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>
18 #include <net/tcp_states.h>
19 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
20 #include <net/transp_v6.h>
22 #include <net/mptcp.h>
23 #include <net/hotdata.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 const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk)
60 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
61 if (sk->sk_prot == &tcpv6_prot)
62 return &inet6_stream_ops;
64 WARN_ON_ONCE(sk->sk_prot != &tcp_prot);
65 return &inet_stream_ops;
68 static int __mptcp_socket_create(struct mptcp_sock *msk)
70 struct mptcp_subflow_context *subflow;
71 struct sock *sk = (struct sock *)msk;
75 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
79 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
80 WRITE_ONCE(msk->first, ssock->sk);
81 subflow = mptcp_subflow_ctx(ssock->sk);
82 list_add(&subflow->node, &msk->conn_list);
84 subflow->request_mptcp = 1;
85 subflow->subflow_id = msk->subflow_id++;
87 /* This is the first subflow, always with id 0 */
88 WRITE_ONCE(subflow->local_id, 0);
89 mptcp_sock_graft(msk->first, sk->sk_socket);
90 iput(SOCK_INODE(ssock));
95 /* If the MPC handshake is not started, returns the first subflow,
96 * eventually allocating it.
98 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
100 struct sock *sk = (struct sock *)msk;
103 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
104 return ERR_PTR(-EINVAL);
107 ret = __mptcp_socket_create(msk);
115 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
117 sk_drops_add(sk, skb);
121 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
123 WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
124 mptcp_sk(sk)->rmem_fwd_alloc + size);
127 static void mptcp_rmem_charge(struct sock *sk, int size)
129 mptcp_rmem_fwd_alloc_add(sk, -size);
132 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
133 struct sk_buff *from)
138 if (MPTCP_SKB_CB(from)->offset ||
139 !skb_try_coalesce(to, from, &fragstolen, &delta))
142 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
143 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
144 to->len, MPTCP_SKB_CB(from)->end_seq);
145 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
147 /* note the fwd memory can reach a negative value after accounting
148 * for the delta, but the later skb free will restore a non
151 atomic_add(delta, &sk->sk_rmem_alloc);
152 mptcp_rmem_charge(sk, delta);
153 kfree_skb_partial(from, fragstolen);
158 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
159 struct sk_buff *from)
161 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
164 return mptcp_try_coalesce((struct sock *)msk, to, from);
167 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
169 amount >>= PAGE_SHIFT;
170 mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
171 __sk_mem_reduce_allocated(sk, amount);
174 static void mptcp_rmem_uncharge(struct sock *sk, int size)
176 struct mptcp_sock *msk = mptcp_sk(sk);
179 mptcp_rmem_fwd_alloc_add(sk, size);
180 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
182 /* see sk_mem_uncharge() for the rationale behind the following schema */
183 if (unlikely(reclaimable >= PAGE_SIZE))
184 __mptcp_rmem_reclaim(sk, reclaimable);
187 static void mptcp_rfree(struct sk_buff *skb)
189 unsigned int len = skb->truesize;
190 struct sock *sk = skb->sk;
192 atomic_sub(len, &sk->sk_rmem_alloc);
193 mptcp_rmem_uncharge(sk, len);
196 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
200 skb->destructor = mptcp_rfree;
201 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
202 mptcp_rmem_charge(sk, skb->truesize);
205 /* "inspired" by tcp_data_queue_ofo(), main differences:
207 * - don't cope with sacks
209 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
211 struct sock *sk = (struct sock *)msk;
212 struct rb_node **p, *parent;
213 u64 seq, end_seq, max_seq;
214 struct sk_buff *skb1;
216 seq = MPTCP_SKB_CB(skb)->map_seq;
217 end_seq = MPTCP_SKB_CB(skb)->end_seq;
218 max_seq = atomic64_read(&msk->rcv_wnd_sent);
220 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
221 RB_EMPTY_ROOT(&msk->out_of_order_queue));
222 if (after64(end_seq, max_seq)) {
225 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
226 (unsigned long long)end_seq - (unsigned long)max_seq,
227 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
228 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
232 p = &msk->out_of_order_queue.rb_node;
233 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
234 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
235 rb_link_node(&skb->rbnode, NULL, p);
236 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
237 msk->ooo_last_skb = skb;
241 /* with 2 subflows, adding at end of ooo queue is quite likely
242 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
244 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
245 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
250 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
251 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
252 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
253 parent = &msk->ooo_last_skb->rbnode;
254 p = &parent->rb_right;
258 /* Find place to insert this segment. Handle overlaps on the way. */
262 skb1 = rb_to_skb(parent);
263 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
264 p = &parent->rb_left;
267 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
268 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
269 /* All the bits are present. Drop. */
271 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
274 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
278 * continue traversing
281 /* skb's seq == skb1's seq and skb covers skb1.
282 * Replace skb1 with skb.
284 rb_replace_node(&skb1->rbnode, &skb->rbnode,
285 &msk->out_of_order_queue);
286 mptcp_drop(sk, skb1);
287 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
290 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
291 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
294 p = &parent->rb_right;
298 /* Insert segment into RB tree. */
299 rb_link_node(&skb->rbnode, parent, p);
300 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
303 /* Remove other segments covered by skb. */
304 while ((skb1 = skb_rb_next(skb)) != NULL) {
305 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
307 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
308 mptcp_drop(sk, skb1);
309 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
311 /* If there is no skb after us, we are the last_skb ! */
313 msk->ooo_last_skb = skb;
317 mptcp_set_owner_r(skb, sk);
320 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
322 struct mptcp_sock *msk = mptcp_sk(sk);
325 if (size <= msk->rmem_fwd_alloc)
328 size -= msk->rmem_fwd_alloc;
329 amt = sk_mem_pages(size);
330 amount = amt << PAGE_SHIFT;
331 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
334 mptcp_rmem_fwd_alloc_add(sk, amount);
338 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
339 struct sk_buff *skb, unsigned int offset,
342 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
343 struct sock *sk = (struct sock *)msk;
344 struct sk_buff *tail;
347 __skb_unlink(skb, &ssk->sk_receive_queue);
352 /* try to fetch required memory from subflow */
353 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
356 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
358 /* the skb map_seq accounts for the skb offset:
359 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
362 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
363 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
364 MPTCP_SKB_CB(skb)->offset = offset;
365 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
367 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
369 msk->bytes_received += copy_len;
370 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
371 tail = skb_peek_tail(&sk->sk_receive_queue);
372 if (tail && mptcp_try_coalesce(sk, tail, skb))
375 mptcp_set_owner_r(skb, sk);
376 __skb_queue_tail(&sk->sk_receive_queue, skb);
378 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
379 mptcp_data_queue_ofo(msk, skb);
383 /* old data, keep it simple and drop the whole pkt, sender
384 * will retransmit as needed, if needed.
386 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
392 static void mptcp_stop_rtx_timer(struct sock *sk)
394 struct inet_connection_sock *icsk = inet_csk(sk);
396 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
397 mptcp_sk(sk)->timer_ival = 0;
400 static void mptcp_close_wake_up(struct sock *sk)
402 if (sock_flag(sk, SOCK_DEAD))
405 sk->sk_state_change(sk);
406 if (sk->sk_shutdown == SHUTDOWN_MASK ||
407 sk->sk_state == TCP_CLOSE)
408 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
410 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
413 /* called under the msk socket lock */
414 static bool mptcp_pending_data_fin_ack(struct sock *sk)
416 struct mptcp_sock *msk = mptcp_sk(sk);
418 return ((1 << sk->sk_state) &
419 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
420 msk->write_seq == READ_ONCE(msk->snd_una);
423 static void mptcp_check_data_fin_ack(struct sock *sk)
425 struct mptcp_sock *msk = mptcp_sk(sk);
427 /* Look for an acknowledged DATA_FIN */
428 if (mptcp_pending_data_fin_ack(sk)) {
429 WRITE_ONCE(msk->snd_data_fin_enable, 0);
431 switch (sk->sk_state) {
433 mptcp_set_state(sk, TCP_FIN_WAIT2);
437 mptcp_set_state(sk, TCP_CLOSE);
441 mptcp_close_wake_up(sk);
445 /* can be called with no lock acquired */
446 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
448 struct mptcp_sock *msk = mptcp_sk(sk);
450 if (READ_ONCE(msk->rcv_data_fin) &&
451 ((1 << inet_sk_state_load(sk)) &
452 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
453 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
455 if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) {
457 *seq = rcv_data_fin_seq;
466 static void mptcp_set_datafin_timeout(struct sock *sk)
468 struct inet_connection_sock *icsk = inet_csk(sk);
471 retransmits = min_t(u32, icsk->icsk_retransmits,
472 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
474 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
477 static void __mptcp_set_timeout(struct sock *sk, long tout)
479 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
482 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
484 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
486 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
487 inet_csk(ssk)->icsk_timeout - jiffies : 0;
490 static void mptcp_set_timeout(struct sock *sk)
492 struct mptcp_subflow_context *subflow;
495 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
496 tout = max(tout, mptcp_timeout_from_subflow(subflow));
497 __mptcp_set_timeout(sk, tout);
500 static inline bool tcp_can_send_ack(const struct sock *ssk)
502 return !((1 << inet_sk_state_load(ssk)) &
503 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
506 void __mptcp_subflow_send_ack(struct sock *ssk)
508 if (tcp_can_send_ack(ssk))
512 static void mptcp_subflow_send_ack(struct sock *ssk)
516 slow = lock_sock_fast(ssk);
517 __mptcp_subflow_send_ack(ssk);
518 unlock_sock_fast(ssk, slow);
521 static void mptcp_send_ack(struct mptcp_sock *msk)
523 struct mptcp_subflow_context *subflow;
525 mptcp_for_each_subflow(msk, subflow)
526 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
529 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
533 slow = lock_sock_fast(ssk);
534 if (tcp_can_send_ack(ssk))
535 tcp_cleanup_rbuf(ssk, 1);
536 unlock_sock_fast(ssk, slow);
539 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
541 const struct inet_connection_sock *icsk = inet_csk(ssk);
542 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
543 const struct tcp_sock *tp = tcp_sk(ssk);
545 return (ack_pending & ICSK_ACK_SCHED) &&
546 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
547 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
548 (rx_empty && ack_pending &
549 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
552 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
554 int old_space = READ_ONCE(msk->old_wspace);
555 struct mptcp_subflow_context *subflow;
556 struct sock *sk = (struct sock *)msk;
557 int space = __mptcp_space(sk);
558 bool cleanup, rx_empty;
560 cleanup = (space > 0) && (space >= (old_space << 1));
561 rx_empty = !__mptcp_rmem(sk);
563 mptcp_for_each_subflow(msk, subflow) {
564 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
566 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
567 mptcp_subflow_cleanup_rbuf(ssk);
571 static bool mptcp_check_data_fin(struct sock *sk)
573 struct mptcp_sock *msk = mptcp_sk(sk);
574 u64 rcv_data_fin_seq;
577 /* Need to ack a DATA_FIN received from a peer while this side
578 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
579 * msk->rcv_data_fin was set when parsing the incoming options
580 * at the subflow level and the msk lock was not held, so this
581 * is the first opportunity to act on the DATA_FIN and change
584 * If we are caught up to the sequence number of the incoming
585 * DATA_FIN, send the DATA_ACK now and do state transition. If
586 * not caught up, do nothing and let the recv code send DATA_ACK
590 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
591 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
592 WRITE_ONCE(msk->rcv_data_fin, 0);
594 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
595 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
597 switch (sk->sk_state) {
598 case TCP_ESTABLISHED:
599 mptcp_set_state(sk, TCP_CLOSE_WAIT);
602 mptcp_set_state(sk, TCP_CLOSING);
605 mptcp_set_state(sk, TCP_CLOSE);
608 /* Other states not expected */
614 if (!__mptcp_check_fallback(msk))
616 mptcp_close_wake_up(sk);
621 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
625 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
626 struct sock *sk = (struct sock *)msk;
627 unsigned int moved = 0;
628 bool more_data_avail;
633 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
635 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
636 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
638 if (unlikely(ssk_rbuf > sk_rbuf)) {
639 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
644 pr_debug("msk=%p ssk=%p", msk, ssk);
647 u32 map_remaining, offset;
648 u32 seq = tp->copied_seq;
652 /* try to move as much data as available */
653 map_remaining = subflow->map_data_len -
654 mptcp_subflow_get_map_offset(subflow);
656 skb = skb_peek(&ssk->sk_receive_queue);
658 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
659 * a different CPU can have already processed the pending
660 * data, stop here or we can enter an infinite loop
667 if (__mptcp_check_fallback(msk)) {
668 /* Under fallback skbs have no MPTCP extension and TCP could
669 * collapse them between the dummy map creation and the
670 * current dequeue. Be sure to adjust the map size.
672 map_remaining = skb->len;
673 subflow->map_data_len = skb->len;
676 offset = seq - TCP_SKB_CB(skb)->seq;
677 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
683 if (offset < skb->len) {
684 size_t len = skb->len - offset;
689 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
693 if (WARN_ON_ONCE(map_remaining < len))
697 sk_eat_skb(ssk, skb);
701 WRITE_ONCE(tp->copied_seq, seq);
702 more_data_avail = mptcp_subflow_data_available(ssk);
704 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
708 } while (more_data_avail);
711 msk->last_data_recv = tcp_jiffies32;
716 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
718 struct sock *sk = (struct sock *)msk;
719 struct sk_buff *skb, *tail;
724 p = rb_first(&msk->out_of_order_queue);
725 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
728 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
732 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
734 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
737 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
741 end_seq = MPTCP_SKB_CB(skb)->end_seq;
742 tail = skb_peek_tail(&sk->sk_receive_queue);
743 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
744 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
746 /* skip overlapping data, if any */
747 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
748 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
750 MPTCP_SKB_CB(skb)->offset += delta;
751 MPTCP_SKB_CB(skb)->map_seq += delta;
752 __skb_queue_tail(&sk->sk_receive_queue, skb);
754 msk->bytes_received += end_seq - msk->ack_seq;
755 WRITE_ONCE(msk->ack_seq, end_seq);
761 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
763 int err = sock_error(ssk);
769 /* only propagate errors on fallen-back sockets or
772 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
775 /* We need to propagate only transition to CLOSE state.
776 * Orphaned socket will see such state change via
777 * subflow_sched_work_if_closed() and that path will properly
778 * destroy the msk as needed.
780 ssk_state = inet_sk_state_load(ssk);
781 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
782 mptcp_set_state(sk, ssk_state);
783 WRITE_ONCE(sk->sk_err, -err);
785 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
791 void __mptcp_error_report(struct sock *sk)
793 struct mptcp_subflow_context *subflow;
794 struct mptcp_sock *msk = mptcp_sk(sk);
796 mptcp_for_each_subflow(msk, subflow)
797 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
801 /* In most cases we will be able to lock the mptcp socket. If its already
802 * owned, we need to defer to the work queue to avoid ABBA deadlock.
804 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
806 struct sock *sk = (struct sock *)msk;
807 unsigned int moved = 0;
809 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
810 __mptcp_ofo_queue(msk);
811 if (unlikely(ssk->sk_err)) {
812 if (!sock_owned_by_user(sk))
813 __mptcp_error_report(sk);
815 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
818 /* If the moves have caught up with the DATA_FIN sequence number
819 * it's time to ack the DATA_FIN and change socket state, but
820 * this is not a good place to change state. Let the workqueue
823 if (mptcp_pending_data_fin(sk, NULL))
824 mptcp_schedule_work(sk);
828 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
830 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
831 struct mptcp_sock *msk = mptcp_sk(sk);
832 int sk_rbuf, ssk_rbuf;
834 /* The peer can send data while we are shutting down this
835 * subflow at msk destruction time, but we must avoid enqueuing
836 * more data to the msk receive queue
838 if (unlikely(subflow->disposable))
841 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
842 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
843 if (unlikely(ssk_rbuf > sk_rbuf))
846 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
847 if (__mptcp_rmem(sk) > sk_rbuf) {
848 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
852 /* Wake-up the reader only for in-sequence data */
854 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
855 sk->sk_data_ready(sk);
856 mptcp_data_unlock(sk);
859 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
861 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
862 WRITE_ONCE(msk->allow_infinite_fallback, false);
863 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
866 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
868 struct sock *sk = (struct sock *)msk;
870 if (sk->sk_state != TCP_ESTABLISHED)
873 /* attach to msk socket only after we are sure we will deal with it
876 if (sk->sk_socket && !ssk->sk_socket)
877 mptcp_sock_graft(ssk, sk->sk_socket);
879 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
880 mptcp_sockopt_sync_locked(msk, ssk);
881 mptcp_subflow_joined(msk, ssk);
882 mptcp_stop_tout_timer(sk);
883 __mptcp_propagate_sndbuf(sk, ssk);
887 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
889 struct mptcp_subflow_context *tmp, *subflow;
890 struct mptcp_sock *msk = mptcp_sk(sk);
892 list_for_each_entry_safe(subflow, tmp, join_list, node) {
893 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
894 bool slow = lock_sock_fast(ssk);
896 list_move_tail(&subflow->node, &msk->conn_list);
897 if (!__mptcp_finish_join(msk, ssk))
898 mptcp_subflow_reset(ssk);
899 unlock_sock_fast(ssk, slow);
903 static bool mptcp_rtx_timer_pending(struct sock *sk)
905 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
908 static void mptcp_reset_rtx_timer(struct sock *sk)
910 struct inet_connection_sock *icsk = inet_csk(sk);
913 /* prevent rescheduling on close */
914 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
917 tout = mptcp_sk(sk)->timer_ival;
918 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
921 bool mptcp_schedule_work(struct sock *sk)
923 if (inet_sk_state_load(sk) != TCP_CLOSE &&
924 schedule_work(&mptcp_sk(sk)->work)) {
925 /* each subflow already holds a reference to the sk, and the
926 * workqueue is invoked by a subflow, so sk can't go away here.
934 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
936 struct mptcp_subflow_context *subflow;
938 msk_owned_by_me(msk);
940 mptcp_for_each_subflow(msk, subflow) {
941 if (READ_ONCE(subflow->data_avail))
942 return mptcp_subflow_tcp_sock(subflow);
948 static bool mptcp_skb_can_collapse_to(u64 write_seq,
949 const struct sk_buff *skb,
950 const struct mptcp_ext *mpext)
952 if (!tcp_skb_can_collapse_to(skb))
955 /* can collapse only if MPTCP level sequence is in order and this
956 * mapping has not been xmitted yet
958 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
962 /* we can append data to the given data frag if:
963 * - there is space available in the backing page_frag
964 * - the data frag tail matches the current page_frag free offset
965 * - the data frag end sequence number matches the current write seq
967 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
968 const struct page_frag *pfrag,
969 const struct mptcp_data_frag *df)
971 return df && pfrag->page == df->page &&
972 pfrag->size - pfrag->offset > 0 &&
973 pfrag->offset == (df->offset + df->data_len) &&
974 df->data_seq + df->data_len == msk->write_seq;
977 static void dfrag_uncharge(struct sock *sk, int len)
979 sk_mem_uncharge(sk, len);
980 sk_wmem_queued_add(sk, -len);
983 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
985 int len = dfrag->data_len + dfrag->overhead;
987 list_del(&dfrag->list);
988 dfrag_uncharge(sk, len);
989 put_page(dfrag->page);
992 /* called under both the msk socket lock and the data lock */
993 static void __mptcp_clean_una(struct sock *sk)
995 struct mptcp_sock *msk = mptcp_sk(sk);
996 struct mptcp_data_frag *dtmp, *dfrag;
999 snd_una = msk->snd_una;
1000 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1001 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1004 if (unlikely(dfrag == msk->first_pending)) {
1005 /* in recovery mode can see ack after the current snd head */
1006 if (WARN_ON_ONCE(!msk->recovery))
1009 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1012 dfrag_clear(sk, dfrag);
1015 dfrag = mptcp_rtx_head(sk);
1016 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1017 u64 delta = snd_una - dfrag->data_seq;
1019 /* prevent wrap around in recovery mode */
1020 if (unlikely(delta > dfrag->already_sent)) {
1021 if (WARN_ON_ONCE(!msk->recovery))
1023 if (WARN_ON_ONCE(delta > dfrag->data_len))
1025 dfrag->already_sent += delta - dfrag->already_sent;
1028 dfrag->data_seq += delta;
1029 dfrag->offset += delta;
1030 dfrag->data_len -= delta;
1031 dfrag->already_sent -= delta;
1033 dfrag_uncharge(sk, delta);
1036 /* all retransmitted data acked, recovery completed */
1037 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1038 msk->recovery = false;
1041 if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) {
1042 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1043 mptcp_stop_rtx_timer(sk);
1045 mptcp_reset_rtx_timer(sk);
1048 if (mptcp_pending_data_fin_ack(sk))
1049 mptcp_schedule_work(sk);
1052 static void __mptcp_clean_una_wakeup(struct sock *sk)
1054 lockdep_assert_held_once(&sk->sk_lock.slock);
1056 __mptcp_clean_una(sk);
1057 mptcp_write_space(sk);
1060 static void mptcp_clean_una_wakeup(struct sock *sk)
1062 mptcp_data_lock(sk);
1063 __mptcp_clean_una_wakeup(sk);
1064 mptcp_data_unlock(sk);
1067 static void mptcp_enter_memory_pressure(struct sock *sk)
1069 struct mptcp_subflow_context *subflow;
1070 struct mptcp_sock *msk = mptcp_sk(sk);
1073 mptcp_for_each_subflow(msk, subflow) {
1074 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1077 tcp_enter_memory_pressure(ssk);
1078 sk_stream_moderate_sndbuf(ssk);
1082 __mptcp_sync_sndbuf(sk);
1085 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1088 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1090 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1091 pfrag, sk->sk_allocation)))
1094 mptcp_enter_memory_pressure(sk);
1098 static struct mptcp_data_frag *
1099 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1102 int offset = ALIGN(orig_offset, sizeof(long));
1103 struct mptcp_data_frag *dfrag;
1105 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1106 dfrag->data_len = 0;
1107 dfrag->data_seq = msk->write_seq;
1108 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1109 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1110 dfrag->already_sent = 0;
1111 dfrag->page = pfrag->page;
1116 struct mptcp_sendmsg_info {
1122 bool data_lock_held;
1125 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1126 u64 data_seq, int avail_size)
1128 u64 window_end = mptcp_wnd_end(msk);
1131 if (__mptcp_check_fallback(msk))
1134 mptcp_snd_wnd = window_end - data_seq;
1135 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1137 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1138 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1139 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1145 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1147 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1151 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1155 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1157 struct sk_buff *skb;
1159 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1161 if (likely(__mptcp_add_ext(skb, gfp))) {
1162 skb_reserve(skb, MAX_TCP_HEADER);
1163 skb->ip_summed = CHECKSUM_PARTIAL;
1164 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1169 mptcp_enter_memory_pressure(sk);
1174 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1176 struct sk_buff *skb;
1178 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1182 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1183 tcp_skb_entail(ssk, skb);
1186 tcp_skb_tsorted_anchor_cleanup(skb);
1191 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1193 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1195 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1198 /* note: this always recompute the csum on the whole skb, even
1199 * if we just appended a single frag. More status info needed
1201 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1203 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1204 __wsum csum = ~csum_unfold(mpext->csum);
1205 int offset = skb->len - added;
1207 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1210 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1212 struct mptcp_ext *mpext)
1217 mpext->infinite_map = 1;
1218 mpext->data_len = 0;
1220 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1221 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1223 mptcp_do_fallback(ssk);
1226 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1228 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1229 struct mptcp_data_frag *dfrag,
1230 struct mptcp_sendmsg_info *info)
1232 u64 data_seq = dfrag->data_seq + info->sent;
1233 int offset = dfrag->offset + info->sent;
1234 struct mptcp_sock *msk = mptcp_sk(sk);
1235 bool zero_window_probe = false;
1236 struct mptcp_ext *mpext = NULL;
1237 bool can_coalesce = false;
1238 bool reuse_skb = true;
1239 struct sk_buff *skb;
1243 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1244 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1246 if (WARN_ON_ONCE(info->sent > info->limit ||
1247 info->limit > dfrag->data_len))
1250 if (unlikely(!__tcp_can_send(ssk)))
1253 /* compute send limit */
1254 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1255 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1256 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1257 copy = info->size_goal;
1259 skb = tcp_write_queue_tail(ssk);
1260 if (skb && copy > skb->len) {
1261 /* Limit the write to the size available in the
1262 * current skb, if any, so that we create at most a new skb.
1263 * Explicitly tells TCP internals to avoid collapsing on later
1264 * queue management operation, to avoid breaking the ext <->
1265 * SSN association set here
1267 mpext = mptcp_get_ext(skb);
1268 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1269 TCP_SKB_CB(skb)->eor = 1;
1270 tcp_mark_push(tcp_sk(ssk), skb);
1274 i = skb_shinfo(skb)->nr_frags;
1275 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1276 if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1277 tcp_mark_push(tcp_sk(ssk), skb);
1284 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1288 i = skb_shinfo(skb)->nr_frags;
1290 mpext = mptcp_get_ext(skb);
1293 /* Zero window and all data acked? Probe. */
1294 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1296 u64 snd_una = READ_ONCE(msk->snd_una);
1298 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1299 tcp_remove_empty_skb(ssk);
1303 zero_window_probe = true;
1304 data_seq = snd_una - 1;
1308 copy = min_t(size_t, copy, info->limit - info->sent);
1309 if (!sk_wmem_schedule(ssk, copy)) {
1310 tcp_remove_empty_skb(ssk);
1315 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1317 get_page(dfrag->page);
1318 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1322 skb->data_len += copy;
1323 skb->truesize += copy;
1324 sk_wmem_queued_add(ssk, copy);
1325 sk_mem_charge(ssk, copy);
1326 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1327 TCP_SKB_CB(skb)->end_seq += copy;
1328 tcp_skb_pcount_set(skb, 0);
1330 /* on skb reuse we just need to update the DSS len */
1332 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1333 mpext->data_len += copy;
1337 memset(mpext, 0, sizeof(*mpext));
1338 mpext->data_seq = data_seq;
1339 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1340 mpext->data_len = copy;
1344 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1345 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1348 if (zero_window_probe) {
1349 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1351 if (READ_ONCE(msk->csum_enabled))
1352 mptcp_update_data_checksum(skb, copy);
1353 tcp_push_pending_frames(ssk);
1357 if (READ_ONCE(msk->csum_enabled))
1358 mptcp_update_data_checksum(skb, copy);
1359 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1360 mptcp_update_infinite_map(msk, ssk, mpext);
1361 trace_mptcp_sendmsg_frag(mpext);
1362 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1366 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1367 sizeof(struct tcphdr) - \
1368 MAX_TCP_OPTION_SPACE - \
1369 sizeof(struct ipv6hdr) - \
1370 sizeof(struct frag_hdr))
1372 struct subflow_send_info {
1377 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1379 if (!subflow->stale)
1383 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1386 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1388 if (unlikely(subflow->stale)) {
1389 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1391 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1394 mptcp_subflow_set_active(subflow);
1396 return __mptcp_subflow_active(subflow);
1399 #define SSK_MODE_ACTIVE 0
1400 #define SSK_MODE_BACKUP 1
1401 #define SSK_MODE_MAX 2
1403 /* implement the mptcp packet scheduler;
1404 * returns the subflow that will transmit the next DSS
1405 * additionally updates the rtx timeout
1407 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1409 struct subflow_send_info send_info[SSK_MODE_MAX];
1410 struct mptcp_subflow_context *subflow;
1411 struct sock *sk = (struct sock *)msk;
1412 u32 pace, burst, wmem;
1413 int i, nr_active = 0;
1418 /* pick the subflow with the lower wmem/wspace ratio */
1419 for (i = 0; i < SSK_MODE_MAX; ++i) {
1420 send_info[i].ssk = NULL;
1421 send_info[i].linger_time = -1;
1424 mptcp_for_each_subflow(msk, subflow) {
1425 trace_mptcp_subflow_get_send(subflow);
1426 ssk = mptcp_subflow_tcp_sock(subflow);
1427 if (!mptcp_subflow_active(subflow))
1430 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1431 nr_active += !subflow->backup;
1432 pace = subflow->avg_pacing_rate;
1433 if (unlikely(!pace)) {
1434 /* init pacing rate from socket */
1435 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1436 pace = subflow->avg_pacing_rate;
1441 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1442 if (linger_time < send_info[subflow->backup].linger_time) {
1443 send_info[subflow->backup].ssk = ssk;
1444 send_info[subflow->backup].linger_time = linger_time;
1447 __mptcp_set_timeout(sk, tout);
1449 /* pick the best backup if no other subflow is active */
1451 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1453 /* According to the blest algorithm, to avoid HoL blocking for the
1454 * faster flow, we need to:
1455 * - estimate the faster flow linger time
1456 * - use the above to estimate the amount of byte transferred
1457 * by the faster flow
1458 * - check that the amount of queued data is greter than the above,
1459 * otherwise do not use the picked, slower, subflow
1460 * We select the subflow with the shorter estimated time to flush
1461 * the queued mem, which basically ensure the above. We just need
1462 * to check that subflow has a non empty cwin.
1464 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1465 if (!ssk || !sk_stream_memory_free(ssk))
1468 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1469 wmem = READ_ONCE(ssk->sk_wmem_queued);
1473 subflow = mptcp_subflow_ctx(ssk);
1474 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1475 READ_ONCE(ssk->sk_pacing_rate) * burst,
1477 msk->snd_burst = burst;
1481 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1483 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1487 static void mptcp_update_post_push(struct mptcp_sock *msk,
1488 struct mptcp_data_frag *dfrag,
1491 u64 snd_nxt_new = dfrag->data_seq;
1493 dfrag->already_sent += sent;
1495 msk->snd_burst -= sent;
1497 snd_nxt_new += dfrag->already_sent;
1499 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1500 * is recovering after a failover. In that event, this re-sends
1503 * Thus compute snd_nxt_new candidate based on
1504 * the dfrag->data_seq that was sent and the data
1505 * that has been handed to the subflow for transmission
1506 * and skip update in case it was old dfrag.
1508 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1509 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1510 WRITE_ONCE(msk->snd_nxt, snd_nxt_new);
1514 void mptcp_check_and_set_pending(struct sock *sk)
1516 if (mptcp_send_head(sk)) {
1517 mptcp_data_lock(sk);
1518 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1519 mptcp_data_unlock(sk);
1523 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1524 struct mptcp_sendmsg_info *info)
1526 struct mptcp_sock *msk = mptcp_sk(sk);
1527 struct mptcp_data_frag *dfrag;
1528 int len, copied = 0, err = 0;
1530 while ((dfrag = mptcp_send_head(sk))) {
1531 info->sent = dfrag->already_sent;
1532 info->limit = dfrag->data_len;
1533 len = dfrag->data_len - dfrag->already_sent;
1537 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1539 err = copied ? : ret;
1547 mptcp_update_post_push(msk, dfrag, ret);
1549 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1551 if (msk->snd_burst <= 0 ||
1552 !sk_stream_memory_free(ssk) ||
1553 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1557 mptcp_set_timeout(sk);
1563 msk->last_data_sent = tcp_jiffies32;
1567 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1569 struct sock *prev_ssk = NULL, *ssk = NULL;
1570 struct mptcp_sock *msk = mptcp_sk(sk);
1571 struct mptcp_sendmsg_info info = {
1574 bool do_check_data_fin = false;
1577 while (mptcp_send_head(sk) && (push_count > 0)) {
1578 struct mptcp_subflow_context *subflow;
1581 if (mptcp_sched_get_send(msk))
1586 mptcp_for_each_subflow(msk, subflow) {
1587 if (READ_ONCE(subflow->scheduled)) {
1588 mptcp_subflow_set_scheduled(subflow, false);
1591 ssk = mptcp_subflow_tcp_sock(subflow);
1592 if (ssk != prev_ssk) {
1593 /* First check. If the ssk has changed since
1594 * the last round, release prev_ssk
1597 mptcp_push_release(prev_ssk, &info);
1599 /* Need to lock the new subflow only if different
1600 * from the previous one, otherwise we are still
1601 * helding the relevant lock
1608 ret = __subflow_push_pending(sk, ssk, &info);
1610 if (ret != -EAGAIN ||
1611 (1 << ssk->sk_state) &
1612 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1616 do_check_data_fin = true;
1621 /* at this point we held the socket lock for the last subflow we used */
1623 mptcp_push_release(ssk, &info);
1625 /* ensure the rtx timer is running */
1626 if (!mptcp_rtx_timer_pending(sk))
1627 mptcp_reset_rtx_timer(sk);
1628 if (do_check_data_fin)
1629 mptcp_check_send_data_fin(sk);
1632 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1634 struct mptcp_sock *msk = mptcp_sk(sk);
1635 struct mptcp_sendmsg_info info = {
1636 .data_lock_held = true,
1638 bool keep_pushing = true;
1639 struct sock *xmit_ssk;
1643 while (mptcp_send_head(sk) && keep_pushing) {
1644 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1647 /* check for a different subflow usage only after
1648 * spooling the first chunk of data
1651 mptcp_subflow_set_scheduled(subflow, false);
1652 ret = __subflow_push_pending(sk, ssk, &info);
1660 if (mptcp_sched_get_send(msk))
1663 if (READ_ONCE(subflow->scheduled)) {
1664 mptcp_subflow_set_scheduled(subflow, false);
1665 ret = __subflow_push_pending(sk, ssk, &info);
1667 keep_pushing = false;
1671 mptcp_for_each_subflow(msk, subflow) {
1672 if (READ_ONCE(subflow->scheduled)) {
1673 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1674 if (xmit_ssk != ssk) {
1675 mptcp_subflow_delegate(subflow,
1676 MPTCP_DELEGATE_SEND);
1677 keep_pushing = false;
1684 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1685 * not going to flush it via release_sock()
1688 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1690 if (!mptcp_rtx_timer_pending(sk))
1691 mptcp_reset_rtx_timer(sk);
1693 if (msk->snd_data_fin_enable &&
1694 msk->snd_nxt + 1 == msk->write_seq)
1695 mptcp_schedule_work(sk);
1699 static int mptcp_disconnect(struct sock *sk, int flags);
1701 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1702 size_t len, int *copied_syn)
1704 unsigned int saved_flags = msg->msg_flags;
1705 struct mptcp_sock *msk = mptcp_sk(sk);
1709 /* on flags based fastopen the mptcp is supposed to create the
1710 * first subflow right now. Otherwise we are in the defer_connect
1711 * path, and the first subflow must be already present.
1712 * Since the defer_connect flag is cleared after the first succsful
1713 * fastopen attempt, no need to check for additional subflow status.
1715 if (msg->msg_flags & MSG_FASTOPEN) {
1716 ssk = __mptcp_nmpc_sk(msk);
1718 return PTR_ERR(ssk);
1726 msg->msg_flags |= MSG_DONTWAIT;
1727 msk->fastopening = 1;
1728 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1729 msk->fastopening = 0;
1730 msg->msg_flags = saved_flags;
1733 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1734 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1735 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1736 msg->msg_namelen, msg->msg_flags, 1);
1738 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1739 * case of any error, except timeout or signal
1741 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1743 } else if (ret && ret != -EINPROGRESS) {
1744 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1745 * __inet_stream_connect() can fail, due to looking check,
1746 * see mptcp_disconnect().
1747 * Attempt it again outside the problematic scope.
1749 if (!mptcp_disconnect(sk, 0))
1750 sk->sk_socket->state = SS_UNCONNECTED;
1752 inet_clear_bit(DEFER_CONNECT, sk);
1757 static int do_copy_data_nocache(struct sock *sk, int copy,
1758 struct iov_iter *from, char *to)
1760 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1761 if (!copy_from_iter_full_nocache(to, copy, from))
1763 } else if (!copy_from_iter_full(to, copy, from)) {
1769 /* open-code sk_stream_memory_free() plus sent limit computation to
1770 * avoid indirect calls in fast-path.
1771 * Called under the msk socket lock, so we can avoid a bunch of ONCE
1774 static u32 mptcp_send_limit(const struct sock *sk)
1776 const struct mptcp_sock *msk = mptcp_sk(sk);
1777 u32 limit, not_sent;
1779 if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
1782 limit = mptcp_notsent_lowat(sk);
1783 if (limit == UINT_MAX)
1786 not_sent = msk->write_seq - msk->snd_nxt;
1787 if (not_sent >= limit)
1790 return limit - not_sent;
1793 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1795 struct mptcp_sock *msk = mptcp_sk(sk);
1796 struct page_frag *pfrag;
1801 /* silently ignore everything else */
1802 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1806 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1807 msg->msg_flags & MSG_FASTOPEN)) {
1810 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1811 copied += copied_syn;
1812 if (ret == -EINPROGRESS && copied_syn > 0)
1818 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1820 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1821 ret = sk_stream_wait_connect(sk, &timeo);
1827 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1830 pfrag = sk_page_frag(sk);
1832 while (msg_data_left(msg)) {
1833 int total_ts, frag_truesize = 0;
1834 struct mptcp_data_frag *dfrag;
1835 bool dfrag_collapsed;
1836 size_t psize, offset;
1839 /* ensure fitting the notsent_lowat() constraint */
1840 copy_limit = mptcp_send_limit(sk);
1842 goto wait_for_memory;
1844 /* reuse tail pfrag, if possible, or carve a new one from the
1847 dfrag = mptcp_pending_tail(sk);
1848 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1849 if (!dfrag_collapsed) {
1850 if (!mptcp_page_frag_refill(sk, pfrag))
1851 goto wait_for_memory;
1853 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1854 frag_truesize = dfrag->overhead;
1857 /* we do not bound vs wspace, to allow a single packet.
1858 * memory accounting will prevent execessive memory usage
1861 offset = dfrag->offset + dfrag->data_len;
1862 psize = pfrag->size - offset;
1863 psize = min_t(size_t, psize, msg_data_left(msg));
1864 psize = min_t(size_t, psize, copy_limit);
1865 total_ts = psize + frag_truesize;
1867 if (!sk_wmem_schedule(sk, total_ts))
1868 goto wait_for_memory;
1870 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1871 page_address(dfrag->page) + offset);
1875 /* data successfully copied into the write queue */
1876 sk_forward_alloc_add(sk, -total_ts);
1878 dfrag->data_len += psize;
1879 frag_truesize += psize;
1880 pfrag->offset += frag_truesize;
1881 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1883 /* charge data on mptcp pending queue to the msk socket
1884 * Note: we charge such data both to sk and ssk
1886 sk_wmem_queued_add(sk, frag_truesize);
1887 if (!dfrag_collapsed) {
1888 get_page(dfrag->page);
1889 list_add_tail(&dfrag->list, &msk->rtx_queue);
1890 if (!msk->first_pending)
1891 WRITE_ONCE(msk->first_pending, dfrag);
1893 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1894 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1900 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1901 __mptcp_push_pending(sk, msg->msg_flags);
1902 ret = sk_stream_wait_memory(sk, &timeo);
1908 __mptcp_push_pending(sk, msg->msg_flags);
1918 copied = sk_stream_error(sk, msg->msg_flags, ret);
1922 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1924 size_t len, int flags,
1925 struct scm_timestamping_internal *tss,
1928 struct sk_buff *skb, *tmp;
1931 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1932 u32 offset = MPTCP_SKB_CB(skb)->offset;
1933 u32 data_len = skb->len - offset;
1934 u32 count = min_t(size_t, len - copied, data_len);
1937 if (!(flags & MSG_TRUNC)) {
1938 err = skb_copy_datagram_msg(skb, offset, msg, count);
1939 if (unlikely(err < 0)) {
1946 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1947 tcp_update_recv_tstamps(skb, tss);
1948 *cmsg_flags |= MPTCP_CMSG_TS;
1953 if (count < data_len) {
1954 if (!(flags & MSG_PEEK)) {
1955 MPTCP_SKB_CB(skb)->offset += count;
1956 MPTCP_SKB_CB(skb)->map_seq += count;
1957 msk->bytes_consumed += count;
1962 if (!(flags & MSG_PEEK)) {
1963 /* we will bulk release the skb memory later */
1964 skb->destructor = NULL;
1965 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1966 __skb_unlink(skb, &msk->receive_queue);
1968 msk->bytes_consumed += count;
1978 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1980 * Only difference: Use highest rtt estimate of the subflows in use.
1982 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1984 struct mptcp_subflow_context *subflow;
1985 struct sock *sk = (struct sock *)msk;
1986 u8 scaling_ratio = U8_MAX;
1987 u32 time, advmss = 1;
1990 msk_owned_by_me(msk);
1995 if (!msk->rcvspace_init)
1996 mptcp_rcv_space_init(msk, msk->first);
1998 msk->rcvq_space.copied += copied;
2000 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
2001 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
2003 rtt_us = msk->rcvq_space.rtt_us;
2004 if (rtt_us && time < (rtt_us >> 3))
2008 mptcp_for_each_subflow(msk, subflow) {
2009 const struct tcp_sock *tp;
2013 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
2015 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
2016 sf_advmss = READ_ONCE(tp->advmss);
2018 rtt_us = max(sf_rtt_us, rtt_us);
2019 advmss = max(sf_advmss, advmss);
2020 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
2023 msk->rcvq_space.rtt_us = rtt_us;
2024 msk->scaling_ratio = scaling_ratio;
2025 if (time < (rtt_us >> 3) || rtt_us == 0)
2028 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2031 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2032 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2036 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2038 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2040 do_div(grow, msk->rcvq_space.space);
2041 rcvwin += (grow << 1);
2043 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2044 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2046 if (rcvbuf > sk->sk_rcvbuf) {
2049 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2050 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2052 /* Make subflows follow along. If we do not do this, we
2053 * get drops at subflow level if skbs can't be moved to
2054 * the mptcp rx queue fast enough (announced rcv_win can
2055 * exceed ssk->sk_rcvbuf).
2057 mptcp_for_each_subflow(msk, subflow) {
2061 ssk = mptcp_subflow_tcp_sock(subflow);
2062 slow = lock_sock_fast(ssk);
2063 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2064 WRITE_ONCE(tcp_sk(ssk)->window_clamp, window_clamp);
2065 tcp_cleanup_rbuf(ssk, 1);
2066 unlock_sock_fast(ssk, slow);
2071 msk->rcvq_space.space = msk->rcvq_space.copied;
2073 msk->rcvq_space.copied = 0;
2074 msk->rcvq_space.time = mstamp;
2077 static void __mptcp_update_rmem(struct sock *sk)
2079 struct mptcp_sock *msk = mptcp_sk(sk);
2081 if (!msk->rmem_released)
2084 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2085 mptcp_rmem_uncharge(sk, msk->rmem_released);
2086 WRITE_ONCE(msk->rmem_released, 0);
2089 static void __mptcp_splice_receive_queue(struct sock *sk)
2091 struct mptcp_sock *msk = mptcp_sk(sk);
2093 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2096 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2098 struct sock *sk = (struct sock *)msk;
2099 unsigned int moved = 0;
2103 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2106 /* we can have data pending in the subflows only if the msk
2107 * receive buffer was full at subflow_data_ready() time,
2108 * that is an unlikely slow path.
2113 slowpath = lock_sock_fast(ssk);
2114 mptcp_data_lock(sk);
2115 __mptcp_update_rmem(sk);
2116 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2117 mptcp_data_unlock(sk);
2119 if (unlikely(ssk->sk_err))
2120 __mptcp_error_report(sk);
2121 unlock_sock_fast(ssk, slowpath);
2124 /* acquire the data lock only if some input data is pending */
2126 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2127 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2128 mptcp_data_lock(sk);
2129 __mptcp_update_rmem(sk);
2130 ret |= __mptcp_ofo_queue(msk);
2131 __mptcp_splice_receive_queue(sk);
2132 mptcp_data_unlock(sk);
2135 mptcp_check_data_fin((struct sock *)msk);
2136 return !skb_queue_empty(&msk->receive_queue);
2139 static unsigned int mptcp_inq_hint(const struct sock *sk)
2141 const struct mptcp_sock *msk = mptcp_sk(sk);
2142 const struct sk_buff *skb;
2144 skb = skb_peek(&msk->receive_queue);
2146 u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq;
2148 if (hint_val >= INT_MAX)
2151 return (unsigned int)hint_val;
2154 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2160 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2161 int flags, int *addr_len)
2163 struct mptcp_sock *msk = mptcp_sk(sk);
2164 struct scm_timestamping_internal tss;
2165 int copied = 0, cmsg_flags = 0;
2169 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2170 if (unlikely(flags & MSG_ERRQUEUE))
2171 return inet_recv_error(sk, msg, len, addr_len);
2174 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2179 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2181 len = min_t(size_t, len, INT_MAX);
2182 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2184 if (unlikely(msk->recvmsg_inq))
2185 cmsg_flags = MPTCP_CMSG_INQ;
2187 while (copied < len) {
2190 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2191 if (unlikely(bytes_read < 0)) {
2193 copied = bytes_read;
2197 copied += bytes_read;
2199 /* be sure to advertise window change */
2200 mptcp_cleanup_rbuf(msk);
2202 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2205 /* only the master socket status is relevant here. The exit
2206 * conditions mirror closely tcp_recvmsg()
2208 if (copied >= target)
2213 sk->sk_state == TCP_CLOSE ||
2214 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2216 signal_pending(current))
2220 copied = sock_error(sk);
2224 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2225 /* race breaker: the shutdown could be after the
2226 * previous receive queue check
2228 if (__mptcp_move_skbs(msk))
2233 if (sk->sk_state == TCP_CLOSE) {
2243 if (signal_pending(current)) {
2244 copied = sock_intr_errno(timeo);
2249 pr_debug("block timeout %ld", timeo);
2250 sk_wait_data(sk, &timeo, NULL);
2254 if (cmsg_flags && copied >= 0) {
2255 if (cmsg_flags & MPTCP_CMSG_TS)
2256 tcp_recv_timestamp(msg, sk, &tss);
2258 if (cmsg_flags & MPTCP_CMSG_INQ) {
2259 unsigned int inq = mptcp_inq_hint(sk);
2261 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2265 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2266 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2267 skb_queue_empty(&msk->receive_queue), copied);
2268 if (!(flags & MSG_PEEK))
2269 mptcp_rcv_space_adjust(msk, copied);
2275 static void mptcp_retransmit_timer(struct timer_list *t)
2277 struct inet_connection_sock *icsk = from_timer(icsk, t,
2278 icsk_retransmit_timer);
2279 struct sock *sk = &icsk->icsk_inet.sk;
2280 struct mptcp_sock *msk = mptcp_sk(sk);
2283 if (!sock_owned_by_user(sk)) {
2284 /* we need a process context to retransmit */
2285 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2286 mptcp_schedule_work(sk);
2288 /* delegate our work to tcp_release_cb() */
2289 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2295 static void mptcp_tout_timer(struct timer_list *t)
2297 struct sock *sk = from_timer(sk, t, sk_timer);
2299 mptcp_schedule_work(sk);
2303 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2306 * A backup subflow is returned only if that is the only kind available.
2308 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2310 struct sock *backup = NULL, *pick = NULL;
2311 struct mptcp_subflow_context *subflow;
2312 int min_stale_count = INT_MAX;
2314 mptcp_for_each_subflow(msk, subflow) {
2315 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2317 if (!__mptcp_subflow_active(subflow))
2320 /* still data outstanding at TCP level? skip this */
2321 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2322 mptcp_pm_subflow_chk_stale(msk, ssk);
2323 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2327 if (subflow->backup) {
2340 /* use backup only if there are no progresses anywhere */
2341 return min_stale_count > 1 ? backup : NULL;
2344 bool __mptcp_retransmit_pending_data(struct sock *sk)
2346 struct mptcp_data_frag *cur, *rtx_head;
2347 struct mptcp_sock *msk = mptcp_sk(sk);
2349 if (__mptcp_check_fallback(msk))
2352 /* the closing socket has some data untransmitted and/or unacked:
2353 * some data in the mptcp rtx queue has not really xmitted yet.
2354 * keep it simple and re-inject the whole mptcp level rtx queue
2356 mptcp_data_lock(sk);
2357 __mptcp_clean_una_wakeup(sk);
2358 rtx_head = mptcp_rtx_head(sk);
2360 mptcp_data_unlock(sk);
2364 msk->recovery_snd_nxt = msk->snd_nxt;
2365 msk->recovery = true;
2366 mptcp_data_unlock(sk);
2368 msk->first_pending = rtx_head;
2371 /* be sure to clear the "sent status" on all re-injected fragments */
2372 list_for_each_entry(cur, &msk->rtx_queue, list) {
2373 if (!cur->already_sent)
2375 cur->already_sent = 0;
2381 /* flags for __mptcp_close_ssk() */
2382 #define MPTCP_CF_PUSH BIT(1)
2383 #define MPTCP_CF_FASTCLOSE BIT(2)
2385 /* be sure to send a reset only if the caller asked for it, also
2386 * clean completely the subflow status when the subflow reaches
2389 static void __mptcp_subflow_disconnect(struct sock *ssk,
2390 struct mptcp_subflow_context *subflow,
2393 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2394 (flags & MPTCP_CF_FASTCLOSE)) {
2395 /* The MPTCP code never wait on the subflow sockets, TCP-level
2396 * disconnect should never fail
2398 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2399 mptcp_subflow_ctx_reset(subflow);
2401 tcp_shutdown(ssk, SEND_SHUTDOWN);
2405 /* subflow sockets can be either outgoing (connect) or incoming
2408 * Outgoing subflows use in-kernel sockets.
2409 * Incoming subflows do not have their own 'struct socket' allocated,
2410 * so we need to use tcp_close() after detaching them from the mptcp
2413 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2414 struct mptcp_subflow_context *subflow,
2417 struct mptcp_sock *msk = mptcp_sk(sk);
2418 bool dispose_it, need_push = false;
2420 /* If the first subflow moved to a close state before accept, e.g. due
2421 * to an incoming reset or listener shutdown, the subflow socket is
2422 * already deleted by inet_child_forget() and the mptcp socket can't
2425 if (msk->in_accept_queue && msk->first == ssk &&
2426 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2427 /* ensure later check in mptcp_worker() will dispose the msk */
2428 sock_set_flag(sk, SOCK_DEAD);
2429 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2430 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2431 mptcp_subflow_drop_ctx(ssk);
2435 dispose_it = msk->free_first || ssk != msk->first;
2437 list_del(&subflow->node);
2439 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2441 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2442 /* be sure to force the tcp_close path
2443 * to generate the egress reset
2445 ssk->sk_lingertime = 0;
2446 sock_set_flag(ssk, SOCK_LINGER);
2447 subflow->send_fastclose = 1;
2450 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2452 __mptcp_subflow_disconnect(ssk, subflow, flags);
2458 subflow->disposable = 1;
2460 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2461 * the ssk has been already destroyed, we just need to release the
2462 * reference owned by msk;
2464 if (!inet_csk(ssk)->icsk_ulp_ops) {
2465 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2466 kfree_rcu(subflow, rcu);
2468 /* otherwise tcp will dispose of the ssk and subflow ctx */
2469 __tcp_close(ssk, 0);
2471 /* close acquired an extra ref */
2476 __mptcp_subflow_error_report(sk, ssk);
2481 if (ssk == msk->first)
2482 WRITE_ONCE(msk->first, NULL);
2485 __mptcp_sync_sndbuf(sk);
2487 __mptcp_push_pending(sk, 0);
2489 /* Catch every 'all subflows closed' scenario, including peers silently
2490 * closing them, e.g. due to timeout.
2491 * For established sockets, allow an additional timeout before closing,
2492 * as the protocol can still create more subflows.
2494 if (list_is_singular(&msk->conn_list) && msk->first &&
2495 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2496 if (sk->sk_state != TCP_ESTABLISHED ||
2497 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2498 mptcp_set_state(sk, TCP_CLOSE);
2499 mptcp_close_wake_up(sk);
2501 mptcp_start_tout_timer(sk);
2506 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2507 struct mptcp_subflow_context *subflow)
2509 if (sk->sk_state == TCP_ESTABLISHED)
2510 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2512 /* subflow aborted before reaching the fully_established status
2513 * attempt the creation of the next subflow
2515 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2517 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2520 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2525 static void __mptcp_close_subflow(struct sock *sk)
2527 struct mptcp_subflow_context *subflow, *tmp;
2528 struct mptcp_sock *msk = mptcp_sk(sk);
2532 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2533 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2535 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2538 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2539 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2542 mptcp_close_ssk(sk, ssk, subflow);
2547 static bool mptcp_close_tout_expired(const struct sock *sk)
2549 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2550 sk->sk_state == TCP_CLOSE)
2553 return time_after32(tcp_jiffies32,
2554 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2557 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2559 struct mptcp_subflow_context *subflow, *tmp;
2560 struct sock *sk = (struct sock *)msk;
2562 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2565 mptcp_token_destroy(msk);
2567 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2568 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2571 slow = lock_sock_fast(tcp_sk);
2572 if (tcp_sk->sk_state != TCP_CLOSE) {
2573 mptcp_send_active_reset_reason(tcp_sk);
2574 tcp_set_state(tcp_sk, TCP_CLOSE);
2576 unlock_sock_fast(tcp_sk, slow);
2579 /* Mirror the tcp_reset() error propagation */
2580 switch (sk->sk_state) {
2582 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2584 case TCP_CLOSE_WAIT:
2585 WRITE_ONCE(sk->sk_err, EPIPE);
2590 WRITE_ONCE(sk->sk_err, ECONNRESET);
2593 mptcp_set_state(sk, TCP_CLOSE);
2594 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2595 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2596 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2598 /* the calling mptcp_worker will properly destroy the socket */
2599 if (sock_flag(sk, SOCK_DEAD))
2602 sk->sk_state_change(sk);
2603 sk_error_report(sk);
2606 static void __mptcp_retrans(struct sock *sk)
2608 struct mptcp_sock *msk = mptcp_sk(sk);
2609 struct mptcp_subflow_context *subflow;
2610 struct mptcp_sendmsg_info info = {};
2611 struct mptcp_data_frag *dfrag;
2616 mptcp_clean_una_wakeup(sk);
2618 /* first check ssk: need to kick "stale" logic */
2619 err = mptcp_sched_get_retrans(msk);
2620 dfrag = mptcp_rtx_head(sk);
2622 if (mptcp_data_fin_enabled(msk)) {
2623 struct inet_connection_sock *icsk = inet_csk(sk);
2625 icsk->icsk_retransmits++;
2626 mptcp_set_datafin_timeout(sk);
2627 mptcp_send_ack(msk);
2632 if (!mptcp_send_head(sk))
2641 mptcp_for_each_subflow(msk, subflow) {
2642 if (READ_ONCE(subflow->scheduled)) {
2645 mptcp_subflow_set_scheduled(subflow, false);
2647 ssk = mptcp_subflow_tcp_sock(subflow);
2651 /* limit retransmission to the bytes already sent on some subflows */
2653 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2654 dfrag->already_sent;
2655 while (info.sent < info.limit) {
2656 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2660 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2665 len = max(copied, len);
2666 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2668 WRITE_ONCE(msk->allow_infinite_fallback, false);
2675 msk->bytes_retrans += len;
2676 dfrag->already_sent = max(dfrag->already_sent, len);
2679 mptcp_check_and_set_pending(sk);
2681 if (!mptcp_rtx_timer_pending(sk))
2682 mptcp_reset_rtx_timer(sk);
2685 /* schedule the timeout timer for the relevant event: either close timeout
2686 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2688 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2690 struct sock *sk = (struct sock *)msk;
2691 unsigned long timeout, close_timeout;
2693 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2696 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2697 mptcp_close_timeout(sk);
2699 /* the close timeout takes precedence on the fail one, and here at least one of
2702 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2704 sk_reset_timer(sk, &sk->sk_timer, timeout);
2707 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2709 struct sock *ssk = msk->first;
2715 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2717 slow = lock_sock_fast(ssk);
2718 mptcp_subflow_reset(ssk);
2719 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2720 unlock_sock_fast(ssk, slow);
2723 static void mptcp_do_fastclose(struct sock *sk)
2725 struct mptcp_subflow_context *subflow, *tmp;
2726 struct mptcp_sock *msk = mptcp_sk(sk);
2728 mptcp_set_state(sk, TCP_CLOSE);
2729 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2730 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2731 subflow, MPTCP_CF_FASTCLOSE);
2734 static void mptcp_worker(struct work_struct *work)
2736 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2737 struct sock *sk = (struct sock *)msk;
2738 unsigned long fail_tout;
2742 state = sk->sk_state;
2743 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2746 mptcp_check_fastclose(msk);
2748 mptcp_pm_nl_work(msk);
2750 mptcp_check_send_data_fin(sk);
2751 mptcp_check_data_fin_ack(sk);
2752 mptcp_check_data_fin(sk);
2754 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2755 __mptcp_close_subflow(sk);
2757 if (mptcp_close_tout_expired(sk)) {
2758 mptcp_do_fastclose(sk);
2759 mptcp_close_wake_up(sk);
2762 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2763 __mptcp_destroy_sock(sk);
2767 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2768 __mptcp_retrans(sk);
2770 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2771 if (fail_tout && time_after(jiffies, fail_tout))
2772 mptcp_mp_fail_no_response(msk);
2779 static void __mptcp_init_sock(struct sock *sk)
2781 struct mptcp_sock *msk = mptcp_sk(sk);
2783 INIT_LIST_HEAD(&msk->conn_list);
2784 INIT_LIST_HEAD(&msk->join_list);
2785 INIT_LIST_HEAD(&msk->rtx_queue);
2786 INIT_WORK(&msk->work, mptcp_worker);
2787 __skb_queue_head_init(&msk->receive_queue);
2788 msk->out_of_order_queue = RB_ROOT;
2789 msk->first_pending = NULL;
2790 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
2791 WRITE_ONCE(msk->rmem_released, 0);
2792 msk->timer_ival = TCP_RTO_MIN;
2793 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2795 WRITE_ONCE(msk->first, NULL);
2796 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2797 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2798 WRITE_ONCE(msk->allow_infinite_fallback, true);
2799 msk->recovery = false;
2800 msk->subflow_id = 1;
2801 msk->last_data_sent = tcp_jiffies32;
2802 msk->last_data_recv = tcp_jiffies32;
2803 msk->last_ack_recv = tcp_jiffies32;
2805 mptcp_pm_data_init(msk);
2807 /* re-use the csk retrans timer for MPTCP-level retrans */
2808 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2809 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2812 static void mptcp_ca_reset(struct sock *sk)
2814 struct inet_connection_sock *icsk = inet_csk(sk);
2816 tcp_assign_congestion_control(sk);
2817 strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name,
2818 sizeof(mptcp_sk(sk)->ca_name));
2820 /* no need to keep a reference to the ops, the name will suffice */
2821 tcp_cleanup_congestion_control(sk);
2822 icsk->icsk_ca_ops = NULL;
2825 static int mptcp_init_sock(struct sock *sk)
2827 struct net *net = sock_net(sk);
2830 __mptcp_init_sock(sk);
2832 if (!mptcp_is_enabled(net))
2833 return -ENOPROTOOPT;
2835 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2838 ret = mptcp_init_sched(mptcp_sk(sk),
2839 mptcp_sched_find(mptcp_get_scheduler(net)));
2843 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2845 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2846 * propagate the correct value
2850 sk_sockets_allocated_inc(sk);
2851 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2852 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2857 static void __mptcp_clear_xmit(struct sock *sk)
2859 struct mptcp_sock *msk = mptcp_sk(sk);
2860 struct mptcp_data_frag *dtmp, *dfrag;
2862 WRITE_ONCE(msk->first_pending, NULL);
2863 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2864 dfrag_clear(sk, dfrag);
2867 void mptcp_cancel_work(struct sock *sk)
2869 struct mptcp_sock *msk = mptcp_sk(sk);
2871 if (cancel_work_sync(&msk->work))
2875 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2879 switch (ssk->sk_state) {
2881 if (!(how & RCV_SHUTDOWN))
2885 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2888 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2889 pr_debug("Fallback");
2890 ssk->sk_shutdown |= how;
2891 tcp_shutdown(ssk, how);
2893 /* simulate the data_fin ack reception to let the state
2894 * machine move forward
2896 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2897 mptcp_schedule_work(sk);
2899 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2901 if (!mptcp_rtx_timer_pending(sk))
2902 mptcp_reset_rtx_timer(sk);
2910 void mptcp_set_state(struct sock *sk, int state)
2912 int oldstate = sk->sk_state;
2915 case TCP_ESTABLISHED:
2916 if (oldstate != TCP_ESTABLISHED)
2917 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2921 if (oldstate == TCP_ESTABLISHED)
2922 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2925 inet_sk_state_store(sk, state);
2928 static const unsigned char new_state[16] = {
2929 /* current state: new state: action: */
2930 [0 /* (Invalid) */] = TCP_CLOSE,
2931 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2932 [TCP_SYN_SENT] = TCP_CLOSE,
2933 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2934 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2935 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2936 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2937 [TCP_CLOSE] = TCP_CLOSE,
2938 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2939 [TCP_LAST_ACK] = TCP_LAST_ACK,
2940 [TCP_LISTEN] = TCP_CLOSE,
2941 [TCP_CLOSING] = TCP_CLOSING,
2942 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2945 static int mptcp_close_state(struct sock *sk)
2947 int next = (int)new_state[sk->sk_state];
2948 int ns = next & TCP_STATE_MASK;
2950 mptcp_set_state(sk, ns);
2952 return next & TCP_ACTION_FIN;
2955 static void mptcp_check_send_data_fin(struct sock *sk)
2957 struct mptcp_subflow_context *subflow;
2958 struct mptcp_sock *msk = mptcp_sk(sk);
2960 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2961 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2962 msk->snd_nxt, msk->write_seq);
2964 /* we still need to enqueue subflows or not really shutting down,
2967 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2968 mptcp_send_head(sk))
2971 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2973 mptcp_for_each_subflow(msk, subflow) {
2974 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2976 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2980 static void __mptcp_wr_shutdown(struct sock *sk)
2982 struct mptcp_sock *msk = mptcp_sk(sk);
2984 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2985 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2986 !!mptcp_send_head(sk));
2988 /* will be ignored by fallback sockets */
2989 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2990 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2992 mptcp_check_send_data_fin(sk);
2995 static void __mptcp_destroy_sock(struct sock *sk)
2997 struct mptcp_sock *msk = mptcp_sk(sk);
2999 pr_debug("msk=%p", msk);
3003 mptcp_stop_rtx_timer(sk);
3004 sk_stop_timer(sk, &sk->sk_timer);
3006 mptcp_release_sched(msk);
3008 sk->sk_prot->destroy(sk);
3010 WARN_ON_ONCE(READ_ONCE(msk->rmem_fwd_alloc));
3011 WARN_ON_ONCE(msk->rmem_released);
3012 sk_stream_kill_queues(sk);
3013 xfrm_sk_free_policy(sk);
3018 void __mptcp_unaccepted_force_close(struct sock *sk)
3020 sock_set_flag(sk, SOCK_DEAD);
3021 mptcp_do_fastclose(sk);
3022 __mptcp_destroy_sock(sk);
3025 static __poll_t mptcp_check_readable(struct sock *sk)
3027 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
3030 static void mptcp_check_listen_stop(struct sock *sk)
3034 if (inet_sk_state_load(sk) != TCP_LISTEN)
3037 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3038 ssk = mptcp_sk(sk)->first;
3039 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3042 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3043 tcp_set_state(ssk, TCP_CLOSE);
3044 mptcp_subflow_queue_clean(sk, ssk);
3045 inet_csk_listen_stop(ssk);
3046 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3050 bool __mptcp_close(struct sock *sk, long timeout)
3052 struct mptcp_subflow_context *subflow;
3053 struct mptcp_sock *msk = mptcp_sk(sk);
3054 bool do_cancel_work = false;
3055 int subflows_alive = 0;
3057 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3059 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3060 mptcp_check_listen_stop(sk);
3061 mptcp_set_state(sk, TCP_CLOSE);
3065 if (mptcp_data_avail(msk) || timeout < 0) {
3066 /* If the msk has read data, or the caller explicitly ask it,
3067 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3069 mptcp_do_fastclose(sk);
3071 } else if (mptcp_close_state(sk)) {
3072 __mptcp_wr_shutdown(sk);
3075 sk_stream_wait_close(sk, timeout);
3078 /* orphan all the subflows */
3079 mptcp_for_each_subflow(msk, subflow) {
3080 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3081 bool slow = lock_sock_fast_nested(ssk);
3083 subflows_alive += ssk->sk_state != TCP_CLOSE;
3085 /* since the close timeout takes precedence on the fail one,
3088 if (ssk == msk->first)
3089 subflow->fail_tout = 0;
3091 /* detach from the parent socket, but allow data_ready to
3092 * push incoming data into the mptcp stack, to properly ack it
3094 ssk->sk_socket = NULL;
3096 unlock_sock_fast(ssk, slow);
3100 /* all the subflows are closed, only timeout can change the msk
3101 * state, let's not keep resources busy for no reasons
3103 if (subflows_alive == 0)
3104 mptcp_set_state(sk, TCP_CLOSE);
3107 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3109 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3111 if (sk->sk_state == TCP_CLOSE) {
3112 __mptcp_destroy_sock(sk);
3113 do_cancel_work = true;
3115 mptcp_start_tout_timer(sk);
3118 return do_cancel_work;
3121 static void mptcp_close(struct sock *sk, long timeout)
3123 bool do_cancel_work;
3127 do_cancel_work = __mptcp_close(sk, timeout);
3130 mptcp_cancel_work(sk);
3135 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3137 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3138 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3139 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3141 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3142 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3145 msk6->saddr = ssk6->saddr;
3146 msk6->flow_label = ssk6->flow_label;
3150 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3151 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3152 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3153 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3154 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3155 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3158 static int mptcp_disconnect(struct sock *sk, int flags)
3160 struct mptcp_sock *msk = mptcp_sk(sk);
3162 /* We are on the fastopen error path. We can't call straight into the
3163 * subflows cleanup code due to lock nesting (we are already under
3164 * msk->firstsocket lock).
3166 if (msk->fastopening)
3169 mptcp_check_listen_stop(sk);
3170 mptcp_set_state(sk, TCP_CLOSE);
3172 mptcp_stop_rtx_timer(sk);
3173 mptcp_stop_tout_timer(sk);
3176 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3178 /* msk->subflow is still intact, the following will not free the first
3181 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3182 WRITE_ONCE(msk->flags, 0);
3184 msk->recovery = false;
3185 WRITE_ONCE(msk->can_ack, false);
3186 WRITE_ONCE(msk->fully_established, false);
3187 WRITE_ONCE(msk->rcv_data_fin, false);
3188 WRITE_ONCE(msk->snd_data_fin_enable, false);
3189 WRITE_ONCE(msk->rcv_fastclose, false);
3190 WRITE_ONCE(msk->use_64bit_ack, false);
3191 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3192 mptcp_pm_data_reset(msk);
3194 msk->bytes_consumed = 0;
3195 msk->bytes_acked = 0;
3196 msk->bytes_received = 0;
3197 msk->bytes_sent = 0;
3198 msk->bytes_retrans = 0;
3199 msk->rcvspace_init = 0;
3201 WRITE_ONCE(sk->sk_shutdown, 0);
3202 sk_error_report(sk);
3206 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3207 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3209 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3211 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3214 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3216 const struct ipv6_pinfo *np = inet6_sk(sk);
3217 struct ipv6_txoptions *opt;
3218 struct ipv6_pinfo *newnp;
3220 newnp = inet6_sk(newsk);
3223 opt = rcu_dereference(np->opt);
3225 opt = ipv6_dup_options(newsk, opt);
3227 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3229 RCU_INIT_POINTER(newnp->opt, opt);
3234 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3236 struct ip_options_rcu *inet_opt, *newopt = NULL;
3237 const struct inet_sock *inet = inet_sk(sk);
3238 struct inet_sock *newinet;
3240 newinet = inet_sk(newsk);
3243 inet_opt = rcu_dereference(inet->inet_opt);
3245 newopt = sock_kmalloc(newsk, sizeof(*inet_opt) +
3246 inet_opt->opt.optlen, GFP_ATOMIC);
3248 memcpy(newopt, inet_opt, sizeof(*inet_opt) +
3249 inet_opt->opt.optlen);
3251 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3253 RCU_INIT_POINTER(newinet->inet_opt, newopt);
3257 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3258 const struct mptcp_options_received *mp_opt,
3260 struct request_sock *req)
3262 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3263 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3264 struct mptcp_subflow_context *subflow;
3265 struct mptcp_sock *msk;
3270 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3271 if (nsk->sk_family == AF_INET6)
3272 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3275 __mptcp_init_sock(nsk);
3277 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3278 if (nsk->sk_family == AF_INET6)
3279 mptcp_copy_ip6_options(nsk, sk);
3282 mptcp_copy_ip_options(nsk, sk);
3284 msk = mptcp_sk(nsk);
3285 WRITE_ONCE(msk->local_key, subflow_req->local_key);
3286 WRITE_ONCE(msk->token, subflow_req->token);
3287 msk->in_accept_queue = 1;
3288 WRITE_ONCE(msk->fully_established, false);
3289 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3290 WRITE_ONCE(msk->csum_enabled, true);
3292 WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1);
3293 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3294 WRITE_ONCE(msk->snd_una, msk->write_seq);
3295 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3296 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3297 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3299 /* passive msk is created after the first/MPC subflow */
3300 msk->subflow_id = 2;
3302 sock_reset_flag(nsk, SOCK_RCU_FREE);
3303 security_inet_csk_clone(nsk, req);
3305 /* this can't race with mptcp_close(), as the msk is
3306 * not yet exposted to user-space
3308 mptcp_set_state(nsk, TCP_ESTABLISHED);
3310 /* The msk maintain a ref to each subflow in the connections list */
3311 WRITE_ONCE(msk->first, ssk);
3312 subflow = mptcp_subflow_ctx(ssk);
3313 list_add(&subflow->node, &msk->conn_list);
3316 /* new mpc subflow takes ownership of the newly
3317 * created mptcp socket
3319 mptcp_token_accept(subflow_req, msk);
3321 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3322 * uses the correct data
3324 mptcp_copy_inaddrs(nsk, ssk);
3325 __mptcp_propagate_sndbuf(nsk, ssk);
3327 mptcp_rcv_space_init(msk, ssk);
3329 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3330 __mptcp_subflow_fully_established(msk, subflow, mp_opt);
3331 bh_unlock_sock(nsk);
3333 /* note: the newly allocated socket refcount is 2 now */
3337 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3339 const struct tcp_sock *tp = tcp_sk(ssk);
3341 msk->rcvspace_init = 1;
3342 msk->rcvq_space.copied = 0;
3343 msk->rcvq_space.rtt_us = 0;
3345 msk->rcvq_space.time = tp->tcp_mstamp;
3347 /* initial rcv_space offering made to peer */
3348 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3349 TCP_INIT_CWND * tp->advmss);
3350 if (msk->rcvq_space.space == 0)
3351 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3354 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3356 struct mptcp_subflow_context *subflow, *tmp;
3357 struct sock *sk = (struct sock *)msk;
3359 __mptcp_clear_xmit(sk);
3361 /* join list will be eventually flushed (with rst) at sock lock release time */
3362 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3363 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3365 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3366 mptcp_data_lock(sk);
3367 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3368 __skb_queue_purge(&sk->sk_receive_queue);
3369 skb_rbtree_purge(&msk->out_of_order_queue);
3370 mptcp_data_unlock(sk);
3372 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3373 * inet_sock_destruct() will dispose it
3375 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3376 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3377 mptcp_token_destroy(msk);
3378 mptcp_pm_free_anno_list(msk);
3379 mptcp_free_local_addr_list(msk);
3382 static void mptcp_destroy(struct sock *sk)
3384 struct mptcp_sock *msk = mptcp_sk(sk);
3386 /* allow the following to close even the initial subflow */
3387 msk->free_first = 1;
3388 mptcp_destroy_common(msk, 0);
3389 sk_sockets_allocated_dec(sk);
3392 void __mptcp_data_acked(struct sock *sk)
3394 if (!sock_owned_by_user(sk))
3395 __mptcp_clean_una(sk);
3397 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3400 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3402 if (!mptcp_send_head(sk))
3405 if (!sock_owned_by_user(sk))
3406 __mptcp_subflow_push_pending(sk, ssk, false);
3408 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3411 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3412 BIT(MPTCP_RETRANSMIT) | \
3413 BIT(MPTCP_FLUSH_JOIN_LIST))
3415 /* processes deferred events and flush wmem */
3416 static void mptcp_release_cb(struct sock *sk)
3417 __must_hold(&sk->sk_lock.slock)
3419 struct mptcp_sock *msk = mptcp_sk(sk);
3422 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3423 struct list_head join_list;
3428 INIT_LIST_HEAD(&join_list);
3429 list_splice_init(&msk->join_list, &join_list);
3431 /* the following actions acquire the subflow socket lock
3433 * 1) can't be invoked in atomic scope
3434 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3435 * datapath acquires the msk socket spinlock while helding
3436 * the subflow socket lock
3438 msk->cb_flags &= ~flags;
3439 spin_unlock_bh(&sk->sk_lock.slock);
3441 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3442 __mptcp_flush_join_list(sk, &join_list);
3443 if (flags & BIT(MPTCP_PUSH_PENDING))
3444 __mptcp_push_pending(sk, 0);
3445 if (flags & BIT(MPTCP_RETRANSMIT))
3446 __mptcp_retrans(sk);
3449 spin_lock_bh(&sk->sk_lock.slock);
3452 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3453 __mptcp_clean_una_wakeup(sk);
3454 if (unlikely(msk->cb_flags)) {
3455 /* be sure to sync the msk state before taking actions
3456 * depending on sk_state (MPTCP_ERROR_REPORT)
3457 * On sk release avoid actions depending on the first subflow
3459 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3460 __mptcp_sync_state(sk, msk->pending_state);
3461 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3462 __mptcp_error_report(sk);
3463 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3464 __mptcp_sync_sndbuf(sk);
3467 __mptcp_update_rmem(sk);
3470 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3471 * TCP can't schedule delack timer before the subflow is fully established.
3472 * MPTCP uses the delack timer to do 3rd ack retransmissions
3474 static void schedule_3rdack_retransmission(struct sock *ssk)
3476 struct inet_connection_sock *icsk = inet_csk(ssk);
3477 struct tcp_sock *tp = tcp_sk(ssk);
3478 unsigned long timeout;
3480 if (mptcp_subflow_ctx(ssk)->fully_established)
3483 /* reschedule with a timeout above RTT, as we must look only for drop */
3485 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3487 timeout = TCP_TIMEOUT_INIT;
3490 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3491 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3492 icsk->icsk_ack.timeout = timeout;
3493 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3496 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3498 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3499 struct sock *sk = subflow->conn;
3501 if (status & BIT(MPTCP_DELEGATE_SEND)) {
3502 mptcp_data_lock(sk);
3503 if (!sock_owned_by_user(sk))
3504 __mptcp_subflow_push_pending(sk, ssk, true);
3506 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3507 mptcp_data_unlock(sk);
3509 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3510 mptcp_data_lock(sk);
3511 if (!sock_owned_by_user(sk))
3512 __mptcp_sync_sndbuf(sk);
3514 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3515 mptcp_data_unlock(sk);
3517 if (status & BIT(MPTCP_DELEGATE_ACK))
3518 schedule_3rdack_retransmission(ssk);
3521 static int mptcp_hash(struct sock *sk)
3523 /* should never be called,
3524 * we hash the TCP subflows not the master socket
3530 static void mptcp_unhash(struct sock *sk)
3532 /* called from sk_common_release(), but nothing to do here */
3535 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3537 struct mptcp_sock *msk = mptcp_sk(sk);
3539 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3540 if (WARN_ON_ONCE(!msk->first))
3543 return inet_csk_get_port(msk->first, snum);
3546 void mptcp_finish_connect(struct sock *ssk)
3548 struct mptcp_subflow_context *subflow;
3549 struct mptcp_sock *msk;
3552 subflow = mptcp_subflow_ctx(ssk);
3556 pr_debug("msk=%p, token=%u", sk, subflow->token);
3558 subflow->map_seq = subflow->iasn;
3559 subflow->map_subflow_seq = 1;
3561 /* the socket is not connected yet, no msk/subflow ops can access/race
3562 * accessing the field below
3564 WRITE_ONCE(msk->local_key, subflow->local_key);
3566 mptcp_pm_new_connection(msk, ssk, 0);
3569 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3571 write_lock_bh(&sk->sk_callback_lock);
3572 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3573 sk_set_socket(sk, parent);
3574 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3575 write_unlock_bh(&sk->sk_callback_lock);
3578 bool mptcp_finish_join(struct sock *ssk)
3580 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3581 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3582 struct sock *parent = (void *)msk;
3585 pr_debug("msk=%p, subflow=%p", msk, subflow);
3587 /* mptcp socket already closing? */
3588 if (!mptcp_is_fully_established(parent)) {
3589 subflow->reset_reason = MPTCP_RST_EMPTCP;
3593 /* active subflow, already present inside the conn_list */
3594 if (!list_empty(&subflow->node)) {
3595 mptcp_subflow_joined(msk, ssk);
3596 mptcp_propagate_sndbuf(parent, ssk);
3600 if (!mptcp_pm_allow_new_subflow(msk))
3601 goto err_prohibited;
3603 /* If we can't acquire msk socket lock here, let the release callback
3606 mptcp_data_lock(parent);
3607 if (!sock_owned_by_user(parent)) {
3608 ret = __mptcp_finish_join(msk, ssk);
3611 list_add_tail(&subflow->node, &msk->conn_list);
3615 list_add_tail(&subflow->node, &msk->join_list);
3616 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3618 mptcp_data_unlock(parent);
3622 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3629 static void mptcp_shutdown(struct sock *sk, int how)
3631 pr_debug("sk=%p, how=%d", sk, how);
3633 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3634 __mptcp_wr_shutdown(sk);
3637 static int mptcp_forward_alloc_get(const struct sock *sk)
3639 return READ_ONCE(sk->sk_forward_alloc) +
3640 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3643 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3645 const struct sock *sk = (void *)msk;
3648 if (sk->sk_state == TCP_LISTEN)
3651 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3654 delta = msk->write_seq - v;
3655 if (__mptcp_check_fallback(msk) && msk->first) {
3656 struct tcp_sock *tp = tcp_sk(msk->first);
3658 /* the first subflow is disconnected after close - see
3659 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3660 * so ignore that status, too.
3662 if (!((1 << msk->first->sk_state) &
3663 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3664 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3666 if (delta > INT_MAX)
3672 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3674 struct mptcp_sock *msk = mptcp_sk(sk);
3679 if (sk->sk_state == TCP_LISTEN)
3683 __mptcp_move_skbs(msk);
3684 *karg = mptcp_inq_hint(sk);
3688 slow = lock_sock_fast(sk);
3689 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3690 unlock_sock_fast(sk, slow);
3693 slow = lock_sock_fast(sk);
3694 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3695 unlock_sock_fast(sk, slow);
3698 return -ENOIOCTLCMD;
3704 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3705 struct mptcp_subflow_context *subflow)
3707 subflow->request_mptcp = 0;
3708 __mptcp_do_fallback(msk);
3711 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3713 struct mptcp_subflow_context *subflow;
3714 struct mptcp_sock *msk = mptcp_sk(sk);
3718 ssk = __mptcp_nmpc_sk(msk);
3720 return PTR_ERR(ssk);
3722 mptcp_set_state(sk, TCP_SYN_SENT);
3723 subflow = mptcp_subflow_ctx(ssk);
3724 #ifdef CONFIG_TCP_MD5SIG
3725 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3728 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3729 mptcp_subflow_early_fallback(msk, subflow);
3731 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3732 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3733 mptcp_subflow_early_fallback(msk, subflow);
3736 WRITE_ONCE(msk->write_seq, subflow->idsn);
3737 WRITE_ONCE(msk->snd_nxt, subflow->idsn);
3738 if (likely(!__mptcp_check_fallback(msk)))
3739 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3741 /* if reaching here via the fastopen/sendmsg path, the caller already
3742 * acquired the subflow socket lock, too.
3744 if (!msk->fastopening)
3747 /* the following mirrors closely a very small chunk of code from
3748 * __inet_stream_connect()
3750 if (ssk->sk_state != TCP_CLOSE)
3753 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3754 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3759 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3763 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3766 if (!msk->fastopening)
3769 /* on successful connect, the msk state will be moved to established by
3770 * subflow_finish_connect()
3772 if (unlikely(err)) {
3773 /* avoid leaving a dangling token in an unconnected socket */
3774 mptcp_token_destroy(msk);
3775 mptcp_set_state(sk, TCP_CLOSE);
3779 mptcp_copy_inaddrs(sk, ssk);
3783 static struct proto mptcp_prot = {
3785 .owner = THIS_MODULE,
3786 .init = mptcp_init_sock,
3787 .connect = mptcp_connect,
3788 .disconnect = mptcp_disconnect,
3789 .close = mptcp_close,
3790 .setsockopt = mptcp_setsockopt,
3791 .getsockopt = mptcp_getsockopt,
3792 .shutdown = mptcp_shutdown,
3793 .destroy = mptcp_destroy,
3794 .sendmsg = mptcp_sendmsg,
3795 .ioctl = mptcp_ioctl,
3796 .recvmsg = mptcp_recvmsg,
3797 .release_cb = mptcp_release_cb,
3799 .unhash = mptcp_unhash,
3800 .get_port = mptcp_get_port,
3801 .forward_alloc_get = mptcp_forward_alloc_get,
3802 .stream_memory_free = mptcp_stream_memory_free,
3803 .sockets_allocated = &mptcp_sockets_allocated,
3805 .memory_allocated = &tcp_memory_allocated,
3806 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3808 .memory_pressure = &tcp_memory_pressure,
3809 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3810 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3811 .sysctl_mem = sysctl_tcp_mem,
3812 .obj_size = sizeof(struct mptcp_sock),
3813 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3814 .no_autobind = true,
3817 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3819 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3820 struct sock *ssk, *sk = sock->sk;
3824 ssk = __mptcp_nmpc_sk(msk);
3830 if (sk->sk_family == AF_INET)
3831 err = inet_bind_sk(ssk, uaddr, addr_len);
3832 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3833 else if (sk->sk_family == AF_INET6)
3834 err = inet6_bind_sk(ssk, uaddr, addr_len);
3837 mptcp_copy_inaddrs(sk, ssk);
3844 static int mptcp_listen(struct socket *sock, int backlog)
3846 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3847 struct sock *sk = sock->sk;
3851 pr_debug("msk=%p", msk);
3856 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3859 ssk = __mptcp_nmpc_sk(msk);
3865 mptcp_set_state(sk, TCP_LISTEN);
3866 sock_set_flag(sk, SOCK_RCU_FREE);
3869 err = __inet_listen_sk(ssk, backlog);
3871 mptcp_set_state(sk, inet_sk_state_load(ssk));
3874 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3875 mptcp_copy_inaddrs(sk, ssk);
3876 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3884 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3885 struct proto_accept_arg *arg)
3887 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3888 struct sock *ssk, *newsk;
3890 pr_debug("msk=%p", msk);
3892 /* Buggy applications can call accept on socket states other then LISTEN
3893 * but no need to allocate the first subflow just to error out.
3895 ssk = READ_ONCE(msk->first);
3899 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3900 newsk = inet_csk_accept(ssk, arg);
3904 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3905 if (sk_is_mptcp(newsk)) {
3906 struct mptcp_subflow_context *subflow;
3907 struct sock *new_mptcp_sock;
3909 subflow = mptcp_subflow_ctx(newsk);
3910 new_mptcp_sock = subflow->conn;
3912 /* is_mptcp should be false if subflow->conn is missing, see
3913 * subflow_syn_recv_sock()
3915 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3916 tcp_sk(newsk)->is_mptcp = 0;
3920 newsk = new_mptcp_sock;
3921 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3923 newsk->sk_kern_sock = arg->kern;
3925 __inet_accept(sock, newsock, newsk);
3927 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3928 msk = mptcp_sk(newsk);
3929 msk->in_accept_queue = 0;
3931 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3932 * This is needed so NOSPACE flag can be set from tcp stack.
3934 mptcp_for_each_subflow(msk, subflow) {
3935 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3937 if (!ssk->sk_socket)
3938 mptcp_sock_graft(ssk, newsock);
3941 /* Do late cleanup for the first subflow as necessary. Also
3942 * deal with bad peers not doing a complete shutdown.
3944 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3945 __mptcp_close_ssk(newsk, msk->first,
3946 mptcp_subflow_ctx(msk->first), 0);
3947 if (unlikely(list_is_singular(&msk->conn_list)))
3948 mptcp_set_state(newsk, TCP_CLOSE);
3952 newsk->sk_kern_sock = arg->kern;
3954 __inet_accept(sock, newsock, newsk);
3955 /* we are being invoked after accepting a non-mp-capable
3956 * flow: sk is a tcp_sk, not an mptcp one.
3958 * Hand the socket over to tcp so all further socket ops
3961 WRITE_ONCE(newsock->sk->sk_socket->ops,
3962 mptcp_fallback_tcp_ops(newsock->sk));
3964 release_sock(newsk);
3969 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3971 struct sock *sk = (struct sock *)msk;
3973 if (__mptcp_stream_is_writeable(sk, 1))
3974 return EPOLLOUT | EPOLLWRNORM;
3976 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
3977 smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
3978 if (__mptcp_stream_is_writeable(sk, 1))
3979 return EPOLLOUT | EPOLLWRNORM;
3984 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3985 struct poll_table_struct *wait)
3987 struct sock *sk = sock->sk;
3988 struct mptcp_sock *msk;
3994 sock_poll_wait(file, sock, wait);
3996 state = inet_sk_state_load(sk);
3997 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3998 if (state == TCP_LISTEN) {
3999 struct sock *ssk = READ_ONCE(msk->first);
4001 if (WARN_ON_ONCE(!ssk))
4004 return inet_csk_listen_poll(ssk);
4007 shutdown = READ_ONCE(sk->sk_shutdown);
4008 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
4010 if (shutdown & RCV_SHUTDOWN)
4011 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
4013 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
4014 mask |= mptcp_check_readable(sk);
4015 if (shutdown & SEND_SHUTDOWN)
4016 mask |= EPOLLOUT | EPOLLWRNORM;
4018 mask |= mptcp_check_writeable(msk);
4019 } else if (state == TCP_SYN_SENT &&
4020 inet_test_bit(DEFER_CONNECT, sk)) {
4021 /* cf tcp_poll() note about TFO */
4022 mask |= EPOLLOUT | EPOLLWRNORM;
4025 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
4027 if (READ_ONCE(sk->sk_err))
4033 static const struct proto_ops mptcp_stream_ops = {
4035 .owner = THIS_MODULE,
4036 .release = inet_release,
4038 .connect = inet_stream_connect,
4039 .socketpair = sock_no_socketpair,
4040 .accept = mptcp_stream_accept,
4041 .getname = inet_getname,
4043 .ioctl = inet_ioctl,
4044 .gettstamp = sock_gettstamp,
4045 .listen = mptcp_listen,
4046 .shutdown = inet_shutdown,
4047 .setsockopt = sock_common_setsockopt,
4048 .getsockopt = sock_common_getsockopt,
4049 .sendmsg = inet_sendmsg,
4050 .recvmsg = inet_recvmsg,
4051 .mmap = sock_no_mmap,
4052 .set_rcvlowat = mptcp_set_rcvlowat,
4055 static struct inet_protosw mptcp_protosw = {
4056 .type = SOCK_STREAM,
4057 .protocol = IPPROTO_MPTCP,
4058 .prot = &mptcp_prot,
4059 .ops = &mptcp_stream_ops,
4060 .flags = INET_PROTOSW_ICSK,
4063 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4065 struct mptcp_delegated_action *delegated;
4066 struct mptcp_subflow_context *subflow;
4069 delegated = container_of(napi, struct mptcp_delegated_action, napi);
4070 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4071 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4073 bh_lock_sock_nested(ssk);
4074 if (!sock_owned_by_user(ssk)) {
4075 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4077 /* tcp_release_cb_override already processed
4078 * the action or will do at next release_sock().
4079 * In both case must dequeue the subflow here - on the same
4080 * CPU that scheduled it.
4083 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4085 bh_unlock_sock(ssk);
4088 if (++work_done == budget)
4092 /* always provide a 0 'work_done' argument, so that napi_complete_done
4093 * will not try accessing the NULL napi->dev ptr
4095 napi_complete_done(napi, 0);
4099 void __init mptcp_proto_init(void)
4101 struct mptcp_delegated_action *delegated;
4104 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4106 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4107 panic("Failed to allocate MPTCP pcpu counter\n");
4109 init_dummy_netdev(&mptcp_napi_dev);
4110 for_each_possible_cpu(cpu) {
4111 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4112 INIT_LIST_HEAD(&delegated->head);
4113 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4115 napi_enable(&delegated->napi);
4118 mptcp_subflow_init();
4123 if (proto_register(&mptcp_prot, 1) != 0)
4124 panic("Failed to register MPTCP proto.\n");
4126 inet_register_protosw(&mptcp_protosw);
4128 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4131 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4132 static const struct proto_ops mptcp_v6_stream_ops = {
4134 .owner = THIS_MODULE,
4135 .release = inet6_release,
4137 .connect = inet_stream_connect,
4138 .socketpair = sock_no_socketpair,
4139 .accept = mptcp_stream_accept,
4140 .getname = inet6_getname,
4142 .ioctl = inet6_ioctl,
4143 .gettstamp = sock_gettstamp,
4144 .listen = mptcp_listen,
4145 .shutdown = inet_shutdown,
4146 .setsockopt = sock_common_setsockopt,
4147 .getsockopt = sock_common_getsockopt,
4148 .sendmsg = inet6_sendmsg,
4149 .recvmsg = inet6_recvmsg,
4150 .mmap = sock_no_mmap,
4151 #ifdef CONFIG_COMPAT
4152 .compat_ioctl = inet6_compat_ioctl,
4154 .set_rcvlowat = mptcp_set_rcvlowat,
4157 static struct proto mptcp_v6_prot;
4159 static struct inet_protosw mptcp_v6_protosw = {
4160 .type = SOCK_STREAM,
4161 .protocol = IPPROTO_MPTCP,
4162 .prot = &mptcp_v6_prot,
4163 .ops = &mptcp_v6_stream_ops,
4164 .flags = INET_PROTOSW_ICSK,
4167 int __init mptcp_proto_v6_init(void)
4171 mptcp_v6_prot = mptcp_prot;
4172 strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name));
4173 mptcp_v6_prot.slab = NULL;
4174 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4175 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4177 err = proto_register(&mptcp_v6_prot, 1);
4181 err = inet6_register_protosw(&mptcp_v6_protosw);
4183 proto_unregister(&mptcp_v6_prot);