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	#define CREATE_TRACE_POINTS
32	#include <trace/events/mptcp.h>
33
34	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
35	struct mptcp6_sock {
36	struct mptcp_sock msk;
37	struct ipv6_pinfo np;
38	};
39	#endif
40
41	enum {
42	MPTCP_CMSG_TS = BIT(0),
43	MPTCP_CMSG_INQ = BIT(1),
44	};
45
46	static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
47
48	static void __mptcp_destroy_sock(struct sock *sk);
49	static void mptcp_check_send_data_fin(struct sock *sk);
50
51	DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions) = {
52	.bh_lock = INIT_LOCAL_LOCK(bh_lock),
53	};
54	static struct net_device *mptcp_napi_dev;
55
56	/* Returns end sequence number of the receiver's advertised window */
57	static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
58	{
59	return READ_ONCE(msk->wnd_end);
60	}
61
62	static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk)
63	{
64	unsigned short family = READ_ONCE(sk->sk_family);
65
66	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
67	if (family == AF_INET6)
68	return &inet6_stream_ops;
69	#endif
70	WARN_ON_ONCE(family != AF_INET);
71	return &inet_stream_ops;
72	}
73
74	bool __mptcp_try_fallback(struct mptcp_sock *msk, int fb_mib)
75	{
76	struct net *net = sock_net(sk: (struct sock *)msk);
77
78	if (__mptcp_check_fallback(msk))
79	return true;
80
81	/* The caller possibly is not holding the msk socket lock, but
82	* in the fallback case only the current subflow is touching
83	* the OoO queue.
84	*/
85	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue))
86	return false;
87
88	spin_lock_bh(lock: &msk->fallback_lock);
89	if (!msk->allow_infinite_fallback) {
90	spin_unlock_bh(lock: &msk->fallback_lock);
91	return false;
92	}
93
94	msk->allow_subflows = false;
95	set_bit(MPTCP_FALLBACK_DONE, addr: &msk->flags);
96	__MPTCP_INC_STATS(net, field: fb_mib);
97	spin_unlock_bh(lock: &msk->fallback_lock);
98	return true;
99	}
100
101	static int __mptcp_socket_create(struct mptcp_sock *msk)
102	{
103	struct mptcp_subflow_context *subflow;
104	struct sock *sk = (struct sock *)msk;
105	struct socket *ssock;
106	int err;
107
108	err = mptcp_subflow_create_socket(sk, family: sk->sk_family, new_sock: &ssock);
109	if (err)
110	return err;
111
112	msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
113	WRITE_ONCE(msk->first, ssock->sk);
114	subflow = mptcp_subflow_ctx(sk: ssock->sk);
115	list_add(new: &subflow->node, head: &msk->conn_list);
116	sock_hold(sk: ssock->sk);
117	subflow->request_mptcp = 1;
118	subflow->subflow_id = msk->subflow_id++;
119
120	/* This is the first subflow, always with id 0 */
121	WRITE_ONCE(subflow->local_id, 0);
122	mptcp_sock_graft(sk: msk->first, parent: sk->sk_socket);
123	iput(SOCK_INODE(socket: ssock));
124
125	return 0;
126	}
127
128	/* If the MPC handshake is not started, returns the first subflow,
129	* eventually allocating it.
130	*/
131	struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
132	{
133	struct sock *sk = (struct sock *)msk;
134	int ret;
135
136	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
137	return ERR_PTR(error: -EINVAL);
138
139	if (!msk->first) {
140	ret = __mptcp_socket_create(msk);
141	if (ret)
142	return ERR_PTR(error: ret);
143	}
144
145	return msk->first;
146	}
147
148	static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
149	{
150	sk_drops_skbadd(sk, skb);
151	__kfree_skb(skb);
152	}
153
154	static bool __mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
155	struct sk_buff *from, bool *fragstolen,
156	int *delta)
157	{
158	int limit = READ_ONCE(sk->sk_rcvbuf);
159
160	if (unlikely(MPTCP_SKB_CB(to)->cant_coalesce) ||
161	MPTCP_SKB_CB(from)->offset ||
162	((to->len + from->len) > (limit >> 3)) ||
163	!skb_try_coalesce(to, from, fragstolen, delta_truesize: delta))
164	return false;
165
166	pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n",
167	MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
168	to->len, MPTCP_SKB_CB(from)->end_seq);
169	MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
170	return true;
171	}
172
173	static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
174	struct sk_buff *from)
175	{
176	bool fragstolen;
177	int delta;
178
179	if (!__mptcp_try_coalesce(sk, to, from, fragstolen: &fragstolen, delta: &delta))
180	return false;
181
182	/* note the fwd memory can reach a negative value after accounting
183	* for the delta, but the later skb free will restore a non
184	* negative one
185	*/
186	atomic_add(i: delta, v: &sk->sk_rmem_alloc);
187	sk_mem_charge(sk, size: delta);
188	kfree_skb_partial(skb: from, head_stolen: fragstolen);
189
190	return true;
191	}
192
193	static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
194	struct sk_buff *from)
195	{
196	if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
197	return false;
198
199	return mptcp_try_coalesce(sk: (struct sock *)msk, to, from);
200	}
201
202	/* "inspired" by tcp_rcvbuf_grow(), main difference:
203	* - mptcp does not maintain a msk-level window clamp
204	* - returns true when the receive buffer is actually updated
205	*/
206	static bool mptcp_rcvbuf_grow(struct sock *sk, u32 newval)
207	{
208	struct mptcp_sock *msk = mptcp_sk(sk);
209	const struct net *net = sock_net(sk);
210	u32 rcvwin, rcvbuf, cap, oldval;
211	u64 grow;
212
213	oldval = msk->rcvq_space.space;
214	msk->rcvq_space.space = newval;
215	if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) ||
216	(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
217	return false;
218
219	/* DRS is always one RTT late. */
220	rcvwin = newval << 1;
221
222	/* slow start: allow the sender to double its rate. */
223	grow = (u64)rcvwin * (newval - oldval);
224	do_div(grow, oldval);
225	rcvwin += grow << 1;
226
227	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue))
228	rcvwin += MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq - msk->ack_seq;
229
230	cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
231
232	rcvbuf = min_t(u32, mptcp_space_from_win(sk, rcvwin), cap);
233	if (rcvbuf > sk->sk_rcvbuf) {
234	WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
235	return true;
236	}
237	return false;
238	}
239
240	/* "inspired" by tcp_data_queue_ofo(), main differences:
241	* - use mptcp seqs
242	* - don't cope with sacks
243	*/
244	static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
245	{
246	struct sock *sk = (struct sock *)msk;
247	struct rb_node **p, *parent;
248	u64 seq, end_seq, max_seq;
249	struct sk_buff *skb1;
250
251	seq = MPTCP_SKB_CB(skb)->map_seq;
252	end_seq = MPTCP_SKB_CB(skb)->end_seq;
253	max_seq = atomic64_read(v: &msk->rcv_wnd_sent);
254
255	pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq,
256	RB_EMPTY_ROOT(&msk->out_of_order_queue));
257	if (after64(end_seq, max_seq)) {
258	/* out of window */
259	mptcp_drop(sk, skb);
260	pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
261	(unsigned long long)end_seq - (unsigned long)max_seq,
262	(unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
263	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_NODSSWINDOW);
264	return;
265	}
266
267	p = &msk->out_of_order_queue.rb_node;
268	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_OFOQUEUE);
269	if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
270	rb_link_node(node: &skb->rbnode, NULL, rb_link: p);
271	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
272	msk->ooo_last_skb = skb;
273	goto end;
274	}
275
276	/* with 2 subflows, adding at end of ooo queue is quite likely
277	* Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
278	*/
279	if (mptcp_ooo_try_coalesce(msk, to: msk->ooo_last_skb, from: skb)) {
280	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_OFOMERGE);
281	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_OFOQUEUETAIL);
282	return;
283	}
284
285	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
286	if (!before64(seq1: seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
287	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_OFOQUEUETAIL);
288	parent = &msk->ooo_last_skb->rbnode;
289	p = &parent->rb_right;
290	goto insert;
291	}
292
293	/* Find place to insert this segment. Handle overlaps on the way. */
294	parent = NULL;
295	while (*p) {
296	parent = *p;
297	skb1 = rb_to_skb(parent);
298	if (before64(seq1: seq, MPTCP_SKB_CB(skb1)->map_seq)) {
299	p = &parent->rb_left;
300	continue;
301	}
302	if (before64(seq1: seq, MPTCP_SKB_CB(skb1)->end_seq)) {
303	if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
304	/* All the bits are present. Drop. */
305	mptcp_drop(sk, skb);
306	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_DUPDATA);
307	return;
308	}
309	if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
310	/* partial overlap:
311	* | skb |
312	* | skb1 |
313	* continue traversing
314	*/
315	} else {
316	/* skb's seq == skb1's seq and skb covers skb1.
317	* Replace skb1 with skb.
318	*/
319	rb_replace_node(victim: &skb1->rbnode, new: &skb->rbnode,
320	root: &msk->out_of_order_queue);
321	mptcp_drop(sk, skb: skb1);
322	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_DUPDATA);
323	goto merge_right;
324	}
325	} else if (mptcp_ooo_try_coalesce(msk, to: skb1, from: skb)) {
326	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_OFOMERGE);
327	return;
328	}
329	p = &parent->rb_right;
330	}
331
332	insert:
333	/* Insert segment into RB tree. */
334	rb_link_node(node: &skb->rbnode, parent, rb_link: p);
335	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
336
337	merge_right:
338	/* Remove other segments covered by skb. */
339	while ((skb1 = skb_rb_next(skb)) != NULL) {
340	if (before64(seq1: end_seq, MPTCP_SKB_CB(skb1)->end_seq))
341	break;
342	rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
343	mptcp_drop(sk, skb: skb1);
344	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_DUPDATA);
345	}
346	/* If there is no skb after us, we are the last_skb ! */
347	if (!skb1)
348	msk->ooo_last_skb = skb;
349
350	end:
351	skb_condense(skb);
352	skb_set_owner_r(skb, sk);
353	/* do not grow rcvbuf for not-yet-accepted or orphaned sockets. */
354	if (sk->sk_socket)
355	mptcp_rcvbuf_grow(sk, newval: msk->rcvq_space.space);
356	}
357
358	static void mptcp_init_skb(struct sock *ssk, struct sk_buff *skb, int offset,
359	int copy_len)
360	{
361	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk: ssk);
362	bool has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
363
364	/* the skb map_seq accounts for the skb offset:
365	* mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
366	* value
367	*/
368	MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
369	MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
370	MPTCP_SKB_CB(skb)->offset = offset;
371	MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
372	MPTCP_SKB_CB(skb)->cant_coalesce = 0;
373
374	__skb_unlink(skb, list: &ssk->sk_receive_queue);
375
376	skb_ext_reset(skb);
377	skb_dst_drop(skb);
378	}
379
380	static bool __mptcp_move_skb(struct sock *sk, struct sk_buff *skb)
381	{
382	u64 copy_len = MPTCP_SKB_CB(skb)->end_seq - MPTCP_SKB_CB(skb)->map_seq;
383	struct mptcp_sock *msk = mptcp_sk(sk);
384	struct sk_buff *tail;
385
386	mptcp_borrow_fwdmem(sk, skb);
387
388	if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
389	/* in sequence */
390	msk->bytes_received += copy_len;
391	WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
392	tail = skb_peek_tail(list_: &sk->sk_receive_queue);
393	if (tail && mptcp_try_coalesce(sk, to: tail, from: skb))
394	return true;
395
396	skb_set_owner_r(skb, sk);
397	__skb_queue_tail(list: &sk->sk_receive_queue, newsk: skb);
398	return true;
399	} else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
400	mptcp_data_queue_ofo(msk, skb);
401	return false;
402	}
403
404	/* old data, keep it simple and drop the whole pkt, sender
405	* will retransmit as needed, if needed.
406	*/
407	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_DUPDATA);
408	mptcp_drop(sk, skb);
409	return false;
410	}
411
412	static void mptcp_stop_rtx_timer(struct sock *sk)
413	{
414	sk_stop_timer(sk, timer: &sk->mptcp_retransmit_timer);
415	mptcp_sk(sk)->timer_ival = 0;
416	}
417
418	static void mptcp_close_wake_up(struct sock *sk)
419	{
420	if (sock_flag(sk, flag: SOCK_DEAD))
421	return;
422
423	sk->sk_state_change(sk);
424	if (sk->sk_shutdown == SHUTDOWN_MASK ||
425	sk->sk_state == TCP_CLOSE)
426	sk_wake_async(sk, how: SOCK_WAKE_WAITD, POLL_HUP);
427	else
428	sk_wake_async(sk, how: SOCK_WAKE_WAITD, POLL_IN);
429	}
430
431	static void mptcp_shutdown_subflows(struct mptcp_sock *msk)
432	{
433	struct mptcp_subflow_context *subflow;
434
435	mptcp_for_each_subflow(msk, subflow) {
436	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
437	bool slow;
438
439	slow = lock_sock_fast(sk: ssk);
440	tcp_shutdown(sk: ssk, SEND_SHUTDOWN);
441	unlock_sock_fast(sk: ssk, slow);
442	}
443	}
444
445	/* called under the msk socket lock */
446	static bool mptcp_pending_data_fin_ack(struct sock *sk)
447	{
448	struct mptcp_sock *msk = mptcp_sk(sk);
449
450	return ((1 << sk->sk_state) &
451	(TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
452	msk->write_seq == READ_ONCE(msk->snd_una);
453	}
454
455	static void mptcp_check_data_fin_ack(struct sock *sk)
456	{
457	struct mptcp_sock *msk = mptcp_sk(sk);
458
459	/* Look for an acknowledged DATA_FIN */
460	if (mptcp_pending_data_fin_ack(sk)) {
461	WRITE_ONCE(msk->snd_data_fin_enable, 0);
462
463	switch (sk->sk_state) {
464	case TCP_FIN_WAIT1:
465	mptcp_set_state(sk, state: TCP_FIN_WAIT2);
466	break;
467	case TCP_CLOSING:
468	case TCP_LAST_ACK:
469	mptcp_shutdown_subflows(msk);
470	mptcp_set_state(sk, state: TCP_CLOSE);
471	break;
472	}
473
474	mptcp_close_wake_up(sk);
475	}
476	}
477
478	/* can be called with no lock acquired */
479	static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
480	{
481	struct mptcp_sock *msk = mptcp_sk(sk);
482
483	if (READ_ONCE(msk->rcv_data_fin) &&
484	((1 << inet_sk_state_load(sk)) &
485	(TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
486	u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
487
488	if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) {
489	if (seq)
490	*seq = rcv_data_fin_seq;
491
492	return true;
493	}
494	}
495
496	return false;
497	}
498
499	static void mptcp_set_datafin_timeout(struct sock *sk)
500	{
501	struct inet_connection_sock *icsk = inet_csk(sk);
502	u32 retransmits;
503
504	retransmits = min_t(u32, icsk->icsk_retransmits,
505	ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
506
507	mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
508	}
509
510	static void __mptcp_set_timeout(struct sock *sk, long tout)
511	{
512	mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
513	}
514
515	static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
516	{
517	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
518
519	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
520	tcp_timeout_expires(sk: ssk) - jiffies : 0;
521	}
522
523	static void mptcp_set_timeout(struct sock *sk)
524	{
525	struct mptcp_subflow_context *subflow;
526	long tout = 0;
527
528	mptcp_for_each_subflow(mptcp_sk(sk), subflow)
529	tout = max(tout, mptcp_timeout_from_subflow(subflow));
530	__mptcp_set_timeout(sk, tout);
531	}
532
533	static inline bool tcp_can_send_ack(const struct sock *ssk)
534	{
535	return !((1 << inet_sk_state_load(sk: ssk)) &
536	(TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
537	}
538
539	void __mptcp_subflow_send_ack(struct sock *ssk)
540	{
541	if (tcp_can_send_ack(ssk))
542	tcp_send_ack(sk: ssk);
543	}
544
545	static void mptcp_subflow_send_ack(struct sock *ssk)
546	{
547	bool slow;
548
549	slow = lock_sock_fast(sk: ssk);
550	__mptcp_subflow_send_ack(ssk);
551	unlock_sock_fast(sk: ssk, slow);
552	}
553
554	static void mptcp_send_ack(struct mptcp_sock *msk)
555	{
556	struct mptcp_subflow_context *subflow;
557
558	mptcp_for_each_subflow(msk, subflow)
559	mptcp_subflow_send_ack(ssk: mptcp_subflow_tcp_sock(subflow));
560	}
561
562	static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied)
563	{
564	bool slow;
565
566	slow = lock_sock_fast(sk: ssk);
567	if (tcp_can_send_ack(ssk))
568	tcp_cleanup_rbuf(sk: ssk, copied);
569	unlock_sock_fast(sk: ssk, slow);
570	}
571
572	static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
573	{
574	const struct inet_connection_sock *icsk = inet_csk(ssk);
575	u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
576	const struct tcp_sock *tp = tcp_sk(ssk);
577
578	return (ack_pending & ICSK_ACK_SCHED) &&
579	((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
580	READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
581	(rx_empty && ack_pending &
582	(ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
583	}
584
585	static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied)
586	{
587	int old_space = READ_ONCE(msk->old_wspace);
588	struct mptcp_subflow_context *subflow;
589	struct sock *sk = (struct sock *)msk;
590	int space = __mptcp_space(sk);
591	bool cleanup, rx_empty;
592
593	cleanup = (space > 0) && (space >= (old_space << 1)) && copied;
594	rx_empty = !sk_rmem_alloc_get(sk) && copied;
595
596	mptcp_for_each_subflow(msk, subflow) {
597	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
598
599	if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
600	mptcp_subflow_cleanup_rbuf(ssk, copied);
601	}
602	}
603
604	static void mptcp_check_data_fin(struct sock *sk)
605	{
606	struct mptcp_sock *msk = mptcp_sk(sk);
607	u64 rcv_data_fin_seq;
608
609	/* Need to ack a DATA_FIN received from a peer while this side
610	* of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
611	* msk->rcv_data_fin was set when parsing the incoming options
612	* at the subflow level and the msk lock was not held, so this
613	* is the first opportunity to act on the DATA_FIN and change
614	* the msk state.
615	*
616	* If we are caught up to the sequence number of the incoming
617	* DATA_FIN, send the DATA_ACK now and do state transition. If
618	* not caught up, do nothing and let the recv code send DATA_ACK
619	* when catching up.
620	*/
621
622	if (mptcp_pending_data_fin(sk, seq: &rcv_data_fin_seq)) {
623	WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
624	WRITE_ONCE(msk->rcv_data_fin, 0);
625
626	WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
627	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
628
629	switch (sk->sk_state) {
630	case TCP_ESTABLISHED:
631	mptcp_set_state(sk, state: TCP_CLOSE_WAIT);
632	break;
633	case TCP_FIN_WAIT1:
634	mptcp_set_state(sk, state: TCP_CLOSING);
635	break;
636	case TCP_FIN_WAIT2:
637	mptcp_shutdown_subflows(msk);
638	mptcp_set_state(sk, state: TCP_CLOSE);
639	break;
640	default:
641	/* Other states not expected */
642	WARN_ON_ONCE(1);
643	break;
644	}
645
646	if (!__mptcp_check_fallback(msk))
647	mptcp_send_ack(msk);
648	mptcp_close_wake_up(sk);
649	}
650	}
651
652	static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk)
653	{
654	if (!mptcp_try_fallback(ssk, fb_mib: MPTCP_MIB_DSSCORRUPTIONFALLBACK)) {
655	MPTCP_INC_STATS(net: sock_net(sk: ssk), field: MPTCP_MIB_DSSCORRUPTIONRESET);
656	mptcp_subflow_reset(ssk);
657	}
658	}
659
660	static void __mptcp_add_backlog(struct sock *sk,
661	struct mptcp_subflow_context *subflow,
662	struct sk_buff *skb)
663	{
664	struct mptcp_sock *msk = mptcp_sk(sk);
665	struct sk_buff *tail = NULL;
666	struct sock *ssk = skb->sk;
667	bool fragstolen;
668	int delta;
669
670	if (unlikely(sk->sk_state == TCP_CLOSE)) {
671	kfree_skb_reason(skb, reason: SKB_DROP_REASON_SOCKET_CLOSE);
672	return;
673	}
674
675	/* Try to coalesce with the last skb in our backlog */
676	if (!list_empty(head: &msk->backlog_list))
677	tail = list_last_entry(&msk->backlog_list, struct sk_buff, list);
678
679	if (tail && MPTCP_SKB_CB(skb)->map_seq == MPTCP_SKB_CB(tail)->end_seq &&
680	ssk == tail->sk &&
681	__mptcp_try_coalesce(sk, to: tail, from: skb, fragstolen: &fragstolen, delta: &delta)) {
682	skb->truesize -= delta;
683	kfree_skb_partial(skb, head_stolen: fragstolen);
684	__mptcp_subflow_lend_fwdmem(subflow, size: delta);
685	goto account;
686	}
687
688	list_add_tail(new: &skb->list, head: &msk->backlog_list);
689	mptcp_subflow_lend_fwdmem(subflow, skb);
690	delta = skb->truesize;
691
692	account:
693	WRITE_ONCE(msk->backlog_len, msk->backlog_len + delta);
694
695	/* Possibly not accept()ed yet, keep track of memory not CG
696	* accounted, mptcp_graft_subflows() will handle it.
697	*/
698	if (!mem_cgroup_from_sk(sk: ssk))
699	msk->backlog_unaccounted += delta;
700	}
701
702	static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
703	struct sock *ssk, bool own_msk)
704	{
705	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk: ssk);
706	struct sock *sk = (struct sock *)msk;
707	bool more_data_avail;
708	struct tcp_sock *tp;
709	bool ret = false;
710
711	pr_debug("msk=%p ssk=%p\n", msk, ssk);
712	tp = tcp_sk(ssk);
713	do {
714	u32 map_remaining, offset;
715	u32 seq = tp->copied_seq;
716	struct sk_buff *skb;
717	bool fin;
718
719	/* try to move as much data as available */
720	map_remaining = subflow->map_data_len -
721	mptcp_subflow_get_map_offset(subflow);
722
723	skb = skb_peek(list_: &ssk->sk_receive_queue);
724	if (unlikely(!skb))
725	break;
726
727	if (__mptcp_check_fallback(msk)) {
728	/* Under fallback skbs have no MPTCP extension and TCP could
729	* collapse them between the dummy map creation and the
730	* current dequeue. Be sure to adjust the map size.
731	*/
732	map_remaining = skb->len;
733	subflow->map_data_len = skb->len;
734	}
735
736	offset = seq - TCP_SKB_CB(skb)->seq;
737	fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
738	if (fin)
739	seq++;
740
741	if (offset < skb->len) {
742	size_t len = skb->len - offset;
743
744	mptcp_init_skb(ssk, skb, offset, copy_len: len);
745
746	if (own_msk && sk_rmem_alloc_get(sk) < sk->sk_rcvbuf) {
747	mptcp_subflow_lend_fwdmem(subflow, skb);
748	ret |= __mptcp_move_skb(sk, skb);
749	} else {
750	__mptcp_add_backlog(sk, subflow, skb);
751	}
752	seq += len;
753
754	if (unlikely(map_remaining < len)) {
755	DEBUG_NET_WARN_ON_ONCE(1);
756	mptcp_dss_corruption(msk, ssk);
757	}
758	} else {
759	if (unlikely(!fin)) {
760	DEBUG_NET_WARN_ON_ONCE(1);
761	mptcp_dss_corruption(msk, ssk);
762	}
763
764	sk_eat_skb(sk: ssk, skb);
765	}
766
767	WRITE_ONCE(tp->copied_seq, seq);
768	more_data_avail = mptcp_subflow_data_available(sk: ssk);
769
770	} while (more_data_avail);
771
772	if (ret)
773	msk->last_data_recv = tcp_jiffies32;
774	return ret;
775	}
776
777	static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
778	{
779	struct sock *sk = (struct sock *)msk;
780	struct sk_buff *skb, *tail;
781	bool moved = false;
782	struct rb_node *p;
783	u64 end_seq;
784
785	p = rb_first(root: &msk->out_of_order_queue);
786	pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
787	while (p) {
788	skb = rb_to_skb(p);
789	if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
790	break;
791
792	p = rb_next(p);
793	rb_erase(&skb->rbnode, &msk->out_of_order_queue);
794
795	if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
796	msk->ack_seq))) {
797	mptcp_drop(sk, skb);
798	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_DUPDATA);
799	continue;
800	}
801
802	end_seq = MPTCP_SKB_CB(skb)->end_seq;
803	tail = skb_peek_tail(list_: &sk->sk_receive_queue);
804	if (!tail || !mptcp_ooo_try_coalesce(msk, to: tail, from: skb)) {
805	int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
806
807	/* skip overlapping data, if any */
808	pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n",
809	MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
810	delta);
811	MPTCP_SKB_CB(skb)->offset += delta;
812	MPTCP_SKB_CB(skb)->map_seq += delta;
813	__skb_queue_tail(list: &sk->sk_receive_queue, newsk: skb);
814	}
815	msk->bytes_received += end_seq - msk->ack_seq;
816	WRITE_ONCE(msk->ack_seq, end_seq);
817	moved = true;
818	}
819	return moved;
820	}
821
822	static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
823	{
824	int ssk_state;
825	int err;
826
827	/* only propagate errors on fallen-back sockets or
828	* on MPC connect
829	*/
830	if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
831	return false;
832
833	err = sock_error(sk: ssk);
834	if (!err)
835	return false;
836
837	/* We need to propagate only transition to CLOSE state.
838	* Orphaned socket will see such state change via
839	* subflow_sched_work_if_closed() and that path will properly
840	* destroy the msk as needed.
841	*/
842	ssk_state = inet_sk_state_load(sk: ssk);
843	if (ssk_state == TCP_CLOSE && !sock_flag(sk, flag: SOCK_DEAD))
844	mptcp_set_state(sk, state: ssk_state);
845	WRITE_ONCE(sk->sk_err, -err);
846
847	/* This barrier is coupled with smp_rmb() in mptcp_poll() */
848	smp_wmb();
849	sk_error_report(sk);
850	return true;
851	}
852
853	void __mptcp_error_report(struct sock *sk)
854	{
855	struct mptcp_subflow_context *subflow;
856	struct mptcp_sock *msk = mptcp_sk(sk);
857
858	mptcp_for_each_subflow(msk, subflow)
859	if (__mptcp_subflow_error_report(sk, ssk: mptcp_subflow_tcp_sock(subflow)))
860	break;
861	}
862
863	/* In most cases we will be able to lock the mptcp socket. If its already
864	* owned, we need to defer to the work queue to avoid ABBA deadlock.
865	*/
866	static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
867	{
868	struct sock *sk = (struct sock *)msk;
869	bool moved;
870
871	moved = __mptcp_move_skbs_from_subflow(msk, ssk, own_msk: true);
872	__mptcp_ofo_queue(msk);
873	if (unlikely(ssk->sk_err))
874	__mptcp_subflow_error_report(sk, ssk);
875
876	/* If the moves have caught up with the DATA_FIN sequence number
877	* it's time to ack the DATA_FIN and change socket state, but
878	* this is not a good place to change state. Let the workqueue
879	* do it.
880	*/
881	if (mptcp_pending_data_fin(sk, NULL))
882	mptcp_schedule_work(sk);
883	return moved;
884	}
885
886	void mptcp_data_ready(struct sock *sk, struct sock *ssk)
887	{
888	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk: ssk);
889	struct mptcp_sock *msk = mptcp_sk(sk);
890
891	/* The peer can send data while we are shutting down this
892	* subflow at subflow destruction time, but we must avoid enqueuing
893	* more data to the msk receive queue
894	*/
895	if (unlikely(subflow->closing))
896	return;
897
898	mptcp_data_lock(sk);
899	if (!sock_owned_by_user(sk)) {
900	/* Wake-up the reader only for in-sequence data */
901	if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
902	sk->sk_data_ready(sk);
903	} else {
904	__mptcp_move_skbs_from_subflow(msk, ssk, own_msk: false);
905	}
906	mptcp_data_unlock(sk);
907	}
908
909	static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
910	{
911	mptcp_subflow_ctx(sk: ssk)->map_seq = READ_ONCE(msk->ack_seq);
912	msk->allow_infinite_fallback = false;
913	mptcp_event(type: MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
914	}
915
916	static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
917	{
918	struct sock *sk = (struct sock *)msk;
919
920	if (sk->sk_state != TCP_ESTABLISHED)
921	return false;
922
923	spin_lock_bh(lock: &msk->fallback_lock);
924	if (!msk->allow_subflows) {
925	spin_unlock_bh(lock: &msk->fallback_lock);
926	return false;
927	}
928	mptcp_subflow_joined(msk, ssk);
929	spin_unlock_bh(lock: &msk->fallback_lock);
930
931	mptcp_subflow_ctx(sk: ssk)->subflow_id = msk->subflow_id++;
932	mptcp_sockopt_sync_locked(msk, ssk);
933	mptcp_stop_tout_timer(sk);
934	__mptcp_propagate_sndbuf(sk, ssk);
935	return true;
936	}
937
938	static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
939	{
940	struct mptcp_subflow_context *tmp, *subflow;
941	struct mptcp_sock *msk = mptcp_sk(sk);
942
943	list_for_each_entry_safe(subflow, tmp, join_list, node) {
944	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
945	bool slow = lock_sock_fast(sk: ssk);
946
947	list_move_tail(list: &subflow->node, head: &msk->conn_list);
948	if (!__mptcp_finish_join(msk, ssk))
949	mptcp_subflow_reset(ssk);
950	unlock_sock_fast(sk: ssk, slow);
951	}
952	}
953
954	static bool mptcp_rtx_timer_pending(struct sock *sk)
955	{
956	return timer_pending(timer: &sk->mptcp_retransmit_timer);
957	}
958
959	static void mptcp_reset_rtx_timer(struct sock *sk)
960	{
961	unsigned long tout;
962
963	/* prevent rescheduling on close */
964	if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
965	return;
966
967	tout = mptcp_sk(sk)->timer_ival;
968	sk_reset_timer(sk, timer: &sk->mptcp_retransmit_timer, expires: jiffies + tout);
969	}
970
971	bool mptcp_schedule_work(struct sock *sk)
972	{
973	if (inet_sk_state_load(sk) == TCP_CLOSE)
974	return false;
975
976	/* Get a reference on this socket, mptcp_worker() will release it.
977	* As mptcp_worker() might complete before us, we can not avoid
978	* a sock_hold()/sock_put() if schedule_work() returns false.
979	*/
980	sock_hold(sk);
981
982	if (schedule_work(work: &mptcp_sk(sk)->work))
983	return true;
984
985	sock_put(sk);
986	return false;
987	}
988
989	static bool mptcp_skb_can_collapse_to(u64 write_seq,
990	const struct sk_buff *skb,
991	const struct mptcp_ext *mpext)
992	{
993	if (!tcp_skb_can_collapse_to(skb))
994	return false;
995
996	/* can collapse only if MPTCP level sequence is in order and this
997	* mapping has not been xmitted yet
998	*/
999	return mpext && mpext->data_seq + mpext->data_len == write_seq &&
1000	!mpext->frozen;
1001	}
1002
1003	/* we can append data to the given data frag if:
1004	* - there is space available in the backing page_frag
1005	* - the data frag tail matches the current page_frag free offset
1006	* - the data frag end sequence number matches the current write seq
1007	*/
1008	static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
1009	const struct page_frag *pfrag,
1010	const struct mptcp_data_frag *df)
1011	{
1012	return df && pfrag->page == df->page &&
1013	pfrag->size - pfrag->offset > 0 &&
1014	pfrag->offset == (df->offset + df->data_len) &&
1015	df->data_seq + df->data_len == msk->write_seq;
1016	}
1017
1018	static void dfrag_uncharge(struct sock *sk, int len)
1019	{
1020	sk_mem_uncharge(sk, size: len);
1021	sk_wmem_queued_add(sk, val: -len);
1022	}
1023
1024	static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1025	{
1026	int len = dfrag->data_len + dfrag->overhead;
1027
1028	list_del(entry: &dfrag->list);
1029	dfrag_uncharge(sk, len);
1030	put_page(page: dfrag->page);
1031	}
1032
1033	/* called under both the msk socket lock and the data lock */
1034	static void __mptcp_clean_una(struct sock *sk)
1035	{
1036	struct mptcp_sock *msk = mptcp_sk(sk);
1037	struct mptcp_data_frag *dtmp, *dfrag;
1038	u64 snd_una;
1039
1040	snd_una = msk->snd_una;
1041	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1042	if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1043	break;
1044
1045	if (unlikely(dfrag == msk->first_pending)) {
1046	/* in recovery mode can see ack after the current snd head */
1047	if (WARN_ON_ONCE(!msk->recovery))
1048	break;
1049
1050	msk->first_pending = mptcp_send_next(sk);
1051	}
1052
1053	dfrag_clear(sk, dfrag);
1054	}
1055
1056	dfrag = mptcp_rtx_head(sk);
1057	if (dfrag && after64(snd_una, dfrag->data_seq)) {
1058	u64 delta = snd_una - dfrag->data_seq;
1059
1060	/* prevent wrap around in recovery mode */
1061	if (unlikely(delta > dfrag->already_sent)) {
1062	if (WARN_ON_ONCE(!msk->recovery))
1063	goto out;
1064	if (WARN_ON_ONCE(delta > dfrag->data_len))
1065	goto out;
1066	dfrag->already_sent += delta - dfrag->already_sent;
1067	}
1068
1069	dfrag->data_seq += delta;
1070	dfrag->offset += delta;
1071	dfrag->data_len -= delta;
1072	dfrag->already_sent -= delta;
1073
1074	dfrag_uncharge(sk, len: delta);
1075	}
1076
1077	/* all retransmitted data acked, recovery completed */
1078	if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1079	msk->recovery = false;
1080
1081	out:
1082	if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) {
1083	if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1084	mptcp_stop_rtx_timer(sk);
1085	} else {
1086	mptcp_reset_rtx_timer(sk);
1087	}
1088
1089	if (mptcp_pending_data_fin_ack(sk))
1090	mptcp_schedule_work(sk);
1091	}
1092
1093	static void __mptcp_clean_una_wakeup(struct sock *sk)
1094	{
1095	lockdep_assert_held_once(&sk->sk_lock.slock);
1096
1097	__mptcp_clean_una(sk);
1098	mptcp_write_space(sk);
1099	}
1100
1101	static void mptcp_clean_una_wakeup(struct sock *sk)
1102	{
1103	mptcp_data_lock(sk);
1104	__mptcp_clean_una_wakeup(sk);
1105	mptcp_data_unlock(sk);
1106	}
1107
1108	static void mptcp_enter_memory_pressure(struct sock *sk)
1109	{
1110	struct mptcp_subflow_context *subflow;
1111	struct mptcp_sock *msk = mptcp_sk(sk);
1112	bool first = true;
1113
1114	mptcp_for_each_subflow(msk, subflow) {
1115	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1116
1117	if (first && !ssk->sk_bypass_prot_mem) {
1118	tcp_enter_memory_pressure(sk: ssk);
1119	first = false;
1120	}
1121
1122	sk_stream_moderate_sndbuf(sk: ssk);
1123	}
1124	__mptcp_sync_sndbuf(sk);
1125	}
1126
1127	/* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1128	* data
1129	*/
1130	static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1131	{
1132	if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1133	pfrag, sk->sk_allocation)))
1134	return true;
1135
1136	mptcp_enter_memory_pressure(sk);
1137	return false;
1138	}
1139
1140	static struct mptcp_data_frag *
1141	mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1142	int orig_offset)
1143	{
1144	int offset = ALIGN(orig_offset, sizeof(long));
1145	struct mptcp_data_frag *dfrag;
1146
1147	dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1148	dfrag->data_len = 0;
1149	dfrag->data_seq = msk->write_seq;
1150	dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1151	dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1152	dfrag->already_sent = 0;
1153	dfrag->page = pfrag->page;
1154
1155	return dfrag;
1156	}
1157
1158	struct mptcp_sendmsg_info {
1159	int mss_now;
1160	int size_goal;
1161	u16 limit;
1162	u16 sent;
1163	unsigned int flags;
1164	bool data_lock_held;
1165	};
1166
1167	static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1168	u64 data_seq, int avail_size)
1169	{
1170	u64 window_end = mptcp_wnd_end(msk);
1171	u64 mptcp_snd_wnd;
1172
1173	if (__mptcp_check_fallback(msk))
1174	return avail_size;
1175
1176	mptcp_snd_wnd = window_end - data_seq;
1177	avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1178
1179	if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1180	tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1181	MPTCP_INC_STATS(net: sock_net(sk: ssk), field: MPTCP_MIB_SNDWNDSHARED);
1182	}
1183
1184	return avail_size;
1185	}
1186
1187	static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1188	{
1189	struct skb_ext *mpext = __skb_ext_alloc(flags: gfp);
1190
1191	if (!mpext)
1192	return false;
1193	__skb_ext_set(skb, id: SKB_EXT_MPTCP, ext: mpext);
1194	return true;
1195	}
1196
1197	static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1198	{
1199	struct sk_buff *skb;
1200
1201	skb = alloc_skb_fclone(MAX_TCP_HEADER, priority: gfp);
1202	if (likely(skb)) {
1203	if (likely(__mptcp_add_ext(skb, gfp))) {
1204	skb_reserve(skb, MAX_TCP_HEADER);
1205	skb->ip_summed = CHECKSUM_PARTIAL;
1206	INIT_LIST_HEAD(list: &skb->tcp_tsorted_anchor);
1207	return skb;
1208	}
1209	__kfree_skb(skb);
1210	} else {
1211	mptcp_enter_memory_pressure(sk);
1212	}
1213	return NULL;
1214	}
1215
1216	static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1217	{
1218	struct sk_buff *skb;
1219
1220	skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1221	if (!skb)
1222	return NULL;
1223
1224	if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1225	tcp_skb_entail(sk: ssk, skb);
1226	return skb;
1227	}
1228	tcp_skb_tsorted_anchor_cleanup(skb);
1229	kfree_skb(skb);
1230	return NULL;
1231	}
1232
1233	static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1234	{
1235	gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1236
1237	return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1238	}
1239
1240	/* note: this always recompute the csum on the whole skb, even
1241	* if we just appended a single frag. More status info needed
1242	*/
1243	static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1244	{
1245	struct mptcp_ext *mpext = mptcp_get_ext(skb);
1246	__wsum csum = ~csum_unfold(n: mpext->csum);
1247	int offset = skb->len - added;
1248
1249	mpext->csum = csum_fold(csum: csum_block_add(csum, csum2: skb_checksum(skb, offset, len: added, csum: 0), offset));
1250	}
1251
1252	static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1253	struct sock *ssk,
1254	struct mptcp_ext *mpext)
1255	{
1256	if (!mpext)
1257	return;
1258
1259	mpext->infinite_map = 1;
1260	mpext->data_len = 0;
1261
1262	if (!mptcp_try_fallback(ssk, fb_mib: MPTCP_MIB_INFINITEMAPTX)) {
1263	MPTCP_INC_STATS(net: sock_net(sk: ssk), field: MPTCP_MIB_FALLBACKFAILED);
1264	mptcp_subflow_reset(ssk);
1265	return;
1266	}
1267
1268	mptcp_subflow_ctx(sk: ssk)->send_infinite_map = 0;
1269	}
1270
1271	#define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1272
1273	static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1274	struct mptcp_data_frag *dfrag,
1275	struct mptcp_sendmsg_info *info)
1276	{
1277	u64 data_seq = dfrag->data_seq + info->sent;
1278	int offset = dfrag->offset + info->sent;
1279	struct mptcp_sock *msk = mptcp_sk(sk);
1280	bool zero_window_probe = false;
1281	struct mptcp_ext *mpext = NULL;
1282	bool can_coalesce = false;
1283	bool reuse_skb = true;
1284	struct sk_buff *skb;
1285	size_t copy;
1286	int i;
1287
1288	pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n",
1289	msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1290
1291	if (WARN_ON_ONCE(info->sent > info->limit ||
1292	info->limit > dfrag->data_len))
1293	return 0;
1294
1295	if (unlikely(!__tcp_can_send(ssk)))
1296	return -EAGAIN;
1297
1298	/* compute send limit */
1299	if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1300	ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1301	info->mss_now = tcp_send_mss(sk: ssk, size_goal: &info->size_goal, flags: info->flags);
1302	copy = info->size_goal;
1303
1304	skb = tcp_write_queue_tail(sk: ssk);
1305	if (skb && copy > skb->len) {
1306	/* Limit the write to the size available in the
1307	* current skb, if any, so that we create at most a new skb.
1308	* Explicitly tells TCP internals to avoid collapsing on later
1309	* queue management operation, to avoid breaking the ext <->
1310	* SSN association set here
1311	*/
1312	mpext = mptcp_get_ext(skb);
1313	if (!mptcp_skb_can_collapse_to(write_seq: data_seq, skb, mpext)) {
1314	TCP_SKB_CB(skb)->eor = 1;
1315	tcp_mark_push(tcp_sk(ssk), skb);
1316	goto alloc_skb;
1317	}
1318
1319	i = skb_shinfo(skb)->nr_frags;
1320	can_coalesce = skb_can_coalesce(skb, i, page: dfrag->page, off: offset);
1321	if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1322	tcp_mark_push(tcp_sk(ssk), skb);
1323	goto alloc_skb;
1324	}
1325
1326	copy -= skb->len;
1327	} else {
1328	alloc_skb:
1329	skb = mptcp_alloc_tx_skb(sk, ssk, data_lock_held: info->data_lock_held);
1330	if (!skb)
1331	return -ENOMEM;
1332
1333	i = skb_shinfo(skb)->nr_frags;
1334	reuse_skb = false;
1335	mpext = mptcp_get_ext(skb);
1336	}
1337
1338	/* Zero window and all data acked? Probe. */
1339	copy = mptcp_check_allowed_size(msk, ssk, data_seq, avail_size: copy);
1340	if (copy == 0) {
1341	u64 snd_una = READ_ONCE(msk->snd_una);
1342
1343	/* No need for zero probe if there are any data pending
1344	* either at the msk or ssk level; skb is the current write
1345	* queue tail and can be empty at this point.
1346	*/
1347	if (snd_una != msk->snd_nxt || skb->len ||
1348	skb != tcp_send_head(sk: ssk)) {
1349	tcp_remove_empty_skb(sk: ssk);
1350	return 0;
1351	}
1352
1353	zero_window_probe = true;
1354	data_seq = snd_una - 1;
1355	copy = 1;
1356	}
1357
1358	copy = min_t(size_t, copy, info->limit - info->sent);
1359	if (!sk_wmem_schedule(sk: ssk, size: copy)) {
1360	tcp_remove_empty_skb(sk: ssk);
1361	return -ENOMEM;
1362	}
1363
1364	if (can_coalesce) {
1365	skb_frag_size_add(frag: &skb_shinfo(skb)->frags[i - 1], delta: copy);
1366	} else {
1367	get_page(page: dfrag->page);
1368	skb_fill_page_desc(skb, i, page: dfrag->page, off: offset, size: copy);
1369	}
1370
1371	skb->len += copy;
1372	skb->data_len += copy;
1373	skb->truesize += copy;
1374	sk_wmem_queued_add(sk: ssk, val: copy);
1375	sk_mem_charge(sk: ssk, size: copy);
1376	WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1377	TCP_SKB_CB(skb)->end_seq += copy;
1378	tcp_skb_pcount_set(skb, segs: 0);
1379
1380	/* on skb reuse we just need to update the DSS len */
1381	if (reuse_skb) {
1382	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1383	mpext->data_len += copy;
1384	goto out;
1385	}
1386
1387	memset(mpext, 0, sizeof(*mpext));
1388	mpext->data_seq = data_seq;
1389	mpext->subflow_seq = mptcp_subflow_ctx(sk: ssk)->rel_write_seq;
1390	mpext->data_len = copy;
1391	mpext->use_map = 1;
1392	mpext->dsn64 = 1;
1393
1394	pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n",
1395	mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1396	mpext->dsn64);
1397
1398	if (zero_window_probe) {
1399	MPTCP_INC_STATS(net: sock_net(sk: ssk), field: MPTCP_MIB_WINPROBE);
1400	mptcp_subflow_ctx(sk: ssk)->rel_write_seq += copy;
1401	mpext->frozen = 1;
1402	if (READ_ONCE(msk->csum_enabled))
1403	mptcp_update_data_checksum(skb, added: copy);
1404	tcp_push_pending_frames(sk: ssk);
1405	return 0;
1406	}
1407	out:
1408	if (READ_ONCE(msk->csum_enabled))
1409	mptcp_update_data_checksum(skb, added: copy);
1410	if (mptcp_subflow_ctx(sk: ssk)->send_infinite_map)
1411	mptcp_update_infinite_map(msk, ssk, mpext);
1412	trace_mptcp_sendmsg_frag(mpext);
1413	mptcp_subflow_ctx(sk: ssk)->rel_write_seq += copy;
1414	return copy;
1415	}
1416
1417	#define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1418	sizeof(struct tcphdr) - \
1419	MAX_TCP_OPTION_SPACE - \
1420	sizeof(struct ipv6hdr) - \
1421	sizeof(struct frag_hdr))
1422
1423	struct subflow_send_info {
1424	struct sock *ssk;
1425	u64 linger_time;
1426	};
1427
1428	void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1429	{
1430	if (!subflow->stale)
1431	return;
1432
1433	subflow->stale = 0;
1434	MPTCP_INC_STATS(net: sock_net(sk: mptcp_subflow_tcp_sock(subflow)), field: MPTCP_MIB_SUBFLOWRECOVER);
1435	}
1436
1437	bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1438	{
1439	if (unlikely(subflow->stale)) {
1440	u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1441
1442	if (subflow->stale_rcv_tstamp == rcv_tstamp)
1443	return false;
1444
1445	mptcp_subflow_set_active(subflow);
1446	}
1447	return __mptcp_subflow_active(subflow);
1448	}
1449
1450	#define SSK_MODE_ACTIVE 0
1451	#define SSK_MODE_BACKUP 1
1452	#define SSK_MODE_MAX 2
1453
1454	/* implement the mptcp packet scheduler;
1455	* returns the subflow that will transmit the next DSS
1456	* additionally updates the rtx timeout
1457	*/
1458	struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1459	{
1460	struct subflow_send_info send_info[SSK_MODE_MAX];
1461	struct mptcp_subflow_context *subflow;
1462	struct sock *sk = (struct sock *)msk;
1463	u32 pace, burst, wmem;
1464	int i, nr_active = 0;
1465	struct sock *ssk;
1466	u64 linger_time;
1467	long tout = 0;
1468
1469	/* pick the subflow with the lower wmem/wspace ratio */
1470	for (i = 0; i < SSK_MODE_MAX; ++i) {
1471	send_info[i].ssk = NULL;
1472	send_info[i].linger_time = -1;
1473	}
1474
1475	mptcp_for_each_subflow(msk, subflow) {
1476	bool backup = subflow->backup || subflow->request_bkup;
1477
1478	trace_mptcp_subflow_get_send(subflow);
1479	ssk = mptcp_subflow_tcp_sock(subflow);
1480	if (!mptcp_subflow_active(subflow))
1481	continue;
1482
1483	tout = max(tout, mptcp_timeout_from_subflow(subflow));
1484	nr_active += !backup;
1485	pace = subflow->avg_pacing_rate;
1486	if (unlikely(!pace)) {
1487	/* init pacing rate from socket */
1488	subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1489	pace = subflow->avg_pacing_rate;
1490	if (!pace)
1491	continue;
1492	}
1493
1494	linger_time = div_u64(dividend: (u64)READ_ONCE(ssk->sk_wmem_queued) << 32, divisor: pace);
1495	if (linger_time < send_info[backup].linger_time) {
1496	send_info[backup].ssk = ssk;
1497	send_info[backup].linger_time = linger_time;
1498	}
1499	}
1500	__mptcp_set_timeout(sk, tout);
1501
1502	/* pick the best backup if no other subflow is active */
1503	if (!nr_active)
1504	send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1505
1506	/* According to the blest algorithm, to avoid HoL blocking for the
1507	* faster flow, we need to:
1508	* - estimate the faster flow linger time
1509	* - use the above to estimate the amount of byte transferred
1510	* by the faster flow
1511	* - check that the amount of queued data is greater than the above,
1512	* otherwise do not use the picked, slower, subflow
1513	* We select the subflow with the shorter estimated time to flush
1514	* the queued mem, which basically ensure the above. We just need
1515	* to check that subflow has a non empty cwin.
1516	*/
1517	ssk = send_info[SSK_MODE_ACTIVE].ssk;
1518	if (!ssk || !sk_stream_memory_free(sk: ssk))
1519	return NULL;
1520
1521	burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1522	wmem = READ_ONCE(ssk->sk_wmem_queued);
1523	if (!burst)
1524	return ssk;
1525
1526	subflow = mptcp_subflow_ctx(sk: ssk);
1527	subflow->avg_pacing_rate = div_u64(dividend: (u64)subflow->avg_pacing_rate * wmem +
1528	READ_ONCE(ssk->sk_pacing_rate) * burst,
1529	divisor: burst + wmem);
1530	msk->snd_burst = burst;
1531	return ssk;
1532	}
1533
1534	static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1535	{
1536	tcp_push(sk: ssk, flags: 0, mss_now: info->mss_now, tcp_sk(ssk)->nonagle, size_goal: info->size_goal);
1537	release_sock(sk: ssk);
1538	}
1539
1540	static void mptcp_update_post_push(struct mptcp_sock *msk,
1541	struct mptcp_data_frag *dfrag,
1542	u32 sent)
1543	{
1544	u64 snd_nxt_new = dfrag->data_seq;
1545
1546	dfrag->already_sent += sent;
1547
1548	msk->snd_burst -= sent;
1549
1550	snd_nxt_new += dfrag->already_sent;
1551
1552	/* snd_nxt_new can be smaller than snd_nxt in case mptcp
1553	* is recovering after a failover. In that event, this re-sends
1554	* old segments.
1555	*
1556	* Thus compute snd_nxt_new candidate based on
1557	* the dfrag->data_seq that was sent and the data
1558	* that has been handed to the subflow for transmission
1559	* and skip update in case it was old dfrag.
1560	*/
1561	if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1562	msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1563	WRITE_ONCE(msk->snd_nxt, snd_nxt_new);
1564	}
1565	}
1566
1567	void mptcp_check_and_set_pending(struct sock *sk)
1568	{
1569	if (mptcp_send_head(sk)) {
1570	mptcp_data_lock(sk);
1571	mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1572	mptcp_data_unlock(sk);
1573	}
1574	}
1575
1576	static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1577	struct mptcp_sendmsg_info *info)
1578	{
1579	struct mptcp_sock *msk = mptcp_sk(sk);
1580	struct mptcp_data_frag *dfrag;
1581	int len, copied = 0, err = 0;
1582
1583	while ((dfrag = mptcp_send_head(sk))) {
1584	info->sent = dfrag->already_sent;
1585	info->limit = dfrag->data_len;
1586	len = dfrag->data_len - dfrag->already_sent;
1587	while (len > 0) {
1588	int ret = 0;
1589
1590	ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1591	if (ret <= 0) {
1592	err = copied ? : ret;
1593	goto out;
1594	}
1595
1596	info->sent += ret;
1597	copied += ret;
1598	len -= ret;
1599
1600	mptcp_update_post_push(msk, dfrag, sent: ret);
1601	}
1602	msk->first_pending = mptcp_send_next(sk);
1603
1604	if (msk->snd_burst <= 0 ||
1605	!sk_stream_memory_free(sk: ssk) ||
1606	!mptcp_subflow_active(subflow: mptcp_subflow_ctx(sk: ssk))) {
1607	err = copied;
1608	goto out;
1609	}
1610	mptcp_set_timeout(sk);
1611	}
1612	err = copied;
1613
1614	out:
1615	if (err > 0)
1616	msk->last_data_sent = tcp_jiffies32;
1617	return err;
1618	}
1619
1620	void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1621	{
1622	struct sock *prev_ssk = NULL, *ssk = NULL;
1623	struct mptcp_sock *msk = mptcp_sk(sk);
1624	struct mptcp_sendmsg_info info = {
1625	.flags = flags,
1626	};
1627	bool copied = false;
1628	int push_count = 1;
1629
1630	while (mptcp_send_head(sk) && (push_count > 0)) {
1631	struct mptcp_subflow_context *subflow;
1632	int ret = 0;
1633
1634	if (mptcp_sched_get_send(msk))
1635	break;
1636
1637	push_count = 0;
1638
1639	mptcp_for_each_subflow(msk, subflow) {
1640	if (READ_ONCE(subflow->scheduled)) {
1641	mptcp_subflow_set_scheduled(subflow, scheduled: false);
1642
1643	prev_ssk = ssk;
1644	ssk = mptcp_subflow_tcp_sock(subflow);
1645	if (ssk != prev_ssk) {
1646	/* First check. If the ssk has changed since
1647	* the last round, release prev_ssk
1648	*/
1649	if (prev_ssk)
1650	mptcp_push_release(ssk: prev_ssk, info: &info);
1651
1652	/* Need to lock the new subflow only if different
1653	* from the previous one, otherwise we are still
1654	* helding the relevant lock
1655	*/
1656	lock_sock(sk: ssk);
1657	}
1658
1659	push_count++;
1660
1661	ret = __subflow_push_pending(sk, ssk, info: &info);
1662	if (ret <= 0) {
1663	if (ret != -EAGAIN ||
1664	(1 << ssk->sk_state) &
1665	(TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1666	push_count--;
1667	continue;
1668	}
1669	copied = true;
1670	}
1671	}
1672	}
1673
1674	/* at this point we held the socket lock for the last subflow we used */
1675	if (ssk)
1676	mptcp_push_release(ssk, info: &info);
1677
1678	/* Avoid scheduling the rtx timer if no data has been pushed; the timer
1679	* will be updated on positive acks by __mptcp_cleanup_una().
1680	*/
1681	if (copied) {
1682	if (!mptcp_rtx_timer_pending(sk))
1683	mptcp_reset_rtx_timer(sk);
1684	mptcp_check_send_data_fin(sk);
1685	}
1686	}
1687
1688	static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1689	{
1690	struct mptcp_sock *msk = mptcp_sk(sk);
1691	struct mptcp_sendmsg_info info = {
1692	.data_lock_held = true,
1693	};
1694	bool keep_pushing = true;
1695	struct sock *xmit_ssk;
1696	int copied = 0;
1697
1698	info.flags = 0;
1699	while (mptcp_send_head(sk) && keep_pushing) {
1700	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk: ssk);
1701	int ret = 0;
1702
1703	/* check for a different subflow usage only after
1704	* spooling the first chunk of data
1705	*/
1706	if (first) {
1707	mptcp_subflow_set_scheduled(subflow, scheduled: false);
1708	ret = __subflow_push_pending(sk, ssk, info: &info);
1709	first = false;
1710	if (ret <= 0)
1711	break;
1712	copied += ret;
1713	continue;
1714	}
1715
1716	if (mptcp_sched_get_send(msk))
1717	goto out;
1718
1719	if (READ_ONCE(subflow->scheduled)) {
1720	mptcp_subflow_set_scheduled(subflow, scheduled: false);
1721	ret = __subflow_push_pending(sk, ssk, info: &info);
1722	if (ret <= 0)
1723	keep_pushing = false;
1724	copied += ret;
1725	}
1726
1727	mptcp_for_each_subflow(msk, subflow) {
1728	if (READ_ONCE(subflow->scheduled)) {
1729	xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1730	if (xmit_ssk != ssk) {
1731	mptcp_subflow_delegate(subflow,
1732	MPTCP_DELEGATE_SEND);
1733	keep_pushing = false;
1734	}
1735	}
1736	}
1737	}
1738
1739	out:
1740	/* __mptcp_alloc_tx_skb could have released some wmem and we are
1741	* not going to flush it via release_sock()
1742	*/
1743	if (copied) {
1744	tcp_push(sk: ssk, flags: 0, mss_now: info.mss_now, tcp_sk(ssk)->nonagle,
1745	size_goal: info.size_goal);
1746	if (!mptcp_rtx_timer_pending(sk))
1747	mptcp_reset_rtx_timer(sk);
1748
1749	if (msk->snd_data_fin_enable &&
1750	msk->snd_nxt + 1 == msk->write_seq)
1751	mptcp_schedule_work(sk);
1752	}
1753	}
1754
1755	static int mptcp_disconnect(struct sock *sk, int flags);
1756
1757	static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1758	size_t len, int *copied_syn)
1759	{
1760	unsigned int saved_flags = msg->msg_flags;
1761	struct mptcp_sock *msk = mptcp_sk(sk);
1762	struct sock *ssk;
1763	int ret;
1764
1765	/* on flags based fastopen the mptcp is supposed to create the
1766	* first subflow right now. Otherwise we are in the defer_connect
1767	* path, and the first subflow must be already present.
1768	* Since the defer_connect flag is cleared after the first succsful
1769	* fastopen attempt, no need to check for additional subflow status.
1770	*/
1771	if (msg->msg_flags & MSG_FASTOPEN) {
1772	ssk = __mptcp_nmpc_sk(msk);
1773	if (IS_ERR(ptr: ssk))
1774	return PTR_ERR(ptr: ssk);
1775	}
1776	if (!msk->first)
1777	return -EINVAL;
1778
1779	ssk = msk->first;
1780
1781	lock_sock(sk: ssk);
1782	msg->msg_flags |= MSG_DONTWAIT;
1783	msk->fastopening = 1;
1784	ret = tcp_sendmsg_fastopen(sk: ssk, msg, copied: copied_syn, size: len, NULL);
1785	msk->fastopening = 0;
1786	msg->msg_flags = saved_flags;
1787	release_sock(sk: ssk);
1788
1789	/* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1790	if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1791	ret = __inet_stream_connect(sock: sk->sk_socket, uaddr: msg->msg_name,
1792	addr_len: msg->msg_namelen, flags: msg->msg_flags, is_sendmsg: 1);
1793
1794	/* Keep the same behaviour of plain TCP: zero the copied bytes in
1795	* case of any error, except timeout or signal
1796	*/
1797	if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1798	*copied_syn = 0;
1799	} else if (ret && ret != -EINPROGRESS) {
1800	/* The disconnect() op called by tcp_sendmsg_fastopen()/
1801	* __inet_stream_connect() can fail, due to looking check,
1802	* see mptcp_disconnect().
1803	* Attempt it again outside the problematic scope.
1804	*/
1805	if (!mptcp_disconnect(sk, flags: 0)) {
1806	sk->sk_disconnects++;
1807	sk->sk_socket->state = SS_UNCONNECTED;
1808	}
1809	}
1810	inet_clear_bit(DEFER_CONNECT, sk);
1811
1812	return ret;
1813	}
1814
1815	static int do_copy_data_nocache(struct sock *sk, int copy,
1816	struct iov_iter *from, char *to)
1817	{
1818	if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1819	if (!copy_from_iter_full_nocache(addr: to, bytes: copy, i: from))
1820	return -EFAULT;
1821	} else if (!copy_from_iter_full(addr: to, bytes: copy, i: from)) {
1822	return -EFAULT;
1823	}
1824	return 0;
1825	}
1826
1827	/* open-code sk_stream_memory_free() plus sent limit computation to
1828	* avoid indirect calls in fast-path.
1829	* Called under the msk socket lock, so we can avoid a bunch of ONCE
1830	* annotations.
1831	*/
1832	static u32 mptcp_send_limit(const struct sock *sk)
1833	{
1834	const struct mptcp_sock *msk = mptcp_sk(sk);
1835	u32 limit, not_sent;
1836
1837	if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
1838	return 0;
1839
1840	limit = mptcp_notsent_lowat(sk);
1841	if (limit == UINT_MAX)
1842	return UINT_MAX;
1843
1844	not_sent = msk->write_seq - msk->snd_nxt;
1845	if (not_sent >= limit)
1846	return 0;
1847
1848	return limit - not_sent;
1849	}
1850
1851	static void mptcp_rps_record_subflows(const struct mptcp_sock *msk)
1852	{
1853	struct mptcp_subflow_context *subflow;
1854
1855	if (!rfs_is_needed())
1856	return;
1857
1858	mptcp_for_each_subflow(msk, subflow) {
1859	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1860
1861	sock_rps_record_flow(sk: ssk);
1862	}
1863	}
1864
1865	static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1866	{
1867	struct mptcp_sock *msk = mptcp_sk(sk);
1868	struct page_frag *pfrag;
1869	size_t copied = 0;
1870	int ret = 0;
1871	long timeo;
1872
1873	/* silently ignore everything else */
1874	msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1875
1876	lock_sock(sk);
1877
1878	mptcp_rps_record_subflows(msk);
1879
1880	if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1881	msg->msg_flags & MSG_FASTOPEN)) {
1882	int copied_syn = 0;
1883
1884	ret = mptcp_sendmsg_fastopen(sk, msg, len, copied_syn: &copied_syn);
1885	copied += copied_syn;
1886	if (ret == -EINPROGRESS && copied_syn > 0)
1887	goto out;
1888	else if (ret)
1889	goto do_error;
1890	}
1891
1892	timeo = sock_sndtimeo(sk, noblock: msg->msg_flags & MSG_DONTWAIT);
1893
1894	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1895	ret = sk_stream_wait_connect(sk, timeo_p: &timeo);
1896	if (ret)
1897	goto do_error;
1898	}
1899
1900	ret = -EPIPE;
1901	if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1902	goto do_error;
1903
1904	pfrag = sk_page_frag(sk);
1905
1906	while (msg_data_left(msg)) {
1907	int total_ts, frag_truesize = 0;
1908	struct mptcp_data_frag *dfrag;
1909	bool dfrag_collapsed;
1910	size_t psize, offset;
1911	u32 copy_limit;
1912
1913	/* ensure fitting the notsent_lowat() constraint */
1914	copy_limit = mptcp_send_limit(sk);
1915	if (!copy_limit)
1916	goto wait_for_memory;
1917
1918	/* reuse tail pfrag, if possible, or carve a new one from the
1919	* page allocator
1920	*/
1921	dfrag = mptcp_pending_tail(sk);
1922	dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, df: dfrag);
1923	if (!dfrag_collapsed) {
1924	if (!mptcp_page_frag_refill(sk, pfrag))
1925	goto wait_for_memory;
1926
1927	dfrag = mptcp_carve_data_frag(msk, pfrag, orig_offset: pfrag->offset);
1928	frag_truesize = dfrag->overhead;
1929	}
1930
1931	/* we do not bound vs wspace, to allow a single packet.
1932	* memory accounting will prevent execessive memory usage
1933	* anyway
1934	*/
1935	offset = dfrag->offset + dfrag->data_len;
1936	psize = pfrag->size - offset;
1937	psize = min_t(size_t, psize, msg_data_left(msg));
1938	psize = min_t(size_t, psize, copy_limit);
1939	total_ts = psize + frag_truesize;
1940
1941	if (!sk_wmem_schedule(sk, size: total_ts))
1942	goto wait_for_memory;
1943
1944	ret = do_copy_data_nocache(sk, copy: psize, from: &msg->msg_iter,
1945	page_address(dfrag->page) + offset);
1946	if (ret)
1947	goto do_error;
1948
1949	/* data successfully copied into the write queue */
1950	sk_forward_alloc_add(sk, val: -total_ts);
1951	copied += psize;
1952	dfrag->data_len += psize;
1953	frag_truesize += psize;
1954	pfrag->offset += frag_truesize;
1955	WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1956
1957	/* charge data on mptcp pending queue to the msk socket
1958	* Note: we charge such data both to sk and ssk
1959	*/
1960	sk_wmem_queued_add(sk, val: frag_truesize);
1961	if (!dfrag_collapsed) {
1962	get_page(page: dfrag->page);
1963	list_add_tail(new: &dfrag->list, head: &msk->rtx_queue);
1964	if (!msk->first_pending)
1965	msk->first_pending = dfrag;
1966	}
1967	pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk,
1968	dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1969	!dfrag_collapsed);
1970
1971	continue;
1972
1973	wait_for_memory:
1974	set_bit(nr: SOCK_NOSPACE, addr: &sk->sk_socket->flags);
1975	__mptcp_push_pending(sk, flags: msg->msg_flags);
1976	ret = sk_stream_wait_memory(sk, timeo_p: &timeo);
1977	if (ret)
1978	goto do_error;
1979	}
1980
1981	if (copied)
1982	__mptcp_push_pending(sk, flags: msg->msg_flags);
1983
1984	out:
1985	release_sock(sk);
1986	return copied;
1987
1988	do_error:
1989	if (copied)
1990	goto out;
1991
1992	copied = sk_stream_error(sk, flags: msg->msg_flags, err: ret);
1993	goto out;
1994	}
1995
1996	static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied);
1997
1998	static int __mptcp_recvmsg_mskq(struct sock *sk, struct msghdr *msg,
1999	size_t len, int flags, int copied_total,
2000	struct scm_timestamping_internal *tss,
2001	int *cmsg_flags)
2002	{
2003	struct mptcp_sock *msk = mptcp_sk(sk);
2004	struct sk_buff *skb, *tmp;
2005	int total_data_len = 0;
2006	int copied = 0;
2007
2008	skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) {
2009	u32 delta, offset = MPTCP_SKB_CB(skb)->offset;
2010	u32 data_len = skb->len - offset;
2011	u32 count;
2012	int err;
2013
2014	if (flags & MSG_PEEK) {
2015	/* skip already peeked skbs */
2016	if (total_data_len + data_len <= copied_total) {
2017	total_data_len += data_len;
2018	continue;
2019	}
2020
2021	/* skip the already peeked data in the current skb */
2022	delta = copied_total - total_data_len;
2023	offset += delta;
2024	data_len -= delta;
2025	}
2026
2027	count = min_t(size_t, len - copied, data_len);
2028	if (!(flags & MSG_TRUNC)) {
2029	err = skb_copy_datagram_msg(from: skb, offset, msg, size: count);
2030	if (unlikely(err < 0)) {
2031	if (!copied)
2032	return err;
2033	break;
2034	}
2035	}
2036
2037	if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
2038	tcp_update_recv_tstamps(skb, tss);
2039	*cmsg_flags |= MPTCP_CMSG_TS;
2040	}
2041
2042	copied += count;
2043
2044	if (!(flags & MSG_PEEK)) {
2045	msk->bytes_consumed += count;
2046	if (count < data_len) {
2047	MPTCP_SKB_CB(skb)->offset += count;
2048	MPTCP_SKB_CB(skb)->map_seq += count;
2049	break;
2050	}
2051
2052	/* avoid the indirect call, we know the destructor is sock_rfree */
2053	skb->destructor = NULL;
2054	skb->sk = NULL;
2055	atomic_sub(i: skb->truesize, v: &sk->sk_rmem_alloc);
2056	sk_mem_uncharge(sk, size: skb->truesize);
2057	__skb_unlink(skb, list: &sk->sk_receive_queue);
2058	skb_attempt_defer_free(skb);
2059	}
2060
2061	if (copied >= len)
2062	break;
2063	}
2064
2065	mptcp_rcv_space_adjust(msk, copied);
2066	return copied;
2067	}
2068
2069	/* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
2070	*
2071	* Only difference: Use highest rtt estimate of the subflows in use.
2072	*/
2073	static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
2074	{
2075	struct mptcp_subflow_context *subflow;
2076	struct sock *sk = (struct sock *)msk;
2077	u8 scaling_ratio = U8_MAX;
2078	u32 time, advmss = 1;
2079	u64 rtt_us, mstamp;
2080
2081	msk_owned_by_me(msk);
2082
2083	if (copied <= 0)
2084	return;
2085
2086	if (!msk->rcvspace_init)
2087	mptcp_rcv_space_init(msk, ssk: msk->first);
2088
2089	msk->rcvq_space.copied += copied;
2090
2091	mstamp = div_u64(dividend: tcp_clock_ns(), NSEC_PER_USEC);
2092	time = tcp_stamp_us_delta(t1: mstamp, t0: msk->rcvq_space.time);
2093
2094	rtt_us = msk->rcvq_space.rtt_us;
2095	if (rtt_us && time < (rtt_us >> 3))
2096	return;
2097
2098	rtt_us = 0;
2099	mptcp_for_each_subflow(msk, subflow) {
2100	const struct tcp_sock *tp;
2101	u64 sf_rtt_us;
2102	u32 sf_advmss;
2103
2104	tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
2105
2106	sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
2107	sf_advmss = READ_ONCE(tp->advmss);
2108
2109	rtt_us = max(sf_rtt_us, rtt_us);
2110	advmss = max(sf_advmss, advmss);
2111	scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
2112	}
2113
2114	msk->rcvq_space.rtt_us = rtt_us;
2115	msk->scaling_ratio = scaling_ratio;
2116	if (time < (rtt_us >> 3) || rtt_us == 0)
2117	return;
2118
2119	if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2120	goto new_measure;
2121
2122	if (mptcp_rcvbuf_grow(sk, newval: msk->rcvq_space.copied)) {
2123	/* Make subflows follow along. If we do not do this, we
2124	* get drops at subflow level if skbs can't be moved to
2125	* the mptcp rx queue fast enough (announced rcv_win can
2126	* exceed ssk->sk_rcvbuf).
2127	*/
2128	mptcp_for_each_subflow(msk, subflow) {
2129	struct sock *ssk;
2130	bool slow;
2131
2132	ssk = mptcp_subflow_tcp_sock(subflow);
2133	slow = lock_sock_fast(sk: ssk);
2134	/* subflows can be added before tcp_init_transfer() */
2135	if (tcp_sk(ssk)->rcvq_space.space)
2136	tcp_rcvbuf_grow(sk: ssk, newval: msk->rcvq_space.copied);
2137	unlock_sock_fast(sk: ssk, slow);
2138	}
2139	}
2140
2141	new_measure:
2142	msk->rcvq_space.copied = 0;
2143	msk->rcvq_space.time = mstamp;
2144	}
2145
2146	static bool __mptcp_move_skbs(struct sock *sk, struct list_head *skbs, u32 *delta)
2147	{
2148	struct sk_buff *skb = list_first_entry(skbs, struct sk_buff, list);
2149	struct mptcp_sock *msk = mptcp_sk(sk);
2150	bool moved = false;
2151
2152	*delta = 0;
2153	while (1) {
2154	/* If the msk recvbuf is full stop, don't drop */
2155	if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf)
2156	break;
2157
2158	prefetch(skb->next);
2159	list_del(entry: &skb->list);
2160	*delta += skb->truesize;
2161
2162	moved |= __mptcp_move_skb(sk, skb);
2163	if (list_empty(head: skbs))
2164	break;
2165
2166	skb = list_first_entry(skbs, struct sk_buff, list);
2167	}
2168
2169	__mptcp_ofo_queue(msk);
2170	if (moved)
2171	mptcp_check_data_fin(sk: (struct sock *)msk);
2172	return moved;
2173	}
2174
2175	static bool mptcp_can_spool_backlog(struct sock *sk, struct list_head *skbs)
2176	{
2177	struct mptcp_sock *msk = mptcp_sk(sk);
2178
2179	/* After CG initialization, subflows should never add skb before
2180	* gaining the CG themself.
2181	*/
2182	DEBUG_NET_WARN_ON_ONCE(msk->backlog_unaccounted && sk->sk_socket &&
2183	mem_cgroup_from_sk(sk));
2184
2185	/* Don't spool the backlog if the rcvbuf is full. */
2186	if (list_empty(head: &msk->backlog_list) ||
2187	sk_rmem_alloc_get(sk) > sk->sk_rcvbuf)
2188	return false;
2189
2190	INIT_LIST_HEAD(list: skbs);
2191	list_splice_init(list: &msk->backlog_list, head: skbs);
2192	return true;
2193	}
2194
2195	static void mptcp_backlog_spooled(struct sock *sk, u32 moved,
2196	struct list_head *skbs)
2197	{
2198	struct mptcp_sock *msk = mptcp_sk(sk);
2199
2200	WRITE_ONCE(msk->backlog_len, msk->backlog_len - moved);
2201	list_splice(list: skbs, head: &msk->backlog_list);
2202	}
2203
2204	static bool mptcp_move_skbs(struct sock *sk)
2205	{
2206	struct list_head skbs;
2207	bool enqueued = false;
2208	u32 moved;
2209
2210	mptcp_data_lock(sk);
2211	while (mptcp_can_spool_backlog(sk, skbs: &skbs)) {
2212	mptcp_data_unlock(sk);
2213	enqueued |= __mptcp_move_skbs(sk, skbs: &skbs, delta: &moved);
2214
2215	mptcp_data_lock(sk);
2216	mptcp_backlog_spooled(sk, moved, skbs: &skbs);
2217	}
2218	mptcp_data_unlock(sk);
2219	return enqueued;
2220	}
2221
2222	static unsigned int mptcp_inq_hint(const struct sock *sk)
2223	{
2224	const struct mptcp_sock *msk = mptcp_sk(sk);
2225	const struct sk_buff *skb;
2226
2227	skb = skb_peek(list_: &sk->sk_receive_queue);
2228	if (skb) {
2229	u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq;
2230
2231	if (hint_val >= INT_MAX)
2232	return INT_MAX;
2233
2234	return (unsigned int)hint_val;
2235	}
2236
2237	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2238	return 1;
2239
2240	return 0;
2241	}
2242
2243	static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2244	int flags, int *addr_len)
2245	{
2246	struct mptcp_sock *msk = mptcp_sk(sk);
2247	struct scm_timestamping_internal tss;
2248	int copied = 0, cmsg_flags = 0;
2249	int target;
2250	long timeo;
2251
2252	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2253	if (unlikely(flags & MSG_ERRQUEUE))
2254	return inet_recv_error(sk, msg, len, addr_len);
2255
2256	lock_sock(sk);
2257	if (unlikely(sk->sk_state == TCP_LISTEN)) {
2258	copied = -ENOTCONN;
2259	goto out_err;
2260	}
2261
2262	mptcp_rps_record_subflows(msk);
2263
2264	timeo = sock_rcvtimeo(sk, noblock: flags & MSG_DONTWAIT);
2265
2266	len = min_t(size_t, len, INT_MAX);
2267	target = sock_rcvlowat(sk, waitall: flags & MSG_WAITALL, len);
2268
2269	if (unlikely(msk->recvmsg_inq))
2270	cmsg_flags = MPTCP_CMSG_INQ;
2271
2272	while (copied < len) {
2273	int err, bytes_read;
2274
2275	bytes_read = __mptcp_recvmsg_mskq(sk, msg, len: len - copied, flags,
2276	copied_total: copied, tss: &tss, cmsg_flags: &cmsg_flags);
2277	if (unlikely(bytes_read < 0)) {
2278	if (!copied)
2279	copied = bytes_read;
2280	goto out_err;
2281	}
2282
2283	copied += bytes_read;
2284
2285	if (!list_empty(head: &msk->backlog_list) && mptcp_move_skbs(sk))
2286	continue;
2287
2288	/* only the MPTCP socket status is relevant here. The exit
2289	* conditions mirror closely tcp_recvmsg()
2290	*/
2291	if (copied >= target)
2292	break;
2293
2294	if (copied) {
2295	if (sk->sk_err ||
2296	sk->sk_state == TCP_CLOSE ||
2297	(sk->sk_shutdown & RCV_SHUTDOWN) ||
2298	!timeo ||
2299	signal_pending(current))
2300	break;
2301	} else {
2302	if (sk->sk_err) {
2303	copied = sock_error(sk);
2304	break;
2305	}
2306
2307	if (sk->sk_shutdown & RCV_SHUTDOWN)
2308	break;
2309
2310	if (sk->sk_state == TCP_CLOSE) {
2311	copied = -ENOTCONN;
2312	break;
2313	}
2314
2315	if (!timeo) {
2316	copied = -EAGAIN;
2317	break;
2318	}
2319
2320	if (signal_pending(current)) {
2321	copied = sock_intr_errno(timeo);
2322	break;
2323	}
2324	}
2325
2326	pr_debug("block timeout %ld\n", timeo);
2327	mptcp_cleanup_rbuf(msk, copied);
2328	err = sk_wait_data(sk, timeo: &timeo, NULL);
2329	if (err < 0) {
2330	err = copied ? : err;
2331	goto out_err;
2332	}
2333	}
2334
2335	mptcp_cleanup_rbuf(msk, copied);
2336
2337	out_err:
2338	if (cmsg_flags && copied >= 0) {
2339	if (cmsg_flags & MPTCP_CMSG_TS)
2340	tcp_recv_timestamp(msg, sk, tss: &tss);
2341
2342	if (cmsg_flags & MPTCP_CMSG_INQ) {
2343	unsigned int inq = mptcp_inq_hint(sk);
2344
2345	put_cmsg(msg, SOL_TCP, TCP_CM_INQ, len: sizeof(inq), data: &inq);
2346	}
2347	}
2348
2349	pr_debug("msk=%p rx queue empty=%d copied=%d\n",
2350	msk, skb_queue_empty(&sk->sk_receive_queue), copied);
2351
2352	release_sock(sk);
2353	return copied;
2354	}
2355
2356	static void mptcp_retransmit_timer(struct timer_list *t)
2357	{
2358	struct sock *sk = timer_container_of(sk, t, mptcp_retransmit_timer);
2359	struct mptcp_sock *msk = mptcp_sk(sk);
2360
2361	bh_lock_sock(sk);
2362	if (!sock_owned_by_user(sk)) {
2363	/* we need a process context to retransmit */
2364	if (!test_and_set_bit(MPTCP_WORK_RTX, addr: &msk->flags))
2365	mptcp_schedule_work(sk);
2366	} else {
2367	/* delegate our work to tcp_release_cb() */
2368	__set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2369	}
2370	bh_unlock_sock(sk);
2371	sock_put(sk);
2372	}
2373
2374	static void mptcp_tout_timer(struct timer_list *t)
2375	{
2376	struct inet_connection_sock *icsk =
2377	timer_container_of(icsk, t, mptcp_tout_timer);
2378	struct sock *sk = &icsk->icsk_inet.sk;
2379
2380	mptcp_schedule_work(sk);
2381	sock_put(sk);
2382	}
2383
2384	/* Find an idle subflow. Return NULL if there is unacked data at tcp
2385	* level.
2386	*
2387	* A backup subflow is returned only if that is the only kind available.
2388	*/
2389	struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2390	{
2391	struct sock *backup = NULL, *pick = NULL;
2392	struct mptcp_subflow_context *subflow;
2393	int min_stale_count = INT_MAX;
2394
2395	mptcp_for_each_subflow(msk, subflow) {
2396	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2397
2398	if (!__mptcp_subflow_active(subflow))
2399	continue;
2400
2401	/* still data outstanding at TCP level? skip this */
2402	if (!tcp_rtx_and_write_queues_empty(sk: ssk)) {
2403	mptcp_pm_subflow_chk_stale(msk, ssk);
2404	min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2405	continue;
2406	}
2407
2408	if (subflow->backup || subflow->request_bkup) {
2409	if (!backup)
2410	backup = ssk;
2411	continue;
2412	}
2413
2414	if (!pick)
2415	pick = ssk;
2416	}
2417
2418	if (pick)
2419	return pick;
2420
2421	/* use backup only if there are no progresses anywhere */
2422	return min_stale_count > 1 ? backup : NULL;
2423	}
2424
2425	bool __mptcp_retransmit_pending_data(struct sock *sk)
2426	{
2427	struct mptcp_data_frag *cur, *rtx_head;
2428	struct mptcp_sock *msk = mptcp_sk(sk);
2429
2430	if (__mptcp_check_fallback(msk))
2431	return false;
2432
2433	/* the closing socket has some data untransmitted and/or unacked:
2434	* some data in the mptcp rtx queue has not really xmitted yet.
2435	* keep it simple and re-inject the whole mptcp level rtx queue
2436	*/
2437	mptcp_data_lock(sk);
2438	__mptcp_clean_una_wakeup(sk);
2439	rtx_head = mptcp_rtx_head(sk);
2440	if (!rtx_head) {
2441	mptcp_data_unlock(sk);
2442	return false;
2443	}
2444
2445	msk->recovery_snd_nxt = msk->snd_nxt;
2446	msk->recovery = true;
2447	mptcp_data_unlock(sk);
2448
2449	msk->first_pending = rtx_head;
2450	msk->snd_burst = 0;
2451
2452	/* be sure to clear the "sent status" on all re-injected fragments */
2453	list_for_each_entry(cur, &msk->rtx_queue, list) {
2454	if (!cur->already_sent)
2455	break;
2456	cur->already_sent = 0;
2457	}
2458
2459	return true;
2460	}
2461
2462	/* flags for __mptcp_close_ssk() */
2463	#define MPTCP_CF_PUSH BIT(1)
2464
2465	/* be sure to send a reset only if the caller asked for it, also
2466	* clean completely the subflow status when the subflow reaches
2467	* TCP_CLOSE state
2468	*/
2469	static void __mptcp_subflow_disconnect(struct sock *ssk,
2470	struct mptcp_subflow_context *subflow,
2471	bool fastclosing)
2472	{
2473	if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2474	fastclosing) {
2475	/* The MPTCP code never wait on the subflow sockets, TCP-level
2476	* disconnect should never fail
2477	*/
2478	WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2479	mptcp_subflow_ctx_reset(subflow);
2480	} else {
2481	tcp_shutdown(sk: ssk, SEND_SHUTDOWN);
2482	}
2483	}
2484
2485	/* subflow sockets can be either outgoing (connect) or incoming
2486	* (accept).
2487	*
2488	* Outgoing subflows use in-kernel sockets.
2489	* Incoming subflows do not have their own 'struct socket' allocated,
2490	* so we need to use tcp_close() after detaching them from the mptcp
2491	* parent socket.
2492	*/
2493	static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2494	struct mptcp_subflow_context *subflow,
2495	unsigned int flags)
2496	{
2497	struct mptcp_sock *msk = mptcp_sk(sk);
2498	bool dispose_it, need_push = false;
2499	int fwd_remaining;
2500
2501	/* Do not pass RX data to the msk, even if the subflow socket is not
2502	* going to be freed (i.e. even for the first subflow on graceful
2503	* subflow close.
2504	*/
2505	lock_sock_nested(sk: ssk, SINGLE_DEPTH_NESTING);
2506	subflow->closing = 1;
2507
2508	/* Borrow the fwd allocated page left-over; fwd memory for the subflow
2509	* could be negative at this point, but will be reach zero soon - when
2510	* the data allocated using such fragment will be freed.
2511	*/
2512	if (subflow->lent_mem_frag) {
2513	fwd_remaining = PAGE_SIZE - subflow->lent_mem_frag;
2514	sk_forward_alloc_add(sk, val: fwd_remaining);
2515	sk_forward_alloc_add(sk: ssk, val: -fwd_remaining);
2516	subflow->lent_mem_frag = 0;
2517	}
2518
2519	/* If the first subflow moved to a close state before accept, e.g. due
2520	* to an incoming reset or listener shutdown, the subflow socket is
2521	* already deleted by inet_child_forget() and the mptcp socket can't
2522	* survive too.
2523	*/
2524	if (msk->in_accept_queue && msk->first == ssk &&
2525	(sock_flag(sk, flag: SOCK_DEAD) || sock_flag(sk: ssk, flag: SOCK_DEAD))) {
2526	/* ensure later check in mptcp_worker() will dispose the msk */
2527	sock_set_flag(sk, flag: SOCK_DEAD);
2528	mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2529	mptcp_subflow_drop_ctx(ssk);
2530	goto out_release;
2531	}
2532
2533	dispose_it = msk->free_first || ssk != msk->first;
2534	if (dispose_it)
2535	list_del(entry: &subflow->node);
2536
2537	if (subflow->send_fastclose && ssk->sk_state != TCP_CLOSE)
2538	tcp_set_state(sk: ssk, state: TCP_CLOSE);
2539
2540	need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2541	if (!dispose_it) {
2542	__mptcp_subflow_disconnect(ssk, subflow, fastclosing: msk->fastclosing);
2543	release_sock(sk: ssk);
2544
2545	goto out;
2546	}
2547
2548	subflow->disposable = 1;
2549
2550	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2551	* the ssk has been already destroyed, we just need to release the
2552	* reference owned by msk;
2553	*/
2554	if (!inet_csk(ssk)->icsk_ulp_ops) {
2555	WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2556	kfree_rcu(subflow, rcu);
2557	} else {
2558	/* otherwise tcp will dispose of the ssk and subflow ctx */
2559	__tcp_close(sk: ssk, timeout: 0);
2560
2561	/* close acquired an extra ref */
2562	__sock_put(sk: ssk);
2563	}
2564
2565	out_release:
2566	__mptcp_subflow_error_report(sk, ssk);
2567	release_sock(sk: ssk);
2568
2569	sock_put(sk: ssk);
2570
2571	if (ssk == msk->first)
2572	WRITE_ONCE(msk->first, NULL);
2573
2574	out:
2575	__mptcp_sync_sndbuf(sk);
2576	if (need_push)
2577	__mptcp_push_pending(sk, flags: 0);
2578
2579	/* Catch every 'all subflows closed' scenario, including peers silently
2580	* closing them, e.g. due to timeout.
2581	* For established sockets, allow an additional timeout before closing,
2582	* as the protocol can still create more subflows.
2583	*/
2584	if (list_is_singular(head: &msk->conn_list) && msk->first &&
2585	inet_sk_state_load(sk: msk->first) == TCP_CLOSE) {
2586	if (sk->sk_state != TCP_ESTABLISHED ||
2587	msk->in_accept_queue || sock_flag(sk, flag: SOCK_DEAD)) {
2588	mptcp_set_state(sk, state: TCP_CLOSE);
2589	mptcp_close_wake_up(sk);
2590	} else {
2591	mptcp_start_tout_timer(sk);
2592	}
2593	}
2594	}
2595
2596	void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2597	struct mptcp_subflow_context *subflow)
2598	{
2599	struct mptcp_sock *msk = mptcp_sk(sk);
2600	struct sk_buff *skb;
2601
2602	/* The first subflow can already be closed or disconnected */
2603	if (subflow->close_event_done || READ_ONCE(subflow->local_id) < 0)
2604	return;
2605
2606	subflow->close_event_done = true;
2607
2608	if (sk->sk_state == TCP_ESTABLISHED)
2609	mptcp_event(type: MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2610
2611	/* Remove any reference from the backlog to this ssk; backlog skbs consume
2612	* space in the msk receive queue, no need to touch sk->sk_rmem_alloc
2613	*/
2614	list_for_each_entry(skb, &msk->backlog_list, list) {
2615	if (skb->sk != ssk)
2616	continue;
2617
2618	atomic_sub(i: skb->truesize, v: &skb->sk->sk_rmem_alloc);
2619	skb->sk = NULL;
2620	}
2621
2622	/* subflow aborted before reaching the fully_established status
2623	* attempt the creation of the next subflow
2624	*/
2625	mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2626
2627	__mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2628	}
2629
2630	static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2631	{
2632	return 0;
2633	}
2634
2635	static void __mptcp_close_subflow(struct sock *sk)
2636	{
2637	struct mptcp_subflow_context *subflow, *tmp;
2638	struct mptcp_sock *msk = mptcp_sk(sk);
2639
2640	might_sleep();
2641
2642	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2643	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2644	int ssk_state = inet_sk_state_load(sk: ssk);
2645
2646	if (ssk_state != TCP_CLOSE &&
2647	(ssk_state != TCP_CLOSE_WAIT ||
2648	inet_sk_state_load(sk) != TCP_ESTABLISHED ||
2649	__mptcp_check_fallback(msk)))
2650	continue;
2651
2652	/* 'subflow_data_ready' will re-sched once rx queue is empty */
2653	if (!skb_queue_empty_lockless(list: &ssk->sk_receive_queue))
2654	continue;
2655
2656	mptcp_close_ssk(sk, ssk, subflow);
2657	}
2658
2659	}
2660
2661	static bool mptcp_close_tout_expired(const struct sock *sk)
2662	{
2663	if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2664	sk->sk_state == TCP_CLOSE)
2665	return false;
2666
2667	return time_after32(tcp_jiffies32,
2668	inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2669	}
2670
2671	static void mptcp_check_fastclose(struct mptcp_sock *msk)
2672	{
2673	struct mptcp_subflow_context *subflow, *tmp;
2674	struct sock *sk = (struct sock *)msk;
2675
2676	if (likely(!READ_ONCE(msk->rcv_fastclose)))
2677	return;
2678
2679	mptcp_token_destroy(msk);
2680
2681	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2682	struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2683	bool slow;
2684
2685	slow = lock_sock_fast(sk: tcp_sk);
2686	if (tcp_sk->sk_state != TCP_CLOSE) {
2687	mptcp_send_active_reset_reason(sk: tcp_sk);
2688	tcp_set_state(sk: tcp_sk, state: TCP_CLOSE);
2689	}
2690	unlock_sock_fast(sk: tcp_sk, slow);
2691	}
2692
2693	/* Mirror the tcp_reset() error propagation */
2694	switch (sk->sk_state) {
2695	case TCP_SYN_SENT:
2696	WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2697	break;
2698	case TCP_CLOSE_WAIT:
2699	WRITE_ONCE(sk->sk_err, EPIPE);
2700	break;
2701	case TCP_CLOSE:
2702	return;
2703	default:
2704	WRITE_ONCE(sk->sk_err, ECONNRESET);
2705	}
2706
2707	mptcp_set_state(sk, state: TCP_CLOSE);
2708	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2709	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2710	set_bit(MPTCP_WORK_CLOSE_SUBFLOW, addr: &msk->flags);
2711
2712	/* the calling mptcp_worker will properly destroy the socket */
2713	if (sock_flag(sk, flag: SOCK_DEAD))
2714	return;
2715
2716	sk->sk_state_change(sk);
2717	sk_error_report(sk);
2718	}
2719
2720	static void __mptcp_retrans(struct sock *sk)
2721	{
2722	struct mptcp_sendmsg_info info = { .data_lock_held = true, };
2723	struct mptcp_sock *msk = mptcp_sk(sk);
2724	struct mptcp_subflow_context *subflow;
2725	struct mptcp_data_frag *dfrag;
2726	struct sock *ssk;
2727	int ret, err;
2728	u16 len = 0;
2729
2730	mptcp_clean_una_wakeup(sk);
2731
2732	/* first check ssk: need to kick "stale" logic */
2733	err = mptcp_sched_get_retrans(msk);
2734	dfrag = mptcp_rtx_head(sk);
2735	if (!dfrag) {
2736	if (mptcp_data_fin_enabled(msk)) {
2737	struct inet_connection_sock *icsk = inet_csk(sk);
2738
2739	WRITE_ONCE(icsk->icsk_retransmits,
2740	icsk->icsk_retransmits + 1);
2741	mptcp_set_datafin_timeout(sk);
2742	mptcp_send_ack(msk);
2743
2744	goto reset_timer;
2745	}
2746
2747	if (!mptcp_send_head(sk))
2748	goto clear_scheduled;
2749
2750	goto reset_timer;
2751	}
2752
2753	if (err)
2754	goto reset_timer;
2755
2756	mptcp_for_each_subflow(msk, subflow) {
2757	if (READ_ONCE(subflow->scheduled)) {
2758	u16 copied = 0;
2759
2760	mptcp_subflow_set_scheduled(subflow, scheduled: false);
2761
2762	ssk = mptcp_subflow_tcp_sock(subflow);
2763
2764	lock_sock(sk: ssk);
2765
2766	/* limit retransmission to the bytes already sent on some subflows */
2767	info.sent = 0;
2768	info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2769	dfrag->already_sent;
2770
2771	/*
2772	* make the whole retrans decision, xmit, disallow
2773	* fallback atomic, note that we can't retrans even
2774	* when an infinite fallback is in progress, i.e. new
2775	* subflows are disallowed.
2776	*/
2777	spin_lock_bh(lock: &msk->fallback_lock);
2778	if (__mptcp_check_fallback(msk) ||
2779	!msk->allow_subflows) {
2780	spin_unlock_bh(lock: &msk->fallback_lock);
2781	release_sock(sk: ssk);
2782	goto clear_scheduled;
2783	}
2784
2785	while (info.sent < info.limit) {
2786	ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info: &info);
2787	if (ret <= 0)
2788	break;
2789
2790	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_RETRANSSEGS);
2791	copied += ret;
2792	info.sent += ret;
2793	}
2794	if (copied) {
2795	len = max(copied, len);
2796	tcp_push(sk: ssk, flags: 0, mss_now: info.mss_now, tcp_sk(ssk)->nonagle,
2797	size_goal: info.size_goal);
2798	msk->allow_infinite_fallback = false;
2799	}
2800	spin_unlock_bh(lock: &msk->fallback_lock);
2801
2802	release_sock(sk: ssk);
2803	}
2804	}
2805
2806	msk->bytes_retrans += len;
2807	dfrag->already_sent = max(dfrag->already_sent, len);
2808
2809	reset_timer:
2810	mptcp_check_and_set_pending(sk);
2811
2812	if (!mptcp_rtx_timer_pending(sk))
2813	mptcp_reset_rtx_timer(sk);
2814
2815	clear_scheduled:
2816	/* If no rtx data was available or in case of fallback, there
2817	* could be left-over scheduled subflows; clear them all
2818	* or later xmit could use bad ones
2819	*/
2820	mptcp_for_each_subflow(msk, subflow)
2821	if (READ_ONCE(subflow->scheduled))
2822	mptcp_subflow_set_scheduled(subflow, scheduled: false);
2823	}
2824
2825	/* schedule the timeout timer for the relevant event: either close timeout
2826	* or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2827	*/
2828	void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2829	{
2830	struct sock *sk = (struct sock *)msk;
2831	unsigned long timeout, close_timeout;
2832
2833	if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2834	return;
2835
2836	close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp -
2837	tcp_jiffies32 + jiffies + mptcp_close_timeout(sk);
2838
2839	/* the close timeout takes precedence on the fail one, and here at least one of
2840	* them is active
2841	*/
2842	timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2843
2844	sk_reset_timer(sk, timer: &inet_csk(sk)->mptcp_tout_timer, expires: timeout);
2845	}
2846
2847	static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2848	{
2849	struct sock *ssk = msk->first;
2850	bool slow;
2851
2852	if (!ssk)
2853	return;
2854
2855	pr_debug("MP_FAIL doesn't respond, reset the subflow\n");
2856
2857	slow = lock_sock_fast(sk: ssk);
2858	mptcp_subflow_reset(ssk);
2859	WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2860	unlock_sock_fast(sk: ssk, slow);
2861	}
2862
2863	static void mptcp_backlog_purge(struct sock *sk)
2864	{
2865	struct mptcp_sock *msk = mptcp_sk(sk);
2866	struct sk_buff *tmp, *skb;
2867	LIST_HEAD(backlog);
2868
2869	mptcp_data_lock(sk);
2870	list_splice_init(list: &msk->backlog_list, head: &backlog);
2871	msk->backlog_len = 0;
2872	mptcp_data_unlock(sk);
2873
2874	list_for_each_entry_safe(skb, tmp, &backlog, list) {
2875	mptcp_borrow_fwdmem(sk, skb);
2876	kfree_skb_reason(skb, reason: SKB_DROP_REASON_SOCKET_CLOSE);
2877	}
2878	sk_mem_reclaim(sk);
2879	}
2880
2881	static void mptcp_do_fastclose(struct sock *sk)
2882	{
2883	struct mptcp_subflow_context *subflow, *tmp;
2884	struct mptcp_sock *msk = mptcp_sk(sk);
2885
2886	mptcp_set_state(sk, state: TCP_CLOSE);
2887	mptcp_backlog_purge(sk);
2888	msk->fastclosing = 1;
2889
2890	/* Explicitly send the fastclose reset as need */
2891	if (__mptcp_check_fallback(msk))
2892	return;
2893
2894	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2895	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2896
2897	lock_sock(sk: ssk);
2898
2899	/* Some subflow socket states don't allow/need a reset.*/
2900	if ((1 << ssk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
2901	goto unlock;
2902
2903	subflow->send_fastclose = 1;
2904
2905	/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2906	* issue in __tcp_select_window(), see tcp_disconnect().
2907	*/
2908	inet_csk(ssk)->icsk_ack.rcv_mss = TCP_MIN_MSS;
2909
2910	tcp_send_active_reset(sk: ssk, priority: ssk->sk_allocation,
2911	reason: SK_RST_REASON_TCP_ABORT_ON_CLOSE);
2912	unlock:
2913	release_sock(sk: ssk);
2914	}
2915	}
2916
2917	static void mptcp_worker(struct work_struct *work)
2918	{
2919	struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2920	struct sock *sk = (struct sock *)msk;
2921	unsigned long fail_tout;
2922	int state;
2923
2924	lock_sock(sk);
2925	state = sk->sk_state;
2926	if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2927	goto unlock;
2928
2929	mptcp_check_fastclose(msk);
2930
2931	mptcp_pm_worker(msk);
2932
2933	mptcp_check_send_data_fin(sk);
2934	mptcp_check_data_fin_ack(sk);
2935	mptcp_check_data_fin(sk);
2936
2937	if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, addr: &msk->flags))
2938	__mptcp_close_subflow(sk);
2939
2940	if (mptcp_close_tout_expired(sk)) {
2941	struct mptcp_subflow_context *subflow, *tmp;
2942
2943	mptcp_do_fastclose(sk);
2944	mptcp_for_each_subflow_safe(msk, subflow, tmp)
2945	__mptcp_close_ssk(sk, ssk: subflow->tcp_sock, subflow, flags: 0);
2946	mptcp_close_wake_up(sk);
2947	}
2948
2949	if (sock_flag(sk, flag: SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2950	__mptcp_destroy_sock(sk);
2951	goto unlock;
2952	}
2953
2954	if (test_and_clear_bit(MPTCP_WORK_RTX, addr: &msk->flags))
2955	__mptcp_retrans(sk);
2956
2957	fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2958	if (fail_tout && time_after(jiffies, fail_tout))
2959	mptcp_mp_fail_no_response(msk);
2960
2961	unlock:
2962	release_sock(sk);
2963	sock_put(sk);
2964	}
2965
2966	static void __mptcp_init_sock(struct sock *sk)
2967	{
2968	struct mptcp_sock *msk = mptcp_sk(sk);
2969
2970	INIT_LIST_HEAD(list: &msk->conn_list);
2971	INIT_LIST_HEAD(list: &msk->join_list);
2972	INIT_LIST_HEAD(list: &msk->rtx_queue);
2973	INIT_LIST_HEAD(list: &msk->backlog_list);
2974	INIT_WORK(&msk->work, mptcp_worker);
2975	msk->out_of_order_queue = RB_ROOT;
2976	msk->first_pending = NULL;
2977	msk->timer_ival = TCP_RTO_MIN;
2978	msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2979	msk->backlog_len = 0;
2980
2981	WRITE_ONCE(msk->first, NULL);
2982	inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2983	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2984	msk->allow_infinite_fallback = true;
2985	msk->allow_subflows = true;
2986	msk->recovery = false;
2987	msk->subflow_id = 1;
2988	msk->last_data_sent = tcp_jiffies32;
2989	msk->last_data_recv = tcp_jiffies32;
2990	msk->last_ack_recv = tcp_jiffies32;
2991
2992	mptcp_pm_data_init(msk);
2993	spin_lock_init(&msk->fallback_lock);
2994
2995	/* re-use the csk retrans timer for MPTCP-level retrans */
2996	timer_setup(&sk->mptcp_retransmit_timer, mptcp_retransmit_timer, 0);
2997	timer_setup(&msk->sk.mptcp_tout_timer, mptcp_tout_timer, 0);
2998	}
2999
3000	static void mptcp_ca_reset(struct sock *sk)
3001	{
3002	struct inet_connection_sock *icsk = inet_csk(sk);
3003
3004	tcp_assign_congestion_control(sk);
3005	strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name,
3006	sizeof(mptcp_sk(sk)->ca_name));
3007
3008	/* no need to keep a reference to the ops, the name will suffice */
3009	tcp_cleanup_congestion_control(sk);
3010	icsk->icsk_ca_ops = NULL;
3011	}
3012
3013	static int mptcp_init_sock(struct sock *sk)
3014	{
3015	struct net *net = sock_net(sk);
3016	int ret;
3017
3018	__mptcp_init_sock(sk);
3019
3020	if (!mptcp_is_enabled(net))
3021	return -ENOPROTOOPT;
3022
3023	if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
3024	return -ENOMEM;
3025
3026	rcu_read_lock();
3027	ret = mptcp_init_sched(mptcp_sk(sk),
3028	sched: mptcp_sched_find(name: mptcp_get_scheduler(net)));
3029	rcu_read_unlock();
3030	if (ret)
3031	return ret;
3032
3033	set_bit(nr: SOCK_CUSTOM_SOCKOPT, addr: &sk->sk_socket->flags);
3034
3035	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
3036	* propagate the correct value
3037	*/
3038	mptcp_ca_reset(sk);
3039
3040	sk_sockets_allocated_inc(sk);
3041	sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
3042	sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
3043
3044	return 0;
3045	}
3046
3047	static void __mptcp_clear_xmit(struct sock *sk)
3048	{
3049	struct mptcp_sock *msk = mptcp_sk(sk);
3050	struct mptcp_data_frag *dtmp, *dfrag;
3051
3052	msk->first_pending = NULL;
3053	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
3054	dfrag_clear(sk, dfrag);
3055	}
3056
3057	void mptcp_cancel_work(struct sock *sk)
3058	{
3059	struct mptcp_sock *msk = mptcp_sk(sk);
3060
3061	if (cancel_work_sync(work: &msk->work))
3062	__sock_put(sk);
3063	}
3064
3065	void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
3066	{
3067	lock_sock(sk: ssk);
3068
3069	switch (ssk->sk_state) {
3070	case TCP_LISTEN:
3071	if (!(how & RCV_SHUTDOWN))
3072	break;
3073	fallthrough;
3074	case TCP_SYN_SENT:
3075	WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
3076	break;
3077	default:
3078	if (__mptcp_check_fallback(mptcp_sk(sk))) {
3079	pr_debug("Fallback\n");
3080	ssk->sk_shutdown |= how;
3081	tcp_shutdown(sk: ssk, how);
3082
3083	/* simulate the data_fin ack reception to let the state
3084	* machine move forward
3085	*/
3086	WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
3087	mptcp_schedule_work(sk);
3088	} else {
3089	pr_debug("Sending DATA_FIN on subflow %p\n", ssk);
3090	tcp_send_ack(sk: ssk);
3091	if (!mptcp_rtx_timer_pending(sk))
3092	mptcp_reset_rtx_timer(sk);
3093	}
3094	break;
3095	}
3096
3097	release_sock(sk: ssk);
3098	}
3099
3100	void mptcp_set_state(struct sock *sk, int state)
3101	{
3102	int oldstate = sk->sk_state;
3103
3104	switch (state) {
3105	case TCP_ESTABLISHED:
3106	if (oldstate != TCP_ESTABLISHED)
3107	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_CURRESTAB);
3108	break;
3109	case TCP_CLOSE_WAIT:
3110	/* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state:
3111	* MPTCP "accepted" sockets will be created later on. So no
3112	* transition from TCP_SYN_RECV to TCP_CLOSE_WAIT.
3113	*/
3114	break;
3115	default:
3116	if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
3117	MPTCP_DEC_STATS(net: sock_net(sk), field: MPTCP_MIB_CURRESTAB);
3118	}
3119
3120	inet_sk_state_store(sk, newstate: state);
3121	}
3122
3123	static const unsigned char new_state[16] = {
3124	/* current state: new state: action: */
3125	[0 /* (Invalid) */] = TCP_CLOSE,
3126	[TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
3127	[TCP_SYN_SENT] = TCP_CLOSE,
3128	[TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
3129	[TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
3130	[TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
3131	[TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
3132	[TCP_CLOSE] = TCP_CLOSE,
3133	[TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
3134	[TCP_LAST_ACK] = TCP_LAST_ACK,
3135	[TCP_LISTEN] = TCP_CLOSE,
3136	[TCP_CLOSING] = TCP_CLOSING,
3137	[TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
3138	};
3139
3140	static int mptcp_close_state(struct sock *sk)
3141	{
3142	int next = (int)new_state[sk->sk_state];
3143	int ns = next & TCP_STATE_MASK;
3144
3145	mptcp_set_state(sk, state: ns);
3146
3147	return next & TCP_ACTION_FIN;
3148	}
3149
3150	static void mptcp_check_send_data_fin(struct sock *sk)
3151	{
3152	struct mptcp_subflow_context *subflow;
3153	struct mptcp_sock *msk = mptcp_sk(sk);
3154
3155	pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n",
3156	msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
3157	msk->snd_nxt, msk->write_seq);
3158
3159	/* we still need to enqueue subflows or not really shutting down,
3160	* skip this
3161	*/
3162	if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
3163	mptcp_send_head(sk))
3164	return;
3165
3166	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3167
3168	mptcp_for_each_subflow(msk, subflow) {
3169	struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
3170
3171	mptcp_subflow_shutdown(sk, ssk: tcp_sk, SEND_SHUTDOWN);
3172	}
3173	}
3174
3175	static void __mptcp_wr_shutdown(struct sock *sk)
3176	{
3177	struct mptcp_sock *msk = mptcp_sk(sk);
3178
3179	pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n",
3180	msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
3181	!!mptcp_send_head(sk));
3182
3183	/* will be ignored by fallback sockets */
3184	WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
3185	WRITE_ONCE(msk->snd_data_fin_enable, 1);
3186
3187	mptcp_check_send_data_fin(sk);
3188	}
3189
3190	static void __mptcp_destroy_sock(struct sock *sk)
3191	{
3192	struct mptcp_sock *msk = mptcp_sk(sk);
3193
3194	pr_debug("msk=%p\n", msk);
3195
3196	might_sleep();
3197
3198	mptcp_stop_rtx_timer(sk);
3199	sk_stop_timer(sk, timer: &inet_csk(sk)->mptcp_tout_timer);
3200	msk->pm.status = 0;
3201	mptcp_release_sched(msk);
3202
3203	sk->sk_prot->destroy(sk);
3204
3205	sk_stream_kill_queues(sk);
3206	xfrm_sk_free_policy(sk);
3207
3208	sock_put(sk);
3209	}
3210
3211	void __mptcp_unaccepted_force_close(struct sock *sk)
3212	{
3213	sock_set_flag(sk, flag: SOCK_DEAD);
3214	mptcp_do_fastclose(sk);
3215	__mptcp_destroy_sock(sk);
3216	}
3217
3218	static __poll_t mptcp_check_readable(struct sock *sk)
3219	{
3220	return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
3221	}
3222
3223	static void mptcp_check_listen_stop(struct sock *sk)
3224	{
3225	struct sock *ssk;
3226
3227	if (inet_sk_state_load(sk) != TCP_LISTEN)
3228	return;
3229
3230	sock_prot_inuse_add(net: sock_net(sk), prot: sk->sk_prot, val: -1);
3231	ssk = mptcp_sk(sk)->first;
3232	if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3233	return;
3234
3235	lock_sock_nested(sk: ssk, SINGLE_DEPTH_NESTING);
3236	tcp_set_state(sk: ssk, state: TCP_CLOSE);
3237	mptcp_subflow_queue_clean(sk, ssk);
3238	inet_csk_listen_stop(sk: ssk);
3239	mptcp_event_pm_listener(ssk, event: MPTCP_EVENT_LISTENER_CLOSED);
3240	release_sock(sk: ssk);
3241	}
3242
3243	bool __mptcp_close(struct sock *sk, long timeout)
3244	{
3245	struct mptcp_subflow_context *subflow;
3246	struct mptcp_sock *msk = mptcp_sk(sk);
3247	bool do_cancel_work = false;
3248	int subflows_alive = 0;
3249
3250	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3251
3252	if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3253	mptcp_check_listen_stop(sk);
3254	mptcp_set_state(sk, state: TCP_CLOSE);
3255	goto cleanup;
3256	}
3257
3258	if (mptcp_data_avail(msk) || timeout < 0) {
3259	/* If the msk has read data, or the caller explicitly ask it,
3260	* do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3261	*/
3262	mptcp_do_fastclose(sk);
3263	timeout = 0;
3264	} else if (mptcp_close_state(sk)) {
3265	__mptcp_wr_shutdown(sk);
3266	}
3267
3268	sk_stream_wait_close(sk, timeo_p: timeout);
3269
3270	cleanup:
3271	/* orphan all the subflows */
3272	mptcp_for_each_subflow(msk, subflow) {
3273	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3274	bool slow = lock_sock_fast_nested(sk: ssk);
3275
3276	subflows_alive += ssk->sk_state != TCP_CLOSE;
3277
3278	/* since the close timeout takes precedence on the fail one,
3279	* cancel the latter
3280	*/
3281	if (ssk == msk->first)
3282	subflow->fail_tout = 0;
3283
3284	/* detach from the parent socket, but allow data_ready to
3285	* push incoming data into the mptcp stack, to properly ack it
3286	*/
3287	ssk->sk_socket = NULL;
3288	ssk->sk_wq = NULL;
3289	unlock_sock_fast(sk: ssk, slow);
3290	}
3291	sock_orphan(sk);
3292
3293	/* all the subflows are closed, only timeout can change the msk
3294	* state, let's not keep resources busy for no reasons
3295	*/
3296	if (subflows_alive == 0)
3297	mptcp_set_state(sk, state: TCP_CLOSE);
3298
3299	sock_hold(sk);
3300	pr_debug("msk=%p state=%d\n", sk, sk->sk_state);
3301	mptcp_pm_connection_closed(msk);
3302
3303	if (sk->sk_state == TCP_CLOSE) {
3304	__mptcp_destroy_sock(sk);
3305	do_cancel_work = true;
3306	} else {
3307	mptcp_start_tout_timer(sk);
3308	}
3309
3310	return do_cancel_work;
3311	}
3312
3313	static void mptcp_close(struct sock *sk, long timeout)
3314	{
3315	bool do_cancel_work;
3316
3317	lock_sock(sk);
3318
3319	do_cancel_work = __mptcp_close(sk, timeout);
3320	release_sock(sk);
3321	if (do_cancel_work)
3322	mptcp_cancel_work(sk);
3323
3324	sock_put(sk);
3325	}
3326
3327	static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3328	{
3329	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3330	const struct ipv6_pinfo *ssk6 = inet6_sk(sk: ssk);
3331	struct ipv6_pinfo *msk6 = inet6_sk(sk: msk);
3332
3333	msk->sk_v6_daddr = ssk->sk_v6_daddr;
3334	msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3335
3336	if (msk6 && ssk6) {
3337	msk6->saddr = ssk6->saddr;
3338	msk6->flow_label = ssk6->flow_label;
3339	}
3340	#endif
3341
3342	inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3343	inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3344	inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3345	inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3346	inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3347	inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3348	}
3349
3350	static void mptcp_destroy_common(struct mptcp_sock *msk)
3351	{
3352	struct mptcp_subflow_context *subflow, *tmp;
3353	struct sock *sk = (struct sock *)msk;
3354
3355	__mptcp_clear_xmit(sk);
3356	mptcp_backlog_purge(sk);
3357
3358	/* join list will be eventually flushed (with rst) at sock lock release time */
3359	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3360	__mptcp_close_ssk(sk, ssk: mptcp_subflow_tcp_sock(subflow), subflow, flags: 0);
3361
3362	__skb_queue_purge(list: &sk->sk_receive_queue);
3363	skb_rbtree_purge(root: &msk->out_of_order_queue);
3364
3365	/* move all the rx fwd alloc into the sk_mem_reclaim_final in
3366	* inet_sock_destruct() will dispose it
3367	*/
3368	mptcp_token_destroy(msk);
3369	mptcp_pm_destroy(msk);
3370	}
3371
3372	static int mptcp_disconnect(struct sock *sk, int flags)
3373	{
3374	struct mptcp_sock *msk = mptcp_sk(sk);
3375
3376	/* We are on the fastopen error path. We can't call straight into the
3377	* subflows cleanup code due to lock nesting (we are already under
3378	* msk->firstsocket lock).
3379	*/
3380	if (msk->fastopening)
3381	return -EBUSY;
3382
3383	mptcp_check_listen_stop(sk);
3384	mptcp_set_state(sk, state: TCP_CLOSE);
3385
3386	mptcp_stop_rtx_timer(sk);
3387	mptcp_stop_tout_timer(sk);
3388
3389	mptcp_pm_connection_closed(msk);
3390
3391	/* msk->subflow is still intact, the following will not free the first
3392	* subflow
3393	*/
3394	mptcp_do_fastclose(sk);
3395	mptcp_destroy_common(msk);
3396
3397	/* The first subflow is already in TCP_CLOSE status, the following
3398	* can't overlap with a fallback anymore
3399	*/
3400	spin_lock_bh(lock: &msk->fallback_lock);
3401	msk->allow_subflows = true;
3402	msk->allow_infinite_fallback = true;
3403	WRITE_ONCE(msk->flags, 0);
3404	spin_unlock_bh(lock: &msk->fallback_lock);
3405
3406	msk->cb_flags = 0;
3407	msk->recovery = false;
3408	WRITE_ONCE(msk->can_ack, false);
3409	WRITE_ONCE(msk->fully_established, false);
3410	WRITE_ONCE(msk->rcv_data_fin, false);
3411	WRITE_ONCE(msk->snd_data_fin_enable, false);
3412	WRITE_ONCE(msk->rcv_fastclose, false);
3413	WRITE_ONCE(msk->use_64bit_ack, false);
3414	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3415	mptcp_pm_data_reset(msk);
3416	mptcp_ca_reset(sk);
3417	msk->bytes_consumed = 0;
3418	msk->bytes_acked = 0;
3419	msk->bytes_received = 0;
3420	msk->bytes_sent = 0;
3421	msk->bytes_retrans = 0;
3422	msk->rcvspace_init = 0;
3423	msk->fastclosing = 0;
3424
3425	/* for fallback's sake */
3426	WRITE_ONCE(msk->ack_seq, 0);
3427
3428	WRITE_ONCE(sk->sk_shutdown, 0);
3429	sk_error_report(sk);
3430	return 0;
3431	}
3432
3433	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3434	static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3435	{
3436	struct mptcp6_sock *msk6 = container_of(mptcp_sk(sk), struct mptcp6_sock, msk);
3437
3438	return &msk6->np;
3439	}
3440
3441	static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3442	{
3443	const struct ipv6_pinfo *np = inet6_sk(sk: sk);
3444	struct ipv6_txoptions *opt;
3445	struct ipv6_pinfo *newnp;
3446
3447	newnp = inet6_sk(sk: newsk);
3448
3449	rcu_read_lock();
3450	opt = rcu_dereference(np->opt);
3451	if (opt) {
3452	opt = ipv6_dup_options(sk: newsk, opt);
3453	if (!opt)
3454	net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3455	}
3456	RCU_INIT_POINTER(newnp->opt, opt);
3457	rcu_read_unlock();
3458	}
3459	#endif
3460
3461	static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3462	{
3463	struct ip_options_rcu *inet_opt, *newopt = NULL;
3464	const struct inet_sock *inet = inet_sk(sk);
3465	struct inet_sock *newinet;
3466
3467	newinet = inet_sk(newsk);
3468
3469	rcu_read_lock();
3470	inet_opt = rcu_dereference(inet->inet_opt);
3471	if (inet_opt) {
3472	newopt = sock_kmemdup(sk: newsk, src: inet_opt, size: sizeof(*inet_opt) +
3473	inet_opt->opt.optlen, GFP_ATOMIC);
3474	if (!newopt)
3475	net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3476	}
3477	RCU_INIT_POINTER(newinet->inet_opt, newopt);
3478	rcu_read_unlock();
3479	}
3480
3481	struct sock *mptcp_sk_clone_init(const struct sock *sk,
3482	const struct mptcp_options_received *mp_opt,
3483	struct sock *ssk,
3484	struct request_sock *req)
3485	{
3486	struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(rsk: req);
3487	struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3488	struct mptcp_subflow_context *subflow;
3489	struct mptcp_sock *msk;
3490
3491	if (!nsk)
3492	return NULL;
3493
3494	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3495	if (nsk->sk_family == AF_INET6)
3496	inet_sk(nsk)->pinet6 = mptcp_inet6_sk(sk: nsk);
3497	#endif
3498
3499	__mptcp_init_sock(sk: nsk);
3500
3501	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3502	if (nsk->sk_family == AF_INET6)
3503	mptcp_copy_ip6_options(newsk: nsk, sk);
3504	else
3505	#endif
3506	mptcp_copy_ip_options(newsk: nsk, sk);
3507
3508	msk = mptcp_sk(nsk);
3509	WRITE_ONCE(msk->local_key, subflow_req->local_key);
3510	WRITE_ONCE(msk->token, subflow_req->token);
3511	msk->in_accept_queue = 1;
3512	WRITE_ONCE(msk->fully_established, false);
3513	if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3514	WRITE_ONCE(msk->csum_enabled, true);
3515
3516	WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1);
3517	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3518	WRITE_ONCE(msk->snd_una, msk->write_seq);
3519	WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3520	msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3521	mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3522
3523	/* passive msk is created after the first/MPC subflow */
3524	msk->subflow_id = 2;
3525
3526	sock_reset_flag(sk: nsk, flag: SOCK_RCU_FREE);
3527	security_inet_csk_clone(newsk: nsk, req);
3528
3529	/* this can't race with mptcp_close(), as the msk is
3530	* not yet exposted to user-space
3531	*/
3532	mptcp_set_state(sk: nsk, state: TCP_ESTABLISHED);
3533
3534	/* The msk maintain a ref to each subflow in the connections list */
3535	WRITE_ONCE(msk->first, ssk);
3536	subflow = mptcp_subflow_ctx(sk: ssk);
3537	list_add(new: &subflow->node, head: &msk->conn_list);
3538	sock_hold(sk: ssk);
3539
3540	/* new mpc subflow takes ownership of the newly
3541	* created mptcp socket
3542	*/
3543	mptcp_token_accept(r: subflow_req, msk);
3544
3545	/* set msk addresses early to ensure mptcp_pm_get_local_id()
3546	* uses the correct data
3547	*/
3548	mptcp_copy_inaddrs(msk: nsk, ssk);
3549	__mptcp_propagate_sndbuf(sk: nsk, ssk);
3550
3551	mptcp_rcv_space_init(msk, ssk);
3552
3553	if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3554	__mptcp_subflow_fully_established(msk, subflow, mp_opt);
3555	bh_unlock_sock(nsk);
3556
3557	/* note: the newly allocated socket refcount is 2 now */
3558	return nsk;
3559	}
3560
3561	void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3562	{
3563	const struct tcp_sock *tp = tcp_sk(ssk);
3564
3565	msk->rcvspace_init = 1;
3566	msk->rcvq_space.copied = 0;
3567	msk->rcvq_space.rtt_us = 0;
3568
3569	msk->rcvq_space.time = tp->tcp_mstamp;
3570
3571	/* initial rcv_space offering made to peer */
3572	msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3573	TCP_INIT_CWND * tp->advmss);
3574	if (msk->rcvq_space.space == 0)
3575	msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3576	}
3577
3578	static void mptcp_destroy(struct sock *sk)
3579	{
3580	struct mptcp_sock *msk = mptcp_sk(sk);
3581
3582	/* allow the following to close even the initial subflow */
3583	msk->free_first = 1;
3584	mptcp_destroy_common(msk);
3585	sk_sockets_allocated_dec(sk);
3586	}
3587
3588	void __mptcp_data_acked(struct sock *sk)
3589	{
3590	if (!sock_owned_by_user(sk))
3591	__mptcp_clean_una(sk);
3592	else
3593	__set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3594	}
3595
3596	void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3597	{
3598	if (!sock_owned_by_user(sk))
3599	__mptcp_subflow_push_pending(sk, ssk, first: false);
3600	else
3601	__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3602	}
3603
3604	#define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3605	BIT(MPTCP_RETRANSMIT) | \
3606	BIT(MPTCP_FLUSH_JOIN_LIST))
3607
3608	/* processes deferred events and flush wmem */
3609	static void mptcp_release_cb(struct sock *sk)
3610	__must_hold(&sk->sk_lock.slock)
3611	{
3612	struct mptcp_sock *msk = mptcp_sk(sk);
3613
3614	for (;;) {
3615	unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3616	struct list_head join_list, skbs;
3617	bool spool_bl;
3618	u32 moved;
3619
3620	spool_bl = mptcp_can_spool_backlog(sk, skbs: &skbs);
3621	if (!flags && !spool_bl)
3622	break;
3623
3624	INIT_LIST_HEAD(list: &join_list);
3625	list_splice_init(list: &msk->join_list, head: &join_list);
3626
3627	/* the following actions acquire the subflow socket lock
3628	*
3629	* 1) can't be invoked in atomic scope
3630	* 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3631	* datapath acquires the msk socket spinlock while helding
3632	* the subflow socket lock
3633	*/
3634	msk->cb_flags &= ~flags;
3635	spin_unlock_bh(lock: &sk->sk_lock.slock);
3636
3637	if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3638	__mptcp_flush_join_list(sk, join_list: &join_list);
3639	if (flags & BIT(MPTCP_PUSH_PENDING))
3640	__mptcp_push_pending(sk, flags: 0);
3641	if (flags & BIT(MPTCP_RETRANSMIT))
3642	__mptcp_retrans(sk);
3643	if (spool_bl && __mptcp_move_skbs(sk, skbs: &skbs, delta: &moved)) {
3644	/* notify ack seq update */
3645	mptcp_cleanup_rbuf(msk, copied: 0);
3646	sk->sk_data_ready(sk);
3647	}
3648
3649	cond_resched();
3650	spin_lock_bh(lock: &sk->sk_lock.slock);
3651	if (spool_bl)
3652	mptcp_backlog_spooled(sk, moved, skbs: &skbs);
3653	}
3654
3655	if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3656	__mptcp_clean_una_wakeup(sk);
3657	if (unlikely(msk->cb_flags)) {
3658	/* be sure to sync the msk state before taking actions
3659	* depending on sk_state (MPTCP_ERROR_REPORT)
3660	* On sk release avoid actions depending on the first subflow
3661	*/
3662	if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3663	__mptcp_sync_state(sk, state: msk->pending_state);
3664	if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3665	__mptcp_error_report(sk);
3666	if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3667	__mptcp_sync_sndbuf(sk);
3668	}
3669	}
3670
3671	/* MP_JOIN client subflow must wait for 4th ack before sending any data:
3672	* TCP can't schedule delack timer before the subflow is fully established.
3673	* MPTCP uses the delack timer to do 3rd ack retransmissions
3674	*/
3675	static void schedule_3rdack_retransmission(struct sock *ssk)
3676	{
3677	struct inet_connection_sock *icsk = inet_csk(ssk);
3678	struct tcp_sock *tp = tcp_sk(ssk);
3679	unsigned long timeout;
3680
3681	if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established))
3682	return;
3683
3684	/* reschedule with a timeout above RTT, as we must look only for drop */
3685	if (tp->srtt_us)
3686	timeout = usecs_to_jiffies(u: tp->srtt_us >> (3 - 1));
3687	else
3688	timeout = TCP_TIMEOUT_INIT;
3689	timeout += jiffies;
3690
3691	WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3692	smp_store_release(&icsk->icsk_ack.pending,
3693	icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
3694	sk_reset_timer(sk: ssk, timer: &icsk->icsk_delack_timer, expires: timeout);
3695	}
3696
3697	void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3698	{
3699	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk: ssk);
3700	struct sock *sk = subflow->conn;
3701
3702	if (status & BIT(MPTCP_DELEGATE_SEND)) {
3703	mptcp_data_lock(sk);
3704	if (!sock_owned_by_user(sk))
3705	__mptcp_subflow_push_pending(sk, ssk, first: true);
3706	else
3707	__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3708	mptcp_data_unlock(sk);
3709	}
3710	if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3711	mptcp_data_lock(sk);
3712	if (!sock_owned_by_user(sk))
3713	__mptcp_sync_sndbuf(sk);
3714	else
3715	__set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3716	mptcp_data_unlock(sk);
3717	}
3718	if (status & BIT(MPTCP_DELEGATE_ACK))
3719	schedule_3rdack_retransmission(ssk);
3720	}
3721
3722	static int mptcp_hash(struct sock *sk)
3723	{
3724	/* should never be called,
3725	* we hash the TCP subflows not the MPTCP socket
3726	*/
3727	WARN_ON_ONCE(1);
3728	return 0;
3729	}
3730
3731	static void mptcp_unhash(struct sock *sk)
3732	{
3733	/* called from sk_common_release(), but nothing to do here */
3734	}
3735
3736	static int mptcp_get_port(struct sock *sk, unsigned short snum)
3737	{
3738	struct mptcp_sock *msk = mptcp_sk(sk);
3739
3740	pr_debug("msk=%p, ssk=%p\n", msk, msk->first);
3741	if (WARN_ON_ONCE(!msk->first))
3742	return -EINVAL;
3743
3744	return inet_csk_get_port(sk: msk->first, snum);
3745	}
3746
3747	void mptcp_finish_connect(struct sock *ssk)
3748	{
3749	struct mptcp_subflow_context *subflow;
3750	struct mptcp_sock *msk;
3751	struct sock *sk;
3752
3753	subflow = mptcp_subflow_ctx(sk: ssk);
3754	sk = subflow->conn;
3755	msk = mptcp_sk(sk);
3756
3757	pr_debug("msk=%p, token=%u\n", sk, subflow->token);
3758
3759	subflow->map_seq = subflow->iasn;
3760	subflow->map_subflow_seq = 1;
3761
3762	/* the socket is not connected yet, no msk/subflow ops can access/race
3763	* accessing the field below
3764	*/
3765	WRITE_ONCE(msk->local_key, subflow->local_key);
3766
3767	mptcp_pm_new_connection(msk, ssk, server_side: 0);
3768	}
3769
3770	void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3771	{
3772	write_lock_bh(&sk->sk_callback_lock);
3773	rcu_assign_pointer(sk->sk_wq, &parent->wq);
3774	sk_set_socket(sk, sock: parent);
3775	write_unlock_bh(&sk->sk_callback_lock);
3776	}
3777
3778	/* Can be called without holding the msk socket lock; use the callback lock
3779	* to avoid {READ_,WRITE_}ONCE annotations on sk_socket.
3780	*/
3781	static void mptcp_sock_check_graft(struct sock *sk, struct sock *ssk)
3782	{
3783	struct socket *sock;
3784
3785	write_lock_bh(&sk->sk_callback_lock);
3786	sock = sk->sk_socket;
3787	write_unlock_bh(&sk->sk_callback_lock);
3788	if (sock) {
3789	mptcp_sock_graft(sk: ssk, parent: sock);
3790	__mptcp_inherit_cgrp_data(sk, ssk);
3791	__mptcp_inherit_memcg(sk, ssk, GFP_ATOMIC);
3792	}
3793	}
3794
3795	bool mptcp_finish_join(struct sock *ssk)
3796	{
3797	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk: ssk);
3798	struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3799	struct sock *parent = (void *)msk;
3800	bool ret = true;
3801
3802	pr_debug("msk=%p, subflow=%p\n", msk, subflow);
3803
3804	/* mptcp socket already closing? */
3805	if (!mptcp_is_fully_established(sk: parent)) {
3806	subflow->reset_reason = MPTCP_RST_EMPTCP;
3807	return false;
3808	}
3809
3810	/* Active subflow, already present inside the conn_list; is grafted
3811	* either by __mptcp_subflow_connect() or accept.
3812	*/
3813	if (!list_empty(head: &subflow->node)) {
3814	spin_lock_bh(lock: &msk->fallback_lock);
3815	if (!msk->allow_subflows) {
3816	spin_unlock_bh(lock: &msk->fallback_lock);
3817	return false;
3818	}
3819	mptcp_subflow_joined(msk, ssk);
3820	spin_unlock_bh(lock: &msk->fallback_lock);
3821	mptcp_propagate_sndbuf(sk: parent, ssk);
3822	return true;
3823	}
3824
3825	if (!mptcp_pm_allow_new_subflow(msk)) {
3826	MPTCP_INC_STATS(net: sock_net(sk: ssk), field: MPTCP_MIB_JOINREJECTED);
3827	goto err_prohibited;
3828	}
3829
3830	/* If we can't acquire msk socket lock here, let the release callback
3831	* handle it
3832	*/
3833	mptcp_data_lock(parent);
3834	if (!sock_owned_by_user(sk: parent)) {
3835	ret = __mptcp_finish_join(msk, ssk);
3836	if (ret) {
3837	sock_hold(sk: ssk);
3838	list_add_tail(new: &subflow->node, head: &msk->conn_list);
3839	mptcp_sock_check_graft(sk: parent, ssk);
3840	}
3841	} else {
3842	sock_hold(sk: ssk);
3843	list_add_tail(new: &subflow->node, head: &msk->join_list);
3844	__set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3845
3846	/* In case of later failures, __mptcp_flush_join_list() will
3847	* properly orphan the ssk via mptcp_close_ssk().
3848	*/
3849	mptcp_sock_check_graft(sk: parent, ssk);
3850	}
3851	mptcp_data_unlock(parent);
3852
3853	if (!ret) {
3854	err_prohibited:
3855	subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3856	return false;
3857	}
3858
3859	return true;
3860	}
3861
3862	static void mptcp_shutdown(struct sock *sk, int how)
3863	{
3864	pr_debug("sk=%p, how=%d\n", sk, how);
3865
3866	if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3867	__mptcp_wr_shutdown(sk);
3868	}
3869
3870	static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3871	{
3872	const struct sock *sk = (void *)msk;
3873	u64 delta;
3874
3875	if (sk->sk_state == TCP_LISTEN)
3876	return -EINVAL;
3877
3878	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3879	return 0;
3880
3881	delta = msk->write_seq - v;
3882	if (__mptcp_check_fallback(msk) && msk->first) {
3883	struct tcp_sock *tp = tcp_sk(msk->first);
3884
3885	/* the first subflow is disconnected after close - see
3886	* __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3887	* so ignore that status, too.
3888	*/
3889	if (!((1 << msk->first->sk_state) &
3890	(TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3891	delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3892	}
3893	if (delta > INT_MAX)
3894	delta = INT_MAX;
3895
3896	return (int)delta;
3897	}
3898
3899	static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3900	{
3901	struct mptcp_sock *msk = mptcp_sk(sk);
3902	bool slow;
3903
3904	switch (cmd) {
3905	case SIOCINQ:
3906	if (sk->sk_state == TCP_LISTEN)
3907	return -EINVAL;
3908
3909	lock_sock(sk);
3910	if (mptcp_move_skbs(sk))
3911	mptcp_cleanup_rbuf(msk, copied: 0);
3912	*karg = mptcp_inq_hint(sk);
3913	release_sock(sk);
3914	break;
3915	case SIOCOUTQ:
3916	slow = lock_sock_fast(sk);
3917	*karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3918	unlock_sock_fast(sk, slow);
3919	break;
3920	case SIOCOUTQNSD:
3921	slow = lock_sock_fast(sk);
3922	*karg = mptcp_ioctl_outq(msk, v: msk->snd_nxt);
3923	unlock_sock_fast(sk, slow);
3924	break;
3925	default:
3926	return -ENOIOCTLCMD;
3927	}
3928
3929	return 0;
3930	}
3931
3932	static int mptcp_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
3933	int addr_len)
3934	{
3935	struct mptcp_subflow_context *subflow;
3936	struct mptcp_sock *msk = mptcp_sk(sk);
3937	int err = -EINVAL;
3938	struct sock *ssk;
3939
3940	ssk = __mptcp_nmpc_sk(msk);
3941	if (IS_ERR(ptr: ssk))
3942	return PTR_ERR(ptr: ssk);
3943
3944	mptcp_set_state(sk, state: TCP_SYN_SENT);
3945	subflow = mptcp_subflow_ctx(sk: ssk);
3946	#ifdef CONFIG_TCP_MD5SIG
3947	/* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3948	* TCP option space.
3949	*/
3950	if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3951	mptcp_early_fallback(msk, subflow, fb_mib: MPTCP_MIB_MD5SIGFALLBACK);
3952	#endif
3953	if (subflow->request_mptcp) {
3954	if (mptcp_active_should_disable(ssk: sk))
3955	mptcp_early_fallback(msk, subflow,
3956	fb_mib: MPTCP_MIB_MPCAPABLEACTIVEDISABLED);
3957	else if (mptcp_token_new_connect(ssk) < 0)
3958	mptcp_early_fallback(msk, subflow,
3959	fb_mib: MPTCP_MIB_TOKENFALLBACKINIT);
3960	}
3961
3962	WRITE_ONCE(msk->write_seq, subflow->idsn);
3963	WRITE_ONCE(msk->snd_nxt, subflow->idsn);
3964	WRITE_ONCE(msk->snd_una, subflow->idsn);
3965	if (likely(!__mptcp_check_fallback(msk)))
3966	MPTCP_INC_STATS(net: sock_net(sk), field: MPTCP_MIB_MPCAPABLEACTIVE);
3967
3968	/* if reaching here via the fastopen/sendmsg path, the caller already
3969	* acquired the subflow socket lock, too.
3970	*/
3971	if (!msk->fastopening)
3972	lock_sock(sk: ssk);
3973
3974	/* the following mirrors closely a very small chunk of code from
3975	* __inet_stream_connect()
3976	*/
3977	if (ssk->sk_state != TCP_CLOSE)
3978	goto out;
3979
3980	if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3981	err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3982	if (err)
3983	goto out;
3984	}
3985
3986	err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3987	if (err < 0)
3988	goto out;
3989
3990	inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3991
3992	out:
3993	if (!msk->fastopening)
3994	release_sock(sk: ssk);
3995
3996	/* on successful connect, the msk state will be moved to established by
3997	* subflow_finish_connect()
3998	*/
3999	if (unlikely(err)) {
4000	/* avoid leaving a dangling token in an unconnected socket */
4001	mptcp_token_destroy(msk);
4002	mptcp_set_state(sk, state: TCP_CLOSE);
4003	return err;
4004	}
4005
4006	mptcp_copy_inaddrs(msk: sk, ssk);
4007	return 0;
4008	}
4009
4010	static struct proto mptcp_prot = {
4011	.name = "MPTCP",
4012	.owner = THIS_MODULE,
4013	.init = mptcp_init_sock,
4014	.connect = mptcp_connect,
4015	.disconnect = mptcp_disconnect,
4016	.close = mptcp_close,
4017	.setsockopt = mptcp_setsockopt,
4018	.getsockopt = mptcp_getsockopt,
4019	.shutdown = mptcp_shutdown,
4020	.destroy = mptcp_destroy,
4021	.sendmsg = mptcp_sendmsg,
4022	.ioctl = mptcp_ioctl,
4023	.recvmsg = mptcp_recvmsg,
4024	.release_cb = mptcp_release_cb,
4025	.hash = mptcp_hash,
4026	.unhash = mptcp_unhash,
4027	.get_port = mptcp_get_port,
4028	.stream_memory_free = mptcp_stream_memory_free,
4029	.sockets_allocated = &mptcp_sockets_allocated,
4030
4031	.memory_allocated = &net_aligned_data.tcp_memory_allocated,
4032	.per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
4033
4034	.memory_pressure = &tcp_memory_pressure,
4035	.sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
4036	.sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
4037	.sysctl_mem = sysctl_tcp_mem,
4038	.obj_size = sizeof(struct mptcp_sock),
4039	.slab_flags = SLAB_TYPESAFE_BY_RCU,
4040	.no_autobind = true,
4041	};
4042
4043	static int mptcp_bind(struct socket *sock, struct sockaddr_unsized *uaddr, int addr_len)
4044	{
4045	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4046	struct sock *ssk, *sk = sock->sk;
4047	int err = -EINVAL;
4048
4049	lock_sock(sk);
4050	ssk = __mptcp_nmpc_sk(msk);
4051	if (IS_ERR(ptr: ssk)) {
4052	err = PTR_ERR(ptr: ssk);
4053	goto unlock;
4054	}
4055
4056	if (sk->sk_family == AF_INET)
4057	err = inet_bind_sk(sk: ssk, uaddr, addr_len);
4058	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4059	else if (sk->sk_family == AF_INET6)
4060	err = inet6_bind_sk(sk: ssk, uaddr, addr_len);
4061	#endif
4062	if (!err)
4063	mptcp_copy_inaddrs(msk: sk, ssk);
4064
4065	unlock:
4066	release_sock(sk);
4067	return err;
4068	}
4069
4070	static int mptcp_listen(struct socket *sock, int backlog)
4071	{
4072	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4073	struct sock *sk = sock->sk;
4074	struct sock *ssk;
4075	int err;
4076
4077	pr_debug("msk=%p\n", msk);
4078
4079	lock_sock(sk);
4080
4081	err = -EINVAL;
4082	if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
4083	goto unlock;
4084
4085	ssk = __mptcp_nmpc_sk(msk);
4086	if (IS_ERR(ptr: ssk)) {
4087	err = PTR_ERR(ptr: ssk);
4088	goto unlock;
4089	}
4090
4091	mptcp_set_state(sk, state: TCP_LISTEN);
4092	sock_set_flag(sk, flag: SOCK_RCU_FREE);
4093
4094	lock_sock(sk: ssk);
4095	err = __inet_listen_sk(sk: ssk, backlog);
4096	release_sock(sk: ssk);
4097	mptcp_set_state(sk, state: inet_sk_state_load(sk: ssk));
4098
4099	if (!err) {
4100	sock_prot_inuse_add(net: sock_net(sk), prot: sk->sk_prot, val: 1);
4101	mptcp_copy_inaddrs(msk: sk, ssk);
4102	mptcp_event_pm_listener(ssk, event: MPTCP_EVENT_LISTENER_CREATED);
4103	}
4104
4105	unlock:
4106	release_sock(sk);
4107	return err;
4108	}
4109
4110	static void mptcp_graft_subflows(struct sock *sk)
4111	{
4112	struct mptcp_subflow_context *subflow;
4113	struct mptcp_sock *msk = mptcp_sk(sk);
4114
4115	if (mem_cgroup_sockets_enabled) {
4116	LIST_HEAD(join_list);
4117
4118	/* Subflows joining after __inet_accept() will get the
4119	* mem CG properly initialized at mptcp_finish_join() time,
4120	* but subflows pending in join_list need explicit
4121	* initialization before flushing `backlog_unaccounted`
4122	* or MPTCP can later unexpectedly observe unaccounted memory.
4123	*/
4124	mptcp_data_lock(sk);
4125	list_splice_init(list: &msk->join_list, head: &join_list);
4126	mptcp_data_unlock(sk);
4127
4128	__mptcp_flush_join_list(sk, join_list: &join_list);
4129	}
4130
4131	mptcp_for_each_subflow(msk, subflow) {
4132	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4133
4134	lock_sock(sk: ssk);
4135
4136	/* Set ssk->sk_socket of accept()ed flows to mptcp socket.
4137	* This is needed so NOSPACE flag can be set from tcp stack.
4138	*/
4139	if (!ssk->sk_socket)
4140	mptcp_sock_graft(sk: ssk, parent: sk->sk_socket);
4141
4142	if (!mem_cgroup_sk_enabled(sk))
4143	goto unlock;
4144
4145	__mptcp_inherit_cgrp_data(sk, ssk);
4146	__mptcp_inherit_memcg(sk, ssk, GFP_KERNEL);
4147
4148	unlock:
4149	release_sock(sk: ssk);
4150	}
4151
4152	if (mem_cgroup_sk_enabled(sk)) {
4153	gfp_t gfp = GFP_KERNEL | __GFP_NOFAIL;
4154	int amt;
4155
4156	/* Account the backlog memory; prior accept() is aware of
4157	* fwd and rmem only.
4158	*/
4159	mptcp_data_lock(sk);
4160	amt = sk_mem_pages(amt: sk->sk_forward_alloc +
4161	msk->backlog_unaccounted +
4162	atomic_read(v: &sk->sk_rmem_alloc)) -
4163	sk_mem_pages(amt: sk->sk_forward_alloc +
4164	atomic_read(v: &sk->sk_rmem_alloc));
4165	msk->backlog_unaccounted = 0;
4166	mptcp_data_unlock(sk);
4167
4168	if (amt)
4169	mem_cgroup_sk_charge(sk, nr_pages: amt, gfp_mask: gfp);
4170	}
4171	}
4172
4173	static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
4174	struct proto_accept_arg *arg)
4175	{
4176	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4177	struct sock *ssk, *newsk;
4178
4179	pr_debug("msk=%p\n", msk);
4180
4181	/* Buggy applications can call accept on socket states other then LISTEN
4182	* but no need to allocate the first subflow just to error out.
4183	*/
4184	ssk = READ_ONCE(msk->first);
4185	if (!ssk)
4186	return -EINVAL;
4187
4188	pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk));
4189	newsk = inet_csk_accept(sk: ssk, arg);
4190	if (!newsk)
4191	return arg->err;
4192
4193	pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk));
4194	if (sk_is_mptcp(sk: newsk)) {
4195	struct mptcp_subflow_context *subflow;
4196	struct sock *new_mptcp_sock;
4197
4198	subflow = mptcp_subflow_ctx(sk: newsk);
4199	new_mptcp_sock = subflow->conn;
4200
4201	/* is_mptcp should be false if subflow->conn is missing, see
4202	* subflow_syn_recv_sock()
4203	*/
4204	if (WARN_ON_ONCE(!new_mptcp_sock)) {
4205	tcp_sk(newsk)->is_mptcp = 0;
4206	goto tcpfallback;
4207	}
4208
4209	newsk = new_mptcp_sock;
4210	MPTCP_INC_STATS(net: sock_net(sk: ssk), field: MPTCP_MIB_MPCAPABLEPASSIVEACK);
4211
4212	newsk->sk_kern_sock = arg->kern;
4213	lock_sock(sk: newsk);
4214	__inet_accept(sock, newsock, newsk);
4215
4216	set_bit(nr: SOCK_CUSTOM_SOCKOPT, addr: &newsock->flags);
4217	msk = mptcp_sk(newsk);
4218	msk->in_accept_queue = 0;
4219
4220	mptcp_graft_subflows(sk: newsk);
4221	mptcp_rps_record_subflows(msk);
4222
4223	/* Do late cleanup for the first subflow as necessary. Also
4224	* deal with bad peers not doing a complete shutdown.
4225	*/
4226	if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
4227	if (unlikely(list_is_singular(&msk->conn_list)))
4228	mptcp_set_state(sk: newsk, state: TCP_CLOSE);
4229	mptcp_close_ssk(sk: newsk, ssk: msk->first,
4230	subflow: mptcp_subflow_ctx(sk: msk->first));
4231	}
4232	} else {
4233	tcpfallback:
4234	newsk->sk_kern_sock = arg->kern;
4235	lock_sock(sk: newsk);
4236	__inet_accept(sock, newsock, newsk);
4237	/* we are being invoked after accepting a non-mp-capable
4238	* flow: sk is a tcp_sk, not an mptcp one.
4239	*
4240	* Hand the socket over to tcp so all further socket ops
4241	* bypass mptcp.
4242	*/
4243	WRITE_ONCE(newsock->sk->sk_socket->ops,
4244	mptcp_fallback_tcp_ops(newsock->sk));
4245	}
4246	release_sock(sk: newsk);
4247
4248	return 0;
4249	}
4250
4251	static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
4252	{
4253	struct sock *sk = (struct sock *)msk;
4254
4255	if (__mptcp_stream_is_writeable(sk, wake: 1))
4256	return EPOLLOUT | EPOLLWRNORM;
4257
4258	set_bit(nr: SOCK_NOSPACE, addr: &sk->sk_socket->flags);
4259	smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
4260	if (__mptcp_stream_is_writeable(sk, wake: 1))
4261	return EPOLLOUT | EPOLLWRNORM;
4262
4263	return 0;
4264	}
4265
4266	static __poll_t mptcp_poll(struct file *file, struct socket *sock,
4267	struct poll_table_struct *wait)
4268	{
4269	struct sock *sk = sock->sk;
4270	struct mptcp_sock *msk;
4271	__poll_t mask = 0;
4272	u8 shutdown;
4273	int state;
4274
4275	msk = mptcp_sk(sk);
4276	sock_poll_wait(filp: file, sock, p: wait);
4277
4278	state = inet_sk_state_load(sk);
4279	pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags);
4280	if (state == TCP_LISTEN) {
4281	struct sock *ssk = READ_ONCE(msk->first);
4282
4283	if (WARN_ON_ONCE(!ssk))
4284	return 0;
4285
4286	return inet_csk_listen_poll(sk: ssk);
4287	}
4288
4289	shutdown = READ_ONCE(sk->sk_shutdown);
4290	if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
4291	mask |= EPOLLHUP;
4292	if (shutdown & RCV_SHUTDOWN)
4293	mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
4294
4295	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
4296	mask |= mptcp_check_readable(sk);
4297	if (shutdown & SEND_SHUTDOWN)
4298	mask |= EPOLLOUT | EPOLLWRNORM;
4299	else
4300	mask |= mptcp_check_writeable(msk);
4301	} else if (state == TCP_SYN_SENT &&
4302	inet_test_bit(DEFER_CONNECT, sk)) {
4303	/* cf tcp_poll() note about TFO */
4304	mask |= EPOLLOUT | EPOLLWRNORM;
4305	}
4306
4307	/* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
4308	smp_rmb();
4309	if (READ_ONCE(sk->sk_err))
4310	mask |= EPOLLERR;
4311
4312	return mask;
4313	}
4314
4315	static const struct proto_ops mptcp_stream_ops = {
4316	.family = PF_INET,
4317	.owner = THIS_MODULE,
4318	.release = inet_release,
4319	.bind = mptcp_bind,
4320	.connect = inet_stream_connect,
4321	.socketpair = sock_no_socketpair,
4322	.accept = mptcp_stream_accept,
4323	.getname = inet_getname,
4324	.poll = mptcp_poll,
4325	.ioctl = inet_ioctl,
4326	.gettstamp = sock_gettstamp,
4327	.listen = mptcp_listen,
4328	.shutdown = inet_shutdown,
4329	.setsockopt = sock_common_setsockopt,
4330	.getsockopt = sock_common_getsockopt,
4331	.sendmsg = inet_sendmsg,
4332	.recvmsg = inet_recvmsg,
4333	.mmap = sock_no_mmap,
4334	.set_rcvlowat = mptcp_set_rcvlowat,
4335	};
4336
4337	static struct inet_protosw mptcp_protosw = {
4338	.type = SOCK_STREAM,
4339	.protocol = IPPROTO_MPTCP,
4340	.prot = &mptcp_prot,
4341	.ops = &mptcp_stream_ops,
4342	.flags = INET_PROTOSW_ICSK,
4343	};
4344
4345	static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4346	{
4347	struct mptcp_delegated_action *delegated;
4348	struct mptcp_subflow_context *subflow;
4349	int work_done = 0;
4350
4351	delegated = container_of(napi, struct mptcp_delegated_action, napi);
4352	while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4353	struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4354
4355	bh_lock_sock_nested(ssk);
4356	if (!sock_owned_by_user(sk: ssk)) {
4357	mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4358	} else {
4359	/* tcp_release_cb_override already processed
4360	* the action or will do at next release_sock().
4361	* In both case must dequeue the subflow here - on the same
4362	* CPU that scheduled it.
4363	*/
4364	smp_wmb();
4365	clear_bit(MPTCP_DELEGATE_SCHEDULED, addr: &subflow->delegated_status);
4366	}
4367	bh_unlock_sock(ssk);
4368	sock_put(sk: ssk);
4369
4370	if (++work_done == budget)
4371	return budget;
4372	}
4373
4374	/* always provide a 0 'work_done' argument, so that napi_complete_done
4375	* will not try accessing the NULL napi->dev ptr
4376	*/
4377	napi_complete_done(n: napi, work_done: 0);
4378	return work_done;
4379	}
4380
4381	void __init mptcp_proto_init(void)
4382	{
4383	struct mptcp_delegated_action *delegated;
4384	int cpu;
4385
4386	mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4387
4388	if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4389	panic(fmt: "Failed to allocate MPTCP pcpu counter\n");
4390
4391	mptcp_napi_dev = alloc_netdev_dummy(sizeof_priv: 0);
4392	if (!mptcp_napi_dev)
4393	panic(fmt: "Failed to allocate MPTCP dummy netdev\n");
4394	for_each_possible_cpu(cpu) {
4395	delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4396	INIT_LIST_HEAD(list: &delegated->head);
4397	netif_napi_add_tx(dev: mptcp_napi_dev, napi: &delegated->napi,
4398	poll: mptcp_napi_poll);
4399	napi_enable(n: &delegated->napi);
4400	}
4401
4402	mptcp_subflow_init();
4403	mptcp_pm_init();
4404	mptcp_sched_init();
4405	mptcp_token_init();
4406
4407	if (proto_register(prot: &mptcp_prot, alloc_slab: 1) != 0)
4408	panic(fmt: "Failed to register MPTCP proto.\n");
4409
4410	inet_register_protosw(p: &mptcp_protosw);
4411
4412	BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4413	}
4414
4415	#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4416	static const struct proto_ops mptcp_v6_stream_ops = {
4417	.family = PF_INET6,
4418	.owner = THIS_MODULE,
4419	.release = inet6_release,
4420	.bind = mptcp_bind,
4421	.connect = inet_stream_connect,
4422	.socketpair = sock_no_socketpair,
4423	.accept = mptcp_stream_accept,
4424	.getname = inet6_getname,
4425	.poll = mptcp_poll,
4426	.ioctl = inet6_ioctl,
4427	.gettstamp = sock_gettstamp,
4428	.listen = mptcp_listen,
4429	.shutdown = inet_shutdown,
4430	.setsockopt = sock_common_setsockopt,
4431	.getsockopt = sock_common_getsockopt,
4432	.sendmsg = inet6_sendmsg,
4433	.recvmsg = inet6_recvmsg,
4434	.mmap = sock_no_mmap,
4435	#ifdef CONFIG_COMPAT
4436	.compat_ioctl = inet6_compat_ioctl,
4437	#endif
4438	.set_rcvlowat = mptcp_set_rcvlowat,
4439	};
4440
4441	static struct proto mptcp_v6_prot;
4442
4443	static struct inet_protosw mptcp_v6_protosw = {
4444	.type = SOCK_STREAM,
4445	.protocol = IPPROTO_MPTCP,
4446	.prot = &mptcp_v6_prot,
4447	.ops = &mptcp_v6_stream_ops,
4448	.flags = INET_PROTOSW_ICSK,
4449	};
4450
4451	int __init mptcp_proto_v6_init(void)
4452	{
4453	int err;
4454
4455	mptcp_v6_prot = mptcp_prot;
4456	strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name));
4457	mptcp_v6_prot.slab = NULL;
4458	mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4459	mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4460
4461	err = proto_register(prot: &mptcp_v6_prot, alloc_slab: 1);
4462	if (err)
4463	return err;
4464
4465	err = inet6_register_protosw(p: &mptcp_v6_protosw);
4466	if (err)
4467	proto_unregister(prot: &mptcp_v6_prot);
4468
4469	return err;
4470	}
4471	#endif
4472