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_charge(struct sock *sk, int size)
136 mptcp_sk(sk)->rmem_fwd_alloc -= size;
139 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
140 struct sk_buff *from)
145 if (MPTCP_SKB_CB(from)->offset ||
146 !skb_try_coalesce(to, from, &fragstolen, &delta))
149 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
150 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
151 to->len, MPTCP_SKB_CB(from)->end_seq);
152 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
154 /* note the fwd memory can reach a negative value after accounting
155 * for the delta, but the later skb free will restore a non
158 atomic_add(delta, &sk->sk_rmem_alloc);
159 mptcp_rmem_charge(sk, delta);
160 kfree_skb_partial(from, fragstolen);
165 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
166 struct sk_buff *from)
168 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
171 return mptcp_try_coalesce((struct sock *)msk, to, from);
174 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
176 amount >>= PAGE_SHIFT;
177 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
178 __sk_mem_reduce_allocated(sk, amount);
181 static void mptcp_rmem_uncharge(struct sock *sk, int size)
183 struct mptcp_sock *msk = mptcp_sk(sk);
186 msk->rmem_fwd_alloc += size;
187 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
189 /* see sk_mem_uncharge() for the rationale behind the following schema */
190 if (unlikely(reclaimable >= PAGE_SIZE))
191 __mptcp_rmem_reclaim(sk, reclaimable);
194 static void mptcp_rfree(struct sk_buff *skb)
196 unsigned int len = skb->truesize;
197 struct sock *sk = skb->sk;
199 atomic_sub(len, &sk->sk_rmem_alloc);
200 mptcp_rmem_uncharge(sk, len);
203 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
207 skb->destructor = mptcp_rfree;
208 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
209 mptcp_rmem_charge(sk, skb->truesize);
212 /* "inspired" by tcp_data_queue_ofo(), main differences:
214 * - don't cope with sacks
216 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
218 struct sock *sk = (struct sock *)msk;
219 struct rb_node **p, *parent;
220 u64 seq, end_seq, max_seq;
221 struct sk_buff *skb1;
223 seq = MPTCP_SKB_CB(skb)->map_seq;
224 end_seq = MPTCP_SKB_CB(skb)->end_seq;
225 max_seq = atomic64_read(&msk->rcv_wnd_sent);
227 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
228 RB_EMPTY_ROOT(&msk->out_of_order_queue));
229 if (after64(end_seq, max_seq)) {
232 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
233 (unsigned long long)end_seq - (unsigned long)max_seq,
234 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
235 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
239 p = &msk->out_of_order_queue.rb_node;
240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
241 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
242 rb_link_node(&skb->rbnode, NULL, p);
243 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
244 msk->ooo_last_skb = skb;
248 /* with 2 subflows, adding at end of ooo queue is quite likely
249 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
251 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
252 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
253 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
257 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
258 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
259 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
260 parent = &msk->ooo_last_skb->rbnode;
261 p = &parent->rb_right;
265 /* Find place to insert this segment. Handle overlaps on the way. */
269 skb1 = rb_to_skb(parent);
270 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
271 p = &parent->rb_left;
274 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
275 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
276 /* All the bits are present. Drop. */
278 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
281 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
285 * continue traversing
288 /* skb's seq == skb1's seq and skb covers skb1.
289 * Replace skb1 with skb.
291 rb_replace_node(&skb1->rbnode, &skb->rbnode,
292 &msk->out_of_order_queue);
293 mptcp_drop(sk, skb1);
294 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
297 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
298 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
301 p = &parent->rb_right;
305 /* Insert segment into RB tree. */
306 rb_link_node(&skb->rbnode, parent, p);
307 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
310 /* Remove other segments covered by skb. */
311 while ((skb1 = skb_rb_next(skb)) != NULL) {
312 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
314 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
315 mptcp_drop(sk, skb1);
316 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
318 /* If there is no skb after us, we are the last_skb ! */
320 msk->ooo_last_skb = skb;
324 mptcp_set_owner_r(skb, sk);
327 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
329 struct mptcp_sock *msk = mptcp_sk(sk);
332 if (size <= msk->rmem_fwd_alloc)
335 size -= msk->rmem_fwd_alloc;
336 amt = sk_mem_pages(size);
337 amount = amt << PAGE_SHIFT;
338 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
341 msk->rmem_fwd_alloc += amount;
345 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
346 struct sk_buff *skb, unsigned int offset,
349 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
350 struct sock *sk = (struct sock *)msk;
351 struct sk_buff *tail;
354 __skb_unlink(skb, &ssk->sk_receive_queue);
359 /* try to fetch required memory from subflow */
360 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
363 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
365 /* the skb map_seq accounts for the skb offset:
366 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
369 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
370 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
371 MPTCP_SKB_CB(skb)->offset = offset;
372 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
374 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
376 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
377 tail = skb_peek_tail(&sk->sk_receive_queue);
378 if (tail && mptcp_try_coalesce(sk, tail, skb))
381 mptcp_set_owner_r(skb, sk);
382 __skb_queue_tail(&sk->sk_receive_queue, skb);
384 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
385 mptcp_data_queue_ofo(msk, skb);
389 /* old data, keep it simple and drop the whole pkt, sender
390 * will retransmit as needed, if needed.
392 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
398 static void mptcp_stop_timer(struct sock *sk)
400 struct inet_connection_sock *icsk = inet_csk(sk);
402 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
403 mptcp_sk(sk)->timer_ival = 0;
406 static void mptcp_close_wake_up(struct sock *sk)
408 if (sock_flag(sk, SOCK_DEAD))
411 sk->sk_state_change(sk);
412 if (sk->sk_shutdown == SHUTDOWN_MASK ||
413 sk->sk_state == TCP_CLOSE)
414 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
416 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
419 static bool mptcp_pending_data_fin_ack(struct sock *sk)
421 struct mptcp_sock *msk = mptcp_sk(sk);
423 return ((1 << sk->sk_state) &
424 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
425 msk->write_seq == READ_ONCE(msk->snd_una);
428 static void mptcp_check_data_fin_ack(struct sock *sk)
430 struct mptcp_sock *msk = mptcp_sk(sk);
432 /* Look for an acknowledged DATA_FIN */
433 if (mptcp_pending_data_fin_ack(sk)) {
434 WRITE_ONCE(msk->snd_data_fin_enable, 0);
436 switch (sk->sk_state) {
438 inet_sk_state_store(sk, TCP_FIN_WAIT2);
442 inet_sk_state_store(sk, TCP_CLOSE);
446 mptcp_close_wake_up(sk);
450 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
452 struct mptcp_sock *msk = mptcp_sk(sk);
454 if (READ_ONCE(msk->rcv_data_fin) &&
455 ((1 << sk->sk_state) &
456 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
457 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
459 if (msk->ack_seq == rcv_data_fin_seq) {
461 *seq = rcv_data_fin_seq;
470 static void mptcp_set_datafin_timeout(const struct sock *sk)
472 struct inet_connection_sock *icsk = inet_csk(sk);
475 retransmits = min_t(u32, icsk->icsk_retransmits,
476 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
478 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
481 static void __mptcp_set_timeout(struct sock *sk, long tout)
483 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
486 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
488 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
490 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
491 inet_csk(ssk)->icsk_timeout - jiffies : 0;
494 static void mptcp_set_timeout(struct sock *sk)
496 struct mptcp_subflow_context *subflow;
499 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
500 tout = max(tout, mptcp_timeout_from_subflow(subflow));
501 __mptcp_set_timeout(sk, tout);
504 static inline bool tcp_can_send_ack(const struct sock *ssk)
506 return !((1 << inet_sk_state_load(ssk)) &
507 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
510 void __mptcp_subflow_send_ack(struct sock *ssk)
512 if (tcp_can_send_ack(ssk))
516 static void mptcp_subflow_send_ack(struct sock *ssk)
520 slow = lock_sock_fast(ssk);
521 __mptcp_subflow_send_ack(ssk);
522 unlock_sock_fast(ssk, slow);
525 static void mptcp_send_ack(struct mptcp_sock *msk)
527 struct mptcp_subflow_context *subflow;
529 mptcp_for_each_subflow(msk, subflow)
530 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
533 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
537 slow = lock_sock_fast(ssk);
538 if (tcp_can_send_ack(ssk))
539 tcp_cleanup_rbuf(ssk, 1);
540 unlock_sock_fast(ssk, slow);
543 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
545 const struct inet_connection_sock *icsk = inet_csk(ssk);
546 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
547 const struct tcp_sock *tp = tcp_sk(ssk);
549 return (ack_pending & ICSK_ACK_SCHED) &&
550 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
551 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
552 (rx_empty && ack_pending &
553 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
556 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
558 int old_space = READ_ONCE(msk->old_wspace);
559 struct mptcp_subflow_context *subflow;
560 struct sock *sk = (struct sock *)msk;
561 int space = __mptcp_space(sk);
562 bool cleanup, rx_empty;
564 cleanup = (space > 0) && (space >= (old_space << 1));
565 rx_empty = !__mptcp_rmem(sk);
567 mptcp_for_each_subflow(msk, subflow) {
568 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
570 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
571 mptcp_subflow_cleanup_rbuf(ssk);
575 static bool mptcp_check_data_fin(struct sock *sk)
577 struct mptcp_sock *msk = mptcp_sk(sk);
578 u64 rcv_data_fin_seq;
581 /* Need to ack a DATA_FIN received from a peer while this side
582 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
583 * msk->rcv_data_fin was set when parsing the incoming options
584 * at the subflow level and the msk lock was not held, so this
585 * is the first opportunity to act on the DATA_FIN and change
588 * If we are caught up to the sequence number of the incoming
589 * DATA_FIN, send the DATA_ACK now and do state transition. If
590 * not caught up, do nothing and let the recv code send DATA_ACK
594 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
595 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
596 WRITE_ONCE(msk->rcv_data_fin, 0);
598 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
599 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
601 switch (sk->sk_state) {
602 case TCP_ESTABLISHED:
603 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
606 inet_sk_state_store(sk, TCP_CLOSING);
609 inet_sk_state_store(sk, TCP_CLOSE);
612 /* Other states not expected */
618 if (!__mptcp_check_fallback(msk))
620 mptcp_close_wake_up(sk);
625 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
629 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
630 struct sock *sk = (struct sock *)msk;
631 unsigned int moved = 0;
632 bool more_data_avail;
637 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
639 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
640 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
642 if (unlikely(ssk_rbuf > sk_rbuf)) {
643 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
648 pr_debug("msk=%p ssk=%p", msk, ssk);
651 u32 map_remaining, offset;
652 u32 seq = tp->copied_seq;
656 /* try to move as much data as available */
657 map_remaining = subflow->map_data_len -
658 mptcp_subflow_get_map_offset(subflow);
660 skb = skb_peek(&ssk->sk_receive_queue);
662 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
663 * a different CPU can have already processed the pending
664 * data, stop here or we can enter an infinite loop
671 if (__mptcp_check_fallback(msk)) {
672 /* Under fallback skbs have no MPTCP extension and TCP could
673 * collapse them between the dummy map creation and the
674 * current dequeue. Be sure to adjust the map size.
676 map_remaining = skb->len;
677 subflow->map_data_len = skb->len;
680 offset = seq - TCP_SKB_CB(skb)->seq;
681 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
687 if (offset < skb->len) {
688 size_t len = skb->len - offset;
693 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
697 if (WARN_ON_ONCE(map_remaining < len))
701 sk_eat_skb(ssk, skb);
705 WRITE_ONCE(tp->copied_seq, seq);
706 more_data_avail = mptcp_subflow_data_available(ssk);
708 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
712 } while (more_data_avail);
718 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
720 struct sock *sk = (struct sock *)msk;
721 struct sk_buff *skb, *tail;
726 p = rb_first(&msk->out_of_order_queue);
727 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
730 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
734 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
736 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
739 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
743 end_seq = MPTCP_SKB_CB(skb)->end_seq;
744 tail = skb_peek_tail(&sk->sk_receive_queue);
745 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
746 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
748 /* skip overlapping data, if any */
749 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
750 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
752 MPTCP_SKB_CB(skb)->offset += delta;
753 MPTCP_SKB_CB(skb)->map_seq += delta;
754 __skb_queue_tail(&sk->sk_receive_queue, skb);
756 msk->ack_seq = end_seq;
762 /* In most cases we will be able to lock the mptcp socket. If its already
763 * owned, we need to defer to the work queue to avoid ABBA deadlock.
765 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
767 struct sock *sk = (struct sock *)msk;
768 unsigned int moved = 0;
770 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
771 __mptcp_ofo_queue(msk);
772 if (unlikely(ssk->sk_err)) {
773 if (!sock_owned_by_user(sk))
774 __mptcp_error_report(sk);
776 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
779 /* If the moves have caught up with the DATA_FIN sequence number
780 * it's time to ack the DATA_FIN and change socket state, but
781 * this is not a good place to change state. Let the workqueue
784 if (mptcp_pending_data_fin(sk, NULL))
785 mptcp_schedule_work(sk);
789 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
791 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
792 struct mptcp_sock *msk = mptcp_sk(sk);
793 int sk_rbuf, ssk_rbuf;
795 /* The peer can send data while we are shutting down this
796 * subflow at msk destruction time, but we must avoid enqueuing
797 * more data to the msk receive queue
799 if (unlikely(subflow->disposable))
802 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
803 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
804 if (unlikely(ssk_rbuf > sk_rbuf))
807 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
808 if (__mptcp_rmem(sk) > sk_rbuf) {
809 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
813 /* Wake-up the reader only for in-sequence data */
815 if (move_skbs_to_msk(msk, ssk))
816 sk->sk_data_ready(sk);
818 mptcp_data_unlock(sk);
821 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
823 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
824 WRITE_ONCE(msk->allow_infinite_fallback, false);
825 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
828 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
830 struct sock *sk = (struct sock *)msk;
832 if (sk->sk_state != TCP_ESTABLISHED)
835 /* attach to msk socket only after we are sure we will deal with it
838 if (sk->sk_socket && !ssk->sk_socket)
839 mptcp_sock_graft(ssk, sk->sk_socket);
841 mptcp_sockopt_sync_locked(msk, ssk);
842 mptcp_subflow_joined(msk, ssk);
846 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
848 struct mptcp_subflow_context *tmp, *subflow;
849 struct mptcp_sock *msk = mptcp_sk(sk);
851 list_for_each_entry_safe(subflow, tmp, join_list, node) {
852 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
853 bool slow = lock_sock_fast(ssk);
855 list_move_tail(&subflow->node, &msk->conn_list);
856 if (!__mptcp_finish_join(msk, ssk))
857 mptcp_subflow_reset(ssk);
858 unlock_sock_fast(ssk, slow);
862 static bool mptcp_timer_pending(struct sock *sk)
864 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
867 static void mptcp_reset_timer(struct sock *sk)
869 struct inet_connection_sock *icsk = inet_csk(sk);
872 /* prevent rescheduling on close */
873 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
876 tout = mptcp_sk(sk)->timer_ival;
877 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
880 bool mptcp_schedule_work(struct sock *sk)
882 if (inet_sk_state_load(sk) != TCP_CLOSE &&
883 schedule_work(&mptcp_sk(sk)->work)) {
884 /* each subflow already holds a reference to the sk, and the
885 * workqueue is invoked by a subflow, so sk can't go away here.
893 void mptcp_subflow_eof(struct sock *sk)
895 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
896 mptcp_schedule_work(sk);
899 static void mptcp_check_for_eof(struct mptcp_sock *msk)
901 struct mptcp_subflow_context *subflow;
902 struct sock *sk = (struct sock *)msk;
905 mptcp_for_each_subflow(msk, subflow)
906 receivers += !subflow->rx_eof;
910 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
911 /* hopefully temporary hack: propagate shutdown status
912 * to msk, when all subflows agree on it
914 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
916 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
917 sk->sk_data_ready(sk);
920 switch (sk->sk_state) {
921 case TCP_ESTABLISHED:
922 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
925 inet_sk_state_store(sk, TCP_CLOSING);
928 inet_sk_state_store(sk, TCP_CLOSE);
933 mptcp_close_wake_up(sk);
936 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
938 struct mptcp_subflow_context *subflow;
939 struct sock *sk = (struct sock *)msk;
941 sock_owned_by_me(sk);
943 mptcp_for_each_subflow(msk, subflow) {
944 if (READ_ONCE(subflow->data_avail))
945 return mptcp_subflow_tcp_sock(subflow);
951 static bool mptcp_skb_can_collapse_to(u64 write_seq,
952 const struct sk_buff *skb,
953 const struct mptcp_ext *mpext)
955 if (!tcp_skb_can_collapse_to(skb))
958 /* can collapse only if MPTCP level sequence is in order and this
959 * mapping has not been xmitted yet
961 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
965 /* we can append data to the given data frag if:
966 * - there is space available in the backing page_frag
967 * - the data frag tail matches the current page_frag free offset
968 * - the data frag end sequence number matches the current write seq
970 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
971 const struct page_frag *pfrag,
972 const struct mptcp_data_frag *df)
974 return df && pfrag->page == df->page &&
975 pfrag->size - pfrag->offset > 0 &&
976 pfrag->offset == (df->offset + df->data_len) &&
977 df->data_seq + df->data_len == msk->write_seq;
980 static void dfrag_uncharge(struct sock *sk, int len)
982 sk_mem_uncharge(sk, len);
983 sk_wmem_queued_add(sk, -len);
986 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
988 int len = dfrag->data_len + dfrag->overhead;
990 list_del(&dfrag->list);
991 dfrag_uncharge(sk, len);
992 put_page(dfrag->page);
995 static void __mptcp_clean_una(struct sock *sk)
997 struct mptcp_sock *msk = mptcp_sk(sk);
998 struct mptcp_data_frag *dtmp, *dfrag;
1001 /* on fallback we just need to ignore snd_una, as this is really
1004 if (__mptcp_check_fallback(msk))
1005 msk->snd_una = READ_ONCE(msk->snd_nxt);
1007 snd_una = msk->snd_una;
1008 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1009 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1012 if (unlikely(dfrag == msk->first_pending)) {
1013 /* in recovery mode can see ack after the current snd head */
1014 if (WARN_ON_ONCE(!msk->recovery))
1017 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1020 dfrag_clear(sk, dfrag);
1023 dfrag = mptcp_rtx_head(sk);
1024 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1025 u64 delta = snd_una - dfrag->data_seq;
1027 /* prevent wrap around in recovery mode */
1028 if (unlikely(delta > dfrag->already_sent)) {
1029 if (WARN_ON_ONCE(!msk->recovery))
1031 if (WARN_ON_ONCE(delta > dfrag->data_len))
1033 dfrag->already_sent += delta - dfrag->already_sent;
1036 dfrag->data_seq += delta;
1037 dfrag->offset += delta;
1038 dfrag->data_len -= delta;
1039 dfrag->already_sent -= delta;
1041 dfrag_uncharge(sk, delta);
1044 /* all retransmitted data acked, recovery completed */
1045 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1046 msk->recovery = false;
1049 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1050 snd_una == READ_ONCE(msk->write_seq)) {
1051 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1052 mptcp_stop_timer(sk);
1054 mptcp_reset_timer(sk);
1058 static void __mptcp_clean_una_wakeup(struct sock *sk)
1060 lockdep_assert_held_once(&sk->sk_lock.slock);
1062 __mptcp_clean_una(sk);
1063 mptcp_write_space(sk);
1066 static void mptcp_clean_una_wakeup(struct sock *sk)
1068 mptcp_data_lock(sk);
1069 __mptcp_clean_una_wakeup(sk);
1070 mptcp_data_unlock(sk);
1073 static void mptcp_enter_memory_pressure(struct sock *sk)
1075 struct mptcp_subflow_context *subflow;
1076 struct mptcp_sock *msk = mptcp_sk(sk);
1079 sk_stream_moderate_sndbuf(sk);
1080 mptcp_for_each_subflow(msk, subflow) {
1081 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1084 tcp_enter_memory_pressure(ssk);
1085 sk_stream_moderate_sndbuf(ssk);
1090 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1093 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1095 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1096 pfrag, sk->sk_allocation)))
1099 mptcp_enter_memory_pressure(sk);
1103 static struct mptcp_data_frag *
1104 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1107 int offset = ALIGN(orig_offset, sizeof(long));
1108 struct mptcp_data_frag *dfrag;
1110 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1111 dfrag->data_len = 0;
1112 dfrag->data_seq = msk->write_seq;
1113 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1114 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1115 dfrag->already_sent = 0;
1116 dfrag->page = pfrag->page;
1121 struct mptcp_sendmsg_info {
1127 bool data_lock_held;
1130 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1131 u64 data_seq, int avail_size)
1133 u64 window_end = mptcp_wnd_end(msk);
1136 if (__mptcp_check_fallback(msk))
1139 mptcp_snd_wnd = window_end - data_seq;
1140 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1142 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1143 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1144 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1150 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1152 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1156 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1160 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1162 struct sk_buff *skb;
1164 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1166 if (likely(__mptcp_add_ext(skb, gfp))) {
1167 skb_reserve(skb, MAX_TCP_HEADER);
1168 skb->ip_summed = CHECKSUM_PARTIAL;
1169 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1174 mptcp_enter_memory_pressure(sk);
1179 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1181 struct sk_buff *skb;
1183 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1187 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1188 tcp_skb_entail(ssk, skb);
1191 tcp_skb_tsorted_anchor_cleanup(skb);
1196 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1198 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1200 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1203 /* note: this always recompute the csum on the whole skb, even
1204 * if we just appended a single frag. More status info needed
1206 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1208 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1209 __wsum csum = ~csum_unfold(mpext->csum);
1210 int offset = skb->len - added;
1212 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1215 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1217 struct mptcp_ext *mpext)
1222 mpext->infinite_map = 1;
1223 mpext->data_len = 0;
1225 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1226 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1228 mptcp_do_fallback(ssk);
1231 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1232 struct mptcp_data_frag *dfrag,
1233 struct mptcp_sendmsg_info *info)
1235 u64 data_seq = dfrag->data_seq + info->sent;
1236 int offset = dfrag->offset + info->sent;
1237 struct mptcp_sock *msk = mptcp_sk(sk);
1238 bool zero_window_probe = false;
1239 struct mptcp_ext *mpext = NULL;
1240 bool can_coalesce = false;
1241 bool reuse_skb = true;
1242 struct sk_buff *skb;
1246 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1247 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1249 if (WARN_ON_ONCE(info->sent > info->limit ||
1250 info->limit > dfrag->data_len))
1253 if (unlikely(!__tcp_can_send(ssk)))
1256 /* compute send limit */
1257 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1258 copy = info->size_goal;
1260 skb = tcp_write_queue_tail(ssk);
1261 if (skb && copy > skb->len) {
1262 /* Limit the write to the size available in the
1263 * current skb, if any, so that we create at most a new skb.
1264 * Explicitly tells TCP internals to avoid collapsing on later
1265 * queue management operation, to avoid breaking the ext <->
1266 * SSN association set here
1268 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1269 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1270 TCP_SKB_CB(skb)->eor = 1;
1274 i = skb_shinfo(skb)->nr_frags;
1275 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1276 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1277 tcp_mark_push(tcp_sk(ssk), skb);
1284 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1288 i = skb_shinfo(skb)->nr_frags;
1290 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1293 /* Zero window and all data acked? Probe. */
1294 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1296 u64 snd_una = READ_ONCE(msk->snd_una);
1298 if (snd_una != msk->snd_nxt) {
1299 tcp_remove_empty_skb(ssk);
1303 zero_window_probe = true;
1304 data_seq = snd_una - 1;
1307 /* all mptcp-level data is acked, no skbs should be present into the
1310 WARN_ON_ONCE(reuse_skb);
1313 copy = min_t(size_t, copy, info->limit - info->sent);
1314 if (!sk_wmem_schedule(ssk, copy)) {
1315 tcp_remove_empty_skb(ssk);
1320 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1322 get_page(dfrag->page);
1323 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1327 skb->data_len += copy;
1328 skb->truesize += copy;
1329 sk_wmem_queued_add(ssk, copy);
1330 sk_mem_charge(ssk, copy);
1331 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1332 TCP_SKB_CB(skb)->end_seq += copy;
1333 tcp_skb_pcount_set(skb, 0);
1335 /* on skb reuse we just need to update the DSS len */
1337 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1338 mpext->data_len += copy;
1339 WARN_ON_ONCE(zero_window_probe);
1343 memset(mpext, 0, sizeof(*mpext));
1344 mpext->data_seq = data_seq;
1345 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1346 mpext->data_len = copy;
1350 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1351 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1354 if (zero_window_probe) {
1355 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1357 if (READ_ONCE(msk->csum_enabled))
1358 mptcp_update_data_checksum(skb, copy);
1359 tcp_push_pending_frames(ssk);
1363 if (READ_ONCE(msk->csum_enabled))
1364 mptcp_update_data_checksum(skb, copy);
1365 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1366 mptcp_update_infinite_map(msk, ssk, mpext);
1367 trace_mptcp_sendmsg_frag(mpext);
1368 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1372 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1373 sizeof(struct tcphdr) - \
1374 MAX_TCP_OPTION_SPACE - \
1375 sizeof(struct ipv6hdr) - \
1376 sizeof(struct frag_hdr))
1378 struct subflow_send_info {
1383 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1385 if (!subflow->stale)
1389 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1392 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1394 if (unlikely(subflow->stale)) {
1395 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1397 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1400 mptcp_subflow_set_active(subflow);
1402 return __mptcp_subflow_active(subflow);
1405 #define SSK_MODE_ACTIVE 0
1406 #define SSK_MODE_BACKUP 1
1407 #define SSK_MODE_MAX 2
1409 /* implement the mptcp packet scheduler;
1410 * returns the subflow that will transmit the next DSS
1411 * additionally updates the rtx timeout
1413 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1415 struct subflow_send_info send_info[SSK_MODE_MAX];
1416 struct mptcp_subflow_context *subflow;
1417 struct sock *sk = (struct sock *)msk;
1418 u32 pace, burst, wmem;
1419 int i, nr_active = 0;
1424 sock_owned_by_me(sk);
1426 if (__mptcp_check_fallback(msk)) {
1429 return __tcp_can_send(msk->first) &&
1430 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1433 /* re-use last subflow, if the burst allow that */
1434 if (msk->last_snd && msk->snd_burst > 0 &&
1435 sk_stream_memory_free(msk->last_snd) &&
1436 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1437 mptcp_set_timeout(sk);
1438 return msk->last_snd;
1441 /* pick the subflow with the lower wmem/wspace ratio */
1442 for (i = 0; i < SSK_MODE_MAX; ++i) {
1443 send_info[i].ssk = NULL;
1444 send_info[i].linger_time = -1;
1447 mptcp_for_each_subflow(msk, subflow) {
1448 trace_mptcp_subflow_get_send(subflow);
1449 ssk = mptcp_subflow_tcp_sock(subflow);
1450 if (!mptcp_subflow_active(subflow))
1453 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1454 nr_active += !subflow->backup;
1455 pace = subflow->avg_pacing_rate;
1456 if (unlikely(!pace)) {
1457 /* init pacing rate from socket */
1458 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1459 pace = subflow->avg_pacing_rate;
1464 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1465 if (linger_time < send_info[subflow->backup].linger_time) {
1466 send_info[subflow->backup].ssk = ssk;
1467 send_info[subflow->backup].linger_time = linger_time;
1470 __mptcp_set_timeout(sk, tout);
1472 /* pick the best backup if no other subflow is active */
1474 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1476 /* According to the blest algorithm, to avoid HoL blocking for the
1477 * faster flow, we need to:
1478 * - estimate the faster flow linger time
1479 * - use the above to estimate the amount of byte transferred
1480 * by the faster flow
1481 * - check that the amount of queued data is greter than the above,
1482 * otherwise do not use the picked, slower, subflow
1483 * We select the subflow with the shorter estimated time to flush
1484 * the queued mem, which basically ensure the above. We just need
1485 * to check that subflow has a non empty cwin.
1487 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1488 if (!ssk || !sk_stream_memory_free(ssk))
1491 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1492 wmem = READ_ONCE(ssk->sk_wmem_queued);
1494 msk->last_snd = NULL;
1498 subflow = mptcp_subflow_ctx(ssk);
1499 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1500 READ_ONCE(ssk->sk_pacing_rate) * burst,
1502 msk->last_snd = ssk;
1503 msk->snd_burst = burst;
1507 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1509 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1513 static void mptcp_update_post_push(struct mptcp_sock *msk,
1514 struct mptcp_data_frag *dfrag,
1517 u64 snd_nxt_new = dfrag->data_seq;
1519 dfrag->already_sent += sent;
1521 msk->snd_burst -= sent;
1523 snd_nxt_new += dfrag->already_sent;
1525 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1526 * is recovering after a failover. In that event, this re-sends
1529 * Thus compute snd_nxt_new candidate based on
1530 * the dfrag->data_seq that was sent and the data
1531 * that has been handed to the subflow for transmission
1532 * and skip update in case it was old dfrag.
1534 if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1535 msk->snd_nxt = snd_nxt_new;
1538 void mptcp_check_and_set_pending(struct sock *sk)
1540 if (mptcp_send_head(sk))
1541 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1544 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1546 struct sock *prev_ssk = NULL, *ssk = NULL;
1547 struct mptcp_sock *msk = mptcp_sk(sk);
1548 struct mptcp_sendmsg_info info = {
1551 bool do_check_data_fin = false;
1552 struct mptcp_data_frag *dfrag;
1555 while ((dfrag = mptcp_send_head(sk))) {
1556 info.sent = dfrag->already_sent;
1557 info.limit = dfrag->data_len;
1558 len = dfrag->data_len - dfrag->already_sent;
1563 ssk = mptcp_subflow_get_send(msk);
1565 /* First check. If the ssk has changed since
1566 * the last round, release prev_ssk
1568 if (ssk != prev_ssk && prev_ssk)
1569 mptcp_push_release(prev_ssk, &info);
1573 /* Need to lock the new subflow only if different
1574 * from the previous one, otherwise we are still
1575 * helding the relevant lock
1577 if (ssk != prev_ssk)
1580 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1584 mptcp_push_release(ssk, &info);
1588 do_check_data_fin = true;
1592 mptcp_update_post_push(msk, dfrag, ret);
1594 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1597 /* at this point we held the socket lock for the last subflow we used */
1599 mptcp_push_release(ssk, &info);
1602 /* ensure the rtx timer is running */
1603 if (!mptcp_timer_pending(sk))
1604 mptcp_reset_timer(sk);
1605 if (do_check_data_fin)
1606 mptcp_check_send_data_fin(sk);
1609 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1611 struct mptcp_sock *msk = mptcp_sk(sk);
1612 struct mptcp_sendmsg_info info = {
1613 .data_lock_held = true,
1615 struct mptcp_data_frag *dfrag;
1616 struct sock *xmit_ssk;
1617 int len, copied = 0;
1621 while ((dfrag = mptcp_send_head(sk))) {
1622 info.sent = dfrag->already_sent;
1623 info.limit = dfrag->data_len;
1624 len = dfrag->data_len - dfrag->already_sent;
1628 /* the caller already invoked the packet scheduler,
1629 * check for a different subflow usage only after
1630 * spooling the first chunk of data
1632 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1635 if (xmit_ssk != ssk) {
1636 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1637 MPTCP_DELEGATE_SEND);
1641 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1650 mptcp_update_post_push(msk, dfrag, ret);
1652 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1656 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1657 * not going to flush it via release_sock()
1660 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1662 if (!mptcp_timer_pending(sk))
1663 mptcp_reset_timer(sk);
1665 if (msk->snd_data_fin_enable &&
1666 msk->snd_nxt + 1 == msk->write_seq)
1667 mptcp_schedule_work(sk);
1671 static void mptcp_set_nospace(struct sock *sk)
1673 /* enable autotune */
1674 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1676 /* will be cleared on avail space */
1677 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1680 static int mptcp_disconnect(struct sock *sk, int flags);
1682 static int mptcp_sendmsg_fastopen(struct sock *sk, struct sock *ssk, struct msghdr *msg,
1683 size_t len, int *copied_syn)
1685 unsigned int saved_flags = msg->msg_flags;
1686 struct mptcp_sock *msk = mptcp_sk(sk);
1690 msg->msg_flags |= MSG_DONTWAIT;
1691 msk->fastopening = 1;
1692 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1693 msk->fastopening = 0;
1694 msg->msg_flags = saved_flags;
1697 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1698 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1699 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1700 msg->msg_namelen, msg->msg_flags, 1);
1702 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1703 * case of any error, except timeout or signal
1705 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1707 } else if (ret && ret != -EINPROGRESS) {
1708 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1709 * __inet_stream_connect() can fail, due to looking check,
1710 * see mptcp_disconnect().
1711 * Attempt it again outside the problematic scope.
1713 if (!mptcp_disconnect(sk, 0))
1714 sk->sk_socket->state = SS_UNCONNECTED;
1720 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1722 struct mptcp_sock *msk = mptcp_sk(sk);
1723 struct page_frag *pfrag;
1724 struct socket *ssock;
1729 /* we don't support FASTOPEN yet */
1730 if (msg->msg_flags & MSG_FASTOPEN)
1733 /* silently ignore everything else */
1734 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1738 ssock = __mptcp_nmpc_socket(msk);
1739 if (unlikely(ssock && inet_sk(ssock->sk)->defer_connect)) {
1742 ret = mptcp_sendmsg_fastopen(sk, ssock->sk, msg, len, &copied_syn);
1743 copied += copied_syn;
1744 if (ret == -EINPROGRESS && copied_syn > 0)
1750 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1752 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1753 ret = sk_stream_wait_connect(sk, &timeo);
1759 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1762 pfrag = sk_page_frag(sk);
1764 while (msg_data_left(msg)) {
1765 int total_ts, frag_truesize = 0;
1766 struct mptcp_data_frag *dfrag;
1767 bool dfrag_collapsed;
1768 size_t psize, offset;
1770 /* reuse tail pfrag, if possible, or carve a new one from the
1773 dfrag = mptcp_pending_tail(sk);
1774 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1775 if (!dfrag_collapsed) {
1776 if (!sk_stream_memory_free(sk))
1777 goto wait_for_memory;
1779 if (!mptcp_page_frag_refill(sk, pfrag))
1780 goto wait_for_memory;
1782 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1783 frag_truesize = dfrag->overhead;
1786 /* we do not bound vs wspace, to allow a single packet.
1787 * memory accounting will prevent execessive memory usage
1790 offset = dfrag->offset + dfrag->data_len;
1791 psize = pfrag->size - offset;
1792 psize = min_t(size_t, psize, msg_data_left(msg));
1793 total_ts = psize + frag_truesize;
1795 if (!sk_wmem_schedule(sk, total_ts))
1796 goto wait_for_memory;
1798 if (copy_page_from_iter(dfrag->page, offset, psize,
1799 &msg->msg_iter) != psize) {
1804 /* data successfully copied into the write queue */
1805 sk->sk_forward_alloc -= total_ts;
1807 dfrag->data_len += psize;
1808 frag_truesize += psize;
1809 pfrag->offset += frag_truesize;
1810 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1812 /* charge data on mptcp pending queue to the msk socket
1813 * Note: we charge such data both to sk and ssk
1815 sk_wmem_queued_add(sk, frag_truesize);
1816 if (!dfrag_collapsed) {
1817 get_page(dfrag->page);
1818 list_add_tail(&dfrag->list, &msk->rtx_queue);
1819 if (!msk->first_pending)
1820 WRITE_ONCE(msk->first_pending, dfrag);
1822 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1823 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1829 mptcp_set_nospace(sk);
1830 __mptcp_push_pending(sk, msg->msg_flags);
1831 ret = sk_stream_wait_memory(sk, &timeo);
1837 __mptcp_push_pending(sk, msg->msg_flags);
1847 copied = sk_stream_error(sk, msg->msg_flags, ret);
1851 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1853 size_t len, int flags,
1854 struct scm_timestamping_internal *tss,
1857 struct sk_buff *skb, *tmp;
1860 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1861 u32 offset = MPTCP_SKB_CB(skb)->offset;
1862 u32 data_len = skb->len - offset;
1863 u32 count = min_t(size_t, len - copied, data_len);
1866 if (!(flags & MSG_TRUNC)) {
1867 err = skb_copy_datagram_msg(skb, offset, msg, count);
1868 if (unlikely(err < 0)) {
1875 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1876 tcp_update_recv_tstamps(skb, tss);
1877 *cmsg_flags |= MPTCP_CMSG_TS;
1882 if (count < data_len) {
1883 if (!(flags & MSG_PEEK)) {
1884 MPTCP_SKB_CB(skb)->offset += count;
1885 MPTCP_SKB_CB(skb)->map_seq += count;
1890 if (!(flags & MSG_PEEK)) {
1891 /* we will bulk release the skb memory later */
1892 skb->destructor = NULL;
1893 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1894 __skb_unlink(skb, &msk->receive_queue);
1905 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1907 * Only difference: Use highest rtt estimate of the subflows in use.
1909 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1911 struct mptcp_subflow_context *subflow;
1912 struct sock *sk = (struct sock *)msk;
1913 u32 time, advmss = 1;
1916 sock_owned_by_me(sk);
1921 msk->rcvq_space.copied += copied;
1923 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1924 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1926 rtt_us = msk->rcvq_space.rtt_us;
1927 if (rtt_us && time < (rtt_us >> 3))
1931 mptcp_for_each_subflow(msk, subflow) {
1932 const struct tcp_sock *tp;
1936 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1938 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1939 sf_advmss = READ_ONCE(tp->advmss);
1941 rtt_us = max(sf_rtt_us, rtt_us);
1942 advmss = max(sf_advmss, advmss);
1945 msk->rcvq_space.rtt_us = rtt_us;
1946 if (time < (rtt_us >> 3) || rtt_us == 0)
1949 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1952 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1953 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1957 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1959 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1961 do_div(grow, msk->rcvq_space.space);
1962 rcvwin += (grow << 1);
1964 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1965 while (tcp_win_from_space(sk, rcvmem) < advmss)
1968 do_div(rcvwin, advmss);
1969 rcvbuf = min_t(u64, rcvwin * rcvmem,
1970 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1972 if (rcvbuf > sk->sk_rcvbuf) {
1975 window_clamp = tcp_win_from_space(sk, rcvbuf);
1976 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1978 /* Make subflows follow along. If we do not do this, we
1979 * get drops at subflow level if skbs can't be moved to
1980 * the mptcp rx queue fast enough (announced rcv_win can
1981 * exceed ssk->sk_rcvbuf).
1983 mptcp_for_each_subflow(msk, subflow) {
1987 ssk = mptcp_subflow_tcp_sock(subflow);
1988 slow = lock_sock_fast(ssk);
1989 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1990 tcp_sk(ssk)->window_clamp = window_clamp;
1991 tcp_cleanup_rbuf(ssk, 1);
1992 unlock_sock_fast(ssk, slow);
1997 msk->rcvq_space.space = msk->rcvq_space.copied;
1999 msk->rcvq_space.copied = 0;
2000 msk->rcvq_space.time = mstamp;
2003 static void __mptcp_update_rmem(struct sock *sk)
2005 struct mptcp_sock *msk = mptcp_sk(sk);
2007 if (!msk->rmem_released)
2010 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2011 mptcp_rmem_uncharge(sk, msk->rmem_released);
2012 WRITE_ONCE(msk->rmem_released, 0);
2015 static void __mptcp_splice_receive_queue(struct sock *sk)
2017 struct mptcp_sock *msk = mptcp_sk(sk);
2019 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2022 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2024 struct sock *sk = (struct sock *)msk;
2025 unsigned int moved = 0;
2029 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2032 /* we can have data pending in the subflows only if the msk
2033 * receive buffer was full at subflow_data_ready() time,
2034 * that is an unlikely slow path.
2039 slowpath = lock_sock_fast(ssk);
2040 mptcp_data_lock(sk);
2041 __mptcp_update_rmem(sk);
2042 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2043 mptcp_data_unlock(sk);
2045 if (unlikely(ssk->sk_err))
2046 __mptcp_error_report(sk);
2047 unlock_sock_fast(ssk, slowpath);
2050 /* acquire the data lock only if some input data is pending */
2052 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2053 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2054 mptcp_data_lock(sk);
2055 __mptcp_update_rmem(sk);
2056 ret |= __mptcp_ofo_queue(msk);
2057 __mptcp_splice_receive_queue(sk);
2058 mptcp_data_unlock(sk);
2061 mptcp_check_data_fin((struct sock *)msk);
2062 return !skb_queue_empty(&msk->receive_queue);
2065 static unsigned int mptcp_inq_hint(const struct sock *sk)
2067 const struct mptcp_sock *msk = mptcp_sk(sk);
2068 const struct sk_buff *skb;
2070 skb = skb_peek(&msk->receive_queue);
2072 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2074 if (hint_val >= INT_MAX)
2077 return (unsigned int)hint_val;
2080 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2086 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2087 int flags, int *addr_len)
2089 struct mptcp_sock *msk = mptcp_sk(sk);
2090 struct scm_timestamping_internal tss;
2091 int copied = 0, cmsg_flags = 0;
2095 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2096 if (unlikely(flags & MSG_ERRQUEUE))
2097 return inet_recv_error(sk, msg, len, addr_len);
2100 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2105 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2107 len = min_t(size_t, len, INT_MAX);
2108 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2110 if (unlikely(msk->recvmsg_inq))
2111 cmsg_flags = MPTCP_CMSG_INQ;
2113 while (copied < len) {
2116 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2117 if (unlikely(bytes_read < 0)) {
2119 copied = bytes_read;
2123 copied += bytes_read;
2125 /* be sure to advertise window change */
2126 mptcp_cleanup_rbuf(msk);
2128 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2131 /* only the master socket status is relevant here. The exit
2132 * conditions mirror closely tcp_recvmsg()
2134 if (copied >= target)
2139 sk->sk_state == TCP_CLOSE ||
2140 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2142 signal_pending(current))
2146 copied = sock_error(sk);
2150 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2151 mptcp_check_for_eof(msk);
2153 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2154 /* race breaker: the shutdown could be after the
2155 * previous receive queue check
2157 if (__mptcp_move_skbs(msk))
2162 if (sk->sk_state == TCP_CLOSE) {
2172 if (signal_pending(current)) {
2173 copied = sock_intr_errno(timeo);
2178 pr_debug("block timeout %ld", timeo);
2179 sk_wait_data(sk, &timeo, NULL);
2183 if (cmsg_flags && copied >= 0) {
2184 if (cmsg_flags & MPTCP_CMSG_TS)
2185 tcp_recv_timestamp(msg, sk, &tss);
2187 if (cmsg_flags & MPTCP_CMSG_INQ) {
2188 unsigned int inq = mptcp_inq_hint(sk);
2190 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2194 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2195 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2196 skb_queue_empty(&msk->receive_queue), copied);
2197 if (!(flags & MSG_PEEK))
2198 mptcp_rcv_space_adjust(msk, copied);
2204 static void mptcp_retransmit_timer(struct timer_list *t)
2206 struct inet_connection_sock *icsk = from_timer(icsk, t,
2207 icsk_retransmit_timer);
2208 struct sock *sk = &icsk->icsk_inet.sk;
2209 struct mptcp_sock *msk = mptcp_sk(sk);
2212 if (!sock_owned_by_user(sk)) {
2213 /* we need a process context to retransmit */
2214 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2215 mptcp_schedule_work(sk);
2217 /* delegate our work to tcp_release_cb() */
2218 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2224 static void mptcp_timeout_timer(struct timer_list *t)
2226 struct sock *sk = from_timer(sk, t, sk_timer);
2228 mptcp_schedule_work(sk);
2232 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2235 * A backup subflow is returned only if that is the only kind available.
2237 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2239 struct sock *backup = NULL, *pick = NULL;
2240 struct mptcp_subflow_context *subflow;
2241 int min_stale_count = INT_MAX;
2243 sock_owned_by_me((const struct sock *)msk);
2245 if (__mptcp_check_fallback(msk))
2248 mptcp_for_each_subflow(msk, subflow) {
2249 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2251 if (!__mptcp_subflow_active(subflow))
2254 /* still data outstanding at TCP level? skip this */
2255 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2256 mptcp_pm_subflow_chk_stale(msk, ssk);
2257 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2261 if (subflow->backup) {
2274 /* use backup only if there are no progresses anywhere */
2275 return min_stale_count > 1 ? backup : NULL;
2278 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2281 iput(SOCK_INODE(msk->subflow));
2282 WRITE_ONCE(msk->subflow, NULL);
2286 bool __mptcp_retransmit_pending_data(struct sock *sk)
2288 struct mptcp_data_frag *cur, *rtx_head;
2289 struct mptcp_sock *msk = mptcp_sk(sk);
2291 if (__mptcp_check_fallback(mptcp_sk(sk)))
2294 if (tcp_rtx_and_write_queues_empty(sk))
2297 /* the closing socket has some data untransmitted and/or unacked:
2298 * some data in the mptcp rtx queue has not really xmitted yet.
2299 * keep it simple and re-inject the whole mptcp level rtx queue
2301 mptcp_data_lock(sk);
2302 __mptcp_clean_una_wakeup(sk);
2303 rtx_head = mptcp_rtx_head(sk);
2305 mptcp_data_unlock(sk);
2309 msk->recovery_snd_nxt = msk->snd_nxt;
2310 msk->recovery = true;
2311 mptcp_data_unlock(sk);
2313 msk->first_pending = rtx_head;
2316 /* be sure to clear the "sent status" on all re-injected fragments */
2317 list_for_each_entry(cur, &msk->rtx_queue, list) {
2318 if (!cur->already_sent)
2320 cur->already_sent = 0;
2326 /* flags for __mptcp_close_ssk() */
2327 #define MPTCP_CF_PUSH BIT(1)
2328 #define MPTCP_CF_FASTCLOSE BIT(2)
2330 /* subflow sockets can be either outgoing (connect) or incoming
2333 * Outgoing subflows use in-kernel sockets.
2334 * Incoming subflows do not have their own 'struct socket' allocated,
2335 * so we need to use tcp_close() after detaching them from the mptcp
2338 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2339 struct mptcp_subflow_context *subflow,
2342 struct mptcp_sock *msk = mptcp_sk(sk);
2343 bool dispose_it, need_push = false;
2345 /* If the first subflow moved to a close state before accept, e.g. due
2346 * to an incoming reset, mptcp either:
2347 * - if either the subflow or the msk are dead, destroy the context
2348 * (the subflow socket is deleted by inet_child_forget) and the msk
2349 * - otherwise do nothing at the moment and take action at accept and/or
2350 * listener shutdown - user-space must be able to accept() the closed
2353 if (msk->in_accept_queue && msk->first == ssk) {
2354 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2357 /* ensure later check in mptcp_worker() will dispose the msk */
2358 sock_set_flag(sk, SOCK_DEAD);
2359 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2360 mptcp_subflow_drop_ctx(ssk);
2364 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2366 list_del(&subflow->node);
2368 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2370 if (flags & MPTCP_CF_FASTCLOSE) {
2371 /* be sure to force the tcp_disconnect() path,
2372 * to generate the egress reset
2374 ssk->sk_lingertime = 0;
2375 sock_set_flag(ssk, SOCK_LINGER);
2376 subflow->send_fastclose = 1;
2379 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2381 /* The MPTCP code never wait on the subflow sockets, TCP-level
2382 * disconnect should never fail
2384 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2385 msk->subflow->state = SS_UNCONNECTED;
2386 mptcp_subflow_ctx_reset(subflow);
2392 subflow->disposable = 1;
2394 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2395 * the ssk has been already destroyed, we just need to release the
2396 * reference owned by msk;
2398 if (!inet_csk(ssk)->icsk_ulp_ops) {
2399 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2400 kfree_rcu(subflow, rcu);
2402 /* otherwise tcp will dispose of the ssk and subflow ctx */
2403 if (ssk->sk_state == TCP_LISTEN) {
2404 tcp_set_state(ssk, TCP_CLOSE);
2405 mptcp_subflow_queue_clean(sk, ssk);
2406 inet_csk_listen_stop(ssk);
2409 __tcp_close(ssk, 0);
2411 /* close acquired an extra ref */
2420 if (ssk == msk->first)
2421 WRITE_ONCE(msk->first, NULL);
2424 if (ssk == msk->last_snd)
2425 msk->last_snd = NULL;
2428 __mptcp_push_pending(sk, 0);
2431 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2432 struct mptcp_subflow_context *subflow)
2434 if (sk->sk_state == TCP_ESTABLISHED)
2435 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2437 /* subflow aborted before reaching the fully_established status
2438 * attempt the creation of the next subflow
2440 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2442 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2445 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2450 static void __mptcp_close_subflow(struct sock *sk)
2452 struct mptcp_subflow_context *subflow, *tmp;
2453 struct mptcp_sock *msk = mptcp_sk(sk);
2457 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2458 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2460 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2463 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2464 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2467 mptcp_close_ssk(sk, ssk, subflow);
2472 static bool mptcp_should_close(const struct sock *sk)
2474 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2475 struct mptcp_subflow_context *subflow;
2477 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2480 /* if all subflows are in closed status don't bother with additional
2483 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2484 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2491 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2493 struct mptcp_subflow_context *subflow, *tmp;
2494 struct sock *sk = &msk->sk.icsk_inet.sk;
2496 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2499 mptcp_token_destroy(msk);
2501 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2502 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2505 slow = lock_sock_fast(tcp_sk);
2506 if (tcp_sk->sk_state != TCP_CLOSE) {
2507 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2508 tcp_set_state(tcp_sk, TCP_CLOSE);
2510 unlock_sock_fast(tcp_sk, slow);
2513 /* Mirror the tcp_reset() error propagation */
2514 switch (sk->sk_state) {
2516 sk->sk_err = ECONNREFUSED;
2518 case TCP_CLOSE_WAIT:
2524 sk->sk_err = ECONNRESET;
2527 inet_sk_state_store(sk, TCP_CLOSE);
2528 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2529 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2530 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2532 /* the calling mptcp_worker will properly destroy the socket */
2533 if (sock_flag(sk, SOCK_DEAD))
2536 sk->sk_state_change(sk);
2537 sk_error_report(sk);
2540 static void __mptcp_retrans(struct sock *sk)
2542 struct mptcp_sock *msk = mptcp_sk(sk);
2543 struct mptcp_sendmsg_info info = {};
2544 struct mptcp_data_frag *dfrag;
2549 mptcp_clean_una_wakeup(sk);
2551 /* first check ssk: need to kick "stale" logic */
2552 ssk = mptcp_subflow_get_retrans(msk);
2553 dfrag = mptcp_rtx_head(sk);
2555 if (mptcp_data_fin_enabled(msk)) {
2556 struct inet_connection_sock *icsk = inet_csk(sk);
2558 icsk->icsk_retransmits++;
2559 mptcp_set_datafin_timeout(sk);
2560 mptcp_send_ack(msk);
2565 if (!mptcp_send_head(sk))
2576 /* limit retransmission to the bytes already sent on some subflows */
2578 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2579 while (info.sent < info.limit) {
2580 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2584 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2589 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2590 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2592 WRITE_ONCE(msk->allow_infinite_fallback, false);
2598 mptcp_check_and_set_pending(sk);
2600 if (!mptcp_timer_pending(sk))
2601 mptcp_reset_timer(sk);
2604 /* schedule the timeout timer for the relevant event: either close timeout
2605 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2607 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2609 struct sock *sk = (struct sock *)msk;
2610 unsigned long timeout, close_timeout;
2612 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2615 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2617 /* the close timeout takes precedence on the fail one, and here at least one of
2620 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2622 sk_reset_timer(sk, &sk->sk_timer, timeout);
2625 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2627 struct sock *ssk = msk->first;
2633 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2635 slow = lock_sock_fast(ssk);
2636 mptcp_subflow_reset(ssk);
2637 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2638 unlock_sock_fast(ssk, slow);
2640 mptcp_reset_timeout(msk, 0);
2643 static void mptcp_do_fastclose(struct sock *sk)
2645 struct mptcp_subflow_context *subflow, *tmp;
2646 struct mptcp_sock *msk = mptcp_sk(sk);
2648 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2649 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2650 subflow, MPTCP_CF_FASTCLOSE);
2653 static void mptcp_worker(struct work_struct *work)
2655 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2656 struct sock *sk = &msk->sk.icsk_inet.sk;
2657 unsigned long fail_tout;
2661 state = sk->sk_state;
2662 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2665 mptcp_check_fastclose(msk);
2667 mptcp_pm_nl_work(msk);
2669 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2670 mptcp_check_for_eof(msk);
2672 mptcp_check_send_data_fin(sk);
2673 mptcp_check_data_fin_ack(sk);
2674 mptcp_check_data_fin(sk);
2676 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2677 __mptcp_close_subflow(sk);
2679 /* There is no point in keeping around an orphaned sk timedout or
2680 * closed, but we need the msk around to reply to incoming DATA_FIN,
2681 * even if it is orphaned and in FIN_WAIT2 state
2683 if (sock_flag(sk, SOCK_DEAD)) {
2684 if (mptcp_should_close(sk)) {
2685 inet_sk_state_store(sk, TCP_CLOSE);
2686 mptcp_do_fastclose(sk);
2688 if (sk->sk_state == TCP_CLOSE) {
2689 __mptcp_destroy_sock(sk);
2694 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2695 __mptcp_retrans(sk);
2697 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2698 if (fail_tout && time_after(jiffies, fail_tout))
2699 mptcp_mp_fail_no_response(msk);
2706 static int __mptcp_init_sock(struct sock *sk)
2708 struct mptcp_sock *msk = mptcp_sk(sk);
2710 INIT_LIST_HEAD(&msk->conn_list);
2711 INIT_LIST_HEAD(&msk->join_list);
2712 INIT_LIST_HEAD(&msk->rtx_queue);
2713 INIT_WORK(&msk->work, mptcp_worker);
2714 __skb_queue_head_init(&msk->receive_queue);
2715 msk->out_of_order_queue = RB_ROOT;
2716 msk->first_pending = NULL;
2717 msk->rmem_fwd_alloc = 0;
2718 WRITE_ONCE(msk->rmem_released, 0);
2719 msk->timer_ival = TCP_RTO_MIN;
2721 WRITE_ONCE(msk->first, NULL);
2722 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2723 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2724 WRITE_ONCE(msk->allow_infinite_fallback, true);
2725 msk->recovery = false;
2727 mptcp_pm_data_init(msk);
2729 /* re-use the csk retrans timer for MPTCP-level retrans */
2730 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2731 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2736 static void mptcp_ca_reset(struct sock *sk)
2738 struct inet_connection_sock *icsk = inet_csk(sk);
2740 tcp_assign_congestion_control(sk);
2741 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2743 /* no need to keep a reference to the ops, the name will suffice */
2744 tcp_cleanup_congestion_control(sk);
2745 icsk->icsk_ca_ops = NULL;
2748 static int mptcp_init_sock(struct sock *sk)
2750 struct net *net = sock_net(sk);
2753 ret = __mptcp_init_sock(sk);
2757 if (!mptcp_is_enabled(net))
2758 return -ENOPROTOOPT;
2760 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2763 ret = __mptcp_socket_create(mptcp_sk(sk));
2767 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2768 * propagate the correct value
2772 sk_sockets_allocated_inc(sk);
2773 sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2774 sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2779 static void __mptcp_clear_xmit(struct sock *sk)
2781 struct mptcp_sock *msk = mptcp_sk(sk);
2782 struct mptcp_data_frag *dtmp, *dfrag;
2784 WRITE_ONCE(msk->first_pending, NULL);
2785 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2786 dfrag_clear(sk, dfrag);
2789 void mptcp_cancel_work(struct sock *sk)
2791 struct mptcp_sock *msk = mptcp_sk(sk);
2793 if (cancel_work_sync(&msk->work))
2797 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2801 switch (ssk->sk_state) {
2803 if (!(how & RCV_SHUTDOWN))
2807 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2810 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2811 pr_debug("Fallback");
2812 ssk->sk_shutdown |= how;
2813 tcp_shutdown(ssk, how);
2815 /* simulate the data_fin ack reception to let the state
2816 * machine move forward
2818 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2819 mptcp_schedule_work(sk);
2821 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2823 if (!mptcp_timer_pending(sk))
2824 mptcp_reset_timer(sk);
2832 static const unsigned char new_state[16] = {
2833 /* current state: new state: action: */
2834 [0 /* (Invalid) */] = TCP_CLOSE,
2835 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2836 [TCP_SYN_SENT] = TCP_CLOSE,
2837 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2838 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2839 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2840 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2841 [TCP_CLOSE] = TCP_CLOSE,
2842 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2843 [TCP_LAST_ACK] = TCP_LAST_ACK,
2844 [TCP_LISTEN] = TCP_CLOSE,
2845 [TCP_CLOSING] = TCP_CLOSING,
2846 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2849 static int mptcp_close_state(struct sock *sk)
2851 int next = (int)new_state[sk->sk_state];
2852 int ns = next & TCP_STATE_MASK;
2854 inet_sk_state_store(sk, ns);
2856 return next & TCP_ACTION_FIN;
2859 static void mptcp_check_send_data_fin(struct sock *sk)
2861 struct mptcp_subflow_context *subflow;
2862 struct mptcp_sock *msk = mptcp_sk(sk);
2864 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2865 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2866 msk->snd_nxt, msk->write_seq);
2868 /* we still need to enqueue subflows or not really shutting down,
2871 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2872 mptcp_send_head(sk))
2875 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2877 mptcp_for_each_subflow(msk, subflow) {
2878 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2880 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2884 static void __mptcp_wr_shutdown(struct sock *sk)
2886 struct mptcp_sock *msk = mptcp_sk(sk);
2888 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2889 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2890 !!mptcp_send_head(sk));
2892 /* will be ignored by fallback sockets */
2893 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2894 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2896 mptcp_check_send_data_fin(sk);
2899 static void __mptcp_destroy_sock(struct sock *sk)
2901 struct mptcp_sock *msk = mptcp_sk(sk);
2903 pr_debug("msk=%p", msk);
2907 mptcp_stop_timer(sk);
2908 sk_stop_timer(sk, &sk->sk_timer);
2911 sk->sk_prot->destroy(sk);
2913 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2914 WARN_ON_ONCE(msk->rmem_released);
2915 sk_stream_kill_queues(sk);
2916 xfrm_sk_free_policy(sk);
2918 sk_refcnt_debug_release(sk);
2922 void __mptcp_unaccepted_force_close(struct sock *sk)
2924 sock_set_flag(sk, SOCK_DEAD);
2925 inet_sk_state_store(sk, TCP_CLOSE);
2926 mptcp_do_fastclose(sk);
2927 __mptcp_destroy_sock(sk);
2930 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2932 /* Concurrent splices from sk_receive_queue into receive_queue will
2933 * always show at least one non-empty queue when checked in this order.
2935 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2936 skb_queue_empty_lockless(&msk->receive_queue))
2939 return EPOLLIN | EPOLLRDNORM;
2942 bool __mptcp_close(struct sock *sk, long timeout)
2944 struct mptcp_subflow_context *subflow;
2945 struct mptcp_sock *msk = mptcp_sk(sk);
2946 bool do_cancel_work = false;
2947 int subflows_alive = 0;
2949 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2951 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2952 inet_sk_state_store(sk, TCP_CLOSE);
2956 if (mptcp_check_readable(msk)) {
2957 /* the msk has read data, do the MPTCP equivalent of TCP reset */
2958 inet_sk_state_store(sk, TCP_CLOSE);
2959 mptcp_do_fastclose(sk);
2960 } else if (mptcp_close_state(sk)) {
2961 __mptcp_wr_shutdown(sk);
2964 sk_stream_wait_close(sk, timeout);
2967 /* orphan all the subflows */
2968 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2969 mptcp_for_each_subflow(msk, subflow) {
2970 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2971 bool slow = lock_sock_fast_nested(ssk);
2973 subflows_alive += ssk->sk_state != TCP_CLOSE;
2975 /* since the close timeout takes precedence on the fail one,
2978 if (ssk == msk->first)
2979 subflow->fail_tout = 0;
2981 /* detach from the parent socket, but allow data_ready to
2982 * push incoming data into the mptcp stack, to properly ack it
2984 ssk->sk_socket = NULL;
2986 unlock_sock_fast(ssk, slow);
2990 /* all the subflows are closed, only timeout can change the msk
2991 * state, let's not keep resources busy for no reasons
2993 if (subflows_alive == 0)
2994 inet_sk_state_store(sk, TCP_CLOSE);
2997 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2998 if (mptcp_sk(sk)->token)
2999 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3001 if (sk->sk_state == TCP_CLOSE) {
3002 __mptcp_destroy_sock(sk);
3003 do_cancel_work = true;
3005 mptcp_reset_timeout(msk, 0);
3008 return do_cancel_work;
3011 static void mptcp_close(struct sock *sk, long timeout)
3013 bool do_cancel_work;
3017 do_cancel_work = __mptcp_close(sk, timeout);
3020 mptcp_cancel_work(sk);
3025 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3027 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3028 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3029 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3031 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3032 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3035 msk6->saddr = ssk6->saddr;
3036 msk6->flow_label = ssk6->flow_label;
3040 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3041 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3042 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3043 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3044 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3045 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3048 static int mptcp_disconnect(struct sock *sk, int flags)
3050 struct mptcp_sock *msk = mptcp_sk(sk);
3052 /* Deny disconnect if other threads are blocked in sk_wait_event()
3053 * or inet_wait_for_connect().
3055 if (sk->sk_wait_pending)
3058 /* We are on the fastopen error path. We can't call straight into the
3059 * subflows cleanup code due to lock nesting (we are already under
3060 * msk->firstsocket lock).
3062 if (msk->fastopening)
3065 inet_sk_state_store(sk, TCP_CLOSE);
3067 mptcp_stop_timer(sk);
3068 sk_stop_timer(sk, &sk->sk_timer);
3070 if (mptcp_sk(sk)->token)
3071 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
3073 /* msk->subflow is still intact, the following will not free the first
3076 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3077 msk->last_snd = NULL;
3078 WRITE_ONCE(msk->flags, 0);
3080 msk->push_pending = 0;
3081 msk->recovery = false;
3082 msk->can_ack = false;
3083 msk->fully_established = false;
3084 msk->rcv_data_fin = false;
3085 msk->snd_data_fin_enable = false;
3086 msk->rcv_fastclose = false;
3087 msk->use_64bit_ack = false;
3088 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3089 mptcp_pm_data_reset(msk);
3092 WRITE_ONCE(sk->sk_shutdown, 0);
3093 sk_error_report(sk);
3097 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3098 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3100 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3102 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3106 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3107 const struct mptcp_options_received *mp_opt,
3109 struct request_sock *req)
3111 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3112 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3113 struct mptcp_sock *msk;
3119 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3120 if (nsk->sk_family == AF_INET6)
3121 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3124 nsk->sk_wait_pending = 0;
3125 __mptcp_init_sock(nsk);
3127 msk = mptcp_sk(nsk);
3128 msk->local_key = subflow_req->local_key;
3129 msk->token = subflow_req->token;
3130 WRITE_ONCE(msk->subflow, NULL);
3131 msk->in_accept_queue = 1;
3132 WRITE_ONCE(msk->fully_established, false);
3133 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3134 WRITE_ONCE(msk->csum_enabled, true);
3136 msk->write_seq = subflow_req->idsn + 1;
3137 msk->snd_nxt = msk->write_seq;
3138 msk->snd_una = msk->write_seq;
3139 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3140 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3142 if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
3143 msk->can_ack = true;
3144 msk->remote_key = mp_opt->sndr_key;
3145 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
3147 WRITE_ONCE(msk->ack_seq, ack_seq);
3148 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3151 sock_reset_flag(nsk, SOCK_RCU_FREE);
3152 security_inet_csk_clone(nsk, req);
3154 /* this can't race with mptcp_close(), as the msk is
3155 * not yet exposted to user-space
3157 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3159 /* The msk maintain a ref to each subflow in the connections list */
3160 WRITE_ONCE(msk->first, ssk);
3161 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3164 /* new mpc subflow takes ownership of the newly
3165 * created mptcp socket
3167 mptcp_token_accept(subflow_req, msk);
3169 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3170 * uses the correct data
3172 mptcp_copy_inaddrs(nsk, ssk);
3173 mptcp_propagate_sndbuf(nsk, ssk);
3175 mptcp_rcv_space_init(msk, ssk);
3176 bh_unlock_sock(nsk);
3178 /* note: the newly allocated socket refcount is 2 now */
3182 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3184 const struct tcp_sock *tp = tcp_sk(ssk);
3186 msk->rcvq_space.copied = 0;
3187 msk->rcvq_space.rtt_us = 0;
3189 msk->rcvq_space.time = tp->tcp_mstamp;
3191 /* initial rcv_space offering made to peer */
3192 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3193 TCP_INIT_CWND * tp->advmss);
3194 if (msk->rcvq_space.space == 0)
3195 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3197 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3200 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3203 struct mptcp_sock *msk = mptcp_sk(sk);
3204 struct socket *listener;
3207 listener = READ_ONCE(msk->subflow);
3208 if (WARN_ON_ONCE(!listener)) {
3213 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3214 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3218 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3219 if (sk_is_mptcp(newsk)) {
3220 struct mptcp_subflow_context *subflow;
3221 struct sock *new_mptcp_sock;
3223 subflow = mptcp_subflow_ctx(newsk);
3224 new_mptcp_sock = subflow->conn;
3226 /* is_mptcp should be false if subflow->conn is missing, see
3227 * subflow_syn_recv_sock()
3229 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3230 tcp_sk(newsk)->is_mptcp = 0;
3234 newsk = new_mptcp_sock;
3235 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3237 MPTCP_INC_STATS(sock_net(sk),
3238 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3242 newsk->sk_kern_sock = kern;
3246 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3248 struct mptcp_subflow_context *subflow, *tmp;
3249 struct sock *sk = (struct sock *)msk;
3251 __mptcp_clear_xmit(sk);
3253 /* join list will be eventually flushed (with rst) at sock lock release time */
3254 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3255 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3257 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3258 mptcp_data_lock(sk);
3259 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3260 __skb_queue_purge(&sk->sk_receive_queue);
3261 skb_rbtree_purge(&msk->out_of_order_queue);
3262 mptcp_data_unlock(sk);
3264 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3265 * inet_sock_destruct() will dispose it
3267 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3268 msk->rmem_fwd_alloc = 0;
3269 mptcp_token_destroy(msk);
3270 mptcp_pm_free_anno_list(msk);
3271 mptcp_free_local_addr_list(msk);
3274 static void mptcp_destroy(struct sock *sk)
3276 struct mptcp_sock *msk = mptcp_sk(sk);
3278 /* clears msk->subflow, allowing the following to close
3279 * even the initial subflow
3281 mptcp_dispose_initial_subflow(msk);
3282 mptcp_destroy_common(msk, 0);
3283 sk_sockets_allocated_dec(sk);
3286 void __mptcp_data_acked(struct sock *sk)
3288 if (!sock_owned_by_user(sk))
3289 __mptcp_clean_una(sk);
3291 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3293 if (mptcp_pending_data_fin_ack(sk))
3294 mptcp_schedule_work(sk);
3297 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3299 if (!mptcp_send_head(sk))
3302 if (!sock_owned_by_user(sk)) {
3303 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3305 if (xmit_ssk == ssk)
3306 __mptcp_subflow_push_pending(sk, ssk);
3308 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3310 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3314 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3315 BIT(MPTCP_RETRANSMIT) | \
3316 BIT(MPTCP_FLUSH_JOIN_LIST))
3318 /* processes deferred events and flush wmem */
3319 static void mptcp_release_cb(struct sock *sk)
3320 __must_hold(&sk->sk_lock.slock)
3322 struct mptcp_sock *msk = mptcp_sk(sk);
3325 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3327 struct list_head join_list;
3332 INIT_LIST_HEAD(&join_list);
3333 list_splice_init(&msk->join_list, &join_list);
3335 /* the following actions acquire the subflow socket lock
3337 * 1) can't be invoked in atomic scope
3338 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3339 * datapath acquires the msk socket spinlock while helding
3340 * the subflow socket lock
3342 msk->push_pending = 0;
3343 msk->cb_flags &= ~flags;
3344 spin_unlock_bh(&sk->sk_lock.slock);
3346 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3347 __mptcp_flush_join_list(sk, &join_list);
3348 if (flags & BIT(MPTCP_PUSH_PENDING))
3349 __mptcp_push_pending(sk, 0);
3350 if (flags & BIT(MPTCP_RETRANSMIT))
3351 __mptcp_retrans(sk);
3354 spin_lock_bh(&sk->sk_lock.slock);
3357 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3358 __mptcp_clean_una_wakeup(sk);
3359 if (unlikely(&msk->cb_flags)) {
3360 /* be sure to set the current sk state before tacking actions
3361 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3363 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3364 __mptcp_set_connected(sk);
3365 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3366 __mptcp_error_report(sk);
3367 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3368 msk->last_snd = NULL;
3371 __mptcp_update_rmem(sk);
3374 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3375 * TCP can't schedule delack timer before the subflow is fully established.
3376 * MPTCP uses the delack timer to do 3rd ack retransmissions
3378 static void schedule_3rdack_retransmission(struct sock *ssk)
3380 struct inet_connection_sock *icsk = inet_csk(ssk);
3381 struct tcp_sock *tp = tcp_sk(ssk);
3382 unsigned long timeout;
3384 if (mptcp_subflow_ctx(ssk)->fully_established)
3387 /* reschedule with a timeout above RTT, as we must look only for drop */
3389 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3391 timeout = TCP_TIMEOUT_INIT;
3394 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3395 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3396 icsk->icsk_ack.timeout = timeout;
3397 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3400 void mptcp_subflow_process_delegated(struct sock *ssk)
3402 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3403 struct sock *sk = subflow->conn;
3405 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3406 mptcp_data_lock(sk);
3407 if (!sock_owned_by_user(sk))
3408 __mptcp_subflow_push_pending(sk, ssk);
3410 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3411 mptcp_data_unlock(sk);
3412 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3414 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3415 schedule_3rdack_retransmission(ssk);
3416 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3420 static int mptcp_hash(struct sock *sk)
3422 /* should never be called,
3423 * we hash the TCP subflows not the master socket
3429 static void mptcp_unhash(struct sock *sk)
3431 /* called from sk_common_release(), but nothing to do here */
3434 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3436 struct mptcp_sock *msk = mptcp_sk(sk);
3437 struct socket *ssock;
3439 ssock = msk->subflow;
3440 pr_debug("msk=%p, subflow=%p", msk, ssock);
3441 if (WARN_ON_ONCE(!ssock))
3444 return inet_csk_get_port(ssock->sk, snum);
3447 void mptcp_finish_connect(struct sock *ssk)
3449 struct mptcp_subflow_context *subflow;
3450 struct mptcp_sock *msk;
3454 subflow = mptcp_subflow_ctx(ssk);
3458 pr_debug("msk=%p, token=%u", sk, subflow->token);
3460 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3462 subflow->map_seq = ack_seq;
3463 subflow->map_subflow_seq = 1;
3465 /* the socket is not connected yet, no msk/subflow ops can access/race
3466 * accessing the field below
3468 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3469 WRITE_ONCE(msk->local_key, subflow->local_key);
3470 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3471 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3472 WRITE_ONCE(msk->ack_seq, ack_seq);
3473 WRITE_ONCE(msk->can_ack, 1);
3474 WRITE_ONCE(msk->snd_una, msk->write_seq);
3475 atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3477 mptcp_pm_new_connection(msk, ssk, 0);
3479 mptcp_rcv_space_init(msk, ssk);
3482 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3484 write_lock_bh(&sk->sk_callback_lock);
3485 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3486 sk_set_socket(sk, parent);
3487 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3488 write_unlock_bh(&sk->sk_callback_lock);
3491 bool mptcp_finish_join(struct sock *ssk)
3493 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3494 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3495 struct sock *parent = (void *)msk;
3498 pr_debug("msk=%p, subflow=%p", msk, subflow);
3500 /* mptcp socket already closing? */
3501 if (!mptcp_is_fully_established(parent)) {
3502 subflow->reset_reason = MPTCP_RST_EMPTCP;
3506 /* active subflow, already present inside the conn_list */
3507 if (!list_empty(&subflow->node)) {
3508 mptcp_subflow_joined(msk, ssk);
3512 if (!mptcp_pm_allow_new_subflow(msk))
3513 goto err_prohibited;
3515 /* If we can't acquire msk socket lock here, let the release callback
3518 mptcp_data_lock(parent);
3519 if (!sock_owned_by_user(parent)) {
3520 ret = __mptcp_finish_join(msk, ssk);
3523 list_add_tail(&subflow->node, &msk->conn_list);
3527 list_add_tail(&subflow->node, &msk->join_list);
3528 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3530 mptcp_data_unlock(parent);
3534 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3541 static void mptcp_shutdown(struct sock *sk, int how)
3543 pr_debug("sk=%p, how=%d", sk, how);
3545 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3546 __mptcp_wr_shutdown(sk);
3549 static int mptcp_forward_alloc_get(const struct sock *sk)
3551 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3554 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3556 const struct sock *sk = (void *)msk;
3559 if (sk->sk_state == TCP_LISTEN)
3562 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3565 delta = msk->write_seq - v;
3566 if (__mptcp_check_fallback(msk) && msk->first) {
3567 struct tcp_sock *tp = tcp_sk(msk->first);
3569 /* the first subflow is disconnected after close - see
3570 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3571 * so ignore that status, too.
3573 if (!((1 << msk->first->sk_state) &
3574 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3575 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3577 if (delta > INT_MAX)
3583 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3585 struct mptcp_sock *msk = mptcp_sk(sk);
3591 if (sk->sk_state == TCP_LISTEN)
3595 __mptcp_move_skbs(msk);
3596 answ = mptcp_inq_hint(sk);
3600 slow = lock_sock_fast(sk);
3601 answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3602 unlock_sock_fast(sk, slow);
3605 slow = lock_sock_fast(sk);
3606 answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3607 unlock_sock_fast(sk, slow);
3610 return -ENOIOCTLCMD;
3613 return put_user(answ, (int __user *)arg);
3616 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3617 struct mptcp_subflow_context *subflow)
3619 subflow->request_mptcp = 0;
3620 __mptcp_do_fallback(msk);
3623 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3625 struct mptcp_subflow_context *subflow;
3626 struct mptcp_sock *msk = mptcp_sk(sk);
3627 struct socket *ssock;
3630 ssock = __mptcp_nmpc_socket(msk);
3634 mptcp_token_destroy(msk);
3635 inet_sk_state_store(sk, TCP_SYN_SENT);
3636 subflow = mptcp_subflow_ctx(ssock->sk);
3637 #ifdef CONFIG_TCP_MD5SIG
3638 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3641 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3642 mptcp_subflow_early_fallback(msk, subflow);
3644 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3645 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3646 mptcp_subflow_early_fallback(msk, subflow);
3648 if (likely(!__mptcp_check_fallback(msk)))
3649 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3651 /* if reaching here via the fastopen/sendmsg path, the caller already
3652 * acquired the subflow socket lock, too.
3654 if (msk->fastopening)
3655 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3657 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3658 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3660 /* on successful connect, the msk state will be moved to established by
3661 * subflow_finish_connect()
3663 if (unlikely(err && err != -EINPROGRESS)) {
3664 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3668 mptcp_copy_inaddrs(sk, ssock->sk);
3670 /* silence EINPROGRESS and let the caller inet_stream_connect
3671 * handle the connection in progress
3676 static struct proto mptcp_prot = {
3678 .owner = THIS_MODULE,
3679 .init = mptcp_init_sock,
3680 .connect = mptcp_connect,
3681 .disconnect = mptcp_disconnect,
3682 .close = mptcp_close,
3683 .accept = mptcp_accept,
3684 .setsockopt = mptcp_setsockopt,
3685 .getsockopt = mptcp_getsockopt,
3686 .shutdown = mptcp_shutdown,
3687 .destroy = mptcp_destroy,
3688 .sendmsg = mptcp_sendmsg,
3689 .ioctl = mptcp_ioctl,
3690 .recvmsg = mptcp_recvmsg,
3691 .release_cb = mptcp_release_cb,
3693 .unhash = mptcp_unhash,
3694 .get_port = mptcp_get_port,
3695 .forward_alloc_get = mptcp_forward_alloc_get,
3696 .sockets_allocated = &mptcp_sockets_allocated,
3698 .memory_allocated = &tcp_memory_allocated,
3699 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3701 .memory_pressure = &tcp_memory_pressure,
3702 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3703 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3704 .sysctl_mem = sysctl_tcp_mem,
3705 .obj_size = sizeof(struct mptcp_sock),
3706 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3707 .no_autobind = true,
3710 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3712 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3713 struct socket *ssock;
3716 lock_sock(sock->sk);
3717 ssock = __mptcp_nmpc_socket(msk);
3723 err = ssock->ops->bind(ssock, uaddr, addr_len);
3725 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3728 release_sock(sock->sk);
3732 static int mptcp_listen(struct socket *sock, int backlog)
3734 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3735 struct socket *ssock;
3738 pr_debug("msk=%p", msk);
3740 lock_sock(sock->sk);
3741 ssock = __mptcp_nmpc_socket(msk);
3747 mptcp_token_destroy(msk);
3748 inet_sk_state_store(sock->sk, TCP_LISTEN);
3749 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3751 err = ssock->ops->listen(ssock, backlog);
3752 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3754 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3757 release_sock(sock->sk);
3761 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3762 int flags, bool kern)
3764 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3765 struct socket *ssock;
3768 pr_debug("msk=%p", msk);
3770 /* Buggy applications can call accept on socket states other then LISTEN
3771 * but no need to allocate the first subflow just to error out.
3773 ssock = READ_ONCE(msk->subflow);
3777 err = ssock->ops->accept(sock, newsock, flags, kern);
3778 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3779 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3780 struct mptcp_subflow_context *subflow;
3781 struct sock *newsk = newsock->sk;
3783 msk->in_accept_queue = 0;
3787 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3788 * This is needed so NOSPACE flag can be set from tcp stack.
3790 mptcp_for_each_subflow(msk, subflow) {
3791 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3793 if (!ssk->sk_socket)
3794 mptcp_sock_graft(ssk, newsock);
3797 /* Do late cleanup for the first subflow as necessary. Also
3798 * deal with bad peers not doing a complete shutdown.
3801 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3802 __mptcp_close_ssk(newsk, msk->first,
3803 mptcp_subflow_ctx(msk->first), 0);
3804 if (unlikely(list_empty(&msk->conn_list)))
3805 inet_sk_state_store(newsk, TCP_CLOSE);
3808 release_sock(newsk);
3814 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3816 struct sock *sk = (struct sock *)msk;
3818 if (sk_stream_is_writeable(sk))
3819 return EPOLLOUT | EPOLLWRNORM;
3821 mptcp_set_nospace(sk);
3822 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3823 if (sk_stream_is_writeable(sk))
3824 return EPOLLOUT | EPOLLWRNORM;
3829 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3830 struct poll_table_struct *wait)
3832 struct sock *sk = sock->sk;
3833 struct mptcp_sock *msk;
3839 sock_poll_wait(file, sock, wait);
3841 state = inet_sk_state_load(sk);
3842 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3843 if (state == TCP_LISTEN) {
3844 struct socket *ssock = READ_ONCE(msk->subflow);
3846 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3849 return inet_csk_listen_poll(ssock->sk);
3852 shutdown = READ_ONCE(sk->sk_shutdown);
3853 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3855 if (shutdown & RCV_SHUTDOWN)
3856 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3858 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3859 mask |= mptcp_check_readable(msk);
3860 if (shutdown & SEND_SHUTDOWN)
3861 mask |= EPOLLOUT | EPOLLWRNORM;
3863 mask |= mptcp_check_writeable(msk);
3864 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3865 /* cf tcp_poll() note about TFO */
3866 mask |= EPOLLOUT | EPOLLWRNORM;
3869 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3877 static const struct proto_ops mptcp_stream_ops = {
3879 .owner = THIS_MODULE,
3880 .release = inet_release,
3882 .connect = inet_stream_connect,
3883 .socketpair = sock_no_socketpair,
3884 .accept = mptcp_stream_accept,
3885 .getname = inet_getname,
3887 .ioctl = inet_ioctl,
3888 .gettstamp = sock_gettstamp,
3889 .listen = mptcp_listen,
3890 .shutdown = inet_shutdown,
3891 .setsockopt = sock_common_setsockopt,
3892 .getsockopt = sock_common_getsockopt,
3893 .sendmsg = inet_sendmsg,
3894 .recvmsg = inet_recvmsg,
3895 .mmap = sock_no_mmap,
3896 .sendpage = inet_sendpage,
3899 static struct inet_protosw mptcp_protosw = {
3900 .type = SOCK_STREAM,
3901 .protocol = IPPROTO_MPTCP,
3902 .prot = &mptcp_prot,
3903 .ops = &mptcp_stream_ops,
3904 .flags = INET_PROTOSW_ICSK,
3907 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3909 struct mptcp_delegated_action *delegated;
3910 struct mptcp_subflow_context *subflow;
3913 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3914 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3915 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3917 bh_lock_sock_nested(ssk);
3918 if (!sock_owned_by_user(ssk) &&
3919 mptcp_subflow_has_delegated_action(subflow))
3920 mptcp_subflow_process_delegated(ssk);
3921 /* ... elsewhere tcp_release_cb_override already processed
3922 * the action or will do at next release_sock().
3923 * In both case must dequeue the subflow here - on the same
3924 * CPU that scheduled it.
3926 bh_unlock_sock(ssk);
3929 if (++work_done == budget)
3933 /* always provide a 0 'work_done' argument, so that napi_complete_done
3934 * will not try accessing the NULL napi->dev ptr
3936 napi_complete_done(napi, 0);
3940 void __init mptcp_proto_init(void)
3942 struct mptcp_delegated_action *delegated;
3945 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3947 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3948 panic("Failed to allocate MPTCP pcpu counter\n");
3950 init_dummy_netdev(&mptcp_napi_dev);
3951 for_each_possible_cpu(cpu) {
3952 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3953 INIT_LIST_HEAD(&delegated->head);
3954 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3956 napi_enable(&delegated->napi);
3959 mptcp_subflow_init();
3963 if (proto_register(&mptcp_prot, 1) != 0)
3964 panic("Failed to register MPTCP proto.\n");
3966 inet_register_protosw(&mptcp_protosw);
3968 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3971 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3972 static const struct proto_ops mptcp_v6_stream_ops = {
3974 .owner = THIS_MODULE,
3975 .release = inet6_release,
3977 .connect = inet_stream_connect,
3978 .socketpair = sock_no_socketpair,
3979 .accept = mptcp_stream_accept,
3980 .getname = inet6_getname,
3982 .ioctl = inet6_ioctl,
3983 .gettstamp = sock_gettstamp,
3984 .listen = mptcp_listen,
3985 .shutdown = inet_shutdown,
3986 .setsockopt = sock_common_setsockopt,
3987 .getsockopt = sock_common_getsockopt,
3988 .sendmsg = inet6_sendmsg,
3989 .recvmsg = inet6_recvmsg,
3990 .mmap = sock_no_mmap,
3991 .sendpage = inet_sendpage,
3992 #ifdef CONFIG_COMPAT
3993 .compat_ioctl = inet6_compat_ioctl,
3997 static struct proto mptcp_v6_prot;
3999 static struct inet_protosw mptcp_v6_protosw = {
4000 .type = SOCK_STREAM,
4001 .protocol = IPPROTO_MPTCP,
4002 .prot = &mptcp_v6_prot,
4003 .ops = &mptcp_v6_stream_ops,
4004 .flags = INET_PROTOSW_ICSK,
4007 int __init mptcp_proto_v6_init(void)
4011 mptcp_v6_prot = mptcp_prot;
4012 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4013 mptcp_v6_prot.slab = NULL;
4014 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4016 err = proto_register(&mptcp_v6_prot, 1);
4020 err = inet6_register_protosw(&mptcp_v6_protosw);
4022 proto_unregister(&mptcp_v6_prot);