp_sk(newsk)->is_mptcp = 0;
3293 newsk = new_mptcp_sock;
3294 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3296 MPTCP_INC_STATS(sock_net(sk),
3297 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3301 newsk->sk_kern_sock = kern;
3305 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3307 struct mptcp_subflow_context *subflow, *tmp;
3308 struct sock *sk = (struct sock *)msk;
3310 __mptcp_clear_xmit(sk);
3312 /* join list will be eventually flushed (with rst) at sock lock release time */
3313 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3314 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3316 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3317 mptcp_data_lock(sk);
3318 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3319 __skb_queue_purge(&sk->sk_receive_queue);
3320 skb_rbtree_purge(&msk->out_of_order_queue);
3321 mptcp_data_unlock(sk);
3323 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3324 * inet_sock_destruct() will dispose it
3326 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3327 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3328 mptcp_token_destroy(msk);
3329 mptcp_pm_free_anno_list(msk);
3330 mptcp_free_local_addr_list(msk);
3333 static void mptcp_destroy(struct sock *sk)
3335 struct mptcp_sock *msk = mptcp_sk(sk);
3337 /* clears msk->subflow, allowing the following to close
3338 * even the initial subflow
3340 mptcp_dispose_initial_subflow(msk);
3341 mptcp_destroy_common(msk, 0);
3342 sk_sockets_allocated_dec(sk);
3345 void __mptcp_data_acked(struct sock *sk)
3347 if (!sock_owned_by_user(sk))
3348 __mptcp_clean_una(sk);
3350 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3352 if (mptcp_pending_data_fin_ack(sk))
3353 mptcp_schedule_work(sk);
3356 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3358 if (!mptcp_send_head(sk))
3361 if (!sock_owned_by_user(sk)) {
3362 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3364 if (xmit_ssk == ssk)
3365 __mptcp_subflow_push_pending(sk, ssk);
3367 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3369 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3373 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3374 BIT(MPTCP_RETRANSMIT) | \
3375 BIT(MPTCP_FLUSH_JOIN_LIST))
3377 /* processes deferred events and flush wmem */
3378 static void mptcp_release_cb(struct sock *sk)
3379 __must_hold(&sk->sk_lock.slock)
3381 struct mptcp_sock *msk = mptcp_sk(sk);
3384 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3386 struct list_head join_list;
3391 INIT_LIST_HEAD(&join_list);
3392 list_splice_init(&msk->join_list, &join_list);
3394 /* the following actions acquire the subflow socket lock
3396 * 1) can't be invoked in atomic scope
3397 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3398 * datapath acquires the msk socket spinlock while helding
3399 * the subflow socket lock
3401 msk->push_pending = 0;
3402 msk->cb_flags &= ~flags;
3403 spin_unlock_bh(&sk->sk_lock.slock);
3405 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3406 __mptcp_flush_join_list(sk, &join_list);
3407 if (flags & BIT(MPTCP_PUSH_PENDING))
3408 __mptcp_push_pending(sk, 0);
3409 if (flags & BIT(MPTCP_RETRANSMIT))
3410 __mptcp_retrans(sk);
3413 spin_lock_bh(&sk->sk_lock.slock);
3416 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3417 __mptcp_clean_una_wakeup(sk);
3418 if (unlikely(msk->cb_flags)) {
3419 /* be sure to set the current sk state before tacking actions
3420 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3422 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3423 __mptcp_set_connected(sk);
3424 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3425 __mptcp_error_report(sk);
3426 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3427 msk->last_snd = NULL;
3430 __mptcp_update_rmem(sk);
3433 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3434 * TCP can't schedule delack timer before the subflow is fully established.
3435 * MPTCP uses the delack timer to do 3rd ack retransmissions
3437 static void schedule_3rdack_retransmission(struct sock *ssk)
3439 struct inet_connection_sock *icsk = inet_csk(ssk);
3440 struct tcp_sock *tp = tcp_sk(ssk);
3441 unsigned long timeout;
3443 if (mptcp_subflow_ctx(ssk)->fully_established)
3446 /* reschedule with a timeout above RTT, as we must look only for drop */
3448 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3450 timeout = TCP_TIMEOUT_INIT;
3453 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3454 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3455 icsk->icsk_ack.timeout = timeout;
3456 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3459 void mptcp_subflow_process_delegated(struct sock *ssk)
3461 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3462 struct sock *sk = subflow->conn;
3464 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3465 mptcp_data_lock(sk);
3466 if (!sock_owned_by_user(sk))
3467 __mptcp_subflow_push_pending(sk, ssk);
3469 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3470 mptcp_data_unlock(sk);
3471 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3473 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3474 schedule_3rdack_retransmission(ssk);
3475 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3479 static int mptcp_hash(struct sock *sk)
3481 /* should never be called,
3482 * we hash the TCP subflows not the master socket
3488 static void mptcp_unhash(struct sock *sk)
3490 /* called from sk_common_release(), but nothing to do here */
3493 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3495 struct mptcp_sock *msk = mptcp_sk(sk);
3496 struct socket *ssock;
3498 ssock = msk->subflow;
3499 pr_debug("msk=%p, subflow=%p", msk, ssock);
3500 if (WARN_ON_ONCE(!ssock))
3503 return inet_csk_get_port(ssock->sk, snum);
3506 void mptcp_finish_connect(struct sock *ssk)
3508 struct mptcp_subflow_context *subflow;
3509 struct mptcp_sock *msk;
3513 subflow = mptcp_subflow_ctx(ssk);
3517 pr_debug("msk=%p, token=%u", sk, subflow->token);
3519 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3521 subflow->map_seq = ack_seq;
3522 subflow->map_subflow_seq = 1;
3524 /* the socket is not connected yet, no msk/subflow ops can access/race
3525 * accessing the field below
3527 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3528 WRITE_ONCE(msk->local_key, subflow->local_key);
3529 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3530 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3531 WRITE_ONCE(msk->ack_seq, ack_seq);
3532 WRITE_ONCE(msk->can_ack, 1);
3533 WRITE_ONCE(msk->snd_una, msk->write_seq);
3534 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3536 mptcp_pm_new_connection(msk, ssk, 0);
3538 mptcp_rcv_space_init(msk, ssk);
3541 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3543 write_lock_bh(&sk->sk_callback_lock);
3544 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3545 sk_set_socket(sk, parent);
3546 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3547 write_unlock_bh(&sk->sk_callback_lock);
3550 bool mptcp_finish_join(struct sock *ssk)
3552 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3553 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3554 struct sock *parent = (void *)msk;
3557 pr_debug("msk=%p, subflow=%p", msk, subflow);
3559 /* mptcp socket already closing? */
3560 if (!mptcp_is_fully_established(parent)) {
3561 subflow->reset_reason = MPTCP_RST_EMPTCP;
3565 /* active subflow, already present inside the conn_list */
3566 if (!list_empty(&subflow->node)) {
3567 mptcp_subflow_joined(msk, ssk);
3571 if (!mptcp_pm_allow_new_subflow(msk))
3572 goto err_prohibited;
3574 /* If we can't acquire msk socket lock here, let the release callback
3577 mptcp_data_lock(parent);
3578 if (!sock_owned_by_user(parent)) {
3579 ret = __mptcp_finish_join(msk, ssk);
3582 list_add_tail(&subflow->node, &msk->conn_list);
3586 list_add_tail(&subflow->node, &msk->join_list);
3587 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3589 mptcp_data_unlock(parent);
3593 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3600 static void mptcp_shutdown(struct sock *sk, int how)
3602 pr_debug("sk=%p, how=%d", sk, how);
3604 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3605 __mptcp_wr_shutdown(sk);
3608 static int mptcp_forward_alloc_get(const struct sock *sk)
3610 return READ_ONCE(sk->sk_forward_alloc) +
3611 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3614 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3616 const struct sock *sk = (void *)msk;
3619 if (sk->sk_state == TCP_LISTEN)
3622 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3625 delta = msk->write_seq - v;
3626 if (__mptcp_check_fallback(msk) && msk->first) {
3627 struct tcp_sock *tp = tcp_sk(msk->first);
3629 /* the first subflow is disconnected after close - see
3630 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3631 * so ignore that status, too.
3633 if (!((1 << msk->first->sk_state) &
3634 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3635 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3637 if (delta > INT_MAX)
3643 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3645 struct mptcp_sock *msk = mptcp_sk(sk);
3651 if (sk->sk_state == TCP_LISTEN)
3655 __mptcp_move_skbs(msk);
3656 answ = mptcp_inq_hint(sk);
3660 slow = lock_sock_fast(sk);
3661 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3662 unlock_sock_fast(sk, slow);
3665 slow = lock_sock_fast(sk);
3666 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3667 unlock_sock_fast(sk, slow);
3670 return -ENOIOCTLCMD;
3673 return put_user(answ, (int __user *)arg);
3676 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3677 struct mptcp_subflow_context *subflow)
3679 subflow->request_mptcp = 0;
3680 __mptcp_do_fallback(msk);
3683 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3685 struct mptcp_subflow_context *subflow;
3686 struct mptcp_sock *msk = mptcp_sk(sk);
3687 struct socket *ssock;
3690 ssock = __mptcp_nmpc_socket(msk);
3694 mptcp_token_destroy(msk);
3695 inet_sk_state_store(sk, TCP_SYN_SENT);
3696 subflow = mptcp_subflow_ctx(ssock->sk);
3697 #ifdef CONFIG_TCP_MD5SIG
3698 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3701 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3702 mptcp_subflow_early_fallback(msk, subflow);
3704 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3705 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3706 mptcp_subflow_early_fallback(msk, subflow);
3708 if (likely(!__mptcp_check_fallback(msk)))
3709 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3711 /* if reaching here via the fastopen/sendmsg path, the caller already
3712 * acquired the subflow socket lock, too.
3714 if (msk->fastopening)
3715 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3717 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3718 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3720 /* on successful connect, the msk state will be moved to established by
3721 * subflow_finish_connect()
3723 if (unlikely(err && err != -EINPROGRESS)) {
3724 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3728 mptcp_copy_inaddrs(sk, ssock->sk);
3730 /* silence EINPROGRESS and let the caller inet_stream_connect
3731 * handle the connection in progress
3736 static struct proto mptcp_prot = {
3738 .owner = THIS_MODULE,
3739 .init = mptcp_init_sock,
3740 .connect = mptcp_connect,
3741 .disconnect = mptcp_disconnect,
3742 .close = mptcp_close,
3743 .accept = mptcp_accept,
3744 .setsockopt = mptcp_setsockopt,
3745 .getsockopt = mptcp_getsockopt,
3746 .shutdown = mptcp_shutdown,
3747 .destroy = mptcp_destroy,
3748 .sendmsg = mptcp_sendmsg,
3749 .ioctl = mptcp_ioctl,
3750 .recvmsg = mptcp_recvmsg,
3751 .release_cb = mptcp_release_cb,
3753 .unhash = mptcp_unhash,
3754 .get_port = mptcp_get_port,
3755 .forward_alloc_get = mptcp_forward_alloc_get,
3756 .sockets_allocated = &mptcp_sockets_allocated,
3758 .memory_allocated = &tcp_memory_allocated,
3759 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3761 .memory_pressure = &tcp_memory_pressure,
3762 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3763 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3764 .sysctl_mem = sysctl_tcp_mem,
3765 .obj_size = sizeof(struct mptcp_sock),
3766 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3767 .no_autobind = true,
3770 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3772 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3773 struct socket *ssock;
3776 lock_sock(sock->sk);
3777 ssock = __mptcp_nmpc_socket(msk);
3783 err = ssock->ops->bind(ssock, uaddr, addr_len);
3785 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3788 release_sock(sock->sk);
3792 static int mptcp_listen(struct socket *sock, int backlog)
3794 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3795 struct sock *sk = sock->sk;
3796 struct socket *ssock;
3799 pr_debug("msk=%p", msk);
3804 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3807 ssock = __mptcp_nmpc_socket(msk);
3813 mptcp_token_destroy(msk);
3814 inet_sk_state_store(sk, TCP_LISTEN);
3815 sock_set_flag(sk, SOCK_RCU_FREE);
3817 err = ssock->ops->listen(ssock, backlog);
3818 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3820 mptcp_copy_inaddrs(sk, ssock->sk);
3827 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3828 int flags, bool kern)
3830 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3831 struct socket *ssock;
3834 pr_debug("msk=%p", msk);
3836 /* Buggy applications can call accept on socket states other then LISTEN
3837 * but no need to allocate the first subflow just to error out.
3839 ssock = READ_ONCE(msk->subflow);
3843 err = ssock->ops->accept(sock, newsock, flags, kern);
3844 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3845 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3846 struct mptcp_subflow_context *subflow;
3847 struct sock *newsk = newsock->sk;
3849 msk->in_accept_queue = 0;
3853 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3854 * This is needed so NOSPACE flag can be set from tcp stack.
3856 mptcp_for_each_subflow(msk, subflow) {
3857 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3859 if (!ssk->sk_socket)
3860 mptcp_sock_graft(ssk, newsock);
3863 /* Do late cleanup for the first subflow as necessary. Also
3864 * deal with bad peers not doing a complete shutdown.
3867 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3868 __mptcp_close_ssk(newsk, msk->first,
3869 mptcp_subflow_ctx(msk->first), 0);
3870 if (unlikely(list_empty(&msk->conn_list)))
3871 inet_sk_state_store(newsk, TCP_CLOSE);
3874 release_sock(newsk);
3880 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3882 struct sock *sk = (struct sock *)msk;
3884 if (sk_stream_is_writeable(sk))
3885 return EPOLLOUT | EPOLLWRNORM;
3887 mptcp_set_nospace(sk);
3888 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3889 if (sk_stream_is_writeable(sk))
3890 return EPOLLOUT | EPOLLWRNORM;
3895 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3896 struct poll_table_struct *wait)
3898 struct sock *sk = sock->sk;
3899 struct mptcp_sock *msk;
3905 sock_poll_wait(file, sock, wait);
3907 state = inet_sk_state_load(sk);
3908 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3909 if (state == TCP_LISTEN) {
3910 struct socket *ssock = READ_ONCE(msk->subflow);
3912 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3915 return inet_csk_listen_poll(ssock->sk);
3918 shutdown = READ_ONCE(sk->sk_shutdown);
3919 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3921 if (shutdown & RCV_SHUTDOWN)
3922 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3924 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3925 mask |= mptcp_check_readable(msk);
3926 if (shutdown & SEND_SHUTDOWN)
3927 mask |= EPOLLOUT | EPOLLWRNORM;
3929 mask |= mptcp_check_writeable(msk);
3930 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3931 /* cf tcp_poll() note about TFO */
3932 mask |= EPOLLOUT | EPOLLWRNORM;
3935 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3937 if (READ_ONCE(sk->sk_err))
3943 static const struct proto_ops mptcp_stream_ops = {
3945 .owner = THIS_MODULE,
3946 .release = inet_release,
3948 .connect = inet_stream_connect,
3949 .socketpair = sock_no_socketpair,
3950 .accept = mptcp_stream_accept,
3951 .getname = inet_getname,
3953 .ioctl = inet_ioctl,
3954 .gettstamp = sock_gettstamp,
3955 .listen = mptcp_listen,
3956 .shutdown = inet_shutdown,
3957 .setsockopt = sock_common_setsockopt,
3958 .getsockopt = sock_common_getsockopt,
3959 .sendmsg = inet_sendmsg,
3960 .recvmsg = inet_recvmsg,
3961 .mmap = sock_no_mmap,
3962 .sendpage = inet_sendpage,
3965 static struct inet_protosw mptcp_protosw = {
3966 .type = SOCK_STREAM,
3967 .protocol = IPPROTO_MPTCP,
3968 .prot = &mptcp_prot,
3969 .ops = &mptcp_stream_ops,
3970 .flags = INET_PROTOSW_ICSK,
3973 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3975 struct mptcp_delegated_action *delegated;
3976 struct mptcp_subflow_context *subflow;
3979 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3980 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3981 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3983 bh_lock_sock_nested(ssk);
3984 if (!sock_owned_by_user(ssk) &&
3985 mptcp_subflow_has_delegated_action(subflow))
3986 mptcp_subflow_process_delegated(ssk);
3987 /* ... elsewhere tcp_release_cb_override already processed
3988 * the action or will do at next release_sock().
3989 * In both case must dequeue the subflow here - on the same
3990 * CPU that scheduled it.
3992 bh_unlock_sock(ssk);
3995 if (++work_done == budget)
3999 /* always provide a 0 'work_done' argument, so that napi_complete_done
4000 * will not try accessing the NULL napi->dev ptr
4002 napi_complete_done(napi, 0);
4006 void __init mptcp_proto_init(void)
4008 struct mptcp_delegated_action *delegated;
4011 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4013 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4014 panic("Failed to allocate MPTCP pcpu counter\n");
4016 init_dummy_netdev(&mptcp_napi_dev);
4017 for_each_possible_cpu(cpu) {
4018 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4019 INIT_LIST_HEAD(&delegated->head);
4020 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4022 napi_enable(&delegated->napi);
4025 mptcp_subflow_init();
4029 if (proto_register(&mptcp_prot, 1) != 0)
4030 panic("Failed to register MPTCP proto.\n");
4032 inet_register_protosw(&mptcp_protosw);
4034 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4037 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4038 static const struct proto_ops mptcp_v6_stream_ops = {
4040 .owner = THIS_MODULE,
4041 .release = inet6_release,
4043 .connect = inet_stream_connect,
4044 .socketpair = sock_no_socketpair,
4045 .accept = mptcp_stream_accept,
4046 .getname = inet6_getname,
4048 .ioctl = inet6_ioctl,
4049 .gettstamp = sock_gettstamp,
4050 .listen = mptcp_listen,
4051 .shutdown = inet_shutdown,
4052 .setsockopt = sock_common_setsockopt,
4053 .getsockopt = sock_common_getsockopt,
4054 .sendmsg = inet6_sendmsg,
4055 .recvmsg = inet6_recvmsg,
4056 .mmap = sock_no_mmap,
4057 .sendpage = inet_sendpage,
4058 #ifdef CONFIG_COMPAT
4059 .compat_ioctl = inet6_compat_ioctl,
4063 static struct proto mptcp_v6_prot;
4065 static struct inet_protosw mptcp_v6_protosw = {
4066 .type = SOCK_STREAM,
4067 .protocol = IPPROTO_MPTCP,
4068 .prot = &mptcp_v6_prot,
4069 .ops = &mptcp_v6_stream_ops,
4070 .flags = INET_PROTOSW_ICSK,
4073 int __init mptcp_proto_v6_init(void)
4077 mptcp_v6_prot = mptcp_prot;
4078 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4079 mptcp_v6_prot.slab = NULL;
4080 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4082 err = proto_register(&mptcp_v6_prot, 1);
4086 err = inet6_register_protosw(&mptcp_v6_protosw);
4088 proto_unregister(&mptcp_v6_prot);