1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
40 MPTCP_CMSG_TS = BIT(0),
41 MPTCP_CMSG_INQ = BIT(1),
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void mptcp_check_send_data_fin(struct sock *sk);
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
52 /* Returns end sequence number of the receiver's advertised window */
53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
55 return READ_ONCE(msk->wnd_end);
58 static bool mptcp_is_tcpsk(struct sock *sk)
60 struct socket *sock = sk->sk_socket;
62 if (unlikely(sk->sk_prot == &tcp_prot)) {
63 /* we are being invoked after mptcp_accept() has
64 * accepted a non-mp-capable flow: sk is a tcp_sk,
67 * Hand the socket over to tcp so all further socket ops
70 WRITE_ONCE(sock->ops, &inet_stream_ops);
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 WRITE_ONCE(sock->ops, &inet6_stream_ops);
82 static int __mptcp_socket_create(struct mptcp_sock *msk)
84 struct mptcp_subflow_context *subflow;
85 struct sock *sk = (struct sock *)msk;
89 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
93 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
94 WRITE_ONCE(msk->first, ssock->sk);
95 subflow = mptcp_subflow_ctx(ssock->sk);
96 list_add(&subflow->node, &msk->conn_list);
98 subflow->request_mptcp = 1;
99 subflow->subflow_id = msk->subflow_id++;
101 /* This is the first subflow, always with id 0 */
102 subflow->local_id_valid = 1;
103 mptcp_sock_graft(msk->first, sk->sk_socket);
104 iput(SOCK_INODE(ssock));
109 /* If the MPC handshake is not started, returns the first subflow,
110 * eventually allocating it.
112 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
114 struct sock *sk = (struct sock *)msk;
117 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
118 return ERR_PTR(-EINVAL);
121 ret = __mptcp_socket_create(msk);
129 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
131 sk_drops_add(sk, skb);
135 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
137 WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
138 mptcp_sk(sk)->rmem_fwd_alloc + size);
141 static void mptcp_rmem_charge(struct sock *sk, int size)
143 mptcp_rmem_fwd_alloc_add(sk, -size);
146 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
147 struct sk_buff *from)
152 if (MPTCP_SKB_CB(from)->offset ||
153 !skb_try_coalesce(to, from, &fragstolen, &delta))
156 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
157 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
158 to->len, MPTCP_SKB_CB(from)->end_seq);
159 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
161 /* note the fwd memory can reach a negative value after accounting
162 * for the delta, but the later skb free will restore a non
165 atomic_add(delta, &sk->sk_rmem_alloc);
166 mptcp_rmem_charge(sk, delta);
167 kfree_skb_partial(from, fragstolen);
172 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
173 struct sk_buff *from)
175 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
178 return mptcp_try_coalesce((struct sock *)msk, to, from);
181 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
183 amount >>= PAGE_SHIFT;
184 mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
185 __sk_mem_reduce_allocated(sk, amount);
188 static void mptcp_rmem_uncharge(struct sock *sk, int size)
190 struct mptcp_sock *msk = mptcp_sk(sk);
193 mptcp_rmem_fwd_alloc_add(sk, size);
194 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
196 /* see sk_mem_uncharge() for the rationale behind the following schema */
197 if (unlikely(reclaimable >= PAGE_SIZE))
198 __mptcp_rmem_reclaim(sk, reclaimable);
201 static void mptcp_rfree(struct sk_buff *skb)
203 unsigned int len = skb->truesize;
204 struct sock *sk = skb->sk;
206 atomic_sub(len, &sk->sk_rmem_alloc);
207 mptcp_rmem_uncharge(sk, len);
210 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
214 skb->destructor = mptcp_rfree;
215 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
216 mptcp_rmem_charge(sk, skb->truesize);
219 /* "inspired" by tcp_data_queue_ofo(), main differences:
221 * - don't cope with sacks
223 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
225 struct sock *sk = (struct sock *)msk;
226 struct rb_node **p, *parent;
227 u64 seq, end_seq, max_seq;
228 struct sk_buff *skb1;
230 seq = MPTCP_SKB_CB(skb)->map_seq;
231 end_seq = MPTCP_SKB_CB(skb)->end_seq;
232 max_seq = atomic64_read(&msk->rcv_wnd_sent);
234 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
235 RB_EMPTY_ROOT(&msk->out_of_order_queue));
236 if (after64(end_seq, max_seq)) {
239 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
240 (unsigned long long)end_seq - (unsigned long)max_seq,
241 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
242 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
246 p = &msk->out_of_order_queue.rb_node;
247 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
248 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
249 rb_link_node(&skb->rbnode, NULL, p);
250 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
251 msk->ooo_last_skb = skb;
255 /* with 2 subflows, adding at end of ooo queue is quite likely
256 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
258 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
259 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
260 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
264 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
265 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
266 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
267 parent = &msk->ooo_last_skb->rbnode;
268 p = &parent->rb_right;
272 /* Find place to insert this segment. Handle overlaps on the way. */
276 skb1 = rb_to_skb(parent);
277 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
278 p = &parent->rb_left;
281 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
282 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
283 /* All the bits are present. Drop. */
285 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
288 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
292 * continue traversing
295 /* skb's seq == skb1's seq and skb covers skb1.
296 * Replace skb1 with skb.
298 rb_replace_node(&skb1->rbnode, &skb->rbnode,
299 &msk->out_of_order_queue);
300 mptcp_drop(sk, skb1);
301 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
304 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
305 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
308 p = &parent->rb_right;
312 /* Insert segment into RB tree. */
313 rb_link_node(&skb->rbnode, parent, p);
314 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
317 /* Remove other segments covered by skb. */
318 while ((skb1 = skb_rb_next(skb)) != NULL) {
319 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
321 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
322 mptcp_drop(sk, skb1);
323 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
325 /* If there is no skb after us, we are the last_skb ! */
327 msk->ooo_last_skb = skb;
331 mptcp_set_owner_r(skb, sk);
334 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
336 struct mptcp_sock *msk = mptcp_sk(sk);
339 if (size <= msk->rmem_fwd_alloc)
342 size -= msk->rmem_fwd_alloc;
343 amt = sk_mem_pages(size);
344 amount = amt << PAGE_SHIFT;
345 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
348 mptcp_rmem_fwd_alloc_add(sk, amount);
352 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
353 struct sk_buff *skb, unsigned int offset,
356 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
357 struct sock *sk = (struct sock *)msk;
358 struct sk_buff *tail;
361 __skb_unlink(skb, &ssk->sk_receive_queue);
366 /* try to fetch required memory from subflow */
367 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
370 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
372 /* the skb map_seq accounts for the skb offset:
373 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
376 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
377 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
378 MPTCP_SKB_CB(skb)->offset = offset;
379 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
381 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
383 msk->bytes_received += copy_len;
384 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
385 tail = skb_peek_tail(&sk->sk_receive_queue);
386 if (tail && mptcp_try_coalesce(sk, tail, skb))
389 mptcp_set_owner_r(skb, sk);
390 __skb_queue_tail(&sk->sk_receive_queue, skb);
392 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
393 mptcp_data_queue_ofo(msk, skb);
397 /* old data, keep it simple and drop the whole pkt, sender
398 * will retransmit as needed, if needed.
400 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
406 static void mptcp_stop_rtx_timer(struct sock *sk)
408 struct inet_connection_sock *icsk = inet_csk(sk);
410 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
411 mptcp_sk(sk)->timer_ival = 0;
414 static void mptcp_close_wake_up(struct sock *sk)
416 if (sock_flag(sk, SOCK_DEAD))
419 sk->sk_state_change(sk);
420 if (sk->sk_shutdown == SHUTDOWN_MASK ||
421 sk->sk_state == TCP_CLOSE)
422 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
424 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
427 static bool mptcp_pending_data_fin_ack(struct sock *sk)
429 struct mptcp_sock *msk = mptcp_sk(sk);
431 return ((1 << sk->sk_state) &
432 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
433 msk->write_seq == READ_ONCE(msk->snd_una);
436 static void mptcp_check_data_fin_ack(struct sock *sk)
438 struct mptcp_sock *msk = mptcp_sk(sk);
440 /* Look for an acknowledged DATA_FIN */
441 if (mptcp_pending_data_fin_ack(sk)) {
442 WRITE_ONCE(msk->snd_data_fin_enable, 0);
444 switch (sk->sk_state) {
446 inet_sk_state_store(sk, TCP_FIN_WAIT2);
450 inet_sk_state_store(sk, TCP_CLOSE);
454 mptcp_close_wake_up(sk);
458 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
460 struct mptcp_sock *msk = mptcp_sk(sk);
462 if (READ_ONCE(msk->rcv_data_fin) &&
463 ((1 << sk->sk_state) &
464 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
465 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
467 if (msk->ack_seq == rcv_data_fin_seq) {
469 *seq = rcv_data_fin_seq;
478 static void mptcp_set_datafin_timeout(struct sock *sk)
480 struct inet_connection_sock *icsk = inet_csk(sk);
483 retransmits = min_t(u32, icsk->icsk_retransmits,
484 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
486 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
489 static void __mptcp_set_timeout(struct sock *sk, long tout)
491 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
494 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
496 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
498 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
499 inet_csk(ssk)->icsk_timeout - jiffies : 0;
502 static void mptcp_set_timeout(struct sock *sk)
504 struct mptcp_subflow_context *subflow;
507 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
508 tout = max(tout, mptcp_timeout_from_subflow(subflow));
509 __mptcp_set_timeout(sk, tout);
512 static inline bool tcp_can_send_ack(const struct sock *ssk)
514 return !((1 << inet_sk_state_load(ssk)) &
515 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
518 void __mptcp_subflow_send_ack(struct sock *ssk)
520 if (tcp_can_send_ack(ssk))
524 static void mptcp_subflow_send_ack(struct sock *ssk)
528 slow = lock_sock_fast(ssk);
529 __mptcp_subflow_send_ack(ssk);
530 unlock_sock_fast(ssk, slow);
533 static void mptcp_send_ack(struct mptcp_sock *msk)
535 struct mptcp_subflow_context *subflow;
537 mptcp_for_each_subflow(msk, subflow)
538 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
541 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
545 slow = lock_sock_fast(ssk);
546 if (tcp_can_send_ack(ssk))
547 tcp_cleanup_rbuf(ssk, 1);
548 unlock_sock_fast(ssk, slow);
551 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
553 const struct inet_connection_sock *icsk = inet_csk(ssk);
554 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
555 const struct tcp_sock *tp = tcp_sk(ssk);
557 return (ack_pending & ICSK_ACK_SCHED) &&
558 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
559 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
560 (rx_empty && ack_pending &
561 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
564 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
566 int old_space = READ_ONCE(msk->old_wspace);
567 struct mptcp_subflow_context *subflow;
568 struct sock *sk = (struct sock *)msk;
569 int space = __mptcp_space(sk);
570 bool cleanup, rx_empty;
572 cleanup = (space > 0) && (space >= (old_space << 1));
573 rx_empty = !__mptcp_rmem(sk);
575 mptcp_for_each_subflow(msk, subflow) {
576 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
578 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
579 mptcp_subflow_cleanup_rbuf(ssk);
583 static bool mptcp_check_data_fin(struct sock *sk)
585 struct mptcp_sock *msk = mptcp_sk(sk);
586 u64 rcv_data_fin_seq;
589 /* Need to ack a DATA_FIN received from a peer while this side
590 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
591 * msk->rcv_data_fin was set when parsing the incoming options
592 * at the subflow level and the msk lock was not held, so this
593 * is the first opportunity to act on the DATA_FIN and change
596 * If we are caught up to the sequence number of the incoming
597 * DATA_FIN, send the DATA_ACK now and do state transition. If
598 * not caught up, do nothing and let the recv code send DATA_ACK
602 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
603 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
604 WRITE_ONCE(msk->rcv_data_fin, 0);
606 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
607 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
609 switch (sk->sk_state) {
610 case TCP_ESTABLISHED:
611 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
614 inet_sk_state_store(sk, TCP_CLOSING);
617 inet_sk_state_store(sk, TCP_CLOSE);
620 /* Other states not expected */
626 if (!__mptcp_check_fallback(msk))
628 mptcp_close_wake_up(sk);
633 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
637 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
638 struct sock *sk = (struct sock *)msk;
639 unsigned int moved = 0;
640 bool more_data_avail;
645 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
647 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
648 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
650 if (unlikely(ssk_rbuf > sk_rbuf)) {
651 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
656 pr_debug("msk=%p ssk=%p", msk, ssk);
659 u32 map_remaining, offset;
660 u32 seq = tp->copied_seq;
664 /* try to move as much data as available */
665 map_remaining = subflow->map_data_len -
666 mptcp_subflow_get_map_offset(subflow);
668 skb = skb_peek(&ssk->sk_receive_queue);
670 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
671 * a different CPU can have already processed the pending
672 * data, stop here or we can enter an infinite loop
679 if (__mptcp_check_fallback(msk)) {
680 /* Under fallback skbs have no MPTCP extension and TCP could
681 * collapse them between the dummy map creation and the
682 * current dequeue. Be sure to adjust the map size.
684 map_remaining = skb->len;
685 subflow->map_data_len = skb->len;
688 offset = seq - TCP_SKB_CB(skb)->seq;
689 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
695 if (offset < skb->len) {
696 size_t len = skb->len - offset;
701 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
705 if (WARN_ON_ONCE(map_remaining < len))
709 sk_eat_skb(ssk, skb);
713 WRITE_ONCE(tp->copied_seq, seq);
714 more_data_avail = mptcp_subflow_data_available(ssk);
716 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
720 } while (more_data_avail);
726 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
728 struct sock *sk = (struct sock *)msk;
729 struct sk_buff *skb, *tail;
734 p = rb_first(&msk->out_of_order_queue);
735 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
738 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
742 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
744 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
747 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
751 end_seq = MPTCP_SKB_CB(skb)->end_seq;
752 tail = skb_peek_tail(&sk->sk_receive_queue);
753 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
754 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
756 /* skip overlapping data, if any */
757 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
758 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
760 MPTCP_SKB_CB(skb)->offset += delta;
761 MPTCP_SKB_CB(skb)->map_seq += delta;
762 __skb_queue_tail(&sk->sk_receive_queue, skb);
764 msk->bytes_received += end_seq - msk->ack_seq;
765 msk->ack_seq = end_seq;
771 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
773 int err = sock_error(ssk);
779 /* only propagate errors on fallen-back sockets or
782 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
785 /* We need to propagate only transition to CLOSE state.
786 * Orphaned socket will see such state change via
787 * subflow_sched_work_if_closed() and that path will properly
788 * destroy the msk as needed.
790 ssk_state = inet_sk_state_load(ssk);
791 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
792 inet_sk_state_store(sk, ssk_state);
793 WRITE_ONCE(sk->sk_err, -err);
795 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
801 void __mptcp_error_report(struct sock *sk)
803 struct mptcp_subflow_context *subflow;
804 struct mptcp_sock *msk = mptcp_sk(sk);
806 mptcp_for_each_subflow(msk, subflow)
807 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
811 /* In most cases we will be able to lock the mptcp socket. If its already
812 * owned, we need to defer to the work queue to avoid ABBA deadlock.
814 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
816 struct sock *sk = (struct sock *)msk;
817 unsigned int moved = 0;
819 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
820 __mptcp_ofo_queue(msk);
821 if (unlikely(ssk->sk_err)) {
822 if (!sock_owned_by_user(sk))
823 __mptcp_error_report(sk);
825 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
828 /* If the moves have caught up with the DATA_FIN sequence number
829 * it's time to ack the DATA_FIN and change socket state, but
830 * this is not a good place to change state. Let the workqueue
833 if (mptcp_pending_data_fin(sk, NULL))
834 mptcp_schedule_work(sk);
838 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
840 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
841 struct mptcp_sock *msk = mptcp_sk(sk);
842 int sk_rbuf, ssk_rbuf;
844 /* The peer can send data while we are shutting down this
845 * subflow at msk destruction time, but we must avoid enqueuing
846 * more data to the msk receive queue
848 if (unlikely(subflow->disposable))
851 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
852 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
853 if (unlikely(ssk_rbuf > sk_rbuf))
856 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
857 if (__mptcp_rmem(sk) > sk_rbuf) {
858 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
862 /* Wake-up the reader only for in-sequence data */
864 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
865 sk->sk_data_ready(sk);
866 mptcp_data_unlock(sk);
869 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
871 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
872 WRITE_ONCE(msk->allow_infinite_fallback, false);
873 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
876 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
878 struct sock *sk = (struct sock *)msk;
880 if (sk->sk_state != TCP_ESTABLISHED)
883 /* attach to msk socket only after we are sure we will deal with it
886 if (sk->sk_socket && !ssk->sk_socket)
887 mptcp_sock_graft(ssk, sk->sk_socket);
889 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
890 mptcp_sockopt_sync_locked(msk, ssk);
891 mptcp_subflow_joined(msk, ssk);
892 mptcp_stop_tout_timer(sk);
893 __mptcp_propagate_sndbuf(sk, ssk);
897 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
899 struct mptcp_subflow_context *tmp, *subflow;
900 struct mptcp_sock *msk = mptcp_sk(sk);
902 list_for_each_entry_safe(subflow, tmp, join_list, node) {
903 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
904 bool slow = lock_sock_fast(ssk);
906 list_move_tail(&subflow->node, &msk->conn_list);
907 if (!__mptcp_finish_join(msk, ssk))
908 mptcp_subflow_reset(ssk);
909 unlock_sock_fast(ssk, slow);
913 static bool mptcp_rtx_timer_pending(struct sock *sk)
915 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
918 static void mptcp_reset_rtx_timer(struct sock *sk)
920 struct inet_connection_sock *icsk = inet_csk(sk);
923 /* prevent rescheduling on close */
924 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
927 tout = mptcp_sk(sk)->timer_ival;
928 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
931 bool mptcp_schedule_work(struct sock *sk)
933 if (inet_sk_state_load(sk) != TCP_CLOSE &&
934 schedule_work(&mptcp_sk(sk)->work)) {
935 /* each subflow already holds a reference to the sk, and the
936 * workqueue is invoked by a subflow, so sk can't go away here.
944 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
946 struct mptcp_subflow_context *subflow;
948 msk_owned_by_me(msk);
950 mptcp_for_each_subflow(msk, subflow) {
951 if (READ_ONCE(subflow->data_avail))
952 return mptcp_subflow_tcp_sock(subflow);
958 static bool mptcp_skb_can_collapse_to(u64 write_seq,
959 const struct sk_buff *skb,
960 const struct mptcp_ext *mpext)
962 if (!tcp_skb_can_collapse_to(skb))
965 /* can collapse only if MPTCP level sequence is in order and this
966 * mapping has not been xmitted yet
968 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
972 /* we can append data to the given data frag if:
973 * - there is space available in the backing page_frag
974 * - the data frag tail matches the current page_frag free offset
975 * - the data frag end sequence number matches the current write seq
977 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
978 const struct page_frag *pfrag,
979 const struct mptcp_data_frag *df)
981 return df && pfrag->page == df->page &&
982 pfrag->size - pfrag->offset > 0 &&
983 pfrag->offset == (df->offset + df->data_len) &&
984 df->data_seq + df->data_len == msk->write_seq;
987 static void dfrag_uncharge(struct sock *sk, int len)
989 sk_mem_uncharge(sk, len);
990 sk_wmem_queued_add(sk, -len);
993 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
995 int len = dfrag->data_len + dfrag->overhead;
997 list_del(&dfrag->list);
998 dfrag_uncharge(sk, len);
999 put_page(dfrag->page);
1002 static void __mptcp_clean_una(struct sock *sk)
1004 struct mptcp_sock *msk = mptcp_sk(sk);
1005 struct mptcp_data_frag *dtmp, *dfrag;
1008 snd_una = msk->snd_una;
1009 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1010 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1013 if (unlikely(dfrag == msk->first_pending)) {
1014 /* in recovery mode can see ack after the current snd head */
1015 if (WARN_ON_ONCE(!msk->recovery))
1018 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1021 dfrag_clear(sk, dfrag);
1024 dfrag = mptcp_rtx_head(sk);
1025 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1026 u64 delta = snd_una - dfrag->data_seq;
1028 /* prevent wrap around in recovery mode */
1029 if (unlikely(delta > dfrag->already_sent)) {
1030 if (WARN_ON_ONCE(!msk->recovery))
1032 if (WARN_ON_ONCE(delta > dfrag->data_len))
1034 dfrag->already_sent += delta - dfrag->already_sent;
1037 dfrag->data_seq += delta;
1038 dfrag->offset += delta;
1039 dfrag->data_len -= delta;
1040 dfrag->already_sent -= delta;
1042 dfrag_uncharge(sk, delta);
1045 /* all retransmitted data acked, recovery completed */
1046 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1047 msk->recovery = false;
1050 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1051 snd_una == READ_ONCE(msk->write_seq)) {
1052 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1053 mptcp_stop_rtx_timer(sk);
1055 mptcp_reset_rtx_timer(sk);
1059 static void __mptcp_clean_una_wakeup(struct sock *sk)
1061 lockdep_assert_held_once(&sk->sk_lock.slock);
1063 __mptcp_clean_una(sk);
1064 mptcp_write_space(sk);
1067 static void mptcp_clean_una_wakeup(struct sock *sk)
1069 mptcp_data_lock(sk);
1070 __mptcp_clean_una_wakeup(sk);
1071 mptcp_data_unlock(sk);
1074 static void mptcp_enter_memory_pressure(struct sock *sk)
1076 struct mptcp_subflow_context *subflow;
1077 struct mptcp_sock *msk = mptcp_sk(sk);
1080 mptcp_for_each_subflow(msk, subflow) {
1081 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1084 tcp_enter_memory_pressure(ssk);
1085 sk_stream_moderate_sndbuf(ssk);
1089 __mptcp_sync_sndbuf(sk);
1092 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1095 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1097 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1098 pfrag, sk->sk_allocation)))
1101 mptcp_enter_memory_pressure(sk);
1105 static struct mptcp_data_frag *
1106 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1109 int offset = ALIGN(orig_offset, sizeof(long));
1110 struct mptcp_data_frag *dfrag;
1112 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1113 dfrag->data_len = 0;
1114 dfrag->data_seq = msk->write_seq;
1115 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1116 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1117 dfrag->already_sent = 0;
1118 dfrag->page = pfrag->page;
1123 struct mptcp_sendmsg_info {
1129 bool data_lock_held;
1132 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1133 u64 data_seq, int avail_size)
1135 u64 window_end = mptcp_wnd_end(msk);
1138 if (__mptcp_check_fallback(msk))
1141 mptcp_snd_wnd = window_end - data_seq;
1142 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1144 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1145 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1146 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1152 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1154 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1158 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1162 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1164 struct sk_buff *skb;
1166 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1168 if (likely(__mptcp_add_ext(skb, gfp))) {
1169 skb_reserve(skb, MAX_TCP_HEADER);
1170 skb->ip_summed = CHECKSUM_PARTIAL;
1171 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1176 mptcp_enter_memory_pressure(sk);
1181 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1183 struct sk_buff *skb;
1185 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1189 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1190 tcp_skb_entail(ssk, skb);
1193 tcp_skb_tsorted_anchor_cleanup(skb);
1198 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1200 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1202 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1205 /* note: this always recompute the csum on the whole skb, even
1206 * if we just appended a single frag. More status info needed
1208 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1210 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1211 __wsum csum = ~csum_unfold(mpext->csum);
1212 int offset = skb->len - added;
1214 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1217 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1219 struct mptcp_ext *mpext)
1224 mpext->infinite_map = 1;
1225 mpext->data_len = 0;
1227 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1228 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1230 mptcp_do_fallback(ssk);
1233 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1235 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1236 struct mptcp_data_frag *dfrag,
1237 struct mptcp_sendmsg_info *info)
1239 u64 data_seq = dfrag->data_seq + info->sent;
1240 int offset = dfrag->offset + info->sent;
1241 struct mptcp_sock *msk = mptcp_sk(sk);
1242 bool zero_window_probe = false;
1243 struct mptcp_ext *mpext = NULL;
1244 bool can_coalesce = false;
1245 bool reuse_skb = true;
1246 struct sk_buff *skb;
1250 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1251 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1253 if (WARN_ON_ONCE(info->sent > info->limit ||
1254 info->limit > dfrag->data_len))
1257 if (unlikely(!__tcp_can_send(ssk)))
1260 /* compute send limit */
1261 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1262 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1263 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1264 copy = info->size_goal;
1266 skb = tcp_write_queue_tail(ssk);
1267 if (skb && copy > skb->len) {
1268 /* Limit the write to the size available in the
1269 * current skb, if any, so that we create at most a new skb.
1270 * Explicitly tells TCP internals to avoid collapsing on later
1271 * queue management operation, to avoid breaking the ext <->
1272 * SSN association set here
1274 mpext = mptcp_get_ext(skb);
1275 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1276 TCP_SKB_CB(skb)->eor = 1;
1280 i = skb_shinfo(skb)->nr_frags;
1281 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1282 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1283 tcp_mark_push(tcp_sk(ssk), skb);
1290 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1294 i = skb_shinfo(skb)->nr_frags;
1296 mpext = mptcp_get_ext(skb);
1299 /* Zero window and all data acked? Probe. */
1300 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1302 u64 snd_una = READ_ONCE(msk->snd_una);
1304 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1305 tcp_remove_empty_skb(ssk);
1309 zero_window_probe = true;
1310 data_seq = snd_una - 1;
1314 copy = min_t(size_t, copy, info->limit - info->sent);
1315 if (!sk_wmem_schedule(ssk, copy)) {
1316 tcp_remove_empty_skb(ssk);
1321 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1323 get_page(dfrag->page);
1324 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1328 skb->data_len += copy;
1329 skb->truesize += copy;
1330 sk_wmem_queued_add(ssk, copy);
1331 sk_mem_charge(ssk, copy);
1332 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1333 TCP_SKB_CB(skb)->end_seq += copy;
1334 tcp_skb_pcount_set(skb, 0);
1336 /* on skb reuse we just need to update the DSS len */
1338 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1339 mpext->data_len += copy;
1343 memset(mpext, 0, sizeof(*mpext));
1344 mpext->data_seq = data_seq;
1345 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1346 mpext->data_len = copy;
1350 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1351 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1354 if (zero_window_probe) {
1355 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1357 if (READ_ONCE(msk->csum_enabled))
1358 mptcp_update_data_checksum(skb, copy);
1359 tcp_push_pending_frames(ssk);
1363 if (READ_ONCE(msk->csum_enabled))
1364 mptcp_update_data_checksum(skb, copy);
1365 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1366 mptcp_update_infinite_map(msk, ssk, mpext);
1367 trace_mptcp_sendmsg_frag(mpext);
1368 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1372 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1373 sizeof(struct tcphdr) - \
1374 MAX_TCP_OPTION_SPACE - \
1375 sizeof(struct ipv6hdr) - \
1376 sizeof(struct frag_hdr))
1378 struct subflow_send_info {
1383 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1385 if (!subflow->stale)
1389 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1392 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1394 if (unlikely(subflow->stale)) {
1395 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1397 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1400 mptcp_subflow_set_active(subflow);
1402 return __mptcp_subflow_active(subflow);
1405 #define SSK_MODE_ACTIVE 0
1406 #define SSK_MODE_BACKUP 1
1407 #define SSK_MODE_MAX 2
1409 /* implement the mptcp packet scheduler;
1410 * returns the subflow that will transmit the next DSS
1411 * additionally updates the rtx timeout
1413 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1415 struct subflow_send_info send_info[SSK_MODE_MAX];
1416 struct mptcp_subflow_context *subflow;
1417 struct sock *sk = (struct sock *)msk;
1418 u32 pace, burst, wmem;
1419 int i, nr_active = 0;
1424 /* pick the subflow with the lower wmem/wspace ratio */
1425 for (i = 0; i < SSK_MODE_MAX; ++i) {
1426 send_info[i].ssk = NULL;
1427 send_info[i].linger_time = -1;
1430 mptcp_for_each_subflow(msk, subflow) {
1431 trace_mptcp_subflow_get_send(subflow);
1432 ssk = mptcp_subflow_tcp_sock(subflow);
1433 if (!mptcp_subflow_active(subflow))
1436 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1437 nr_active += !subflow->backup;
1438 pace = subflow->avg_pacing_rate;
1439 if (unlikely(!pace)) {
1440 /* init pacing rate from socket */
1441 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1442 pace = subflow->avg_pacing_rate;
1447 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1448 if (linger_time < send_info[subflow->backup].linger_time) {
1449 send_info[subflow->backup].ssk = ssk;
1450 send_info[subflow->backup].linger_time = linger_time;
1453 __mptcp_set_timeout(sk, tout);
1455 /* pick the best backup if no other subflow is active */
1457 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1459 /* According to the blest algorithm, to avoid HoL blocking for the
1460 * faster flow, we need to:
1461 * - estimate the faster flow linger time
1462 * - use the above to estimate the amount of byte transferred
1463 * by the faster flow
1464 * - check that the amount of queued data is greter than the above,
1465 * otherwise do not use the picked, slower, subflow
1466 * We select the subflow with the shorter estimated time to flush
1467 * the queued mem, which basically ensure the above. We just need
1468 * to check that subflow has a non empty cwin.
1470 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1471 if (!ssk || !sk_stream_memory_free(ssk))
1474 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1475 wmem = READ_ONCE(ssk->sk_wmem_queued);
1479 subflow = mptcp_subflow_ctx(ssk);
1480 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1481 READ_ONCE(ssk->sk_pacing_rate) * burst,
1483 msk->snd_burst = burst;
1487 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1489 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1493 static void mptcp_update_post_push(struct mptcp_sock *msk,
1494 struct mptcp_data_frag *dfrag,
1497 u64 snd_nxt_new = dfrag->data_seq;
1499 dfrag->already_sent += sent;
1501 msk->snd_burst -= sent;
1503 snd_nxt_new += dfrag->already_sent;
1505 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1506 * is recovering after a failover. In that event, this re-sends
1509 * Thus compute snd_nxt_new candidate based on
1510 * the dfrag->data_seq that was sent and the data
1511 * that has been handed to the subflow for transmission
1512 * and skip update in case it was old dfrag.
1514 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1515 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1516 msk->snd_nxt = snd_nxt_new;
1520 void mptcp_check_and_set_pending(struct sock *sk)
1522 if (mptcp_send_head(sk))
1523 mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1526 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1527 struct mptcp_sendmsg_info *info)
1529 struct mptcp_sock *msk = mptcp_sk(sk);
1530 struct mptcp_data_frag *dfrag;
1531 int len, copied = 0, err = 0;
1533 while ((dfrag = mptcp_send_head(sk))) {
1534 info->sent = dfrag->already_sent;
1535 info->limit = dfrag->data_len;
1536 len = dfrag->data_len - dfrag->already_sent;
1540 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1542 err = copied ? : ret;
1550 mptcp_update_post_push(msk, dfrag, ret);
1552 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1554 if (msk->snd_burst <= 0 ||
1555 !sk_stream_memory_free(ssk) ||
1556 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1560 mptcp_set_timeout(sk);
1568 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1570 struct sock *prev_ssk = NULL, *ssk = NULL;
1571 struct mptcp_sock *msk = mptcp_sk(sk);
1572 struct mptcp_sendmsg_info info = {
1575 bool do_check_data_fin = false;
1578 while (mptcp_send_head(sk) && (push_count > 0)) {
1579 struct mptcp_subflow_context *subflow;
1582 if (mptcp_sched_get_send(msk))
1587 mptcp_for_each_subflow(msk, subflow) {
1588 if (READ_ONCE(subflow->scheduled)) {
1589 mptcp_subflow_set_scheduled(subflow, false);
1592 ssk = mptcp_subflow_tcp_sock(subflow);
1593 if (ssk != prev_ssk) {
1594 /* First check. If the ssk has changed since
1595 * the last round, release prev_ssk
1598 mptcp_push_release(prev_ssk, &info);
1600 /* Need to lock the new subflow only if different
1601 * from the previous one, otherwise we are still
1602 * helding the relevant lock
1609 ret = __subflow_push_pending(sk, ssk, &info);
1611 if (ret != -EAGAIN ||
1612 (1 << ssk->sk_state) &
1613 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1617 do_check_data_fin = true;
1622 /* at this point we held the socket lock for the last subflow we used */
1624 mptcp_push_release(ssk, &info);
1626 /* ensure the rtx timer is running */
1627 if (!mptcp_rtx_timer_pending(sk))
1628 mptcp_reset_rtx_timer(sk);
1629 if (do_check_data_fin)
1630 mptcp_check_send_data_fin(sk);
1633 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1635 struct mptcp_sock *msk = mptcp_sk(sk);
1636 struct mptcp_sendmsg_info info = {
1637 .data_lock_held = true,
1639 bool keep_pushing = true;
1640 struct sock *xmit_ssk;
1644 while (mptcp_send_head(sk) && keep_pushing) {
1645 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1648 /* check for a different subflow usage only after
1649 * spooling the first chunk of data
1652 mptcp_subflow_set_scheduled(subflow, false);
1653 ret = __subflow_push_pending(sk, ssk, &info);
1661 if (mptcp_sched_get_send(msk))
1664 if (READ_ONCE(subflow->scheduled)) {
1665 mptcp_subflow_set_scheduled(subflow, false);
1666 ret = __subflow_push_pending(sk, ssk, &info);
1668 keep_pushing = false;
1672 mptcp_for_each_subflow(msk, subflow) {
1673 if (READ_ONCE(subflow->scheduled)) {
1674 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1675 if (xmit_ssk != ssk) {
1676 mptcp_subflow_delegate(subflow,
1677 MPTCP_DELEGATE_SEND);
1678 keep_pushing = false;
1685 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1686 * not going to flush it via release_sock()
1689 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1691 if (!mptcp_rtx_timer_pending(sk))
1692 mptcp_reset_rtx_timer(sk);
1694 if (msk->snd_data_fin_enable &&
1695 msk->snd_nxt + 1 == msk->write_seq)
1696 mptcp_schedule_work(sk);
1700 static void mptcp_set_nospace(struct sock *sk)
1702 /* enable autotune */
1703 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1705 /* will be cleared on avail space */
1706 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1709 static int mptcp_disconnect(struct sock *sk, int flags);
1711 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1712 size_t len, int *copied_syn)
1714 unsigned int saved_flags = msg->msg_flags;
1715 struct mptcp_sock *msk = mptcp_sk(sk);
1719 /* on flags based fastopen the mptcp is supposed to create the
1720 * first subflow right now. Otherwise we are in the defer_connect
1721 * path, and the first subflow must be already present.
1722 * Since the defer_connect flag is cleared after the first succsful
1723 * fastopen attempt, no need to check for additional subflow status.
1725 if (msg->msg_flags & MSG_FASTOPEN) {
1726 ssk = __mptcp_nmpc_sk(msk);
1728 return PTR_ERR(ssk);
1736 msg->msg_flags |= MSG_DONTWAIT;
1737 msk->fastopening = 1;
1738 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1739 msk->fastopening = 0;
1740 msg->msg_flags = saved_flags;
1743 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1744 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1745 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1746 msg->msg_namelen, msg->msg_flags, 1);
1748 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1749 * case of any error, except timeout or signal
1751 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1753 } else if (ret && ret != -EINPROGRESS) {
1754 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1755 * __inet_stream_connect() can fail, due to looking check,
1756 * see mptcp_disconnect().
1757 * Attempt it again outside the problematic scope.
1759 if (!mptcp_disconnect(sk, 0))
1760 sk->sk_socket->state = SS_UNCONNECTED;
1762 inet_clear_bit(DEFER_CONNECT, sk);
1767 static int do_copy_data_nocache(struct sock *sk, int copy,
1768 struct iov_iter *from, char *to)
1770 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1771 if (!copy_from_iter_full_nocache(to, copy, from))
1773 } else if (!copy_from_iter_full(to, copy, from)) {
1779 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1781 struct mptcp_sock *msk = mptcp_sk(sk);
1782 struct page_frag *pfrag;
1787 /* silently ignore everything else */
1788 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1792 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1793 msg->msg_flags & MSG_FASTOPEN)) {
1796 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1797 copied += copied_syn;
1798 if (ret == -EINPROGRESS && copied_syn > 0)
1804 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1806 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1807 ret = sk_stream_wait_connect(sk, &timeo);
1813 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1816 pfrag = sk_page_frag(sk);
1818 while (msg_data_left(msg)) {
1819 int total_ts, frag_truesize = 0;
1820 struct mptcp_data_frag *dfrag;
1821 bool dfrag_collapsed;
1822 size_t psize, offset;
1824 /* reuse tail pfrag, if possible, or carve a new one from the
1827 dfrag = mptcp_pending_tail(sk);
1828 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1829 if (!dfrag_collapsed) {
1830 if (!sk_stream_memory_free(sk))
1831 goto wait_for_memory;
1833 if (!mptcp_page_frag_refill(sk, pfrag))
1834 goto wait_for_memory;
1836 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1837 frag_truesize = dfrag->overhead;
1840 /* we do not bound vs wspace, to allow a single packet.
1841 * memory accounting will prevent execessive memory usage
1844 offset = dfrag->offset + dfrag->data_len;
1845 psize = pfrag->size - offset;
1846 psize = min_t(size_t, psize, msg_data_left(msg));
1847 total_ts = psize + frag_truesize;
1849 if (!sk_wmem_schedule(sk, total_ts))
1850 goto wait_for_memory;
1852 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1853 page_address(dfrag->page) + offset);
1857 /* data successfully copied into the write queue */
1858 sk_forward_alloc_add(sk, -total_ts);
1860 dfrag->data_len += psize;
1861 frag_truesize += psize;
1862 pfrag->offset += frag_truesize;
1863 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1865 /* charge data on mptcp pending queue to the msk socket
1866 * Note: we charge such data both to sk and ssk
1868 sk_wmem_queued_add(sk, frag_truesize);
1869 if (!dfrag_collapsed) {
1870 get_page(dfrag->page);
1871 list_add_tail(&dfrag->list, &msk->rtx_queue);
1872 if (!msk->first_pending)
1873 WRITE_ONCE(msk->first_pending, dfrag);
1875 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1876 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1882 mptcp_set_nospace(sk);
1883 __mptcp_push_pending(sk, msg->msg_flags);
1884 ret = sk_stream_wait_memory(sk, &timeo);
1890 __mptcp_push_pending(sk, msg->msg_flags);
1900 copied = sk_stream_error(sk, msg->msg_flags, ret);
1904 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1906 size_t len, int flags,
1907 struct scm_timestamping_internal *tss,
1910 struct sk_buff *skb, *tmp;
1913 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1914 u32 offset = MPTCP_SKB_CB(skb)->offset;
1915 u32 data_len = skb->len - offset;
1916 u32 count = min_t(size_t, len - copied, data_len);
1919 if (!(flags & MSG_TRUNC)) {
1920 err = skb_copy_datagram_msg(skb, offset, msg, count);
1921 if (unlikely(err < 0)) {
1928 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1929 tcp_update_recv_tstamps(skb, tss);
1930 *cmsg_flags |= MPTCP_CMSG_TS;
1935 if (count < data_len) {
1936 if (!(flags & MSG_PEEK)) {
1937 MPTCP_SKB_CB(skb)->offset += count;
1938 MPTCP_SKB_CB(skb)->map_seq += count;
1939 msk->bytes_consumed += count;
1944 if (!(flags & MSG_PEEK)) {
1945 /* we will bulk release the skb memory later */
1946 skb->destructor = NULL;
1947 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1948 __skb_unlink(skb, &msk->receive_queue);
1950 msk->bytes_consumed += count;
1960 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1962 * Only difference: Use highest rtt estimate of the subflows in use.
1964 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1966 struct mptcp_subflow_context *subflow;
1967 struct sock *sk = (struct sock *)msk;
1968 u8 scaling_ratio = U8_MAX;
1969 u32 time, advmss = 1;
1972 msk_owned_by_me(msk);
1977 msk->rcvq_space.copied += copied;
1979 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1980 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1982 rtt_us = msk->rcvq_space.rtt_us;
1983 if (rtt_us && time < (rtt_us >> 3))
1987 mptcp_for_each_subflow(msk, subflow) {
1988 const struct tcp_sock *tp;
1992 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1994 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1995 sf_advmss = READ_ONCE(tp->advmss);
1997 rtt_us = max(sf_rtt_us, rtt_us);
1998 advmss = max(sf_advmss, advmss);
1999 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
2002 msk->rcvq_space.rtt_us = rtt_us;
2003 msk->scaling_ratio = scaling_ratio;
2004 if (time < (rtt_us >> 3) || rtt_us == 0)
2007 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2010 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2011 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2015 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2017 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2019 do_div(grow, msk->rcvq_space.space);
2020 rcvwin += (grow << 1);
2022 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2023 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2025 if (rcvbuf > sk->sk_rcvbuf) {
2028 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2029 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2031 /* Make subflows follow along. If we do not do this, we
2032 * get drops at subflow level if skbs can't be moved to
2033 * the mptcp rx queue fast enough (announced rcv_win can
2034 * exceed ssk->sk_rcvbuf).
2036 mptcp_for_each_subflow(msk, subflow) {
2040 ssk = mptcp_subflow_tcp_sock(subflow);
2041 slow = lock_sock_fast(ssk);
2042 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2043 tcp_sk(ssk)->window_clamp = window_clamp;
2044 tcp_cleanup_rbuf(ssk, 1);
2045 unlock_sock_fast(ssk, slow);
2050 msk->rcvq_space.space = msk->rcvq_space.copied;
2052 msk->rcvq_space.copied = 0;
2053 msk->rcvq_space.time = mstamp;
2056 static void __mptcp_update_rmem(struct sock *sk)
2058 struct mptcp_sock *msk = mptcp_sk(sk);
2060 if (!msk->rmem_released)
2063 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2064 mptcp_rmem_uncharge(sk, msk->rmem_released);
2065 WRITE_ONCE(msk->rmem_released, 0);
2068 static void __mptcp_splice_receive_queue(struct sock *sk)
2070 struct mptcp_sock *msk = mptcp_sk(sk);
2072 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2075 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2077 struct sock *sk = (struct sock *)msk;
2078 unsigned int moved = 0;
2082 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2085 /* we can have data pending in the subflows only if the msk
2086 * receive buffer was full at subflow_data_ready() time,
2087 * that is an unlikely slow path.
2092 slowpath = lock_sock_fast(ssk);
2093 mptcp_data_lock(sk);
2094 __mptcp_update_rmem(sk);
2095 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2096 mptcp_data_unlock(sk);
2098 if (unlikely(ssk->sk_err))
2099 __mptcp_error_report(sk);
2100 unlock_sock_fast(ssk, slowpath);
2103 /* acquire the data lock only if some input data is pending */
2105 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2106 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2107 mptcp_data_lock(sk);
2108 __mptcp_update_rmem(sk);
2109 ret |= __mptcp_ofo_queue(msk);
2110 __mptcp_splice_receive_queue(sk);
2111 mptcp_data_unlock(sk);
2114 mptcp_check_data_fin((struct sock *)msk);
2115 return !skb_queue_empty(&msk->receive_queue);
2118 static unsigned int mptcp_inq_hint(const struct sock *sk)
2120 const struct mptcp_sock *msk = mptcp_sk(sk);
2121 const struct sk_buff *skb;
2123 skb = skb_peek(&msk->receive_queue);
2125 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2127 if (hint_val >= INT_MAX)
2130 return (unsigned int)hint_val;
2133 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2139 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2140 int flags, int *addr_len)
2142 struct mptcp_sock *msk = mptcp_sk(sk);
2143 struct scm_timestamping_internal tss;
2144 int copied = 0, cmsg_flags = 0;
2148 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2149 if (unlikely(flags & MSG_ERRQUEUE))
2150 return inet_recv_error(sk, msg, len, addr_len);
2153 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2158 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2160 len = min_t(size_t, len, INT_MAX);
2161 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2163 if (unlikely(msk->recvmsg_inq))
2164 cmsg_flags = MPTCP_CMSG_INQ;
2166 while (copied < len) {
2169 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2170 if (unlikely(bytes_read < 0)) {
2172 copied = bytes_read;
2176 copied += bytes_read;
2178 /* be sure to advertise window change */
2179 mptcp_cleanup_rbuf(msk);
2181 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2184 /* only the master socket status is relevant here. The exit
2185 * conditions mirror closely tcp_recvmsg()
2187 if (copied >= target)
2192 sk->sk_state == TCP_CLOSE ||
2193 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2195 signal_pending(current))
2199 copied = sock_error(sk);
2203 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2204 /* race breaker: the shutdown could be after the
2205 * previous receive queue check
2207 if (__mptcp_move_skbs(msk))
2212 if (sk->sk_state == TCP_CLOSE) {
2222 if (signal_pending(current)) {
2223 copied = sock_intr_errno(timeo);
2228 pr_debug("block timeout %ld", timeo);
2229 sk_wait_data(sk, &timeo, NULL);
2233 if (cmsg_flags && copied >= 0) {
2234 if (cmsg_flags & MPTCP_CMSG_TS)
2235 tcp_recv_timestamp(msg, sk, &tss);
2237 if (cmsg_flags & MPTCP_CMSG_INQ) {
2238 unsigned int inq = mptcp_inq_hint(sk);
2240 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2244 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2245 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2246 skb_queue_empty(&msk->receive_queue), copied);
2247 if (!(flags & MSG_PEEK))
2248 mptcp_rcv_space_adjust(msk, copied);
2254 static void mptcp_retransmit_timer(struct timer_list *t)
2256 struct inet_connection_sock *icsk = from_timer(icsk, t,
2257 icsk_retransmit_timer);
2258 struct sock *sk = &icsk->icsk_inet.sk;
2259 struct mptcp_sock *msk = mptcp_sk(sk);
2262 if (!sock_owned_by_user(sk)) {
2263 /* we need a process context to retransmit */
2264 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2265 mptcp_schedule_work(sk);
2267 /* delegate our work to tcp_release_cb() */
2268 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2274 static void mptcp_tout_timer(struct timer_list *t)
2276 struct sock *sk = from_timer(sk, t, sk_timer);
2278 mptcp_schedule_work(sk);
2282 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2285 * A backup subflow is returned only if that is the only kind available.
2287 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2289 struct sock *backup = NULL, *pick = NULL;
2290 struct mptcp_subflow_context *subflow;
2291 int min_stale_count = INT_MAX;
2293 mptcp_for_each_subflow(msk, subflow) {
2294 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2296 if (!__mptcp_subflow_active(subflow))
2299 /* still data outstanding at TCP level? skip this */
2300 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2301 mptcp_pm_subflow_chk_stale(msk, ssk);
2302 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2306 if (subflow->backup) {
2319 /* use backup only if there are no progresses anywhere */
2320 return min_stale_count > 1 ? backup : NULL;
2323 bool __mptcp_retransmit_pending_data(struct sock *sk)
2325 struct mptcp_data_frag *cur, *rtx_head;
2326 struct mptcp_sock *msk = mptcp_sk(sk);
2328 if (__mptcp_check_fallback(msk))
2331 if (tcp_rtx_and_write_queues_empty(sk))
2334 /* the closing socket has some data untransmitted and/or unacked:
2335 * some data in the mptcp rtx queue has not really xmitted yet.
2336 * keep it simple and re-inject the whole mptcp level rtx queue
2338 mptcp_data_lock(sk);
2339 __mptcp_clean_una_wakeup(sk);
2340 rtx_head = mptcp_rtx_head(sk);
2342 mptcp_data_unlock(sk);
2346 msk->recovery_snd_nxt = msk->snd_nxt;
2347 msk->recovery = true;
2348 mptcp_data_unlock(sk);
2350 msk->first_pending = rtx_head;
2353 /* be sure to clear the "sent status" on all re-injected fragments */
2354 list_for_each_entry(cur, &msk->rtx_queue, list) {
2355 if (!cur->already_sent)
2357 cur->already_sent = 0;
2363 /* flags for __mptcp_close_ssk() */
2364 #define MPTCP_CF_PUSH BIT(1)
2365 #define MPTCP_CF_FASTCLOSE BIT(2)
2367 /* be sure to send a reset only if the caller asked for it, also
2368 * clean completely the subflow status when the subflow reaches
2371 static void __mptcp_subflow_disconnect(struct sock *ssk,
2372 struct mptcp_subflow_context *subflow,
2375 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2376 (flags & MPTCP_CF_FASTCLOSE)) {
2377 /* The MPTCP code never wait on the subflow sockets, TCP-level
2378 * disconnect should never fail
2380 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2381 mptcp_subflow_ctx_reset(subflow);
2383 tcp_shutdown(ssk, SEND_SHUTDOWN);
2387 /* subflow sockets can be either outgoing (connect) or incoming
2390 * Outgoing subflows use in-kernel sockets.
2391 * Incoming subflows do not have their own 'struct socket' allocated,
2392 * so we need to use tcp_close() after detaching them from the mptcp
2395 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2396 struct mptcp_subflow_context *subflow,
2399 struct mptcp_sock *msk = mptcp_sk(sk);
2400 bool dispose_it, need_push = false;
2402 /* If the first subflow moved to a close state before accept, e.g. due
2403 * to an incoming reset or listener shutdown, the subflow socket is
2404 * already deleted by inet_child_forget() and the mptcp socket can't
2407 if (msk->in_accept_queue && msk->first == ssk &&
2408 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2409 /* ensure later check in mptcp_worker() will dispose the msk */
2410 sock_set_flag(sk, SOCK_DEAD);
2411 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2412 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2413 mptcp_subflow_drop_ctx(ssk);
2417 dispose_it = msk->free_first || ssk != msk->first;
2419 list_del(&subflow->node);
2421 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2423 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2424 /* be sure to force the tcp_close path
2425 * to generate the egress reset
2427 ssk->sk_lingertime = 0;
2428 sock_set_flag(ssk, SOCK_LINGER);
2429 subflow->send_fastclose = 1;
2432 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2434 __mptcp_subflow_disconnect(ssk, subflow, flags);
2440 subflow->disposable = 1;
2442 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2443 * the ssk has been already destroyed, we just need to release the
2444 * reference owned by msk;
2446 if (!inet_csk(ssk)->icsk_ulp_ops) {
2447 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2448 kfree_rcu(subflow, rcu);
2450 /* otherwise tcp will dispose of the ssk and subflow ctx */
2451 __tcp_close(ssk, 0);
2453 /* close acquired an extra ref */
2458 __mptcp_subflow_error_report(sk, ssk);
2463 if (ssk == msk->first)
2464 WRITE_ONCE(msk->first, NULL);
2467 __mptcp_sync_sndbuf(sk);
2469 __mptcp_push_pending(sk, 0);
2471 /* Catch every 'all subflows closed' scenario, including peers silently
2472 * closing them, e.g. due to timeout.
2473 * For established sockets, allow an additional timeout before closing,
2474 * as the protocol can still create more subflows.
2476 if (list_is_singular(&msk->conn_list) && msk->first &&
2477 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2478 if (sk->sk_state != TCP_ESTABLISHED ||
2479 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2480 inet_sk_state_store(sk, TCP_CLOSE);
2481 mptcp_close_wake_up(sk);
2483 mptcp_start_tout_timer(sk);
2488 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2489 struct mptcp_subflow_context *subflow)
2491 if (sk->sk_state == TCP_ESTABLISHED)
2492 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2494 /* subflow aborted before reaching the fully_established status
2495 * attempt the creation of the next subflow
2497 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2499 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2502 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2507 static void __mptcp_close_subflow(struct sock *sk)
2509 struct mptcp_subflow_context *subflow, *tmp;
2510 struct mptcp_sock *msk = mptcp_sk(sk);
2514 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2515 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2517 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2520 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2521 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2524 mptcp_close_ssk(sk, ssk, subflow);
2529 static bool mptcp_close_tout_expired(const struct sock *sk)
2531 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2532 sk->sk_state == TCP_CLOSE)
2535 return time_after32(tcp_jiffies32,
2536 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2539 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2541 struct mptcp_subflow_context *subflow, *tmp;
2542 struct sock *sk = (struct sock *)msk;
2544 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2547 mptcp_token_destroy(msk);
2549 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2550 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2553 slow = lock_sock_fast(tcp_sk);
2554 if (tcp_sk->sk_state != TCP_CLOSE) {
2555 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2556 tcp_set_state(tcp_sk, TCP_CLOSE);
2558 unlock_sock_fast(tcp_sk, slow);
2561 /* Mirror the tcp_reset() error propagation */
2562 switch (sk->sk_state) {
2564 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2566 case TCP_CLOSE_WAIT:
2567 WRITE_ONCE(sk->sk_err, EPIPE);
2572 WRITE_ONCE(sk->sk_err, ECONNRESET);
2575 inet_sk_state_store(sk, TCP_CLOSE);
2576 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2577 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2578 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2580 /* the calling mptcp_worker will properly destroy the socket */
2581 if (sock_flag(sk, SOCK_DEAD))
2584 sk->sk_state_change(sk);
2585 sk_error_report(sk);
2588 static void __mptcp_retrans(struct sock *sk)
2590 struct mptcp_sock *msk = mptcp_sk(sk);
2591 struct mptcp_subflow_context *subflow;
2592 struct mptcp_sendmsg_info info = {};
2593 struct mptcp_data_frag *dfrag;
2598 mptcp_clean_una_wakeup(sk);
2600 /* first check ssk: need to kick "stale" logic */
2601 err = mptcp_sched_get_retrans(msk);
2602 dfrag = mptcp_rtx_head(sk);
2604 if (mptcp_data_fin_enabled(msk)) {
2605 struct inet_connection_sock *icsk = inet_csk(sk);
2607 icsk->icsk_retransmits++;
2608 mptcp_set_datafin_timeout(sk);
2609 mptcp_send_ack(msk);
2614 if (!mptcp_send_head(sk))
2623 mptcp_for_each_subflow(msk, subflow) {
2624 if (READ_ONCE(subflow->scheduled)) {
2627 mptcp_subflow_set_scheduled(subflow, false);
2629 ssk = mptcp_subflow_tcp_sock(subflow);
2633 /* limit retransmission to the bytes already sent on some subflows */
2635 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2636 dfrag->already_sent;
2637 while (info.sent < info.limit) {
2638 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2642 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2647 len = max(copied, len);
2648 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2650 WRITE_ONCE(msk->allow_infinite_fallback, false);
2657 msk->bytes_retrans += len;
2658 dfrag->already_sent = max(dfrag->already_sent, len);
2661 mptcp_check_and_set_pending(sk);
2663 if (!mptcp_rtx_timer_pending(sk))
2664 mptcp_reset_rtx_timer(sk);
2667 /* schedule the timeout timer for the relevant event: either close timeout
2668 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2670 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2672 struct sock *sk = (struct sock *)msk;
2673 unsigned long timeout, close_timeout;
2675 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2678 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2679 mptcp_close_timeout(sk);
2681 /* the close timeout takes precedence on the fail one, and here at least one of
2684 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2686 sk_reset_timer(sk, &sk->sk_timer, timeout);
2689 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2691 struct sock *ssk = msk->first;
2697 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2699 slow = lock_sock_fast(ssk);
2700 mptcp_subflow_reset(ssk);
2701 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2702 unlock_sock_fast(ssk, slow);
2705 static void mptcp_do_fastclose(struct sock *sk)
2707 struct mptcp_subflow_context *subflow, *tmp;
2708 struct mptcp_sock *msk = mptcp_sk(sk);
2710 inet_sk_state_store(sk, TCP_CLOSE);
2711 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2712 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2713 subflow, MPTCP_CF_FASTCLOSE);
2716 static void mptcp_worker(struct work_struct *work)
2718 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2719 struct sock *sk = (struct sock *)msk;
2720 unsigned long fail_tout;
2724 state = sk->sk_state;
2725 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2728 mptcp_check_fastclose(msk);
2730 mptcp_pm_nl_work(msk);
2732 mptcp_check_send_data_fin(sk);
2733 mptcp_check_data_fin_ack(sk);
2734 mptcp_check_data_fin(sk);
2736 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2737 __mptcp_close_subflow(sk);
2739 if (mptcp_close_tout_expired(sk)) {
2740 mptcp_do_fastclose(sk);
2741 mptcp_close_wake_up(sk);
2744 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2745 __mptcp_destroy_sock(sk);
2749 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2750 __mptcp_retrans(sk);
2752 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2753 if (fail_tout && time_after(jiffies, fail_tout))
2754 mptcp_mp_fail_no_response(msk);
2761 static void __mptcp_init_sock(struct sock *sk)
2763 struct mptcp_sock *msk = mptcp_sk(sk);
2765 INIT_LIST_HEAD(&msk->conn_list);
2766 INIT_LIST_HEAD(&msk->join_list);
2767 INIT_LIST_HEAD(&msk->rtx_queue);
2768 INIT_WORK(&msk->work, mptcp_worker);
2769 __skb_queue_head_init(&msk->receive_queue);
2770 msk->out_of_order_queue = RB_ROOT;
2771 msk->first_pending = NULL;
2772 msk->rmem_fwd_alloc = 0;
2773 WRITE_ONCE(msk->rmem_released, 0);
2774 msk->timer_ival = TCP_RTO_MIN;
2775 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2777 WRITE_ONCE(msk->first, NULL);
2778 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2779 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2780 WRITE_ONCE(msk->allow_infinite_fallback, true);
2781 msk->recovery = false;
2782 msk->subflow_id = 1;
2784 mptcp_pm_data_init(msk);
2786 /* re-use the csk retrans timer for MPTCP-level retrans */
2787 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2788 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2791 static void mptcp_ca_reset(struct sock *sk)
2793 struct inet_connection_sock *icsk = inet_csk(sk);
2795 tcp_assign_congestion_control(sk);
2796 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2798 /* no need to keep a reference to the ops, the name will suffice */
2799 tcp_cleanup_congestion_control(sk);
2800 icsk->icsk_ca_ops = NULL;
2803 static int mptcp_init_sock(struct sock *sk)
2805 struct net *net = sock_net(sk);
2808 __mptcp_init_sock(sk);
2810 if (!mptcp_is_enabled(net))
2811 return -ENOPROTOOPT;
2813 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2816 ret = mptcp_init_sched(mptcp_sk(sk),
2817 mptcp_sched_find(mptcp_get_scheduler(net)));
2821 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2823 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2824 * propagate the correct value
2828 sk_sockets_allocated_inc(sk);
2829 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2830 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2835 static void __mptcp_clear_xmit(struct sock *sk)
2837 struct mptcp_sock *msk = mptcp_sk(sk);
2838 struct mptcp_data_frag *dtmp, *dfrag;
2840 WRITE_ONCE(msk->first_pending, NULL);
2841 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2842 dfrag_clear(sk, dfrag);
2845 void mptcp_cancel_work(struct sock *sk)
2847 struct mptcp_sock *msk = mptcp_sk(sk);
2849 if (cancel_work_sync(&msk->work))
2853 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2857 switch (ssk->sk_state) {
2859 if (!(how & RCV_SHUTDOWN))
2863 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2866 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2867 pr_debug("Fallback");
2868 ssk->sk_shutdown |= how;
2869 tcp_shutdown(ssk, how);
2871 /* simulate the data_fin ack reception to let the state
2872 * machine move forward
2874 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2875 mptcp_schedule_work(sk);
2877 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2879 if (!mptcp_rtx_timer_pending(sk))
2880 mptcp_reset_rtx_timer(sk);
2888 static const unsigned char new_state[16] = {
2889 /* current state: new state: action: */
2890 [0 /* (Invalid) */] = TCP_CLOSE,
2891 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2892 [TCP_SYN_SENT] = TCP_CLOSE,
2893 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2894 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2895 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2896 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2897 [TCP_CLOSE] = TCP_CLOSE,
2898 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2899 [TCP_LAST_ACK] = TCP_LAST_ACK,
2900 [TCP_LISTEN] = TCP_CLOSE,
2901 [TCP_CLOSING] = TCP_CLOSING,
2902 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2905 static int mptcp_close_state(struct sock *sk)
2907 int next = (int)new_state[sk->sk_state];
2908 int ns = next & TCP_STATE_MASK;
2910 inet_sk_state_store(sk, ns);
2912 return next & TCP_ACTION_FIN;
2915 static void mptcp_check_send_data_fin(struct sock *sk)
2917 struct mptcp_subflow_context *subflow;
2918 struct mptcp_sock *msk = mptcp_sk(sk);
2920 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2921 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2922 msk->snd_nxt, msk->write_seq);
2924 /* we still need to enqueue subflows or not really shutting down,
2927 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2928 mptcp_send_head(sk))
2931 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2933 mptcp_for_each_subflow(msk, subflow) {
2934 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2936 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2940 static void __mptcp_wr_shutdown(struct sock *sk)
2942 struct mptcp_sock *msk = mptcp_sk(sk);
2944 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2945 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2946 !!mptcp_send_head(sk));
2948 /* will be ignored by fallback sockets */
2949 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2950 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2952 mptcp_check_send_data_fin(sk);
2955 static void __mptcp_destroy_sock(struct sock *sk)
2957 struct mptcp_sock *msk = mptcp_sk(sk);
2959 pr_debug("msk=%p", msk);
2963 mptcp_stop_rtx_timer(sk);
2964 sk_stop_timer(sk, &sk->sk_timer);
2966 mptcp_release_sched(msk);
2968 sk->sk_prot->destroy(sk);
2970 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2971 WARN_ON_ONCE(msk->rmem_released);
2972 sk_stream_kill_queues(sk);
2973 xfrm_sk_free_policy(sk);
2978 void __mptcp_unaccepted_force_close(struct sock *sk)
2980 sock_set_flag(sk, SOCK_DEAD);
2981 mptcp_do_fastclose(sk);
2982 __mptcp_destroy_sock(sk);
2985 static __poll_t mptcp_check_readable(struct sock *sk)
2987 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
2990 static void mptcp_check_listen_stop(struct sock *sk)
2994 if (inet_sk_state_load(sk) != TCP_LISTEN)
2997 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2998 ssk = mptcp_sk(sk)->first;
2999 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3002 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3003 tcp_set_state(ssk, TCP_CLOSE);
3004 mptcp_subflow_queue_clean(sk, ssk);
3005 inet_csk_listen_stop(ssk);
3006 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3010 bool __mptcp_close(struct sock *sk, long timeout)
3012 struct mptcp_subflow_context *subflow;
3013 struct mptcp_sock *msk = mptcp_sk(sk);
3014 bool do_cancel_work = false;
3015 int subflows_alive = 0;
3017 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3019 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3020 mptcp_check_listen_stop(sk);
3021 inet_sk_state_store(sk, TCP_CLOSE);
3025 if (mptcp_data_avail(msk) || timeout < 0) {
3026 /* If the msk has read data, or the caller explicitly ask it,
3027 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3029 mptcp_do_fastclose(sk);
3031 } else if (mptcp_close_state(sk)) {
3032 __mptcp_wr_shutdown(sk);
3035 sk_stream_wait_close(sk, timeout);
3038 /* orphan all the subflows */
3039 mptcp_for_each_subflow(msk, subflow) {
3040 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3041 bool slow = lock_sock_fast_nested(ssk);
3043 subflows_alive += ssk->sk_state != TCP_CLOSE;
3045 /* since the close timeout takes precedence on the fail one,
3048 if (ssk == msk->first)
3049 subflow->fail_tout = 0;
3051 /* detach from the parent socket, but allow data_ready to
3052 * push incoming data into the mptcp stack, to properly ack it
3054 ssk->sk_socket = NULL;
3056 unlock_sock_fast(ssk, slow);
3060 /* all the subflows are closed, only timeout can change the msk
3061 * state, let's not keep resources busy for no reasons
3063 if (subflows_alive == 0)
3064 inet_sk_state_store(sk, TCP_CLOSE);
3067 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3069 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3071 if (sk->sk_state == TCP_CLOSE) {
3072 __mptcp_destroy_sock(sk);
3073 do_cancel_work = true;
3075 mptcp_start_tout_timer(sk);
3078 return do_cancel_work;
3081 static void mptcp_close(struct sock *sk, long timeout)
3083 bool do_cancel_work;
3087 do_cancel_work = __mptcp_close(sk, timeout);
3090 mptcp_cancel_work(sk);
3095 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3097 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3098 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3099 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3101 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3102 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3105 msk6->saddr = ssk6->saddr;
3106 msk6->flow_label = ssk6->flow_label;
3110 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3111 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3112 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3113 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3114 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3115 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3118 static int mptcp_disconnect(struct sock *sk, int flags)
3120 struct mptcp_sock *msk = mptcp_sk(sk);
3122 /* We are on the fastopen error path. We can't call straight into the
3123 * subflows cleanup code due to lock nesting (we are already under
3124 * msk->firstsocket lock).
3126 if (msk->fastopening)
3129 mptcp_check_listen_stop(sk);
3130 inet_sk_state_store(sk, TCP_CLOSE);
3132 mptcp_stop_rtx_timer(sk);
3133 mptcp_stop_tout_timer(sk);
3136 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3138 /* msk->subflow is still intact, the following will not free the first
3141 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3142 WRITE_ONCE(msk->flags, 0);
3144 msk->push_pending = 0;
3145 msk->recovery = false;
3146 msk->can_ack = false;
3147 msk->fully_established = false;
3148 msk->rcv_data_fin = false;
3149 msk->snd_data_fin_enable = false;
3150 msk->rcv_fastclose = false;
3151 msk->use_64bit_ack = false;
3152 msk->bytes_consumed = 0;
3153 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3154 mptcp_pm_data_reset(msk);
3156 msk->bytes_acked = 0;
3157 msk->bytes_received = 0;
3158 msk->bytes_sent = 0;
3159 msk->bytes_retrans = 0;
3161 WRITE_ONCE(sk->sk_shutdown, 0);
3162 sk_error_report(sk);
3166 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3167 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3169 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3171 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3175 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3176 const struct mptcp_options_received *mp_opt,
3178 struct request_sock *req)
3180 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3181 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3182 struct mptcp_sock *msk;
3187 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3188 if (nsk->sk_family == AF_INET6)
3189 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3192 __mptcp_init_sock(nsk);
3194 msk = mptcp_sk(nsk);
3195 msk->local_key = subflow_req->local_key;
3196 msk->token = subflow_req->token;
3197 msk->in_accept_queue = 1;
3198 WRITE_ONCE(msk->fully_established, false);
3199 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3200 WRITE_ONCE(msk->csum_enabled, true);
3202 msk->write_seq = subflow_req->idsn + 1;
3203 msk->snd_nxt = msk->write_seq;
3204 msk->snd_una = msk->write_seq;
3205 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3206 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3207 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3209 /* passive msk is created after the first/MPC subflow */
3210 msk->subflow_id = 2;
3212 sock_reset_flag(nsk, SOCK_RCU_FREE);
3213 security_inet_csk_clone(nsk, req);
3215 /* this can't race with mptcp_close(), as the msk is
3216 * not yet exposted to user-space
3218 inet_sk_state_store(nsk, TCP_ESTABLISHED);
3220 /* The msk maintain a ref to each subflow in the connections list */
3221 WRITE_ONCE(msk->first, ssk);
3222 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3225 /* new mpc subflow takes ownership of the newly
3226 * created mptcp socket
3228 mptcp_token_accept(subflow_req, msk);
3230 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3231 * uses the correct data
3233 mptcp_copy_inaddrs(nsk, ssk);
3234 __mptcp_propagate_sndbuf(nsk, ssk);
3236 mptcp_rcv_space_init(msk, ssk);
3237 bh_unlock_sock(nsk);
3239 /* note: the newly allocated socket refcount is 2 now */
3243 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3245 const struct tcp_sock *tp = tcp_sk(ssk);
3247 msk->rcvq_space.copied = 0;
3248 msk->rcvq_space.rtt_us = 0;
3250 msk->rcvq_space.time = tp->tcp_mstamp;
3252 /* initial rcv_space offering made to peer */
3253 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3254 TCP_INIT_CWND * tp->advmss);
3255 if (msk->rcvq_space.space == 0)
3256 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3258 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3261 static struct sock *mptcp_accept(struct sock *ssk, int flags, int *err,
3266 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3267 newsk = inet_csk_accept(ssk, flags, err, kern);
3271 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3272 if (sk_is_mptcp(newsk)) {
3273 struct mptcp_subflow_context *subflow;
3274 struct sock *new_mptcp_sock;
3276 subflow = mptcp_subflow_ctx(newsk);
3277 new_mptcp_sock = subflow->conn;
3279 /* is_mptcp should be false if subflow->conn is missing, see
3280 * subflow_syn_recv_sock()
3282 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3283 tcp_sk(newsk)->is_mptcp = 0;
3287 newsk = new_mptcp_sock;
3288 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3290 MPTCP_INC_STATS(sock_net(ssk),
3291 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3295 newsk->sk_kern_sock = kern;
3299 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3301 struct mptcp_subflow_context *subflow, *tmp;
3302 struct sock *sk = (struct sock *)msk;
3304 __mptcp_clear_xmit(sk);
3306 /* join list will be eventually flushed (with rst) at sock lock release time */
3307 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3308 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3310 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3311 mptcp_data_lock(sk);
3312 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3313 __skb_queue_purge(&sk->sk_receive_queue);
3314 skb_rbtree_purge(&msk->out_of_order_queue);
3315 mptcp_data_unlock(sk);
3317 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3318 * inet_sock_destruct() will dispose it
3320 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3321 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3322 mptcp_token_destroy(msk);
3323 mptcp_pm_free_anno_list(msk);
3324 mptcp_free_local_addr_list(msk);
3327 static void mptcp_destroy(struct sock *sk)
3329 struct mptcp_sock *msk = mptcp_sk(sk);
3331 /* allow the following to close even the initial subflow */
3332 msk->free_first = 1;
3333 mptcp_destroy_common(msk, 0);
3334 sk_sockets_allocated_dec(sk);
3337 void __mptcp_data_acked(struct sock *sk)
3339 if (!sock_owned_by_user(sk))
3340 __mptcp_clean_una(sk);
3342 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3344 if (mptcp_pending_data_fin_ack(sk))
3345 mptcp_schedule_work(sk);
3348 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3350 if (!mptcp_send_head(sk))
3353 if (!sock_owned_by_user(sk))
3354 __mptcp_subflow_push_pending(sk, ssk, false);
3356 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3359 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3360 BIT(MPTCP_RETRANSMIT) | \
3361 BIT(MPTCP_FLUSH_JOIN_LIST))
3363 /* processes deferred events and flush wmem */
3364 static void mptcp_release_cb(struct sock *sk)
3365 __must_hold(&sk->sk_lock.slock)
3367 struct mptcp_sock *msk = mptcp_sk(sk);
3370 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3372 struct list_head join_list;
3377 INIT_LIST_HEAD(&join_list);
3378 list_splice_init(&msk->join_list, &join_list);
3380 /* the following actions acquire the subflow socket lock
3382 * 1) can't be invoked in atomic scope
3383 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3384 * datapath acquires the msk socket spinlock while helding
3385 * the subflow socket lock
3387 msk->push_pending = 0;
3388 msk->cb_flags &= ~flags;
3389 spin_unlock_bh(&sk->sk_lock.slock);
3391 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3392 __mptcp_flush_join_list(sk, &join_list);
3393 if (flags & BIT(MPTCP_PUSH_PENDING))
3394 __mptcp_push_pending(sk, 0);
3395 if (flags & BIT(MPTCP_RETRANSMIT))
3396 __mptcp_retrans(sk);
3399 spin_lock_bh(&sk->sk_lock.slock);
3402 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3403 __mptcp_clean_una_wakeup(sk);
3404 if (unlikely(msk->cb_flags)) {
3405 /* be sure to sync the msk state before taking actions
3406 * depending on sk_state (MPTCP_ERROR_REPORT)
3407 * On sk release avoid actions depending on the first subflow
3409 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3410 __mptcp_sync_state(sk, msk->pending_state);
3411 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3412 __mptcp_error_report(sk);
3413 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3414 __mptcp_sync_sndbuf(sk);
3417 __mptcp_update_rmem(sk);
3420 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3421 * TCP can't schedule delack timer before the subflow is fully established.
3422 * MPTCP uses the delack timer to do 3rd ack retransmissions
3424 static void schedule_3rdack_retransmission(struct sock *ssk)
3426 struct inet_connection_sock *icsk = inet_csk(ssk);
3427 struct tcp_sock *tp = tcp_sk(ssk);
3428 unsigned long timeout;
3430 if (mptcp_subflow_ctx(ssk)->fully_established)
3433 /* reschedule with a timeout above RTT, as we must look only for drop */
3435 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3437 timeout = TCP_TIMEOUT_INIT;
3440 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3441 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3442 icsk->icsk_ack.timeout = timeout;
3443 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3446 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3448 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3449 struct sock *sk = subflow->conn;
3451 if (status & BIT(MPTCP_DELEGATE_SEND)) {
3452 mptcp_data_lock(sk);
3453 if (!sock_owned_by_user(sk))
3454 __mptcp_subflow_push_pending(sk, ssk, true);
3456 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3457 mptcp_data_unlock(sk);
3459 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3460 mptcp_data_lock(sk);
3461 if (!sock_owned_by_user(sk))
3462 __mptcp_sync_sndbuf(sk);
3464 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3465 mptcp_data_unlock(sk);
3467 if (status & BIT(MPTCP_DELEGATE_ACK))
3468 schedule_3rdack_retransmission(ssk);
3471 static int mptcp_hash(struct sock *sk)
3473 /* should never be called,
3474 * we hash the TCP subflows not the master socket
3480 static void mptcp_unhash(struct sock *sk)
3482 /* called from sk_common_release(), but nothing to do here */
3485 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3487 struct mptcp_sock *msk = mptcp_sk(sk);
3489 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3490 if (WARN_ON_ONCE(!msk->first))
3493 return inet_csk_get_port(msk->first, snum);
3496 void mptcp_finish_connect(struct sock *ssk)
3498 struct mptcp_subflow_context *subflow;
3499 struct mptcp_sock *msk;
3502 subflow = mptcp_subflow_ctx(ssk);
3506 pr_debug("msk=%p, token=%u", sk, subflow->token);
3508 subflow->map_seq = subflow->iasn;
3509 subflow->map_subflow_seq = 1;
3511 /* the socket is not connected yet, no msk/subflow ops can access/race
3512 * accessing the field below
3514 WRITE_ONCE(msk->local_key, subflow->local_key);
3515 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3516 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3517 WRITE_ONCE(msk->snd_una, msk->write_seq);
3519 mptcp_pm_new_connection(msk, ssk, 0);
3521 mptcp_rcv_space_init(msk, ssk);
3524 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3526 write_lock_bh(&sk->sk_callback_lock);
3527 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3528 sk_set_socket(sk, parent);
3529 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3530 write_unlock_bh(&sk->sk_callback_lock);
3533 bool mptcp_finish_join(struct sock *ssk)
3535 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3536 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3537 struct sock *parent = (void *)msk;
3540 pr_debug("msk=%p, subflow=%p", msk, subflow);
3542 /* mptcp socket already closing? */
3543 if (!mptcp_is_fully_established(parent)) {
3544 subflow->reset_reason = MPTCP_RST_EMPTCP;
3548 /* active subflow, already present inside the conn_list */
3549 if (!list_empty(&subflow->node)) {
3550 mptcp_subflow_joined(msk, ssk);
3551 mptcp_propagate_sndbuf(parent, ssk);
3555 if (!mptcp_pm_allow_new_subflow(msk))
3556 goto err_prohibited;
3558 /* If we can't acquire msk socket lock here, let the release callback
3561 mptcp_data_lock(parent);
3562 if (!sock_owned_by_user(parent)) {
3563 ret = __mptcp_finish_join(msk, ssk);
3566 list_add_tail(&subflow->node, &msk->conn_list);
3570 list_add_tail(&subflow->node, &msk->join_list);
3571 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3573 mptcp_data_unlock(parent);
3577 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3584 static void mptcp_shutdown(struct sock *sk, int how)
3586 pr_debug("sk=%p, how=%d", sk, how);
3588 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3589 __mptcp_wr_shutdown(sk);
3592 static int mptcp_forward_alloc_get(const struct sock *sk)
3594 return READ_ONCE(sk->sk_forward_alloc) +
3595 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3598 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3600 const struct sock *sk = (void *)msk;
3603 if (sk->sk_state == TCP_LISTEN)
3606 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3609 delta = msk->write_seq - v;
3610 if (__mptcp_check_fallback(msk) && msk->first) {
3611 struct tcp_sock *tp = tcp_sk(msk->first);
3613 /* the first subflow is disconnected after close - see
3614 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3615 * so ignore that status, too.
3617 if (!((1 << msk->first->sk_state) &
3618 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3619 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3621 if (delta > INT_MAX)
3627 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3629 struct mptcp_sock *msk = mptcp_sk(sk);
3634 if (sk->sk_state == TCP_LISTEN)
3638 __mptcp_move_skbs(msk);
3639 *karg = mptcp_inq_hint(sk);
3643 slow = lock_sock_fast(sk);
3644 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3645 unlock_sock_fast(sk, slow);
3648 slow = lock_sock_fast(sk);
3649 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3650 unlock_sock_fast(sk, slow);
3653 return -ENOIOCTLCMD;
3659 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3660 struct mptcp_subflow_context *subflow)
3662 subflow->request_mptcp = 0;
3663 __mptcp_do_fallback(msk);
3666 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3668 struct mptcp_subflow_context *subflow;
3669 struct mptcp_sock *msk = mptcp_sk(sk);
3673 ssk = __mptcp_nmpc_sk(msk);
3675 return PTR_ERR(ssk);
3677 inet_sk_state_store(sk, TCP_SYN_SENT);
3678 subflow = mptcp_subflow_ctx(ssk);
3679 #ifdef CONFIG_TCP_MD5SIG
3680 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3683 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3684 mptcp_subflow_early_fallback(msk, subflow);
3686 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3687 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3688 mptcp_subflow_early_fallback(msk, subflow);
3690 if (likely(!__mptcp_check_fallback(msk)))
3691 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3693 /* if reaching here via the fastopen/sendmsg path, the caller already
3694 * acquired the subflow socket lock, too.
3696 if (!msk->fastopening)
3699 /* the following mirrors closely a very small chunk of code from
3700 * __inet_stream_connect()
3702 if (ssk->sk_state != TCP_CLOSE)
3705 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3706 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3711 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3715 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3718 if (!msk->fastopening)
3721 /* on successful connect, the msk state will be moved to established by
3722 * subflow_finish_connect()
3724 if (unlikely(err)) {
3725 /* avoid leaving a dangling token in an unconnected socket */
3726 mptcp_token_destroy(msk);
3727 inet_sk_state_store(sk, TCP_CLOSE);
3731 mptcp_copy_inaddrs(sk, ssk);
3735 static struct proto mptcp_prot = {
3737 .owner = THIS_MODULE,
3738 .init = mptcp_init_sock,
3739 .connect = mptcp_connect,
3740 .disconnect = mptcp_disconnect,
3741 .close = mptcp_close,
3742 .accept = mptcp_accept,
3743 .setsockopt = mptcp_setsockopt,
3744 .getsockopt = mptcp_getsockopt,
3745 .shutdown = mptcp_shutdown,
3746 .destroy = mptcp_destroy,
3747 .sendmsg = mptcp_sendmsg,
3748 .ioctl = mptcp_ioctl,
3749 .recvmsg = mptcp_recvmsg,
3750 .release_cb = mptcp_release_cb,
3752 .unhash = mptcp_unhash,
3753 .get_port = mptcp_get_port,
3754 .forward_alloc_get = mptcp_forward_alloc_get,
3755 .sockets_allocated = &mptcp_sockets_allocated,
3757 .memory_allocated = &tcp_memory_allocated,
3758 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3760 .memory_pressure = &tcp_memory_pressure,
3761 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3762 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3763 .sysctl_mem = sysctl_tcp_mem,
3764 .obj_size = sizeof(struct mptcp_sock),
3765 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3766 .no_autobind = true,
3769 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3771 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3772 struct sock *ssk, *sk = sock->sk;
3776 ssk = __mptcp_nmpc_sk(msk);
3782 if (sk->sk_family == AF_INET)
3783 err = inet_bind_sk(ssk, uaddr, addr_len);
3784 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3785 else if (sk->sk_family == AF_INET6)
3786 err = inet6_bind_sk(ssk, uaddr, addr_len);
3789 mptcp_copy_inaddrs(sk, ssk);
3796 static int mptcp_listen(struct socket *sock, int backlog)
3798 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3799 struct sock *sk = sock->sk;
3803 pr_debug("msk=%p", msk);
3808 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3811 ssk = __mptcp_nmpc_sk(msk);
3817 inet_sk_state_store(sk, TCP_LISTEN);
3818 sock_set_flag(sk, SOCK_RCU_FREE);
3821 err = __inet_listen_sk(ssk, backlog);
3823 inet_sk_state_store(sk, inet_sk_state_load(ssk));
3826 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3827 mptcp_copy_inaddrs(sk, ssk);
3828 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3836 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3837 int flags, bool kern)
3839 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3840 struct sock *ssk, *newsk;
3843 pr_debug("msk=%p", msk);
3845 /* Buggy applications can call accept on socket states other then LISTEN
3846 * but no need to allocate the first subflow just to error out.
3848 ssk = READ_ONCE(msk->first);
3852 newsk = mptcp_accept(ssk, flags, &err, kern);
3858 __inet_accept(sock, newsock, newsk);
3859 if (!mptcp_is_tcpsk(newsock->sk)) {
3860 struct mptcp_sock *msk = mptcp_sk(newsk);
3861 struct mptcp_subflow_context *subflow;
3863 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3864 msk->in_accept_queue = 0;
3866 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3867 * This is needed so NOSPACE flag can be set from tcp stack.
3869 mptcp_for_each_subflow(msk, subflow) {
3870 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3872 if (!ssk->sk_socket)
3873 mptcp_sock_graft(ssk, newsock);
3876 /* Do late cleanup for the first subflow as necessary. Also
3877 * deal with bad peers not doing a complete shutdown.
3879 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3880 __mptcp_close_ssk(newsk, msk->first,
3881 mptcp_subflow_ctx(msk->first), 0);
3882 if (unlikely(list_is_singular(&msk->conn_list)))
3883 inet_sk_state_store(newsk, TCP_CLOSE);
3886 release_sock(newsk);
3891 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3893 struct sock *sk = (struct sock *)msk;
3895 if (sk_stream_is_writeable(sk))
3896 return EPOLLOUT | EPOLLWRNORM;
3898 mptcp_set_nospace(sk);
3899 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3900 if (sk_stream_is_writeable(sk))
3901 return EPOLLOUT | EPOLLWRNORM;
3906 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3907 struct poll_table_struct *wait)
3909 struct sock *sk = sock->sk;
3910 struct mptcp_sock *msk;
3916 sock_poll_wait(file, sock, wait);
3918 state = inet_sk_state_load(sk);
3919 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3920 if (state == TCP_LISTEN) {
3921 struct sock *ssk = READ_ONCE(msk->first);
3923 if (WARN_ON_ONCE(!ssk))
3926 return inet_csk_listen_poll(ssk);
3929 shutdown = READ_ONCE(sk->sk_shutdown);
3930 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3932 if (shutdown & RCV_SHUTDOWN)
3933 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3935 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3936 mask |= mptcp_check_readable(sk);
3937 if (shutdown & SEND_SHUTDOWN)
3938 mask |= EPOLLOUT | EPOLLWRNORM;
3940 mask |= mptcp_check_writeable(msk);
3941 } else if (state == TCP_SYN_SENT &&
3942 inet_test_bit(DEFER_CONNECT, sk)) {
3943 /* cf tcp_poll() note about TFO */
3944 mask |= EPOLLOUT | EPOLLWRNORM;
3947 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3949 if (READ_ONCE(sk->sk_err))
3955 static const struct proto_ops mptcp_stream_ops = {
3957 .owner = THIS_MODULE,
3958 .release = inet_release,
3960 .connect = inet_stream_connect,
3961 .socketpair = sock_no_socketpair,
3962 .accept = mptcp_stream_accept,
3963 .getname = inet_getname,
3965 .ioctl = inet_ioctl,
3966 .gettstamp = sock_gettstamp,
3967 .listen = mptcp_listen,
3968 .shutdown = inet_shutdown,
3969 .setsockopt = sock_common_setsockopt,
3970 .getsockopt = sock_common_getsockopt,
3971 .sendmsg = inet_sendmsg,
3972 .recvmsg = inet_recvmsg,
3973 .mmap = sock_no_mmap,
3974 .set_rcvlowat = mptcp_set_rcvlowat,
3977 static struct inet_protosw mptcp_protosw = {
3978 .type = SOCK_STREAM,
3979 .protocol = IPPROTO_MPTCP,
3980 .prot = &mptcp_prot,
3981 .ops = &mptcp_stream_ops,
3982 .flags = INET_PROTOSW_ICSK,
3985 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3987 struct mptcp_delegated_action *delegated;
3988 struct mptcp_subflow_context *subflow;
3991 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3992 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3993 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3995 bh_lock_sock_nested(ssk);
3996 if (!sock_owned_by_user(ssk)) {
3997 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
3999 /* tcp_release_cb_override already processed
4000 * the action or will do at next release_sock().
4001 * In both case must dequeue the subflow here - on the same
4002 * CPU that scheduled it.
4005 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4007 bh_unlock_sock(ssk);
4010 if (++work_done == budget)
4014 /* always provide a 0 'work_done' argument, so that napi_complete_done
4015 * will not try accessing the NULL napi->dev ptr
4017 napi_complete_done(napi, 0);
4021 void __init mptcp_proto_init(void)
4023 struct mptcp_delegated_action *delegated;
4026 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4028 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4029 panic("Failed to allocate MPTCP pcpu counter\n");
4031 init_dummy_netdev(&mptcp_napi_dev);
4032 for_each_possible_cpu(cpu) {
4033 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4034 INIT_LIST_HEAD(&delegated->head);
4035 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4037 napi_enable(&delegated->napi);
4040 mptcp_subflow_init();
4045 if (proto_register(&mptcp_prot, 1) != 0)
4046 panic("Failed to register MPTCP proto.\n");
4048 inet_register_protosw(&mptcp_protosw);
4050 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4053 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4054 static const struct proto_ops mptcp_v6_stream_ops = {
4056 .owner = THIS_MODULE,
4057 .release = inet6_release,
4059 .connect = inet_stream_connect,
4060 .socketpair = sock_no_socketpair,
4061 .accept = mptcp_stream_accept,
4062 .getname = inet6_getname,
4064 .ioctl = inet6_ioctl,
4065 .gettstamp = sock_gettstamp,
4066 .listen = mptcp_listen,
4067 .shutdown = inet_shutdown,
4068 .setsockopt = sock_common_setsockopt,
4069 .getsockopt = sock_common_getsockopt,
4070 .sendmsg = inet6_sendmsg,
4071 .recvmsg = inet6_recvmsg,
4072 .mmap = sock_no_mmap,
4073 #ifdef CONFIG_COMPAT
4074 .compat_ioctl = inet6_compat_ioctl,
4076 .set_rcvlowat = mptcp_set_rcvlowat,
4079 static struct proto mptcp_v6_prot;
4081 static struct inet_protosw mptcp_v6_protosw = {
4082 .type = SOCK_STREAM,
4083 .protocol = IPPROTO_MPTCP,
4084 .prot = &mptcp_v6_prot,
4085 .ops = &mptcp_v6_stream_ops,
4086 .flags = INET_PROTOSW_ICSK,
4089 int __init mptcp_proto_v6_init(void)
4093 mptcp_v6_prot = mptcp_prot;
4094 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4095 mptcp_v6_prot.slab = NULL;
4096 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4097 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4099 err = proto_register(&mptcp_v6_prot, 1);
4103 err = inet6_register_protosw(&mptcp_v6_protosw);
4105 proto_unregister(&mptcp_v6_prot);