1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright (c) 2017 - 2019, Intel Corporation.
7 #define pr_fmt(fmt) "MPTCP: " fmt
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
23 #include <net/mptcp.h>
25 #include <asm/ioctls.h>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
34 struct mptcp_sock msk;
40 MPTCP_CMSG_TS = BIT(0),
41 MPTCP_CMSG_INQ = BIT(1),
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void mptcp_check_send_data_fin(struct sock *sk);
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
52 /* Returns end sequence number of the receiver's advertised window */
53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
55 return READ_ONCE(msk->wnd_end);
58 static 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 static bool mptcp_pending_data_fin_ack(struct sock *sk)
415 struct mptcp_sock *msk = mptcp_sk(sk);
417 return ((1 << sk->sk_state) &
418 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
419 msk->write_seq == READ_ONCE(msk->snd_una);
422 static void mptcp_check_data_fin_ack(struct sock *sk)
424 struct mptcp_sock *msk = mptcp_sk(sk);
426 /* Look for an acknowledged DATA_FIN */
427 if (mptcp_pending_data_fin_ack(sk)) {
428 WRITE_ONCE(msk->snd_data_fin_enable, 0);
430 switch (sk->sk_state) {
432 mptcp_set_state(sk, TCP_FIN_WAIT2);
436 mptcp_set_state(sk, TCP_CLOSE);
440 mptcp_close_wake_up(sk);
444 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
446 struct mptcp_sock *msk = mptcp_sk(sk);
448 if (READ_ONCE(msk->rcv_data_fin) &&
449 ((1 << sk->sk_state) &
450 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
451 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
453 if (msk->ack_seq == rcv_data_fin_seq) {
455 *seq = rcv_data_fin_seq;
464 static void mptcp_set_datafin_timeout(struct sock *sk)
466 struct inet_connection_sock *icsk = inet_csk(sk);
469 retransmits = min_t(u32, icsk->icsk_retransmits,
470 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
472 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
475 static void __mptcp_set_timeout(struct sock *sk, long tout)
477 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
480 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
482 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
484 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
485 inet_csk(ssk)->icsk_timeout - jiffies : 0;
488 static void mptcp_set_timeout(struct sock *sk)
490 struct mptcp_subflow_context *subflow;
493 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
494 tout = max(tout, mptcp_timeout_from_subflow(subflow));
495 __mptcp_set_timeout(sk, tout);
498 static inline bool tcp_can_send_ack(const struct sock *ssk)
500 return !((1 << inet_sk_state_load(ssk)) &
501 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
504 void __mptcp_subflow_send_ack(struct sock *ssk)
506 if (tcp_can_send_ack(ssk))
510 static void mptcp_subflow_send_ack(struct sock *ssk)
514 slow = lock_sock_fast(ssk);
515 __mptcp_subflow_send_ack(ssk);
516 unlock_sock_fast(ssk, slow);
519 static void mptcp_send_ack(struct mptcp_sock *msk)
521 struct mptcp_subflow_context *subflow;
523 mptcp_for_each_subflow(msk, subflow)
524 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
527 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
531 slow = lock_sock_fast(ssk);
532 if (tcp_can_send_ack(ssk))
533 tcp_cleanup_rbuf(ssk, 1);
534 unlock_sock_fast(ssk, slow);
537 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
539 const struct inet_connection_sock *icsk = inet_csk(ssk);
540 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
541 const struct tcp_sock *tp = tcp_sk(ssk);
543 return (ack_pending & ICSK_ACK_SCHED) &&
544 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
545 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
546 (rx_empty && ack_pending &
547 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
550 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
552 int old_space = READ_ONCE(msk->old_wspace);
553 struct mptcp_subflow_context *subflow;
554 struct sock *sk = (struct sock *)msk;
555 int space = __mptcp_space(sk);
556 bool cleanup, rx_empty;
558 cleanup = (space > 0) && (space >= (old_space << 1));
559 rx_empty = !__mptcp_rmem(sk);
561 mptcp_for_each_subflow(msk, subflow) {
562 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
564 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
565 mptcp_subflow_cleanup_rbuf(ssk);
569 static bool mptcp_check_data_fin(struct sock *sk)
571 struct mptcp_sock *msk = mptcp_sk(sk);
572 u64 rcv_data_fin_seq;
575 /* Need to ack a DATA_FIN received from a peer while this side
576 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
577 * msk->rcv_data_fin was set when parsing the incoming options
578 * at the subflow level and the msk lock was not held, so this
579 * is the first opportunity to act on the DATA_FIN and change
582 * If we are caught up to the sequence number of the incoming
583 * DATA_FIN, send the DATA_ACK now and do state transition. If
584 * not caught up, do nothing and let the recv code send DATA_ACK
588 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
589 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
590 WRITE_ONCE(msk->rcv_data_fin, 0);
592 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
593 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
595 switch (sk->sk_state) {
596 case TCP_ESTABLISHED:
597 mptcp_set_state(sk, TCP_CLOSE_WAIT);
600 mptcp_set_state(sk, TCP_CLOSING);
603 mptcp_set_state(sk, TCP_CLOSE);
606 /* Other states not expected */
612 if (!__mptcp_check_fallback(msk))
614 mptcp_close_wake_up(sk);
619 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
623 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
624 struct sock *sk = (struct sock *)msk;
625 unsigned int moved = 0;
626 bool more_data_avail;
631 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
633 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
634 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
636 if (unlikely(ssk_rbuf > sk_rbuf)) {
637 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
642 pr_debug("msk=%p ssk=%p", msk, ssk);
645 u32 map_remaining, offset;
646 u32 seq = tp->copied_seq;
650 /* try to move as much data as available */
651 map_remaining = subflow->map_data_len -
652 mptcp_subflow_get_map_offset(subflow);
654 skb = skb_peek(&ssk->sk_receive_queue);
656 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
657 * a different CPU can have already processed the pending
658 * data, stop here or we can enter an infinite loop
665 if (__mptcp_check_fallback(msk)) {
666 /* Under fallback skbs have no MPTCP extension and TCP could
667 * collapse them between the dummy map creation and the
668 * current dequeue. Be sure to adjust the map size.
670 map_remaining = skb->len;
671 subflow->map_data_len = skb->len;
674 offset = seq - TCP_SKB_CB(skb)->seq;
675 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
681 if (offset < skb->len) {
682 size_t len = skb->len - offset;
687 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
691 if (WARN_ON_ONCE(map_remaining < len))
695 sk_eat_skb(ssk, skb);
699 WRITE_ONCE(tp->copied_seq, seq);
700 more_data_avail = mptcp_subflow_data_available(ssk);
702 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
706 } while (more_data_avail);
712 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
714 struct sock *sk = (struct sock *)msk;
715 struct sk_buff *skb, *tail;
720 p = rb_first(&msk->out_of_order_queue);
721 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
724 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
728 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
730 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
733 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
737 end_seq = MPTCP_SKB_CB(skb)->end_seq;
738 tail = skb_peek_tail(&sk->sk_receive_queue);
739 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
740 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
742 /* skip overlapping data, if any */
743 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
744 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
746 MPTCP_SKB_CB(skb)->offset += delta;
747 MPTCP_SKB_CB(skb)->map_seq += delta;
748 __skb_queue_tail(&sk->sk_receive_queue, skb);
750 msk->bytes_received += end_seq - msk->ack_seq;
751 msk->ack_seq = end_seq;
757 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
759 int err = sock_error(ssk);
765 /* only propagate errors on fallen-back sockets or
768 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
771 /* We need to propagate only transition to CLOSE state.
772 * Orphaned socket will see such state change via
773 * subflow_sched_work_if_closed() and that path will properly
774 * destroy the msk as needed.
776 ssk_state = inet_sk_state_load(ssk);
777 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
778 mptcp_set_state(sk, ssk_state);
779 WRITE_ONCE(sk->sk_err, -err);
781 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
787 void __mptcp_error_report(struct sock *sk)
789 struct mptcp_subflow_context *subflow;
790 struct mptcp_sock *msk = mptcp_sk(sk);
792 mptcp_for_each_subflow(msk, subflow)
793 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
797 /* In most cases we will be able to lock the mptcp socket. If its already
798 * owned, we need to defer to the work queue to avoid ABBA deadlock.
800 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
802 struct sock *sk = (struct sock *)msk;
803 unsigned int moved = 0;
805 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
806 __mptcp_ofo_queue(msk);
807 if (unlikely(ssk->sk_err)) {
808 if (!sock_owned_by_user(sk))
809 __mptcp_error_report(sk);
811 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
814 /* If the moves have caught up with the DATA_FIN sequence number
815 * it's time to ack the DATA_FIN and change socket state, but
816 * this is not a good place to change state. Let the workqueue
819 if (mptcp_pending_data_fin(sk, NULL))
820 mptcp_schedule_work(sk);
824 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
826 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
827 struct mptcp_sock *msk = mptcp_sk(sk);
828 int sk_rbuf, ssk_rbuf;
830 /* The peer can send data while we are shutting down this
831 * subflow at msk destruction time, but we must avoid enqueuing
832 * more data to the msk receive queue
834 if (unlikely(subflow->disposable))
837 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
838 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
839 if (unlikely(ssk_rbuf > sk_rbuf))
842 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
843 if (__mptcp_rmem(sk) > sk_rbuf) {
844 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
848 /* Wake-up the reader only for in-sequence data */
850 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
851 sk->sk_data_ready(sk);
852 mptcp_data_unlock(sk);
855 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
857 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
858 WRITE_ONCE(msk->allow_infinite_fallback, false);
859 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
862 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
864 struct sock *sk = (struct sock *)msk;
866 if (sk->sk_state != TCP_ESTABLISHED)
869 /* attach to msk socket only after we are sure we will deal with it
872 if (sk->sk_socket && !ssk->sk_socket)
873 mptcp_sock_graft(ssk, sk->sk_socket);
875 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
876 mptcp_sockopt_sync_locked(msk, ssk);
877 mptcp_subflow_joined(msk, ssk);
878 mptcp_stop_tout_timer(sk);
879 __mptcp_propagate_sndbuf(sk, ssk);
883 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
885 struct mptcp_subflow_context *tmp, *subflow;
886 struct mptcp_sock *msk = mptcp_sk(sk);
888 list_for_each_entry_safe(subflow, tmp, join_list, node) {
889 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
890 bool slow = lock_sock_fast(ssk);
892 list_move_tail(&subflow->node, &msk->conn_list);
893 if (!__mptcp_finish_join(msk, ssk))
894 mptcp_subflow_reset(ssk);
895 unlock_sock_fast(ssk, slow);
899 static bool mptcp_rtx_timer_pending(struct sock *sk)
901 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
904 static void mptcp_reset_rtx_timer(struct sock *sk)
906 struct inet_connection_sock *icsk = inet_csk(sk);
909 /* prevent rescheduling on close */
910 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
913 tout = mptcp_sk(sk)->timer_ival;
914 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
917 bool mptcp_schedule_work(struct sock *sk)
919 if (inet_sk_state_load(sk) != TCP_CLOSE &&
920 schedule_work(&mptcp_sk(sk)->work)) {
921 /* each subflow already holds a reference to the sk, and the
922 * workqueue is invoked by a subflow, so sk can't go away here.
930 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
932 struct mptcp_subflow_context *subflow;
934 msk_owned_by_me(msk);
936 mptcp_for_each_subflow(msk, subflow) {
937 if (READ_ONCE(subflow->data_avail))
938 return mptcp_subflow_tcp_sock(subflow);
944 static bool mptcp_skb_can_collapse_to(u64 write_seq,
945 const struct sk_buff *skb,
946 const struct mptcp_ext *mpext)
948 if (!tcp_skb_can_collapse_to(skb))
951 /* can collapse only if MPTCP level sequence is in order and this
952 * mapping has not been xmitted yet
954 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
958 /* we can append data to the given data frag if:
959 * - there is space available in the backing page_frag
960 * - the data frag tail matches the current page_frag free offset
961 * - the data frag end sequence number matches the current write seq
963 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
964 const struct page_frag *pfrag,
965 const struct mptcp_data_frag *df)
967 return df && pfrag->page == df->page &&
968 pfrag->size - pfrag->offset > 0 &&
969 pfrag->offset == (df->offset + df->data_len) &&
970 df->data_seq + df->data_len == msk->write_seq;
973 static void dfrag_uncharge(struct sock *sk, int len)
975 sk_mem_uncharge(sk, len);
976 sk_wmem_queued_add(sk, -len);
979 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
981 int len = dfrag->data_len + dfrag->overhead;
983 list_del(&dfrag->list);
984 dfrag_uncharge(sk, len);
985 put_page(dfrag->page);
988 static void __mptcp_clean_una(struct sock *sk)
990 struct mptcp_sock *msk = mptcp_sk(sk);
991 struct mptcp_data_frag *dtmp, *dfrag;
994 snd_una = msk->snd_una;
995 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
996 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
999 if (unlikely(dfrag == msk->first_pending)) {
1000 /* in recovery mode can see ack after the current snd head */
1001 if (WARN_ON_ONCE(!msk->recovery))
1004 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1007 dfrag_clear(sk, dfrag);
1010 dfrag = mptcp_rtx_head(sk);
1011 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1012 u64 delta = snd_una - dfrag->data_seq;
1014 /* prevent wrap around in recovery mode */
1015 if (unlikely(delta > dfrag->already_sent)) {
1016 if (WARN_ON_ONCE(!msk->recovery))
1018 if (WARN_ON_ONCE(delta > dfrag->data_len))
1020 dfrag->already_sent += delta - dfrag->already_sent;
1023 dfrag->data_seq += delta;
1024 dfrag->offset += delta;
1025 dfrag->data_len -= delta;
1026 dfrag->already_sent -= delta;
1028 dfrag_uncharge(sk, delta);
1031 /* all retransmitted data acked, recovery completed */
1032 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1033 msk->recovery = false;
1036 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1037 snd_una == READ_ONCE(msk->write_seq)) {
1038 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1039 mptcp_stop_rtx_timer(sk);
1041 mptcp_reset_rtx_timer(sk);
1045 static void __mptcp_clean_una_wakeup(struct sock *sk)
1047 lockdep_assert_held_once(&sk->sk_lock.slock);
1049 __mptcp_clean_una(sk);
1050 mptcp_write_space(sk);
1053 static void mptcp_clean_una_wakeup(struct sock *sk)
1055 mptcp_data_lock(sk);
1056 __mptcp_clean_una_wakeup(sk);
1057 mptcp_data_unlock(sk);
1060 static void mptcp_enter_memory_pressure(struct sock *sk)
1062 struct mptcp_subflow_context *subflow;
1063 struct mptcp_sock *msk = mptcp_sk(sk);
1066 mptcp_for_each_subflow(msk, subflow) {
1067 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1070 tcp_enter_memory_pressure(ssk);
1071 sk_stream_moderate_sndbuf(ssk);
1075 __mptcp_sync_sndbuf(sk);
1078 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1081 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1083 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1084 pfrag, sk->sk_allocation)))
1087 mptcp_enter_memory_pressure(sk);
1091 static struct mptcp_data_frag *
1092 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1095 int offset = ALIGN(orig_offset, sizeof(long));
1096 struct mptcp_data_frag *dfrag;
1098 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1099 dfrag->data_len = 0;
1100 dfrag->data_seq = msk->write_seq;
1101 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1102 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1103 dfrag->already_sent = 0;
1104 dfrag->page = pfrag->page;
1109 struct mptcp_sendmsg_info {
1115 bool data_lock_held;
1118 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1119 u64 data_seq, int avail_size)
1121 u64 window_end = mptcp_wnd_end(msk);
1124 if (__mptcp_check_fallback(msk))
1127 mptcp_snd_wnd = window_end - data_seq;
1128 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1130 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1131 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1132 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1138 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1140 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1144 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1148 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1150 struct sk_buff *skb;
1152 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1154 if (likely(__mptcp_add_ext(skb, gfp))) {
1155 skb_reserve(skb, MAX_TCP_HEADER);
1156 skb->ip_summed = CHECKSUM_PARTIAL;
1157 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1162 mptcp_enter_memory_pressure(sk);
1167 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1169 struct sk_buff *skb;
1171 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1175 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1176 tcp_skb_entail(ssk, skb);
1179 tcp_skb_tsorted_anchor_cleanup(skb);
1184 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1186 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1188 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1191 /* note: this always recompute the csum on the whole skb, even
1192 * if we just appended a single frag. More status info needed
1194 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1196 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1197 __wsum csum = ~csum_unfold(mpext->csum);
1198 int offset = skb->len - added;
1200 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1203 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1205 struct mptcp_ext *mpext)
1210 mpext->infinite_map = 1;
1211 mpext->data_len = 0;
1213 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1214 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1216 mptcp_do_fallback(ssk);
1219 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1221 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1222 struct mptcp_data_frag *dfrag,
1223 struct mptcp_sendmsg_info *info)
1225 u64 data_seq = dfrag->data_seq + info->sent;
1226 int offset = dfrag->offset + info->sent;
1227 struct mptcp_sock *msk = mptcp_sk(sk);
1228 bool zero_window_probe = false;
1229 struct mptcp_ext *mpext = NULL;
1230 bool can_coalesce = false;
1231 bool reuse_skb = true;
1232 struct sk_buff *skb;
1236 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1237 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1239 if (WARN_ON_ONCE(info->sent > info->limit ||
1240 info->limit > dfrag->data_len))
1243 if (unlikely(!__tcp_can_send(ssk)))
1246 /* compute send limit */
1247 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1248 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1249 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1250 copy = info->size_goal;
1252 skb = tcp_write_queue_tail(ssk);
1253 if (skb && copy > skb->len) {
1254 /* Limit the write to the size available in the
1255 * current skb, if any, so that we create at most a new skb.
1256 * Explicitly tells TCP internals to avoid collapsing on later
1257 * queue management operation, to avoid breaking the ext <->
1258 * SSN association set here
1260 mpext = mptcp_get_ext(skb);
1261 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1262 TCP_SKB_CB(skb)->eor = 1;
1263 tcp_mark_push(tcp_sk(ssk), skb);
1267 i = skb_shinfo(skb)->nr_frags;
1268 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1269 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1270 tcp_mark_push(tcp_sk(ssk), skb);
1277 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1281 i = skb_shinfo(skb)->nr_frags;
1283 mpext = mptcp_get_ext(skb);
1286 /* Zero window and all data acked? Probe. */
1287 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1289 u64 snd_una = READ_ONCE(msk->snd_una);
1291 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1292 tcp_remove_empty_skb(ssk);
1296 zero_window_probe = true;
1297 data_seq = snd_una - 1;
1301 copy = min_t(size_t, copy, info->limit - info->sent);
1302 if (!sk_wmem_schedule(ssk, copy)) {
1303 tcp_remove_empty_skb(ssk);
1308 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1310 get_page(dfrag->page);
1311 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1315 skb->data_len += copy;
1316 skb->truesize += copy;
1317 sk_wmem_queued_add(ssk, copy);
1318 sk_mem_charge(ssk, copy);
1319 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1320 TCP_SKB_CB(skb)->end_seq += copy;
1321 tcp_skb_pcount_set(skb, 0);
1323 /* on skb reuse we just need to update the DSS len */
1325 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1326 mpext->data_len += copy;
1330 memset(mpext, 0, sizeof(*mpext));
1331 mpext->data_seq = data_seq;
1332 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1333 mpext->data_len = copy;
1337 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1338 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1341 if (zero_window_probe) {
1342 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1344 if (READ_ONCE(msk->csum_enabled))
1345 mptcp_update_data_checksum(skb, copy);
1346 tcp_push_pending_frames(ssk);
1350 if (READ_ONCE(msk->csum_enabled))
1351 mptcp_update_data_checksum(skb, copy);
1352 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1353 mptcp_update_infinite_map(msk, ssk, mpext);
1354 trace_mptcp_sendmsg_frag(mpext);
1355 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1359 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1360 sizeof(struct tcphdr) - \
1361 MAX_TCP_OPTION_SPACE - \
1362 sizeof(struct ipv6hdr) - \
1363 sizeof(struct frag_hdr))
1365 struct subflow_send_info {
1370 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1372 if (!subflow->stale)
1376 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1379 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1381 if (unlikely(subflow->stale)) {
1382 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1384 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1387 mptcp_subflow_set_active(subflow);
1389 return __mptcp_subflow_active(subflow);
1392 #define SSK_MODE_ACTIVE 0
1393 #define SSK_MODE_BACKUP 1
1394 #define SSK_MODE_MAX 2
1396 /* implement the mptcp packet scheduler;
1397 * returns the subflow that will transmit the next DSS
1398 * additionally updates the rtx timeout
1400 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1402 struct subflow_send_info send_info[SSK_MODE_MAX];
1403 struct mptcp_subflow_context *subflow;
1404 struct sock *sk = (struct sock *)msk;
1405 u32 pace, burst, wmem;
1406 int i, nr_active = 0;
1411 /* pick the subflow with the lower wmem/wspace ratio */
1412 for (i = 0; i < SSK_MODE_MAX; ++i) {
1413 send_info[i].ssk = NULL;
1414 send_info[i].linger_time = -1;
1417 mptcp_for_each_subflow(msk, subflow) {
1418 trace_mptcp_subflow_get_send(subflow);
1419 ssk = mptcp_subflow_tcp_sock(subflow);
1420 if (!mptcp_subflow_active(subflow))
1423 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1424 nr_active += !subflow->backup;
1425 pace = subflow->avg_pacing_rate;
1426 if (unlikely(!pace)) {
1427 /* init pacing rate from socket */
1428 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1429 pace = subflow->avg_pacing_rate;
1434 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1435 if (linger_time < send_info[subflow->backup].linger_time) {
1436 send_info[subflow->backup].ssk = ssk;
1437 send_info[subflow->backup].linger_time = linger_time;
1440 __mptcp_set_timeout(sk, tout);
1442 /* pick the best backup if no other subflow is active */
1444 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1446 /* According to the blest algorithm, to avoid HoL blocking for the
1447 * faster flow, we need to:
1448 * - estimate the faster flow linger time
1449 * - use the above to estimate the amount of byte transferred
1450 * by the faster flow
1451 * - check that the amount of queued data is greter than the above,
1452 * otherwise do not use the picked, slower, subflow
1453 * We select the subflow with the shorter estimated time to flush
1454 * the queued mem, which basically ensure the above. We just need
1455 * to check that subflow has a non empty cwin.
1457 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1458 if (!ssk || !sk_stream_memory_free(ssk))
1461 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1462 wmem = READ_ONCE(ssk->sk_wmem_queued);
1466 subflow = mptcp_subflow_ctx(ssk);
1467 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1468 READ_ONCE(ssk->sk_pacing_rate) * burst,
1470 msk->snd_burst = burst;
1474 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1476 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1480 static void mptcp_update_post_push(struct mptcp_sock *msk,
1481 struct mptcp_data_frag *dfrag,
1484 u64 snd_nxt_new = dfrag->data_seq;
1486 dfrag->already_sent += sent;
1488 msk->snd_burst -= sent;
1490 snd_nxt_new += dfrag->already_sent;
1492 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1493 * is recovering after a failover. In that event, this re-sends
1496 * Thus compute snd_nxt_new candidate based on
1497 * the dfrag->data_seq that was sent and the data
1498 * that has been handed to the subflow for transmission
1499 * and skip update in case it was old dfrag.
1501 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1502 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1503 msk->snd_nxt = snd_nxt_new;
1507 void mptcp_check_and_set_pending(struct sock *sk)
1509 if (mptcp_send_head(sk)) {
1510 mptcp_data_lock(sk);
1511 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1512 mptcp_data_unlock(sk);
1516 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1517 struct mptcp_sendmsg_info *info)
1519 struct mptcp_sock *msk = mptcp_sk(sk);
1520 struct mptcp_data_frag *dfrag;
1521 int len, copied = 0, err = 0;
1523 while ((dfrag = mptcp_send_head(sk))) {
1524 info->sent = dfrag->already_sent;
1525 info->limit = dfrag->data_len;
1526 len = dfrag->data_len - dfrag->already_sent;
1530 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1532 err = copied ? : ret;
1540 mptcp_update_post_push(msk, dfrag, ret);
1542 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1544 if (msk->snd_burst <= 0 ||
1545 !sk_stream_memory_free(ssk) ||
1546 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1550 mptcp_set_timeout(sk);
1558 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1560 struct sock *prev_ssk = NULL, *ssk = NULL;
1561 struct mptcp_sock *msk = mptcp_sk(sk);
1562 struct mptcp_sendmsg_info info = {
1565 bool do_check_data_fin = false;
1568 while (mptcp_send_head(sk) && (push_count > 0)) {
1569 struct mptcp_subflow_context *subflow;
1572 if (mptcp_sched_get_send(msk))
1577 mptcp_for_each_subflow(msk, subflow) {
1578 if (READ_ONCE(subflow->scheduled)) {
1579 mptcp_subflow_set_scheduled(subflow, false);
1582 ssk = mptcp_subflow_tcp_sock(subflow);
1583 if (ssk != prev_ssk) {
1584 /* First check. If the ssk has changed since
1585 * the last round, release prev_ssk
1588 mptcp_push_release(prev_ssk, &info);
1590 /* Need to lock the new subflow only if different
1591 * from the previous one, otherwise we are still
1592 * helding the relevant lock
1599 ret = __subflow_push_pending(sk, ssk, &info);
1601 if (ret != -EAGAIN ||
1602 (1 << ssk->sk_state) &
1603 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1607 do_check_data_fin = true;
1612 /* at this point we held the socket lock for the last subflow we used */
1614 mptcp_push_release(ssk, &info);
1616 /* ensure the rtx timer is running */
1617 if (!mptcp_rtx_timer_pending(sk))
1618 mptcp_reset_rtx_timer(sk);
1619 if (do_check_data_fin)
1620 mptcp_check_send_data_fin(sk);
1623 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1625 struct mptcp_sock *msk = mptcp_sk(sk);
1626 struct mptcp_sendmsg_info info = {
1627 .data_lock_held = true,
1629 bool keep_pushing = true;
1630 struct sock *xmit_ssk;
1634 while (mptcp_send_head(sk) && keep_pushing) {
1635 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1638 /* check for a different subflow usage only after
1639 * spooling the first chunk of data
1642 mptcp_subflow_set_scheduled(subflow, false);
1643 ret = __subflow_push_pending(sk, ssk, &info);
1651 if (mptcp_sched_get_send(msk))
1654 if (READ_ONCE(subflow->scheduled)) {
1655 mptcp_subflow_set_scheduled(subflow, false);
1656 ret = __subflow_push_pending(sk, ssk, &info);
1658 keep_pushing = false;
1662 mptcp_for_each_subflow(msk, subflow) {
1663 if (READ_ONCE(subflow->scheduled)) {
1664 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1665 if (xmit_ssk != ssk) {
1666 mptcp_subflow_delegate(subflow,
1667 MPTCP_DELEGATE_SEND);
1668 keep_pushing = false;
1675 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1676 * not going to flush it via release_sock()
1679 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1681 if (!mptcp_rtx_timer_pending(sk))
1682 mptcp_reset_rtx_timer(sk);
1684 if (msk->snd_data_fin_enable &&
1685 msk->snd_nxt + 1 == msk->write_seq)
1686 mptcp_schedule_work(sk);
1690 static void mptcp_set_nospace(struct sock *sk)
1692 /* enable autotune */
1693 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1695 /* will be cleared on avail space */
1696 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
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 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1771 struct mptcp_sock *msk = mptcp_sk(sk);
1772 struct page_frag *pfrag;
1777 /* silently ignore everything else */
1778 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1782 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1783 msg->msg_flags & MSG_FASTOPEN)) {
1786 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1787 copied += copied_syn;
1788 if (ret == -EINPROGRESS && copied_syn > 0)
1794 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1796 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1797 ret = sk_stream_wait_connect(sk, &timeo);
1803 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1806 pfrag = sk_page_frag(sk);
1808 while (msg_data_left(msg)) {
1809 int total_ts, frag_truesize = 0;
1810 struct mptcp_data_frag *dfrag;
1811 bool dfrag_collapsed;
1812 size_t psize, offset;
1814 /* reuse tail pfrag, if possible, or carve a new one from the
1817 dfrag = mptcp_pending_tail(sk);
1818 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1819 if (!dfrag_collapsed) {
1820 if (!sk_stream_memory_free(sk))
1821 goto wait_for_memory;
1823 if (!mptcp_page_frag_refill(sk, pfrag))
1824 goto wait_for_memory;
1826 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1827 frag_truesize = dfrag->overhead;
1830 /* we do not bound vs wspace, to allow a single packet.
1831 * memory accounting will prevent execessive memory usage
1834 offset = dfrag->offset + dfrag->data_len;
1835 psize = pfrag->size - offset;
1836 psize = min_t(size_t, psize, msg_data_left(msg));
1837 total_ts = psize + frag_truesize;
1839 if (!sk_wmem_schedule(sk, total_ts))
1840 goto wait_for_memory;
1842 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1843 page_address(dfrag->page) + offset);
1847 /* data successfully copied into the write queue */
1848 sk_forward_alloc_add(sk, -total_ts);
1850 dfrag->data_len += psize;
1851 frag_truesize += psize;
1852 pfrag->offset += frag_truesize;
1853 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1855 /* charge data on mptcp pending queue to the msk socket
1856 * Note: we charge such data both to sk and ssk
1858 sk_wmem_queued_add(sk, frag_truesize);
1859 if (!dfrag_collapsed) {
1860 get_page(dfrag->page);
1861 list_add_tail(&dfrag->list, &msk->rtx_queue);
1862 if (!msk->first_pending)
1863 WRITE_ONCE(msk->first_pending, dfrag);
1865 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1866 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1872 mptcp_set_nospace(sk);
1873 __mptcp_push_pending(sk, msg->msg_flags);
1874 ret = sk_stream_wait_memory(sk, &timeo);
1880 __mptcp_push_pending(sk, msg->msg_flags);
1890 copied = sk_stream_error(sk, msg->msg_flags, ret);
1894 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1896 size_t len, int flags,
1897 struct scm_timestamping_internal *tss,
1900 struct sk_buff *skb, *tmp;
1903 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1904 u32 offset = MPTCP_SKB_CB(skb)->offset;
1905 u32 data_len = skb->len - offset;
1906 u32 count = min_t(size_t, len - copied, data_len);
1909 if (!(flags & MSG_TRUNC)) {
1910 err = skb_copy_datagram_msg(skb, offset, msg, count);
1911 if (unlikely(err < 0)) {
1918 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1919 tcp_update_recv_tstamps(skb, tss);
1920 *cmsg_flags |= MPTCP_CMSG_TS;
1925 if (count < data_len) {
1926 if (!(flags & MSG_PEEK)) {
1927 MPTCP_SKB_CB(skb)->offset += count;
1928 MPTCP_SKB_CB(skb)->map_seq += count;
1929 msk->bytes_consumed += count;
1934 if (!(flags & MSG_PEEK)) {
1935 /* we will bulk release the skb memory later */
1936 skb->destructor = NULL;
1937 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1938 __skb_unlink(skb, &msk->receive_queue);
1940 msk->bytes_consumed += count;
1950 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1952 * Only difference: Use highest rtt estimate of the subflows in use.
1954 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1956 struct mptcp_subflow_context *subflow;
1957 struct sock *sk = (struct sock *)msk;
1958 u8 scaling_ratio = U8_MAX;
1959 u32 time, advmss = 1;
1962 msk_owned_by_me(msk);
1967 if (!msk->rcvspace_init)
1968 mptcp_rcv_space_init(msk, msk->first);
1970 msk->rcvq_space.copied += copied;
1972 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1973 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1975 rtt_us = msk->rcvq_space.rtt_us;
1976 if (rtt_us && time < (rtt_us >> 3))
1980 mptcp_for_each_subflow(msk, subflow) {
1981 const struct tcp_sock *tp;
1985 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1987 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1988 sf_advmss = READ_ONCE(tp->advmss);
1990 rtt_us = max(sf_rtt_us, rtt_us);
1991 advmss = max(sf_advmss, advmss);
1992 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1995 msk->rcvq_space.rtt_us = rtt_us;
1996 msk->scaling_ratio = scaling_ratio;
1997 if (time < (rtt_us >> 3) || rtt_us == 0)
2000 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2003 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2004 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2008 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2010 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2012 do_div(grow, msk->rcvq_space.space);
2013 rcvwin += (grow << 1);
2015 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2016 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2018 if (rcvbuf > sk->sk_rcvbuf) {
2021 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2022 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2024 /* Make subflows follow along. If we do not do this, we
2025 * get drops at subflow level if skbs can't be moved to
2026 * the mptcp rx queue fast enough (announced rcv_win can
2027 * exceed ssk->sk_rcvbuf).
2029 mptcp_for_each_subflow(msk, subflow) {
2033 ssk = mptcp_subflow_tcp_sock(subflow);
2034 slow = lock_sock_fast(ssk);
2035 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2036 tcp_sk(ssk)->window_clamp = window_clamp;
2037 tcp_cleanup_rbuf(ssk, 1);
2038 unlock_sock_fast(ssk, slow);
2043 msk->rcvq_space.space = msk->rcvq_space.copied;
2045 msk->rcvq_space.copied = 0;
2046 msk->rcvq_space.time = mstamp;
2049 static void __mptcp_update_rmem(struct sock *sk)
2051 struct mptcp_sock *msk = mptcp_sk(sk);
2053 if (!msk->rmem_released)
2056 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2057 mptcp_rmem_uncharge(sk, msk->rmem_released);
2058 WRITE_ONCE(msk->rmem_released, 0);
2061 static void __mptcp_splice_receive_queue(struct sock *sk)
2063 struct mptcp_sock *msk = mptcp_sk(sk);
2065 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2068 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2070 struct sock *sk = (struct sock *)msk;
2071 unsigned int moved = 0;
2075 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2078 /* we can have data pending in the subflows only if the msk
2079 * receive buffer was full at subflow_data_ready() time,
2080 * that is an unlikely slow path.
2085 slowpath = lock_sock_fast(ssk);
2086 mptcp_data_lock(sk);
2087 __mptcp_update_rmem(sk);
2088 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2089 mptcp_data_unlock(sk);
2091 if (unlikely(ssk->sk_err))
2092 __mptcp_error_report(sk);
2093 unlock_sock_fast(ssk, slowpath);
2096 /* acquire the data lock only if some input data is pending */
2098 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2099 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2100 mptcp_data_lock(sk);
2101 __mptcp_update_rmem(sk);
2102 ret |= __mptcp_ofo_queue(msk);
2103 __mptcp_splice_receive_queue(sk);
2104 mptcp_data_unlock(sk);
2107 mptcp_check_data_fin((struct sock *)msk);
2108 return !skb_queue_empty(&msk->receive_queue);
2111 static unsigned int mptcp_inq_hint(const struct sock *sk)
2113 const struct mptcp_sock *msk = mptcp_sk(sk);
2114 const struct sk_buff *skb;
2116 skb = skb_peek(&msk->receive_queue);
2118 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2120 if (hint_val >= INT_MAX)
2123 return (unsigned int)hint_val;
2126 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2132 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2133 int flags, int *addr_len)
2135 struct mptcp_sock *msk = mptcp_sk(sk);
2136 struct scm_timestamping_internal tss;
2137 int copied = 0, cmsg_flags = 0;
2141 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2142 if (unlikely(flags & MSG_ERRQUEUE))
2143 return inet_recv_error(sk, msg, len, addr_len);
2146 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2151 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2153 len = min_t(size_t, len, INT_MAX);
2154 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2156 if (unlikely(msk->recvmsg_inq))
2157 cmsg_flags = MPTCP_CMSG_INQ;
2159 while (copied < len) {
2162 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2163 if (unlikely(bytes_read < 0)) {
2165 copied = bytes_read;
2169 copied += bytes_read;
2171 /* be sure to advertise window change */
2172 mptcp_cleanup_rbuf(msk);
2174 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2177 /* only the master socket status is relevant here. The exit
2178 * conditions mirror closely tcp_recvmsg()
2180 if (copied >= target)
2185 sk->sk_state == TCP_CLOSE ||
2186 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2188 signal_pending(current))
2192 copied = sock_error(sk);
2196 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2197 /* race breaker: the shutdown could be after the
2198 * previous receive queue check
2200 if (__mptcp_move_skbs(msk))
2205 if (sk->sk_state == TCP_CLOSE) {
2215 if (signal_pending(current)) {
2216 copied = sock_intr_errno(timeo);
2221 pr_debug("block timeout %ld", timeo);
2222 sk_wait_data(sk, &timeo, NULL);
2226 if (cmsg_flags && copied >= 0) {
2227 if (cmsg_flags & MPTCP_CMSG_TS)
2228 tcp_recv_timestamp(msg, sk, &tss);
2230 if (cmsg_flags & MPTCP_CMSG_INQ) {
2231 unsigned int inq = mptcp_inq_hint(sk);
2233 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2237 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2238 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2239 skb_queue_empty(&msk->receive_queue), copied);
2240 if (!(flags & MSG_PEEK))
2241 mptcp_rcv_space_adjust(msk, copied);
2247 static void mptcp_retransmit_timer(struct timer_list *t)
2249 struct inet_connection_sock *icsk = from_timer(icsk, t,
2250 icsk_retransmit_timer);
2251 struct sock *sk = &icsk->icsk_inet.sk;
2252 struct mptcp_sock *msk = mptcp_sk(sk);
2255 if (!sock_owned_by_user(sk)) {
2256 /* we need a process context to retransmit */
2257 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2258 mptcp_schedule_work(sk);
2260 /* delegate our work to tcp_release_cb() */
2261 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2267 static void mptcp_tout_timer(struct timer_list *t)
2269 struct sock *sk = from_timer(sk, t, sk_timer);
2271 mptcp_schedule_work(sk);
2275 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2278 * A backup subflow is returned only if that is the only kind available.
2280 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2282 struct sock *backup = NULL, *pick = NULL;
2283 struct mptcp_subflow_context *subflow;
2284 int min_stale_count = INT_MAX;
2286 mptcp_for_each_subflow(msk, subflow) {
2287 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2289 if (!__mptcp_subflow_active(subflow))
2292 /* still data outstanding at TCP level? skip this */
2293 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2294 mptcp_pm_subflow_chk_stale(msk, ssk);
2295 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2299 if (subflow->backup) {
2312 /* use backup only if there are no progresses anywhere */
2313 return min_stale_count > 1 ? backup : NULL;
2316 bool __mptcp_retransmit_pending_data(struct sock *sk)
2318 struct mptcp_data_frag *cur, *rtx_head;
2319 struct mptcp_sock *msk = mptcp_sk(sk);
2321 if (__mptcp_check_fallback(msk))
2324 /* the closing socket has some data untransmitted and/or unacked:
2325 * some data in the mptcp rtx queue has not really xmitted yet.
2326 * keep it simple and re-inject the whole mptcp level rtx queue
2328 mptcp_data_lock(sk);
2329 __mptcp_clean_una_wakeup(sk);
2330 rtx_head = mptcp_rtx_head(sk);
2332 mptcp_data_unlock(sk);
2336 msk->recovery_snd_nxt = msk->snd_nxt;
2337 msk->recovery = true;
2338 mptcp_data_unlock(sk);
2340 msk->first_pending = rtx_head;
2343 /* be sure to clear the "sent status" on all re-injected fragments */
2344 list_for_each_entry(cur, &msk->rtx_queue, list) {
2345 if (!cur->already_sent)
2347 cur->already_sent = 0;
2353 /* flags for __mptcp_close_ssk() */
2354 #define MPTCP_CF_PUSH BIT(1)
2355 #define MPTCP_CF_FASTCLOSE BIT(2)
2357 /* be sure to send a reset only if the caller asked for it, also
2358 * clean completely the subflow status when the subflow reaches
2361 static void __mptcp_subflow_disconnect(struct sock *ssk,
2362 struct mptcp_subflow_context *subflow,
2365 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2366 (flags & MPTCP_CF_FASTCLOSE)) {
2367 /* The MPTCP code never wait on the subflow sockets, TCP-level
2368 * disconnect should never fail
2370 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2371 mptcp_subflow_ctx_reset(subflow);
2373 tcp_shutdown(ssk, SEND_SHUTDOWN);
2377 /* subflow sockets can be either outgoing (connect) or incoming
2380 * Outgoing subflows use in-kernel sockets.
2381 * Incoming subflows do not have their own 'struct socket' allocated,
2382 * so we need to use tcp_close() after detaching them from the mptcp
2385 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2386 struct mptcp_subflow_context *subflow,
2389 struct mptcp_sock *msk = mptcp_sk(sk);
2390 bool dispose_it, need_push = false;
2392 /* If the first subflow moved to a close state before accept, e.g. due
2393 * to an incoming reset or listener shutdown, the subflow socket is
2394 * already deleted by inet_child_forget() and the mptcp socket can't
2397 if (msk->in_accept_queue && msk->first == ssk &&
2398 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2399 /* ensure later check in mptcp_worker() will dispose the msk */
2400 sock_set_flag(sk, SOCK_DEAD);
2401 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2402 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2403 mptcp_subflow_drop_ctx(ssk);
2407 dispose_it = msk->free_first || ssk != msk->first;
2409 list_del(&subflow->node);
2411 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2413 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2414 /* be sure to force the tcp_close path
2415 * to generate the egress reset
2417 ssk->sk_lingertime = 0;
2418 sock_set_flag(ssk, SOCK_LINGER);
2419 subflow->send_fastclose = 1;
2422 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2424 __mptcp_subflow_disconnect(ssk, subflow, flags);
2430 subflow->disposable = 1;
2432 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2433 * the ssk has been already destroyed, we just need to release the
2434 * reference owned by msk;
2436 if (!inet_csk(ssk)->icsk_ulp_ops) {
2437 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2438 kfree_rcu(subflow, rcu);
2440 /* otherwise tcp will dispose of the ssk and subflow ctx */
2441 __tcp_close(ssk, 0);
2443 /* close acquired an extra ref */
2448 __mptcp_subflow_error_report(sk, ssk);
2453 if (ssk == msk->first)
2454 WRITE_ONCE(msk->first, NULL);
2457 __mptcp_sync_sndbuf(sk);
2459 __mptcp_push_pending(sk, 0);
2461 /* Catch every 'all subflows closed' scenario, including peers silently
2462 * closing them, e.g. due to timeout.
2463 * For established sockets, allow an additional timeout before closing,
2464 * as the protocol can still create more subflows.
2466 if (list_is_singular(&msk->conn_list) && msk->first &&
2467 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2468 if (sk->sk_state != TCP_ESTABLISHED ||
2469 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2470 mptcp_set_state(sk, TCP_CLOSE);
2471 mptcp_close_wake_up(sk);
2473 mptcp_start_tout_timer(sk);
2478 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2479 struct mptcp_subflow_context *subflow)
2481 if (sk->sk_state == TCP_ESTABLISHED)
2482 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2484 /* subflow aborted before reaching the fully_established status
2485 * attempt the creation of the next subflow
2487 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2489 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2492 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2497 static void __mptcp_close_subflow(struct sock *sk)
2499 struct mptcp_subflow_context *subflow, *tmp;
2500 struct mptcp_sock *msk = mptcp_sk(sk);
2504 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2505 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2507 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2510 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2511 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2514 mptcp_close_ssk(sk, ssk, subflow);
2519 static bool mptcp_close_tout_expired(const struct sock *sk)
2521 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2522 sk->sk_state == TCP_CLOSE)
2525 return time_after32(tcp_jiffies32,
2526 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2529 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2531 struct mptcp_subflow_context *subflow, *tmp;
2532 struct sock *sk = (struct sock *)msk;
2534 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2537 mptcp_token_destroy(msk);
2539 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2540 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2543 slow = lock_sock_fast(tcp_sk);
2544 if (tcp_sk->sk_state != TCP_CLOSE) {
2545 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2546 tcp_set_state(tcp_sk, TCP_CLOSE);
2548 unlock_sock_fast(tcp_sk, slow);
2551 /* Mirror the tcp_reset() error propagation */
2552 switch (sk->sk_state) {
2554 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2556 case TCP_CLOSE_WAIT:
2557 WRITE_ONCE(sk->sk_err, EPIPE);
2562 WRITE_ONCE(sk->sk_err, ECONNRESET);
2565 mptcp_set_state(sk, TCP_CLOSE);
2566 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2567 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2568 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2570 /* the calling mptcp_worker will properly destroy the socket */
2571 if (sock_flag(sk, SOCK_DEAD))
2574 sk->sk_state_change(sk);
2575 sk_error_report(sk);
2578 static void __mptcp_retrans(struct sock *sk)
2580 struct mptcp_sock *msk = mptcp_sk(sk);
2581 struct mptcp_subflow_context *subflow;
2582 struct mptcp_sendmsg_info info = {};
2583 struct mptcp_data_frag *dfrag;
2588 mptcp_clean_una_wakeup(sk);
2590 /* first check ssk: need to kick "stale" logic */
2591 err = mptcp_sched_get_retrans(msk);
2592 dfrag = mptcp_rtx_head(sk);
2594 if (mptcp_data_fin_enabled(msk)) {
2595 struct inet_connection_sock *icsk = inet_csk(sk);
2597 icsk->icsk_retransmits++;
2598 mptcp_set_datafin_timeout(sk);
2599 mptcp_send_ack(msk);
2604 if (!mptcp_send_head(sk))
2613 mptcp_for_each_subflow(msk, subflow) {
2614 if (READ_ONCE(subflow->scheduled)) {
2617 mptcp_subflow_set_scheduled(subflow, false);
2619 ssk = mptcp_subflow_tcp_sock(subflow);
2623 /* limit retransmission to the bytes already sent on some subflows */
2625 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2626 dfrag->already_sent;
2627 while (info.sent < info.limit) {
2628 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2632 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2637 len = max(copied, len);
2638 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2640 WRITE_ONCE(msk->allow_infinite_fallback, false);
2647 msk->bytes_retrans += len;
2648 dfrag->already_sent = max(dfrag->already_sent, len);
2651 mptcp_check_and_set_pending(sk);
2653 if (!mptcp_rtx_timer_pending(sk))
2654 mptcp_reset_rtx_timer(sk);
2657 /* schedule the timeout timer for the relevant event: either close timeout
2658 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2660 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2662 struct sock *sk = (struct sock *)msk;
2663 unsigned long timeout, close_timeout;
2665 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2668 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2669 mptcp_close_timeout(sk);
2671 /* the close timeout takes precedence on the fail one, and here at least one of
2674 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2676 sk_reset_timer(sk, &sk->sk_timer, timeout);
2679 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2681 struct sock *ssk = msk->first;
2687 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2689 slow = lock_sock_fast(ssk);
2690 mptcp_subflow_reset(ssk);
2691 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2692 unlock_sock_fast(ssk, slow);
2695 static void mptcp_do_fastclose(struct sock *sk)
2697 struct mptcp_subflow_context *subflow, *tmp;
2698 struct mptcp_sock *msk = mptcp_sk(sk);
2700 mptcp_set_state(sk, TCP_CLOSE);
2701 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2702 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2703 subflow, MPTCP_CF_FASTCLOSE);
2706 static void mptcp_worker(struct work_struct *work)
2708 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2709 struct sock *sk = (struct sock *)msk;
2710 unsigned long fail_tout;
2714 state = sk->sk_state;
2715 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2718 mptcp_check_fastclose(msk);
2720 mptcp_pm_nl_work(msk);
2722 mptcp_check_send_data_fin(sk);
2723 mptcp_check_data_fin_ack(sk);
2724 mptcp_check_data_fin(sk);
2726 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2727 __mptcp_close_subflow(sk);
2729 if (mptcp_close_tout_expired(sk)) {
2730 mptcp_do_fastclose(sk);
2731 mptcp_close_wake_up(sk);
2734 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2735 __mptcp_destroy_sock(sk);
2739 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2740 __mptcp_retrans(sk);
2742 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2743 if (fail_tout && time_after(jiffies, fail_tout))
2744 mptcp_mp_fail_no_response(msk);
2751 static void __mptcp_init_sock(struct sock *sk)
2753 struct mptcp_sock *msk = mptcp_sk(sk);
2755 INIT_LIST_HEAD(&msk->conn_list);
2756 INIT_LIST_HEAD(&msk->join_list);
2757 INIT_LIST_HEAD(&msk->rtx_queue);
2758 INIT_WORK(&msk->work, mptcp_worker);
2759 __skb_queue_head_init(&msk->receive_queue);
2760 msk->out_of_order_queue = RB_ROOT;
2761 msk->first_pending = NULL;
2762 msk->rmem_fwd_alloc = 0;
2763 WRITE_ONCE(msk->rmem_released, 0);
2764 msk->timer_ival = TCP_RTO_MIN;
2765 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2767 WRITE_ONCE(msk->first, NULL);
2768 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2769 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2770 WRITE_ONCE(msk->allow_infinite_fallback, true);
2771 msk->recovery = false;
2772 msk->subflow_id = 1;
2774 mptcp_pm_data_init(msk);
2776 /* re-use the csk retrans timer for MPTCP-level retrans */
2777 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2778 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2781 static void mptcp_ca_reset(struct sock *sk)
2783 struct inet_connection_sock *icsk = inet_csk(sk);
2785 tcp_assign_congestion_control(sk);
2786 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2788 /* no need to keep a reference to the ops, the name will suffice */
2789 tcp_cleanup_congestion_control(sk);
2790 icsk->icsk_ca_ops = NULL;
2793 static int mptcp_init_sock(struct sock *sk)
2795 struct net *net = sock_net(sk);
2798 __mptcp_init_sock(sk);
2800 if (!mptcp_is_enabled(net))
2801 return -ENOPROTOOPT;
2803 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2806 ret = mptcp_init_sched(mptcp_sk(sk),
2807 mptcp_sched_find(mptcp_get_scheduler(net)));
2811 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2813 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2814 * propagate the correct value
2818 sk_sockets_allocated_inc(sk);
2819 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2820 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2825 static void __mptcp_clear_xmit(struct sock *sk)
2827 struct mptcp_sock *msk = mptcp_sk(sk);
2828 struct mptcp_data_frag *dtmp, *dfrag;
2830 WRITE_ONCE(msk->first_pending, NULL);
2831 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2832 dfrag_clear(sk, dfrag);
2835 void mptcp_cancel_work(struct sock *sk)
2837 struct mptcp_sock *msk = mptcp_sk(sk);
2839 if (cancel_work_sync(&msk->work))
2843 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2847 switch (ssk->sk_state) {
2849 if (!(how & RCV_SHUTDOWN))
2853 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2856 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2857 pr_debug("Fallback");
2858 ssk->sk_shutdown |= how;
2859 tcp_shutdown(ssk, how);
2861 /* simulate the data_fin ack reception to let the state
2862 * machine move forward
2864 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2865 mptcp_schedule_work(sk);
2867 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2869 if (!mptcp_rtx_timer_pending(sk))
2870 mptcp_reset_rtx_timer(sk);
2878 void mptcp_set_state(struct sock *sk, int state)
2880 int oldstate = sk->sk_state;
2883 case TCP_ESTABLISHED:
2884 if (oldstate != TCP_ESTABLISHED)
2885 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2889 if (oldstate == TCP_ESTABLISHED)
2890 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2893 inet_sk_state_store(sk, state);
2896 static const unsigned char new_state[16] = {
2897 /* current state: new state: action: */
2898 [0 /* (Invalid) */] = TCP_CLOSE,
2899 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2900 [TCP_SYN_SENT] = TCP_CLOSE,
2901 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2902 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2903 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2904 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2905 [TCP_CLOSE] = TCP_CLOSE,
2906 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2907 [TCP_LAST_ACK] = TCP_LAST_ACK,
2908 [TCP_LISTEN] = TCP_CLOSE,
2909 [TCP_CLOSING] = TCP_CLOSING,
2910 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2913 static int mptcp_close_state(struct sock *sk)
2915 int next = (int)new_state[sk->sk_state];
2916 int ns = next & TCP_STATE_MASK;
2918 mptcp_set_state(sk, ns);
2920 return next & TCP_ACTION_FIN;
2923 static void mptcp_check_send_data_fin(struct sock *sk)
2925 struct mptcp_subflow_context *subflow;
2926 struct mptcp_sock *msk = mptcp_sk(sk);
2928 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2929 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2930 msk->snd_nxt, msk->write_seq);
2932 /* we still need to enqueue subflows or not really shutting down,
2935 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2936 mptcp_send_head(sk))
2939 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2941 mptcp_for_each_subflow(msk, subflow) {
2942 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2944 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2948 static void __mptcp_wr_shutdown(struct sock *sk)
2950 struct mptcp_sock *msk = mptcp_sk(sk);
2952 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2953 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2954 !!mptcp_send_head(sk));
2956 /* will be ignored by fallback sockets */
2957 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2958 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2960 mptcp_check_send_data_fin(sk);
2963 static void __mptcp_destroy_sock(struct sock *sk)
2965 struct mptcp_sock *msk = mptcp_sk(sk);
2967 pr_debug("msk=%p", msk);
2971 mptcp_stop_rtx_timer(sk);
2972 sk_stop_timer(sk, &sk->sk_timer);
2974 mptcp_release_sched(msk);
2976 sk->sk_prot->destroy(sk);
2978 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2979 WARN_ON_ONCE(msk->rmem_released);
2980 sk_stream_kill_queues(sk);
2981 xfrm_sk_free_policy(sk);
2986 void __mptcp_unaccepted_force_close(struct sock *sk)
2988 sock_set_flag(sk, SOCK_DEAD);
2989 mptcp_do_fastclose(sk);
2990 __mptcp_destroy_sock(sk);
2993 static __poll_t mptcp_check_readable(struct sock *sk)
2995 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
2998 static void mptcp_check_listen_stop(struct sock *sk)
3002 if (inet_sk_state_load(sk) != TCP_LISTEN)
3005 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3006 ssk = mptcp_sk(sk)->first;
3007 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3010 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3011 tcp_set_state(ssk, TCP_CLOSE);
3012 mptcp_subflow_queue_clean(sk, ssk);
3013 inet_csk_listen_stop(ssk);
3014 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3018 bool __mptcp_close(struct sock *sk, long timeout)
3020 struct mptcp_subflow_context *subflow;
3021 struct mptcp_sock *msk = mptcp_sk(sk);
3022 bool do_cancel_work = false;
3023 int subflows_alive = 0;
3025 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3027 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3028 mptcp_check_listen_stop(sk);
3029 mptcp_set_state(sk, TCP_CLOSE);
3033 if (mptcp_data_avail(msk) || timeout < 0) {
3034 /* If the msk has read data, or the caller explicitly ask it,
3035 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3037 mptcp_do_fastclose(sk);
3039 } else if (mptcp_close_state(sk)) {
3040 __mptcp_wr_shutdown(sk);
3043 sk_stream_wait_close(sk, timeout);
3046 /* orphan all the subflows */
3047 mptcp_for_each_subflow(msk, subflow) {
3048 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3049 bool slow = lock_sock_fast_nested(ssk);
3051 subflows_alive += ssk->sk_state != TCP_CLOSE;
3053 /* since the close timeout takes precedence on the fail one,
3056 if (ssk == msk->first)
3057 subflow->fail_tout = 0;
3059 /* detach from the parent socket, but allow data_ready to
3060 * push incoming data into the mptcp stack, to properly ack it
3062 ssk->sk_socket = NULL;
3064 unlock_sock_fast(ssk, slow);
3068 /* all the subflows are closed, only timeout can change the msk
3069 * state, let's not keep resources busy for no reasons
3071 if (subflows_alive == 0)
3072 mptcp_set_state(sk, TCP_CLOSE);
3075 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3077 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3079 if (sk->sk_state == TCP_CLOSE) {
3080 __mptcp_destroy_sock(sk);
3081 do_cancel_work = true;
3083 mptcp_start_tout_timer(sk);
3086 return do_cancel_work;
3089 static void mptcp_close(struct sock *sk, long timeout)
3091 bool do_cancel_work;
3095 do_cancel_work = __mptcp_close(sk, timeout);
3098 mptcp_cancel_work(sk);
3103 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3105 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3106 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3107 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3109 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3110 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3113 msk6->saddr = ssk6->saddr;
3114 msk6->flow_label = ssk6->flow_label;
3118 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3119 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3120 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3121 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3122 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3123 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3126 static int mptcp_disconnect(struct sock *sk, int flags)
3128 struct mptcp_sock *msk = mptcp_sk(sk);
3130 /* We are on the fastopen error path. We can't call straight into the
3131 * subflows cleanup code due to lock nesting (we are already under
3132 * msk->firstsocket lock).
3134 if (msk->fastopening)
3137 mptcp_check_listen_stop(sk);
3138 mptcp_set_state(sk, TCP_CLOSE);
3140 mptcp_stop_rtx_timer(sk);
3141 mptcp_stop_tout_timer(sk);
3144 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3146 /* msk->subflow is still intact, the following will not free the first
3149 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3150 WRITE_ONCE(msk->flags, 0);
3152 msk->recovery = false;
3153 msk->can_ack = false;
3154 msk->fully_established = false;
3155 msk->rcv_data_fin = false;
3156 msk->snd_data_fin_enable = false;
3157 msk->rcv_fastclose = false;
3158 msk->use_64bit_ack = false;
3159 msk->bytes_consumed = 0;
3160 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3161 mptcp_pm_data_reset(msk);
3163 msk->bytes_acked = 0;
3164 msk->bytes_received = 0;
3165 msk->bytes_sent = 0;
3166 msk->bytes_retrans = 0;
3167 msk->rcvspace_init = 0;
3169 WRITE_ONCE(sk->sk_shutdown, 0);
3170 sk_error_report(sk);
3174 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3175 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3177 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3179 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3182 static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3184 const struct ipv6_pinfo *np = inet6_sk(sk);
3185 struct ipv6_txoptions *opt;
3186 struct ipv6_pinfo *newnp;
3188 newnp = inet6_sk(newsk);
3191 opt = rcu_dereference(np->opt);
3193 opt = ipv6_dup_options(newsk, opt);
3195 net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3197 RCU_INIT_POINTER(newnp->opt, opt);
3202 static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3204 struct ip_options_rcu *inet_opt, *newopt = NULL;
3205 const struct inet_sock *inet = inet_sk(sk);
3206 struct inet_sock *newinet;
3208 newinet = inet_sk(newsk);
3211 inet_opt = rcu_dereference(inet->inet_opt);
3213 newopt = sock_kmalloc(newsk, sizeof(*inet_opt) +
3214 inet_opt->opt.optlen, GFP_ATOMIC);
3216 memcpy(newopt, inet_opt, sizeof(*inet_opt) +
3217 inet_opt->opt.optlen);
3219 net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3221 RCU_INIT_POINTER(newinet->inet_opt, newopt);
3225 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3226 const struct mptcp_options_received *mp_opt,
3228 struct request_sock *req)
3230 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3231 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3232 struct mptcp_subflow_context *subflow;
3233 struct mptcp_sock *msk;
3238 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3239 if (nsk->sk_family == AF_INET6)
3240 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3243 __mptcp_init_sock(nsk);
3245 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3246 if (nsk->sk_family == AF_INET6)
3247 mptcp_copy_ip6_options(nsk, sk);
3250 mptcp_copy_ip_options(nsk, sk);
3252 msk = mptcp_sk(nsk);
3253 msk->local_key = subflow_req->local_key;
3254 msk->token = subflow_req->token;
3255 msk->in_accept_queue = 1;
3256 WRITE_ONCE(msk->fully_established, false);
3257 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3258 WRITE_ONCE(msk->csum_enabled, true);
3260 msk->write_seq = subflow_req->idsn + 1;
3261 msk->snd_nxt = msk->write_seq;
3262 msk->snd_una = msk->write_seq;
3263 msk->wnd_end = msk->snd_nxt + tcp_sk(ssk)->snd_wnd;
3264 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3265 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3267 /* passive msk is created after the first/MPC subflow */
3268 msk->subflow_id = 2;
3270 sock_reset_flag(nsk, SOCK_RCU_FREE);
3271 security_inet_csk_clone(nsk, req);
3273 /* this can't race with mptcp_close(), as the msk is
3274 * not yet exposted to user-space
3276 mptcp_set_state(nsk, TCP_ESTABLISHED);
3278 /* The msk maintain a ref to each subflow in the connections list */
3279 WRITE_ONCE(msk->first, ssk);
3280 subflow = mptcp_subflow_ctx(ssk);
3281 list_add(&subflow->node, &msk->conn_list);
3284 /* new mpc subflow takes ownership of the newly
3285 * created mptcp socket
3287 mptcp_token_accept(subflow_req, msk);
3289 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3290 * uses the correct data
3292 mptcp_copy_inaddrs(nsk, ssk);
3293 __mptcp_propagate_sndbuf(nsk, ssk);
3295 mptcp_rcv_space_init(msk, ssk);
3297 if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3298 __mptcp_subflow_fully_established(msk, subflow, mp_opt);
3299 bh_unlock_sock(nsk);
3301 /* note: the newly allocated socket refcount is 2 now */
3305 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3307 const struct tcp_sock *tp = tcp_sk(ssk);
3309 msk->rcvspace_init = 1;
3310 msk->rcvq_space.copied = 0;
3311 msk->rcvq_space.rtt_us = 0;
3313 msk->rcvq_space.time = tp->tcp_mstamp;
3315 /* initial rcv_space offering made to peer */
3316 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3317 TCP_INIT_CWND * tp->advmss);
3318 if (msk->rcvq_space.space == 0)
3319 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3322 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3324 struct mptcp_subflow_context *subflow, *tmp;
3325 struct sock *sk = (struct sock *)msk;
3327 __mptcp_clear_xmit(sk);
3329 /* join list will be eventually flushed (with rst) at sock lock release time */
3330 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3331 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3333 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3334 mptcp_data_lock(sk);
3335 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3336 __skb_queue_purge(&sk->sk_receive_queue);
3337 skb_rbtree_purge(&msk->out_of_order_queue);
3338 mptcp_data_unlock(sk);
3340 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3341 * inet_sock_destruct() will dispose it
3343 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3344 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3345 mptcp_token_destroy(msk);
3346 mptcp_pm_free_anno_list(msk);
3347 mptcp_free_local_addr_list(msk);
3350 static void mptcp_destroy(struct sock *sk)
3352 struct mptcp_sock *msk = mptcp_sk(sk);
3354 /* allow the following to close even the initial subflow */
3355 msk->free_first = 1;
3356 mptcp_destroy_common(msk, 0);
3357 sk_sockets_allocated_dec(sk);
3360 void __mptcp_data_acked(struct sock *sk)
3362 if (!sock_owned_by_user(sk))
3363 __mptcp_clean_una(sk);
3365 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3367 if (mptcp_pending_data_fin_ack(sk))
3368 mptcp_schedule_work(sk);
3371 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3373 if (!mptcp_send_head(sk))
3376 if (!sock_owned_by_user(sk))
3377 __mptcp_subflow_push_pending(sk, ssk, false);
3379 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3382 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3383 BIT(MPTCP_RETRANSMIT) | \
3384 BIT(MPTCP_FLUSH_JOIN_LIST))
3386 /* processes deferred events and flush wmem */
3387 static void mptcp_release_cb(struct sock *sk)
3388 __must_hold(&sk->sk_lock.slock)
3390 struct mptcp_sock *msk = mptcp_sk(sk);
3393 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3394 struct list_head join_list;
3399 INIT_LIST_HEAD(&join_list);
3400 list_splice_init(&msk->join_list, &join_list);
3402 /* the following actions acquire the subflow socket lock
3404 * 1) can't be invoked in atomic scope
3405 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3406 * datapath acquires the msk socket spinlock while helding
3407 * the subflow socket lock
3409 msk->cb_flags &= ~flags;
3410 spin_unlock_bh(&sk->sk_lock.slock);
3412 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3413 __mptcp_flush_join_list(sk, &join_list);
3414 if (flags & BIT(MPTCP_PUSH_PENDING))
3415 __mptcp_push_pending(sk, 0);
3416 if (flags & BIT(MPTCP_RETRANSMIT))
3417 __mptcp_retrans(sk);
3420 spin_lock_bh(&sk->sk_lock.slock);
3423 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3424 __mptcp_clean_una_wakeup(sk);
3425 if (unlikely(msk->cb_flags)) {
3426 /* be sure to sync the msk state before taking actions
3427 * depending on sk_state (MPTCP_ERROR_REPORT)
3428 * On sk release avoid actions depending on the first subflow
3430 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3431 __mptcp_sync_state(sk, msk->pending_state);
3432 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3433 __mptcp_error_report(sk);
3434 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3435 __mptcp_sync_sndbuf(sk);
3438 __mptcp_update_rmem(sk);
3441 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3442 * TCP can't schedule delack timer before the subflow is fully established.
3443 * MPTCP uses the delack timer to do 3rd ack retransmissions
3445 static void schedule_3rdack_retransmission(struct sock *ssk)
3447 struct inet_connection_sock *icsk = inet_csk(ssk);
3448 struct tcp_sock *tp = tcp_sk(ssk);
3449 unsigned long timeout;
3451 if (mptcp_subflow_ctx(ssk)->fully_established)
3454 /* reschedule with a timeout above RTT, as we must look only for drop */
3456 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3458 timeout = TCP_TIMEOUT_INIT;
3461 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3462 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3463 icsk->icsk_ack.timeout = timeout;
3464 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3467 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3469 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3470 struct sock *sk = subflow->conn;
3472 if (status & BIT(MPTCP_DELEGATE_SEND)) {
3473 mptcp_data_lock(sk);
3474 if (!sock_owned_by_user(sk))
3475 __mptcp_subflow_push_pending(sk, ssk, true);
3477 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3478 mptcp_data_unlock(sk);
3480 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3481 mptcp_data_lock(sk);
3482 if (!sock_owned_by_user(sk))
3483 __mptcp_sync_sndbuf(sk);
3485 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3486 mptcp_data_unlock(sk);
3488 if (status & BIT(MPTCP_DELEGATE_ACK))
3489 schedule_3rdack_retransmission(ssk);
3492 static int mptcp_hash(struct sock *sk)
3494 /* should never be called,
3495 * we hash the TCP subflows not the master socket
3501 static void mptcp_unhash(struct sock *sk)
3503 /* called from sk_common_release(), but nothing to do here */
3506 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3508 struct mptcp_sock *msk = mptcp_sk(sk);
3510 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3511 if (WARN_ON_ONCE(!msk->first))
3514 return inet_csk_get_port(msk->first, snum);
3517 void mptcp_finish_connect(struct sock *ssk)
3519 struct mptcp_subflow_context *subflow;
3520 struct mptcp_sock *msk;
3523 subflow = mptcp_subflow_ctx(ssk);
3527 pr_debug("msk=%p, token=%u", sk, subflow->token);
3529 subflow->map_seq = subflow->iasn;
3530 subflow->map_subflow_seq = 1;
3532 /* the socket is not connected yet, no msk/subflow ops can access/race
3533 * accessing the field below
3535 WRITE_ONCE(msk->local_key, subflow->local_key);
3537 mptcp_pm_new_connection(msk, ssk, 0);
3540 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3542 write_lock_bh(&sk->sk_callback_lock);
3543 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3544 sk_set_socket(sk, parent);
3545 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3546 write_unlock_bh(&sk->sk_callback_lock);
3549 bool mptcp_finish_join(struct sock *ssk)
3551 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3552 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3553 struct sock *parent = (void *)msk;
3556 pr_debug("msk=%p, subflow=%p", msk, subflow);
3558 /* mptcp socket already closing? */
3559 if (!mptcp_is_fully_established(parent)) {
3560 subflow->reset_reason = MPTCP_RST_EMPTCP;
3564 /* active subflow, already present inside the conn_list */
3565 if (!list_empty(&subflow->node)) {
3566 mptcp_subflow_joined(msk, ssk);
3567 mptcp_propagate_sndbuf(parent, ssk);
3571 if (!mptcp_pm_allow_new_subflow(msk))
3572 goto err_prohibited;
3574 /* If we can't acquire msk socket lock here, let the release callback
3577 mptcp_data_lock(parent);
3578 if (!sock_owned_by_user(parent)) {
3579 ret = __mptcp_finish_join(msk, ssk);
3582 list_add_tail(&subflow->node, &msk->conn_list);
3586 list_add_tail(&subflow->node, &msk->join_list);
3587 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3589 mptcp_data_unlock(parent);
3593 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3600 static void mptcp_shutdown(struct sock *sk, int how)
3602 pr_debug("sk=%p, how=%d", sk, how);
3604 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3605 __mptcp_wr_shutdown(sk);
3608 static int mptcp_forward_alloc_get(const struct sock *sk)
3610 return READ_ONCE(sk->sk_forward_alloc) +
3611 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3614 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3616 const struct sock *sk = (void *)msk;
3619 if (sk->sk_state == TCP_LISTEN)
3622 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3625 delta = msk->write_seq - v;
3626 if (__mptcp_check_fallback(msk) && msk->first) {
3627 struct tcp_sock *tp = tcp_sk(msk->first);
3629 /* the first subflow is disconnected after close - see
3630 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3631 * so ignore that status, too.
3633 if (!((1 << msk->first->sk_state) &
3634 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3635 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3637 if (delta > INT_MAX)
3643 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3645 struct mptcp_sock *msk = mptcp_sk(sk);
3650 if (sk->sk_state == TCP_LISTEN)
3654 __mptcp_move_skbs(msk);
3655 *karg = mptcp_inq_hint(sk);
3659 slow = lock_sock_fast(sk);
3660 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3661 unlock_sock_fast(sk, slow);
3664 slow = lock_sock_fast(sk);
3665 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3666 unlock_sock_fast(sk, slow);
3669 return -ENOIOCTLCMD;
3675 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3676 struct mptcp_subflow_context *subflow)
3678 subflow->request_mptcp = 0;
3679 __mptcp_do_fallback(msk);
3682 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3684 struct mptcp_subflow_context *subflow;
3685 struct mptcp_sock *msk = mptcp_sk(sk);
3689 ssk = __mptcp_nmpc_sk(msk);
3691 return PTR_ERR(ssk);
3693 mptcp_set_state(sk, TCP_SYN_SENT);
3694 subflow = mptcp_subflow_ctx(ssk);
3695 #ifdef CONFIG_TCP_MD5SIG
3696 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3699 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3700 mptcp_subflow_early_fallback(msk, subflow);
3702 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3703 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3704 mptcp_subflow_early_fallback(msk, subflow);
3706 if (likely(!__mptcp_check_fallback(msk)))
3707 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3709 /* if reaching here via the fastopen/sendmsg path, the caller already
3710 * acquired the subflow socket lock, too.
3712 if (!msk->fastopening)
3715 /* the following mirrors closely a very small chunk of code from
3716 * __inet_stream_connect()
3718 if (ssk->sk_state != TCP_CLOSE)
3721 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3722 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3727 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3731 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3734 if (!msk->fastopening)
3737 /* on successful connect, the msk state will be moved to established by
3738 * subflow_finish_connect()
3740 if (unlikely(err)) {
3741 /* avoid leaving a dangling token in an unconnected socket */
3742 mptcp_token_destroy(msk);
3743 mptcp_set_state(sk, TCP_CLOSE);
3747 mptcp_copy_inaddrs(sk, ssk);
3751 static struct proto mptcp_prot = {
3753 .owner = THIS_MODULE,
3754 .init = mptcp_init_sock,
3755 .connect = mptcp_connect,
3756 .disconnect = mptcp_disconnect,
3757 .close = mptcp_close,
3758 .setsockopt = mptcp_setsockopt,
3759 .getsockopt = mptcp_getsockopt,
3760 .shutdown = mptcp_shutdown,
3761 .destroy = mptcp_destroy,
3762 .sendmsg = mptcp_sendmsg,
3763 .ioctl = mptcp_ioctl,
3764 .recvmsg = mptcp_recvmsg,
3765 .release_cb = mptcp_release_cb,
3767 .unhash = mptcp_unhash,
3768 .get_port = mptcp_get_port,
3769 .forward_alloc_get = mptcp_forward_alloc_get,
3770 .sockets_allocated = &mptcp_sockets_allocated,
3772 .memory_allocated = &tcp_memory_allocated,
3773 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3775 .memory_pressure = &tcp_memory_pressure,
3776 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3777 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3778 .sysctl_mem = sysctl_tcp_mem,
3779 .obj_size = sizeof(struct mptcp_sock),
3780 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3781 .no_autobind = true,
3784 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3786 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3787 struct sock *ssk, *sk = sock->sk;
3791 ssk = __mptcp_nmpc_sk(msk);
3797 if (sk->sk_family == AF_INET)
3798 err = inet_bind_sk(ssk, uaddr, addr_len);
3799 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3800 else if (sk->sk_family == AF_INET6)
3801 err = inet6_bind_sk(ssk, uaddr, addr_len);
3804 mptcp_copy_inaddrs(sk, ssk);
3811 static int mptcp_listen(struct socket *sock, int backlog)
3813 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3814 struct sock *sk = sock->sk;
3818 pr_debug("msk=%p", msk);
3823 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3826 ssk = __mptcp_nmpc_sk(msk);
3832 mptcp_set_state(sk, TCP_LISTEN);
3833 sock_set_flag(sk, SOCK_RCU_FREE);
3836 err = __inet_listen_sk(ssk, backlog);
3838 mptcp_set_state(sk, inet_sk_state_load(ssk));
3841 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3842 mptcp_copy_inaddrs(sk, ssk);
3843 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3851 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3852 int flags, bool kern)
3854 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3855 struct sock *ssk, *newsk;
3858 pr_debug("msk=%p", msk);
3860 /* Buggy applications can call accept on socket states other then LISTEN
3861 * but no need to allocate the first subflow just to error out.
3863 ssk = READ_ONCE(msk->first);
3867 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3868 newsk = inet_csk_accept(ssk, flags, &err, kern);
3872 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3873 if (sk_is_mptcp(newsk)) {
3874 struct mptcp_subflow_context *subflow;
3875 struct sock *new_mptcp_sock;
3877 subflow = mptcp_subflow_ctx(newsk);
3878 new_mptcp_sock = subflow->conn;
3880 /* is_mptcp should be false if subflow->conn is missing, see
3881 * subflow_syn_recv_sock()
3883 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3884 tcp_sk(newsk)->is_mptcp = 0;
3888 newsk = new_mptcp_sock;
3889 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3891 newsk->sk_kern_sock = kern;
3893 __inet_accept(sock, newsock, newsk);
3895 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3896 msk = mptcp_sk(newsk);
3897 msk->in_accept_queue = 0;
3899 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3900 * This is needed so NOSPACE flag can be set from tcp stack.
3902 mptcp_for_each_subflow(msk, subflow) {
3903 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3905 if (!ssk->sk_socket)
3906 mptcp_sock_graft(ssk, newsock);
3909 /* Do late cleanup for the first subflow as necessary. Also
3910 * deal with bad peers not doing a complete shutdown.
3912 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3913 __mptcp_close_ssk(newsk, msk->first,
3914 mptcp_subflow_ctx(msk->first), 0);
3915 if (unlikely(list_is_singular(&msk->conn_list)))
3916 mptcp_set_state(newsk, TCP_CLOSE);
3920 newsk->sk_kern_sock = kern;
3922 __inet_accept(sock, newsock, newsk);
3923 /* we are being invoked after accepting a non-mp-capable
3924 * flow: sk is a tcp_sk, not an mptcp one.
3926 * Hand the socket over to tcp so all further socket ops
3929 WRITE_ONCE(newsock->sk->sk_socket->ops,
3930 mptcp_fallback_tcp_ops(newsock->sk));
3932 release_sock(newsk);
3937 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3939 struct sock *sk = (struct sock *)msk;
3941 if (sk_stream_is_writeable(sk))
3942 return EPOLLOUT | EPOLLWRNORM;
3944 mptcp_set_nospace(sk);
3945 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3946 if (sk_stream_is_writeable(sk))
3947 return EPOLLOUT | EPOLLWRNORM;
3952 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3953 struct poll_table_struct *wait)
3955 struct sock *sk = sock->sk;
3956 struct mptcp_sock *msk;
3962 sock_poll_wait(file, sock, wait);
3964 state = inet_sk_state_load(sk);
3965 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3966 if (state == TCP_LISTEN) {
3967 struct sock *ssk = READ_ONCE(msk->first);
3969 if (WARN_ON_ONCE(!ssk))
3972 return inet_csk_listen_poll(ssk);
3975 shutdown = READ_ONCE(sk->sk_shutdown);
3976 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3978 if (shutdown & RCV_SHUTDOWN)
3979 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3981 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3982 mask |= mptcp_check_readable(sk);
3983 if (shutdown & SEND_SHUTDOWN)
3984 mask |= EPOLLOUT | EPOLLWRNORM;
3986 mask |= mptcp_check_writeable(msk);
3987 } else if (state == TCP_SYN_SENT &&
3988 inet_test_bit(DEFER_CONNECT, sk)) {
3989 /* cf tcp_poll() note about TFO */
3990 mask |= EPOLLOUT | EPOLLWRNORM;
3993 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3995 if (READ_ONCE(sk->sk_err))
4001 static const struct proto_ops mptcp_stream_ops = {
4003 .owner = THIS_MODULE,
4004 .release = inet_release,
4006 .connect = inet_stream_connect,
4007 .socketpair = sock_no_socketpair,
4008 .accept = mptcp_stream_accept,
4009 .getname = inet_getname,
4011 .ioctl = inet_ioctl,
4012 .gettstamp = sock_gettstamp,
4013 .listen = mptcp_listen,
4014 .shutdown = inet_shutdown,
4015 .setsockopt = sock_common_setsockopt,
4016 .getsockopt = sock_common_getsockopt,
4017 .sendmsg = inet_sendmsg,
4018 .recvmsg = inet_recvmsg,
4019 .mmap = sock_no_mmap,
4020 .set_rcvlowat = mptcp_set_rcvlowat,
4023 static struct inet_protosw mptcp_protosw = {
4024 .type = SOCK_STREAM,
4025 .protocol = IPPROTO_MPTCP,
4026 .prot = &mptcp_prot,
4027 .ops = &mptcp_stream_ops,
4028 .flags = INET_PROTOSW_ICSK,
4031 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4033 struct mptcp_delegated_action *delegated;
4034 struct mptcp_subflow_context *subflow;
4037 delegated = container_of(napi, struct mptcp_delegated_action, napi);
4038 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4039 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4041 bh_lock_sock_nested(ssk);
4042 if (!sock_owned_by_user(ssk)) {
4043 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4045 /* tcp_release_cb_override already processed
4046 * the action or will do at next release_sock().
4047 * In both case must dequeue the subflow here - on the same
4048 * CPU that scheduled it.
4051 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4053 bh_unlock_sock(ssk);
4056 if (++work_done == budget)
4060 /* always provide a 0 'work_done' argument, so that napi_complete_done
4061 * will not try accessing the NULL napi->dev ptr
4063 napi_complete_done(napi, 0);
4067 void __init mptcp_proto_init(void)
4069 struct mptcp_delegated_action *delegated;
4072 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4074 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4075 panic("Failed to allocate MPTCP pcpu counter\n");
4077 init_dummy_netdev(&mptcp_napi_dev);
4078 for_each_possible_cpu(cpu) {
4079 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4080 INIT_LIST_HEAD(&delegated->head);
4081 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4083 napi_enable(&delegated->napi);
4086 mptcp_subflow_init();
4091 if (proto_register(&mptcp_prot, 1) != 0)
4092 panic("Failed to register MPTCP proto.\n");
4094 inet_register_protosw(&mptcp_protosw);
4096 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4099 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4100 static const struct proto_ops mptcp_v6_stream_ops = {
4102 .owner = THIS_MODULE,
4103 .release = inet6_release,
4105 .connect = inet_stream_connect,
4106 .socketpair = sock_no_socketpair,
4107 .accept = mptcp_stream_accept,
4108 .getname = inet6_getname,
4110 .ioctl = inet6_ioctl,
4111 .gettstamp = sock_gettstamp,
4112 .listen = mptcp_listen,
4113 .shutdown = inet_shutdown,
4114 .setsockopt = sock_common_setsockopt,
4115 .getsockopt = sock_common_getsockopt,
4116 .sendmsg = inet6_sendmsg,
4117 .recvmsg = inet6_recvmsg,
4118 .mmap = sock_no_mmap,
4119 #ifdef CONFIG_COMPAT
4120 .compat_ioctl = inet6_compat_ioctl,
4122 .set_rcvlowat = mptcp_set_rcvlowat,
4125 static struct proto mptcp_v6_prot;
4127 static struct inet_protosw mptcp_v6_protosw = {
4128 .type = SOCK_STREAM,
4129 .protocol = IPPROTO_MPTCP,
4130 .prot = &mptcp_v6_prot,
4131 .ops = &mptcp_v6_stream_ops,
4132 .flags = INET_PROTOSW_ICSK,
4135 int __init mptcp_proto_v6_init(void)
4139 mptcp_v6_prot = mptcp_prot;
4140 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4141 mptcp_v6_prot.slab = NULL;
4142 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4143 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4145 err = proto_register(&mptcp_v6_prot, 1);
4149 err = inet6_register_protosw(&mptcp_v6_protosw);
4151 proto_unregister(&mptcp_v6_prot);