1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
40 MPTCP_CMSG_TS = BIT(0),
41 MPTCP_CMSG_INQ = BIT(1),
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void mptcp_check_send_data_fin(struct sock *sk);
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
52 /* Returns end sequence number of the receiver's advertised window */
53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
55 return READ_ONCE(msk->wnd_end);
58 static bool mptcp_is_tcpsk(struct sock *sk)
60 struct socket *sock = sk->sk_socket;
62 if (unlikely(sk->sk_prot == &tcp_prot)) {
63 /* we are being invoked after mptcp_accept() has
64 * accepted a non-mp-capable flow: sk is a tcp_sk,
67 * Hand the socket over to tcp so all further socket ops
70 WRITE_ONCE(sock->ops, &inet_stream_ops);
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 WRITE_ONCE(sock->ops, &inet6_stream_ops);
82 static int __mptcp_socket_create(struct mptcp_sock *msk)
84 struct mptcp_subflow_context *subflow;
85 struct sock *sk = (struct sock *)msk;
89 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
93 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
94 WRITE_ONCE(msk->first, ssock->sk);
95 subflow = mptcp_subflow_ctx(ssock->sk);
96 list_add(&subflow->node, &msk->conn_list);
98 subflow->request_mptcp = 1;
99 subflow->subflow_id = msk->subflow_id++;
101 /* This is the first subflow, always with id 0 */
102 subflow->local_id_valid = 1;
103 mptcp_sock_graft(msk->first, sk->sk_socket);
104 iput(SOCK_INODE(ssock));
109 /* If the MPC handshake is not started, returns the first subflow,
110 * eventually allocating it.
112 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
114 struct sock *sk = (struct sock *)msk;
117 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
118 return ERR_PTR(-EINVAL);
121 ret = __mptcp_socket_create(msk);
125 mptcp_sockopt_sync(msk, msk->first);
131 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
133 sk_drops_add(sk, skb);
137 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
139 WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
140 mptcp_sk(sk)->rmem_fwd_alloc + size);
143 static void mptcp_rmem_charge(struct sock *sk, int size)
145 mptcp_rmem_fwd_alloc_add(sk, -size);
148 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
149 struct sk_buff *from)
154 if (MPTCP_SKB_CB(from)->offset ||
155 !skb_try_coalesce(to, from, &fragstolen, &delta))
158 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
159 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
160 to->len, MPTCP_SKB_CB(from)->end_seq);
161 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
163 /* note the fwd memory can reach a negative value after accounting
164 * for the delta, but the later skb free will restore a non
167 atomic_add(delta, &sk->sk_rmem_alloc);
168 mptcp_rmem_charge(sk, delta);
169 kfree_skb_partial(from, fragstolen);
174 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
175 struct sk_buff *from)
177 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
180 return mptcp_try_coalesce((struct sock *)msk, to, from);
183 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
185 amount >>= PAGE_SHIFT;
186 mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
187 __sk_mem_reduce_allocated(sk, amount);
190 static void mptcp_rmem_uncharge(struct sock *sk, int size)
192 struct mptcp_sock *msk = mptcp_sk(sk);
195 mptcp_rmem_fwd_alloc_add(sk, size);
196 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
198 /* see sk_mem_uncharge() for the rationale behind the following schema */
199 if (unlikely(reclaimable >= PAGE_SIZE))
200 __mptcp_rmem_reclaim(sk, reclaimable);
203 static void mptcp_rfree(struct sk_buff *skb)
205 unsigned int len = skb->truesize;
206 struct sock *sk = skb->sk;
208 atomic_sub(len, &sk->sk_rmem_alloc);
209 mptcp_rmem_uncharge(sk, len);
212 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
216 skb->destructor = mptcp_rfree;
217 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
218 mptcp_rmem_charge(sk, skb->truesize);
221 /* "inspired" by tcp_data_queue_ofo(), main differences:
223 * - don't cope with sacks
225 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
227 struct sock *sk = (struct sock *)msk;
228 struct rb_node **p, *parent;
229 u64 seq, end_seq, max_seq;
230 struct sk_buff *skb1;
232 seq = MPTCP_SKB_CB(skb)->map_seq;
233 end_seq = MPTCP_SKB_CB(skb)->end_seq;
234 max_seq = atomic64_read(&msk->rcv_wnd_sent);
236 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
237 RB_EMPTY_ROOT(&msk->out_of_order_queue));
238 if (after64(end_seq, max_seq)) {
241 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
242 (unsigned long long)end_seq - (unsigned long)max_seq,
243 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
248 p = &msk->out_of_order_queue.rb_node;
249 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
250 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
251 rb_link_node(&skb->rbnode, NULL, p);
252 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
253 msk->ooo_last_skb = skb;
257 /* with 2 subflows, adding at end of ooo queue is quite likely
258 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
260 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
261 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
266 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
267 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
268 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
269 parent = &msk->ooo_last_skb->rbnode;
270 p = &parent->rb_right;
274 /* Find place to insert this segment. Handle overlaps on the way. */
278 skb1 = rb_to_skb(parent);
279 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
280 p = &parent->rb_left;
283 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
284 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
285 /* All the bits are present. Drop. */
287 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
290 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
294 * continue traversing
297 /* skb's seq == skb1's seq and skb covers skb1.
298 * Replace skb1 with skb.
300 rb_replace_node(&skb1->rbnode, &skb->rbnode,
301 &msk->out_of_order_queue);
302 mptcp_drop(sk, skb1);
303 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
306 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
307 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
310 p = &parent->rb_right;
314 /* Insert segment into RB tree. */
315 rb_link_node(&skb->rbnode, parent, p);
316 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
319 /* Remove other segments covered by skb. */
320 while ((skb1 = skb_rb_next(skb)) != NULL) {
321 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
323 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
324 mptcp_drop(sk, skb1);
325 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
327 /* If there is no skb after us, we are the last_skb ! */
329 msk->ooo_last_skb = skb;
333 mptcp_set_owner_r(skb, sk);
336 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
338 struct mptcp_sock *msk = mptcp_sk(sk);
341 if (size <= msk->rmem_fwd_alloc)
344 size -= msk->rmem_fwd_alloc;
345 amt = sk_mem_pages(size);
346 amount = amt << PAGE_SHIFT;
347 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
350 mptcp_rmem_fwd_alloc_add(sk, amount);
354 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
355 struct sk_buff *skb, unsigned int offset,
358 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
359 struct sock *sk = (struct sock *)msk;
360 struct sk_buff *tail;
363 __skb_unlink(skb, &ssk->sk_receive_queue);
368 /* try to fetch required memory from subflow */
369 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
372 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
374 /* the skb map_seq accounts for the skb offset:
375 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
378 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
379 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
380 MPTCP_SKB_CB(skb)->offset = offset;
381 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
383 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
385 msk->bytes_received += copy_len;
386 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
387 tail = skb_peek_tail(&sk->sk_receive_queue);
388 if (tail && mptcp_try_coalesce(sk, tail, skb))
391 mptcp_set_owner_r(skb, sk);
392 __skb_queue_tail(&sk->sk_receive_queue, skb);
394 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
395 mptcp_data_queue_ofo(msk, skb);
399 /* old data, keep it simple and drop the whole pkt, sender
400 * will retransmit as needed, if needed.
402 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
408 static void mptcp_stop_timer(struct sock *sk)
410 struct inet_connection_sock *icsk = inet_csk(sk);
412 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
413 mptcp_sk(sk)->timer_ival = 0;
416 static void mptcp_close_wake_up(struct sock *sk)
418 if (sock_flag(sk, SOCK_DEAD))
421 sk->sk_state_change(sk);
422 if (sk->sk_shutdown == SHUTDOWN_MASK ||
423 sk->sk_state == TCP_CLOSE)
424 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
426 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
429 static bool mptcp_pending_data_fin_ack(struct sock *sk)
431 struct mptcp_sock *msk = mptcp_sk(sk);
433 return ((1 << sk->sk_state) &
434 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
435 msk->write_seq == READ_ONCE(msk->snd_una);
438 static void mptcp_check_data_fin_ack(struct sock *sk)
440 struct mptcp_sock *msk = mptcp_sk(sk);
442 /* Look for an acknowledged DATA_FIN */
443 if (mptcp_pending_data_fin_ack(sk)) {
444 WRITE_ONCE(msk->snd_data_fin_enable, 0);
446 switch (sk->sk_state) {
448 inet_sk_state_store(sk, TCP_FIN_WAIT2);
452 inet_sk_state_store(sk, TCP_CLOSE);
456 mptcp_close_wake_up(sk);
460 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
462 struct mptcp_sock *msk = mptcp_sk(sk);
464 if (READ_ONCE(msk->rcv_data_fin) &&
465 ((1 << sk->sk_state) &
466 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
467 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
469 if (msk->ack_seq == rcv_data_fin_seq) {
471 *seq = rcv_data_fin_seq;
480 static void mptcp_set_datafin_timeout(struct sock *sk)
482 struct inet_connection_sock *icsk = inet_csk(sk);
485 retransmits = min_t(u32, icsk->icsk_retransmits,
486 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
488 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
491 static void __mptcp_set_timeout(struct sock *sk, long tout)
493 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
496 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
498 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
500 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
501 inet_csk(ssk)->icsk_timeout - jiffies : 0;
504 static void mptcp_set_timeout(struct sock *sk)
506 struct mptcp_subflow_context *subflow;
509 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
510 tout = max(tout, mptcp_timeout_from_subflow(subflow));
511 __mptcp_set_timeout(sk, tout);
514 static inline bool tcp_can_send_ack(const struct sock *ssk)
516 return !((1 << inet_sk_state_load(ssk)) &
517 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
520 void __mptcp_subflow_send_ack(struct sock *ssk)
522 if (tcp_can_send_ack(ssk))
526 static void mptcp_subflow_send_ack(struct sock *ssk)
530 slow = lock_sock_fast(ssk);
531 __mptcp_subflow_send_ack(ssk);
532 unlock_sock_fast(ssk, slow);
535 static void mptcp_send_ack(struct mptcp_sock *msk)
537 struct mptcp_subflow_context *subflow;
539 mptcp_for_each_subflow(msk, subflow)
540 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
543 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
547 slow = lock_sock_fast(ssk);
548 if (tcp_can_send_ack(ssk))
549 tcp_cleanup_rbuf(ssk, 1);
550 unlock_sock_fast(ssk, slow);
553 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
555 const struct inet_connection_sock *icsk = inet_csk(ssk);
556 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
557 const struct tcp_sock *tp = tcp_sk(ssk);
559 return (ack_pending & ICSK_ACK_SCHED) &&
560 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
561 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
562 (rx_empty && ack_pending &
563 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
566 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
568 int old_space = READ_ONCE(msk->old_wspace);
569 struct mptcp_subflow_context *subflow;
570 struct sock *sk = (struct sock *)msk;
571 int space = __mptcp_space(sk);
572 bool cleanup, rx_empty;
574 cleanup = (space > 0) && (space >= (old_space << 1));
575 rx_empty = !__mptcp_rmem(sk);
577 mptcp_for_each_subflow(msk, subflow) {
578 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
580 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
581 mptcp_subflow_cleanup_rbuf(ssk);
585 static bool mptcp_check_data_fin(struct sock *sk)
587 struct mptcp_sock *msk = mptcp_sk(sk);
588 u64 rcv_data_fin_seq;
591 /* Need to ack a DATA_FIN received from a peer while this side
592 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
593 * msk->rcv_data_fin was set when parsing the incoming options
594 * at the subflow level and the msk lock was not held, so this
595 * is the first opportunity to act on the DATA_FIN and change
598 * If we are caught up to the sequence number of the incoming
599 * DATA_FIN, send the DATA_ACK now and do state transition. If
600 * not caught up, do nothing and let the recv code send DATA_ACK
604 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
605 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
606 WRITE_ONCE(msk->rcv_data_fin, 0);
608 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
609 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
611 switch (sk->sk_state) {
612 case TCP_ESTABLISHED:
613 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
616 inet_sk_state_store(sk, TCP_CLOSING);
619 inet_sk_state_store(sk, TCP_CLOSE);
622 /* Other states not expected */
628 if (!__mptcp_check_fallback(msk))
630 mptcp_close_wake_up(sk);
635 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
639 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
640 struct sock *sk = (struct sock *)msk;
641 unsigned int moved = 0;
642 bool more_data_avail;
647 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
649 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
650 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
652 if (unlikely(ssk_rbuf > sk_rbuf)) {
653 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
658 pr_debug("msk=%p ssk=%p", msk, ssk);
661 u32 map_remaining, offset;
662 u32 seq = tp->copied_seq;
666 /* try to move as much data as available */
667 map_remaining = subflow->map_data_len -
668 mptcp_subflow_get_map_offset(subflow);
670 skb = skb_peek(&ssk->sk_receive_queue);
672 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
673 * a different CPU can have already processed the pending
674 * data, stop here or we can enter an infinite loop
681 if (__mptcp_check_fallback(msk)) {
682 /* Under fallback skbs have no MPTCP extension and TCP could
683 * collapse them between the dummy map creation and the
684 * current dequeue. Be sure to adjust the map size.
686 map_remaining = skb->len;
687 subflow->map_data_len = skb->len;
690 offset = seq - TCP_SKB_CB(skb)->seq;
691 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
697 if (offset < skb->len) {
698 size_t len = skb->len - offset;
703 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
707 if (WARN_ON_ONCE(map_remaining < len))
711 sk_eat_skb(ssk, skb);
715 WRITE_ONCE(tp->copied_seq, seq);
716 more_data_avail = mptcp_subflow_data_available(ssk);
718 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
722 } while (more_data_avail);
728 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
730 struct sock *sk = (struct sock *)msk;
731 struct sk_buff *skb, *tail;
736 p = rb_first(&msk->out_of_order_queue);
737 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
740 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
744 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
746 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
749 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
753 end_seq = MPTCP_SKB_CB(skb)->end_seq;
754 tail = skb_peek_tail(&sk->sk_receive_queue);
755 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
756 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
758 /* skip overlapping data, if any */
759 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
760 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
762 MPTCP_SKB_CB(skb)->offset += delta;
763 MPTCP_SKB_CB(skb)->map_seq += delta;
764 __skb_queue_tail(&sk->sk_receive_queue, skb);
766 msk->bytes_received += end_seq - msk->ack_seq;
767 msk->ack_seq = end_seq;
773 /* In most cases we will be able to lock the mptcp socket. If its already
774 * owned, we need to defer to the work queue to avoid ABBA deadlock.
776 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
778 struct sock *sk = (struct sock *)msk;
779 unsigned int moved = 0;
781 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
782 __mptcp_ofo_queue(msk);
783 if (unlikely(ssk->sk_err)) {
784 if (!sock_owned_by_user(sk))
785 __mptcp_error_report(sk);
787 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
790 /* If the moves have caught up with the DATA_FIN sequence number
791 * it's time to ack the DATA_FIN and change socket state, but
792 * this is not a good place to change state. Let the workqueue
795 if (mptcp_pending_data_fin(sk, NULL))
796 mptcp_schedule_work(sk);
800 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
802 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
803 struct mptcp_sock *msk = mptcp_sk(sk);
804 int sk_rbuf, ssk_rbuf;
806 /* The peer can send data while we are shutting down this
807 * subflow at msk destruction time, but we must avoid enqueuing
808 * more data to the msk receive queue
810 if (unlikely(subflow->disposable))
813 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
814 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
815 if (unlikely(ssk_rbuf > sk_rbuf))
818 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
819 if (__mptcp_rmem(sk) > sk_rbuf) {
820 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
824 /* Wake-up the reader only for in-sequence data */
826 if (move_skbs_to_msk(msk, ssk))
827 sk->sk_data_ready(sk);
829 mptcp_data_unlock(sk);
832 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
834 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
835 WRITE_ONCE(msk->allow_infinite_fallback, false);
836 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
839 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
841 struct sock *sk = (struct sock *)msk;
843 if (sk->sk_state != TCP_ESTABLISHED)
846 /* attach to msk socket only after we are sure we will deal with it
849 if (sk->sk_socket && !ssk->sk_socket)
850 mptcp_sock_graft(ssk, sk->sk_socket);
852 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
853 mptcp_sockopt_sync_locked(msk, ssk);
854 mptcp_subflow_joined(msk, ssk);
858 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
860 struct mptcp_subflow_context *tmp, *subflow;
861 struct mptcp_sock *msk = mptcp_sk(sk);
863 list_for_each_entry_safe(subflow, tmp, join_list, node) {
864 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
865 bool slow = lock_sock_fast(ssk);
867 list_move_tail(&subflow->node, &msk->conn_list);
868 if (!__mptcp_finish_join(msk, ssk))
869 mptcp_subflow_reset(ssk);
870 unlock_sock_fast(ssk, slow);
874 static bool mptcp_timer_pending(struct sock *sk)
876 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
879 static void mptcp_reset_timer(struct sock *sk)
881 struct inet_connection_sock *icsk = inet_csk(sk);
884 /* prevent rescheduling on close */
885 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
888 tout = mptcp_sk(sk)->timer_ival;
889 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
892 bool mptcp_schedule_work(struct sock *sk)
894 if (inet_sk_state_load(sk) != TCP_CLOSE &&
895 schedule_work(&mptcp_sk(sk)->work)) {
896 /* each subflow already holds a reference to the sk, and the
897 * workqueue is invoked by a subflow, so sk can't go away here.
905 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
907 struct mptcp_subflow_context *subflow;
909 msk_owned_by_me(msk);
911 mptcp_for_each_subflow(msk, subflow) {
912 if (READ_ONCE(subflow->data_avail))
913 return mptcp_subflow_tcp_sock(subflow);
919 static bool mptcp_skb_can_collapse_to(u64 write_seq,
920 const struct sk_buff *skb,
921 const struct mptcp_ext *mpext)
923 if (!tcp_skb_can_collapse_to(skb))
926 /* can collapse only if MPTCP level sequence is in order and this
927 * mapping has not been xmitted yet
929 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
933 /* we can append data to the given data frag if:
934 * - there is space available in the backing page_frag
935 * - the data frag tail matches the current page_frag free offset
936 * - the data frag end sequence number matches the current write seq
938 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
939 const struct page_frag *pfrag,
940 const struct mptcp_data_frag *df)
942 return df && pfrag->page == df->page &&
943 pfrag->size - pfrag->offset > 0 &&
944 pfrag->offset == (df->offset + df->data_len) &&
945 df->data_seq + df->data_len == msk->write_seq;
948 static void dfrag_uncharge(struct sock *sk, int len)
950 sk_mem_uncharge(sk, len);
951 sk_wmem_queued_add(sk, -len);
954 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
956 int len = dfrag->data_len + dfrag->overhead;
958 list_del(&dfrag->list);
959 dfrag_uncharge(sk, len);
960 put_page(dfrag->page);
963 static void __mptcp_clean_una(struct sock *sk)
965 struct mptcp_sock *msk = mptcp_sk(sk);
966 struct mptcp_data_frag *dtmp, *dfrag;
969 snd_una = msk->snd_una;
970 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
971 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
974 if (unlikely(dfrag == msk->first_pending)) {
975 /* in recovery mode can see ack after the current snd head */
976 if (WARN_ON_ONCE(!msk->recovery))
979 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
982 dfrag_clear(sk, dfrag);
985 dfrag = mptcp_rtx_head(sk);
986 if (dfrag && after64(snd_una, dfrag->data_seq)) {
987 u64 delta = snd_una - dfrag->data_seq;
989 /* prevent wrap around in recovery mode */
990 if (unlikely(delta > dfrag->already_sent)) {
991 if (WARN_ON_ONCE(!msk->recovery))
993 if (WARN_ON_ONCE(delta > dfrag->data_len))
995 dfrag->already_sent += delta - dfrag->already_sent;
998 dfrag->data_seq += delta;
999 dfrag->offset += delta;
1000 dfrag->data_len -= delta;
1001 dfrag->already_sent -= delta;
1003 dfrag_uncharge(sk, delta);
1006 /* all retransmitted data acked, recovery completed */
1007 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1008 msk->recovery = false;
1011 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1012 snd_una == READ_ONCE(msk->write_seq)) {
1013 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1014 mptcp_stop_timer(sk);
1016 mptcp_reset_timer(sk);
1020 static void __mptcp_clean_una_wakeup(struct sock *sk)
1022 lockdep_assert_held_once(&sk->sk_lock.slock);
1024 __mptcp_clean_una(sk);
1025 mptcp_write_space(sk);
1028 static void mptcp_clean_una_wakeup(struct sock *sk)
1030 mptcp_data_lock(sk);
1031 __mptcp_clean_una_wakeup(sk);
1032 mptcp_data_unlock(sk);
1035 static void mptcp_enter_memory_pressure(struct sock *sk)
1037 struct mptcp_subflow_context *subflow;
1038 struct mptcp_sock *msk = mptcp_sk(sk);
1041 sk_stream_moderate_sndbuf(sk);
1042 mptcp_for_each_subflow(msk, subflow) {
1043 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1046 tcp_enter_memory_pressure(ssk);
1047 sk_stream_moderate_sndbuf(ssk);
1052 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1055 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1057 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1058 pfrag, sk->sk_allocation)))
1061 mptcp_enter_memory_pressure(sk);
1065 static struct mptcp_data_frag *
1066 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1069 int offset = ALIGN(orig_offset, sizeof(long));
1070 struct mptcp_data_frag *dfrag;
1072 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1073 dfrag->data_len = 0;
1074 dfrag->data_seq = msk->write_seq;
1075 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1076 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1077 dfrag->already_sent = 0;
1078 dfrag->page = pfrag->page;
1083 struct mptcp_sendmsg_info {
1089 bool data_lock_held;
1092 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1093 u64 data_seq, int avail_size)
1095 u64 window_end = mptcp_wnd_end(msk);
1098 if (__mptcp_check_fallback(msk))
1101 mptcp_snd_wnd = window_end - data_seq;
1102 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1104 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1105 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1106 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1112 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1114 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1118 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1122 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1124 struct sk_buff *skb;
1126 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1128 if (likely(__mptcp_add_ext(skb, gfp))) {
1129 skb_reserve(skb, MAX_TCP_HEADER);
1130 skb->ip_summed = CHECKSUM_PARTIAL;
1131 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1136 mptcp_enter_memory_pressure(sk);
1141 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1143 struct sk_buff *skb;
1145 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1149 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1150 tcp_skb_entail(ssk, skb);
1153 tcp_skb_tsorted_anchor_cleanup(skb);
1158 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1160 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1162 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1165 /* note: this always recompute the csum on the whole skb, even
1166 * if we just appended a single frag. More status info needed
1168 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1170 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1171 __wsum csum = ~csum_unfold(mpext->csum);
1172 int offset = skb->len - added;
1174 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1177 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1179 struct mptcp_ext *mpext)
1184 mpext->infinite_map = 1;
1185 mpext->data_len = 0;
1187 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1188 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1190 mptcp_do_fallback(ssk);
1193 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1194 struct mptcp_data_frag *dfrag,
1195 struct mptcp_sendmsg_info *info)
1197 u64 data_seq = dfrag->data_seq + info->sent;
1198 int offset = dfrag->offset + info->sent;
1199 struct mptcp_sock *msk = mptcp_sk(sk);
1200 bool zero_window_probe = false;
1201 struct mptcp_ext *mpext = NULL;
1202 bool can_coalesce = false;
1203 bool reuse_skb = true;
1204 struct sk_buff *skb;
1208 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1209 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1211 if (WARN_ON_ONCE(info->sent > info->limit ||
1212 info->limit > dfrag->data_len))
1215 if (unlikely(!__tcp_can_send(ssk)))
1218 /* compute send limit */
1219 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1220 copy = info->size_goal;
1222 skb = tcp_write_queue_tail(ssk);
1223 if (skb && copy > skb->len) {
1224 /* Limit the write to the size available in the
1225 * current skb, if any, so that we create at most a new skb.
1226 * Explicitly tells TCP internals to avoid collapsing on later
1227 * queue management operation, to avoid breaking the ext <->
1228 * SSN association set here
1230 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1231 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1232 TCP_SKB_CB(skb)->eor = 1;
1236 i = skb_shinfo(skb)->nr_frags;
1237 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1238 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1239 tcp_mark_push(tcp_sk(ssk), skb);
1246 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1250 i = skb_shinfo(skb)->nr_frags;
1252 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1255 /* Zero window and all data acked? Probe. */
1256 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1258 u64 snd_una = READ_ONCE(msk->snd_una);
1260 if (snd_una != msk->snd_nxt) {
1261 tcp_remove_empty_skb(ssk);
1265 zero_window_probe = true;
1266 data_seq = snd_una - 1;
1269 /* all mptcp-level data is acked, no skbs should be present into the
1272 WARN_ON_ONCE(reuse_skb);
1275 copy = min_t(size_t, copy, info->limit - info->sent);
1276 if (!sk_wmem_schedule(ssk, copy)) {
1277 tcp_remove_empty_skb(ssk);
1282 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1284 get_page(dfrag->page);
1285 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1289 skb->data_len += copy;
1290 skb->truesize += copy;
1291 sk_wmem_queued_add(ssk, copy);
1292 sk_mem_charge(ssk, copy);
1293 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1294 TCP_SKB_CB(skb)->end_seq += copy;
1295 tcp_skb_pcount_set(skb, 0);
1297 /* on skb reuse we just need to update the DSS len */
1299 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1300 mpext->data_len += copy;
1301 WARN_ON_ONCE(zero_window_probe);
1305 memset(mpext, 0, sizeof(*mpext));
1306 mpext->data_seq = data_seq;
1307 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1308 mpext->data_len = copy;
1312 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1313 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1316 if (zero_window_probe) {
1317 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1319 if (READ_ONCE(msk->csum_enabled))
1320 mptcp_update_data_checksum(skb, copy);
1321 tcp_push_pending_frames(ssk);
1325 if (READ_ONCE(msk->csum_enabled))
1326 mptcp_update_data_checksum(skb, copy);
1327 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1328 mptcp_update_infinite_map(msk, ssk, mpext);
1329 trace_mptcp_sendmsg_frag(mpext);
1330 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1334 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1335 sizeof(struct tcphdr) - \
1336 MAX_TCP_OPTION_SPACE - \
1337 sizeof(struct ipv6hdr) - \
1338 sizeof(struct frag_hdr))
1340 struct subflow_send_info {
1345 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1347 if (!subflow->stale)
1351 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1354 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1356 if (unlikely(subflow->stale)) {
1357 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1359 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1362 mptcp_subflow_set_active(subflow);
1364 return __mptcp_subflow_active(subflow);
1367 #define SSK_MODE_ACTIVE 0
1368 #define SSK_MODE_BACKUP 1
1369 #define SSK_MODE_MAX 2
1371 /* implement the mptcp packet scheduler;
1372 * returns the subflow that will transmit the next DSS
1373 * additionally updates the rtx timeout
1375 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1377 struct subflow_send_info send_info[SSK_MODE_MAX];
1378 struct mptcp_subflow_context *subflow;
1379 struct sock *sk = (struct sock *)msk;
1380 u32 pace, burst, wmem;
1381 int i, nr_active = 0;
1386 /* pick the subflow with the lower wmem/wspace ratio */
1387 for (i = 0; i < SSK_MODE_MAX; ++i) {
1388 send_info[i].ssk = NULL;
1389 send_info[i].linger_time = -1;
1392 mptcp_for_each_subflow(msk, subflow) {
1393 trace_mptcp_subflow_get_send(subflow);
1394 ssk = mptcp_subflow_tcp_sock(subflow);
1395 if (!mptcp_subflow_active(subflow))
1398 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1399 nr_active += !subflow->backup;
1400 pace = subflow->avg_pacing_rate;
1401 if (unlikely(!pace)) {
1402 /* init pacing rate from socket */
1403 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1404 pace = subflow->avg_pacing_rate;
1409 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1410 if (linger_time < send_info[subflow->backup].linger_time) {
1411 send_info[subflow->backup].ssk = ssk;
1412 send_info[subflow->backup].linger_time = linger_time;
1415 __mptcp_set_timeout(sk, tout);
1417 /* pick the best backup if no other subflow is active */
1419 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1421 /* According to the blest algorithm, to avoid HoL blocking for the
1422 * faster flow, we need to:
1423 * - estimate the faster flow linger time
1424 * - use the above to estimate the amount of byte transferred
1425 * by the faster flow
1426 * - check that the amount of queued data is greter than the above,
1427 * otherwise do not use the picked, slower, subflow
1428 * We select the subflow with the shorter estimated time to flush
1429 * the queued mem, which basically ensure the above. We just need
1430 * to check that subflow has a non empty cwin.
1432 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1433 if (!ssk || !sk_stream_memory_free(ssk))
1436 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1437 wmem = READ_ONCE(ssk->sk_wmem_queued);
1441 subflow = mptcp_subflow_ctx(ssk);
1442 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1443 READ_ONCE(ssk->sk_pacing_rate) * burst,
1445 msk->snd_burst = burst;
1449 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1451 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1455 static void mptcp_update_post_push(struct mptcp_sock *msk,
1456 struct mptcp_data_frag *dfrag,
1459 u64 snd_nxt_new = dfrag->data_seq;
1461 dfrag->already_sent += sent;
1463 msk->snd_burst -= sent;
1465 snd_nxt_new += dfrag->already_sent;
1467 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1468 * is recovering after a failover. In that event, this re-sends
1471 * Thus compute snd_nxt_new candidate based on
1472 * the dfrag->data_seq that was sent and the data
1473 * that has been handed to the subflow for transmission
1474 * and skip update in case it was old dfrag.
1476 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1477 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1478 msk->snd_nxt = snd_nxt_new;
1482 void mptcp_check_and_set_pending(struct sock *sk)
1484 if (mptcp_send_head(sk))
1485 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1488 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1489 struct mptcp_sendmsg_info *info)
1491 struct mptcp_sock *msk = mptcp_sk(sk);
1492 struct mptcp_data_frag *dfrag;
1493 int len, copied = 0, err = 0;
1495 while ((dfrag = mptcp_send_head(sk))) {
1496 info->sent = dfrag->already_sent;
1497 info->limit = dfrag->data_len;
1498 len = dfrag->data_len - dfrag->already_sent;
1502 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1504 err = copied ? : ret;
1512 mptcp_update_post_push(msk, dfrag, ret);
1514 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1516 if (msk->snd_burst <= 0 ||
1517 !sk_stream_memory_free(ssk) ||
1518 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1522 mptcp_set_timeout(sk);
1530 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1532 struct sock *prev_ssk = NULL, *ssk = NULL;
1533 struct mptcp_sock *msk = mptcp_sk(sk);
1534 struct mptcp_sendmsg_info info = {
1537 bool do_check_data_fin = false;
1540 while (mptcp_send_head(sk) && (push_count > 0)) {
1541 struct mptcp_subflow_context *subflow;
1544 if (mptcp_sched_get_send(msk))
1549 mptcp_for_each_subflow(msk, subflow) {
1550 if (READ_ONCE(subflow->scheduled)) {
1551 mptcp_subflow_set_scheduled(subflow, false);
1554 ssk = mptcp_subflow_tcp_sock(subflow);
1555 if (ssk != prev_ssk) {
1556 /* First check. If the ssk has changed since
1557 * the last round, release prev_ssk
1560 mptcp_push_release(prev_ssk, &info);
1562 /* Need to lock the new subflow only if different
1563 * from the previous one, otherwise we are still
1564 * helding the relevant lock
1571 ret = __subflow_push_pending(sk, ssk, &info);
1573 if (ret != -EAGAIN ||
1574 (1 << ssk->sk_state) &
1575 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1579 do_check_data_fin = true;
1584 /* at this point we held the socket lock for the last subflow we used */
1586 mptcp_push_release(ssk, &info);
1588 /* ensure the rtx timer is running */
1589 if (!mptcp_timer_pending(sk))
1590 mptcp_reset_timer(sk);
1591 if (do_check_data_fin)
1592 mptcp_check_send_data_fin(sk);
1595 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1597 struct mptcp_sock *msk = mptcp_sk(sk);
1598 struct mptcp_sendmsg_info info = {
1599 .data_lock_held = true,
1601 bool keep_pushing = true;
1602 struct sock *xmit_ssk;
1606 while (mptcp_send_head(sk) && keep_pushing) {
1607 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1610 /* check for a different subflow usage only after
1611 * spooling the first chunk of data
1614 mptcp_subflow_set_scheduled(subflow, false);
1615 ret = __subflow_push_pending(sk, ssk, &info);
1623 if (mptcp_sched_get_send(msk))
1626 if (READ_ONCE(subflow->scheduled)) {
1627 mptcp_subflow_set_scheduled(subflow, false);
1628 ret = __subflow_push_pending(sk, ssk, &info);
1630 keep_pushing = false;
1634 mptcp_for_each_subflow(msk, subflow) {
1635 if (READ_ONCE(subflow->scheduled)) {
1636 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1637 if (xmit_ssk != ssk) {
1638 mptcp_subflow_delegate(subflow,
1639 MPTCP_DELEGATE_SEND);
1640 keep_pushing = false;
1647 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1648 * not going to flush it via release_sock()
1651 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1653 if (!mptcp_timer_pending(sk))
1654 mptcp_reset_timer(sk);
1656 if (msk->snd_data_fin_enable &&
1657 msk->snd_nxt + 1 == msk->write_seq)
1658 mptcp_schedule_work(sk);
1662 static void mptcp_set_nospace(struct sock *sk)
1664 /* enable autotune */
1665 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1667 /* will be cleared on avail space */
1668 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1671 static int mptcp_disconnect(struct sock *sk, int flags);
1673 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1674 size_t len, int *copied_syn)
1676 unsigned int saved_flags = msg->msg_flags;
1677 struct mptcp_sock *msk = mptcp_sk(sk);
1681 /* on flags based fastopen the mptcp is supposed to create the
1682 * first subflow right now. Otherwise we are in the defer_connect
1683 * path, and the first subflow must be already present.
1684 * Since the defer_connect flag is cleared after the first succsful
1685 * fastopen attempt, no need to check for additional subflow status.
1687 if (msg->msg_flags & MSG_FASTOPEN) {
1688 ssk = __mptcp_nmpc_sk(msk);
1690 return PTR_ERR(ssk);
1698 msg->msg_flags |= MSG_DONTWAIT;
1699 msk->fastopening = 1;
1700 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1701 msk->fastopening = 0;
1702 msg->msg_flags = saved_flags;
1705 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1706 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1707 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1708 msg->msg_namelen, msg->msg_flags, 1);
1710 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1711 * case of any error, except timeout or signal
1713 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1715 } else if (ret && ret != -EINPROGRESS) {
1716 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1717 * __inet_stream_connect() can fail, due to looking check,
1718 * see mptcp_disconnect().
1719 * Attempt it again outside the problematic scope.
1721 if (!mptcp_disconnect(sk, 0))
1722 sk->sk_socket->state = SS_UNCONNECTED;
1724 inet_clear_bit(DEFER_CONNECT, sk);
1729 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1731 struct mptcp_sock *msk = mptcp_sk(sk);
1732 struct page_frag *pfrag;
1737 /* silently ignore everything else */
1738 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1742 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1743 msg->msg_flags & MSG_FASTOPEN)) {
1746 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1747 copied += copied_syn;
1748 if (ret == -EINPROGRESS && copied_syn > 0)
1754 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1756 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1757 ret = sk_stream_wait_connect(sk, &timeo);
1763 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1766 pfrag = sk_page_frag(sk);
1768 while (msg_data_left(msg)) {
1769 int total_ts, frag_truesize = 0;
1770 struct mptcp_data_frag *dfrag;
1771 bool dfrag_collapsed;
1772 size_t psize, offset;
1774 /* reuse tail pfrag, if possible, or carve a new one from the
1777 dfrag = mptcp_pending_tail(sk);
1778 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1779 if (!dfrag_collapsed) {
1780 if (!sk_stream_memory_free(sk))
1781 goto wait_for_memory;
1783 if (!mptcp_page_frag_refill(sk, pfrag))
1784 goto wait_for_memory;
1786 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1787 frag_truesize = dfrag->overhead;
1790 /* we do not bound vs wspace, to allow a single packet.
1791 * memory accounting will prevent execessive memory usage
1794 offset = dfrag->offset + dfrag->data_len;
1795 psize = pfrag->size - offset;
1796 psize = min_t(size_t, psize, msg_data_left(msg));
1797 total_ts = psize + frag_truesize;
1799 if (!sk_wmem_schedule(sk, total_ts))
1800 goto wait_for_memory;
1802 if (copy_page_from_iter(dfrag->page, offset, psize,
1803 &msg->msg_iter) != psize) {
1808 /* data successfully copied into the write queue */
1809 sk_forward_alloc_add(sk, -total_ts);
1811 dfrag->data_len += psize;
1812 frag_truesize += psize;
1813 pfrag->offset += frag_truesize;
1814 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1816 /* charge data on mptcp pending queue to the msk socket
1817 * Note: we charge such data both to sk and ssk
1819 sk_wmem_queued_add(sk, frag_truesize);
1820 if (!dfrag_collapsed) {
1821 get_page(dfrag->page);
1822 list_add_tail(&dfrag->list, &msk->rtx_queue);
1823 if (!msk->first_pending)
1824 WRITE_ONCE(msk->first_pending, dfrag);
1826 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1827 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1833 mptcp_set_nospace(sk);
1834 __mptcp_push_pending(sk, msg->msg_flags);
1835 ret = sk_stream_wait_memory(sk, &timeo);
1841 __mptcp_push_pending(sk, msg->msg_flags);
1851 copied = sk_stream_error(sk, msg->msg_flags, ret);
1855 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1857 size_t len, int flags,
1858 struct scm_timestamping_internal *tss,
1861 struct sk_buff *skb, *tmp;
1864 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1865 u32 offset = MPTCP_SKB_CB(skb)->offset;
1866 u32 data_len = skb->len - offset;
1867 u32 count = min_t(size_t, len - copied, data_len);
1870 if (!(flags & MSG_TRUNC)) {
1871 err = skb_copy_datagram_msg(skb, offset, msg, count);
1872 if (unlikely(err < 0)) {
1879 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1880 tcp_update_recv_tstamps(skb, tss);
1881 *cmsg_flags |= MPTCP_CMSG_TS;
1886 if (count < data_len) {
1887 if (!(flags & MSG_PEEK)) {
1888 MPTCP_SKB_CB(skb)->offset += count;
1889 MPTCP_SKB_CB(skb)->map_seq += count;
1894 if (!(flags & MSG_PEEK)) {
1895 /* we will bulk release the skb memory later */
1896 skb->destructor = NULL;
1897 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1898 __skb_unlink(skb, &msk->receive_queue);
1909 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1911 * Only difference: Use highest rtt estimate of the subflows in use.
1913 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1915 struct mptcp_subflow_context *subflow;
1916 struct sock *sk = (struct sock *)msk;
1917 u8 scaling_ratio = U8_MAX;
1918 u32 time, advmss = 1;
1921 msk_owned_by_me(msk);
1926 msk->rcvq_space.copied += copied;
1928 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1929 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1931 rtt_us = msk->rcvq_space.rtt_us;
1932 if (rtt_us && time < (rtt_us >> 3))
1936 mptcp_for_each_subflow(msk, subflow) {
1937 const struct tcp_sock *tp;
1941 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1943 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1944 sf_advmss = READ_ONCE(tp->advmss);
1946 rtt_us = max(sf_rtt_us, rtt_us);
1947 advmss = max(sf_advmss, advmss);
1948 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1951 msk->rcvq_space.rtt_us = rtt_us;
1952 msk->scaling_ratio = scaling_ratio;
1953 if (time < (rtt_us >> 3) || rtt_us == 0)
1956 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1959 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1960 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1964 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1966 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1968 do_div(grow, msk->rcvq_space.space);
1969 rcvwin += (grow << 1);
1971 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
1972 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1974 if (rcvbuf > sk->sk_rcvbuf) {
1977 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
1978 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1980 /* Make subflows follow along. If we do not do this, we
1981 * get drops at subflow level if skbs can't be moved to
1982 * the mptcp rx queue fast enough (announced rcv_win can
1983 * exceed ssk->sk_rcvbuf).
1985 mptcp_for_each_subflow(msk, subflow) {
1989 ssk = mptcp_subflow_tcp_sock(subflow);
1990 slow = lock_sock_fast(ssk);
1991 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1992 tcp_sk(ssk)->window_clamp = window_clamp;
1993 tcp_cleanup_rbuf(ssk, 1);
1994 unlock_sock_fast(ssk, slow);
1999 msk->rcvq_space.space = msk->rcvq_space.copied;
2001 msk->rcvq_space.copied = 0;
2002 msk->rcvq_space.time = mstamp;
2005 static void __mptcp_update_rmem(struct sock *sk)
2007 struct mptcp_sock *msk = mptcp_sk(sk);
2009 if (!msk->rmem_released)
2012 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2013 mptcp_rmem_uncharge(sk, msk->rmem_released);
2014 WRITE_ONCE(msk->rmem_released, 0);
2017 static void __mptcp_splice_receive_queue(struct sock *sk)
2019 struct mptcp_sock *msk = mptcp_sk(sk);
2021 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2024 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2026 struct sock *sk = (struct sock *)msk;
2027 unsigned int moved = 0;
2031 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2034 /* we can have data pending in the subflows only if the msk
2035 * receive buffer was full at subflow_data_ready() time,
2036 * that is an unlikely slow path.
2041 slowpath = lock_sock_fast(ssk);
2042 mptcp_data_lock(sk);
2043 __mptcp_update_rmem(sk);
2044 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2045 mptcp_data_unlock(sk);
2047 if (unlikely(ssk->sk_err))
2048 __mptcp_error_report(sk);
2049 unlock_sock_fast(ssk, slowpath);
2052 /* acquire the data lock only if some input data is pending */
2054 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2055 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2056 mptcp_data_lock(sk);
2057 __mptcp_update_rmem(sk);
2058 ret |= __mptcp_ofo_queue(msk);
2059 __mptcp_splice_receive_queue(sk);
2060 mptcp_data_unlock(sk);
2063 mptcp_check_data_fin((struct sock *)msk);
2064 return !skb_queue_empty(&msk->receive_queue);
2067 static unsigned int mptcp_inq_hint(const struct sock *sk)
2069 const struct mptcp_sock *msk = mptcp_sk(sk);
2070 const struct sk_buff *skb;
2072 skb = skb_peek(&msk->receive_queue);
2074 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2076 if (hint_val >= INT_MAX)
2079 return (unsigned int)hint_val;
2082 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2088 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2089 int flags, int *addr_len)
2091 struct mptcp_sock *msk = mptcp_sk(sk);
2092 struct scm_timestamping_internal tss;
2093 int copied = 0, cmsg_flags = 0;
2097 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2098 if (unlikely(flags & MSG_ERRQUEUE))
2099 return inet_recv_error(sk, msg, len, addr_len);
2102 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2107 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2109 len = min_t(size_t, len, INT_MAX);
2110 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2112 if (unlikely(msk->recvmsg_inq))
2113 cmsg_flags = MPTCP_CMSG_INQ;
2115 while (copied < len) {
2118 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2119 if (unlikely(bytes_read < 0)) {
2121 copied = bytes_read;
2125 copied += bytes_read;
2127 /* be sure to advertise window change */
2128 mptcp_cleanup_rbuf(msk);
2130 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2133 /* only the master socket status is relevant here. The exit
2134 * conditions mirror closely tcp_recvmsg()
2136 if (copied >= target)
2141 sk->sk_state == TCP_CLOSE ||
2142 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2144 signal_pending(current))
2148 copied = sock_error(sk);
2152 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2153 /* race breaker: the shutdown could be after the
2154 * previous receive queue check
2156 if (__mptcp_move_skbs(msk))
2161 if (sk->sk_state == TCP_CLOSE) {
2171 if (signal_pending(current)) {
2172 copied = sock_intr_errno(timeo);
2177 pr_debug("block timeout %ld", timeo);
2178 sk_wait_data(sk, &timeo, NULL);
2182 if (cmsg_flags && copied >= 0) {
2183 if (cmsg_flags & MPTCP_CMSG_TS)
2184 tcp_recv_timestamp(msg, sk, &tss);
2186 if (cmsg_flags & MPTCP_CMSG_INQ) {
2187 unsigned int inq = mptcp_inq_hint(sk);
2189 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2193 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2194 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2195 skb_queue_empty(&msk->receive_queue), copied);
2196 if (!(flags & MSG_PEEK))
2197 mptcp_rcv_space_adjust(msk, copied);
2203 static void mptcp_retransmit_timer(struct timer_list *t)
2205 struct inet_connection_sock *icsk = from_timer(icsk, t,
2206 icsk_retransmit_timer);
2207 struct sock *sk = &icsk->icsk_inet.sk;
2208 struct mptcp_sock *msk = mptcp_sk(sk);
2211 if (!sock_owned_by_user(sk)) {
2212 /* we need a process context to retransmit */
2213 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2214 mptcp_schedule_work(sk);
2216 /* delegate our work to tcp_release_cb() */
2217 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2223 static void mptcp_timeout_timer(struct timer_list *t)
2225 struct sock *sk = from_timer(sk, t, sk_timer);
2227 mptcp_schedule_work(sk);
2231 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2234 * A backup subflow is returned only if that is the only kind available.
2236 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2238 struct sock *backup = NULL, *pick = NULL;
2239 struct mptcp_subflow_context *subflow;
2240 int min_stale_count = INT_MAX;
2242 mptcp_for_each_subflow(msk, subflow) {
2243 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2245 if (!__mptcp_subflow_active(subflow))
2248 /* still data outstanding at TCP level? skip this */
2249 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2250 mptcp_pm_subflow_chk_stale(msk, ssk);
2251 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2255 if (subflow->backup) {
2268 /* use backup only if there are no progresses anywhere */
2269 return min_stale_count > 1 ? backup : NULL;
2272 bool __mptcp_retransmit_pending_data(struct sock *sk)
2274 struct mptcp_data_frag *cur, *rtx_head;
2275 struct mptcp_sock *msk = mptcp_sk(sk);
2277 if (__mptcp_check_fallback(msk))
2280 if (tcp_rtx_and_write_queues_empty(sk))
2283 /* the closing socket has some data untransmitted and/or unacked:
2284 * some data in the mptcp rtx queue has not really xmitted yet.
2285 * keep it simple and re-inject the whole mptcp level rtx queue
2287 mptcp_data_lock(sk);
2288 __mptcp_clean_una_wakeup(sk);
2289 rtx_head = mptcp_rtx_head(sk);
2291 mptcp_data_unlock(sk);
2295 msk->recovery_snd_nxt = msk->snd_nxt;
2296 msk->recovery = true;
2297 mptcp_data_unlock(sk);
2299 msk->first_pending = rtx_head;
2302 /* be sure to clear the "sent status" on all re-injected fragments */
2303 list_for_each_entry(cur, &msk->rtx_queue, list) {
2304 if (!cur->already_sent)
2306 cur->already_sent = 0;
2312 /* flags for __mptcp_close_ssk() */
2313 #define MPTCP_CF_PUSH BIT(1)
2314 #define MPTCP_CF_FASTCLOSE BIT(2)
2316 /* subflow sockets can be either outgoing (connect) or incoming
2319 * Outgoing subflows use in-kernel sockets.
2320 * Incoming subflows do not have their own 'struct socket' allocated,
2321 * so we need to use tcp_close() after detaching them from the mptcp
2324 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2325 struct mptcp_subflow_context *subflow,
2328 struct mptcp_sock *msk = mptcp_sk(sk);
2329 bool dispose_it, need_push = false;
2331 /* If the first subflow moved to a close state before accept, e.g. due
2332 * to an incoming reset, mptcp either:
2333 * - if either the subflow or the msk are dead, destroy the context
2334 * (the subflow socket is deleted by inet_child_forget) and the msk
2335 * - otherwise do nothing at the moment and take action at accept and/or
2336 * listener shutdown - user-space must be able to accept() the closed
2339 if (msk->in_accept_queue && msk->first == ssk) {
2340 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2343 /* ensure later check in mptcp_worker() will dispose the msk */
2344 sock_set_flag(sk, SOCK_DEAD);
2345 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2346 mptcp_subflow_drop_ctx(ssk);
2350 dispose_it = msk->free_first || ssk != msk->first;
2352 list_del(&subflow->node);
2354 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2356 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2357 /* be sure to force the tcp_disconnect() path,
2358 * to generate the egress reset
2360 ssk->sk_lingertime = 0;
2361 sock_set_flag(ssk, SOCK_LINGER);
2362 subflow->send_fastclose = 1;
2365 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2367 /* The MPTCP code never wait on the subflow sockets, TCP-level
2368 * disconnect should never fail
2370 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2371 mptcp_subflow_ctx_reset(subflow);
2377 subflow->disposable = 1;
2379 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2380 * the ssk has been already destroyed, we just need to release the
2381 * reference owned by msk;
2383 if (!inet_csk(ssk)->icsk_ulp_ops) {
2384 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2385 kfree_rcu(subflow, rcu);
2387 /* otherwise tcp will dispose of the ssk and subflow ctx */
2388 __tcp_close(ssk, 0);
2390 /* close acquired an extra ref */
2399 if (ssk == msk->first)
2400 WRITE_ONCE(msk->first, NULL);
2404 __mptcp_push_pending(sk, 0);
2407 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2408 struct mptcp_subflow_context *subflow)
2410 if (sk->sk_state == TCP_ESTABLISHED)
2411 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2413 /* subflow aborted before reaching the fully_established status
2414 * attempt the creation of the next subflow
2416 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2418 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2421 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2426 static void __mptcp_close_subflow(struct sock *sk)
2428 struct mptcp_subflow_context *subflow, *tmp;
2429 struct mptcp_sock *msk = mptcp_sk(sk);
2433 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2434 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2436 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2439 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2440 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2443 mptcp_close_ssk(sk, ssk, subflow);
2448 static bool mptcp_should_close(const struct sock *sk)
2450 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2451 struct mptcp_subflow_context *subflow;
2453 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2456 /* if all subflows are in closed status don't bother with additional
2459 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2460 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2467 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2469 struct mptcp_subflow_context *subflow, *tmp;
2470 struct sock *sk = (struct sock *)msk;
2472 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2475 mptcp_token_destroy(msk);
2477 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2478 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2481 slow = lock_sock_fast(tcp_sk);
2482 if (tcp_sk->sk_state != TCP_CLOSE) {
2483 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2484 tcp_set_state(tcp_sk, TCP_CLOSE);
2486 unlock_sock_fast(tcp_sk, slow);
2489 /* Mirror the tcp_reset() error propagation */
2490 switch (sk->sk_state) {
2492 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2494 case TCP_CLOSE_WAIT:
2495 WRITE_ONCE(sk->sk_err, EPIPE);
2500 WRITE_ONCE(sk->sk_err, ECONNRESET);
2503 inet_sk_state_store(sk, TCP_CLOSE);
2504 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2505 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2506 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2508 /* the calling mptcp_worker will properly destroy the socket */
2509 if (sock_flag(sk, SOCK_DEAD))
2512 sk->sk_state_change(sk);
2513 sk_error_report(sk);
2516 static void __mptcp_retrans(struct sock *sk)
2518 struct mptcp_sock *msk = mptcp_sk(sk);
2519 struct mptcp_subflow_context *subflow;
2520 struct mptcp_sendmsg_info info = {};
2521 struct mptcp_data_frag *dfrag;
2526 mptcp_clean_una_wakeup(sk);
2528 /* first check ssk: need to kick "stale" logic */
2529 err = mptcp_sched_get_retrans(msk);
2530 dfrag = mptcp_rtx_head(sk);
2532 if (mptcp_data_fin_enabled(msk)) {
2533 struct inet_connection_sock *icsk = inet_csk(sk);
2535 icsk->icsk_retransmits++;
2536 mptcp_set_datafin_timeout(sk);
2537 mptcp_send_ack(msk);
2542 if (!mptcp_send_head(sk))
2551 mptcp_for_each_subflow(msk, subflow) {
2552 if (READ_ONCE(subflow->scheduled)) {
2555 mptcp_subflow_set_scheduled(subflow, false);
2557 ssk = mptcp_subflow_tcp_sock(subflow);
2561 /* limit retransmission to the bytes already sent on some subflows */
2563 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2564 dfrag->already_sent;
2565 while (info.sent < info.limit) {
2566 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2570 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2575 len = max(copied, len);
2576 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2578 WRITE_ONCE(msk->allow_infinite_fallback, false);
2585 msk->bytes_retrans += len;
2586 dfrag->already_sent = max(dfrag->already_sent, len);
2589 mptcp_check_and_set_pending(sk);
2591 if (!mptcp_timer_pending(sk))
2592 mptcp_reset_timer(sk);
2595 /* schedule the timeout timer for the relevant event: either close timeout
2596 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2598 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2600 struct sock *sk = (struct sock *)msk;
2601 unsigned long timeout, close_timeout;
2603 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2606 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2608 /* the close timeout takes precedence on the fail one, and here at least one of
2611 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2613 sk_reset_timer(sk, &sk->sk_timer, timeout);
2616 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2618 struct sock *ssk = msk->first;
2624 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2626 slow = lock_sock_fast(ssk);
2627 mptcp_subflow_reset(ssk);
2628 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2629 unlock_sock_fast(ssk, slow);
2631 mptcp_reset_timeout(msk, 0);
2634 static void mptcp_do_fastclose(struct sock *sk)
2636 struct mptcp_subflow_context *subflow, *tmp;
2637 struct mptcp_sock *msk = mptcp_sk(sk);
2639 inet_sk_state_store(sk, TCP_CLOSE);
2640 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2641 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2642 subflow, MPTCP_CF_FASTCLOSE);
2645 static void mptcp_worker(struct work_struct *work)
2647 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2648 struct sock *sk = (struct sock *)msk;
2649 unsigned long fail_tout;
2653 state = sk->sk_state;
2654 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2657 mptcp_check_fastclose(msk);
2659 mptcp_pm_nl_work(msk);
2661 mptcp_check_send_data_fin(sk);
2662 mptcp_check_data_fin_ack(sk);
2663 mptcp_check_data_fin(sk);
2665 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2666 __mptcp_close_subflow(sk);
2668 /* There is no point in keeping around an orphaned sk timedout or
2669 * closed, but we need the msk around to reply to incoming DATA_FIN,
2670 * even if it is orphaned and in FIN_WAIT2 state
2672 if (sock_flag(sk, SOCK_DEAD)) {
2673 if (mptcp_should_close(sk))
2674 mptcp_do_fastclose(sk);
2676 if (sk->sk_state == TCP_CLOSE) {
2677 __mptcp_destroy_sock(sk);
2682 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2683 __mptcp_retrans(sk);
2685 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2686 if (fail_tout && time_after(jiffies, fail_tout))
2687 mptcp_mp_fail_no_response(msk);
2694 static void __mptcp_init_sock(struct sock *sk)
2696 struct mptcp_sock *msk = mptcp_sk(sk);
2698 INIT_LIST_HEAD(&msk->conn_list);
2699 INIT_LIST_HEAD(&msk->join_list);
2700 INIT_LIST_HEAD(&msk->rtx_queue);
2701 INIT_WORK(&msk->work, mptcp_worker);
2702 __skb_queue_head_init(&msk->receive_queue);
2703 msk->out_of_order_queue = RB_ROOT;
2704 msk->first_pending = NULL;
2705 msk->rmem_fwd_alloc = 0;
2706 WRITE_ONCE(msk->rmem_released, 0);
2707 msk->timer_ival = TCP_RTO_MIN;
2709 WRITE_ONCE(msk->first, NULL);
2710 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2711 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2712 WRITE_ONCE(msk->allow_infinite_fallback, true);
2713 msk->recovery = false;
2714 msk->subflow_id = 1;
2716 mptcp_pm_data_init(msk);
2718 /* re-use the csk retrans timer for MPTCP-level retrans */
2719 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2720 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2723 static void mptcp_ca_reset(struct sock *sk)
2725 struct inet_connection_sock *icsk = inet_csk(sk);
2727 tcp_assign_congestion_control(sk);
2728 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2730 /* no need to keep a reference to the ops, the name will suffice */
2731 tcp_cleanup_congestion_control(sk);
2732 icsk->icsk_ca_ops = NULL;
2735 static int mptcp_init_sock(struct sock *sk)
2737 struct net *net = sock_net(sk);
2740 __mptcp_init_sock(sk);
2742 if (!mptcp_is_enabled(net))
2743 return -ENOPROTOOPT;
2745 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2748 ret = mptcp_init_sched(mptcp_sk(sk),
2749 mptcp_sched_find(mptcp_get_scheduler(net)));
2753 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2755 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2756 * propagate the correct value
2760 sk_sockets_allocated_inc(sk);
2761 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2762 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2767 static void __mptcp_clear_xmit(struct sock *sk)
2769 struct mptcp_sock *msk = mptcp_sk(sk);
2770 struct mptcp_data_frag *dtmp, *dfrag;
2772 WRITE_ONCE(msk->first_pending, NULL);
2773 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2774 dfrag_clear(sk, dfrag);
2777 void mptcp_cancel_work(struct sock *sk)
2779 struct mptcp_sock *msk = mptcp_sk(sk);
2781 if (cancel_work_sync(&msk->work))
2785 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2789 switch (ssk->sk_state) {
2791 if (!(how & RCV_SHUTDOWN))
2795 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2798 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2799 pr_debug("Fallback");
2800 ssk->sk_shutdown |= how;
2801 tcp_shutdown(ssk, how);
2803 /* simulate the data_fin ack reception to let the state
2804 * machine move forward
2806 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2807 mptcp_schedule_work(sk);
2809 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2811 if (!mptcp_timer_pending(sk))
2812 mptcp_reset_timer(sk);
2820 static const unsigned char new_state[16] = {
2821 /* current state: new state: action: */
2822 [0 /* (Invalid) */] = TCP_CLOSE,
2823 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2824 [TCP_SYN_SENT] = TCP_CLOSE,
2825 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2826 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2827 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2828 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2829 [TCP_CLOSE] = TCP_CLOSE,
2830 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2831 [TCP_LAST_ACK] = TCP_LAST_ACK,
2832 [TCP_LISTEN] = TCP_CLOSE,
2833 [TCP_CLOSING] = TCP_CLOSING,
2834 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2837 static int mptcp_close_state(struct sock *sk)
2839 int next = (int)new_state[sk->sk_state];
2840 int ns = next & TCP_STATE_MASK;
2842 inet_sk_state_store(sk, ns);
2844 return next & TCP_ACTION_FIN;
2847 static void mptcp_check_send_data_fin(struct sock *sk)
2849 struct mptcp_subflow_context *subflow;
2850 struct mptcp_sock *msk = mptcp_sk(sk);
2852 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2853 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2854 msk->snd_nxt, msk->write_seq);
2856 /* we still need to enqueue subflows or not really shutting down,
2859 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2860 mptcp_send_head(sk))
2863 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2865 mptcp_for_each_subflow(msk, subflow) {
2866 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2868 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2872 static void __mptcp_wr_shutdown(struct sock *sk)
2874 struct mptcp_sock *msk = mptcp_sk(sk);
2876 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2877 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2878 !!mptcp_send_head(sk));
2880 /* will be ignored by fallback sockets */
2881 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2882 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2884 mptcp_check_send_data_fin(sk);
2887 static void __mptcp_destroy_sock(struct sock *sk)
2889 struct mptcp_sock *msk = mptcp_sk(sk);
2891 pr_debug("msk=%p", msk);
2895 mptcp_stop_timer(sk);
2896 sk_stop_timer(sk, &sk->sk_timer);
2898 mptcp_release_sched(msk);
2900 sk->sk_prot->destroy(sk);
2902 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2903 WARN_ON_ONCE(msk->rmem_released);
2904 sk_stream_kill_queues(sk);
2905 xfrm_sk_free_policy(sk);
2910 void __mptcp_unaccepted_force_close(struct sock *sk)
2912 sock_set_flag(sk, SOCK_DEAD);
2913 mptcp_do_fastclose(sk);
2914 __mptcp_destroy_sock(sk);
2917 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2919 /* Concurrent splices from sk_receive_queue into receive_queue will
2920 * always show at least one non-empty queue when checked in this order.
2922 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2923 skb_queue_empty_lockless(&msk->receive_queue))
2926 return EPOLLIN | EPOLLRDNORM;
2929 static void mptcp_check_listen_stop(struct sock *sk)
2933 if (inet_sk_state_load(sk) != TCP_LISTEN)
2936 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2937 ssk = mptcp_sk(sk)->first;
2938 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2941 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2942 tcp_set_state(ssk, TCP_CLOSE);
2943 mptcp_subflow_queue_clean(sk, ssk);
2944 inet_csk_listen_stop(ssk);
2945 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2949 bool __mptcp_close(struct sock *sk, long timeout)
2951 struct mptcp_subflow_context *subflow;
2952 struct mptcp_sock *msk = mptcp_sk(sk);
2953 bool do_cancel_work = false;
2954 int subflows_alive = 0;
2956 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2958 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2959 mptcp_check_listen_stop(sk);
2960 inet_sk_state_store(sk, TCP_CLOSE);
2964 if (mptcp_check_readable(msk) || timeout < 0) {
2965 /* If the msk has read data, or the caller explicitly ask it,
2966 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
2968 mptcp_do_fastclose(sk);
2970 } else if (mptcp_close_state(sk)) {
2971 __mptcp_wr_shutdown(sk);
2974 sk_stream_wait_close(sk, timeout);
2977 /* orphan all the subflows */
2978 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2979 mptcp_for_each_subflow(msk, subflow) {
2980 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2981 bool slow = lock_sock_fast_nested(ssk);
2983 subflows_alive += ssk->sk_state != TCP_CLOSE;
2985 /* since the close timeout takes precedence on the fail one,
2988 if (ssk == msk->first)
2989 subflow->fail_tout = 0;
2991 /* detach from the parent socket, but allow data_ready to
2992 * push incoming data into the mptcp stack, to properly ack it
2994 ssk->sk_socket = NULL;
2996 unlock_sock_fast(ssk, slow);
3000 /* all the subflows are closed, only timeout can change the msk
3001 * state, let's not keep resources busy for no reasons
3003 if (subflows_alive == 0)
3004 inet_sk_state_store(sk, TCP_CLOSE);
3007 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3009 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3011 if (sk->sk_state == TCP_CLOSE) {
3012 __mptcp_destroy_sock(sk);
3013 do_cancel_work = true;
3015 mptcp_reset_timeout(msk, 0);
3018 return do_cancel_work;
3021 static void mptcp_close(struct sock *sk, long timeout)
3023 bool do_cancel_work;
3027 do_cancel_work = __mptcp_close(sk, timeout);
3030 mptcp_cancel_work(sk);
3035 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3037 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3038 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3039 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3041 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3042 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3045 msk6->saddr = ssk6->saddr;
3046 msk6->flow_label = ssk6->flow_label;
3050 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3051 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3052 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3053 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3054 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3055 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3058 static int mptcp_disconnect(struct sock *sk, int flags)
3060 struct mptcp_sock *msk = mptcp_sk(sk);
3062 /* Deny disconnect if other threads are blocked in sk_wait_event()
3063 * or inet_wait_for_connect().
3065 if (sk->sk_wait_pending)
3068 /* We are on the fastopen error path. We can't call straight into the
3069 * subflows cleanup code due to lock nesting (we are already under
3070 * msk->firstsocket lock).
3072 if (msk->fastopening)
3075 mptcp_check_listen_stop(sk);
3076 inet_sk_state_store(sk, TCP_CLOSE);
3078 mptcp_stop_timer(sk);
3079 sk_stop_timer(sk, &sk->sk_timer);
3082 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3084 /* msk->subflow is still intact, the following will not free the first
3087 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3088 WRITE_ONCE(msk->flags, 0);
3090 msk->push_pending = 0;
3091 msk->recovery = false;
3092 msk->can_ack = false;
3093 msk->fully_established = false;
3094 msk->rcv_data_fin = false;
3095 msk->snd_data_fin_enable = false;
3096 msk->rcv_fastclose = false;
3097 msk->use_64bit_ack = false;
3098 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3099 mptcp_pm_data_reset(msk);
3101 msk->bytes_acked = 0;
3102 msk->bytes_received = 0;
3103 msk->bytes_sent = 0;
3104 msk->bytes_retrans = 0;
3106 WRITE_ONCE(sk->sk_shutdown, 0);
3107 sk_error_report(sk);
3111 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3112 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3114 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3116 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3120 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3121 const struct mptcp_options_received *mp_opt,
3123 struct request_sock *req)
3125 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3126 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3127 struct mptcp_sock *msk;
3132 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3133 if (nsk->sk_family == AF_INET6)
3134 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3137 nsk->sk_wait_pending = 0;
3138 __mptcp_init_sock(nsk);
3140 msk = mptcp_sk(nsk);
3141 msk->local_key = subflow_req->local_key;
3142 msk->token = subflow_req->token;
3143 msk->in_accept_queue = 1;
3144 WRITE_ONCE(msk->fully_established, false);
3145 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3146 WRITE_ONCE(msk->csum_enabled, true);
3148 msk->write_seq = subflow_req->idsn + 1;
3149 msk->snd_nxt = msk->write_seq;
3150 msk->snd_una = msk->write_seq;
3151 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3152 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3153 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3155 /* passive msk is created after the first/MPC subflow */
3156 msk->subflow_id = 2;
3158 sock_reset_flag(nsk, SOCK_RCU_FREE);
3159 security_inet_csk_clone(nsk, req);
3161 /* this can't race with mptcp_close(), as the msk is
3162 * not yet exposted to user-space
3164 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3166 /* The msk maintain a ref to each subflow in the connections list */
3167 WRITE_ONCE(msk->first, ssk);
3168 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3171 /* new mpc subflow takes ownership of the newly
3172 * created mptcp socket
3174 mptcp_token_accept(subflow_req, msk);
3176 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3177 * uses the correct data
3179 mptcp_copy_inaddrs(nsk, ssk);
3180 mptcp_propagate_sndbuf(nsk, ssk);
3182 mptcp_rcv_space_init(msk, ssk);
3183 bh_unlock_sock(nsk);
3185 /* note: the newly allocated socket refcount is 2 now */
3189 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3191 const struct tcp_sock *tp = tcp_sk(ssk);
3193 msk->rcvq_space.copied = 0;
3194 msk->rcvq_space.rtt_us = 0;
3196 msk->rcvq_space.time = tp->tcp_mstamp;
3198 /* initial rcv_space offering made to peer */
3199 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3200 TCP_INIT_CWND * tp->advmss);
3201 if (msk->rcvq_space.space == 0)
3202 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3204 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3207 static struct sock *mptcp_accept(struct sock *ssk, int flags, int *err,
3212 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3213 newsk = inet_csk_accept(ssk, flags, err, kern);
3217 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3218 if (sk_is_mptcp(newsk)) {
3219 struct mptcp_subflow_context *subflow;
3220 struct sock *new_mptcp_sock;
3222 subflow = mptcp_subflow_ctx(newsk);
3223 new_mptcp_sock = subflow->conn;
3225 /* is_mptcp should be false if subflow->conn is missing, see
3226 * subflow_syn_recv_sock()
3228 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3229 tcp_sk(newsk)->is_mptcp = 0;
3233 newsk = new_mptcp_sock;
3234 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3236 MPTCP_INC_STATS(sock_net(ssk),
3237 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3241 newsk->sk_kern_sock = kern;
3245 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3247 struct mptcp_subflow_context *subflow, *tmp;
3248 struct sock *sk = (struct sock *)msk;
3250 __mptcp_clear_xmit(sk);
3252 /* join list will be eventually flushed (with rst) at sock lock release time */
3253 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3254 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3256 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3257 mptcp_data_lock(sk);
3258 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3259 __skb_queue_purge(&sk->sk_receive_queue);
3260 skb_rbtree_purge(&msk->out_of_order_queue);
3261 mptcp_data_unlock(sk);
3263 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3264 * inet_sock_destruct() will dispose it
3266 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3267 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3268 mptcp_token_destroy(msk);
3269 mptcp_pm_free_anno_list(msk);
3270 mptcp_free_local_addr_list(msk);
3273 static void mptcp_destroy(struct sock *sk)
3275 struct mptcp_sock *msk = mptcp_sk(sk);
3277 /* allow the following to close even the initial subflow */
3278 msk->free_first = 1;
3279 mptcp_destroy_common(msk, 0);
3280 sk_sockets_allocated_dec(sk);
3283 void __mptcp_data_acked(struct sock *sk)
3285 if (!sock_owned_by_user(sk))
3286 __mptcp_clean_una(sk);
3288 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3290 if (mptcp_pending_data_fin_ack(sk))
3291 mptcp_schedule_work(sk);
3294 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3296 if (!mptcp_send_head(sk))
3299 if (!sock_owned_by_user(sk))
3300 __mptcp_subflow_push_pending(sk, ssk, false);
3302 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3305 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3306 BIT(MPTCP_RETRANSMIT) | \
3307 BIT(MPTCP_FLUSH_JOIN_LIST))
3309 /* processes deferred events and flush wmem */
3310 static void mptcp_release_cb(struct sock *sk)
3311 __must_hold(&sk->sk_lock.slock)
3313 struct mptcp_sock *msk = mptcp_sk(sk);
3316 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3318 struct list_head join_list;
3323 INIT_LIST_HEAD(&join_list);
3324 list_splice_init(&msk->join_list, &join_list);
3326 /* the following actions acquire the subflow socket lock
3328 * 1) can't be invoked in atomic scope
3329 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3330 * datapath acquires the msk socket spinlock while helding
3331 * the subflow socket lock
3333 msk->push_pending = 0;
3334 msk->cb_flags &= ~flags;
3335 spin_unlock_bh(&sk->sk_lock.slock);
3337 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3338 __mptcp_flush_join_list(sk, &join_list);
3339 if (flags & BIT(MPTCP_PUSH_PENDING))
3340 __mptcp_push_pending(sk, 0);
3341 if (flags & BIT(MPTCP_RETRANSMIT))
3342 __mptcp_retrans(sk);
3345 spin_lock_bh(&sk->sk_lock.slock);
3348 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3349 __mptcp_clean_una_wakeup(sk);
3350 if (unlikely(msk->cb_flags)) {
3351 /* be sure to set the current sk state before tacking actions
3352 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3354 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3355 __mptcp_set_connected(sk);
3356 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3357 __mptcp_error_report(sk);
3360 __mptcp_update_rmem(sk);
3363 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3364 * TCP can't schedule delack timer before the subflow is fully established.
3365 * MPTCP uses the delack timer to do 3rd ack retransmissions
3367 static void schedule_3rdack_retransmission(struct sock *ssk)
3369 struct inet_connection_sock *icsk = inet_csk(ssk);
3370 struct tcp_sock *tp = tcp_sk(ssk);
3371 unsigned long timeout;
3373 if (mptcp_subflow_ctx(ssk)->fully_established)
3376 /* reschedule with a timeout above RTT, as we must look only for drop */
3378 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3380 timeout = TCP_TIMEOUT_INIT;
3383 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3384 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3385 icsk->icsk_ack.timeout = timeout;
3386 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3389 void mptcp_subflow_process_delegated(struct sock *ssk)
3391 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3392 struct sock *sk = subflow->conn;
3394 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3395 mptcp_data_lock(sk);
3396 if (!sock_owned_by_user(sk))
3397 __mptcp_subflow_push_pending(sk, ssk, true);
3399 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3400 mptcp_data_unlock(sk);
3401 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3403 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3404 schedule_3rdack_retransmission(ssk);
3405 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3409 static int mptcp_hash(struct sock *sk)
3411 /* should never be called,
3412 * we hash the TCP subflows not the master socket
3418 static void mptcp_unhash(struct sock *sk)
3420 /* called from sk_common_release(), but nothing to do here */
3423 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3425 struct mptcp_sock *msk = mptcp_sk(sk);
3427 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3428 if (WARN_ON_ONCE(!msk->first))
3431 return inet_csk_get_port(msk->first, snum);
3434 void mptcp_finish_connect(struct sock *ssk)
3436 struct mptcp_subflow_context *subflow;
3437 struct mptcp_sock *msk;
3440 subflow = mptcp_subflow_ctx(ssk);
3444 pr_debug("msk=%p, token=%u", sk, subflow->token);
3446 subflow->map_seq = subflow->iasn;
3447 subflow->map_subflow_seq = 1;
3449 /* the socket is not connected yet, no msk/subflow ops can access/race
3450 * accessing the field below
3452 WRITE_ONCE(msk->local_key, subflow->local_key);
3453 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3454 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3455 WRITE_ONCE(msk->snd_una, msk->write_seq);
3457 mptcp_pm_new_connection(msk, ssk, 0);
3459 mptcp_rcv_space_init(msk, ssk);
3462 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3464 write_lock_bh(&sk->sk_callback_lock);
3465 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3466 sk_set_socket(sk, parent);
3467 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3468 write_unlock_bh(&sk->sk_callback_lock);
3471 bool mptcp_finish_join(struct sock *ssk)
3473 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3474 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3475 struct sock *parent = (void *)msk;
3478 pr_debug("msk=%p, subflow=%p", msk, subflow);
3480 /* mptcp socket already closing? */
3481 if (!mptcp_is_fully_established(parent)) {
3482 subflow->reset_reason = MPTCP_RST_EMPTCP;
3486 /* active subflow, already present inside the conn_list */
3487 if (!list_empty(&subflow->node)) {
3488 mptcp_subflow_joined(msk, ssk);
3492 if (!mptcp_pm_allow_new_subflow(msk))
3493 goto err_prohibited;
3495 /* If we can't acquire msk socket lock here, let the release callback
3498 mptcp_data_lock(parent);
3499 if (!sock_owned_by_user(parent)) {
3500 ret = __mptcp_finish_join(msk, ssk);
3503 list_add_tail(&subflow->node, &msk->conn_list);
3507 list_add_tail(&subflow->node, &msk->join_list);
3508 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3510 mptcp_data_unlock(parent);
3514 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3521 static void mptcp_shutdown(struct sock *sk, int how)
3523 pr_debug("sk=%p, how=%d", sk, how);
3525 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3526 __mptcp_wr_shutdown(sk);
3529 static int mptcp_forward_alloc_get(const struct sock *sk)
3531 return READ_ONCE(sk->sk_forward_alloc) +
3532 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3535 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3537 const struct sock *sk = (void *)msk;
3540 if (sk->sk_state == TCP_LISTEN)
3543 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3546 delta = msk->write_seq - v;
3547 if (__mptcp_check_fallback(msk) && msk->first) {
3548 struct tcp_sock *tp = tcp_sk(msk->first);
3550 /* the first subflow is disconnected after close - see
3551 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3552 * so ignore that status, too.
3554 if (!((1 << msk->first->sk_state) &
3555 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3556 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3558 if (delta > INT_MAX)
3564 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3566 struct mptcp_sock *msk = mptcp_sk(sk);
3571 if (sk->sk_state == TCP_LISTEN)
3575 __mptcp_move_skbs(msk);
3576 *karg = mptcp_inq_hint(sk);
3580 slow = lock_sock_fast(sk);
3581 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3582 unlock_sock_fast(sk, slow);
3585 slow = lock_sock_fast(sk);
3586 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3587 unlock_sock_fast(sk, slow);
3590 return -ENOIOCTLCMD;
3596 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3597 struct mptcp_subflow_context *subflow)
3599 subflow->request_mptcp = 0;
3600 __mptcp_do_fallback(msk);
3603 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3605 struct mptcp_subflow_context *subflow;
3606 struct mptcp_sock *msk = mptcp_sk(sk);
3610 ssk = __mptcp_nmpc_sk(msk);
3612 return PTR_ERR(ssk);
3614 inet_sk_state_store(sk, TCP_SYN_SENT);
3615 subflow = mptcp_subflow_ctx(ssk);
3616 #ifdef CONFIG_TCP_MD5SIG
3617 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3620 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3621 mptcp_subflow_early_fallback(msk, subflow);
3623 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3624 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3625 mptcp_subflow_early_fallback(msk, subflow);
3627 if (likely(!__mptcp_check_fallback(msk)))
3628 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3630 /* if reaching here via the fastopen/sendmsg path, the caller already
3631 * acquired the subflow socket lock, too.
3633 if (!msk->fastopening)
3636 /* the following mirrors closely a very small chunk of code from
3637 * __inet_stream_connect()
3639 if (ssk->sk_state != TCP_CLOSE)
3642 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3643 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3648 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3652 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3655 if (!msk->fastopening)
3658 /* on successful connect, the msk state will be moved to established by
3659 * subflow_finish_connect()
3661 if (unlikely(err)) {
3662 /* avoid leaving a dangling token in an unconnected socket */
3663 mptcp_token_destroy(msk);
3664 inet_sk_state_store(sk, TCP_CLOSE);
3668 mptcp_copy_inaddrs(sk, ssk);
3672 static struct proto mptcp_prot = {
3674 .owner = THIS_MODULE,
3675 .init = mptcp_init_sock,
3676 .connect = mptcp_connect,
3677 .disconnect = mptcp_disconnect,
3678 .close = mptcp_close,
3679 .accept = mptcp_accept,
3680 .setsockopt = mptcp_setsockopt,
3681 .getsockopt = mptcp_getsockopt,
3682 .shutdown = mptcp_shutdown,
3683 .destroy = mptcp_destroy,
3684 .sendmsg = mptcp_sendmsg,
3685 .ioctl = mptcp_ioctl,
3686 .recvmsg = mptcp_recvmsg,
3687 .release_cb = mptcp_release_cb,
3689 .unhash = mptcp_unhash,
3690 .get_port = mptcp_get_port,
3691 .forward_alloc_get = mptcp_forward_alloc_get,
3692 .sockets_allocated = &mptcp_sockets_allocated,
3694 .memory_allocated = &tcp_memory_allocated,
3695 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3697 .memory_pressure = &tcp_memory_pressure,
3698 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3699 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3700 .sysctl_mem = sysctl_tcp_mem,
3701 .obj_size = sizeof(struct mptcp_sock),
3702 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3703 .no_autobind = true,
3706 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3708 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3709 struct sock *ssk, *sk = sock->sk;
3713 ssk = __mptcp_nmpc_sk(msk);
3719 if (sk->sk_family == AF_INET)
3720 err = inet_bind_sk(ssk, uaddr, addr_len);
3721 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3722 else if (sk->sk_family == AF_INET6)
3723 err = inet6_bind_sk(ssk, uaddr, addr_len);
3726 mptcp_copy_inaddrs(sk, ssk);
3733 static int mptcp_listen(struct socket *sock, int backlog)
3735 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3736 struct sock *sk = sock->sk;
3740 pr_debug("msk=%p", msk);
3745 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3748 ssk = __mptcp_nmpc_sk(msk);
3754 inet_sk_state_store(sk, TCP_LISTEN);
3755 sock_set_flag(sk, SOCK_RCU_FREE);
3758 err = __inet_listen_sk(ssk, backlog);
3760 inet_sk_state_store(sk, inet_sk_state_load(ssk));
3763 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3764 mptcp_copy_inaddrs(sk, ssk);
3765 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3773 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3774 int flags, bool kern)
3776 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3777 struct sock *ssk, *newsk;
3780 pr_debug("msk=%p", msk);
3782 /* Buggy applications can call accept on socket states other then LISTEN
3783 * but no need to allocate the first subflow just to error out.
3785 ssk = READ_ONCE(msk->first);
3789 newsk = mptcp_accept(ssk, flags, &err, kern);
3795 __inet_accept(sock, newsock, newsk);
3796 if (!mptcp_is_tcpsk(newsock->sk)) {
3797 struct mptcp_sock *msk = mptcp_sk(newsk);
3798 struct mptcp_subflow_context *subflow;
3800 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3801 msk->in_accept_queue = 0;
3803 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3804 * This is needed so NOSPACE flag can be set from tcp stack.
3806 mptcp_for_each_subflow(msk, subflow) {
3807 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3809 if (!ssk->sk_socket)
3810 mptcp_sock_graft(ssk, newsock);
3813 /* Do late cleanup for the first subflow as necessary. Also
3814 * deal with bad peers not doing a complete shutdown.
3816 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3817 __mptcp_close_ssk(newsk, msk->first,
3818 mptcp_subflow_ctx(msk->first), 0);
3819 if (unlikely(list_is_singular(&msk->conn_list)))
3820 inet_sk_state_store(newsk, TCP_CLOSE);
3823 release_sock(newsk);
3828 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3830 struct sock *sk = (struct sock *)msk;
3832 if (sk_stream_is_writeable(sk))
3833 return EPOLLOUT | EPOLLWRNORM;
3835 mptcp_set_nospace(sk);
3836 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3837 if (sk_stream_is_writeable(sk))
3838 return EPOLLOUT | EPOLLWRNORM;
3843 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3844 struct poll_table_struct *wait)
3846 struct sock *sk = sock->sk;
3847 struct mptcp_sock *msk;
3853 sock_poll_wait(file, sock, wait);
3855 state = inet_sk_state_load(sk);
3856 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3857 if (state == TCP_LISTEN) {
3858 struct sock *ssk = READ_ONCE(msk->first);
3860 if (WARN_ON_ONCE(!ssk))
3863 return inet_csk_listen_poll(ssk);
3866 shutdown = READ_ONCE(sk->sk_shutdown);
3867 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3869 if (shutdown & RCV_SHUTDOWN)
3870 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3872 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3873 mask |= mptcp_check_readable(msk);
3874 if (shutdown & SEND_SHUTDOWN)
3875 mask |= EPOLLOUT | EPOLLWRNORM;
3877 mask |= mptcp_check_writeable(msk);
3878 } else if (state == TCP_SYN_SENT &&
3879 inet_test_bit(DEFER_CONNECT, sk)) {
3880 /* cf tcp_poll() note about TFO */
3881 mask |= EPOLLOUT | EPOLLWRNORM;
3884 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3886 if (READ_ONCE(sk->sk_err))
3892 static const struct proto_ops mptcp_stream_ops = {
3894 .owner = THIS_MODULE,
3895 .release = inet_release,
3897 .connect = inet_stream_connect,
3898 .socketpair = sock_no_socketpair,
3899 .accept = mptcp_stream_accept,
3900 .getname = inet_getname,
3902 .ioctl = inet_ioctl,
3903 .gettstamp = sock_gettstamp,
3904 .listen = mptcp_listen,
3905 .shutdown = inet_shutdown,
3906 .setsockopt = sock_common_setsockopt,
3907 .getsockopt = sock_common_getsockopt,
3908 .sendmsg = inet_sendmsg,
3909 .recvmsg = inet_recvmsg,
3910 .mmap = sock_no_mmap,
3913 static struct inet_protosw mptcp_protosw = {
3914 .type = SOCK_STREAM,
3915 .protocol = IPPROTO_MPTCP,
3916 .prot = &mptcp_prot,
3917 .ops = &mptcp_stream_ops,
3918 .flags = INET_PROTOSW_ICSK,
3921 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3923 struct mptcp_delegated_action *delegated;
3924 struct mptcp_subflow_context *subflow;
3927 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3928 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3929 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3931 bh_lock_sock_nested(ssk);
3932 if (!sock_owned_by_user(ssk) &&
3933 mptcp_subflow_has_delegated_action(subflow))
3934 mptcp_subflow_process_delegated(ssk);
3935 /* ... elsewhere tcp_release_cb_override already processed
3936 * the action or will do at next release_sock().
3937 * In both case must dequeue the subflow here - on the same
3938 * CPU that scheduled it.
3940 bh_unlock_sock(ssk);
3943 if (++work_done == budget)
3947 /* always provide a 0 'work_done' argument, so that napi_complete_done
3948 * will not try accessing the NULL napi->dev ptr
3950 napi_complete_done(napi, 0);
3954 void __init mptcp_proto_init(void)
3956 struct mptcp_delegated_action *delegated;
3959 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3961 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3962 panic("Failed to allocate MPTCP pcpu counter\n");
3964 init_dummy_netdev(&mptcp_napi_dev);
3965 for_each_possible_cpu(cpu) {
3966 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3967 INIT_LIST_HEAD(&delegated->head);
3968 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3970 napi_enable(&delegated->napi);
3973 mptcp_subflow_init();
3978 if (proto_register(&mptcp_prot, 1) != 0)
3979 panic("Failed to register MPTCP proto.\n");
3981 inet_register_protosw(&mptcp_protosw);
3983 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3986 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3987 static const struct proto_ops mptcp_v6_stream_ops = {
3989 .owner = THIS_MODULE,
3990 .release = inet6_release,
3992 .connect = inet_stream_connect,
3993 .socketpair = sock_no_socketpair,
3994 .accept = mptcp_stream_accept,
3995 .getname = inet6_getname,
3997 .ioctl = inet6_ioctl,
3998 .gettstamp = sock_gettstamp,
3999 .listen = mptcp_listen,
4000 .shutdown = inet_shutdown,
4001 .setsockopt = sock_common_setsockopt,
4002 .getsockopt = sock_common_getsockopt,
4003 .sendmsg = inet6_sendmsg,
4004 .recvmsg = inet6_recvmsg,
4005 .mmap = sock_no_mmap,
4006 #ifdef CONFIG_COMPAT
4007 .compat_ioctl = inet6_compat_ioctl,
4011 static struct proto mptcp_v6_prot;
4013 static struct inet_protosw mptcp_v6_protosw = {
4014 .type = SOCK_STREAM,
4015 .protocol = IPPROTO_MPTCP,
4016 .prot = &mptcp_v6_prot,
4017 .ops = &mptcp_v6_stream_ops,
4018 .flags = INET_PROTOSW_ICSK,
4021 int __init mptcp_proto_v6_init(void)
4025 mptcp_v6_prot = mptcp_prot;
4026 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4027 mptcp_v6_prot.slab = NULL;
4028 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4029 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4031 err = proto_register(&mptcp_v6_prot, 1);
4035 err = inet6_register_protosw(&mptcp_v6_protosw);
4037 proto_unregister(&mptcp_v6_prot);