linux - net/mptcp/protocol.c

4895 lines · 124.2 KiB

1	// SPDX-License-Identifier: GPL-2.0
2	/* Multipath TCP
3	*
4	* Copyright (c) 2017 - 2019, Intel Corporation.
5	*/
6
7	#define pr_fmt(fmt) "MPTCP: " fmt
8
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>
14	#include <net/aligned_data.h>
15	#include <net/rps.h>
16	#include <net/sock.h>
17	#include <net/inet_common.h>
18	#include <net/inet_hashtables.h>
19	#include <net/protocol.h>
20	#include <net/tcp_states.h>
21	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
22	#include <net/transp_v6.h>
23	#endif
24	#include <net/mptcp.h>
25	#include <net/hotdata.h>
26	#include <net/xfrm.h>
27	#include <asm/ioctls.h>
28	#include "protocol.h"
29	#include "mib.h"
30
31	static unsigned int mptcp_inq_hint(const struct sock *sk);
32
33	#define CREATE_TRACE_POINTS
34	#include <trace/events/mptcp.h>
35
36	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
37	struct mptcp6_sock {
38	struct mptcp_sock msk;
39	struct ipv6_pinfo np;
40	};
41	#endif
42
43	enum {
44	MPTCP_CMSG_TS = BIT(0),
45	MPTCP_CMSG_INQ = BIT(1),
46	};
47
48	static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
49
50	static void __mptcp_destroy_sock(struct sock *sk);
51	static void mptcp_check_send_data_fin(struct sock *sk);
52
53	DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions) = {
54	.bh_lock = INIT_LOCAL_LOCK(bh_lock),
55	};
56	static struct net_device *mptcp_napi_dev;
57
58	/* Returns end sequence number of the receiver's advertised window */
59	static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
60	{
61	return READ_ONCE(msk->wnd_end);
62	}
63
64	static const struct proto_ops mptcp_fallback_tcp_ops(const struct sock sk)
65	{
66	unsigned short family = READ_ONCE(sk->sk_family);
67
68	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
69	if (family == AF_INET6)
70	return &inet6_stream_ops;
71	#endif
72	WARN_ON_ONCE(family != AF_INET);
73	return &inet_stream_ops;
74	}
75
76	bool __mptcp_try_fallback(struct mptcp_sock *msk, int fb_mib)
77	{
78	struct net net = sock_net((struct sock )msk);
79
80	if (__mptcp_check_fallback(msk))
81	return true;
82
83	/* The caller possibly is not holding the msk socket lock, but
84	* in the fallback case only the current subflow is touching
85	* the OoO queue.
86	*/
87	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue))
88	return false;
89
90	spin_lock_bh(&msk->fallback_lock);
91	if (!msk->allow_infinite_fallback) {
92	spin_unlock_bh(&msk->fallback_lock);
93	return false;
94	}
95
96	msk->allow_subflows = false;
97	set_bit(MPTCP_FALLBACK_DONE, &msk->flags);
98	clear_bit(MPTCP_RTX_ENABLED, &msk->flags);
99	__MPTCP_INC_STATS(net, fb_mib);
100	spin_unlock_bh(&msk->fallback_lock);
101	return true;
102	}
103
104	static int __mptcp_socket_create(struct mptcp_sock *msk)
105	{
106	struct mptcp_subflow_context *subflow;
107	struct sock sk = (struct sock )msk;
108	struct socket *ssock;
109	int err;
110
111	err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
112	if (err)
113	return err;
114
115	msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
116	WRITE_ONCE(msk->first, ssock->sk);
117	subflow = mptcp_subflow_ctx(ssock->sk);
118	list_add(&subflow->node, &msk->conn_list);
119	sock_hold(ssock->sk);
120	subflow->request_mptcp = 1;
121	subflow->subflow_id = msk->subflow_id++;
122
123	/* This is the first subflow, always with id 0 */
124	WRITE_ONCE(subflow->local_id, 0);
125	mptcp_sock_graft(msk->first, sk->sk_socket);
126	iput(SOCK_INODE(ssock));
127
128	return 0;
129	}
130
131	/* If the MPC handshake is not started, returns the first subflow,
132	* eventually allocating it.
133	*/
134	struct sock __mptcp_nmpc_sk(struct mptcp_sock msk)
135	{
136	struct sock sk = (struct sock )msk;
137	int ret;
138
139	if (!((1 << sk->sk_state) & (TCPF_CLOSE \| TCPF_LISTEN)))
140	return ERR_PTR(-EINVAL);
141
142	if (!msk->first) {
143	ret = __mptcp_socket_create(msk);
144	if (ret)
145	return ERR_PTR(ret);
146	}
147
148	return msk->first;
149	}
150
151	static void mptcp_drop(struct sock sk, struct sk_buff skb)
152	{
153	/* The skb forward memory was already transferred to sk by
154	* mptcp_borrow_fwdmem(), even before setting the destructor.
155	*/
156	if (!skb->destructor)
157	sk_mem_reclaim(sk);
158
159	sk_drops_skbadd(sk, skb);
160	__kfree_skb(skb);
161	}
162
163	static bool __mptcp_try_coalesce(struct sock sk, struct sk_buff to,
164	struct sk_buff from, bool fragstolen,
165	int *delta)
166	{
167	int limit = READ_ONCE(sk->sk_rcvbuf);
168
169	if (unlikely(MPTCP_SKB_CB(to)->cant_coalesce) \|\|
170	MPTCP_SKB_CB(from)->offset \|\|
171	((to->len + from->len) > (limit >> 3)) \|\|
172	!skb_try_coalesce(to, from, fragstolen, delta))
173	return false;
174
175	pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n",
176	MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
177	to->len, MPTCP_SKB_CB(from)->end_seq);
178	MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
179	return true;
180	}
181
182	static bool mptcp_try_coalesce(struct sock sk, struct sk_buff to,
183	struct sk_buff *from)
184	{
185	bool fragstolen;
186	int delta;
187
188	if (!__mptcp_try_coalesce(sk, to, from, &fragstolen, &delta))
189	return false;
190
191	/* note the fwd memory can reach a negative value after accounting
192	* for the delta, but the later skb free will restore a non
193	* negative one
194	*/
195	atomic_add(delta, &sk->sk_rmem_alloc);
196	sk_mem_charge(sk, delta);
197	kfree_skb_partial(from, fragstolen);
198
199	return true;
200	}
201
202	static bool mptcp_ooo_try_coalesce(struct mptcp_sock msk, struct sk_buff to,
203	struct sk_buff *from)
204	{
205	if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
206	return false;
207
208	return mptcp_try_coalesce((struct sock *)msk, to, from);
209	}
210
211	/* "inspired" by tcp_rcvbuf_grow(), main difference:
212	* - mptcp does not maintain a msk-level window clamp
213	* - returns true when the receive buffer is actually updated
214	*/
215	static bool mptcp_rcvbuf_grow(struct sock *sk, u32 newval)
216	{
217	struct mptcp_sock *msk = mptcp_sk(sk);
218	const struct net *net = sock_net(sk);
219	u32 rcvwin, rcvbuf, cap, oldval;
220	u64 grow;
221
222	oldval = msk->rcvq_space.space;
223	msk->rcvq_space.space = newval;
224	if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) \|\|
225	(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
226	return false;
227
228	/* DRS is always one RTT late. */
229	rcvwin = newval << 1;
230
231	/* slow start: allow the sender to double its rate. */
232	grow = (u64)rcvwin * (newval - oldval);
233	do_div(grow, oldval);
234	rcvwin += grow << 1;
235
236	cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
237
238	rcvbuf = min_t(u32, mptcp_space_from_win(sk, rcvwin), cap);
239	if (rcvbuf > sk->sk_rcvbuf) {
240	WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
241	return true;
242	}
243	return false;
244	}
245
246	/* "Inspired" from the TCP version; main difference: stop as soon as the MPTCP
247	* socket is under memory limit.
248	*/
249	static void mptcp_prune_ofo_queue(struct sock *sk,
250	const struct sk_buff *in_skb)
251	{
252	struct mptcp_sock *msk = mptcp_sk(sk);
253	struct rb_node node, prev;
254	bool pruned = false;
255	u64 mem;
256
257	if (RB_EMPTY_ROOT(&msk->out_of_order_queue))
258	return;
259
260	node = &msk->ooo_last_skb->rbnode;
261
262	do {
263	struct sk_buff *skb = rb_to_skb(node);
264
265	/* Stop pruning if the incoming skb would land in OoO tail. */
266	if (after64(MPTCP_SKB_CB(in_skb)->map_seq,
267	MPTCP_SKB_CB(skb)->map_seq))
268	break;
269
270	pruned = true;
271	prev = rb_prev(node);
272	rb_erase(node, &msk->out_of_order_queue);
273	mptcp_drop(sk, skb);
274	msk->ooo_last_skb = rb_to_skb(prev);
275
276	mem = (unsigned int)sk_rmem_alloc_get(sk);
277	if (mem <= sk->sk_rcvbuf)
278	break;
279
280	node = prev;
281	} while (node);
282
283	if (pruned)
284	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOPRUNED);
285	}
286
287	/* The stack can't drop packets for fallback socket at the msk level, or the
288	* stream will break.
289	*/
290	static bool mptcp_can_ingest(const struct sock *sk)
291	{
292	return likely(sk_rmem_alloc_get(sk) <= READ_ONCE(sk->sk_rcvbuf)) \|\|
293	__mptcp_check_fallback(mptcp_sk(sk));
294	}
295
296	static bool mptcp_try_rmem_schedule(struct sock sk, const struct sk_buff skb)
297	{
298	if (!mptcp_can_ingest(sk)) {
299	mptcp_prune_ofo_queue(sk, skb);
300	return mptcp_can_ingest(sk);
301	}
302	return true;
303	}
304
305	/* "inspired" by tcp_data_queue_ofo(), main differences:
306	* - use mptcp seqs
307	* - don't cope with sacks
308	*/
309	static void mptcp_data_queue_ofo(struct mptcp_sock msk, struct sk_buff skb)
310	{
311	struct sock sk = (struct sock )msk;
312	struct rb_node *p, parent;
313	u64 seq, end_seq, max_seq;
314	struct sk_buff *skb1;
315
316	seq = MPTCP_SKB_CB(skb)->map_seq;
317	end_seq = MPTCP_SKB_CB(skb)->end_seq;
318	max_seq = atomic64_read(&msk->rcv_wnd_sent);
319
320	pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq,
321	RB_EMPTY_ROOT(&msk->out_of_order_queue));
322	if (after64(end_seq, max_seq)) {
323	/* out of window */
324	mptcp_drop(sk, skb);
325	pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
326	(unsigned long long)end_seq - (unsigned long)max_seq,
327	(unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
328	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
329	return;
330	}
331
332	if (!mptcp_try_rmem_schedule(sk, skb)) {
333	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
334	mptcp_drop(sk, skb);
335	return;
336	}
337
338	p = &msk->out_of_order_queue.rb_node;
339	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
340	if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
341	rb_link_node(&skb->rbnode, NULL, p);
342	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
343	msk->ooo_last_skb = skb;
344	goto end;
345	}
346
347	/* with 2 subflows, adding at end of ooo queue is quite likely
348	* Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
349	*/
350	if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
351	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
352	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
353	return;
354	}
355
356	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
357	if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
358	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
359	parent = &msk->ooo_last_skb->rbnode;
360	p = &parent->rb_right;
361	goto insert;
362	}
363
364	/* Find place to insert this segment. Handle overlaps on the way. */
365	parent = NULL;
366	while (*p) {
367	parent = *p;
368	skb1 = rb_to_skb(parent);
369	if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
370	p = &parent->rb_left;
371	continue;
372	}
373	if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
374	if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
375	/* All the bits are present. Drop. */
376	mptcp_drop(sk, skb);
377	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
378	return;
379	}
380	if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
381	/* partial overlap:
382	* \| skb \|
383	* \| skb1 \|
384	* continue traversing
385	*/
386	} else {
387	/* skb's seq == skb1's seq and skb covers skb1.
388	* Replace skb1 with skb.
389	*/
390	rb_replace_node(&skb1->rbnode, &skb->rbnode,
391	&msk->out_of_order_queue);
392	mptcp_drop(sk, skb1);
393	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
394	goto merge_right;
395	}
396	} else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
397	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
398	return;
399	}
400	p = &parent->rb_right;
401	}
402
403	insert:
404	/* Insert segment into RB tree. */
405	rb_link_node(&skb->rbnode, parent, p);
406	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
407
408	merge_right:
409	/* Remove other segments covered by skb. */
410	while ((skb1 = skb_rb_next(skb)) != NULL) {
411	if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
412	break;
413	rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
414	mptcp_drop(sk, skb1);
415	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
416	}
417	/* If there is no skb after us, we are the last_skb ! */
418	if (!skb1)
419	msk->ooo_last_skb = skb;
420
421	end:
422	skb_condense(skb);
423	skb_set_owner_r(skb, sk);
424	}
425
426	static void mptcp_init_skb(struct sock ssk, struct sk_buff skb, int offset,
427	int copy_len)
428	{
429	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
430	bool has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
431
432	/* the skb map_seq accounts for the skb offset:
433	* mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
434	* value
435	*/
436	MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
437	MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
438	MPTCP_SKB_CB(skb)->offset = offset;
439	MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
440	MPTCP_SKB_CB(skb)->cant_coalesce = 0;
441
442	__skb_unlink(skb, &ssk->sk_receive_queue);
443
444	skb_ext_reset(skb);
445	skb_dst_drop(skb);
446	}
447
448	static bool __mptcp_move_skb(struct sock sk, struct sk_buff skb)
449	{
450	u64 copy_len = MPTCP_SKB_CB(skb)->end_seq - MPTCP_SKB_CB(skb)->map_seq;
451	struct mptcp_sock *msk = mptcp_sk(sk);
452	struct sk_buff *tail;
453
454	mptcp_borrow_fwdmem(sk, skb);
455
456	if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
457	/* in sequence */
458	insert:
459	if (!mptcp_try_rmem_schedule(sk, skb)) {
460	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
461	mptcp_drop(sk, skb);
462	return false;
463	}
464
465	msk->bytes_received += copy_len;
466	WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
467	tail = skb_peek_tail(&sk->sk_receive_queue);
468	if (tail && mptcp_try_coalesce(sk, tail, skb))
469	return true;
470
471	skb_set_owner_r(skb, sk);
472	__skb_queue_tail(&sk->sk_receive_queue, skb);
473	return true;
474	} else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
475	mptcp_data_queue_ofo(msk, skb);
476	return false;
477	}
478
479	/* Partial packet */
480	if (after64(MPTCP_SKB_CB(skb)->end_seq, msk->ack_seq)) {
481	copy_len = MPTCP_SKB_CB(skb)->end_seq - msk->ack_seq;
482	MPTCP_SKB_CB(skb)->offset += msk->ack_seq -
483	MPTCP_SKB_CB(skb)->map_seq;
484	MPTCP_SKB_CB(skb)->map_seq += msk->ack_seq -
485	MPTCP_SKB_CB(skb)->map_seq;
486	goto insert;
487	}
488
489	/* Completely old data */
490	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
491	mptcp_drop(sk, skb);
492	return false;
493	}
494
495	static void mptcp_stop_rtx_timer(struct sock *sk)
496	{
497	sk_stop_timer(sk, &sk->mptcp_retransmit_timer);
498	mptcp_sk(sk)->timer_ival = 0;
499	}
500
501	static void mptcp_close_wake_up(struct sock *sk)
502	{
503	if (sock_flag(sk, SOCK_DEAD))
504	return;
505
506	sk->sk_state_change(sk);
507	if (sk->sk_shutdown == SHUTDOWN_MASK \|\|
508	sk->sk_state == TCP_CLOSE)
509	sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
510	else
511	sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
512	}
513
514	static void mptcp_shutdown_subflows(struct mptcp_sock *msk)
515	{
516	struct mptcp_subflow_context *subflow;
517
518	mptcp_for_each_subflow(msk, subflow) {
519	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
520	bool slow;
521
522	slow = lock_sock_fast(ssk);
523	tcp_shutdown(ssk, SEND_SHUTDOWN);
524	unlock_sock_fast(ssk, slow);
525	}
526	}
527
528	/* called under the msk socket lock */
529	static bool mptcp_pending_data_fin_ack(struct sock *sk)
530	{
531	struct mptcp_sock *msk = mptcp_sk(sk);
532
533	return ((1 << sk->sk_state) &
534	(TCPF_FIN_WAIT1 \| TCPF_CLOSING \| TCPF_LAST_ACK)) &&
535	msk->write_seq == READ_ONCE(msk->snd_una);
536	}
537
538	static void mptcp_check_data_fin_ack(struct sock *sk)
539	{
540	struct mptcp_sock *msk = mptcp_sk(sk);
541
542	/* Look for an acknowledged DATA_FIN */
543	if (mptcp_pending_data_fin_ack(sk)) {
544	WRITE_ONCE(msk->snd_data_fin_enable, 0);
545
546	switch (sk->sk_state) {
547	case TCP_FIN_WAIT1:
548	mptcp_set_state(sk, TCP_FIN_WAIT2);
549	break;
550	case TCP_CLOSING:
551	case TCP_LAST_ACK:
552	mptcp_shutdown_subflows(msk);
553	mptcp_set_state(sk, TCP_CLOSE);
554	break;
555	}
556
557	mptcp_close_wake_up(sk);
558	}
559	}
560
561	/* can be called with no lock acquired */
562	static bool mptcp_pending_data_fin(struct sock sk, u64 seq)
563	{
564	struct mptcp_sock *msk = mptcp_sk(sk);
565
566	if (READ_ONCE(msk->rcv_data_fin) &&
567	((1 << inet_sk_state_load(sk)) &
568	(TCPF_ESTABLISHED \| TCPF_FIN_WAIT1 \| TCPF_FIN_WAIT2))) {
569	u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
570
571	if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) {
572	if (seq)
573	*seq = rcv_data_fin_seq;
574
575	return true;
576	}
577	}
578
579	return false;
580	}
581
582	static void mptcp_set_datafin_timeout(struct sock *sk)
583	{
584	struct inet_connection_sock *icsk = inet_csk(sk);
585	u32 rto_min = READ_ONCE(icsk->icsk_rto_min);
586	u32 rto_max = READ_ONCE(icsk->icsk_rto_max);
587	u32 retransmits;
588
589	/* The sysctls are validated independently: rto_min > rto_max is
590	* possible, guard against ilog2(0).
591	*/
592	retransmits = min_t(u32, icsk->icsk_retransmits,
593	ilog2(max_t(u32, rto_max / rto_min, 1)));
594
595	mptcp_sk(sk)->timer_ival = rto_min << retransmits;
596	}
597
598	static void __mptcp_set_timeout(struct sock *sk, long tout)
599	{
600	mptcp_sk(sk)->timer_ival = tout > 0 ? tout :
601	READ_ONCE(inet_csk(sk)->icsk_rto_min);
602	}
603
604	static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
605	{
606	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
607
608	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
609	tcp_timeout_expires(ssk) - jiffies : 0;
610	}
611
612	static void mptcp_set_timeout(struct sock *sk)
613	{
614	struct mptcp_subflow_context *subflow;
615	long tout = 0;
616
617	mptcp_for_each_subflow(mptcp_sk(sk), subflow)
618	tout = max(tout, mptcp_timeout_from_subflow(subflow));
619	__mptcp_set_timeout(sk, tout);
620	}
621
622	static inline bool tcp_can_send_ack(const struct sock *ssk)
623	{
624	return !((1 << inet_sk_state_load(ssk)) &
625	(TCPF_SYN_SENT \| TCPF_SYN_RECV \| TCPF_TIME_WAIT \| TCPF_CLOSE \| TCPF_LISTEN));
626	}
627
628	void __mptcp_subflow_send_ack(struct sock *ssk)
629	{
630	if (tcp_can_send_ack(ssk))
631	tcp_send_ack(ssk);
632	}
633
634	static void mptcp_subflow_send_ack(struct sock *ssk)
635	{
636	bool slow;
637
638	slow = lock_sock_fast(ssk);
639	__mptcp_subflow_send_ack(ssk);
640	unlock_sock_fast(ssk, slow);
641	}
642
643	static void mptcp_send_ack(struct mptcp_sock *msk)
644	{
645	struct mptcp_subflow_context *subflow;
646
647	mptcp_for_each_subflow(msk, subflow)
648	mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
649	}
650
651	static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied)
652	{
653	bool slow;
654
655	slow = lock_sock_fast(ssk);
656	if (tcp_can_send_ack(ssk))
657	tcp_cleanup_rbuf(ssk, copied);
658	unlock_sock_fast(ssk, slow);
659	}
660
661	static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
662	{
663	const struct inet_connection_sock *icsk = inet_csk(ssk);
664	u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
665	const struct tcp_sock *tp = tcp_sk(ssk);
666
667	return (ack_pending & ICSK_ACK_SCHED) &&
668	((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
669	READ_ONCE(icsk->icsk_ack.rcv_mss)) \|\|
670	(rx_empty && ack_pending &
671	(ICSK_ACK_PUSHED2 \| ICSK_ACK_PUSHED)));
672	}
673
674	static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied)
675	{
676	int old_space = READ_ONCE(msk->old_wspace);
677	struct mptcp_subflow_context *subflow;
678	struct sock sk = (struct sock )msk;
679	int space = __mptcp_space(sk);
680	bool cleanup, rx_empty;
681
682	cleanup = (space > 0) && (space >= (old_space << 1)) && copied;
683	rx_empty = !sk_rmem_alloc_get(sk) && copied;
684
685	mptcp_for_each_subflow(msk, subflow) {
686	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
687
688	if (cleanup \|\| mptcp_subflow_could_cleanup(ssk, rx_empty))
689	mptcp_subflow_cleanup_rbuf(ssk, copied);
690	}
691	}
692
693	static void mptcp_check_data_fin(struct sock *sk)
694	{
695	struct mptcp_sock *msk = mptcp_sk(sk);
696	u64 rcv_data_fin_seq;
697
698	/* Need to ack a DATA_FIN received from a peer while this side
699	* of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
700	* msk->rcv_data_fin was set when parsing the incoming options
701	* at the subflow level and the msk lock was not held, so this
702	* is the first opportunity to act on the DATA_FIN and change
703	* the msk state.
704	*
705	* If we are caught up to the sequence number of the incoming
706	* DATA_FIN, send the DATA_ACK now and do state transition. If
707	* not caught up, do nothing and let the recv code send DATA_ACK
708	* when catching up.
709	*/
710
711	if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
712	WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
713	WRITE_ONCE(msk->rcv_data_fin, 0);
714
715	WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown \| RCV_SHUTDOWN);
716	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
717
718	switch (sk->sk_state) {
719	case TCP_ESTABLISHED:
720	mptcp_set_state(sk, TCP_CLOSE_WAIT);
721	break;
722	case TCP_FIN_WAIT1:
723	mptcp_set_state(sk, TCP_CLOSING);
724	break;
725	case TCP_FIN_WAIT2:
726	mptcp_shutdown_subflows(msk);
727	mptcp_set_state(sk, TCP_CLOSE);
728	break;
729	default:
730	/* Other states not expected */
731	WARN_ON_ONCE(1);
732	break;
733	}
734
735	if (!__mptcp_check_fallback(msk))
736	mptcp_send_ack(msk);
737	mptcp_close_wake_up(sk);
738	}
739	}
740
741	static void mptcp_dss_corruption(struct mptcp_sock msk, struct sock ssk)
742	{
743	if (!mptcp_try_fallback(ssk, MPTCP_MIB_DSSCORRUPTIONFALLBACK)) {
744	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET);
745	mptcp_subflow_reset(ssk);
746	}
747	}
748
749	static void __mptcp_add_backlog(struct sock *sk,
750	struct mptcp_subflow_context *subflow,
751	struct sk_buff *skb)
752	{
753	struct mptcp_sock *msk = mptcp_sk(sk);
754	struct sk_buff *tail = NULL;
755	struct sock *ssk = skb->sk;
756	bool fragstolen;
757	u64 limit;
758	int delta;
759
760	if (unlikely(sk->sk_state == TCP_CLOSE)) {
761	kfree_skb_reason(skb, SKB_DROP_REASON_SOCKET_CLOSE);
762	return;
763	}
764
765	/* Similar additional allowance as plain TCP. */
766	limit = READ_ONCE(sk->sk_rcvbuf);
767	limit += (limit >> 1) + 64 * 1024;
768	limit = min_t(u64, limit, UINT_MAX);
769	if (msk->backlog_len > limit && !__mptcp_check_fallback(msk)) {
770	__MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_BACKLOGDROP);
771	kfree_skb_reason(skb, SKB_DROP_REASON_SOCKET_BACKLOG);
772	return;
773	}
774
775	/* Try to coalesce with the last skb in our backlog */
776	if (!list_empty(&msk->backlog_list))
777	tail = list_last_entry(&msk->backlog_list, struct sk_buff, list);
778
779	if (tail && MPTCP_SKB_CB(skb)->map_seq == MPTCP_SKB_CB(tail)->end_seq &&
780	ssk == tail->sk &&
781	__mptcp_try_coalesce(sk, tail, skb, &fragstolen, &delta)) {
782	skb->truesize -= delta;
783	kfree_skb_partial(skb, fragstolen);
784	__mptcp_subflow_lend_fwdmem(subflow, delta);
785	goto account;
786	}
787
788	list_add_tail(&skb->list, &msk->backlog_list);
789	mptcp_subflow_lend_fwdmem(subflow, skb);
790	delta = skb->truesize;
791
792	account:
793	WRITE_ONCE(msk->backlog_len, msk->backlog_len + delta);
794
795	/* Possibly not accept()ed yet, keep track of memory not CG
796	* accounted, mptcp_graft_subflows() will handle it.
797	*/
798	if (!mem_cgroup_from_sk(ssk))
799	msk->backlog_unaccounted += delta;
800	}
801
802	static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
803	struct sock *ssk, bool own_msk)
804	{
805	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
806	struct sock sk = (struct sock )msk;
807	bool more_data_avail;
808	struct tcp_sock *tp;
809	bool ret = false;
810
811	pr_debug("msk=%p ssk=%p\n", msk, ssk);
812	tp = tcp_sk(ssk);
813	do {
814	u32 map_remaining, offset;
815	u32 seq = tp->copied_seq;
816	struct sk_buff *skb;
817	bool fin;
818
819	/* try to move as much data as available */
820	map_remaining = subflow->map_data_len -
821	mptcp_subflow_get_map_offset(subflow);
822
823	skb = skb_peek(&ssk->sk_receive_queue);
824	if (unlikely(!skb))
825	break;
826
827	if (__mptcp_check_fallback(msk)) {
828	/* Under fallback skbs have no MPTCP extension and TCP could
829	* collapse them between the dummy map creation and the
830	* current dequeue. Be sure to adjust the map size.
831	*/
832	map_remaining = skb->len;
833	subflow->map_data_len = skb->len;
834	}
835
836	offset = seq - TCP_SKB_CB(skb)->seq;
837	fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
838	if (fin)
839	seq++;
840
841	if (offset < skb->len) {
842	size_t len = skb->len - offset;
843
844	mptcp_init_skb(ssk, skb, offset, len);
845
846	if (own_msk) {
847	mptcp_subflow_lend_fwdmem(subflow, skb);
848	ret \|= __mptcp_move_skb(sk, skb);
849	} else {
850	__mptcp_add_backlog(sk, subflow, skb);
851	}
852	seq += len;
853
854	if (unlikely(map_remaining < len)) {
855	DEBUG_NET_WARN_ON_ONCE(1);
856	mptcp_dss_corruption(msk, ssk);
857	}
858	} else {
859	sk_eat_skb(ssk, skb);
860
861	if (unlikely(!fin)) {
862	DEBUG_NET_WARN_ON_ONCE(1);
863	mptcp_dss_corruption(msk, ssk);
864	}
865	}
866
867	WRITE_ONCE(tp->copied_seq, seq);
868	more_data_avail = mptcp_subflow_data_available(ssk);
869
870	} while (more_data_avail);
871
872	if (ret)
873	msk->last_data_recv = tcp_jiffies32;
874	return ret;
875	}
876
877	static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
878	{
879	struct sock sk = (struct sock )msk;
880	struct sk_buff skb, tail;
881	bool moved = false;
882	struct rb_node *p;
883	u64 end_seq;
884
885	p = rb_first(&msk->out_of_order_queue);
886	pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
887	while (p) {
888	skb = rb_to_skb(p);
889	if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
890	break;
891
892	p = rb_next(p);
893	rb_erase(&skb->rbnode, &msk->out_of_order_queue);
894
895	if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
896	msk->ack_seq))) {
897	mptcp_drop(sk, skb);
898	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
899	continue;
900	}
901
902	end_seq = MPTCP_SKB_CB(skb)->end_seq;
903	tail = skb_peek_tail(&sk->sk_receive_queue);
904	if (!tail \|\| !mptcp_ooo_try_coalesce(msk, tail, skb)) {
905	int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
906
907	/* skip overlapping data, if any */
908	pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n",
909	MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
910	delta);
911	MPTCP_SKB_CB(skb)->offset += delta;
912	MPTCP_SKB_CB(skb)->map_seq += delta;
913	__skb_queue_tail(&sk->sk_receive_queue, skb);
914	}
915	msk->bytes_received += end_seq - msk->ack_seq;
916	WRITE_ONCE(msk->ack_seq, end_seq);
917	moved = true;
918	}
919	return moved;
920	}
921
922	static bool __mptcp_subflow_error_report(struct sock sk, struct sock ssk)
923	{
924	int ssk_state;
925	int err;
926
927	/* only propagate errors on fallen-back sockets or
928	* on MPC connect
929	*/
930	if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
931	return false;
932
933	err = sock_error(ssk);
934	if (!err)
935	return false;
936
937	/* We need to propagate only transition to CLOSE state.
938	* Orphaned socket will see such state change via
939	* subflow_sched_work_if_closed() and that path will properly
940	* destroy the msk as needed.
941	*/
942	ssk_state = inet_sk_state_load(ssk);
943	if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
944	mptcp_set_state(sk, ssk_state);
945	WRITE_ONCE(sk->sk_err, -err);
946
947	/* This barrier is coupled with smp_rmb() in mptcp_poll() */
948	smp_wmb();
949	sk_error_report(sk);
950	return true;
951	}
952
953	void __mptcp_error_report(struct sock *sk)
954	{
955	struct mptcp_subflow_context *subflow;
956	struct mptcp_sock *msk = mptcp_sk(sk);
957
958	mptcp_for_each_subflow(msk, subflow)
959	if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
960	break;
961	}
962
963	/* In most cases we will be able to lock the mptcp socket. If its already
964	* owned, we need to defer to the work queue to avoid ABBA deadlock.
965	*/
966	static bool move_skbs_to_msk(struct mptcp_sock msk, struct sock ssk)
967	{
968	struct sock sk = (struct sock )msk;
969	bool moved;
970
971	moved = __mptcp_move_skbs_from_subflow(msk, ssk, true);
972	__mptcp_ofo_queue(msk);
973	if (unlikely(ssk->sk_err))
974	__mptcp_subflow_error_report(sk, ssk);
975
976	/* If the moves have caught up with the DATA_FIN sequence number
977	* it's time to ack the DATA_FIN and change socket state, but
978	* this is not a good place to change state. Let the workqueue
979	* do it.
980	*/
981	if (mptcp_pending_data_fin(sk, NULL))
982	mptcp_schedule_work(sk);
983	return moved;
984	}
985
986	static void mptcp_rcv_rtt_update(struct mptcp_sock *msk,
987	struct mptcp_subflow_context *subflow)
988	{
989	const struct tcp_sock *tp = tcp_sk(subflow->tcp_sock);
990	u32 rtt_us = tp->rcv_rtt_est.rtt_us;
991	int id;
992
993	/* Update once per subflow per rcvwnd to avoid touching the msk
994	* too often.
995	*/
996	if (!rtt_us \|\| tp->rcv_rtt_est.seq == subflow->prev_rtt_seq)
997	return;
998
999	subflow->prev_rtt_seq = tp->rcv_rtt_est.seq;
1000
1001	/* Pairs with READ_ONCE() in mptcp_rtt_us_est(). */
1002	id = msk->rcv_rtt_est.next_sample;
1003	WRITE_ONCE(msk->rcv_rtt_est.samples[id], rtt_us);
1004	if (++msk->rcv_rtt_est.next_sample == MPTCP_RTT_SAMPLES)
1005	msk->rcv_rtt_est.next_sample = 0;
1006
1007	/* EWMA among the incoming subflows */
1008	msk->scaling_ratio = ((msk->scaling_ratio << 3) - msk->scaling_ratio +
1009	tp->scaling_ratio) >> 3;
1010	}
1011
1012	void mptcp_data_ready(struct sock sk, struct sock ssk)
1013	{
1014	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1015	struct mptcp_sock *msk = mptcp_sk(sk);
1016
1017	/* The peer can send data while we are shutting down this
1018	* subflow at subflow destruction time, but we must avoid enqueuing
1019	* more data to the msk receive queue
1020	*/
1021	if (unlikely(subflow->closing))
1022	return;
1023
1024	mptcp_data_lock(sk);
1025	mptcp_rcv_rtt_update(msk, subflow);
1026	if (!sock_owned_by_user(sk)) {
1027	/* Wake-up the reader only for in-sequence data */
1028	if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
1029	sk->sk_data_ready(sk);
1030	} else {
1031	__mptcp_move_skbs_from_subflow(msk, ssk, false);
1032	}
1033	mptcp_data_unlock(sk);
1034	}
1035
1036	static void mptcp_subflow_joined(struct mptcp_sock msk, struct sock ssk)
1037	{
1038	mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
1039	msk->allow_infinite_fallback = false;
1040	mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
1041	}
1042
1043	static bool __mptcp_finish_join(struct mptcp_sock msk, struct sock ssk)
1044	{
1045	struct sock sk = (struct sock )msk;
1046
1047	if (sk->sk_state != TCP_ESTABLISHED)
1048	return false;
1049
1050	spin_lock_bh(&msk->fallback_lock);
1051	if (!msk->allow_subflows) {
1052	spin_unlock_bh(&msk->fallback_lock);
1053	return false;
1054	}
1055	mptcp_subflow_joined(msk, ssk);
1056	spin_unlock_bh(&msk->fallback_lock);
1057
1058	mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
1059	mptcp_sockopt_sync_locked(msk, ssk);
1060	mptcp_stop_tout_timer(sk);
1061	__mptcp_propagate_sndbuf(sk, ssk);
1062	return true;
1063	}
1064
1065	static void __mptcp_flush_join_list(struct sock sk, struct list_head join_list)
1066	{
1067	struct mptcp_subflow_context tmp, subflow;
1068	struct mptcp_sock *msk = mptcp_sk(sk);
1069
1070	list_for_each_entry_safe(subflow, tmp, join_list, node) {
1071	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1072	bool slow = lock_sock_fast(ssk);
1073
1074	list_move_tail(&subflow->node, &msk->conn_list);
1075	if (!__mptcp_finish_join(msk, ssk))
1076	mptcp_subflow_reset(ssk);
1077	unlock_sock_fast(ssk, slow);
1078	}
1079	}
1080
1081	static bool mptcp_rtx_timer_pending(struct sock *sk)
1082	{
1083	return timer_pending(&sk->mptcp_retransmit_timer);
1084	}
1085
1086	static void mptcp_reset_rtx_timer(struct sock *sk)
1087	{
1088	struct mptcp_sock *msk = mptcp_sk(sk);
1089	unsigned long tout;
1090
1091	/* Prevent rescheduling on close and in case of fallback. */
1092	if (!test_bit(MPTCP_RTX_ENABLED, &msk->flags))
1093	return;
1094
1095	tout = msk->timer_ival;
1096	sk_reset_timer(sk, &sk->mptcp_retransmit_timer, jiffies + tout);
1097	}
1098
1099	bool mptcp_schedule_work(struct sock *sk)
1100	{
1101	if (inet_sk_state_load(sk) == TCP_CLOSE)
1102	return false;
1103
1104	/* Get a reference on this socket, mptcp_worker() will release it.
1105	* As mptcp_worker() might complete before us, we can not avoid
1106	* a sock_hold()/sock_put() if schedule_work() returns false.
1107	*/
1108	sock_hold(sk);
1109
1110	if (schedule_work(&mptcp_sk(sk)->work))
1111	return true;
1112
1113	sock_put(sk);
1114	return false;
1115	}
1116
1117	static bool mptcp_skb_can_collapse_to(u64 write_seq,
1118	const struct sk_buff *skb,
1119	const struct mptcp_ext *mpext)
1120	{
1121	if (!tcp_skb_can_collapse_to(skb))
1122	return false;
1123
1124	/* can collapse only if MPTCP level sequence is in order and this
1125	* mapping has not been xmitted yet
1126	*/
1127	return mpext && mpext->data_seq + mpext->data_len == write_seq &&
1128	!mpext->frozen;
1129	}
1130
1131	/* we can append data to the given data frag if:
1132	* - there is space available in the backing page_frag
1133	* - the data frag tail matches the current page_frag free offset
1134	* - the data frag end sequence number matches the current write seq
1135	*/
1136	static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
1137	const struct page_frag *pfrag,
1138	const struct mptcp_data_frag *df)
1139	{
1140	return df && !df->eor &&
1141	pfrag->page == df->page &&
1142	pfrag->size - pfrag->offset > 0 &&
1143	pfrag->offset == (df->offset + df->data_len) &&
1144	df->data_seq + df->data_len == msk->write_seq;
1145	}
1146
1147	static void dfrag_uncharge(struct sock *sk, int len)
1148	{
1149	sk_mem_uncharge(sk, len);
1150	sk_wmem_queued_add(sk, -len);
1151	}
1152
1153	static void dfrag_clear(struct sock sk, struct mptcp_data_frag dfrag)
1154	{
1155	int len = dfrag->data_len + dfrag->overhead;
1156
1157	list_del(&dfrag->list);
1158	dfrag_uncharge(sk, len);
1159	put_page(dfrag->page);
1160	}
1161
1162	/* called under both the msk socket lock and the data lock */
1163	static void __mptcp_clean_una(struct sock *sk)
1164	{
1165	struct mptcp_sock *msk = mptcp_sk(sk);
1166	struct mptcp_data_frag dtmp, dfrag;
1167	u64 snd_una;
1168
1169	snd_una = msk->snd_una;
1170	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1171	if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1172	break;
1173
1174	if (unlikely(dfrag == msk->first_pending)) {
1175	/* in recovery mode can see ack after the current snd head */
1176	if (WARN_ON_ONCE(!msk->recovery))
1177	break;
1178
1179	msk->first_pending = mptcp_send_next(sk);
1180	}
1181
1182	dfrag_clear(sk, dfrag);
1183	}
1184
1185	dfrag = mptcp_rtx_head(sk);
1186	if (dfrag && after64(snd_una, dfrag->data_seq)) {
1187	u64 delta = snd_una - dfrag->data_seq;
1188
1189	/* prevent wrap around in recovery mode */
1190	if (unlikely(delta > dfrag->already_sent)) {
1191	if (WARN_ON_ONCE(!msk->recovery))
1192	goto out;
1193	if (WARN_ON_ONCE(delta > dfrag->data_len))
1194	goto out;
1195	dfrag->already_sent += delta - dfrag->already_sent;
1196	}
1197
1198	dfrag->data_seq += delta;
1199	dfrag->offset += delta;
1200	dfrag->data_len -= delta;
1201	dfrag->already_sent -= delta;
1202
1203	dfrag_uncharge(sk, delta);
1204	}
1205
1206	/* all retransmitted data acked, recovery completed */
1207	if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1208	msk->recovery = false;
1209
1210	out:
1211	if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) {
1212	if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1213	mptcp_stop_rtx_timer(sk);
1214	} else {
1215	mptcp_reset_rtx_timer(sk);
1216	}
1217
1218	if (mptcp_pending_data_fin_ack(sk))
1219	mptcp_schedule_work(sk);
1220	}
1221
1222	static void __mptcp_clean_una_wakeup(struct sock *sk)
1223	{
1224	lockdep_assert_held_once(&sk->sk_lock.slock);
1225
1226	__mptcp_clean_una(sk);
1227	mptcp_write_space(sk);
1228	}
1229
1230	static void mptcp_enter_memory_pressure(struct sock *sk)
1231	{
1232	struct mptcp_subflow_context *subflow;
1233	struct mptcp_sock *msk = mptcp_sk(sk);
1234	bool first = true;
1235
1236	mptcp_for_each_subflow(msk, subflow) {
1237	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1238
1239	if (first && !ssk->sk_bypass_prot_mem) {
1240	tcp_enter_memory_pressure(ssk);
1241	first = false;
1242	}
1243
1244	sk_stream_moderate_sndbuf(ssk);
1245	}
1246	__mptcp_sync_sndbuf(sk);
1247	}
1248
1249	/* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1250	* data
1251	*/
1252	static bool mptcp_page_frag_refill(struct sock sk, struct page_frag pfrag)
1253	{
1254	if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1255	pfrag, sk->sk_allocation)))
1256	return true;
1257
1258	mptcp_enter_memory_pressure(sk);
1259	return false;
1260	}
1261
1262	static struct mptcp_data_frag *
1263	mptcp_carve_data_frag(const struct mptcp_sock msk, struct page_frag pfrag,
1264	int orig_offset)
1265	{
1266	int offset = ALIGN(orig_offset, sizeof(long));
1267	struct mptcp_data_frag *dfrag;
1268
1269	dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1270	dfrag->data_len = 0;
1271	dfrag->data_seq = msk->write_seq;
1272	dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1273	dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1274	dfrag->already_sent = 0;
1275	dfrag->page = pfrag->page;
1276	dfrag->eor = 0;
1277
1278	return dfrag;
1279	}
1280
1281	struct mptcp_sendmsg_info {
1282	int mss_now;
1283	int size_goal;
1284	u16 limit;
1285	u16 sent;
1286	unsigned int flags;
1287	bool data_lock_held;
1288	};
1289
1290	static size_t mptcp_check_allowed_size(const struct mptcp_sock *msk,
1291	struct sock *ssk, u64 data_seq,
1292	size_t avail_size)
1293	{
1294	u64 window_end = mptcp_wnd_end(msk);
1295	u64 mptcp_snd_wnd;
1296
1297	if (__mptcp_check_fallback(msk))
1298	return avail_size;
1299
1300	mptcp_snd_wnd = window_end - data_seq;
1301	avail_size = min(mptcp_snd_wnd, avail_size);
1302
1303	if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1304	tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1305	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1306	}
1307
1308	return avail_size;
1309	}
1310
1311	static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1312	{
1313	struct skb_ext *mpext = __skb_ext_alloc(gfp);
1314
1315	if (!mpext)
1316	return false;
1317	__skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1318	return true;
1319	}
1320
1321	static struct sk_buff __mptcp_do_alloc_tx_skb(struct sock sk, gfp_t gfp)
1322	{
1323	struct sk_buff *skb;
1324
1325	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1326	if (likely(skb)) {
1327	if (likely(__mptcp_add_ext(skb, gfp))) {
1328	skb_reserve(skb, MAX_TCP_HEADER);
1329	skb->ip_summed = CHECKSUM_PARTIAL;
1330	INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1331	return skb;
1332	}
1333	__kfree_skb(skb);
1334	} else {
1335	mptcp_enter_memory_pressure(sk);
1336	}
1337	return NULL;
1338	}
1339
1340	static struct sk_buff __mptcp_alloc_tx_skb(struct sock sk, struct sock *ssk, gfp_t gfp)
1341	{
1342	struct sk_buff *skb;
1343
1344	skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1345	if (!skb)
1346	return NULL;
1347
1348	if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1349	tcp_skb_entail(ssk, skb);
1350	return skb;
1351	}
1352	tcp_skb_tsorted_anchor_cleanup(skb);
1353	kfree_skb(skb);
1354	return NULL;
1355	}
1356
1357	static struct sk_buff mptcp_alloc_tx_skb(struct sock sk, struct sock *ssk, bool data_lock_held)
1358	{
1359	gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1360
1361	return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1362	}
1363
1364	/* note: this always recompute the csum on the whole skb, even
1365	* if we just appended a single frag. More status info needed
1366	*/
1367	static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1368	{
1369	struct mptcp_ext *mpext = mptcp_get_ext(skb);
1370	__wsum csum = ~csum_unfold(mpext->csum);
1371	int offset = skb->len - added;
1372
1373	mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1374	}
1375
1376	static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1377	struct sock *ssk,
1378	struct mptcp_ext *mpext)
1379	{
1380	if (!mpext)
1381	return;
1382
1383	mpext->infinite_map = 1;
1384	mpext->data_len = 0;
1385
1386	if (!mptcp_try_fallback(ssk, MPTCP_MIB_INFINITEMAPTX)) {
1387	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_FALLBACKFAILED);
1388	mptcp_subflow_reset(ssk);
1389	return;
1390	}
1391
1392	mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1393	}
1394
1395	#define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1396
1397	static int mptcp_sendmsg_frag(struct sock sk, struct sock ssk,
1398	struct mptcp_data_frag *dfrag,
1399	struct mptcp_sendmsg_info *info)
1400	{
1401	u64 data_seq = dfrag->data_seq + info->sent;
1402	int offset = dfrag->offset + info->sent;
1403	struct mptcp_sock *msk = mptcp_sk(sk);
1404	bool zero_window_probe = false;
1405	struct mptcp_ext *mpext = NULL;
1406	bool can_coalesce = false;
1407	bool reuse_skb = true;
1408	struct sk_buff *skb;
1409	size_t copy;
1410	int i;
1411
1412	pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n",
1413	msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1414
1415	if (WARN_ON_ONCE(info->sent > info->limit \|\|
1416	info->limit > dfrag->data_len))
1417	return 0;
1418
1419	if (unlikely(!__tcp_can_send(ssk)))
1420	return -EAGAIN;
1421
1422	/* compute send limit */
1423	if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1424	ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1425	info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1426	copy = info->size_goal;
1427
1428	skb = tcp_write_queue_tail(ssk);
1429	if (skb && copy > skb->len) {
1430	/* Limit the write to the size available in the
1431	* current skb, if any, so that we create at most a new skb.
1432	* Explicitly tells TCP internals to avoid collapsing on later
1433	* queue management operation, to avoid breaking the ext <->
1434	* SSN association set here
1435	*/
1436	mpext = mptcp_get_ext(skb);
1437	if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1438	TCP_SKB_CB(skb)->eor = 1;
1439	tcp_mark_push(tcp_sk(ssk), skb);
1440	goto alloc_skb;
1441	}
1442
1443	i = skb_shinfo(skb)->nr_frags;
1444	can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1445	if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1446	tcp_mark_push(tcp_sk(ssk), skb);
1447	goto alloc_skb;
1448	}
1449
1450	copy -= skb->len;
1451	} else {
1452	alloc_skb:
1453	skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1454	if (!skb)
1455	return -ENOMEM;
1456
1457	i = skb_shinfo(skb)->nr_frags;
1458	reuse_skb = false;
1459	mpext = mptcp_get_ext(skb);
1460	}
1461
1462	/* Zero window and all data acked? Probe. */
1463	copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1464	if (copy == 0) {
1465	u64 snd_una = READ_ONCE(msk->snd_una);
1466
1467	/* No need for zero probe if there are any data pending
1468	* either at the msk or ssk level; skb is the current write
1469	* queue tail and can be empty at this point.
1470	*/
1471	if (snd_una != msk->snd_nxt \|\| skb->len \|\|
1472	skb != tcp_send_head(ssk)) {
1473	tcp_remove_empty_skb(ssk);
1474	return 0;
1475	}
1476
1477	zero_window_probe = true;
1478	data_seq = snd_una - 1;
1479	copy = 1;
1480	}
1481
1482	copy = min_t(size_t, copy, info->limit - info->sent);
1483	if (!sk_wmem_schedule(ssk, copy)) {
1484	tcp_remove_empty_skb(ssk);
1485	return -ENOMEM;
1486	}
1487
1488	if (can_coalesce) {
1489	skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1490	} else {
1491	get_page(dfrag->page);
1492	skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1493	}
1494
1495	skb->len += copy;
1496	skb->data_len += copy;
1497	skb->truesize += copy;
1498	sk_wmem_queued_add(ssk, copy);
1499	sk_mem_charge(ssk, copy);
1500	WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1501	TCP_SKB_CB(skb)->end_seq += copy;
1502	tcp_skb_pcount_set(skb, 0);
1503
1504	/* on skb reuse we just need to update the DSS len */
1505	if (reuse_skb) {
1506	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1507	mpext->data_len += copy;
1508	goto out;
1509	}
1510
1511	memset(mpext, 0, sizeof(*mpext));
1512	mpext->data_seq = data_seq;
1513	mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1514	mpext->data_len = copy;
1515	mpext->use_map = 1;
1516	mpext->dsn64 = 1;
1517
1518	pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n",
1519	mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1520	mpext->dsn64);
1521
1522	if (zero_window_probe) {
1523	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_WINPROBE);
1524	mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1525	mpext->frozen = 1;
1526	if (READ_ONCE(msk->csum_enabled))
1527	mptcp_update_data_checksum(skb, copy);
1528	tcp_push_pending_frames(ssk);
1529	return 0;
1530	}
1531	out:
1532	if (READ_ONCE(msk->csum_enabled))
1533	mptcp_update_data_checksum(skb, copy);
1534	if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1535	mptcp_update_infinite_map(msk, ssk, mpext);
1536	trace_mptcp_sendmsg_frag(mpext);
1537	mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1538
1539	/* if this is the last chunk of a dfrag with MSG_EOR set,
1540	* mark the skb to prevent coalescing with subsequent data.
1541	*/
1542	if (dfrag->eor && info->sent + copy >= dfrag->data_len)
1543	TCP_SKB_CB(skb)->eor = 1;
1544
1545	return copy;
1546	}
1547
1548	#define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1549	sizeof(struct tcphdr) - \
1550	MAX_TCP_OPTION_SPACE - \
1551	sizeof(struct ipv6hdr) - \
1552	sizeof(struct frag_hdr))
1553
1554	struct subflow_send_info {
1555	struct sock *ssk;
1556	u64 linger_time;
1557	};
1558
1559	void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1560	{
1561	if (!subflow->stale)
1562	return;
1563
1564	subflow->stale = 0;
1565	MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1566	}
1567
1568	bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1569	{
1570	if (unlikely(subflow->stale)) {
1571	u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1572
1573	if (subflow->stale_rcv_tstamp == rcv_tstamp)
1574	return false;
1575
1576	mptcp_subflow_set_active(subflow);
1577	}
1578	return __mptcp_subflow_active(subflow);
1579	}
1580
1581	#define SSK_MODE_ACTIVE 0
1582	#define SSK_MODE_BACKUP 1
1583	#define SSK_MODE_MAX 2
1584
1585	/* implement the mptcp packet scheduler;
1586	* returns the subflow that will transmit the next DSS
1587	* additionally updates the rtx timeout
1588	*/
1589	struct sock mptcp_subflow_get_send(struct mptcp_sock msk)
1590	{
1591	struct subflow_send_info send_info[SSK_MODE_MAX];
1592	struct mptcp_subflow_context *subflow;
1593	struct sock sk = (struct sock )msk;
1594	u32 pace, burst, wmem;
1595	int i, nr_active = 0;
1596	struct sock *ssk;
1597	u64 linger_time;
1598	long tout = 0;
1599
1600	/* pick the subflow with the lower wmem/wspace ratio */
1601	for (i = 0; i < SSK_MODE_MAX; ++i) {
1602	send_info[i].ssk = NULL;
1603	send_info[i].linger_time = -1;
1604	}
1605
1606	mptcp_for_each_subflow(msk, subflow) {
1607	bool backup = subflow->backup \|\| subflow->request_bkup;
1608
1609	trace_mptcp_subflow_get_send(subflow);
1610	ssk = mptcp_subflow_tcp_sock(subflow);
1611	if (!mptcp_subflow_active(subflow))
1612	continue;
1613
1614	tout = max(tout, mptcp_timeout_from_subflow(subflow));
1615	nr_active += !backup;
1616	pace = subflow->avg_pacing_rate;
1617	if (unlikely(!pace)) {
1618	/* init pacing rate from socket */
1619	subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1620	pace = subflow->avg_pacing_rate;
1621	if (!pace)
1622	continue;
1623	}
1624
1625	linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1626	if (linger_time < send_info[backup].linger_time) {
1627	send_info[backup].ssk = ssk;
1628	send_info[backup].linger_time = linger_time;
1629	}
1630	}
1631	__mptcp_set_timeout(sk, tout);
1632
1633	/* pick the best backup if no other subflow is active */
1634	if (!nr_active)
1635	send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1636
1637	/* According to the blest algorithm, to avoid HoL blocking for the
1638	* faster flow, we need to:
1639	* - estimate the faster flow linger time
1640	* - use the above to estimate the amount of byte transferred
1641	* by the faster flow
1642	* - check that the amount of queued data is greater than the above,
1643	* otherwise do not use the picked, slower, subflow
1644	* We select the subflow with the shorter estimated time to flush
1645	* the queued mem, which basically ensure the above. We just need
1646	* to check that subflow has a non empty cwin.
1647	*/
1648	ssk = send_info[SSK_MODE_ACTIVE].ssk;
1649	if (!ssk \|\| !sk_stream_memory_free(ssk))
1650	return NULL;
1651
1652	burst = min(MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1653	wmem = READ_ONCE(ssk->sk_wmem_queued);
1654	if (!burst)
1655	return ssk;
1656
1657	subflow = mptcp_subflow_ctx(ssk);
1658	subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1659	READ_ONCE(ssk->sk_pacing_rate) * burst,
1660	burst + wmem);
1661	msk->snd_burst = burst;
1662	return ssk;
1663	}
1664
1665	static void mptcp_push_release(struct sock ssk, struct mptcp_sendmsg_info info)
1666	{
1667	if (info->mss_now)
1668	tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle,
1669	info->size_goal);
1670	release_sock(ssk);
1671	}
1672
1673	static void mptcp_update_post_push(struct mptcp_sock *msk,
1674	struct mptcp_data_frag *dfrag,
1675	u32 sent)
1676	{
1677	u64 snd_nxt_new = dfrag->data_seq;
1678
1679	dfrag->already_sent += sent;
1680
1681	msk->snd_burst -= sent;
1682
1683	snd_nxt_new += dfrag->already_sent;
1684
1685	/* snd_nxt_new can be smaller than snd_nxt in case mptcp
1686	* is recovering after a failover. In that event, this re-sends
1687	* old segments.
1688	*
1689	* Thus compute snd_nxt_new candidate based on
1690	* the dfrag->data_seq that was sent and the data
1691	* that has been handed to the subflow for transmission
1692	* and skip update in case it was old dfrag.
1693	*/
1694	if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1695	msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1696	WRITE_ONCE(msk->snd_nxt, snd_nxt_new);
1697	}
1698	}
1699
1700	void mptcp_check_and_set_pending(struct sock *sk)
1701	{
1702	if (mptcp_send_head(sk)) {
1703	mptcp_data_lock(sk);
1704	mptcp_sk(sk)->cb_flags \|= BIT(MPTCP_PUSH_PENDING);
1705	mptcp_data_unlock(sk);
1706	}
1707	}
1708
1709	static int __subflow_push_pending(struct sock sk, struct sock ssk,
1710	struct mptcp_sendmsg_info *info)
1711	{
1712	struct mptcp_sock *msk = mptcp_sk(sk);
1713	struct mptcp_data_frag *dfrag;
1714	int len, copied = 0, err = 0;
1715
1716	while ((dfrag = mptcp_send_head(sk))) {
1717	info->sent = dfrag->already_sent;
1718	info->limit = dfrag->data_len;
1719	len = dfrag->data_len - dfrag->already_sent;
1720	while (len > 0) {
1721	int ret = 0;
1722
1723	ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1724	if (ret <= 0) {
1725	err = copied ? : ret;
1726	goto out;
1727	}
1728
1729	info->sent += ret;
1730	copied += ret;
1731	len -= ret;
1732
1733	mptcp_update_post_push(msk, dfrag, ret);
1734	}
1735	msk->first_pending = mptcp_send_next(sk);
1736
1737	if (msk->snd_burst <= 0 \|\|
1738	!sk_stream_memory_free(ssk) \|\|
1739	!mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1740	err = copied;
1741	goto out;
1742	}
1743	mptcp_set_timeout(sk);
1744	}
1745	err = copied;
1746
1747	out:
1748	if (err > 0)
1749	msk->last_data_sent = tcp_jiffies32;
1750	return err;
1751	}
1752
1753	void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1754	{
1755	struct sock prev_ssk = NULL, ssk = NULL;
1756	struct mptcp_sock *msk = mptcp_sk(sk);
1757	struct mptcp_sendmsg_info info = {
1758	.flags = flags,
1759	};
1760	bool copied = false;
1761	int push_count = 1;
1762
1763	while (mptcp_send_head(sk) && (push_count > 0)) {
1764	struct mptcp_subflow_context *subflow;
1765	int ret = 0;
1766
1767	if (mptcp_sched_get_send(msk))
1768	break;
1769
1770	push_count = 0;
1771
1772	mptcp_for_each_subflow(msk, subflow) {
1773	if (READ_ONCE(subflow->scheduled)) {
1774	mptcp_subflow_set_scheduled(subflow, false);
1775
1776	prev_ssk = ssk;
1777	ssk = mptcp_subflow_tcp_sock(subflow);
1778	if (ssk != prev_ssk) {
1779	/* First check. If the ssk has changed since
1780	* the last round, release prev_ssk
1781	*/
1782	if (prev_ssk)
1783	mptcp_push_release(prev_ssk, &info);
1784
1785	/* Need to lock the new subflow only if different
1786	* from the previous one, otherwise we are still
1787	* helding the relevant lock
1788	*/
1789	lock_sock(ssk);
1790	}
1791
1792	push_count++;
1793
1794	ret = __subflow_push_pending(sk, ssk, &info);
1795	if (ret <= 0) {
1796	if (ret != -EAGAIN \|\|
1797	(1 << ssk->sk_state) &
1798	(TCPF_FIN_WAIT1 \| TCPF_FIN_WAIT2 \| TCPF_CLOSE))
1799	push_count--;
1800	continue;
1801	}
1802	copied = true;
1803	}
1804	}
1805	}
1806
1807	/* at this point we held the socket lock for the last subflow we used */
1808	if (ssk)
1809	mptcp_push_release(ssk, &info);
1810
1811	/* Avoid scheduling the rtx timer if no data has been pushed; the timer
1812	* will be updated on positive acks by __mptcp_cleanup_una().
1813	*/
1814	if (copied) {
1815	if (!mptcp_rtx_timer_pending(sk))
1816	mptcp_reset_rtx_timer(sk);
1817	mptcp_check_send_data_fin(sk);
1818	}
1819	}
1820
1821	static void __mptcp_subflow_push_pending(struct sock sk, struct sock ssk, bool first)
1822	{
1823	struct mptcp_sock *msk = mptcp_sk(sk);
1824	struct mptcp_sendmsg_info info = {
1825	.data_lock_held = true,
1826	};
1827	bool keep_pushing = true;
1828	struct sock *xmit_ssk;
1829	int copied = 0;
1830
1831	info.flags = 0;
1832	while (mptcp_send_head(sk) && keep_pushing) {
1833	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1834	int ret = 0;
1835
1836	/* check for a different subflow usage only after
1837	* spooling the first chunk of data
1838	*/
1839	if (first) {
1840	mptcp_subflow_set_scheduled(subflow, false);
1841	ret = __subflow_push_pending(sk, ssk, &info);
1842	first = false;
1843	if (ret <= 0)
1844	break;
1845	copied += ret;
1846	continue;
1847	}
1848
1849	if (mptcp_sched_get_send(msk))
1850	goto out;
1851
1852	if (READ_ONCE(subflow->scheduled)) {
1853	mptcp_subflow_set_scheduled(subflow, false);
1854	ret = __subflow_push_pending(sk, ssk, &info);
1855	if (ret <= 0)
1856	keep_pushing = false;
1857	else
1858	copied += ret;
1859	}
1860
1861	mptcp_for_each_subflow(msk, subflow) {
1862	if (READ_ONCE(subflow->scheduled)) {
1863	xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1864	if (xmit_ssk != ssk) {
1865	mptcp_subflow_delegate(subflow,
1866	MPTCP_DELEGATE_SEND);
1867	keep_pushing = false;
1868	}
1869	}
1870	}
1871	}
1872
1873	out:
1874	/* __mptcp_alloc_tx_skb could have released some wmem and we are
1875	* not going to flush it via release_sock()
1876	*/
1877	if (copied) {
1878	tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1879	info.size_goal);
1880	if (!mptcp_rtx_timer_pending(sk))
1881	mptcp_reset_rtx_timer(sk);
1882
1883	if (msk->snd_data_fin_enable &&
1884	msk->snd_nxt + 1 == msk->write_seq)
1885	mptcp_schedule_work(sk);
1886	}
1887	}
1888
1889	static int mptcp_disconnect(struct sock *sk, int flags);
1890
1891	static int mptcp_sendmsg_fastopen(struct sock sk, struct msghdr msg,
1892	size_t len, int *copied_syn)
1893	{
1894	unsigned int saved_flags = msg->msg_flags;
1895	struct mptcp_sock *msk = mptcp_sk(sk);
1896	struct sock *ssk;
1897	int ret;
1898
1899	/* on flags based fastopen the mptcp is supposed to create the
1900	* first subflow right now. Otherwise we are in the defer_connect
1901	* path, and the first subflow must be already present.
1902	* Since the defer_connect flag is cleared after the first succsful
1903	* fastopen attempt, no need to check for additional subflow status.
1904	*/
1905	if (msg->msg_flags & MSG_FASTOPEN) {
1906	ssk = __mptcp_nmpc_sk(msk);
1907	if (IS_ERR(ssk))
1908	return PTR_ERR(ssk);
1909	}
1910	if (!msk->first)
1911	return -EINVAL;
1912
1913	ssk = msk->first;
1914
1915	lock_sock(ssk);
1916	msg->msg_flags \|= MSG_DONTWAIT;
1917	msk->fastopening = 1;
1918	ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1919	msk->fastopening = 0;
1920	msg->msg_flags = saved_flags;
1921	release_sock(ssk);
1922
1923	/* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1924	if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1925	ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1926	msg->msg_namelen, msg->msg_flags, 1);
1927
1928	/* Keep the same behaviour of plain TCP: zero the copied bytes in
1929	* case of any error, except timeout or signal
1930	*/
1931	if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1932	*copied_syn = 0;
1933	} else if (ret && ret != -EINPROGRESS) {
1934	/* The disconnect() op called by tcp_sendmsg_fastopen()/
1935	* __inet_stream_connect() can fail, due to looking check,
1936	* see mptcp_disconnect().
1937	* Attempt it again outside the problematic scope.
1938	*/
1939	if (!mptcp_disconnect(sk, 0)) {
1940	sk->sk_disconnects++;
1941	sk->sk_socket->state = SS_UNCONNECTED;
1942	}
1943	}
1944	inet_clear_bit(DEFER_CONNECT, sk);
1945
1946	return ret;
1947	}
1948
1949	static int do_copy_data_nocache(struct sock *sk, int copy,
1950	struct iov_iter from, char to)
1951	{
1952	if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1953	if (!copy_from_iter_full_nocache(to, copy, from))
1954	return -EFAULT;
1955	} else if (!copy_from_iter_full(to, copy, from)) {
1956	return -EFAULT;
1957	}
1958	return 0;
1959	}
1960
1961	/* open-code sk_stream_memory_free() plus sent limit computation to
1962	* avoid indirect calls in fast-path.
1963	* Called under the msk socket lock, so we can avoid a bunch of ONCE
1964	* annotations.
1965	*/
1966	static u32 mptcp_send_limit(const struct sock *sk)
1967	{
1968	const struct mptcp_sock *msk = mptcp_sk(sk);
1969	u32 limit, not_sent;
1970
1971	if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
1972	return 0;
1973
1974	limit = mptcp_notsent_lowat(sk);
1975	if (limit == UINT_MAX)
1976	return UINT_MAX;
1977
1978	not_sent = msk->write_seq - msk->snd_nxt;
1979	if (not_sent >= limit)
1980	return 0;
1981
1982	return limit - not_sent;
1983	}
1984
1985	static void mptcp_rps_record_subflows(const struct mptcp_sock *msk)
1986	{
1987	struct mptcp_subflow_context *subflow;
1988
1989	if (!rfs_is_needed())
1990	return;
1991
1992	mptcp_for_each_subflow(msk, subflow) {
1993	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1994
1995	sock_rps_record_flow(ssk);
1996	}
1997	}
1998
1999	static int mptcp_sendmsg(struct sock sk, struct msghdr msg, size_t len)
2000	{
2001	struct mptcp_sock *msk = mptcp_sk(sk);
2002	struct page_frag *pfrag;
2003	size_t copied = 0;
2004	int ret = 0;
2005	long timeo;
2006
2007	/* silently ignore everything else */
2008	msg->msg_flags &= MSG_MORE \| MSG_DONTWAIT \| MSG_NOSIGNAL \|
2009	MSG_FASTOPEN \| MSG_EOR;
2010
2011	lock_sock(sk);
2012
2013	mptcp_rps_record_subflows(msk);
2014
2015	if (unlikely(inet_test_bit(DEFER_CONNECT, sk) \|\|
2016	msg->msg_flags & MSG_FASTOPEN)) {
2017	int copied_syn = 0;
2018
2019	ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
2020	copied += copied_syn;
2021	if (ret == -EINPROGRESS && copied_syn > 0)
2022	goto out;
2023	else if (ret)
2024	goto do_error;
2025	}
2026
2027	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
2028
2029	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED \| TCPF_CLOSE_WAIT)) {
2030	ret = sk_stream_wait_connect(sk, &timeo);
2031	if (ret)
2032	goto do_error;
2033	}
2034
2035	ret = -EPIPE;
2036	if (unlikely(sk->sk_err \|\| (sk->sk_shutdown & SEND_SHUTDOWN)))
2037	goto do_error;
2038
2039	pfrag = sk_page_frag(sk);
2040
2041	while (msg_data_left(msg)) {
2042	int total_ts, frag_truesize = 0;
2043	struct mptcp_data_frag *dfrag;
2044	bool dfrag_collapsed;
2045	size_t psize, offset;
2046	u32 copy_limit;
2047
2048	/* ensure fitting the notsent_lowat() constraint */
2049	copy_limit = mptcp_send_limit(sk);
2050	if (!copy_limit)
2051	goto wait_for_memory;
2052
2053	/* reuse tail pfrag, if possible, or carve a new one from the
2054	* page allocator
2055	*/
2056	dfrag = mptcp_pending_tail(sk);
2057	dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
2058	if (!dfrag_collapsed) {
2059	if (!mptcp_page_frag_refill(sk, pfrag))
2060	goto wait_for_memory;
2061
2062	dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
2063	frag_truesize = dfrag->overhead;
2064	}
2065
2066	/* we do not bound vs wspace, to allow a single packet.
2067	* memory accounting will prevent execessive memory usage
2068	* anyway
2069	*/
2070	offset = dfrag->offset + dfrag->data_len;
2071	psize = pfrag->size - offset;
2072	psize = min_t(size_t, psize, msg_data_left(msg));
2073	psize = min_t(size_t, psize, copy_limit);
2074	total_ts = psize + frag_truesize;
2075
2076	if (!sk_wmem_schedule(sk, total_ts))
2077	goto wait_for_memory;
2078
2079	ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
2080	page_address(dfrag->page) + offset);
2081	if (ret)
2082	goto do_error;
2083
2084	/* data successfully copied into the write queue */
2085	sk_forward_alloc_add(sk, -total_ts);
2086	copied += psize;
2087	dfrag->data_len += psize;
2088	frag_truesize += psize;
2089	pfrag->offset += frag_truesize;
2090	WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
2091
2092	/* charge data on mptcp pending queue to the msk socket
2093	* Note: we charge such data both to sk and ssk
2094	*/
2095	sk_wmem_queued_add(sk, frag_truesize);
2096	if (!dfrag_collapsed) {
2097	get_page(dfrag->page);
2098	list_add_tail(&dfrag->list, &msk->rtx_queue);
2099	if (!msk->first_pending)
2100	msk->first_pending = dfrag;
2101	}
2102	pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk,
2103	dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
2104	!dfrag_collapsed);
2105
2106	continue;
2107
2108	wait_for_memory:
2109	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2110	__mptcp_push_pending(sk, msg->msg_flags);
2111	ret = sk_stream_wait_memory(sk, &timeo);
2112	if (ret)
2113	goto do_error;
2114	}
2115
2116	if (copied) {
2117	/* mark the last dfrag with EOR if MSG_EOR was set */
2118	if (msg->msg_flags & MSG_EOR) {
2119	struct mptcp_data_frag *dfrag = mptcp_pending_tail(sk);
2120
2121	if (dfrag)
2122	dfrag->eor = 1;
2123	}
2124	__mptcp_push_pending(sk, msg->msg_flags);
2125	}
2126
2127	out:
2128	release_sock(sk);
2129	return copied;
2130
2131	do_error:
2132	if (copied)
2133	goto out;
2134
2135	copied = sk_stream_error(sk, msg->msg_flags, ret);
2136	goto out;
2137	}
2138
2139	static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied);
2140
2141	static void mptcp_eat_recv_skb(struct sock sk, struct sk_buff skb)
2142	{
2143	/* avoid the indirect call, we know the destructor is sock_rfree */
2144	skb->destructor = NULL;
2145	skb->sk = NULL;
2146	atomic_sub(skb->truesize, &sk->sk_rmem_alloc);
2147	sk_mem_uncharge(sk, skb->truesize);
2148	__skb_unlink(skb, &sk->sk_receive_queue);
2149	skb_attempt_defer_free(skb);
2150	}
2151
2152	static int __mptcp_recvmsg_mskq(struct sock sk, struct msghdr msg,
2153	size_t len, int flags, int copied_total,
2154	struct scm_timestamping_internal *tss,
2155	int cmsg_flags, struct sk_buff *last)
2156	{
2157	struct mptcp_sock *msk = mptcp_sk(sk);
2158	struct sk_buff skb, tmp;
2159	int total_data_len = 0;
2160	int copied = 0;
2161
2162	skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) {
2163	u32 delta, offset = MPTCP_SKB_CB(skb)->offset;
2164	u32 data_len = skb->len - offset;
2165	u32 count;
2166	int err;
2167
2168	if (flags & MSG_PEEK) {
2169	/* skip already peeked skbs */
2170	if (total_data_len + data_len <= copied_total) {
2171	total_data_len += data_len;
2172	*last = skb;
2173	continue;
2174	}
2175
2176	/* skip the already peeked data in the current skb */
2177	delta = copied_total - total_data_len;
2178	offset += delta;
2179	data_len -= delta;
2180	}
2181
2182	count = min_t(size_t, len - copied, data_len);
2183	if (!(flags & MSG_TRUNC)) {
2184	err = skb_copy_datagram_msg(skb, offset, msg, count);
2185	if (unlikely(err < 0)) {
2186	if (!copied)
2187	return err;
2188	break;
2189	}
2190	}
2191
2192	if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
2193	tcp_update_recv_tstamps(skb, tss);
2194	*cmsg_flags \|= MPTCP_CMSG_TS;
2195	}
2196
2197	copied += count;
2198
2199	if (!(flags & MSG_PEEK)) {
2200	msk->bytes_consumed += count;
2201	if (count < data_len) {
2202	MPTCP_SKB_CB(skb)->offset += count;
2203	MPTCP_SKB_CB(skb)->map_seq += count;
2204	break;
2205	}
2206
2207	mptcp_eat_recv_skb(sk, skb);
2208	} else {
2209	*last = skb;
2210	}
2211
2212	if (copied >= len)
2213	break;
2214	}
2215
2216	mptcp_rcv_space_adjust(msk, copied);
2217	return copied;
2218	}
2219
2220	static void mptcp_rcv_space_init(struct mptcp_sock msk, const struct sock ssk)
2221	{
2222	const struct tcp_sock *tp = tcp_sk(ssk);
2223
2224	msk->rcvspace_init = 1;
2225	msk->rcvq_space.copied = 0;
2226
2227	/* initial rcv_space offering made to peer */
2228	msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2229	TCP_INIT_CWND * tp->advmss);
2230	if (msk->rcvq_space.space == 0)
2231	msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2232	}
2233
2234	/* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
2235	*
2236	* Only difference: Use lowest rtt estimate of the subflows in use, see
2237	* mptcp_rcv_rtt_update() and mptcp_rtt_us_est().
2238	*/
2239	static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
2240	{
2241	struct mptcp_subflow_context *subflow;
2242	struct sock sk = (struct sock )msk;
2243	u32 time, rtt_us;
2244	u64 mstamp;
2245
2246	msk_owned_by_me(msk);
2247
2248	if (copied <= 0)
2249	return;
2250
2251	if (!msk->rcvspace_init)
2252	mptcp_rcv_space_init(msk, msk->first);
2253
2254	msk->rcvq_space.copied += copied;
2255
2256	mstamp = mptcp_stamp();
2257	time = tcp_stamp_us_delta(mstamp, READ_ONCE(msk->rcvq_space.time));
2258
2259	rtt_us = mptcp_rtt_us_est(msk);
2260	if (rtt_us == U32_MAX \|\| time < (rtt_us >> 3))
2261	return;
2262
2263	copied = msk->rcvq_space.copied;
2264	copied -= mptcp_inq_hint(sk);
2265	if (copied <= msk->rcvq_space.space)
2266	goto new_measure;
2267
2268	trace_mptcp_rcvbuf_grow(sk, time);
2269	if (mptcp_rcvbuf_grow(sk, copied)) {
2270	/* Make subflows follow along. If we do not do this, we
2271	* get drops at subflow level if skbs can't be moved to
2272	* the mptcp rx queue fast enough (announced rcv_win can
2273	* exceed ssk->sk_rcvbuf).
2274	*/
2275	mptcp_for_each_subflow(msk, subflow) {
2276	struct sock *ssk;
2277	bool slow;
2278
2279	ssk = mptcp_subflow_tcp_sock(subflow);
2280	slow = lock_sock_fast(ssk);
2281	/* subflows can be added before tcp_init_transfer() */
2282	if (tcp_sk(ssk)->rcvq_space.space)
2283	tcp_rcvbuf_grow(ssk, copied);
2284	unlock_sock_fast(ssk, slow);
2285	}
2286	}
2287
2288	new_measure:
2289	msk->rcvq_space.copied = 0;
2290	msk->rcvq_space.time = mstamp;
2291	}
2292
2293	static bool __mptcp_move_skbs(struct sock sk, struct list_head skbs, u32 *delta)
2294	{
2295	struct sk_buff *skb = list_first_entry(skbs, struct sk_buff, list);
2296	struct mptcp_sock *msk = mptcp_sk(sk);
2297	bool moved = false;
2298
2299	while (1) {
2300	prefetch(skb->next);
2301	list_del(&skb->list);
2302	*delta += skb->truesize;
2303
2304	moved \|= __mptcp_move_skb(sk, skb);
2305	if (list_empty(skbs))
2306	break;
2307
2308	skb = list_first_entry(skbs, struct sk_buff, list);
2309	}
2310
2311	__mptcp_ofo_queue(msk);
2312	if (moved)
2313	mptcp_check_data_fin((struct sock *)msk);
2314	return moved;
2315	}
2316
2317	static bool mptcp_can_spool_backlog(struct sock sk, struct list_head skbs)
2318	{
2319	struct mptcp_sock *msk = mptcp_sk(sk);
2320
2321	/* After CG initialization, subflows should never add skb before
2322	* gaining the CG themself.
2323	*/
2324	DEBUG_NET_WARN_ON_ONCE(msk->backlog_unaccounted && sk->sk_socket &&
2325	mem_cgroup_from_sk(sk));
2326
2327	if (list_empty(&msk->backlog_list))
2328	return false;
2329
2330	INIT_LIST_HEAD(skbs);
2331	list_splice_init(&msk->backlog_list, skbs);
2332	return true;
2333	}
2334
2335	static bool mptcp_move_skbs(struct sock *sk)
2336	{
2337	struct mptcp_sock *msk = mptcp_sk(sk);
2338	struct list_head skbs;
2339	bool enqueued = false;
2340	u32 moved = 0;
2341
2342	mptcp_data_lock(sk);
2343	while (mptcp_can_spool_backlog(sk, &skbs)) {
2344	mptcp_data_unlock(sk);
2345	enqueued \|= __mptcp_move_skbs(sk, &skbs, &moved);
2346
2347	mptcp_data_lock(sk);
2348	}
2349	WRITE_ONCE(msk->backlog_len, msk->backlog_len - moved);
2350	mptcp_data_unlock(sk);
2351
2352	if (enqueued && mptcp_epollin_ready(sk))
2353	sk->sk_data_ready(sk);
2354
2355	return enqueued;
2356	}
2357
2358	static unsigned int mptcp_inq_hint(const struct sock *sk)
2359	{
2360	const struct mptcp_sock *msk = mptcp_sk(sk);
2361	const struct sk_buff *skb;
2362
2363	skb = skb_peek(&sk->sk_receive_queue);
2364	if (skb) {
2365	u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq;
2366
2367	if (hint_val >= INT_MAX)
2368	return INT_MAX;
2369
2370	return (unsigned int)hint_val;
2371	}
2372
2373	if (sk->sk_state == TCP_CLOSE \|\| (sk->sk_shutdown & RCV_SHUTDOWN))
2374	return 1;
2375
2376	return 0;
2377	}
2378
2379	static int mptcp_recvmsg(struct sock sk, struct msghdr msg, size_t len,
2380	int flags)
2381	{
2382	struct mptcp_sock *msk = mptcp_sk(sk);
2383	struct scm_timestamping_internal tss;
2384	int copied = 0, cmsg_flags = 0;
2385	int target;
2386	long timeo;
2387
2388	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2389	if (unlikely(flags & MSG_ERRQUEUE))
2390	return inet_recv_error(sk, msg, len);
2391
2392	lock_sock(sk);
2393	if (unlikely(sk->sk_state == TCP_LISTEN)) {
2394	copied = -ENOTCONN;
2395	goto out_err;
2396	}
2397
2398	mptcp_rps_record_subflows(msk);
2399
2400	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2401
2402	len = min_t(size_t, len, INT_MAX);
2403	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2404
2405	if (unlikely(msk->recvmsg_inq))
2406	cmsg_flags = MPTCP_CMSG_INQ;
2407
2408	while (copied < len) {
2409	struct sk_buff *last = NULL;
2410	int err, bytes_read;
2411
2412	bytes_read = __mptcp_recvmsg_mskq(sk, msg, len - copied, flags,
2413	copied, &tss, &cmsg_flags,
2414	&last);
2415	if (unlikely(bytes_read < 0)) {
2416	if (!copied)
2417	copied = bytes_read;
2418	goto out_err;
2419	}
2420
2421	copied += bytes_read;
2422
2423	if (!list_empty(&msk->backlog_list) && mptcp_move_skbs(sk))
2424	continue;
2425
2426	/* only the MPTCP socket status is relevant here. The exit
2427	* conditions mirror closely tcp_recvmsg()
2428	*/
2429	if (copied >= target)
2430	break;
2431
2432	if (copied) {
2433	if (tcp_recv_should_stop(sk) \|\|
2434	!timeo)
2435	break;
2436	} else {
2437	if (sk->sk_err) {
2438	copied = sock_error(sk);
2439	break;
2440	}
2441
2442	if (sk->sk_shutdown & RCV_SHUTDOWN)
2443	break;
2444
2445	if (sk->sk_state == TCP_CLOSE) {
2446	copied = -ENOTCONN;
2447	break;
2448	}
2449
2450	if (!timeo) {
2451	copied = -EAGAIN;
2452	break;
2453	}
2454
2455	if (signal_pending(current)) {
2456	copied = sock_intr_errno(timeo);
2457	break;
2458	}
2459	}
2460
2461	pr_debug("block timeout %ld\n", timeo);
2462	mptcp_cleanup_rbuf(msk, copied);
2463	err = sk_wait_data(sk, &timeo, last);
2464	if (err < 0) {
2465	copied = copied ? : err;
2466	goto out_err;
2467	}
2468	}
2469
2470	mptcp_cleanup_rbuf(msk, copied);
2471
2472	out_err:
2473	if (cmsg_flags && copied >= 0) {
2474	if (cmsg_flags & MPTCP_CMSG_TS)
2475	tcp_recv_timestamp(msg, sk, &tss);
2476
2477	if (cmsg_flags & MPTCP_CMSG_INQ) {
2478	unsigned int inq = mptcp_inq_hint(sk);
2479
2480	put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2481	}
2482	}
2483
2484	pr_debug("msk=%p rx queue empty=%d copied=%d\n",
2485	msk, skb_queue_empty(&sk->sk_receive_queue), copied);
2486
2487	release_sock(sk);
2488	return copied;
2489	}
2490
2491	static void mptcp_retransmit_timer(struct timer_list *t)
2492	{
2493	struct sock *sk = timer_container_of(sk, t, mptcp_retransmit_timer);
2494	struct mptcp_sock *msk = mptcp_sk(sk);
2495
2496	bh_lock_sock(sk);
2497	if (!sock_owned_by_user(sk)) {
2498	/* we need a process context to retransmit */
2499	if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2500	mptcp_schedule_work(sk);
2501	} else {
2502	/* delegate our work to tcp_release_cb() */
2503	__set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2504	}
2505	bh_unlock_sock(sk);
2506	sock_put(sk);
2507	}
2508
2509	static void mptcp_tout_timer(struct timer_list *t)
2510	{
2511	struct inet_connection_sock *icsk =
2512	timer_container_of(icsk, t, mptcp_tout_timer);
2513	struct sock *sk = &icsk->icsk_inet.sk;
2514
2515	mptcp_schedule_work(sk);
2516	sock_put(sk);
2517	}
2518
2519	/* Find an idle subflow. Return NULL if there is unacked data at tcp
2520	* level.
2521	*
2522	* A backup subflow is returned only if that is the only kind available.
2523	*/
2524	struct sock mptcp_subflow_get_retrans(struct mptcp_sock msk)
2525	{
2526	struct sock backup = NULL, pick = NULL;
2527	struct mptcp_subflow_context *subflow;
2528	int min_stale_count = INT_MAX;
2529
2530	mptcp_for_each_subflow(msk, subflow) {
2531	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2532
2533	if (!__mptcp_subflow_active(subflow))
2534	continue;
2535
2536	/* still data outstanding at TCP level? skip this */
2537	if (!tcp_rtx_and_write_queues_empty(ssk)) {
2538	min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2539	continue;
2540	}
2541
2542	if (subflow->backup \|\| subflow->request_bkup) {
2543	if (!backup)
2544	backup = ssk;
2545	continue;
2546	}
2547
2548	if (!pick)
2549	pick = ssk;
2550	}
2551
2552	if (pick)
2553	return pick;
2554
2555	/* use backup only if there are no progresses anywhere */
2556	return min_stale_count > 1 ? backup : NULL;
2557	}
2558
2559	bool __mptcp_retransmit_pending_data(struct sock *sk)
2560	{
2561	struct mptcp_data_frag cur, rtx_head;
2562	struct mptcp_sock *msk = mptcp_sk(sk);
2563
2564	if (__mptcp_check_fallback(msk))
2565	return false;
2566
2567	/* the closing socket has some data untransmitted and/or unacked:
2568	* some data in the mptcp rtx queue has not really xmitted yet.
2569	* keep it simple and re-inject the whole mptcp level rtx queue
2570	*/
2571	mptcp_data_lock(sk);
2572	__mptcp_clean_una_wakeup(sk);
2573	rtx_head = mptcp_rtx_head(sk);
2574	if (!rtx_head) {
2575	mptcp_data_unlock(sk);
2576	return false;
2577	}
2578
2579	msk->recovery_snd_nxt = msk->snd_nxt;
2580	msk->recovery = true;
2581	mptcp_data_unlock(sk);
2582
2583	msk->first_pending = rtx_head;
2584	msk->snd_burst = 0;
2585
2586	/* be sure to clear the "sent status" on all re-injected fragments */
2587	list_for_each_entry(cur, &msk->rtx_queue, list) {
2588	if (!cur->already_sent)
2589	break;
2590	cur->already_sent = 0;
2591	}
2592
2593	return true;
2594	}
2595
2596	/* flags for __mptcp_close_ssk() */
2597	#define MPTCP_CF_PUSH BIT(1)
2598
2599	/* be sure to send a reset only if the caller asked for it, also
2600	* clean completely the subflow status when the subflow reaches
2601	* TCP_CLOSE state
2602	*/
2603	static void __mptcp_subflow_disconnect(struct sock *ssk,
2604	struct mptcp_subflow_context *subflow,
2605	bool fastclosing)
2606	{
2607	if (((1 << ssk->sk_state) & (TCPF_CLOSE \| TCPF_LISTEN)) \|\|
2608	fastclosing) {
2609	/* The MPTCP code never wait on the subflow sockets, TCP-level
2610	* disconnect should never fail
2611	*/
2612	WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2613	mptcp_subflow_ctx_reset(subflow);
2614	} else {
2615	tcp_shutdown(ssk, SEND_SHUTDOWN);
2616	}
2617	}
2618
2619	static void mptcp_cleanup_ssk_backlog(struct sock sk, struct sock ssk)
2620	{
2621	struct mptcp_sock *msk = mptcp_sk(sk);
2622	struct sk_buff *skb;
2623
2624	mptcp_data_lock(sk);
2625	list_for_each_entry(skb, &msk->backlog_list, list) {
2626	if (skb->sk != ssk)
2627	continue;
2628
2629	atomic_sub(skb->truesize, &skb->sk->sk_rmem_alloc);
2630	skb->sk = NULL;
2631	}
2632	mptcp_data_unlock(sk);
2633	}
2634
2635	/* subflow sockets can be either outgoing (connect) or incoming
2636	* (accept).
2637	*
2638	* Outgoing subflows use in-kernel sockets.
2639	* Incoming subflows do not have their own 'struct socket' allocated,
2640	* so we need to use tcp_close() after detaching them from the mptcp
2641	* parent socket.
2642	*/
2643	static void __mptcp_close_ssk(struct sock sk, struct sock ssk,
2644	struct mptcp_subflow_context *subflow,
2645	unsigned int flags)
2646	{
2647	struct mptcp_sock *msk = mptcp_sk(sk);
2648	bool dispose_it, need_push = false;
2649	int fwd_remaining;
2650
2651	/* Do not pass RX data to the msk, even if the subflow socket is not
2652	* going to be freed (i.e. even for the first subflow on graceful
2653	* subflow close.
2654	*/
2655	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2656	subflow->closing = 1;
2657
2658	if (flags & MPTCP_CF_PUSH)
2659	mptcp_cleanup_ssk_backlog(sk, ssk);
2660
2661	/* Borrow the fwd allocated page left-over; fwd memory for the subflow
2662	* could be negative at this point, but will be reach zero soon - when
2663	* the data allocated using such fragment will be freed.
2664	*/
2665	if (subflow->lent_mem_frag) {
2666	fwd_remaining = PAGE_SIZE - subflow->lent_mem_frag;
2667	sk_forward_alloc_add(sk, fwd_remaining);
2668	sk_forward_alloc_add(ssk, -fwd_remaining);
2669	subflow->lent_mem_frag = 0;
2670	}
2671
2672	/* If the first subflow moved to a close state before accept, e.g. due
2673	* to an incoming reset or listener shutdown, the subflow socket is
2674	* already deleted by inet_child_forget() and the mptcp socket can't
2675	* survive too.
2676	*/
2677	if (msk->in_accept_queue && msk->first == ssk &&
2678	(sock_flag(sk, SOCK_DEAD) \|\| sock_flag(ssk, SOCK_DEAD))) {
2679	/* ensure later check in mptcp_worker() will dispose the msk */
2680	sock_set_flag(sk, SOCK_DEAD);
2681	mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2682	mptcp_subflow_drop_ctx(ssk);
2683	goto out_release;
2684	}
2685
2686	dispose_it = msk->free_first \|\| ssk != msk->first;
2687	if (dispose_it)
2688	list_del(&subflow->node);
2689
2690	if (subflow->send_fastclose && ssk->sk_state != TCP_CLOSE)
2691	tcp_set_state(ssk, TCP_CLOSE);
2692
2693	need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2694	if (!dispose_it) {
2695	__mptcp_subflow_disconnect(ssk, subflow, msk->fastclosing);
2696	release_sock(ssk);
2697
2698	goto out;
2699	}
2700
2701	subflow->disposable = 1;
2702
2703	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2704	* the ssk has been already destroyed, we just need to release the
2705	* reference owned by msk;
2706	*/
2707	if (!inet_csk(ssk)->icsk_ulp_ops) {
2708	WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2709	kfree_rcu(subflow, rcu);
2710	} else {
2711	/* otherwise tcp will dispose of the ssk and subflow ctx */
2712	__tcp_close(ssk, 0);
2713
2714	/* close acquired an extra ref */
2715	__sock_put(ssk);
2716	}
2717
2718	out_release:
2719	__mptcp_subflow_error_report(sk, ssk);
2720	release_sock(ssk);
2721
2722	sock_put(ssk);
2723
2724	if (ssk == msk->first)
2725	WRITE_ONCE(msk->first, NULL);
2726
2727	out:
2728	__mptcp_sync_sndbuf(sk);
2729	if (need_push)
2730	__mptcp_push_pending(sk, 0);
2731
2732	/* Catch every 'all subflows closed' scenario, including peers silently
2733	* closing them, e.g. due to timeout.
2734	* For established sockets, allow an additional timeout before closing,
2735	* as the protocol can still create more subflows.
2736	*/
2737	if (list_is_singular(&msk->conn_list) && msk->first &&
2738	inet_sk_state_load(msk->first) == TCP_CLOSE) {
2739	if (sk->sk_state != TCP_ESTABLISHED \|\|
2740	msk->in_accept_queue \|\| sock_flag(sk, SOCK_DEAD)) {
2741	mptcp_set_state(sk, TCP_CLOSE);
2742	mptcp_close_wake_up(sk);
2743	} else {
2744	mptcp_start_tout_timer(sk);
2745	}
2746	}
2747	}
2748
2749	void mptcp_close_ssk(struct sock sk, struct sock ssk,
2750	struct mptcp_subflow_context *subflow)
2751	{
2752	/* The first subflow can already be closed or disconnected */
2753	if (subflow->close_event_done \|\| READ_ONCE(subflow->local_id) < 0)
2754	return;
2755
2756	subflow->close_event_done = true;
2757
2758	if (sk->sk_state == TCP_ESTABLISHED)
2759	mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2760
2761	/* subflow aborted before reaching the fully_established status
2762	* attempt the creation of the next subflow
2763	*/
2764	mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2765
2766	__mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2767	}
2768
2769	static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2770	{
2771	return 0;
2772	}
2773
2774	static void __mptcp_close_subflow(struct sock *sk)
2775	{
2776	struct mptcp_subflow_context subflow, tmp;
2777	struct mptcp_sock *msk = mptcp_sk(sk);
2778
2779	might_sleep();
2780
2781	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2782	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2783	int ssk_state = inet_sk_state_load(ssk);
2784
2785	if (ssk_state != TCP_CLOSE &&
2786	(ssk_state != TCP_CLOSE_WAIT \|\|
2787	inet_sk_state_load(sk) != TCP_ESTABLISHED \|\|
2788	__mptcp_check_fallback(msk)))
2789	continue;
2790
2791	/* 'subflow_data_ready' will re-sched once rx queue is empty */
2792	if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2793	continue;
2794
2795	mptcp_close_ssk(sk, ssk, subflow);
2796	}
2797
2798	}
2799
2800	static bool mptcp_close_tout_expired(const struct sock *sk)
2801	{
2802	if (!inet_csk(sk)->icsk_mtup.probe_timestamp \|\|
2803	sk->sk_state == TCP_CLOSE)
2804	return false;
2805
2806	return time_after32(tcp_jiffies32,
2807	inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2808	}
2809
2810	static void mptcp_check_fastclose(struct mptcp_sock *msk)
2811	{
2812	struct mptcp_subflow_context subflow, tmp;
2813	struct sock sk = (struct sock )msk;
2814
2815	if (likely(!READ_ONCE(msk->rcv_fastclose)))
2816	return;
2817
2818	mptcp_token_destroy(msk);
2819
2820	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2821	struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2822	bool slow;
2823
2824	slow = lock_sock_fast(tcp_sk);
2825	if (tcp_sk->sk_state != TCP_CLOSE) {
2826	mptcp_send_active_reset_reason(tcp_sk);
2827	tcp_set_state(tcp_sk, TCP_CLOSE);
2828	}
2829	unlock_sock_fast(tcp_sk, slow);
2830	}
2831
2832	/* Mirror the tcp_reset() error propagation */
2833	switch (sk->sk_state) {
2834	case TCP_SYN_SENT:
2835	WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2836	break;
2837	case TCP_CLOSE_WAIT:
2838	WRITE_ONCE(sk->sk_err, EPIPE);
2839	break;
2840	case TCP_CLOSE:
2841	return;
2842	default:
2843	WRITE_ONCE(sk->sk_err, ECONNRESET);
2844	}
2845
2846	mptcp_set_state(sk, TCP_CLOSE);
2847	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2848	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2849	set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2850
2851	/* the calling mptcp_worker will properly destroy the socket */
2852	if (sock_flag(sk, SOCK_DEAD))
2853	return;
2854
2855	sk->sk_state_change(sk);
2856	sk_error_report(sk);
2857	}
2858
2859	/*
2860	* Retransmit the specified data fragment on all the selected subflows,
2861	* starting from the specified sequence
2862	*/
2863	static int __mptcp_push_retrans(struct sock sk, struct mptcp_data_frag dfrag,
2864	u64 sent_seq)
2865	{
2866	struct mptcp_sendmsg_info info = { .data_lock_held = true, };
2867	struct mptcp_sock *msk = mptcp_sk(sk);
2868	struct mptcp_subflow_context *subflow;
2869	struct sock *ssk;
2870	int ret, len = 0;
2871
2872	mptcp_for_each_subflow(msk, subflow) {
2873	if (READ_ONCE(subflow->scheduled)) {
2874	u16 offset = sent_seq - dfrag->data_seq;
2875	u16 copied = 0;
2876
2877	mptcp_subflow_set_scheduled(subflow, false);
2878
2879	ssk = mptcp_subflow_tcp_sock(subflow);
2880
2881	lock_sock(ssk);
2882
2883	/* limit retransmission to the bytes already sent on some subflows */
2884	info.sent = offset;
2885	info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2886	dfrag->already_sent;
2887
2888	/*
2889	* make the whole retrans decision, xmit, disallow
2890	* fallback atomic, note that we can't retrans even
2891	* when an infinite fallback is in progress, i.e. new
2892	* subflows are disallowed.
2893	*/
2894	spin_lock_bh(&msk->fallback_lock);
2895	if (__mptcp_check_fallback(msk) \|\|
2896	!msk->allow_subflows) {
2897	spin_unlock_bh(&msk->fallback_lock);
2898	release_sock(ssk);
2899	return -1;
2900	}
2901
2902	while (info.sent < info.limit) {
2903	ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2904	if (ret <= 0)
2905	break;
2906
2907	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2908	copied += ret;
2909	info.sent += ret;
2910	}
2911	if (copied) {
2912	len = max(copied, len);
2913	tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2914	info.size_goal);
2915	msk->allow_infinite_fallback = false;
2916	}
2917	spin_unlock_bh(&msk->fallback_lock);
2918
2919	release_sock(ssk);
2920	}
2921	}
2922	return len;
2923	}
2924
2925	static void __mptcp_retrans(struct sock *sk)
2926	{
2927	struct mptcp_sock *msk = mptcp_sk(sk);
2928	struct mptcp_subflow_context *subflow;
2929	struct mptcp_data_frag *dfrag;
2930	u64 retrans_seq, sent_seq;
2931	bool need_retrans;
2932	int err, len;
2933
2934	mptcp_pm_chk_stale(msk);
2935
2936	/* Get an updated and consistent rtx queue status. */
2937	mptcp_data_lock(sk);
2938	__mptcp_clean_una_wakeup(sk);
2939	retrans_seq = msk->snd_una;
2940	dfrag = mptcp_rtx_head(sk);
2941	need_retrans = !!dfrag;
2942	mptcp_data_unlock(sk);
2943
2944	for (;;) {
2945	bool already_acked;
2946
2947	err = mptcp_sched_get_retrans(msk);
2948	if (err)
2949	break;
2950
2951	/* `already_sent` can be 0 for `dfrag` belonging to the RTX
2952	* queue due to __mptcp_retransmit_pending_data().
2953	*/
2954	if (!dfrag \|\| !dfrag->already_sent)
2955	break;
2956
2957	/* Can fail only in case of fallback. */
2958	len = __mptcp_push_retrans(sk, dfrag, retrans_seq);
2959	if (len < 0)
2960	goto clear_scheduled;
2961
2962	retrans_seq += len;
2963	msk->bytes_retrans += len;
2964	dfrag->already_sent = max_t(u16, dfrag->already_sent,
2965	retrans_seq - dfrag->data_seq);
2966
2967	/* With csum enabled, retransmission can send new data. */
2968	sent_seq = dfrag->already_sent + dfrag->data_seq;
2969	if (after64(sent_seq, msk->snd_nxt))
2970	WRITE_ONCE(msk->snd_nxt, sent_seq);
2971
2972	/* Attempt the next fragment only if the current one is
2973	* completely retransmitted.
2974	*/
2975	if (before64(retrans_seq, dfrag->data_seq + dfrag->data_len))
2976	break;
2977
2978	dfrag = list_is_last(&dfrag->list, &msk->rtx_queue) ?
2979	NULL : list_next_entry(dfrag, list);
2980	if (!dfrag)
2981	break;
2982
2983	/* Incoming acks can move snd_una after the current dfrag
2984	* across loop iterations, if so start again from RTX head.
2985	*/
2986	mptcp_data_lock(sk);
2987	already_acked = !before64(msk->snd_una, dfrag->data_seq +
2988	dfrag->already_sent);
2989	if (already_acked) {
2990	__mptcp_clean_una_wakeup(sk);
2991	retrans_seq = msk->snd_una;
2992	dfrag = mptcp_rtx_head(sk);
2993	need_retrans = !!dfrag;
2994	} else if (after64(msk->snd_una, retrans_seq)) {
2995	retrans_seq = msk->snd_una;
2996	}
2997	mptcp_data_unlock(sk);
2998	}
2999
3000	/* Attempt data-fin retransmission only when the RTX queue is empty. */
3001	if (!need_retrans) {
3002	if (mptcp_data_fin_enabled(msk)) {
3003	struct inet_connection_sock *icsk = inet_csk(sk);
3004
3005	WRITE_ONCE(icsk->icsk_retransmits,
3006	icsk->icsk_retransmits + 1);
3007	mptcp_set_datafin_timeout(sk);
3008	mptcp_send_ack(msk);
3009	goto reset_timer;
3010	}
3011
3012	if (!mptcp_send_head(sk))
3013	goto clear_scheduled;
3014	}
3015
3016	reset_timer:
3017	mptcp_check_and_set_pending(sk);
3018
3019	if (!mptcp_rtx_timer_pending(sk))
3020	mptcp_reset_rtx_timer(sk);
3021
3022	clear_scheduled:
3023	/* If no rtx data was available or in case of fallback, there
3024	* could be left-over scheduled subflows; clear them all
3025	* or later xmit could use bad ones
3026	*/
3027	mptcp_for_each_subflow(msk, subflow)
3028	if (READ_ONCE(subflow->scheduled))
3029	mptcp_subflow_set_scheduled(subflow, false);
3030	}
3031
3032	/* schedule the timeout timer for the relevant event: either close timeout
3033	* or mp_fail timeout. The close timeout takes precedence on the mp_fail one
3034	*/
3035	void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
3036	{
3037	struct sock sk = (struct sock )msk;
3038	unsigned long timeout, close_timeout;
3039
3040	if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
3041	return;
3042
3043	close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp -
3044	tcp_jiffies32 + jiffies + mptcp_close_timeout(sk);
3045
3046	/* the close timeout takes precedence on the fail one, and here at least one of
3047	* them is active
3048	*/
3049	timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
3050
3051	sk_reset_timer(sk, &inet_csk(sk)->mptcp_tout_timer, timeout);
3052	}
3053
3054	static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
3055	{
3056	struct sock *ssk = msk->first;
3057	bool slow;
3058
3059	if (!ssk)
3060	return;
3061
3062	pr_debug("MP_FAIL doesn't respond, reset the subflow\n");
3063
3064	slow = lock_sock_fast(ssk);
3065	mptcp_subflow_reset(ssk);
3066	WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
3067	unlock_sock_fast(ssk, slow);
3068	}
3069
3070	static void mptcp_backlog_purge(struct sock *sk)
3071	{
3072	struct mptcp_sock *msk = mptcp_sk(sk);
3073	struct sk_buff tmp, skb;
3074	LIST_HEAD(backlog);
3075
3076	mptcp_data_lock(sk);
3077	list_splice_init(&msk->backlog_list, &backlog);
3078	msk->backlog_len = 0;
3079	mptcp_data_unlock(sk);
3080
3081	list_for_each_entry_safe(skb, tmp, &backlog, list) {
3082	mptcp_borrow_fwdmem(sk, skb);
3083	kfree_skb_reason(skb, SKB_DROP_REASON_SOCKET_CLOSE);
3084	}
3085	sk_mem_reclaim(sk);
3086	}
3087
3088	static void mptcp_do_fastclose(struct sock *sk)
3089	{
3090	struct mptcp_subflow_context subflow, tmp;
3091	struct mptcp_sock *msk = mptcp_sk(sk);
3092
3093	mptcp_set_state(sk, TCP_CLOSE);
3094	mptcp_backlog_purge(sk);
3095	msk->fastclosing = 1;
3096
3097	/* Explicitly send the fastclose reset as need */
3098	if (__mptcp_check_fallback(msk))
3099	return;
3100
3101	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
3102	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3103
3104	lock_sock(ssk);
3105
3106	/* Some subflow socket states don't allow/need a reset.*/
3107	if ((1 << ssk->sk_state) & (TCPF_LISTEN \| TCPF_CLOSE))
3108	goto unlock;
3109
3110	subflow->send_fastclose = 1;
3111
3112	/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3113	* issue in __tcp_select_window(), see tcp_disconnect().
3114	*/
3115	inet_csk(ssk)->icsk_ack.rcv_mss = TCP_MIN_MSS;
3116
3117	tcp_send_active_reset(ssk, SK_RST_REASON_TCP_ABORT_ON_CLOSE);
3118	unlock:
3119	release_sock(ssk);
3120	}
3121	}
3122
3123	static void mptcp_worker(struct work_struct *work)
3124	{
3125	struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
3126	struct sock sk = (struct sock )msk;
3127	unsigned long fail_tout;
3128	int state;
3129
3130	lock_sock(sk);
3131	state = sk->sk_state;
3132	if (unlikely((1 << state) & (TCPF_CLOSE \| TCPF_LISTEN)))
3133	goto unlock;
3134
3135	mptcp_check_fastclose(msk);
3136
3137	mptcp_pm_worker(msk);
3138
3139	mptcp_check_send_data_fin(sk);
3140	mptcp_check_data_fin_ack(sk);
3141	mptcp_check_data_fin(sk);
3142
3143	if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
3144	__mptcp_close_subflow(sk);
3145
3146	if (mptcp_close_tout_expired(sk)) {
3147	struct mptcp_subflow_context subflow, tmp;
3148
3149	mptcp_do_fastclose(sk);
3150	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3151	__mptcp_close_ssk(sk, subflow->tcp_sock, subflow, 0);
3152	mptcp_close_wake_up(sk);
3153	}
3154
3155	if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
3156	__mptcp_destroy_sock(sk);
3157	goto unlock;
3158	}
3159
3160	if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
3161	__mptcp_retrans(sk);
3162
3163	fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
3164	if (fail_tout && time_after(jiffies, fail_tout))
3165	mptcp_mp_fail_no_response(msk);
3166
3167	unlock:
3168	release_sock(sk);
3169	sock_put(sk);
3170	}
3171
3172	static void __mptcp_init_sock(struct sock *sk)
3173	{
3174	struct inet_connection_sock *icsk = inet_csk(sk);
3175	struct mptcp_sock *msk = mptcp_sk(sk);
3176	struct net *net = sock_net(sk);
3177
3178	INIT_LIST_HEAD(&msk->conn_list);
3179	INIT_LIST_HEAD(&msk->join_list);
3180	INIT_LIST_HEAD(&msk->rtx_queue);
3181	INIT_LIST_HEAD(&msk->backlog_list);
3182	INIT_WORK(&msk->work, mptcp_worker);
3183	msk->out_of_order_queue = RB_ROOT;
3184	msk->first_pending = NULL;
3185
3186	/* msk does not go through tcp_init_sock(); seed RTO bounds. */
3187	icsk->icsk_rto_min =
3188	usecs_to_jiffies(READ_ONCE(net->ipv4.sysctl_tcp_rto_min_us));
3189	icsk->icsk_rto_max =
3190	msecs_to_jiffies(READ_ONCE(net->ipv4.sysctl_tcp_rto_max_ms));
3191	msk->timer_ival = icsk->icsk_rto_min;
3192	msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
3193	msk->backlog_len = 0;
3194	mptcp_init_rtt_est(msk);
3195
3196	WRITE_ONCE(msk->first, NULL);
3197	inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
3198	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3199	msk->allow_infinite_fallback = true;
3200	msk->allow_subflows = true;
3201	msk->recovery = false;
3202	msk->subflow_id = 1;
3203	msk->last_data_sent = tcp_jiffies32;
3204	msk->last_data_recv = tcp_jiffies32;
3205	msk->last_ack_recv = tcp_jiffies32;
3206
3207	mptcp_pm_data_init(msk);
3208	spin_lock_init(&msk->fallback_lock);
3209
3210	/* re-use the csk retrans timer for MPTCP-level retrans */
3211	timer_setup(&sk->mptcp_retransmit_timer, mptcp_retransmit_timer, 0);
3212	timer_setup(&msk->sk.mptcp_tout_timer, mptcp_tout_timer, 0);
3213	}
3214
3215	static void mptcp_ca_reset(struct sock *sk)
3216	{
3217	struct inet_connection_sock *icsk = inet_csk(sk);
3218
3219	tcp_assign_congestion_control(sk);
3220	strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name,
3221	sizeof(mptcp_sk(sk)->ca_name));
3222
3223	/* no need to keep a reference to the ops, the name will suffice */
3224	tcp_cleanup_congestion_control(sk);
3225	icsk->icsk_ca_ops = NULL;
3226	}
3227
3228	static int mptcp_init_sock(struct sock *sk)
3229	{
3230	struct net *net = sock_net(sk);
3231	int ret;
3232
3233	__mptcp_init_sock(sk);
3234
3235	if (!mptcp_is_enabled(net))
3236	return -ENOPROTOOPT;
3237
3238	if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
3239	return -ENOMEM;
3240
3241	rcu_read_lock();
3242	ret = mptcp_init_sched(mptcp_sk(sk),
3243	mptcp_sched_find(mptcp_get_scheduler(net)));
3244	rcu_read_unlock();
3245	if (ret)
3246	return ret;
3247
3248	set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
3249
3250	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
3251	* propagate the correct value
3252	*/
3253	mptcp_ca_reset(sk);
3254
3255	sk_sockets_allocated_inc(sk);
3256	sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
3257	sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
3258	sk->sk_write_space = sk_stream_write_space;
3259
3260	return 0;
3261	}
3262
3263	static void __mptcp_clear_xmit(struct sock *sk)
3264	{
3265	struct mptcp_sock *msk = mptcp_sk(sk);
3266	struct mptcp_data_frag dtmp, dfrag;
3267
3268	msk->first_pending = NULL;
3269	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
3270	dfrag_clear(sk, dfrag);
3271	}
3272
3273	void mptcp_cancel_work(struct sock *sk)
3274	{
3275	struct mptcp_sock *msk = mptcp_sk(sk);
3276
3277	if (cancel_work_sync(&msk->work))
3278	__sock_put(sk);
3279	}
3280
3281	void mptcp_subflow_shutdown(struct sock sk, struct sock ssk, int how)
3282	{
3283	lock_sock(ssk);
3284
3285	switch (ssk->sk_state) {
3286	case TCP_LISTEN:
3287	if (!(how & RCV_SHUTDOWN))
3288	break;
3289	fallthrough;
3290	case TCP_SYN_SENT:
3291	WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
3292	break;
3293	default:
3294	if (__mptcp_check_fallback(mptcp_sk(sk))) {
3295	pr_debug("Fallback\n");
3296	ssk->sk_shutdown \|= how;
3297	tcp_shutdown(ssk, how);
3298
3299	/* simulate the data_fin ack reception to let the state
3300	* machine move forward
3301	*/
3302	WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
3303	mptcp_schedule_work(sk);
3304	} else {
3305	pr_debug("Sending DATA_FIN on subflow %p\n", ssk);
3306	tcp_send_ack(ssk);
3307	if (!mptcp_rtx_timer_pending(sk))
3308	mptcp_reset_rtx_timer(sk);
3309	}
3310	break;
3311	}
3312
3313	release_sock(ssk);
3314	}
3315
3316	void mptcp_set_state(struct sock *sk, int state)
3317	{
3318	int oldstate = sk->sk_state;
3319
3320	switch (state) {
3321	case TCP_ESTABLISHED:
3322	if (oldstate != TCP_ESTABLISHED)
3323	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
3324	break;
3325	case TCP_CLOSE_WAIT:
3326	/* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state:
3327	* MPTCP "accepted" sockets will be created later on. So no
3328	* transition from TCP_SYN_RECV to TCP_CLOSE_WAIT.
3329	*/
3330	break;
3331	case TCP_CLOSE:
3332	clear_bit(MPTCP_RTX_ENABLED, &mptcp_sk(sk)->flags);
3333	fallthrough;
3334	default:
3335	if (oldstate == TCP_ESTABLISHED \|\| oldstate == TCP_CLOSE_WAIT)
3336	MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
3337	}
3338
3339	inet_sk_state_store(sk, state);
3340	}
3341
3342	static const unsigned char new_state[16] = {
3343	/* current state: new state: action: */
3344	[0 /* (Invalid) */] = TCP_CLOSE,
3345	[TCP_ESTABLISHED] = TCP_FIN_WAIT1 \| TCP_ACTION_FIN,
3346	[TCP_SYN_SENT] = TCP_CLOSE,
3347	[TCP_SYN_RECV] = TCP_FIN_WAIT1 \| TCP_ACTION_FIN,
3348	[TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
3349	[TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
3350	[TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
3351	[TCP_CLOSE] = TCP_CLOSE,
3352	[TCP_CLOSE_WAIT] = TCP_LAST_ACK \| TCP_ACTION_FIN,
3353	[TCP_LAST_ACK] = TCP_LAST_ACK,
3354	[TCP_LISTEN] = TCP_CLOSE,
3355	[TCP_CLOSING] = TCP_CLOSING,
3356	[TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
3357	};
3358
3359	static int mptcp_close_state(struct sock *sk)
3360	{
3361	int next = (int)new_state[sk->sk_state];
3362	int ns = next & TCP_STATE_MASK;
3363
3364	mptcp_set_state(sk, ns);
3365
3366	return next & TCP_ACTION_FIN;
3367	}
3368
3369	static void mptcp_check_send_data_fin(struct sock *sk)
3370	{
3371	struct mptcp_subflow_context *subflow;
3372	struct mptcp_sock *msk = mptcp_sk(sk);
3373
3374	pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n",
3375	msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
3376	msk->snd_nxt, msk->write_seq);
3377
3378	/* we still need to enqueue subflows or not really shutting down,
3379	* skip this
3380	*/
3381	if (!msk->snd_data_fin_enable \|\| msk->snd_nxt + 1 != msk->write_seq \|\|
3382	mptcp_send_head(sk))
3383	return;
3384
3385	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3386
3387	mptcp_for_each_subflow(msk, subflow) {
3388	struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
3389
3390	mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
3391	}
3392	}
3393
3394	static void __mptcp_wr_shutdown(struct sock *sk)
3395	{
3396	struct mptcp_sock *msk = mptcp_sk(sk);
3397
3398	pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n",
3399	msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
3400	!!mptcp_send_head(sk));
3401
3402	/* will be ignored by fallback sockets */
3403	WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
3404	WRITE_ONCE(msk->snd_data_fin_enable, 1);
3405
3406	mptcp_check_send_data_fin(sk);
3407	}
3408
3409	static void __mptcp_destroy_sock(struct sock *sk)
3410	{
3411	struct mptcp_sock *msk = mptcp_sk(sk);
3412
3413	pr_debug("msk=%p\n", msk);
3414
3415	might_sleep();
3416
3417	mptcp_stop_rtx_timer(sk);
3418	sk_stop_timer(sk, &inet_csk(sk)->mptcp_tout_timer);
3419	msk->pm.status = 0;
3420	mptcp_release_sched(msk);
3421
3422	sk->sk_prot->destroy(sk);
3423
3424	sk_stream_kill_queues(sk);
3425	xfrm_sk_free_policy(sk);
3426
3427	sock_put(sk);
3428	}
3429
3430	void __mptcp_unaccepted_force_close(struct sock *sk)
3431	{
3432	sock_set_flag(sk, SOCK_DEAD);
3433	mptcp_do_fastclose(sk);
3434	__mptcp_destroy_sock(sk);
3435	}
3436
3437	static __poll_t mptcp_check_readable(struct sock *sk)
3438	{
3439	return mptcp_epollin_ready(sk) ? EPOLLIN \| EPOLLRDNORM : 0;
3440	}
3441
3442	static void mptcp_check_listen_stop(struct sock *sk)
3443	{
3444	struct sock *ssk;
3445
3446	if (inet_sk_state_load(sk) != TCP_LISTEN)
3447	return;
3448
3449	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3450	ssk = mptcp_sk(sk)->first;
3451	if (WARN_ON_ONCE(!ssk \|\| inet_sk_state_load(ssk) != TCP_LISTEN))
3452	return;
3453
3454	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3455	tcp_set_state(ssk, TCP_CLOSE);
3456	mptcp_subflow_queue_clean(sk, ssk);
3457	inet_csk_listen_stop(ssk);
3458	mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3459	release_sock(ssk);
3460	}
3461
3462	bool __mptcp_close(struct sock *sk, long timeout)
3463	{
3464	struct mptcp_subflow_context *subflow;
3465	struct mptcp_sock *msk = mptcp_sk(sk);
3466	bool do_cancel_work = false;
3467	int subflows_alive = 0;
3468
3469	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3470
3471	if ((1 << sk->sk_state) & (TCPF_LISTEN \| TCPF_CLOSE)) {
3472	mptcp_check_listen_stop(sk);
3473	mptcp_set_state(sk, TCP_CLOSE);
3474	goto cleanup;
3475	}
3476
3477	if (mptcp_data_avail(msk) \|\| timeout < 0 \|\|
3478	(sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime)) {
3479	/* If the msk has read data, or the caller explicitly ask it,
3480	* do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3481	*/
3482	mptcp_do_fastclose(sk);
3483	timeout = 0;
3484	} else if (mptcp_close_state(sk)) {
3485	__mptcp_wr_shutdown(sk);
3486	}
3487
3488	sk_stream_wait_close(sk, timeout);
3489
3490	cleanup:
3491	/* orphan all the subflows */
3492	mptcp_for_each_subflow(msk, subflow) {
3493	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3494	bool slow = lock_sock_fast_nested(ssk);
3495
3496	subflows_alive += ssk->sk_state != TCP_CLOSE;
3497
3498	/* since the close timeout takes precedence on the fail one,
3499	* cancel the latter
3500	*/
3501	if (ssk == msk->first)
3502	subflow->fail_tout = 0;
3503
3504	/* detach from the parent socket, but allow data_ready to
3505	* push incoming data into the mptcp stack, to properly ack it
3506	*/
3507	ssk->sk_socket = NULL;
3508	ssk->sk_wq = NULL;
3509	unlock_sock_fast(ssk, slow);
3510	}
3511	sock_orphan(sk);
3512
3513	/* all the subflows are closed, only timeout can change the msk
3514	* state, let's not keep resources busy for no reasons
3515	*/
3516	if (subflows_alive == 0)
3517	mptcp_set_state(sk, TCP_CLOSE);
3518
3519	sock_hold(sk);
3520	pr_debug("msk=%p state=%d\n", sk, sk->sk_state);
3521	mptcp_pm_connection_closed(msk);
3522
3523	if (sk->sk_state == TCP_CLOSE) {
3524	__mptcp_destroy_sock(sk);
3525	do_cancel_work = true;
3526	} else {
3527	mptcp_start_tout_timer(sk);
3528	}
3529
3530	return do_cancel_work;
3531	}
3532
3533	static void mptcp_close(struct sock *sk, long timeout)
3534	{
3535	bool do_cancel_work;
3536
3537	lock_sock(sk);
3538
3539	do_cancel_work = __mptcp_close(sk, timeout);
3540	release_sock(sk);
3541	if (do_cancel_work)
3542	mptcp_cancel_work(sk);
3543
3544	sock_put(sk);
3545	}
3546
3547	static void mptcp_copy_inaddrs(struct sock msk, const struct sock ssk)
3548	{
3549	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3550	const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3551	struct ipv6_pinfo *msk6 = inet6_sk(msk);
3552
3553	msk->sk_v6_daddr = ssk->sk_v6_daddr;
3554	msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3555
3556	if (msk6 && ssk6) {
3557	msk6->saddr = ssk6->saddr;
3558	msk6->flow_label = ssk6->flow_label;
3559	}
3560	#endif
3561
3562	inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3563	inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3564	inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3565	inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3566	inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3567	inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3568	}
3569
3570	static void mptcp_destroy_common(struct mptcp_sock *msk)
3571	{
3572	struct mptcp_subflow_context subflow, tmp;
3573	struct sock sk = (struct sock )msk;
3574
3575	__mptcp_clear_xmit(sk);
3576	mptcp_backlog_purge(sk);
3577
3578	/* join list will be eventually flushed (with rst) at sock lock release time */
3579	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3580	__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, 0);
3581
3582	__skb_queue_purge(&sk->sk_receive_queue);
3583	skb_rbtree_purge(&msk->out_of_order_queue);
3584
3585	/* move all the rx fwd alloc into the sk_mem_reclaim_final in
3586	* inet_sock_destruct() will dispose it
3587	*/
3588	mptcp_token_destroy(msk);
3589	mptcp_pm_destroy(msk);
3590	}
3591
3592	static int mptcp_disconnect(struct sock *sk, int flags)
3593	{
3594	struct inet_connection_sock *icsk = inet_csk(sk);
3595	struct mptcp_sock *msk = mptcp_sk(sk);
3596
3597	/* We are on the fastopen error path. We can't call straight into the
3598	* subflows cleanup code due to lock nesting (we are already under
3599	* msk->firstsocket lock).
3600	*/
3601	if (msk->fastopening)
3602	return -EBUSY;
3603
3604	mptcp_check_listen_stop(sk);
3605	mptcp_set_state(sk, TCP_CLOSE);
3606
3607	/* The later subflow close can not kick again the tout timer,
3608	* as the msk is already in closed status.
3609	*/
3610	msk->timer_ival = icsk->icsk_rto_min;
3611	sk_stop_timer_sync(sk, &sk->mptcp_retransmit_timer);
3612	icsk->icsk_mtup.probe_timestamp = 0;
3613	sk_stop_timer_sync(sk, &icsk->mptcp_tout_timer);
3614
3615	mptcp_pm_connection_closed(msk);
3616
3617	/* msk->subflow is still intact, the following will not free the first
3618	* subflow
3619	*/
3620	mptcp_do_fastclose(sk);
3621	mptcp_destroy_common(msk);
3622
3623	/* The first subflow is already in TCP_CLOSE status, the following
3624	* can't overlap with a fallback anymore
3625	*/
3626	spin_lock_bh(&msk->fallback_lock);
3627	msk->allow_subflows = true;
3628	msk->allow_infinite_fallback = true;
3629	WRITE_ONCE(msk->flags, 0);
3630	spin_unlock_bh(&msk->fallback_lock);
3631
3632	msk->cb_flags = 0;
3633	msk->recovery = false;
3634	WRITE_ONCE(msk->can_ack, false);
3635	WRITE_ONCE(msk->fully_established, false);
3636	WRITE_ONCE(msk->rcv_data_fin, false);
3637	WRITE_ONCE(msk->snd_data_fin_enable, false);
3638	WRITE_ONCE(msk->rcv_fastclose, false);
3639	WRITE_ONCE(msk->use_64bit_ack, false);
3640	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3641	mptcp_pm_data_reset(msk);
3642	mptcp_ca_reset(sk);
3643	msk->bytes_consumed = 0;
3644	msk->bytes_acked = 0;
3645	msk->bytes_received = 0;
3646	msk->bytes_sent = 0;
3647	msk->bytes_retrans = 0;
3648	msk->rcvspace_init = 0;
3649	msk->fastclosing = 0;
3650	mptcp_init_rtt_est(msk);
3651
3652	/* for fallback's sake */
3653	WRITE_ONCE(msk->ack_seq, 0);
3654	atomic64_set(&msk->rcv_wnd_sent, 0);
3655
3656	WRITE_ONCE(sk->sk_shutdown, 0);
3657	sk_error_report(sk);
3658	return 0;
3659	}
3660
3661	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3662	static struct ipv6_pinfo mptcp_inet6_sk(const struct sock sk)
3663	{
3664	struct mptcp6_sock *msk6 = container_of(mptcp_sk(sk), struct mptcp6_sock, msk);
3665
3666	return &msk6->np;
3667	}
3668
3669	static void mptcp_copy_ip6_options(struct sock newsk, const struct sock sk)
3670	{
3671	const struct ipv6_pinfo *np = inet6_sk(sk);
3672	struct ipv6_txoptions *opt;
3673	struct ipv6_pinfo *newnp;
3674
3675	newnp = inet6_sk(newsk);
3676
3677	rcu_read_lock();
3678	opt = rcu_dereference(np->opt);
3679	if (opt) {
3680	opt = ipv6_dup_options(newsk, opt);
3681	if (!opt)
3682	net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3683	}
3684	RCU_INIT_POINTER(newnp->opt, opt);
3685	rcu_read_unlock();
3686	}
3687	#endif
3688
3689	static void mptcp_copy_ip_options(struct sock newsk, const struct sock sk)
3690	{
3691	struct ip_options_rcu inet_opt, newopt = NULL;
3692	const struct inet_sock *inet = inet_sk(sk);
3693	struct inet_sock *newinet;
3694
3695	newinet = inet_sk(newsk);
3696
3697	rcu_read_lock();
3698	inet_opt = rcu_dereference(inet->inet_opt);
3699	if (inet_opt) {
3700	newopt = sock_kmemdup(newsk, inet_opt, sizeof(*inet_opt) +
3701	inet_opt->opt.optlen, GFP_ATOMIC);
3702	if (!newopt)
3703	net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3704	}
3705	RCU_INIT_POINTER(newinet->inet_opt, newopt);
3706	rcu_read_unlock();
3707	}
3708
3709	struct sock mptcp_sk_clone_init(const struct sock sk,
3710	const struct mptcp_options_received *mp_opt,
3711	struct sock *ssk,
3712	struct request_sock *req)
3713	{
3714	struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3715	struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3716	struct mptcp_subflow_context *subflow;
3717	struct mptcp_sock *msk;
3718
3719	if (!nsk)
3720	return NULL;
3721
3722	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3723	if (nsk->sk_family == AF_INET6)
3724	inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3725	#endif
3726
3727	__mptcp_init_sock(nsk);
3728
3729	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3730	if (nsk->sk_family == AF_INET6)
3731	mptcp_copy_ip6_options(nsk, sk);
3732	else
3733	#endif
3734	mptcp_copy_ip_options(nsk, sk);
3735
3736	msk = mptcp_sk(nsk);
3737	WRITE_ONCE(msk->local_key, subflow_req->local_key);
3738	WRITE_ONCE(msk->token, subflow_req->token);
3739	msk->in_accept_queue = 1;
3740	WRITE_ONCE(msk->fully_established, false);
3741	if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3742	WRITE_ONCE(msk->csum_enabled, true);
3743
3744	WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1);
3745	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3746	WRITE_ONCE(msk->snd_una, msk->write_seq);
3747	WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3748	msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3749	mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3750
3751	/* passive msk is created after the first/MPC subflow */
3752	msk->subflow_id = 2;
3753
3754	sock_reset_flag(nsk, SOCK_RCU_FREE);
3755	security_inet_csk_clone(nsk, req);
3756
3757	/* this can't race with mptcp_close(), as the msk is
3758	* not yet exposted to user-space
3759	*/
3760	mptcp_set_state(nsk, TCP_ESTABLISHED);
3761
3762	/* The msk maintain a ref to each subflow in the connections list */
3763	WRITE_ONCE(msk->first, ssk);
3764	subflow = mptcp_subflow_ctx(ssk);
3765	list_add(&subflow->node, &msk->conn_list);
3766	sock_hold(ssk);
3767
3768	/* new mpc subflow takes ownership of the newly
3769	* created mptcp socket
3770	*/
3771	mptcp_token_accept(subflow_req, msk);
3772
3773	/* set msk addresses early to ensure mptcp_pm_get_local_id()
3774	* uses the correct data
3775	*/
3776	mptcp_copy_inaddrs(nsk, ssk);
3777
3778	mptcp_rcv_space_init(msk, ssk);
3779	msk->rcvq_space.time = mptcp_stamp();
3780
3781	if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3782	__mptcp_subflow_fully_established(msk, subflow, mp_opt);
3783	bh_unlock_sock(nsk);
3784
3785	/* note: the newly allocated socket refcount is 2 now */
3786	return nsk;
3787	}
3788
3789	static void mptcp_destroy(struct sock *sk)
3790	{
3791	struct mptcp_sock *msk = mptcp_sk(sk);
3792
3793	/* allow the following to close even the initial subflow */
3794	msk->free_first = 1;
3795	mptcp_destroy_common(msk);
3796	sk_sockets_allocated_dec(sk);
3797	}
3798
3799	void __mptcp_data_acked(struct sock *sk)
3800	{
3801	if (!sock_owned_by_user(sk))
3802	__mptcp_clean_una(sk);
3803	else
3804	__set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3805	}
3806
3807	void __mptcp_check_push(struct sock sk, struct sock ssk)
3808	{
3809	if (!sock_owned_by_user(sk))
3810	__mptcp_subflow_push_pending(sk, ssk, false);
3811	else
3812	__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3813	}
3814
3815	#define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) \| \
3816	BIT(MPTCP_RETRANSMIT) \| \
3817	BIT(MPTCP_FLUSH_JOIN_LIST))
3818
3819	/* processes deferred events and flush wmem */
3820	static void mptcp_release_cb(struct sock *sk)
3821	__must_hold(&sk->sk_lock.slock)
3822	{
3823	struct mptcp_sock *msk = mptcp_sk(sk);
3824	u32 moved = 0;
3825
3826	for (;;) {
3827	unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3828	struct list_head join_list, skbs;
3829	bool spool_bl;
3830
3831	spool_bl = mptcp_can_spool_backlog(sk, &skbs);
3832	if (!flags && !spool_bl)
3833	break;
3834
3835	INIT_LIST_HEAD(&join_list);
3836	list_splice_init(&msk->join_list, &join_list);
3837
3838	/* the following actions acquire the subflow socket lock
3839	*
3840	* 1) can't be invoked in atomic scope
3841	* 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3842	* datapath acquires the msk socket spinlock while helding
3843	* the subflow socket lock
3844	*/
3845	msk->cb_flags &= ~flags;
3846	spin_unlock_bh(&sk->sk_lock.slock);
3847
3848	if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3849	__mptcp_flush_join_list(sk, &join_list);
3850	if (flags & BIT(MPTCP_PUSH_PENDING))
3851	__mptcp_push_pending(sk, 0);
3852	if (flags & BIT(MPTCP_RETRANSMIT))
3853	__mptcp_retrans(sk);
3854	if (spool_bl && __mptcp_move_skbs(sk, &skbs, &moved)) {
3855	/* notify ack seq update */
3856	mptcp_cleanup_rbuf(msk, 0);
3857	sk->sk_data_ready(sk);
3858	}
3859
3860	cond_resched();
3861	spin_lock_bh(&sk->sk_lock.slock);
3862	}
3863	if (moved)
3864	WRITE_ONCE(msk->backlog_len, msk->backlog_len - moved);
3865
3866	if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3867	__mptcp_clean_una_wakeup(sk);
3868	if (unlikely(msk->cb_flags)) {
3869	/* be sure to sync the msk state before taking actions
3870	* depending on sk_state (MPTCP_ERROR_REPORT)
3871	* On sk release avoid actions depending on the first subflow
3872	*/
3873	if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3874	__mptcp_sync_state(sk, msk->pending_state);
3875	if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3876	__mptcp_error_report(sk);
3877	if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3878	__mptcp_sync_sndbuf(sk);
3879	}
3880	}
3881
3882	/* MP_JOIN client subflow must wait for 4th ack before sending any data:
3883	* TCP can't schedule delack timer before the subflow is fully established.
3884	* MPTCP uses the delack timer to do 3rd ack retransmissions
3885	*/
3886	static void schedule_3rdack_retransmission(struct sock *ssk)
3887	{
3888	struct inet_connection_sock *icsk = inet_csk(ssk);
3889	struct tcp_sock *tp = tcp_sk(ssk);
3890	unsigned long timeout;
3891
3892	if (mptcp_subflow_ctx(ssk)->fully_established)
3893	return;
3894
3895	/* reschedule with a timeout above RTT, as we must look only for drop */
3896	if (tp->srtt_us)
3897	timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3898	else
3899	timeout = TCP_TIMEOUT_INIT;
3900	timeout += jiffies;
3901
3902	WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3903	smp_store_release(&icsk->icsk_ack.pending,
3904	icsk->icsk_ack.pending \| ICSK_ACK_SCHED \| ICSK_ACK_TIMER);
3905	sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3906	}
3907
3908	void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3909	{
3910	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3911	struct sock *sk = subflow->conn;
3912
3913	if (status & BIT(MPTCP_DELEGATE_SEND)) {
3914	mptcp_data_lock(sk);
3915	if (!sock_owned_by_user(sk))
3916	__mptcp_subflow_push_pending(sk, ssk, true);
3917	else
3918	__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3919	mptcp_data_unlock(sk);
3920	}
3921	if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3922	mptcp_data_lock(sk);
3923	if (!sock_owned_by_user(sk))
3924	__mptcp_sync_sndbuf(sk);
3925	else
3926	__set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3927	mptcp_data_unlock(sk);
3928	}
3929	if (status & BIT(MPTCP_DELEGATE_ACK))
3930	schedule_3rdack_retransmission(ssk);
3931	}
3932
3933	static int mptcp_hash(struct sock *sk)
3934	{
3935	/* should never be called,
3936	* we hash the TCP subflows not the MPTCP socket
3937	*/
3938	WARN_ON_ONCE(1);
3939	return 0;
3940	}
3941
3942	static void mptcp_unhash(struct sock *sk)
3943	{
3944	/* called from sk_common_release(), but nothing to do here */
3945	}
3946
3947	static int mptcp_get_port(struct sock *sk, unsigned short snum)
3948	{
3949	struct mptcp_sock *msk = mptcp_sk(sk);
3950
3951	pr_debug("msk=%p, ssk=%p\n", msk, msk->first);
3952	if (WARN_ON_ONCE(!msk->first))
3953	return -EINVAL;
3954
3955	return inet_csk_get_port(msk->first, snum);
3956	}
3957
3958	void mptcp_finish_connect(struct sock *ssk)
3959	{
3960	struct mptcp_subflow_context *subflow;
3961	struct mptcp_sock *msk;
3962	struct sock *sk;
3963
3964	subflow = mptcp_subflow_ctx(ssk);
3965	sk = subflow->conn;
3966	msk = mptcp_sk(sk);
3967
3968	pr_debug("msk=%p, token=%u\n", sk, subflow->token);
3969
3970	subflow->map_seq = subflow->iasn;
3971	subflow->map_subflow_seq = 1;
3972
3973	/* the socket is not connected yet, no msk/subflow ops can access/race
3974	* accessing the field below
3975	*/
3976	WRITE_ONCE(msk->local_key, subflow->local_key);
3977	WRITE_ONCE(msk->rcvq_space.time, mptcp_stamp());
3978
3979	mptcp_pm_new_connection(msk, ssk, 0);
3980	}
3981
3982	void mptcp_sock_graft(struct sock sk, struct socket parent)
3983	{
3984	write_lock_bh(&sk->sk_callback_lock);
3985	rcu_assign_pointer(sk->sk_wq, &parent->wq);
3986	sk_set_socket(sk, parent);
3987	write_unlock_bh(&sk->sk_callback_lock);
3988	}
3989
3990	/* Can be called without holding the msk socket lock; use the callback lock
3991	* to avoid {READ_,WRITE_}ONCE annotations on sk_socket.
3992	*/
3993	static void mptcp_sock_check_graft(struct sock sk, struct sock ssk)
3994	{
3995	struct socket *sock;
3996
3997	write_lock_bh(&sk->sk_callback_lock);
3998	sock = sk->sk_socket;
3999	write_unlock_bh(&sk->sk_callback_lock);
4000	if (sock) {
4001	mptcp_sock_graft(ssk, sock);
4002	__mptcp_inherit_cgrp_data(sk, ssk);
4003	__mptcp_inherit_memcg(sk, ssk, GFP_ATOMIC);
4004	}
4005	}
4006
4007	bool mptcp_finish_join(struct sock *ssk)
4008	{
4009	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
4010	struct mptcp_sock *msk = mptcp_sk(subflow->conn);
4011	struct sock parent = (void )msk;
4012	bool ret = true;
4013
4014	pr_debug("msk=%p, subflow=%p\n", msk, subflow);
4015
4016	/* mptcp socket already closing? */
4017	if (!mptcp_is_fully_established(parent)) {
4018	MPTCP_INC_STATS(sock_net(parent), MPTCP_MIB_MPJOINNOTESTABLISHED);
4019	subflow->reset_reason = MPTCP_RST_EMPTCP;
4020	return false;
4021	}
4022
4023	/* Active subflow, already present inside the conn_list; is grafted
4024	* either by __mptcp_subflow_connect() or accept.
4025	*/
4026	if (!list_empty(&subflow->node)) {
4027	spin_lock_bh(&msk->fallback_lock);
4028	if (!msk->allow_subflows) {
4029	spin_unlock_bh(&msk->fallback_lock);
4030	return false;
4031	}
4032	mptcp_subflow_joined(msk, ssk);
4033	spin_unlock_bh(&msk->fallback_lock);
4034	mptcp_propagate_sndbuf(parent, ssk);
4035	return true;
4036	}
4037
4038	if (!mptcp_pm_allow_new_subflow(msk)) {
4039	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_JOINREJECTED);
4040	goto err_prohibited;
4041	}
4042
4043	/* If we can't acquire msk socket lock here, let the release callback
4044	* handle it
4045	*/
4046	mptcp_data_lock(parent);
4047	if (!sock_owned_by_user(parent)) {
4048	ret = __mptcp_finish_join(msk, ssk);
4049	if (ret) {
4050	sock_hold(ssk);
4051	list_add_tail(&subflow->node, &msk->conn_list);
4052	mptcp_sock_check_graft(parent, ssk);
4053	}
4054	} else {
4055	sock_hold(ssk);
4056	list_add_tail(&subflow->node, &msk->join_list);
4057	__set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
4058
4059	/* In case of later failures, __mptcp_flush_join_list() will
4060	* properly orphan the ssk via mptcp_close_ssk().
4061	*/
4062	mptcp_sock_check_graft(parent, ssk);
4063	}
4064	mptcp_data_unlock(parent);
4065
4066	if (!ret) {
4067	mptcp_pm_close_subflow(msk);
4068	err_prohibited:
4069	subflow->reset_reason = MPTCP_RST_EPROHIBIT;
4070	return false;
4071	}
4072
4073	return true;
4074	}
4075
4076	static void mptcp_shutdown(struct sock *sk, int how)
4077	{
4078	pr_debug("sk=%p, how=%d\n", sk, how);
4079
4080	if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
4081	__mptcp_wr_shutdown(sk);
4082	}
4083
4084	static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
4085	{
4086	const struct sock sk = (void )msk;
4087	u64 delta;
4088
4089	if (sk->sk_state == TCP_LISTEN)
4090	return -EINVAL;
4091
4092	if ((1 << sk->sk_state) & (TCPF_SYN_SENT \| TCPF_SYN_RECV))
4093	return 0;
4094
4095	delta = msk->write_seq - v;
4096	if (__mptcp_check_fallback(msk) && msk->first) {
4097	struct tcp_sock *tp = tcp_sk(msk->first);
4098
4099	/* the first subflow is disconnected after close - see
4100	* __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
4101	* so ignore that status, too.
4102	*/
4103	if (!((1 << msk->first->sk_state) &
4104	(TCPF_SYN_SENT \| TCPF_SYN_RECV \| TCPF_CLOSE)))
4105	delta += READ_ONCE(tp->write_seq) - tp->snd_una;
4106	}
4107	if (delta > INT_MAX)
4108	delta = INT_MAX;
4109
4110	return (int)delta;
4111	}
4112
4113	static int mptcp_ioctl(struct sock sk, int cmd, int karg)
4114	{
4115	struct mptcp_sock *msk = mptcp_sk(sk);
4116	bool slow;
4117
4118	switch (cmd) {
4119	case SIOCINQ:
4120	if (sk->sk_state == TCP_LISTEN)
4121	return -EINVAL;
4122
4123	lock_sock(sk);
4124	if (mptcp_move_skbs(sk))
4125	mptcp_cleanup_rbuf(msk, 0);
4126	*karg = mptcp_inq_hint(sk);
4127	release_sock(sk);
4128	break;
4129	case SIOCOUTQ:
4130	slow = lock_sock_fast(sk);
4131	*karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
4132	unlock_sock_fast(sk, slow);
4133	break;
4134	case SIOCOUTQNSD:
4135	slow = lock_sock_fast(sk);
4136	*karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
4137	unlock_sock_fast(sk, slow);
4138	break;
4139	default:
4140	return -ENOIOCTLCMD;
4141	}
4142
4143	return 0;
4144	}
4145
4146	static int mptcp_connect(struct sock sk, struct sockaddr_unsized uaddr,
4147	int addr_len)
4148	{
4149	struct mptcp_subflow_context *subflow;
4150	struct mptcp_sock *msk = mptcp_sk(sk);
4151	int err = -EINVAL;
4152	struct sock *ssk;
4153
4154	ssk = __mptcp_nmpc_sk(msk);
4155	if (IS_ERR(ssk))
4156	return PTR_ERR(ssk);
4157
4158	set_bit(MPTCP_RTX_ENABLED, &msk->flags);
4159	mptcp_set_state(sk, TCP_SYN_SENT);
4160	subflow = mptcp_subflow_ctx(ssk);
4161	#ifdef CONFIG_TCP_MD5SIG
4162	/* no MPTCP if MD5SIG is enabled on this socket or we may run out of
4163	* TCP option space.
4164	*/
4165	if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
4166	mptcp_early_fallback(msk, subflow, MPTCP_MIB_MD5SIGFALLBACK);
4167	#endif
4168	if (subflow->request_mptcp) {
4169	if (mptcp_active_should_disable(sk))
4170	mptcp_early_fallback(msk, subflow,
4171	MPTCP_MIB_MPCAPABLEACTIVEDISABLED);
4172	else if (mptcp_token_new_connect(ssk) < 0)
4173	mptcp_early_fallback(msk, subflow,
4174	MPTCP_MIB_TOKENFALLBACKINIT);
4175	}
4176
4177	WRITE_ONCE(msk->write_seq, subflow->idsn);
4178	WRITE_ONCE(msk->snd_nxt, subflow->idsn);
4179	WRITE_ONCE(msk->snd_una, subflow->idsn);
4180	if (likely(!__mptcp_check_fallback(msk)))
4181	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
4182
4183	/* if reaching here via the fastopen/sendmsg path, the caller already
4184	* acquired the subflow socket lock, too.
4185	*/
4186	if (!msk->fastopening)
4187	lock_sock(ssk);
4188
4189	/* the following mirrors closely a very small chunk of code from
4190	* __inet_stream_connect()
4191	*/
4192	if (ssk->sk_state != TCP_CLOSE)
4193	goto out;
4194
4195	if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
4196	err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
4197	if (err)
4198	goto out;
4199	}
4200
4201	err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
4202	if (err < 0)
4203	goto out;
4204
4205	inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
4206
4207	out:
4208	if (!msk->fastopening)
4209	release_sock(ssk);
4210
4211	/* on successful connect, the msk state will be moved to established by
4212	* subflow_finish_connect()
4213	*/
4214	if (unlikely(err)) {
4215	/* avoid leaving a dangling token in an unconnected socket */
4216	mptcp_token_destroy(msk);
4217	mptcp_set_state(sk, TCP_CLOSE);
4218	return err;
4219	}
4220
4221	mptcp_copy_inaddrs(sk, ssk);
4222	return 0;
4223	}
4224
4225	static struct proto mptcp_prot = {
4226	.name = "MPTCP",
4227	.owner = THIS_MODULE,
4228	.init = mptcp_init_sock,
4229	.connect = mptcp_connect,
4230	.disconnect = mptcp_disconnect,
4231	.close = mptcp_close,
4232	.setsockopt = mptcp_setsockopt,
4233	.getsockopt = mptcp_getsockopt,
4234	.shutdown = mptcp_shutdown,
4235	.destroy = mptcp_destroy,
4236	.sendmsg = mptcp_sendmsg,
4237	.ioctl = mptcp_ioctl,
4238	.recvmsg = mptcp_recvmsg,
4239	.release_cb = mptcp_release_cb,
4240	.hash = mptcp_hash,
4241	.unhash = mptcp_unhash,
4242	.get_port = mptcp_get_port,
4243	.stream_memory_free = mptcp_stream_memory_free,
4244	.sockets_allocated = &mptcp_sockets_allocated,
4245
4246	.memory_allocated = &net_aligned_data.tcp_memory_allocated,
4247	.per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
4248
4249	.memory_pressure = &tcp_memory_pressure,
4250	.sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
4251	.sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
4252	.sysctl_mem = sysctl_tcp_mem,
4253	.obj_size = sizeof(struct mptcp_sock),
4254	.slab_flags = SLAB_TYPESAFE_BY_RCU,
4255	.no_autobind = true,
4256	};
4257
4258	static int mptcp_bind(struct socket sock, struct sockaddr_unsized uaddr, int addr_len)
4259	{
4260	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4261	struct sock ssk, sk = sock->sk;
4262	int err = -EINVAL;
4263
4264	lock_sock(sk);
4265	ssk = __mptcp_nmpc_sk(msk);
4266	if (IS_ERR(ssk)) {
4267	err = PTR_ERR(ssk);
4268	goto unlock;
4269	}
4270
4271	if (sk->sk_family == AF_INET)
4272	err = inet_bind_sk(ssk, uaddr, addr_len);
4273	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4274	else if (sk->sk_family == AF_INET6)
4275	err = inet6_bind_sk(ssk, uaddr, addr_len);
4276	#endif
4277	if (!err)
4278	mptcp_copy_inaddrs(sk, ssk);
4279
4280	unlock:
4281	release_sock(sk);
4282	return err;
4283	}
4284
4285	static int mptcp_listen(struct socket *sock, int backlog)
4286	{
4287	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4288	struct sock *sk = sock->sk;
4289	struct sock *ssk;
4290	int err;
4291
4292	pr_debug("msk=%p\n", msk);
4293
4294	lock_sock(sk);
4295
4296	err = -EINVAL;
4297	if (sock->state != SS_UNCONNECTED \|\| sock->type != SOCK_STREAM)
4298	goto unlock;
4299
4300	ssk = __mptcp_nmpc_sk(msk);
4301	if (IS_ERR(ssk)) {
4302	err = PTR_ERR(ssk);
4303	goto unlock;
4304	}
4305
4306	set_bit(MPTCP_RTX_ENABLED, &msk->flags);
4307	mptcp_set_state(sk, TCP_LISTEN);
4308	sock_set_flag(sk, SOCK_RCU_FREE);
4309
4310	lock_sock(ssk);
4311	err = __inet_listen_sk(ssk, backlog);
4312	release_sock(ssk);
4313	mptcp_set_state(sk, inet_sk_state_load(ssk));
4314
4315	if (!err) {
4316	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
4317	mptcp_copy_inaddrs(sk, ssk);
4318	mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
4319	}
4320
4321	unlock:
4322	release_sock(sk);
4323	return err;
4324	}
4325
4326	static void mptcp_graft_subflows(struct sock *sk)
4327	{
4328	struct mptcp_subflow_context *subflow;
4329	struct mptcp_sock *msk = mptcp_sk(sk);
4330
4331	if (mem_cgroup_sockets_enabled) {
4332	LIST_HEAD(join_list);
4333
4334	/* Subflows joining after __inet_accept() will get the
4335	* mem CG properly initialized at mptcp_finish_join() time,
4336	* but subflows pending in join_list need explicit
4337	* initialization before flushing `backlog_unaccounted`
4338	* or MPTCP can later unexpectedly observe unaccounted memory.
4339	*/
4340	mptcp_data_lock(sk);
4341	list_splice_init(&msk->join_list, &join_list);
4342	mptcp_data_unlock(sk);
4343
4344	__mptcp_flush_join_list(sk, &join_list);
4345	}
4346
4347	mptcp_for_each_subflow(msk, subflow) {
4348	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4349
4350	lock_sock(ssk);
4351
4352	/* Set ssk->sk_socket of accept()ed flows to mptcp socket.
4353	* This is needed so NOSPACE flag can be set from tcp stack.
4354	*/
4355	if (!ssk->sk_socket)
4356	mptcp_sock_graft(ssk, sk->sk_socket);
4357
4358	if (!mem_cgroup_sk_enabled(sk))
4359	goto unlock;
4360
4361	__mptcp_inherit_cgrp_data(sk, ssk);
4362	__mptcp_inherit_memcg(sk, ssk, GFP_KERNEL);
4363
4364	unlock:
4365	release_sock(ssk);
4366	}
4367
4368	if (mem_cgroup_sk_enabled(sk)) {
4369	gfp_t gfp = GFP_KERNEL \| __GFP_NOFAIL;
4370	int amt;
4371
4372	/* Account the backlog memory; prior accept() is aware of
4373	* fwd and rmem only.
4374	*/
4375	mptcp_data_lock(sk);
4376	amt = sk_mem_pages(sk->sk_forward_alloc +
4377	msk->backlog_unaccounted +
4378	atomic_read(&sk->sk_rmem_alloc)) -
4379	sk_mem_pages(sk->sk_forward_alloc +
4380	atomic_read(&sk->sk_rmem_alloc));
4381	msk->backlog_unaccounted = 0;
4382	mptcp_data_unlock(sk);
4383
4384	if (amt)
4385	mem_cgroup_sk_charge(sk, amt, gfp);
4386	}
4387	}
4388
4389	static int mptcp_stream_accept(struct socket sock, struct socket newsock,
4390	struct proto_accept_arg *arg)
4391	{
4392	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4393	struct sock ssk, newsk;
4394
4395	pr_debug("msk=%p\n", msk);
4396
4397	/* Buggy applications can call accept on socket states other then LISTEN
4398	* but no need to allocate the first subflow just to error out.
4399	*/
4400	ssk = READ_ONCE(msk->first);
4401	if (!ssk)
4402	return -EINVAL;
4403
4404	pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk));
4405	newsk = inet_csk_accept(ssk, arg);
4406	if (!newsk)
4407	return arg->err;
4408
4409	pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk));
4410	if (sk_is_mptcp(newsk)) {
4411	struct mptcp_subflow_context *subflow;
4412	struct sock *new_mptcp_sock;
4413
4414	subflow = mptcp_subflow_ctx(newsk);
4415	new_mptcp_sock = subflow->conn;
4416
4417	/* is_mptcp should be false if subflow->conn is missing, see
4418	* subflow_syn_recv_sock()
4419	*/
4420	if (WARN_ON_ONCE(!new_mptcp_sock)) {
4421	tcp_sk(newsk)->is_mptcp = 0;
4422	goto tcpfallback;
4423	}
4424
4425	newsk = new_mptcp_sock;
4426	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
4427
4428	newsk->sk_kern_sock = arg->kern;
4429	lock_sock(newsk);
4430	__inet_accept(sock, newsock, newsk);
4431
4432	set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
4433	msk = mptcp_sk(newsk);
4434	msk->in_accept_queue = 0;
4435
4436	mptcp_graft_subflows(newsk);
4437	mptcp_rps_record_subflows(msk);
4438	__mptcp_propagate_sndbuf(newsk, mptcp_subflow_tcp_sock(subflow));
4439
4440	/* Do late cleanup for the first subflow as necessary. Also
4441	* deal with bad peers not doing a complete shutdown.
4442	*/
4443	if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
4444	if (unlikely(list_is_singular(&msk->conn_list)))
4445	mptcp_set_state(newsk, TCP_CLOSE);
4446	mptcp_close_ssk(newsk, msk->first,
4447	mptcp_subflow_ctx(msk->first));
4448	}
4449	} else {
4450	tcpfallback:
4451	newsk->sk_kern_sock = arg->kern;
4452	lock_sock(newsk);
4453	__inet_accept(sock, newsock, newsk);
4454	/* we are being invoked after accepting a non-mp-capable
4455	* flow: sk is a tcp_sk, not an mptcp one.
4456	*
4457	* Hand the socket over to tcp so all further socket ops
4458	* bypass mptcp.
4459	*/
4460	WRITE_ONCE(newsock->sk->sk_socket->ops,
4461	mptcp_fallback_tcp_ops(newsock->sk));
4462	}
4463	release_sock(newsk);
4464
4465	return 0;
4466	}
4467
4468	static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
4469	{
4470	struct sock sk = (struct sock )msk;
4471
4472	if (__mptcp_stream_is_writeable(sk, 1))
4473	return EPOLLOUT \| EPOLLWRNORM;
4474
4475	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
4476	smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
4477	if (__mptcp_stream_is_writeable(sk, 1))
4478	return EPOLLOUT \| EPOLLWRNORM;
4479
4480	return 0;
4481	}
4482
4483	static __poll_t mptcp_poll(struct file file, struct socket sock,
4484	struct poll_table_struct *wait)
4485	{
4486	struct sock *sk = sock->sk;
4487	struct mptcp_sock *msk;
4488	__poll_t mask = 0;
4489	u8 shutdown;
4490	int state;
4491
4492	msk = mptcp_sk(sk);
4493	sock_poll_wait(file, sock, wait);
4494
4495	state = inet_sk_state_load(sk);
4496	pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags);
4497	if (state == TCP_LISTEN) {
4498	struct sock *ssk = READ_ONCE(msk->first);
4499
4500	if (WARN_ON_ONCE(!ssk))
4501	return 0;
4502
4503	return inet_csk_listen_poll(ssk);
4504	}
4505
4506	shutdown = READ_ONCE(sk->sk_shutdown);
4507	if (shutdown == SHUTDOWN_MASK \|\| state == TCP_CLOSE)
4508	mask \|= EPOLLHUP;
4509	if (shutdown & RCV_SHUTDOWN)
4510	mask \|= EPOLLIN \| EPOLLRDNORM \| EPOLLRDHUP;
4511
4512	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
4513	mask \|= mptcp_check_readable(sk);
4514	if (shutdown & SEND_SHUTDOWN)
4515	mask \|= EPOLLOUT \| EPOLLWRNORM;
4516	else
4517	mask \|= mptcp_check_writeable(msk);
4518	} else if (state == TCP_SYN_SENT &&
4519	inet_test_bit(DEFER_CONNECT, sk)) {
4520	/* cf tcp_poll() note about TFO */
4521	mask \|= EPOLLOUT \| EPOLLWRNORM;
4522	}
4523
4524	/* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
4525	smp_rmb();
4526	if (READ_ONCE(sk->sk_err))
4527	mask \|= EPOLLERR;
4528
4529	return mask;
4530	}
4531
4532	static struct sk_buff mptcp_recv_skb(struct sock sk, u32 *off)
4533	{
4534	struct mptcp_sock *msk = mptcp_sk(sk);
4535	struct sk_buff *skb;
4536	u32 offset;
4537
4538	if (!list_empty(&msk->backlog_list))
4539	mptcp_move_skbs(sk);
4540
4541	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
4542	offset = MPTCP_SKB_CB(skb)->offset;
4543	if (offset < skb->len) {
4544	*off = offset;
4545	return skb;
4546	}
4547	mptcp_eat_recv_skb(sk, skb);
4548	}
4549	return NULL;
4550	}
4551
4552	/*
4553	* Note:
4554	* - It is assumed that the socket was locked by the caller.
4555	*/
4556	static int __mptcp_read_sock(struct sock sk, read_descriptor_t desc,
4557	sk_read_actor_t recv_actor, bool noack)
4558	{
4559	struct mptcp_sock *msk = mptcp_sk(sk);
4560	struct sk_buff *skb;
4561	int copied = 0;
4562	u32 offset;
4563
4564	msk_owned_by_me(msk);
4565
4566	if (sk->sk_state == TCP_LISTEN)
4567	return -ENOTCONN;
4568	while ((skb = mptcp_recv_skb(sk, &offset)) != NULL) {
4569	u32 data_len = skb->len - offset;
4570	int count;
4571	u32 size;
4572
4573	size = min_t(size_t, data_len, INT_MAX);
4574	count = recv_actor(desc, skb, offset, size);
4575	if (count <= 0) {
4576	if (!copied)
4577	copied = count;
4578	break;
4579	}
4580
4581	copied += count;
4582
4583	msk->bytes_consumed += count;
4584	if (count < data_len) {
4585	MPTCP_SKB_CB(skb)->offset += count;
4586	MPTCP_SKB_CB(skb)->map_seq += count;
4587	break;
4588	}
4589
4590	mptcp_eat_recv_skb(sk, skb);
4591	if (!desc->count)
4592	break;
4593	}
4594
4595	if (noack)
4596	goto out;
4597
4598	mptcp_rcv_space_adjust(msk, copied);
4599
4600	if (copied > 0) {
4601	mptcp_recv_skb(sk, &offset);
4602	mptcp_cleanup_rbuf(msk, copied);
4603	}
4604	out:
4605	return copied;
4606	}
4607
4608	static int mptcp_read_sock(struct sock sk, read_descriptor_t desc,
4609	sk_read_actor_t recv_actor)
4610	{
4611	return __mptcp_read_sock(sk, desc, recv_actor, false);
4612	}
4613
4614	static int __mptcp_splice_read(struct sock sk, struct tcp_splice_state tss)
4615	{
4616	/* Store TCP splice context information in read_descriptor_t. */
4617	read_descriptor_t rd_desc = {
4618	.arg.data = tss,
4619	.count = tss->len,
4620	};
4621
4622	return mptcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
4623	}
4624
4625	/**
4626	* mptcp_splice_read - splice data from MPTCP socket to a pipe
4627	* @sock: socket to splice from
4628	* @ppos: position (not valid)
4629	* @pipe: pipe to splice to
4630	* @len: number of bytes to splice
4631	* @flags: splice modifier flags
4632	*
4633	* Description:
4634	* Will read pages from given socket and fill them into a pipe.
4635	*
4636	* Return:
4637	* Amount of bytes that have been spliced.
4638	*
4639	**/
4640	static ssize_t mptcp_splice_read(struct socket sock, loff_t ppos,
4641	struct pipe_inode_info *pipe, size_t len,
4642	unsigned int flags)
4643	{
4644	struct tcp_splice_state tss = {
4645	.pipe = pipe,
4646	.len = len,
4647	.flags = flags,
4648	};
4649	struct sock *sk = sock->sk;
4650	ssize_t spliced = 0;
4651	int ret = 0;
4652	long timeo;
4653
4654	/*
4655	* We can't seek on a socket input
4656	*/
4657	if (unlikely(*ppos))
4658	return -ESPIPE;
4659
4660	lock_sock(sk);
4661
4662	mptcp_rps_record_subflows(mptcp_sk(sk));
4663
4664	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
4665	while (tss.len) {
4666	ret = __mptcp_splice_read(sk, &tss);
4667	if (ret < 0) {
4668	break;
4669	} else if (!ret) {
4670	if (spliced)
4671	break;
4672	if (sock_flag(sk, SOCK_DONE))
4673	break;
4674	if (sk->sk_err) {
4675	ret = sock_error(sk);
4676	break;
4677	}
4678	if (sk->sk_shutdown & RCV_SHUTDOWN)
4679	break;
4680	if (sk->sk_state == TCP_CLOSE) {
4681	/*
4682	* This occurs when user tries to read
4683	* from never connected socket.
4684	*/
4685	ret = -ENOTCONN;
4686	break;
4687	}
4688	if (!timeo) {
4689	ret = -EAGAIN;
4690	break;
4691	}
4692	/* if __mptcp_splice_read() got nothing while we have
4693	* an skb in receive queue, we do not want to loop.
4694	* This might happen with URG data.
4695	*/
4696	if (!skb_queue_empty(&sk->sk_receive_queue))
4697	break;
4698	ret = sk_wait_data(sk, &timeo, NULL);
4699	if (ret < 0)
4700	break;
4701	if (signal_pending(current)) {
4702	ret = sock_intr_errno(timeo);
4703	break;
4704	}
4705	continue;
4706	}
4707	tss.len -= ret;
4708	spliced += ret;
4709
4710	if (!tss.len \|\| !timeo)
4711	break;
4712	release_sock(sk);
4713	lock_sock(sk);
4714
4715	if (tcp_recv_should_stop(sk))
4716	break;
4717	}
4718
4719	release_sock(sk);
4720
4721	if (spliced)
4722	return spliced;
4723
4724	return ret;
4725	}
4726
4727	static const struct proto_ops mptcp_stream_ops = {
4728	.family = PF_INET,
4729	.owner = THIS_MODULE,
4730	.release = inet_release,
4731	.bind = mptcp_bind,
4732	.connect = inet_stream_connect,
4733	.socketpair = sock_no_socketpair,
4734	.accept = mptcp_stream_accept,
4735	.getname = inet_getname,
4736	.poll = mptcp_poll,
4737	.ioctl = inet_ioctl,
4738	.gettstamp = sock_gettstamp,
4739	.listen = mptcp_listen,
4740	.shutdown = inet_shutdown,
4741	.setsockopt = sock_common_setsockopt,
4742	.getsockopt = sock_common_getsockopt,
4743	.sendmsg = inet_sendmsg,
4744	.recvmsg = inet_recvmsg,
4745	.mmap = sock_no_mmap,
4746	.set_rcvlowat = mptcp_set_rcvlowat,
4747	.read_sock = mptcp_read_sock,
4748	.splice_read = mptcp_splice_read,
4749	};
4750
4751	static struct inet_protosw mptcp_protosw = {
4752	.type = SOCK_STREAM,
4753	.protocol = IPPROTO_MPTCP,
4754	.prot = &mptcp_prot,
4755	.ops = &mptcp_stream_ops,
4756	.flags = INET_PROTOSW_ICSK,
4757	};
4758
4759	static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4760	{
4761	struct mptcp_delegated_action *delegated;
4762	struct mptcp_subflow_context *subflow;
4763	int work_done = 0;
4764
4765	delegated = container_of(napi, struct mptcp_delegated_action, napi);
4766	while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4767	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4768
4769	bh_lock_sock_nested(ssk);
4770	if (!sock_owned_by_user(ssk)) {
4771	mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4772	} else {
4773	/* tcp_release_cb_override already processed
4774	* the action or will do at next release_sock().
4775	* In both case must dequeue the subflow here - on the same
4776	* CPU that scheduled it.
4777	*/
4778	smp_wmb();
4779	clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4780	}
4781	bh_unlock_sock(ssk);
4782	sock_put(ssk);
4783
4784	if (++work_done == budget)
4785	return budget;
4786	}
4787
4788	/* always provide a 0 'work_done' argument, so that napi_complete_done
4789	* will not try accessing the NULL napi->dev ptr
4790	*/
4791	napi_complete_done(napi, 0);
4792	return work_done;
4793	}
4794
4795	void __init mptcp_proto_init(void)
4796	{
4797	struct mptcp_delegated_action *delegated;
4798	int cpu;
4799
4800	mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4801
4802	if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4803	panic("Failed to allocate MPTCP pcpu counter\n");
4804
4805	mptcp_napi_dev = alloc_netdev_dummy(0);
4806	if (!mptcp_napi_dev)
4807	panic("Failed to allocate MPTCP dummy netdev\n");
4808	for_each_possible_cpu(cpu) {
4809	delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4810	INIT_LIST_HEAD(&delegated->head);
4811	netif_napi_add_tx(mptcp_napi_dev, &delegated->napi,
4812	mptcp_napi_poll);
4813	napi_enable(&delegated->napi);
4814	}
4815
4816	mptcp_subflow_init();
4817	mptcp_pm_init();
4818	mptcp_sched_init();
4819	mptcp_token_init();
4820
4821	if (proto_register(&mptcp_prot, 1) != 0)
4822	panic("Failed to register MPTCP proto.\n");
4823
4824	inet_register_protosw(&mptcp_protosw);
4825
4826	BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4827
4828	/* struct mptcp_data_frag: 'overhead' corresponds to the alignment
4829	* (ALIGN(1, sizeof(long)) - 1, so 8-1) + the struct's size
4830	*/
4831	BUILD_BUG_ON(ALIGN(1, sizeof(long)) - 1 + sizeof(struct mptcp_data_frag)
4832	> U8_MAX);
4833	}
4834
4835	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4836	static const struct proto_ops mptcp_v6_stream_ops = {
4837	.family = PF_INET6,
4838	.owner = THIS_MODULE,
4839	.release = inet6_release,
4840	.bind = mptcp_bind,
4841	.connect = inet_stream_connect,
4842	.socketpair = sock_no_socketpair,
4843	.accept = mptcp_stream_accept,
4844	.getname = inet6_getname,
4845	.poll = mptcp_poll,
4846	.ioctl = inet6_ioctl,
4847	.gettstamp = sock_gettstamp,
4848	.listen = mptcp_listen,
4849	.shutdown = inet_shutdown,
4850	.setsockopt = sock_common_setsockopt,
4851	.getsockopt = sock_common_getsockopt,
4852	.sendmsg = inet6_sendmsg,
4853	.recvmsg = inet6_recvmsg,
4854	.mmap = sock_no_mmap,
4855	#ifdef CONFIG_COMPAT
4856	.compat_ioctl = inet6_compat_ioctl,
4857	#endif
4858	.set_rcvlowat = mptcp_set_rcvlowat,
4859	.read_sock = mptcp_read_sock,
4860	.splice_read = mptcp_splice_read,
4861	};
4862
4863	static struct proto mptcp_v6_prot;
4864
4865	static struct inet_protosw mptcp_v6_protosw = {
4866	.type = SOCK_STREAM,
4867	.protocol = IPPROTO_MPTCP,
4868	.prot = &mptcp_v6_prot,
4869	.ops = &mptcp_v6_stream_ops,
4870	.flags = INET_PROTOSW_ICSK,
4871	};
4872
4873	int __init mptcp_proto_v6_init(void)
4874	{
4875	int err;
4876
4877	mptcp_subflow_v6_init();
4878
4879	mptcp_v6_prot = mptcp_prot;
4880	strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name));
4881	mptcp_v6_prot.slab = NULL;
4882	mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4883	mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4884
4885	err = proto_register(&mptcp_v6_prot, 1);
4886	if (err)
4887	return err;
4888
4889	err = inet6_register_protosw(&mptcp_v6_protosw);
4890	if (err)
4891	proto_unregister(&mptcp_v6_prot);
4892
4893	return err;
4894	}
4895	#endif