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 sock->ops = &inet_stream_ops;
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 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 WRITE_ONCE(msk->subflow, ssock);
96 subflow = mptcp_subflow_ctx(ssock->sk);
97 list_add(&subflow->node, &msk->conn_list);
99 subflow->request_mptcp = 1;
100 subflow->subflow_id = msk->subflow_id++;
102 /* This is the first subflow, always with id 0 */
103 subflow->local_id_valid = 1;
104 mptcp_sock_graft(msk->first, sk->sk_socket);
109 /* If the MPC handshake is not started, returns the first subflow,
110 * eventually allocating it.
112 struct socket *__mptcp_nmpc_socket(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);
122 return ERR_PTR(-EINVAL);
124 ret = __mptcp_socket_create(msk);
128 mptcp_sockopt_sync(msk, msk->first);
134 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
136 sk_drops_add(sk, skb);
140 static void mptcp_rmem_charge(struct sock *sk, int size)
142 mptcp_sk(sk)->rmem_fwd_alloc -= size;
145 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
146 struct sk_buff *from)
151 if (MPTCP_SKB_CB(from)->offset ||
152 !skb_try_coalesce(to, from, &fragstolen, &delta))
155 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
156 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
157 to->len, MPTCP_SKB_CB(from)->end_seq);
158 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
160 /* note the fwd memory can reach a negative value after accounting
161 * for the delta, but the later skb free will restore a non
164 atomic_add(delta, &sk->sk_rmem_alloc);
165 mptcp_rmem_charge(sk, delta);
166 kfree_skb_partial(from, fragstolen);
171 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
172 struct sk_buff *from)
174 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
177 return mptcp_try_coalesce((struct sock *)msk, to, from);
180 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
182 amount >>= PAGE_SHIFT;
183 mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
184 __sk_mem_reduce_allocated(sk, amount);
187 static void mptcp_rmem_uncharge(struct sock *sk, int size)
189 struct mptcp_sock *msk = mptcp_sk(sk);
192 msk->rmem_fwd_alloc += size;
193 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
195 /* see sk_mem_uncharge() for the rationale behind the following schema */
196 if (unlikely(reclaimable >= PAGE_SIZE))
197 __mptcp_rmem_reclaim(sk, reclaimable);
200 static void mptcp_rfree(struct sk_buff *skb)
202 unsigned int len = skb->truesize;
203 struct sock *sk = skb->sk;
205 atomic_sub(len, &sk->sk_rmem_alloc);
206 mptcp_rmem_uncharge(sk, len);
209 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
213 skb->destructor = mptcp_rfree;
214 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
215 mptcp_rmem_charge(sk, skb->truesize);
218 /* "inspired" by tcp_data_queue_ofo(), main differences:
220 * - don't cope with sacks
222 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
224 struct sock *sk = (struct sock *)msk;
225 struct rb_node **p, *parent;
226 u64 seq, end_seq, max_seq;
227 struct sk_buff *skb1;
229 seq = MPTCP_SKB_CB(skb)->map_seq;
230 end_seq = MPTCP_SKB_CB(skb)->end_seq;
231 max_seq = atomic64_read(&msk->rcv_wnd_sent);
233 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
234 RB_EMPTY_ROOT(&msk->out_of_order_queue));
235 if (after64(end_seq, max_seq)) {
238 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
239 (unsigned long long)end_seq - (unsigned long)max_seq,
240 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
241 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
245 p = &msk->out_of_order_queue.rb_node;
246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
247 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
248 rb_link_node(&skb->rbnode, NULL, p);
249 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
250 msk->ooo_last_skb = skb;
254 /* with 2 subflows, adding at end of ooo queue is quite likely
255 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
257 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
258 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
259 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
263 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
264 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
265 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
266 parent = &msk->ooo_last_skb->rbnode;
267 p = &parent->rb_right;
271 /* Find place to insert this segment. Handle overlaps on the way. */
275 skb1 = rb_to_skb(parent);
276 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
277 p = &parent->rb_left;
280 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
281 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
282 /* All the bits are present. Drop. */
284 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
287 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
291 * continue traversing
294 /* skb's seq == skb1's seq and skb covers skb1.
295 * Replace skb1 with skb.
297 rb_replace_node(&skb1->rbnode, &skb->rbnode,
298 &msk->out_of_order_queue);
299 mptcp_drop(sk, skb1);
300 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
303 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
304 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
307 p = &parent->rb_right;
311 /* Insert segment into RB tree. */
312 rb_link_node(&skb->rbnode, parent, p);
313 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
316 /* Remove other segments covered by skb. */
317 while ((skb1 = skb_rb_next(skb)) != NULL) {
318 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
320 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
321 mptcp_drop(sk, skb1);
322 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
324 /* If there is no skb after us, we are the last_skb ! */
326 msk->ooo_last_skb = skb;
330 mptcp_set_owner_r(skb, sk);
333 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
335 struct mptcp_sock *msk = mptcp_sk(sk);
338 if (size <= msk->rmem_fwd_alloc)
341 size -= msk->rmem_fwd_alloc;
342 amt = sk_mem_pages(size);
343 amount = amt << PAGE_SHIFT;
344 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
347 msk->rmem_fwd_alloc += amount;
351 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
352 struct sk_buff *skb, unsigned int offset,
355 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
356 struct sock *sk = (struct sock *)msk;
357 struct sk_buff *tail;
360 __skb_unlink(skb, &ssk->sk_receive_queue);
365 /* try to fetch required memory from subflow */
366 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
369 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
371 /* the skb map_seq accounts for the skb offset:
372 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
375 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
376 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
377 MPTCP_SKB_CB(skb)->offset = offset;
378 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
380 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
382 msk->bytes_received += copy_len;
383 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
384 tail = skb_peek_tail(&sk->sk_receive_queue);
385 if (tail && mptcp_try_coalesce(sk, tail, skb))
388 mptcp_set_owner_r(skb, sk);
389 __skb_queue_tail(&sk->sk_receive_queue, skb);
391 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
392 mptcp_data_queue_ofo(msk, skb);
396 /* old data, keep it simple and drop the whole pkt, sender
397 * will retransmit as needed, if needed.
399 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
405 static void mptcp_stop_timer(struct sock *sk)
407 struct inet_connection_sock *icsk = inet_csk(sk);
409 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
410 mptcp_sk(sk)->timer_ival = 0;
413 static void mptcp_close_wake_up(struct sock *sk)
415 if (sock_flag(sk, SOCK_DEAD))
418 sk->sk_state_change(sk);
419 if (sk->sk_shutdown == SHUTDOWN_MASK ||
420 sk->sk_state == TCP_CLOSE)
421 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
423 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
426 static bool mptcp_pending_data_fin_ack(struct sock *sk)
428 struct mptcp_sock *msk = mptcp_sk(sk);
430 return ((1 << sk->sk_state) &
431 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
432 msk->write_seq == READ_ONCE(msk->snd_una);
435 static void mptcp_check_data_fin_ack(struct sock *sk)
437 struct mptcp_sock *msk = mptcp_sk(sk);
439 /* Look for an acknowledged DATA_FIN */
440 if (mptcp_pending_data_fin_ack(sk)) {
441 WRITE_ONCE(msk->snd_data_fin_enable, 0);
443 switch (sk->sk_state) {
445 inet_sk_state_store(sk, TCP_FIN_WAIT2);
449 inet_sk_state_store(sk, TCP_CLOSE);
453 mptcp_close_wake_up(sk);
457 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
459 struct mptcp_sock *msk = mptcp_sk(sk);
461 if (READ_ONCE(msk->rcv_data_fin) &&
462 ((1 << sk->sk_state) &
463 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
464 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
466 if (msk->ack_seq == rcv_data_fin_seq) {
468 *seq = rcv_data_fin_seq;
477 static void mptcp_set_datafin_timeout(struct sock *sk)
479 struct inet_connection_sock *icsk = inet_csk(sk);
482 retransmits = min_t(u32, icsk->icsk_retransmits,
483 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
485 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
488 static void __mptcp_set_timeout(struct sock *sk, long tout)
490 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
493 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
495 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
497 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
498 inet_csk(ssk)->icsk_timeout - jiffies : 0;
501 static void mptcp_set_timeout(struct sock *sk)
503 struct mptcp_subflow_context *subflow;
506 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
507 tout = max(tout, mptcp_timeout_from_subflow(subflow));
508 __mptcp_set_timeout(sk, tout);
511 static inline bool tcp_can_send_ack(const struct sock *ssk)
513 return !((1 << inet_sk_state_load(ssk)) &
514 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
517 void __mptcp_subflow_send_ack(struct sock *ssk)
519 if (tcp_can_send_ack(ssk))
523 static void mptcp_subflow_send_ack(struct sock *ssk)
527 slow = lock_sock_fast(ssk);
528 __mptcp_subflow_send_ack(ssk);
529 unlock_sock_fast(ssk, slow);
532 static void mptcp_send_ack(struct mptcp_sock *msk)
534 struct mptcp_subflow_context *subflow;
536 mptcp_for_each_subflow(msk, subflow)
537 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
540 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
544 slow = lock_sock_fast(ssk);
545 if (tcp_can_send_ack(ssk))
546 tcp_cleanup_rbuf(ssk, 1);
547 unlock_sock_fast(ssk, slow);
550 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
552 const struct inet_connection_sock *icsk = inet_csk(ssk);
553 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
554 const struct tcp_sock *tp = tcp_sk(ssk);
556 return (ack_pending & ICSK_ACK_SCHED) &&
557 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
558 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
559 (rx_empty && ack_pending &
560 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
563 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
565 int old_space = READ_ONCE(msk->old_wspace);
566 struct mptcp_subflow_context *subflow;
567 struct sock *sk = (struct sock *)msk;
568 int space = __mptcp_space(sk);
569 bool cleanup, rx_empty;
571 cleanup = (space > 0) && (space >= (old_space << 1));
572 rx_empty = !__mptcp_rmem(sk);
574 mptcp_for_each_subflow(msk, subflow) {
575 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
577 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
578 mptcp_subflow_cleanup_rbuf(ssk);
582 static bool mptcp_check_data_fin(struct sock *sk)
584 struct mptcp_sock *msk = mptcp_sk(sk);
585 u64 rcv_data_fin_seq;
588 /* Need to ack a DATA_FIN received from a peer while this side
589 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
590 * msk->rcv_data_fin was set when parsing the incoming options
591 * at the subflow level and the msk lock was not held, so this
592 * is the first opportunity to act on the DATA_FIN and change
595 * If we are caught up to the sequence number of the incoming
596 * DATA_FIN, send the DATA_ACK now and do state transition. If
597 * not caught up, do nothing and let the recv code send DATA_ACK
601 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
602 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
603 WRITE_ONCE(msk->rcv_data_fin, 0);
605 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
606 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
608 switch (sk->sk_state) {
609 case TCP_ESTABLISHED:
610 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
613 inet_sk_state_store(sk, TCP_CLOSING);
616 inet_sk_state_store(sk, TCP_CLOSE);
619 /* Other states not expected */
625 if (!__mptcp_check_fallback(msk))
627 mptcp_close_wake_up(sk);
632 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
636 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
637 struct sock *sk = (struct sock *)msk;
638 unsigned int moved = 0;
639 bool more_data_avail;
644 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
646 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
647 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
649 if (unlikely(ssk_rbuf > sk_rbuf)) {
650 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
655 pr_debug("msk=%p ssk=%p", msk, ssk);
658 u32 map_remaining, offset;
659 u32 seq = tp->copied_seq;
663 /* try to move as much data as available */
664 map_remaining = subflow->map_data_len -
665 mptcp_subflow_get_map_offset(subflow);
667 skb = skb_peek(&ssk->sk_receive_queue);
669 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
670 * a different CPU can have already processed the pending
671 * data, stop here or we can enter an infinite loop
678 if (__mptcp_check_fallback(msk)) {
679 /* Under fallback skbs have no MPTCP extension and TCP could
680 * collapse them between the dummy map creation and the
681 * current dequeue. Be sure to adjust the map size.
683 map_remaining = skb->len;
684 subflow->map_data_len = skb->len;
687 offset = seq - TCP_SKB_CB(skb)->seq;
688 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
694 if (offset < skb->len) {
695 size_t len = skb->len - offset;
700 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
704 if (WARN_ON_ONCE(map_remaining < len))
708 sk_eat_skb(ssk, skb);
712 WRITE_ONCE(tp->copied_seq, seq);
713 more_data_avail = mptcp_subflow_data_available(ssk);
715 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
719 } while (more_data_avail);
725 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
727 struct sock *sk = (struct sock *)msk;
728 struct sk_buff *skb, *tail;
733 p = rb_first(&msk->out_of_order_queue);
734 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
737 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
741 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
743 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
746 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
750 end_seq = MPTCP_SKB_CB(skb)->end_seq;
751 tail = skb_peek_tail(&sk->sk_receive_queue);
752 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
753 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
755 /* skip overlapping data, if any */
756 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
757 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
759 MPTCP_SKB_CB(skb)->offset += delta;
760 MPTCP_SKB_CB(skb)->map_seq += delta;
761 __skb_queue_tail(&sk->sk_receive_queue, skb);
763 msk->bytes_received += end_seq - msk->ack_seq;
764 msk->ack_seq = end_seq;
770 /* In most cases we will be able to lock the mptcp socket. If its already
771 * owned, we need to defer to the work queue to avoid ABBA deadlock.
773 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
775 struct sock *sk = (struct sock *)msk;
776 unsigned int moved = 0;
778 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
779 __mptcp_ofo_queue(msk);
780 if (unlikely(ssk->sk_err)) {
781 if (!sock_owned_by_user(sk))
782 __mptcp_error_report(sk);
784 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
787 /* If the moves have caught up with the DATA_FIN sequence number
788 * it's time to ack the DATA_FIN and change socket state, but
789 * this is not a good place to change state. Let the workqueue
792 if (mptcp_pending_data_fin(sk, NULL))
793 mptcp_schedule_work(sk);
797 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
799 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
800 struct mptcp_sock *msk = mptcp_sk(sk);
801 int sk_rbuf, ssk_rbuf;
803 /* The peer can send data while we are shutting down this
804 * subflow at msk destruction time, but we must avoid enqueuing
805 * more data to the msk receive queue
807 if (unlikely(subflow->disposable))
810 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
811 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
812 if (unlikely(ssk_rbuf > sk_rbuf))
815 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
816 if (__mptcp_rmem(sk) > sk_rbuf) {
817 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
821 /* Wake-up the reader only for in-sequence data */
823 if (move_skbs_to_msk(msk, ssk))
824 sk->sk_data_ready(sk);
826 mptcp_data_unlock(sk);
829 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
831 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
832 WRITE_ONCE(msk->allow_infinite_fallback, false);
833 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
836 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
838 struct sock *sk = (struct sock *)msk;
840 if (sk->sk_state != TCP_ESTABLISHED)
843 /* attach to msk socket only after we are sure we will deal with it
846 if (sk->sk_socket && !ssk->sk_socket)
847 mptcp_sock_graft(ssk, sk->sk_socket);
849 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
850 mptcp_sockopt_sync_locked(msk, ssk);
851 mptcp_subflow_joined(msk, ssk);
855 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
857 struct mptcp_subflow_context *tmp, *subflow;
858 struct mptcp_sock *msk = mptcp_sk(sk);
860 list_for_each_entry_safe(subflow, tmp, join_list, node) {
861 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
862 bool slow = lock_sock_fast(ssk);
864 list_move_tail(&subflow->node, &msk->conn_list);
865 if (!__mptcp_finish_join(msk, ssk))
866 mptcp_subflow_reset(ssk);
867 unlock_sock_fast(ssk, slow);
871 static bool mptcp_timer_pending(struct sock *sk)
873 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
876 static void mptcp_reset_timer(struct sock *sk)
878 struct inet_connection_sock *icsk = inet_csk(sk);
881 /* prevent rescheduling on close */
882 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
885 tout = mptcp_sk(sk)->timer_ival;
886 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
889 bool mptcp_schedule_work(struct sock *sk)
891 if (inet_sk_state_load(sk) != TCP_CLOSE &&
892 schedule_work(&mptcp_sk(sk)->work)) {
893 /* each subflow already holds a reference to the sk, and the
894 * workqueue is invoked by a subflow, so sk can't go away here.
902 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
904 struct mptcp_subflow_context *subflow;
906 msk_owned_by_me(msk);
908 mptcp_for_each_subflow(msk, subflow) {
909 if (READ_ONCE(subflow->data_avail))
910 return mptcp_subflow_tcp_sock(subflow);
916 static bool mptcp_skb_can_collapse_to(u64 write_seq,
917 const struct sk_buff *skb,
918 const struct mptcp_ext *mpext)
920 if (!tcp_skb_can_collapse_to(skb))
923 /* can collapse only if MPTCP level sequence is in order and this
924 * mapping has not been xmitted yet
926 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
930 /* we can append data to the given data frag if:
931 * - there is space available in the backing page_frag
932 * - the data frag tail matches the current page_frag free offset
933 * - the data frag end sequence number matches the current write seq
935 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
936 const struct page_frag *pfrag,
937 const struct mptcp_data_frag *df)
939 return df && pfrag->page == df->page &&
940 pfrag->size - pfrag->offset > 0 &&
941 pfrag->offset == (df->offset + df->data_len) &&
942 df->data_seq + df->data_len == msk->write_seq;
945 static void dfrag_uncharge(struct sock *sk, int len)
947 sk_mem_uncharge(sk, len);
948 sk_wmem_queued_add(sk, -len);
951 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
953 int len = dfrag->data_len + dfrag->overhead;
955 list_del(&dfrag->list);
956 dfrag_uncharge(sk, len);
957 put_page(dfrag->page);
960 static void __mptcp_clean_una(struct sock *sk)
962 struct mptcp_sock *msk = mptcp_sk(sk);
963 struct mptcp_data_frag *dtmp, *dfrag;
966 snd_una = msk->snd_una;
967 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
968 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
971 if (unlikely(dfrag == msk->first_pending)) {
972 /* in recovery mode can see ack after the current snd head */
973 if (WARN_ON_ONCE(!msk->recovery))
976 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
979 dfrag_clear(sk, dfrag);
982 dfrag = mptcp_rtx_head(sk);
983 if (dfrag && after64(snd_una, dfrag->data_seq)) {
984 u64 delta = snd_una - dfrag->data_seq;
986 /* prevent wrap around in recovery mode */
987 if (unlikely(delta > dfrag->already_sent)) {
988 if (WARN_ON_ONCE(!msk->recovery))
990 if (WARN_ON_ONCE(delta > dfrag->data_len))
992 dfrag->already_sent += delta - dfrag->already_sent;
995 dfrag->data_seq += delta;
996 dfrag->offset += delta;
997 dfrag->data_len -= delta;
998 dfrag->already_sent -= delta;
1000 dfrag_uncharge(sk, delta);
1003 /* all retransmitted data acked, recovery completed */
1004 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1005 msk->recovery = false;
1008 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1009 snd_una == READ_ONCE(msk->write_seq)) {
1010 if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1011 mptcp_stop_timer(sk);
1013 mptcp_reset_timer(sk);
1017 static void __mptcp_clean_una_wakeup(struct sock *sk)
1019 lockdep_assert_held_once(&sk->sk_lock.slock);
1021 __mptcp_clean_una(sk);
1022 mptcp_write_space(sk);
1025 static void mptcp_clean_una_wakeup(struct sock *sk)
1027 mptcp_data_lock(sk);
1028 __mptcp_clean_una_wakeup(sk);
1029 mptcp_data_unlock(sk);
1032 static void mptcp_enter_memory_pressure(struct sock *sk)
1034 struct mptcp_subflow_context *subflow;
1035 struct mptcp_sock *msk = mptcp_sk(sk);
1038 sk_stream_moderate_sndbuf(sk);
1039 mptcp_for_each_subflow(msk, subflow) {
1040 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1043 tcp_enter_memory_pressure(ssk);
1044 sk_stream_moderate_sndbuf(ssk);
1049 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1052 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1054 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1055 pfrag, sk->sk_allocation)))
1058 mptcp_enter_memory_pressure(sk);
1062 static struct mptcp_data_frag *
1063 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1066 int offset = ALIGN(orig_offset, sizeof(long));
1067 struct mptcp_data_frag *dfrag;
1069 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1070 dfrag->data_len = 0;
1071 dfrag->data_seq = msk->write_seq;
1072 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1073 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1074 dfrag->already_sent = 0;
1075 dfrag->page = pfrag->page;
1080 struct mptcp_sendmsg_info {
1086 bool data_lock_held;
1089 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1090 u64 data_seq, int avail_size)
1092 u64 window_end = mptcp_wnd_end(msk);
1095 if (__mptcp_check_fallback(msk))
1098 mptcp_snd_wnd = window_end - data_seq;
1099 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1101 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1102 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1103 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1109 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1111 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1115 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1119 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1121 struct sk_buff *skb;
1123 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1125 if (likely(__mptcp_add_ext(skb, gfp))) {
1126 skb_reserve(skb, MAX_TCP_HEADER);
1127 skb->ip_summed = CHECKSUM_PARTIAL;
1128 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1133 mptcp_enter_memory_pressure(sk);
1138 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1140 struct sk_buff *skb;
1142 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1146 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1147 tcp_skb_entail(ssk, skb);
1150 tcp_skb_tsorted_anchor_cleanup(skb);
1155 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1157 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1159 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1162 /* note: this always recompute the csum on the whole skb, even
1163 * if we just appended a single frag. More status info needed
1165 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1167 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1168 __wsum csum = ~csum_unfold(mpext->csum);
1169 int offset = skb->len - added;
1171 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1174 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1176 struct mptcp_ext *mpext)
1181 mpext->infinite_map = 1;
1182 mpext->data_len = 0;
1184 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1185 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1187 mptcp_do_fallback(ssk);
1190 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1191 struct mptcp_data_frag *dfrag,
1192 struct mptcp_sendmsg_info *info)
1194 u64 data_seq = dfrag->data_seq + info->sent;
1195 int offset = dfrag->offset + info->sent;
1196 struct mptcp_sock *msk = mptcp_sk(sk);
1197 bool zero_window_probe = false;
1198 struct mptcp_ext *mpext = NULL;
1199 bool can_coalesce = false;
1200 bool reuse_skb = true;
1201 struct sk_buff *skb;
1205 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1206 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1208 if (WARN_ON_ONCE(info->sent > info->limit ||
1209 info->limit > dfrag->data_len))
1212 if (unlikely(!__tcp_can_send(ssk)))
1215 /* compute send limit */
1216 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1217 copy = info->size_goal;
1219 skb = tcp_write_queue_tail(ssk);
1220 if (skb && copy > skb->len) {
1221 /* Limit the write to the size available in the
1222 * current skb, if any, so that we create at most a new skb.
1223 * Explicitly tells TCP internals to avoid collapsing on later
1224 * queue management operation, to avoid breaking the ext <->
1225 * SSN association set here
1227 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1228 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1229 TCP_SKB_CB(skb)->eor = 1;
1233 i = skb_shinfo(skb)->nr_frags;
1234 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1235 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1236 tcp_mark_push(tcp_sk(ssk), skb);
1243 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1247 i = skb_shinfo(skb)->nr_frags;
1249 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1252 /* Zero window and all data acked? Probe. */
1253 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1255 u64 snd_una = READ_ONCE(msk->snd_una);
1257 if (snd_una != msk->snd_nxt) {
1258 tcp_remove_empty_skb(ssk);
1262 zero_window_probe = true;
1263 data_seq = snd_una - 1;
1266 /* all mptcp-level data is acked, no skbs should be present into the
1269 WARN_ON_ONCE(reuse_skb);
1272 copy = min_t(size_t, copy, info->limit - info->sent);
1273 if (!sk_wmem_schedule(ssk, copy)) {
1274 tcp_remove_empty_skb(ssk);
1279 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1281 get_page(dfrag->page);
1282 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1286 skb->data_len += copy;
1287 skb->truesize += copy;
1288 sk_wmem_queued_add(ssk, copy);
1289 sk_mem_charge(ssk, copy);
1290 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1291 TCP_SKB_CB(skb)->end_seq += copy;
1292 tcp_skb_pcount_set(skb, 0);
1294 /* on skb reuse we just need to update the DSS len */
1296 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1297 mpext->data_len += copy;
1298 WARN_ON_ONCE(zero_window_probe);
1302 memset(mpext, 0, sizeof(*mpext));
1303 mpext->data_seq = data_seq;
1304 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1305 mpext->data_len = copy;
1309 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1310 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1313 if (zero_window_probe) {
1314 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1316 if (READ_ONCE(msk->csum_enabled))
1317 mptcp_update_data_checksum(skb, copy);
1318 tcp_push_pending_frames(ssk);
1322 if (READ_ONCE(msk->csum_enabled))
1323 mptcp_update_data_checksum(skb, copy);
1324 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1325 mptcp_update_infinite_map(msk, ssk, mpext);
1326 trace_mptcp_sendmsg_frag(mpext);
1327 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1331 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1332 sizeof(struct tcphdr) - \
1333 MAX_TCP_OPTION_SPACE - \
1334 sizeof(struct ipv6hdr) - \
1335 sizeof(struct frag_hdr))
1337 struct subflow_send_info {
1342 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1344 if (!subflow->stale)
1348 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1351 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1353 if (unlikely(subflow->stale)) {
1354 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1356 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1359 mptcp_subflow_set_active(subflow);
1361 return __mptcp_subflow_active(subflow);
1364 #define SSK_MODE_ACTIVE 0
1365 #define SSK_MODE_BACKUP 1
1366 #define SSK_MODE_MAX 2
1368 /* implement the mptcp packet scheduler;
1369 * returns the subflow that will transmit the next DSS
1370 * additionally updates the rtx timeout
1372 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1374 struct subflow_send_info send_info[SSK_MODE_MAX];
1375 struct mptcp_subflow_context *subflow;
1376 struct sock *sk = (struct sock *)msk;
1377 u32 pace, burst, wmem;
1378 int i, nr_active = 0;
1383 msk_owned_by_me(msk);
1385 if (__mptcp_check_fallback(msk)) {
1388 return __tcp_can_send(msk->first) &&
1389 sk_stream_memory_free(msk->first) ? msk->first : NULL;
1392 /* re-use last subflow, if the burst allow that */
1393 if (msk->last_snd && msk->snd_burst > 0 &&
1394 sk_stream_memory_free(msk->last_snd) &&
1395 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1396 mptcp_set_timeout(sk);
1397 return msk->last_snd;
1400 /* pick the subflow with the lower wmem/wspace ratio */
1401 for (i = 0; i < SSK_MODE_MAX; ++i) {
1402 send_info[i].ssk = NULL;
1403 send_info[i].linger_time = -1;
1406 mptcp_for_each_subflow(msk, subflow) {
1407 trace_mptcp_subflow_get_send(subflow);
1408 ssk = mptcp_subflow_tcp_sock(subflow);
1409 if (!mptcp_subflow_active(subflow))
1412 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1413 nr_active += !subflow->backup;
1414 pace = subflow->avg_pacing_rate;
1415 if (unlikely(!pace)) {
1416 /* init pacing rate from socket */
1417 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1418 pace = subflow->avg_pacing_rate;
1423 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1424 if (linger_time < send_info[subflow->backup].linger_time) {
1425 send_info[subflow->backup].ssk = ssk;
1426 send_info[subflow->backup].linger_time = linger_time;
1429 __mptcp_set_timeout(sk, tout);
1431 /* pick the best backup if no other subflow is active */
1433 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1435 /* According to the blest algorithm, to avoid HoL blocking for the
1436 * faster flow, we need to:
1437 * - estimate the faster flow linger time
1438 * - use the above to estimate the amount of byte transferred
1439 * by the faster flow
1440 * - check that the amount of queued data is greter than the above,
1441 * otherwise do not use the picked, slower, subflow
1442 * We select the subflow with the shorter estimated time to flush
1443 * the queued mem, which basically ensure the above. We just need
1444 * to check that subflow has a non empty cwin.
1446 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1447 if (!ssk || !sk_stream_memory_free(ssk))
1450 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1451 wmem = READ_ONCE(ssk->sk_wmem_queued);
1453 msk->last_snd = NULL;
1457 subflow = mptcp_subflow_ctx(ssk);
1458 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1459 READ_ONCE(ssk->sk_pacing_rate) * burst,
1461 msk->last_snd = ssk;
1462 msk->snd_burst = burst;
1466 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1468 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1472 static void mptcp_update_post_push(struct mptcp_sock *msk,
1473 struct mptcp_data_frag *dfrag,
1476 u64 snd_nxt_new = dfrag->data_seq;
1478 dfrag->already_sent += sent;
1480 msk->snd_burst -= sent;
1482 snd_nxt_new += dfrag->already_sent;
1484 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1485 * is recovering after a failover. In that event, this re-sends
1488 * Thus compute snd_nxt_new candidate based on
1489 * the dfrag->data_seq that was sent and the data
1490 * that has been handed to the subflow for transmission
1491 * and skip update in case it was old dfrag.
1493 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1494 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1495 msk->snd_nxt = snd_nxt_new;
1499 void mptcp_check_and_set_pending(struct sock *sk)
1501 if (mptcp_send_head(sk))
1502 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1505 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1507 struct sock *prev_ssk = NULL, *ssk = NULL;
1508 struct mptcp_sock *msk = mptcp_sk(sk);
1509 struct mptcp_sendmsg_info info = {
1512 bool do_check_data_fin = false;
1513 struct mptcp_data_frag *dfrag;
1516 while ((dfrag = mptcp_send_head(sk))) {
1517 info.sent = dfrag->already_sent;
1518 info.limit = dfrag->data_len;
1519 len = dfrag->data_len - dfrag->already_sent;
1524 ssk = mptcp_subflow_get_send(msk);
1526 /* First check. If the ssk has changed since
1527 * the last round, release prev_ssk
1529 if (ssk != prev_ssk && prev_ssk)
1530 mptcp_push_release(prev_ssk, &info);
1534 /* Need to lock the new subflow only if different
1535 * from the previous one, otherwise we are still
1536 * helding the relevant lock
1538 if (ssk != prev_ssk)
1541 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1545 mptcp_push_release(ssk, &info);
1549 do_check_data_fin = true;
1553 mptcp_update_post_push(msk, dfrag, ret);
1555 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1558 /* at this point we held the socket lock for the last subflow we used */
1560 mptcp_push_release(ssk, &info);
1563 /* ensure the rtx timer is running */
1564 if (!mptcp_timer_pending(sk))
1565 mptcp_reset_timer(sk);
1566 if (do_check_data_fin)
1567 mptcp_check_send_data_fin(sk);
1570 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1572 struct mptcp_sock *msk = mptcp_sk(sk);
1573 struct mptcp_sendmsg_info info = {
1574 .data_lock_held = true,
1576 struct mptcp_data_frag *dfrag;
1577 struct sock *xmit_ssk;
1578 int len, copied = 0;
1581 while ((dfrag = mptcp_send_head(sk))) {
1582 info.sent = dfrag->already_sent;
1583 info.limit = dfrag->data_len;
1584 len = dfrag->data_len - dfrag->already_sent;
1588 /* check for a different subflow usage only after
1589 * spooling the first chunk of data
1591 xmit_ssk = first ? ssk : mptcp_subflow_get_send(msk);
1594 if (xmit_ssk != ssk) {
1595 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1596 MPTCP_DELEGATE_SEND);
1600 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1609 mptcp_update_post_push(msk, dfrag, ret);
1611 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1615 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1616 * not going to flush it via release_sock()
1619 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1621 if (!mptcp_timer_pending(sk))
1622 mptcp_reset_timer(sk);
1624 if (msk->snd_data_fin_enable &&
1625 msk->snd_nxt + 1 == msk->write_seq)
1626 mptcp_schedule_work(sk);
1630 static void mptcp_set_nospace(struct sock *sk)
1632 /* enable autotune */
1633 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1635 /* will be cleared on avail space */
1636 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1639 static int mptcp_disconnect(struct sock *sk, int flags);
1641 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1642 size_t len, int *copied_syn)
1644 unsigned int saved_flags = msg->msg_flags;
1645 struct mptcp_sock *msk = mptcp_sk(sk);
1646 struct socket *ssock;
1650 /* on flags based fastopen the mptcp is supposed to create the
1651 * first subflow right now. Otherwise we are in the defer_connect
1652 * path, and the first subflow must be already present.
1653 * Since the defer_connect flag is cleared after the first succsful
1654 * fastopen attempt, no need to check for additional subflow status.
1656 if (msg->msg_flags & MSG_FASTOPEN) {
1657 ssock = __mptcp_nmpc_socket(msk);
1659 return PTR_ERR(ssock);
1667 msg->msg_flags |= MSG_DONTWAIT;
1668 msk->fastopening = 1;
1669 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1670 msk->fastopening = 0;
1671 msg->msg_flags = saved_flags;
1674 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1675 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1676 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1677 msg->msg_namelen, msg->msg_flags, 1);
1679 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1680 * case of any error, except timeout or signal
1682 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1684 } else if (ret && ret != -EINPROGRESS) {
1685 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1686 * __inet_stream_connect() can fail, due to looking check,
1687 * see mptcp_disconnect().
1688 * Attempt it again outside the problematic scope.
1690 if (!mptcp_disconnect(sk, 0))
1691 sk->sk_socket->state = SS_UNCONNECTED;
1693 inet_sk(sk)->defer_connect = 0;
1698 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1700 struct mptcp_sock *msk = mptcp_sk(sk);
1701 struct page_frag *pfrag;
1706 /* silently ignore everything else */
1707 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1711 if (unlikely(inet_sk(sk)->defer_connect || msg->msg_flags & MSG_FASTOPEN)) {
1714 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1715 copied += copied_syn;
1716 if (ret == -EINPROGRESS && copied_syn > 0)
1722 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1724 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1725 ret = sk_stream_wait_connect(sk, &timeo);
1731 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1734 pfrag = sk_page_frag(sk);
1736 while (msg_data_left(msg)) {
1737 int total_ts, frag_truesize = 0;
1738 struct mptcp_data_frag *dfrag;
1739 bool dfrag_collapsed;
1740 size_t psize, offset;
1742 /* reuse tail pfrag, if possible, or carve a new one from the
1745 dfrag = mptcp_pending_tail(sk);
1746 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1747 if (!dfrag_collapsed) {
1748 if (!sk_stream_memory_free(sk))
1749 goto wait_for_memory;
1751 if (!mptcp_page_frag_refill(sk, pfrag))
1752 goto wait_for_memory;
1754 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1755 frag_truesize = dfrag->overhead;
1758 /* we do not bound vs wspace, to allow a single packet.
1759 * memory accounting will prevent execessive memory usage
1762 offset = dfrag->offset + dfrag->data_len;
1763 psize = pfrag->size - offset;
1764 psize = min_t(size_t, psize, msg_data_left(msg));
1765 total_ts = psize + frag_truesize;
1767 if (!sk_wmem_schedule(sk, total_ts))
1768 goto wait_for_memory;
1770 if (copy_page_from_iter(dfrag->page, offset, psize,
1771 &msg->msg_iter) != psize) {
1776 /* data successfully copied into the write queue */
1777 sk->sk_forward_alloc -= total_ts;
1779 dfrag->data_len += psize;
1780 frag_truesize += psize;
1781 pfrag->offset += frag_truesize;
1782 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1784 /* charge data on mptcp pending queue to the msk socket
1785 * Note: we charge such data both to sk and ssk
1787 sk_wmem_queued_add(sk, frag_truesize);
1788 if (!dfrag_collapsed) {
1789 get_page(dfrag->page);
1790 list_add_tail(&dfrag->list, &msk->rtx_queue);
1791 if (!msk->first_pending)
1792 WRITE_ONCE(msk->first_pending, dfrag);
1794 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1795 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1801 mptcp_set_nospace(sk);
1802 __mptcp_push_pending(sk, msg->msg_flags);
1803 ret = sk_stream_wait_memory(sk, &timeo);
1809 __mptcp_push_pending(sk, msg->msg_flags);
1819 copied = sk_stream_error(sk, msg->msg_flags, ret);
1823 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1825 size_t len, int flags,
1826 struct scm_timestamping_internal *tss,
1829 struct sk_buff *skb, *tmp;
1832 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1833 u32 offset = MPTCP_SKB_CB(skb)->offset;
1834 u32 data_len = skb->len - offset;
1835 u32 count = min_t(size_t, len - copied, data_len);
1838 if (!(flags & MSG_TRUNC)) {
1839 err = skb_copy_datagram_msg(skb, offset, msg, count);
1840 if (unlikely(err < 0)) {
1847 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1848 tcp_update_recv_tstamps(skb, tss);
1849 *cmsg_flags |= MPTCP_CMSG_TS;
1854 if (count < data_len) {
1855 if (!(flags & MSG_PEEK)) {
1856 MPTCP_SKB_CB(skb)->offset += count;
1857 MPTCP_SKB_CB(skb)->map_seq += count;
1862 if (!(flags & MSG_PEEK)) {
1863 /* we will bulk release the skb memory later */
1864 skb->destructor = NULL;
1865 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1866 __skb_unlink(skb, &msk->receive_queue);
1877 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1879 * Only difference: Use highest rtt estimate of the subflows in use.
1881 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1883 struct mptcp_subflow_context *subflow;
1884 struct sock *sk = (struct sock *)msk;
1885 u8 scaling_ratio = U8_MAX;
1886 u32 time, advmss = 1;
1889 msk_owned_by_me(msk);
1894 msk->rcvq_space.copied += copied;
1896 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1897 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1899 rtt_us = msk->rcvq_space.rtt_us;
1900 if (rtt_us && time < (rtt_us >> 3))
1904 mptcp_for_each_subflow(msk, subflow) {
1905 const struct tcp_sock *tp;
1909 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1911 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1912 sf_advmss = READ_ONCE(tp->advmss);
1914 rtt_us = max(sf_rtt_us, rtt_us);
1915 advmss = max(sf_advmss, advmss);
1916 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1919 msk->rcvq_space.rtt_us = rtt_us;
1920 msk->scaling_ratio = scaling_ratio;
1921 if (time < (rtt_us >> 3) || rtt_us == 0)
1924 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1927 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1928 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1932 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1934 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1936 do_div(grow, msk->rcvq_space.space);
1937 rcvwin += (grow << 1);
1939 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
1940 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1942 if (rcvbuf > sk->sk_rcvbuf) {
1945 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
1946 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1948 /* Make subflows follow along. If we do not do this, we
1949 * get drops at subflow level if skbs can't be moved to
1950 * the mptcp rx queue fast enough (announced rcv_win can
1951 * exceed ssk->sk_rcvbuf).
1953 mptcp_for_each_subflow(msk, subflow) {
1957 ssk = mptcp_subflow_tcp_sock(subflow);
1958 slow = lock_sock_fast(ssk);
1959 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1960 tcp_sk(ssk)->window_clamp = window_clamp;
1961 tcp_cleanup_rbuf(ssk, 1);
1962 unlock_sock_fast(ssk, slow);
1967 msk->rcvq_space.space = msk->rcvq_space.copied;
1969 msk->rcvq_space.copied = 0;
1970 msk->rcvq_space.time = mstamp;
1973 static void __mptcp_update_rmem(struct sock *sk)
1975 struct mptcp_sock *msk = mptcp_sk(sk);
1977 if (!msk->rmem_released)
1980 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1981 mptcp_rmem_uncharge(sk, msk->rmem_released);
1982 WRITE_ONCE(msk->rmem_released, 0);
1985 static void __mptcp_splice_receive_queue(struct sock *sk)
1987 struct mptcp_sock *msk = mptcp_sk(sk);
1989 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1992 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1994 struct sock *sk = (struct sock *)msk;
1995 unsigned int moved = 0;
1999 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2002 /* we can have data pending in the subflows only if the msk
2003 * receive buffer was full at subflow_data_ready() time,
2004 * that is an unlikely slow path.
2009 slowpath = lock_sock_fast(ssk);
2010 mptcp_data_lock(sk);
2011 __mptcp_update_rmem(sk);
2012 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2013 mptcp_data_unlock(sk);
2015 if (unlikely(ssk->sk_err))
2016 __mptcp_error_report(sk);
2017 unlock_sock_fast(ssk, slowpath);
2020 /* acquire the data lock only if some input data is pending */
2022 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2023 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2024 mptcp_data_lock(sk);
2025 __mptcp_update_rmem(sk);
2026 ret |= __mptcp_ofo_queue(msk);
2027 __mptcp_splice_receive_queue(sk);
2028 mptcp_data_unlock(sk);
2031 mptcp_check_data_fin((struct sock *)msk);
2032 return !skb_queue_empty(&msk->receive_queue);
2035 static unsigned int mptcp_inq_hint(const struct sock *sk)
2037 const struct mptcp_sock *msk = mptcp_sk(sk);
2038 const struct sk_buff *skb;
2040 skb = skb_peek(&msk->receive_queue);
2042 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2044 if (hint_val >= INT_MAX)
2047 return (unsigned int)hint_val;
2050 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2056 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2057 int flags, int *addr_len)
2059 struct mptcp_sock *msk = mptcp_sk(sk);
2060 struct scm_timestamping_internal tss;
2061 int copied = 0, cmsg_flags = 0;
2065 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2066 if (unlikely(flags & MSG_ERRQUEUE))
2067 return inet_recv_error(sk, msg, len, addr_len);
2070 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2075 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2077 len = min_t(size_t, len, INT_MAX);
2078 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2080 if (unlikely(msk->recvmsg_inq))
2081 cmsg_flags = MPTCP_CMSG_INQ;
2083 while (copied < len) {
2086 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2087 if (unlikely(bytes_read < 0)) {
2089 copied = bytes_read;
2093 copied += bytes_read;
2095 /* be sure to advertise window change */
2096 mptcp_cleanup_rbuf(msk);
2098 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2101 /* only the master socket status is relevant here. The exit
2102 * conditions mirror closely tcp_recvmsg()
2104 if (copied >= target)
2109 sk->sk_state == TCP_CLOSE ||
2110 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2112 signal_pending(current))
2116 copied = sock_error(sk);
2120 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2121 /* race breaker: the shutdown could be after the
2122 * previous receive queue check
2124 if (__mptcp_move_skbs(msk))
2129 if (sk->sk_state == TCP_CLOSE) {
2139 if (signal_pending(current)) {
2140 copied = sock_intr_errno(timeo);
2145 pr_debug("block timeout %ld", timeo);
2146 sk_wait_data(sk, &timeo, NULL);
2150 if (cmsg_flags && copied >= 0) {
2151 if (cmsg_flags & MPTCP_CMSG_TS)
2152 tcp_recv_timestamp(msg, sk, &tss);
2154 if (cmsg_flags & MPTCP_CMSG_INQ) {
2155 unsigned int inq = mptcp_inq_hint(sk);
2157 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2161 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2162 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2163 skb_queue_empty(&msk->receive_queue), copied);
2164 if (!(flags & MSG_PEEK))
2165 mptcp_rcv_space_adjust(msk, copied);
2171 static void mptcp_retransmit_timer(struct timer_list *t)
2173 struct inet_connection_sock *icsk = from_timer(icsk, t,
2174 icsk_retransmit_timer);
2175 struct sock *sk = &icsk->icsk_inet.sk;
2176 struct mptcp_sock *msk = mptcp_sk(sk);
2179 if (!sock_owned_by_user(sk)) {
2180 /* we need a process context to retransmit */
2181 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2182 mptcp_schedule_work(sk);
2184 /* delegate our work to tcp_release_cb() */
2185 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2191 static void mptcp_timeout_timer(struct timer_list *t)
2193 struct sock *sk = from_timer(sk, t, sk_timer);
2195 mptcp_schedule_work(sk);
2199 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2202 * A backup subflow is returned only if that is the only kind available.
2204 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2206 struct sock *backup = NULL, *pick = NULL;
2207 struct mptcp_subflow_context *subflow;
2208 int min_stale_count = INT_MAX;
2210 msk_owned_by_me(msk);
2212 if (__mptcp_check_fallback(msk))
2215 mptcp_for_each_subflow(msk, subflow) {
2216 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2218 if (!__mptcp_subflow_active(subflow))
2221 /* still data outstanding at TCP level? skip this */
2222 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2223 mptcp_pm_subflow_chk_stale(msk, ssk);
2224 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2228 if (subflow->backup) {
2241 /* use backup only if there are no progresses anywhere */
2242 return min_stale_count > 1 ? backup : NULL;
2245 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2248 iput(SOCK_INODE(msk->subflow));
2249 WRITE_ONCE(msk->subflow, NULL);
2253 bool __mptcp_retransmit_pending_data(struct sock *sk)
2255 struct mptcp_data_frag *cur, *rtx_head;
2256 struct mptcp_sock *msk = mptcp_sk(sk);
2258 if (__mptcp_check_fallback(msk))
2261 if (tcp_rtx_and_write_queues_empty(sk))
2264 /* the closing socket has some data untransmitted and/or unacked:
2265 * some data in the mptcp rtx queue has not really xmitted yet.
2266 * keep it simple and re-inject the whole mptcp level rtx queue
2268 mptcp_data_lock(sk);
2269 __mptcp_clean_una_wakeup(sk);
2270 rtx_head = mptcp_rtx_head(sk);
2272 mptcp_data_unlock(sk);
2276 msk->recovery_snd_nxt = msk->snd_nxt;
2277 msk->recovery = true;
2278 mptcp_data_unlock(sk);
2280 msk->first_pending = rtx_head;
2283 /* be sure to clear the "sent status" on all re-injected fragments */
2284 list_for_each_entry(cur, &msk->rtx_queue, list) {
2285 if (!cur->already_sent)
2287 cur->already_sent = 0;
2293 /* flags for __mptcp_close_ssk() */
2294 #define MPTCP_CF_PUSH BIT(1)
2295 #define MPTCP_CF_FASTCLOSE BIT(2)
2297 /* subflow sockets can be either outgoing (connect) or incoming
2300 * Outgoing subflows use in-kernel sockets.
2301 * Incoming subflows do not have their own 'struct socket' allocated,
2302 * so we need to use tcp_close() after detaching them from the mptcp
2305 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2306 struct mptcp_subflow_context *subflow,
2309 struct mptcp_sock *msk = mptcp_sk(sk);
2310 bool dispose_it, need_push = false;
2312 /* If the first subflow moved to a close state before accept, e.g. due
2313 * to an incoming reset, mptcp either:
2314 * - if either the subflow or the msk are dead, destroy the context
2315 * (the subflow socket is deleted by inet_child_forget) and the msk
2316 * - otherwise do nothing at the moment and take action at accept and/or
2317 * listener shutdown - user-space must be able to accept() the closed
2320 if (msk->in_accept_queue && msk->first == ssk) {
2321 if (!sock_flag(sk, SOCK_DEAD) && !sock_flag(ssk, SOCK_DEAD))
2324 /* ensure later check in mptcp_worker() will dispose the msk */
2325 sock_set_flag(sk, SOCK_DEAD);
2326 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2327 mptcp_subflow_drop_ctx(ssk);
2331 dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2333 list_del(&subflow->node);
2335 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2337 if (flags & MPTCP_CF_FASTCLOSE) {
2338 /* be sure to force the tcp_disconnect() path,
2339 * to generate the egress reset
2341 ssk->sk_lingertime = 0;
2342 sock_set_flag(ssk, SOCK_LINGER);
2343 subflow->send_fastclose = 1;
2346 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2348 /* The MPTCP code never wait on the subflow sockets, TCP-level
2349 * disconnect should never fail
2351 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2352 msk->subflow->state = SS_UNCONNECTED;
2353 mptcp_subflow_ctx_reset(subflow);
2359 subflow->disposable = 1;
2361 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2362 * the ssk has been already destroyed, we just need to release the
2363 * reference owned by msk;
2365 if (!inet_csk(ssk)->icsk_ulp_ops) {
2366 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2367 kfree_rcu(subflow, rcu);
2369 /* otherwise tcp will dispose of the ssk and subflow ctx */
2370 __tcp_close(ssk, 0);
2372 /* close acquired an extra ref */
2381 if (ssk == msk->first)
2382 WRITE_ONCE(msk->first, NULL);
2385 if (ssk == msk->last_snd)
2386 msk->last_snd = NULL;
2389 __mptcp_push_pending(sk, 0);
2392 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2393 struct mptcp_subflow_context *subflow)
2395 if (sk->sk_state == TCP_ESTABLISHED)
2396 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2398 /* subflow aborted before reaching the fully_established status
2399 * attempt the creation of the next subflow
2401 mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2403 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2406 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2411 static void __mptcp_close_subflow(struct sock *sk)
2413 struct mptcp_subflow_context *subflow, *tmp;
2414 struct mptcp_sock *msk = mptcp_sk(sk);
2418 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2419 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2421 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2424 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2425 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2428 mptcp_close_ssk(sk, ssk, subflow);
2433 static bool mptcp_should_close(const struct sock *sk)
2435 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2436 struct mptcp_subflow_context *subflow;
2438 if (delta >= TCP_TIMEWAIT_LEN || mptcp_sk(sk)->in_accept_queue)
2441 /* if all subflows are in closed status don't bother with additional
2444 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2445 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2452 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2454 struct mptcp_subflow_context *subflow, *tmp;
2455 struct sock *sk = (struct sock *)msk;
2457 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2460 mptcp_token_destroy(msk);
2462 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2463 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2466 slow = lock_sock_fast(tcp_sk);
2467 if (tcp_sk->sk_state != TCP_CLOSE) {
2468 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2469 tcp_set_state(tcp_sk, TCP_CLOSE);
2471 unlock_sock_fast(tcp_sk, slow);
2474 /* Mirror the tcp_reset() error propagation */
2475 switch (sk->sk_state) {
2477 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2479 case TCP_CLOSE_WAIT:
2480 WRITE_ONCE(sk->sk_err, EPIPE);
2485 WRITE_ONCE(sk->sk_err, ECONNRESET);
2488 inet_sk_state_store(sk, TCP_CLOSE);
2489 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2490 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2491 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2493 /* the calling mptcp_worker will properly destroy the socket */
2494 if (sock_flag(sk, SOCK_DEAD))
2497 sk->sk_state_change(sk);
2498 sk_error_report(sk);
2501 static void __mptcp_retrans(struct sock *sk)
2503 struct mptcp_sock *msk = mptcp_sk(sk);
2504 struct mptcp_sendmsg_info info = {};
2505 struct mptcp_data_frag *dfrag;
2510 mptcp_clean_una_wakeup(sk);
2512 /* first check ssk: need to kick "stale" logic */
2513 ssk = mptcp_subflow_get_retrans(msk);
2514 dfrag = mptcp_rtx_head(sk);
2516 if (mptcp_data_fin_enabled(msk)) {
2517 struct inet_connection_sock *icsk = inet_csk(sk);
2519 icsk->icsk_retransmits++;
2520 mptcp_set_datafin_timeout(sk);
2521 mptcp_send_ack(msk);
2526 if (!mptcp_send_head(sk))
2537 /* limit retransmission to the bytes already sent on some subflows */
2539 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2540 while (info.sent < info.limit) {
2541 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2545 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2550 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2551 msk->bytes_retrans += copied;
2552 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2554 WRITE_ONCE(msk->allow_infinite_fallback, false);
2560 mptcp_check_and_set_pending(sk);
2562 if (!mptcp_timer_pending(sk))
2563 mptcp_reset_timer(sk);
2566 /* schedule the timeout timer for the relevant event: either close timeout
2567 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2569 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2571 struct sock *sk = (struct sock *)msk;
2572 unsigned long timeout, close_timeout;
2574 if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2577 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2579 /* the close timeout takes precedence on the fail one, and here at least one of
2582 timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2584 sk_reset_timer(sk, &sk->sk_timer, timeout);
2587 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2589 struct sock *ssk = msk->first;
2595 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2597 slow = lock_sock_fast(ssk);
2598 mptcp_subflow_reset(ssk);
2599 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2600 unlock_sock_fast(ssk, slow);
2602 mptcp_reset_timeout(msk, 0);
2605 static void mptcp_do_fastclose(struct sock *sk)
2607 struct mptcp_subflow_context *subflow, *tmp;
2608 struct mptcp_sock *msk = mptcp_sk(sk);
2610 inet_sk_state_store(sk, TCP_CLOSE);
2611 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2612 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2613 subflow, MPTCP_CF_FASTCLOSE);
2616 static void mptcp_worker(struct work_struct *work)
2618 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2619 struct sock *sk = (struct sock *)msk;
2620 unsigned long fail_tout;
2624 state = sk->sk_state;
2625 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2628 mptcp_check_fastclose(msk);
2630 mptcp_pm_nl_work(msk);
2632 mptcp_check_send_data_fin(sk);
2633 mptcp_check_data_fin_ack(sk);
2634 mptcp_check_data_fin(sk);
2636 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2637 __mptcp_close_subflow(sk);
2639 /* There is no point in keeping around an orphaned sk timedout or
2640 * closed, but we need the msk around to reply to incoming DATA_FIN,
2641 * even if it is orphaned and in FIN_WAIT2 state
2643 if (sock_flag(sk, SOCK_DEAD)) {
2644 if (mptcp_should_close(sk))
2645 mptcp_do_fastclose(sk);
2647 if (sk->sk_state == TCP_CLOSE) {
2648 __mptcp_destroy_sock(sk);
2653 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2654 __mptcp_retrans(sk);
2656 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2657 if (fail_tout && time_after(jiffies, fail_tout))
2658 mptcp_mp_fail_no_response(msk);
2665 static int __mptcp_init_sock(struct sock *sk)
2667 struct mptcp_sock *msk = mptcp_sk(sk);
2669 INIT_LIST_HEAD(&msk->conn_list);
2670 INIT_LIST_HEAD(&msk->join_list);
2671 INIT_LIST_HEAD(&msk->rtx_queue);
2672 INIT_WORK(&msk->work, mptcp_worker);
2673 __skb_queue_head_init(&msk->receive_queue);
2674 msk->out_of_order_queue = RB_ROOT;
2675 msk->first_pending = NULL;
2676 msk->rmem_fwd_alloc = 0;
2677 WRITE_ONCE(msk->rmem_released, 0);
2678 msk->timer_ival = TCP_RTO_MIN;
2680 WRITE_ONCE(msk->first, NULL);
2681 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2682 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2683 WRITE_ONCE(msk->allow_infinite_fallback, true);
2684 msk->recovery = false;
2685 msk->subflow_id = 1;
2687 mptcp_pm_data_init(msk);
2689 /* re-use the csk retrans timer for MPTCP-level retrans */
2690 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2691 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2696 static void mptcp_ca_reset(struct sock *sk)
2698 struct inet_connection_sock *icsk = inet_csk(sk);
2700 tcp_assign_congestion_control(sk);
2701 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2703 /* no need to keep a reference to the ops, the name will suffice */
2704 tcp_cleanup_congestion_control(sk);
2705 icsk->icsk_ca_ops = NULL;
2708 static int mptcp_init_sock(struct sock *sk)
2710 struct net *net = sock_net(sk);
2713 ret = __mptcp_init_sock(sk);
2717 if (!mptcp_is_enabled(net))
2718 return -ENOPROTOOPT;
2720 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2723 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2725 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2726 * propagate the correct value
2730 sk_sockets_allocated_inc(sk);
2731 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2732 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2737 static void __mptcp_clear_xmit(struct sock *sk)
2739 struct mptcp_sock *msk = mptcp_sk(sk);
2740 struct mptcp_data_frag *dtmp, *dfrag;
2742 WRITE_ONCE(msk->first_pending, NULL);
2743 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2744 dfrag_clear(sk, dfrag);
2747 void mptcp_cancel_work(struct sock *sk)
2749 struct mptcp_sock *msk = mptcp_sk(sk);
2751 if (cancel_work_sync(&msk->work))
2755 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2759 switch (ssk->sk_state) {
2761 if (!(how & RCV_SHUTDOWN))
2765 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2768 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2769 pr_debug("Fallback");
2770 ssk->sk_shutdown |= how;
2771 tcp_shutdown(ssk, how);
2773 /* simulate the data_fin ack reception to let the state
2774 * machine move forward
2776 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2777 mptcp_schedule_work(sk);
2779 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2781 if (!mptcp_timer_pending(sk))
2782 mptcp_reset_timer(sk);
2790 static const unsigned char new_state[16] = {
2791 /* current state: new state: action: */
2792 [0 /* (Invalid) */] = TCP_CLOSE,
2793 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2794 [TCP_SYN_SENT] = TCP_CLOSE,
2795 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2796 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2797 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2798 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2799 [TCP_CLOSE] = TCP_CLOSE,
2800 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2801 [TCP_LAST_ACK] = TCP_LAST_ACK,
2802 [TCP_LISTEN] = TCP_CLOSE,
2803 [TCP_CLOSING] = TCP_CLOSING,
2804 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2807 static int mptcp_close_state(struct sock *sk)
2809 int next = (int)new_state[sk->sk_state];
2810 int ns = next & TCP_STATE_MASK;
2812 inet_sk_state_store(sk, ns);
2814 return next & TCP_ACTION_FIN;
2817 static void mptcp_check_send_data_fin(struct sock *sk)
2819 struct mptcp_subflow_context *subflow;
2820 struct mptcp_sock *msk = mptcp_sk(sk);
2822 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2823 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2824 msk->snd_nxt, msk->write_seq);
2826 /* we still need to enqueue subflows or not really shutting down,
2829 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2830 mptcp_send_head(sk))
2833 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2835 mptcp_for_each_subflow(msk, subflow) {
2836 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2838 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2842 static void __mptcp_wr_shutdown(struct sock *sk)
2844 struct mptcp_sock *msk = mptcp_sk(sk);
2846 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2847 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2848 !!mptcp_send_head(sk));
2850 /* will be ignored by fallback sockets */
2851 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2852 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2854 mptcp_check_send_data_fin(sk);
2857 static void __mptcp_destroy_sock(struct sock *sk)
2859 struct mptcp_sock *msk = mptcp_sk(sk);
2861 pr_debug("msk=%p", msk);
2865 mptcp_stop_timer(sk);
2866 sk_stop_timer(sk, &sk->sk_timer);
2869 sk->sk_prot->destroy(sk);
2871 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2872 WARN_ON_ONCE(msk->rmem_released);
2873 sk_stream_kill_queues(sk);
2874 xfrm_sk_free_policy(sk);
2879 void __mptcp_unaccepted_force_close(struct sock *sk)
2881 sock_set_flag(sk, SOCK_DEAD);
2882 mptcp_do_fastclose(sk);
2883 __mptcp_destroy_sock(sk);
2886 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2888 /* Concurrent splices from sk_receive_queue into receive_queue will
2889 * always show at least one non-empty queue when checked in this order.
2891 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2892 skb_queue_empty_lockless(&msk->receive_queue))
2895 return EPOLLIN | EPOLLRDNORM;
2898 static void mptcp_check_listen_stop(struct sock *sk)
2902 if (inet_sk_state_load(sk) != TCP_LISTEN)
2905 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2906 ssk = mptcp_sk(sk)->first;
2907 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2910 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2911 tcp_set_state(ssk, TCP_CLOSE);
2912 mptcp_subflow_queue_clean(sk, ssk);
2913 inet_csk_listen_stop(ssk);
2914 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
2918 bool __mptcp_close(struct sock *sk, long timeout)
2920 struct mptcp_subflow_context *subflow;
2921 struct mptcp_sock *msk = mptcp_sk(sk);
2922 bool do_cancel_work = false;
2923 int subflows_alive = 0;
2925 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2927 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2928 mptcp_check_listen_stop(sk);
2929 inet_sk_state_store(sk, TCP_CLOSE);
2933 if (mptcp_check_readable(msk) || timeout < 0) {
2934 /* If the msk has read data, or the caller explicitly ask it,
2935 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
2937 mptcp_do_fastclose(sk);
2939 } else if (mptcp_close_state(sk)) {
2940 __mptcp_wr_shutdown(sk);
2943 sk_stream_wait_close(sk, timeout);
2946 /* orphan all the subflows */
2947 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2948 mptcp_for_each_subflow(msk, subflow) {
2949 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2950 bool slow = lock_sock_fast_nested(ssk);
2952 subflows_alive += ssk->sk_state != TCP_CLOSE;
2954 /* since the close timeout takes precedence on the fail one,
2957 if (ssk == msk->first)
2958 subflow->fail_tout = 0;
2960 /* detach from the parent socket, but allow data_ready to
2961 * push incoming data into the mptcp stack, to properly ack it
2963 ssk->sk_socket = NULL;
2965 unlock_sock_fast(ssk, slow);
2969 /* all the subflows are closed, only timeout can change the msk
2970 * state, let's not keep resources busy for no reasons
2972 if (subflows_alive == 0)
2973 inet_sk_state_store(sk, TCP_CLOSE);
2976 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2978 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2980 if (sk->sk_state == TCP_CLOSE) {
2981 __mptcp_destroy_sock(sk);
2982 do_cancel_work = true;
2984 mptcp_reset_timeout(msk, 0);
2987 return do_cancel_work;
2990 static void mptcp_close(struct sock *sk, long timeout)
2992 bool do_cancel_work;
2996 do_cancel_work = __mptcp_close(sk, timeout);
2999 mptcp_cancel_work(sk);
3004 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3006 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3007 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3008 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3010 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3011 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3014 msk6->saddr = ssk6->saddr;
3015 msk6->flow_label = ssk6->flow_label;
3019 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3020 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3021 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3022 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3023 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3024 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3027 static int mptcp_disconnect(struct sock *sk, int flags)
3029 struct mptcp_sock *msk = mptcp_sk(sk);
3031 /* Deny disconnect if other threads are blocked in sk_wait_event()
3032 * or inet_wait_for_connect().
3034 if (sk->sk_wait_pending)
3037 /* We are on the fastopen error path. We can't call straight into the
3038 * subflows cleanup code due to lock nesting (we are already under
3039 * msk->firstsocket lock).
3041 if (msk->fastopening)
3044 mptcp_check_listen_stop(sk);
3045 inet_sk_state_store(sk, TCP_CLOSE);
3047 mptcp_stop_timer(sk);
3048 sk_stop_timer(sk, &sk->sk_timer);
3051 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3053 /* msk->subflow is still intact, the following will not free the first
3056 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3057 msk->last_snd = NULL;
3058 WRITE_ONCE(msk->flags, 0);
3060 msk->push_pending = 0;
3061 msk->recovery = false;
3062 msk->can_ack = false;
3063 msk->fully_established = false;
3064 msk->rcv_data_fin = false;
3065 msk->snd_data_fin_enable = false;
3066 msk->rcv_fastclose = false;
3067 msk->use_64bit_ack = false;
3068 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3069 mptcp_pm_data_reset(msk);
3071 msk->bytes_acked = 0;
3072 msk->bytes_received = 0;
3073 msk->bytes_sent = 0;
3074 msk->bytes_retrans = 0;
3076 WRITE_ONCE(sk->sk_shutdown, 0);
3077 sk_error_report(sk);
3081 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3082 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3084 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3086 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3090 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3091 const struct mptcp_options_received *mp_opt,
3093 struct request_sock *req)
3095 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3096 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3097 struct mptcp_sock *msk;
3102 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3103 if (nsk->sk_family == AF_INET6)
3104 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3107 nsk->sk_wait_pending = 0;
3108 __mptcp_init_sock(nsk);
3110 msk = mptcp_sk(nsk);
3111 msk->local_key = subflow_req->local_key;
3112 msk->token = subflow_req->token;
3113 WRITE_ONCE(msk->subflow, NULL);
3114 msk->in_accept_queue = 1;
3115 WRITE_ONCE(msk->fully_established, false);
3116 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3117 WRITE_ONCE(msk->csum_enabled, true);
3119 msk->write_seq = subflow_req->idsn + 1;
3120 msk->snd_nxt = msk->write_seq;
3121 msk->snd_una = msk->write_seq;
3122 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3123 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3125 /* passive msk is created after the first/MPC subflow */
3126 msk->subflow_id = 2;
3128 sock_reset_flag(nsk, SOCK_RCU_FREE);
3129 security_inet_csk_clone(nsk, req);
3131 /* this can't race with mptcp_close(), as the msk is
3132 * not yet exposted to user-space
3134 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3136 /* The msk maintain a ref to each subflow in the connections list */
3137 WRITE_ONCE(msk->first, ssk);
3138 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3141 /* new mpc subflow takes ownership of the newly
3142 * created mptcp socket
3144 mptcp_token_accept(subflow_req, msk);
3146 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3147 * uses the correct data
3149 mptcp_copy_inaddrs(nsk, ssk);
3150 mptcp_propagate_sndbuf(nsk, ssk);
3152 mptcp_rcv_space_init(msk, ssk);
3153 bh_unlock_sock(nsk);
3155 /* note: the newly allocated socket refcount is 2 now */
3159 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3161 const struct tcp_sock *tp = tcp_sk(ssk);
3163 msk->rcvq_space.copied = 0;
3164 msk->rcvq_space.rtt_us = 0;
3166 msk->rcvq_space.time = tp->tcp_mstamp;
3168 /* initial rcv_space offering made to peer */
3169 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3170 TCP_INIT_CWND * tp->advmss);
3171 if (msk->rcvq_space.space == 0)
3172 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3174 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3177 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
3180 struct mptcp_sock *msk = mptcp_sk(sk);
3181 struct socket *listener;
3184 listener = READ_ONCE(msk->subflow);
3185 if (WARN_ON_ONCE(!listener)) {
3190 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3191 newsk = inet_csk_accept(listener->sk, flags, err, kern);
3195 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3196 if (sk_is_mptcp(newsk)) {
3197 struct mptcp_subflow_context *subflow;
3198 struct sock *new_mptcp_sock;
3200 subflow = mptcp_subflow_ctx(newsk);
3201 new_mptcp_sock = subflow->conn;
3203 /* is_mptcp should be false if subflow->conn is missing, see
3204 * subflow_syn_recv_sock()
3206 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3207 tcp_sk(newsk)->is_mptcp = 0;
3211 newsk = new_mptcp_sock;
3212 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3214 MPTCP_INC_STATS(sock_net(sk),
3215 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3219 newsk->sk_kern_sock = kern;
3223 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3225 struct mptcp_subflow_context *subflow, *tmp;
3226 struct sock *sk = (struct sock *)msk;
3228 __mptcp_clear_xmit(sk);
3230 /* join list will be eventually flushed (with rst) at sock lock release time */
3231 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3232 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3234 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3235 mptcp_data_lock(sk);
3236 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3237 __skb_queue_purge(&sk->sk_receive_queue);
3238 skb_rbtree_purge(&msk->out_of_order_queue);
3239 mptcp_data_unlock(sk);
3241 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3242 * inet_sock_destruct() will dispose it
3244 sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3245 msk->rmem_fwd_alloc = 0;
3246 mptcp_token_destroy(msk);
3247 mptcp_pm_free_anno_list(msk);
3248 mptcp_free_local_addr_list(msk);
3251 static void mptcp_destroy(struct sock *sk)
3253 struct mptcp_sock *msk = mptcp_sk(sk);
3255 /* clears msk->subflow, allowing the following to close
3256 * even the initial subflow
3258 mptcp_dispose_initial_subflow(msk);
3259 mptcp_destroy_common(msk, 0);
3260 sk_sockets_allocated_dec(sk);
3263 void __mptcp_data_acked(struct sock *sk)
3265 if (!sock_owned_by_user(sk))
3266 __mptcp_clean_una(sk);
3268 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3270 if (mptcp_pending_data_fin_ack(sk))
3271 mptcp_schedule_work(sk);
3274 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3276 if (!mptcp_send_head(sk))
3279 if (!sock_owned_by_user(sk))
3280 __mptcp_subflow_push_pending(sk, ssk, false);
3282 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3285 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3286 BIT(MPTCP_RETRANSMIT) | \
3287 BIT(MPTCP_FLUSH_JOIN_LIST))
3289 /* processes deferred events and flush wmem */
3290 static void mptcp_release_cb(struct sock *sk)
3291 __must_hold(&sk->sk_lock.slock)
3293 struct mptcp_sock *msk = mptcp_sk(sk);
3296 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3298 struct list_head join_list;
3303 INIT_LIST_HEAD(&join_list);
3304 list_splice_init(&msk->join_list, &join_list);
3306 /* the following actions acquire the subflow socket lock
3308 * 1) can't be invoked in atomic scope
3309 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3310 * datapath acquires the msk socket spinlock while helding
3311 * the subflow socket lock
3313 msk->push_pending = 0;
3314 msk->cb_flags &= ~flags;
3315 spin_unlock_bh(&sk->sk_lock.slock);
3317 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3318 __mptcp_flush_join_list(sk, &join_list);
3319 if (flags & BIT(MPTCP_PUSH_PENDING))
3320 __mptcp_push_pending(sk, 0);
3321 if (flags & BIT(MPTCP_RETRANSMIT))
3322 __mptcp_retrans(sk);
3325 spin_lock_bh(&sk->sk_lock.slock);
3328 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3329 __mptcp_clean_una_wakeup(sk);
3330 if (unlikely(&msk->cb_flags)) {
3331 /* be sure to set the current sk state before tacking actions
3332 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3334 if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3335 __mptcp_set_connected(sk);
3336 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3337 __mptcp_error_report(sk);
3338 if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3339 msk->last_snd = NULL;
3342 __mptcp_update_rmem(sk);
3345 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3346 * TCP can't schedule delack timer before the subflow is fully established.
3347 * MPTCP uses the delack timer to do 3rd ack retransmissions
3349 static void schedule_3rdack_retransmission(struct sock *ssk)
3351 struct inet_connection_sock *icsk = inet_csk(ssk);
3352 struct tcp_sock *tp = tcp_sk(ssk);
3353 unsigned long timeout;
3355 if (mptcp_subflow_ctx(ssk)->fully_established)
3358 /* reschedule with a timeout above RTT, as we must look only for drop */
3360 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3362 timeout = TCP_TIMEOUT_INIT;
3365 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3366 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3367 icsk->icsk_ack.timeout = timeout;
3368 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3371 void mptcp_subflow_process_delegated(struct sock *ssk)
3373 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3374 struct sock *sk = subflow->conn;
3376 if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3377 mptcp_data_lock(sk);
3378 if (!sock_owned_by_user(sk))
3379 __mptcp_subflow_push_pending(sk, ssk, true);
3381 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3382 mptcp_data_unlock(sk);
3383 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3385 if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3386 schedule_3rdack_retransmission(ssk);
3387 mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3391 static int mptcp_hash(struct sock *sk)
3393 /* should never be called,
3394 * we hash the TCP subflows not the master socket
3400 static void mptcp_unhash(struct sock *sk)
3402 /* called from sk_common_release(), but nothing to do here */
3405 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3407 struct mptcp_sock *msk = mptcp_sk(sk);
3408 struct socket *ssock;
3410 ssock = msk->subflow;
3411 pr_debug("msk=%p, subflow=%p", msk, ssock);
3412 if (WARN_ON_ONCE(!ssock))
3415 return inet_csk_get_port(ssock->sk, snum);
3418 void mptcp_finish_connect(struct sock *ssk)
3420 struct mptcp_subflow_context *subflow;
3421 struct mptcp_sock *msk;
3424 subflow = mptcp_subflow_ctx(ssk);
3428 pr_debug("msk=%p, token=%u", sk, subflow->token);
3430 subflow->map_seq = subflow->iasn;
3431 subflow->map_subflow_seq = 1;
3433 /* the socket is not connected yet, no msk/subflow ops can access/race
3434 * accessing the field below
3436 WRITE_ONCE(msk->local_key, subflow->local_key);
3437 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3438 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3439 WRITE_ONCE(msk->snd_una, msk->write_seq);
3441 mptcp_pm_new_connection(msk, ssk, 0);
3443 mptcp_rcv_space_init(msk, ssk);
3446 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3448 write_lock_bh(&sk->sk_callback_lock);
3449 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3450 sk_set_socket(sk, parent);
3451 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3452 write_unlock_bh(&sk->sk_callback_lock);
3455 bool mptcp_finish_join(struct sock *ssk)
3457 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3458 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3459 struct sock *parent = (void *)msk;
3462 pr_debug("msk=%p, subflow=%p", msk, subflow);
3464 /* mptcp socket already closing? */
3465 if (!mptcp_is_fully_established(parent)) {
3466 subflow->reset_reason = MPTCP_RST_EMPTCP;
3470 /* active subflow, already present inside the conn_list */
3471 if (!list_empty(&subflow->node)) {
3472 mptcp_subflow_joined(msk, ssk);
3476 if (!mptcp_pm_allow_new_subflow(msk))
3477 goto err_prohibited;
3479 /* If we can't acquire msk socket lock here, let the release callback
3482 mptcp_data_lock(parent);
3483 if (!sock_owned_by_user(parent)) {
3484 ret = __mptcp_finish_join(msk, ssk);
3487 list_add_tail(&subflow->node, &msk->conn_list);
3491 list_add_tail(&subflow->node, &msk->join_list);
3492 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3494 mptcp_data_unlock(parent);
3498 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3505 static void mptcp_shutdown(struct sock *sk, int how)
3507 pr_debug("sk=%p, how=%d", sk, how);
3509 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3510 __mptcp_wr_shutdown(sk);
3513 static int mptcp_forward_alloc_get(const struct sock *sk)
3515 return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3518 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3520 const struct sock *sk = (void *)msk;
3523 if (sk->sk_state == TCP_LISTEN)
3526 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3529 delta = msk->write_seq - v;
3530 if (__mptcp_check_fallback(msk) && msk->first) {
3531 struct tcp_sock *tp = tcp_sk(msk->first);
3533 /* the first subflow is disconnected after close - see
3534 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3535 * so ignore that status, too.
3537 if (!((1 << msk->first->sk_state) &
3538 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3539 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3541 if (delta > INT_MAX)
3547 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3549 struct mptcp_sock *msk = mptcp_sk(sk);
3554 if (sk->sk_state == TCP_LISTEN)
3558 __mptcp_move_skbs(msk);
3559 *karg = mptcp_inq_hint(sk);
3563 slow = lock_sock_fast(sk);
3564 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3565 unlock_sock_fast(sk, slow);
3568 slow = lock_sock_fast(sk);
3569 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3570 unlock_sock_fast(sk, slow);
3573 return -ENOIOCTLCMD;
3579 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3580 struct mptcp_subflow_context *subflow)
3582 subflow->request_mptcp = 0;
3583 __mptcp_do_fallback(msk);
3586 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3588 struct mptcp_subflow_context *subflow;
3589 struct mptcp_sock *msk = mptcp_sk(sk);
3590 struct socket *ssock;
3593 ssock = __mptcp_nmpc_socket(msk);
3595 return PTR_ERR(ssock);
3597 mptcp_token_destroy(msk);
3598 inet_sk_state_store(sk, TCP_SYN_SENT);
3599 subflow = mptcp_subflow_ctx(ssock->sk);
3600 #ifdef CONFIG_TCP_MD5SIG
3601 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3604 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3605 mptcp_subflow_early_fallback(msk, subflow);
3607 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3608 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3609 mptcp_subflow_early_fallback(msk, subflow);
3611 if (likely(!__mptcp_check_fallback(msk)))
3612 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3614 /* if reaching here via the fastopen/sendmsg path, the caller already
3615 * acquired the subflow socket lock, too.
3617 if (msk->fastopening)
3618 err = __inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK, 1);
3620 err = inet_stream_connect(ssock, uaddr, addr_len, O_NONBLOCK);
3621 inet_sk(sk)->defer_connect = inet_sk(ssock->sk)->defer_connect;
3623 /* on successful connect, the msk state will be moved to established by
3624 * subflow_finish_connect()
3626 if (unlikely(err && err != -EINPROGRESS)) {
3627 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3631 mptcp_copy_inaddrs(sk, ssock->sk);
3633 /* silence EINPROGRESS and let the caller inet_stream_connect
3634 * handle the connection in progress
3639 static struct proto mptcp_prot = {
3641 .owner = THIS_MODULE,
3642 .init = mptcp_init_sock,
3643 .connect = mptcp_connect,
3644 .disconnect = mptcp_disconnect,
3645 .close = mptcp_close,
3646 .accept = mptcp_accept,
3647 .setsockopt = mptcp_setsockopt,
3648 .getsockopt = mptcp_getsockopt,
3649 .shutdown = mptcp_shutdown,
3650 .destroy = mptcp_destroy,
3651 .sendmsg = mptcp_sendmsg,
3652 .ioctl = mptcp_ioctl,
3653 .recvmsg = mptcp_recvmsg,
3654 .release_cb = mptcp_release_cb,
3656 .unhash = mptcp_unhash,
3657 .get_port = mptcp_get_port,
3658 .forward_alloc_get = mptcp_forward_alloc_get,
3659 .sockets_allocated = &mptcp_sockets_allocated,
3661 .memory_allocated = &tcp_memory_allocated,
3662 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3664 .memory_pressure = &tcp_memory_pressure,
3665 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3666 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3667 .sysctl_mem = sysctl_tcp_mem,
3668 .obj_size = sizeof(struct mptcp_sock),
3669 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3670 .no_autobind = true,
3673 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3675 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3676 struct socket *ssock;
3679 lock_sock(sock->sk);
3680 ssock = __mptcp_nmpc_socket(msk);
3681 if (IS_ERR(ssock)) {
3682 err = PTR_ERR(ssock);
3686 err = ssock->ops->bind(ssock, uaddr, addr_len);
3688 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3691 release_sock(sock->sk);
3695 static int mptcp_listen(struct socket *sock, int backlog)
3697 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3698 struct sock *sk = sock->sk;
3699 struct socket *ssock;
3702 pr_debug("msk=%p", msk);
3707 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3710 ssock = __mptcp_nmpc_socket(msk);
3711 if (IS_ERR(ssock)) {
3712 err = PTR_ERR(ssock);
3716 mptcp_token_destroy(msk);
3717 inet_sk_state_store(sk, TCP_LISTEN);
3718 sock_set_flag(sk, SOCK_RCU_FREE);
3720 err = ssock->ops->listen(ssock, backlog);
3721 inet_sk_state_store(sk, inet_sk_state_load(ssock->sk));
3723 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3724 mptcp_copy_inaddrs(sk, ssock->sk);
3725 mptcp_event_pm_listener(ssock->sk, MPTCP_EVENT_LISTENER_CREATED);
3733 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3734 int flags, bool kern)
3736 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3737 struct socket *ssock;
3741 pr_debug("msk=%p", msk);
3743 /* Buggy applications can call accept on socket states other then LISTEN
3744 * but no need to allocate the first subflow just to error out.
3746 ssock = READ_ONCE(msk->subflow);
3750 newsk = mptcp_accept(sock->sk, flags, &err, kern);
3756 __inet_accept(sock, newsock, newsk);
3757 if (!mptcp_is_tcpsk(newsock->sk)) {
3758 struct mptcp_sock *msk = mptcp_sk(newsk);
3759 struct mptcp_subflow_context *subflow;
3761 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3762 msk->in_accept_queue = 0;
3764 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3765 * This is needed so NOSPACE flag can be set from tcp stack.
3767 mptcp_for_each_subflow(msk, subflow) {
3768 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3770 if (!ssk->sk_socket)
3771 mptcp_sock_graft(ssk, newsock);
3774 /* Do late cleanup for the first subflow as necessary. Also
3775 * deal with bad peers not doing a complete shutdown.
3778 unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3779 __mptcp_close_ssk(newsk, msk->first,
3780 mptcp_subflow_ctx(msk->first), 0);
3781 if (unlikely(list_empty(&msk->conn_list)))
3782 inet_sk_state_store(newsk, TCP_CLOSE);
3785 release_sock(newsk);
3790 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3792 struct sock *sk = (struct sock *)msk;
3794 if (sk_stream_is_writeable(sk))
3795 return EPOLLOUT | EPOLLWRNORM;
3797 mptcp_set_nospace(sk);
3798 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3799 if (sk_stream_is_writeable(sk))
3800 return EPOLLOUT | EPOLLWRNORM;
3805 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3806 struct poll_table_struct *wait)
3808 struct sock *sk = sock->sk;
3809 struct mptcp_sock *msk;
3815 sock_poll_wait(file, sock, wait);
3817 state = inet_sk_state_load(sk);
3818 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3819 if (state == TCP_LISTEN) {
3820 struct socket *ssock = READ_ONCE(msk->subflow);
3822 if (WARN_ON_ONCE(!ssock || !ssock->sk))
3825 return inet_csk_listen_poll(ssock->sk);
3828 shutdown = READ_ONCE(sk->sk_shutdown);
3829 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3831 if (shutdown & RCV_SHUTDOWN)
3832 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3834 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3835 mask |= mptcp_check_readable(msk);
3836 if (shutdown & SEND_SHUTDOWN)
3837 mask |= EPOLLOUT | EPOLLWRNORM;
3839 mask |= mptcp_check_writeable(msk);
3840 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
3841 /* cf tcp_poll() note about TFO */
3842 mask |= EPOLLOUT | EPOLLWRNORM;
3845 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3847 if (READ_ONCE(sk->sk_err))
3853 static const struct proto_ops mptcp_stream_ops = {
3855 .owner = THIS_MODULE,
3856 .release = inet_release,
3858 .connect = inet_stream_connect,
3859 .socketpair = sock_no_socketpair,
3860 .accept = mptcp_stream_accept,
3861 .getname = inet_getname,
3863 .ioctl = inet_ioctl,
3864 .gettstamp = sock_gettstamp,
3865 .listen = mptcp_listen,
3866 .shutdown = inet_shutdown,
3867 .setsockopt = sock_common_setsockopt,
3868 .getsockopt = sock_common_getsockopt,
3869 .sendmsg = inet_sendmsg,
3870 .recvmsg = inet_recvmsg,
3871 .mmap = sock_no_mmap,
3874 static struct inet_protosw mptcp_protosw = {
3875 .type = SOCK_STREAM,
3876 .protocol = IPPROTO_MPTCP,
3877 .prot = &mptcp_prot,
3878 .ops = &mptcp_stream_ops,
3879 .flags = INET_PROTOSW_ICSK,
3882 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3884 struct mptcp_delegated_action *delegated;
3885 struct mptcp_subflow_context *subflow;
3888 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3889 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3890 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3892 bh_lock_sock_nested(ssk);
3893 if (!sock_owned_by_user(ssk) &&
3894 mptcp_subflow_has_delegated_action(subflow))
3895 mptcp_subflow_process_delegated(ssk);
3896 /* ... elsewhere tcp_release_cb_override already processed
3897 * the action or will do at next release_sock().
3898 * In both case must dequeue the subflow here - on the same
3899 * CPU that scheduled it.
3901 bh_unlock_sock(ssk);
3904 if (++work_done == budget)
3908 /* always provide a 0 'work_done' argument, so that napi_complete_done
3909 * will not try accessing the NULL napi->dev ptr
3911 napi_complete_done(napi, 0);
3915 void __init mptcp_proto_init(void)
3917 struct mptcp_delegated_action *delegated;
3920 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3922 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3923 panic("Failed to allocate MPTCP pcpu counter\n");
3925 init_dummy_netdev(&mptcp_napi_dev);
3926 for_each_possible_cpu(cpu) {
3927 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3928 INIT_LIST_HEAD(&delegated->head);
3929 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3931 napi_enable(&delegated->napi);
3934 mptcp_subflow_init();
3938 if (proto_register(&mptcp_prot, 1) != 0)
3939 panic("Failed to register MPTCP proto.\n");
3941 inet_register_protosw(&mptcp_protosw);
3943 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3946 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3947 static const struct proto_ops mptcp_v6_stream_ops = {
3949 .owner = THIS_MODULE,
3950 .release = inet6_release,
3952 .connect = inet_stream_connect,
3953 .socketpair = sock_no_socketpair,
3954 .accept = mptcp_stream_accept,
3955 .getname = inet6_getname,
3957 .ioctl = inet6_ioctl,
3958 .gettstamp = sock_gettstamp,
3959 .listen = mptcp_listen,
3960 .shutdown = inet_shutdown,
3961 .setsockopt = sock_common_setsockopt,
3962 .getsockopt = sock_common_getsockopt,
3963 .sendmsg = inet6_sendmsg,
3964 .recvmsg = inet6_recvmsg,
3965 .mmap = sock_no_mmap,
3966 #ifdef CONFIG_COMPAT
3967 .compat_ioctl = inet6_compat_ioctl,
3971 static struct proto mptcp_v6_prot;
3973 static struct inet_protosw mptcp_v6_protosw = {
3974 .type = SOCK_STREAM,
3975 .protocol = IPPROTO_MPTCP,
3976 .prot = &mptcp_v6_prot,
3977 .ops = &mptcp_v6_stream_ops,
3978 .flags = INET_PROTOSW_ICSK,
3981 int __init mptcp_proto_v6_init(void)
3985 mptcp_v6_prot = mptcp_prot;
3986 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3987 mptcp_v6_prot.slab = NULL;
3988 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3989 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
3991 err = proto_register(&mptcp_v6_prot, 1);
3995 err = inet6_register_protosw(&mptcp_v6_protosw);
3997 proto_unregister(&mptcp_v6_prot);