xref: /linux/net/mptcp/protocol.c (revision 4461568aa4e565de2c336f4875ddf912f26da8a5)
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/sock.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
18 #include <net/tcp.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
22 #endif
23 #include <net/mptcp.h>
24 #include <net/xfrm.h>
25 #include <asm/ioctls.h>
26 #include "protocol.h"
27 #include "mib.h"
28 
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
31 
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
33 struct mptcp6_sock {
34 	struct mptcp_sock msk;
35 	struct ipv6_pinfo np;
36 };
37 #endif
38 
39 struct mptcp_skb_cb {
40 	u64 map_seq;
41 	u64 end_seq;
42 	u32 offset;
43 	u8  has_rxtstamp:1;
44 };
45 
46 #define MPTCP_SKB_CB(__skb)	((struct mptcp_skb_cb *)&((__skb)->cb[0]))
47 
48 enum {
49 	MPTCP_CMSG_TS = BIT(0),
50 	MPTCP_CMSG_INQ = BIT(1),
51 };
52 
53 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
54 
55 static void __mptcp_destroy_sock(struct sock *sk);
56 static void __mptcp_check_send_data_fin(struct sock *sk);
57 
58 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
59 static struct net_device mptcp_napi_dev;
60 
61 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
62  * completed yet or has failed, return the subflow socket.
63  * Otherwise return NULL.
64  */
65 struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
66 {
67 	if (!msk->subflow || READ_ONCE(msk->can_ack))
68 		return NULL;
69 
70 	return msk->subflow;
71 }
72 
73 /* Returns end sequence number of the receiver's advertised window */
74 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
75 {
76 	return READ_ONCE(msk->wnd_end);
77 }
78 
79 static bool mptcp_is_tcpsk(struct sock *sk)
80 {
81 	struct socket *sock = sk->sk_socket;
82 
83 	if (unlikely(sk->sk_prot == &tcp_prot)) {
84 		/* we are being invoked after mptcp_accept() has
85 		 * accepted a non-mp-capable flow: sk is a tcp_sk,
86 		 * not an mptcp one.
87 		 *
88 		 * Hand the socket over to tcp so all further socket ops
89 		 * bypass mptcp.
90 		 */
91 		sock->ops = &inet_stream_ops;
92 		return true;
93 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
94 	} else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
95 		sock->ops = &inet6_stream_ops;
96 		return true;
97 #endif
98 	}
99 
100 	return false;
101 }
102 
103 static int __mptcp_socket_create(struct mptcp_sock *msk)
104 {
105 	struct mptcp_subflow_context *subflow;
106 	struct sock *sk = (struct sock *)msk;
107 	struct socket *ssock;
108 	int err;
109 
110 	err = mptcp_subflow_create_socket(sk, &ssock);
111 	if (err)
112 		return err;
113 
114 	msk->first = ssock->sk;
115 	msk->subflow = ssock;
116 	subflow = mptcp_subflow_ctx(ssock->sk);
117 	list_add(&subflow->node, &msk->conn_list);
118 	sock_hold(ssock->sk);
119 	subflow->request_mptcp = 1;
120 
121 	/* This is the first subflow, always with id 0 */
122 	subflow->local_id_valid = 1;
123 	mptcp_sock_graft(msk->first, sk->sk_socket);
124 
125 	return 0;
126 }
127 
128 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
129 {
130 	sk_drops_add(sk, skb);
131 	__kfree_skb(skb);
132 }
133 
134 static void mptcp_rmem_charge(struct sock *sk, int size)
135 {
136 	mptcp_sk(sk)->rmem_fwd_alloc -= size;
137 }
138 
139 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
140 			       struct sk_buff *from)
141 {
142 	bool fragstolen;
143 	int delta;
144 
145 	if (MPTCP_SKB_CB(from)->offset ||
146 	    !skb_try_coalesce(to, from, &fragstolen, &delta))
147 		return false;
148 
149 	pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
150 		 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
151 		 to->len, MPTCP_SKB_CB(from)->end_seq);
152 	MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
153 	kfree_skb_partial(from, fragstolen);
154 	atomic_add(delta, &sk->sk_rmem_alloc);
155 	mptcp_rmem_charge(sk, delta);
156 	return true;
157 }
158 
159 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
160 				   struct sk_buff *from)
161 {
162 	if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
163 		return false;
164 
165 	return mptcp_try_coalesce((struct sock *)msk, to, from);
166 }
167 
168 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
169 {
170 	amount >>= PAGE_SHIFT;
171 	mptcp_sk(sk)->rmem_fwd_alloc -= amount << PAGE_SHIFT;
172 	__sk_mem_reduce_allocated(sk, amount);
173 }
174 
175 static void mptcp_rmem_uncharge(struct sock *sk, int size)
176 {
177 	struct mptcp_sock *msk = mptcp_sk(sk);
178 	int reclaimable;
179 
180 	msk->rmem_fwd_alloc += size;
181 	reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
182 
183 	/* see sk_mem_uncharge() for the rationale behind the following schema */
184 	if (unlikely(reclaimable >= PAGE_SIZE))
185 		__mptcp_rmem_reclaim(sk, reclaimable);
186 }
187 
188 static void mptcp_rfree(struct sk_buff *skb)
189 {
190 	unsigned int len = skb->truesize;
191 	struct sock *sk = skb->sk;
192 
193 	atomic_sub(len, &sk->sk_rmem_alloc);
194 	mptcp_rmem_uncharge(sk, len);
195 }
196 
197 static void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
198 {
199 	skb_orphan(skb);
200 	skb->sk = sk;
201 	skb->destructor = mptcp_rfree;
202 	atomic_add(skb->truesize, &sk->sk_rmem_alloc);
203 	mptcp_rmem_charge(sk, skb->truesize);
204 }
205 
206 /* "inspired" by tcp_data_queue_ofo(), main differences:
207  * - use mptcp seqs
208  * - don't cope with sacks
209  */
210 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
211 {
212 	struct sock *sk = (struct sock *)msk;
213 	struct rb_node **p, *parent;
214 	u64 seq, end_seq, max_seq;
215 	struct sk_buff *skb1;
216 
217 	seq = MPTCP_SKB_CB(skb)->map_seq;
218 	end_seq = MPTCP_SKB_CB(skb)->end_seq;
219 	max_seq = atomic64_read(&msk->rcv_wnd_sent);
220 
221 	pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
222 		 RB_EMPTY_ROOT(&msk->out_of_order_queue));
223 	if (after64(end_seq, max_seq)) {
224 		/* out of window */
225 		mptcp_drop(sk, skb);
226 		pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
227 			 (unsigned long long)end_seq - (unsigned long)max_seq,
228 			 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
229 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
230 		return;
231 	}
232 
233 	p = &msk->out_of_order_queue.rb_node;
234 	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
235 	if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
236 		rb_link_node(&skb->rbnode, NULL, p);
237 		rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
238 		msk->ooo_last_skb = skb;
239 		goto end;
240 	}
241 
242 	/* with 2 subflows, adding at end of ooo queue is quite likely
243 	 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
244 	 */
245 	if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
246 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
247 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
248 		return;
249 	}
250 
251 	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
252 	if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
253 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
254 		parent = &msk->ooo_last_skb->rbnode;
255 		p = &parent->rb_right;
256 		goto insert;
257 	}
258 
259 	/* Find place to insert this segment. Handle overlaps on the way. */
260 	parent = NULL;
261 	while (*p) {
262 		parent = *p;
263 		skb1 = rb_to_skb(parent);
264 		if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
265 			p = &parent->rb_left;
266 			continue;
267 		}
268 		if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
269 			if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
270 				/* All the bits are present. Drop. */
271 				mptcp_drop(sk, skb);
272 				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
273 				return;
274 			}
275 			if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
276 				/* partial overlap:
277 				 *     |     skb      |
278 				 *  |     skb1    |
279 				 * continue traversing
280 				 */
281 			} else {
282 				/* skb's seq == skb1's seq and skb covers skb1.
283 				 * Replace skb1 with skb.
284 				 */
285 				rb_replace_node(&skb1->rbnode, &skb->rbnode,
286 						&msk->out_of_order_queue);
287 				mptcp_drop(sk, skb1);
288 				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
289 				goto merge_right;
290 			}
291 		} else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
292 			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
293 			return;
294 		}
295 		p = &parent->rb_right;
296 	}
297 
298 insert:
299 	/* Insert segment into RB tree. */
300 	rb_link_node(&skb->rbnode, parent, p);
301 	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
302 
303 merge_right:
304 	/* Remove other segments covered by skb. */
305 	while ((skb1 = skb_rb_next(skb)) != NULL) {
306 		if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
307 			break;
308 		rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
309 		mptcp_drop(sk, skb1);
310 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
311 	}
312 	/* If there is no skb after us, we are the last_skb ! */
313 	if (!skb1)
314 		msk->ooo_last_skb = skb;
315 
316 end:
317 	skb_condense(skb);
318 	mptcp_set_owner_r(skb, sk);
319 }
320 
321 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
322 {
323 	struct mptcp_sock *msk = mptcp_sk(sk);
324 	int amt, amount;
325 
326 	if (size <= msk->rmem_fwd_alloc)
327 		return true;
328 
329 	size -= msk->rmem_fwd_alloc;
330 	amt = sk_mem_pages(size);
331 	amount = amt << PAGE_SHIFT;
332 	if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
333 		return false;
334 
335 	msk->rmem_fwd_alloc += amount;
336 	return true;
337 }
338 
339 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
340 			     struct sk_buff *skb, unsigned int offset,
341 			     size_t copy_len)
342 {
343 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
344 	struct sock *sk = (struct sock *)msk;
345 	struct sk_buff *tail;
346 	bool has_rxtstamp;
347 
348 	__skb_unlink(skb, &ssk->sk_receive_queue);
349 
350 	skb_ext_reset(skb);
351 	skb_orphan(skb);
352 
353 	/* try to fetch required memory from subflow */
354 	if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
355 		goto drop;
356 
357 	has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
358 
359 	/* the skb map_seq accounts for the skb offset:
360 	 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
361 	 * value
362 	 */
363 	MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
364 	MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
365 	MPTCP_SKB_CB(skb)->offset = offset;
366 	MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
367 
368 	if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
369 		/* in sequence */
370 		WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
371 		tail = skb_peek_tail(&sk->sk_receive_queue);
372 		if (tail && mptcp_try_coalesce(sk, tail, skb))
373 			return true;
374 
375 		mptcp_set_owner_r(skb, sk);
376 		__skb_queue_tail(&sk->sk_receive_queue, skb);
377 		return true;
378 	} else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
379 		mptcp_data_queue_ofo(msk, skb);
380 		return false;
381 	}
382 
383 	/* old data, keep it simple and drop the whole pkt, sender
384 	 * will retransmit as needed, if needed.
385 	 */
386 	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
387 drop:
388 	mptcp_drop(sk, skb);
389 	return false;
390 }
391 
392 static void mptcp_stop_timer(struct sock *sk)
393 {
394 	struct inet_connection_sock *icsk = inet_csk(sk);
395 
396 	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
397 	mptcp_sk(sk)->timer_ival = 0;
398 }
399 
400 static void mptcp_close_wake_up(struct sock *sk)
401 {
402 	if (sock_flag(sk, SOCK_DEAD))
403 		return;
404 
405 	sk->sk_state_change(sk);
406 	if (sk->sk_shutdown == SHUTDOWN_MASK ||
407 	    sk->sk_state == TCP_CLOSE)
408 		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
409 	else
410 		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
411 }
412 
413 static bool mptcp_pending_data_fin_ack(struct sock *sk)
414 {
415 	struct mptcp_sock *msk = mptcp_sk(sk);
416 
417 	return !__mptcp_check_fallback(msk) &&
418 	       ((1 << sk->sk_state) &
419 		(TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
420 	       msk->write_seq == READ_ONCE(msk->snd_una);
421 }
422 
423 static void mptcp_check_data_fin_ack(struct sock *sk)
424 {
425 	struct mptcp_sock *msk = mptcp_sk(sk);
426 
427 	/* Look for an acknowledged DATA_FIN */
428 	if (mptcp_pending_data_fin_ack(sk)) {
429 		WRITE_ONCE(msk->snd_data_fin_enable, 0);
430 
431 		switch (sk->sk_state) {
432 		case TCP_FIN_WAIT1:
433 			inet_sk_state_store(sk, TCP_FIN_WAIT2);
434 			break;
435 		case TCP_CLOSING:
436 		case TCP_LAST_ACK:
437 			inet_sk_state_store(sk, TCP_CLOSE);
438 			break;
439 		}
440 
441 		mptcp_close_wake_up(sk);
442 	}
443 }
444 
445 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
446 {
447 	struct mptcp_sock *msk = mptcp_sk(sk);
448 
449 	if (READ_ONCE(msk->rcv_data_fin) &&
450 	    ((1 << sk->sk_state) &
451 	     (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
452 		u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
453 
454 		if (msk->ack_seq == rcv_data_fin_seq) {
455 			if (seq)
456 				*seq = rcv_data_fin_seq;
457 
458 			return true;
459 		}
460 	}
461 
462 	return false;
463 }
464 
465 static void mptcp_set_datafin_timeout(const struct sock *sk)
466 {
467 	struct inet_connection_sock *icsk = inet_csk(sk);
468 	u32 retransmits;
469 
470 	retransmits = min_t(u32, icsk->icsk_retransmits,
471 			    ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
472 
473 	mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
474 }
475 
476 static void __mptcp_set_timeout(struct sock *sk, long tout)
477 {
478 	mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
479 }
480 
481 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
482 {
483 	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
484 
485 	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
486 	       inet_csk(ssk)->icsk_timeout - jiffies : 0;
487 }
488 
489 static void mptcp_set_timeout(struct sock *sk)
490 {
491 	struct mptcp_subflow_context *subflow;
492 	long tout = 0;
493 
494 	mptcp_for_each_subflow(mptcp_sk(sk), subflow)
495 		tout = max(tout, mptcp_timeout_from_subflow(subflow));
496 	__mptcp_set_timeout(sk, tout);
497 }
498 
499 static inline bool tcp_can_send_ack(const struct sock *ssk)
500 {
501 	return !((1 << inet_sk_state_load(ssk)) &
502 	       (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
503 }
504 
505 void __mptcp_subflow_send_ack(struct sock *ssk)
506 {
507 	if (tcp_can_send_ack(ssk))
508 		tcp_send_ack(ssk);
509 }
510 
511 static void mptcp_subflow_send_ack(struct sock *ssk)
512 {
513 	bool slow;
514 
515 	slow = lock_sock_fast(ssk);
516 	__mptcp_subflow_send_ack(ssk);
517 	unlock_sock_fast(ssk, slow);
518 }
519 
520 static void mptcp_send_ack(struct mptcp_sock *msk)
521 {
522 	struct mptcp_subflow_context *subflow;
523 
524 	mptcp_for_each_subflow(msk, subflow)
525 		mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
526 }
527 
528 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
529 {
530 	bool slow;
531 
532 	slow = lock_sock_fast(ssk);
533 	if (tcp_can_send_ack(ssk))
534 		tcp_cleanup_rbuf(ssk, 1);
535 	unlock_sock_fast(ssk, slow);
536 }
537 
538 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
539 {
540 	const struct inet_connection_sock *icsk = inet_csk(ssk);
541 	u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
542 	const struct tcp_sock *tp = tcp_sk(ssk);
543 
544 	return (ack_pending & ICSK_ACK_SCHED) &&
545 		((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
546 		  READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
547 		 (rx_empty && ack_pending &
548 			      (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
549 }
550 
551 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
552 {
553 	int old_space = READ_ONCE(msk->old_wspace);
554 	struct mptcp_subflow_context *subflow;
555 	struct sock *sk = (struct sock *)msk;
556 	int space =  __mptcp_space(sk);
557 	bool cleanup, rx_empty;
558 
559 	cleanup = (space > 0) && (space >= (old_space << 1));
560 	rx_empty = !__mptcp_rmem(sk);
561 
562 	mptcp_for_each_subflow(msk, subflow) {
563 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
564 
565 		if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
566 			mptcp_subflow_cleanup_rbuf(ssk);
567 	}
568 }
569 
570 static bool mptcp_check_data_fin(struct sock *sk)
571 {
572 	struct mptcp_sock *msk = mptcp_sk(sk);
573 	u64 rcv_data_fin_seq;
574 	bool ret = false;
575 
576 	if (__mptcp_check_fallback(msk))
577 		return ret;
578 
579 	/* Need to ack a DATA_FIN received from a peer while this side
580 	 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
581 	 * msk->rcv_data_fin was set when parsing the incoming options
582 	 * at the subflow level and the msk lock was not held, so this
583 	 * is the first opportunity to act on the DATA_FIN and change
584 	 * the msk state.
585 	 *
586 	 * If we are caught up to the sequence number of the incoming
587 	 * DATA_FIN, send the DATA_ACK now and do state transition.  If
588 	 * not caught up, do nothing and let the recv code send DATA_ACK
589 	 * when catching up.
590 	 */
591 
592 	if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
593 		WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
594 		WRITE_ONCE(msk->rcv_data_fin, 0);
595 
596 		sk->sk_shutdown |= RCV_SHUTDOWN;
597 		smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
598 
599 		switch (sk->sk_state) {
600 		case TCP_ESTABLISHED:
601 			inet_sk_state_store(sk, TCP_CLOSE_WAIT);
602 			break;
603 		case TCP_FIN_WAIT1:
604 			inet_sk_state_store(sk, TCP_CLOSING);
605 			break;
606 		case TCP_FIN_WAIT2:
607 			inet_sk_state_store(sk, TCP_CLOSE);
608 			break;
609 		default:
610 			/* Other states not expected */
611 			WARN_ON_ONCE(1);
612 			break;
613 		}
614 
615 		ret = true;
616 		mptcp_send_ack(msk);
617 		mptcp_close_wake_up(sk);
618 	}
619 	return ret;
620 }
621 
622 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
623 					   struct sock *ssk,
624 					   unsigned int *bytes)
625 {
626 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
627 	struct sock *sk = (struct sock *)msk;
628 	unsigned int moved = 0;
629 	bool more_data_avail;
630 	struct tcp_sock *tp;
631 	bool done = false;
632 	int sk_rbuf;
633 
634 	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
635 
636 	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
637 		int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
638 
639 		if (unlikely(ssk_rbuf > sk_rbuf)) {
640 			WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
641 			sk_rbuf = ssk_rbuf;
642 		}
643 	}
644 
645 	pr_debug("msk=%p ssk=%p", msk, ssk);
646 	tp = tcp_sk(ssk);
647 	do {
648 		u32 map_remaining, offset;
649 		u32 seq = tp->copied_seq;
650 		struct sk_buff *skb;
651 		bool fin;
652 
653 		/* try to move as much data as available */
654 		map_remaining = subflow->map_data_len -
655 				mptcp_subflow_get_map_offset(subflow);
656 
657 		skb = skb_peek(&ssk->sk_receive_queue);
658 		if (!skb) {
659 			/* if no data is found, a racing workqueue/recvmsg
660 			 * already processed the new data, stop here or we
661 			 * can enter an infinite loop
662 			 */
663 			if (!moved)
664 				done = true;
665 			break;
666 		}
667 
668 		if (__mptcp_check_fallback(msk)) {
669 			/* if we are running under the workqueue, TCP could have
670 			 * collapsed skbs between dummy map creation and now
671 			 * be sure to adjust the size
672 			 */
673 			map_remaining = skb->len;
674 			subflow->map_data_len = skb->len;
675 		}
676 
677 		offset = seq - TCP_SKB_CB(skb)->seq;
678 		fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
679 		if (fin) {
680 			done = true;
681 			seq++;
682 		}
683 
684 		if (offset < skb->len) {
685 			size_t len = skb->len - offset;
686 
687 			if (tp->urg_data)
688 				done = true;
689 
690 			if (__mptcp_move_skb(msk, ssk, skb, offset, len))
691 				moved += len;
692 			seq += len;
693 
694 			if (WARN_ON_ONCE(map_remaining < len))
695 				break;
696 		} else {
697 			WARN_ON_ONCE(!fin);
698 			sk_eat_skb(ssk, skb);
699 			done = true;
700 		}
701 
702 		WRITE_ONCE(tp->copied_seq, seq);
703 		more_data_avail = mptcp_subflow_data_available(ssk);
704 
705 		if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
706 			done = true;
707 			break;
708 		}
709 	} while (more_data_avail);
710 
711 	*bytes += moved;
712 	return done;
713 }
714 
715 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
716 {
717 	struct sock *sk = (struct sock *)msk;
718 	struct sk_buff *skb, *tail;
719 	bool moved = false;
720 	struct rb_node *p;
721 	u64 end_seq;
722 
723 	p = rb_first(&msk->out_of_order_queue);
724 	pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
725 	while (p) {
726 		skb = rb_to_skb(p);
727 		if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
728 			break;
729 
730 		p = rb_next(p);
731 		rb_erase(&skb->rbnode, &msk->out_of_order_queue);
732 
733 		if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
734 				      msk->ack_seq))) {
735 			mptcp_drop(sk, skb);
736 			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
737 			continue;
738 		}
739 
740 		end_seq = MPTCP_SKB_CB(skb)->end_seq;
741 		tail = skb_peek_tail(&sk->sk_receive_queue);
742 		if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
743 			int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
744 
745 			/* skip overlapping data, if any */
746 			pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
747 				 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
748 				 delta);
749 			MPTCP_SKB_CB(skb)->offset += delta;
750 			MPTCP_SKB_CB(skb)->map_seq += delta;
751 			__skb_queue_tail(&sk->sk_receive_queue, skb);
752 		}
753 		msk->ack_seq = end_seq;
754 		moved = true;
755 	}
756 	return moved;
757 }
758 
759 /* In most cases we will be able to lock the mptcp socket.  If its already
760  * owned, we need to defer to the work queue to avoid ABBA deadlock.
761  */
762 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
763 {
764 	struct sock *sk = (struct sock *)msk;
765 	unsigned int moved = 0;
766 
767 	__mptcp_move_skbs_from_subflow(msk, ssk, &moved);
768 	__mptcp_ofo_queue(msk);
769 	if (unlikely(ssk->sk_err)) {
770 		if (!sock_owned_by_user(sk))
771 			__mptcp_error_report(sk);
772 		else
773 			__set_bit(MPTCP_ERROR_REPORT,  &msk->cb_flags);
774 	}
775 
776 	/* If the moves have caught up with the DATA_FIN sequence number
777 	 * it's time to ack the DATA_FIN and change socket state, but
778 	 * this is not a good place to change state. Let the workqueue
779 	 * do it.
780 	 */
781 	if (mptcp_pending_data_fin(sk, NULL))
782 		mptcp_schedule_work(sk);
783 	return moved > 0;
784 }
785 
786 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
787 {
788 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
789 	struct mptcp_sock *msk = mptcp_sk(sk);
790 	int sk_rbuf, ssk_rbuf;
791 
792 	/* The peer can send data while we are shutting down this
793 	 * subflow at msk destruction time, but we must avoid enqueuing
794 	 * more data to the msk receive queue
795 	 */
796 	if (unlikely(subflow->disposable))
797 		return;
798 
799 	ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
800 	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
801 	if (unlikely(ssk_rbuf > sk_rbuf))
802 		sk_rbuf = ssk_rbuf;
803 
804 	/* over limit? can't append more skbs to msk, Also, no need to wake-up*/
805 	if (__mptcp_rmem(sk) > sk_rbuf) {
806 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
807 		return;
808 	}
809 
810 	/* Wake-up the reader only for in-sequence data */
811 	mptcp_data_lock(sk);
812 	if (move_skbs_to_msk(msk, ssk))
813 		sk->sk_data_ready(sk);
814 
815 	mptcp_data_unlock(sk);
816 }
817 
818 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
819 {
820 	struct sock *sk = (struct sock *)msk;
821 
822 	if (sk->sk_state != TCP_ESTABLISHED)
823 		return false;
824 
825 	/* attach to msk socket only after we are sure we will deal with it
826 	 * at close time
827 	 */
828 	if (sk->sk_socket && !ssk->sk_socket)
829 		mptcp_sock_graft(ssk, sk->sk_socket);
830 
831 	mptcp_propagate_sndbuf((struct sock *)msk, ssk);
832 	mptcp_sockopt_sync_locked(msk, ssk);
833 	return true;
834 }
835 
836 static void __mptcp_flush_join_list(struct sock *sk)
837 {
838 	struct mptcp_subflow_context *tmp, *subflow;
839 	struct mptcp_sock *msk = mptcp_sk(sk);
840 
841 	list_for_each_entry_safe(subflow, tmp, &msk->join_list, node) {
842 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
843 		bool slow = lock_sock_fast(ssk);
844 
845 		list_move_tail(&subflow->node, &msk->conn_list);
846 		if (!__mptcp_finish_join(msk, ssk))
847 			mptcp_subflow_reset(ssk);
848 		unlock_sock_fast(ssk, slow);
849 	}
850 }
851 
852 static bool mptcp_timer_pending(struct sock *sk)
853 {
854 	return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
855 }
856 
857 static void mptcp_reset_timer(struct sock *sk)
858 {
859 	struct inet_connection_sock *icsk = inet_csk(sk);
860 	unsigned long tout;
861 
862 	/* prevent rescheduling on close */
863 	if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
864 		return;
865 
866 	tout = mptcp_sk(sk)->timer_ival;
867 	sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
868 }
869 
870 bool mptcp_schedule_work(struct sock *sk)
871 {
872 	if (inet_sk_state_load(sk) != TCP_CLOSE &&
873 	    schedule_work(&mptcp_sk(sk)->work)) {
874 		/* each subflow already holds a reference to the sk, and the
875 		 * workqueue is invoked by a subflow, so sk can't go away here.
876 		 */
877 		sock_hold(sk);
878 		return true;
879 	}
880 	return false;
881 }
882 
883 void mptcp_subflow_eof(struct sock *sk)
884 {
885 	if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
886 		mptcp_schedule_work(sk);
887 }
888 
889 static void mptcp_check_for_eof(struct mptcp_sock *msk)
890 {
891 	struct mptcp_subflow_context *subflow;
892 	struct sock *sk = (struct sock *)msk;
893 	int receivers = 0;
894 
895 	mptcp_for_each_subflow(msk, subflow)
896 		receivers += !subflow->rx_eof;
897 	if (receivers)
898 		return;
899 
900 	if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
901 		/* hopefully temporary hack: propagate shutdown status
902 		 * to msk, when all subflows agree on it
903 		 */
904 		sk->sk_shutdown |= RCV_SHUTDOWN;
905 
906 		smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
907 		sk->sk_data_ready(sk);
908 	}
909 
910 	switch (sk->sk_state) {
911 	case TCP_ESTABLISHED:
912 		inet_sk_state_store(sk, TCP_CLOSE_WAIT);
913 		break;
914 	case TCP_FIN_WAIT1:
915 		inet_sk_state_store(sk, TCP_CLOSING);
916 		break;
917 	case TCP_FIN_WAIT2:
918 		inet_sk_state_store(sk, TCP_CLOSE);
919 		break;
920 	default:
921 		return;
922 	}
923 	mptcp_close_wake_up(sk);
924 }
925 
926 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
927 {
928 	struct mptcp_subflow_context *subflow;
929 	struct sock *sk = (struct sock *)msk;
930 
931 	sock_owned_by_me(sk);
932 
933 	mptcp_for_each_subflow(msk, subflow) {
934 		if (READ_ONCE(subflow->data_avail))
935 			return mptcp_subflow_tcp_sock(subflow);
936 	}
937 
938 	return NULL;
939 }
940 
941 static bool mptcp_skb_can_collapse_to(u64 write_seq,
942 				      const struct sk_buff *skb,
943 				      const struct mptcp_ext *mpext)
944 {
945 	if (!tcp_skb_can_collapse_to(skb))
946 		return false;
947 
948 	/* can collapse only if MPTCP level sequence is in order and this
949 	 * mapping has not been xmitted yet
950 	 */
951 	return mpext && mpext->data_seq + mpext->data_len == write_seq &&
952 	       !mpext->frozen;
953 }
954 
955 /* we can append data to the given data frag if:
956  * - there is space available in the backing page_frag
957  * - the data frag tail matches the current page_frag free offset
958  * - the data frag end sequence number matches the current write seq
959  */
960 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
961 				       const struct page_frag *pfrag,
962 				       const struct mptcp_data_frag *df)
963 {
964 	return df && pfrag->page == df->page &&
965 		pfrag->size - pfrag->offset > 0 &&
966 		pfrag->offset == (df->offset + df->data_len) &&
967 		df->data_seq + df->data_len == msk->write_seq;
968 }
969 
970 static void dfrag_uncharge(struct sock *sk, int len)
971 {
972 	sk_mem_uncharge(sk, len);
973 	sk_wmem_queued_add(sk, -len);
974 }
975 
976 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
977 {
978 	int len = dfrag->data_len + dfrag->overhead;
979 
980 	list_del(&dfrag->list);
981 	dfrag_uncharge(sk, len);
982 	put_page(dfrag->page);
983 }
984 
985 static void __mptcp_clean_una(struct sock *sk)
986 {
987 	struct mptcp_sock *msk = mptcp_sk(sk);
988 	struct mptcp_data_frag *dtmp, *dfrag;
989 	u64 snd_una;
990 
991 	/* on fallback we just need to ignore snd_una, as this is really
992 	 * plain TCP
993 	 */
994 	if (__mptcp_check_fallback(msk))
995 		msk->snd_una = READ_ONCE(msk->snd_nxt);
996 
997 	snd_una = msk->snd_una;
998 	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
999 		if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1000 			break;
1001 
1002 		if (unlikely(dfrag == msk->first_pending)) {
1003 			/* in recovery mode can see ack after the current snd head */
1004 			if (WARN_ON_ONCE(!msk->recovery))
1005 				break;
1006 
1007 			WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1008 		}
1009 
1010 		dfrag_clear(sk, dfrag);
1011 	}
1012 
1013 	dfrag = mptcp_rtx_head(sk);
1014 	if (dfrag && after64(snd_una, dfrag->data_seq)) {
1015 		u64 delta = snd_una - dfrag->data_seq;
1016 
1017 		/* prevent wrap around in recovery mode */
1018 		if (unlikely(delta > dfrag->already_sent)) {
1019 			if (WARN_ON_ONCE(!msk->recovery))
1020 				goto out;
1021 			if (WARN_ON_ONCE(delta > dfrag->data_len))
1022 				goto out;
1023 			dfrag->already_sent += delta - dfrag->already_sent;
1024 		}
1025 
1026 		dfrag->data_seq += delta;
1027 		dfrag->offset += delta;
1028 		dfrag->data_len -= delta;
1029 		dfrag->already_sent -= delta;
1030 
1031 		dfrag_uncharge(sk, delta);
1032 	}
1033 
1034 	/* all retransmitted data acked, recovery completed */
1035 	if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1036 		msk->recovery = false;
1037 
1038 out:
1039 	if (snd_una == READ_ONCE(msk->snd_nxt) &&
1040 	    snd_una == READ_ONCE(msk->write_seq)) {
1041 		if (mptcp_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1042 			mptcp_stop_timer(sk);
1043 	} else {
1044 		mptcp_reset_timer(sk);
1045 	}
1046 }
1047 
1048 static void __mptcp_clean_una_wakeup(struct sock *sk)
1049 {
1050 	lockdep_assert_held_once(&sk->sk_lock.slock);
1051 
1052 	__mptcp_clean_una(sk);
1053 	mptcp_write_space(sk);
1054 }
1055 
1056 static void mptcp_clean_una_wakeup(struct sock *sk)
1057 {
1058 	mptcp_data_lock(sk);
1059 	__mptcp_clean_una_wakeup(sk);
1060 	mptcp_data_unlock(sk);
1061 }
1062 
1063 static void mptcp_enter_memory_pressure(struct sock *sk)
1064 {
1065 	struct mptcp_subflow_context *subflow;
1066 	struct mptcp_sock *msk = mptcp_sk(sk);
1067 	bool first = true;
1068 
1069 	sk_stream_moderate_sndbuf(sk);
1070 	mptcp_for_each_subflow(msk, subflow) {
1071 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1072 
1073 		if (first)
1074 			tcp_enter_memory_pressure(ssk);
1075 		sk_stream_moderate_sndbuf(ssk);
1076 		first = false;
1077 	}
1078 }
1079 
1080 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1081  * data
1082  */
1083 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1084 {
1085 	if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1086 					pfrag, sk->sk_allocation)))
1087 		return true;
1088 
1089 	mptcp_enter_memory_pressure(sk);
1090 	return false;
1091 }
1092 
1093 static struct mptcp_data_frag *
1094 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1095 		      int orig_offset)
1096 {
1097 	int offset = ALIGN(orig_offset, sizeof(long));
1098 	struct mptcp_data_frag *dfrag;
1099 
1100 	dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1101 	dfrag->data_len = 0;
1102 	dfrag->data_seq = msk->write_seq;
1103 	dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1104 	dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1105 	dfrag->already_sent = 0;
1106 	dfrag->page = pfrag->page;
1107 
1108 	return dfrag;
1109 }
1110 
1111 struct mptcp_sendmsg_info {
1112 	int mss_now;
1113 	int size_goal;
1114 	u16 limit;
1115 	u16 sent;
1116 	unsigned int flags;
1117 	bool data_lock_held;
1118 };
1119 
1120 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1121 				    u64 data_seq, int avail_size)
1122 {
1123 	u64 window_end = mptcp_wnd_end(msk);
1124 	u64 mptcp_snd_wnd;
1125 
1126 	if (__mptcp_check_fallback(msk))
1127 		return avail_size;
1128 
1129 	mptcp_snd_wnd = window_end - data_seq;
1130 	avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1131 
1132 	if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1133 		tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1134 		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1135 	}
1136 
1137 	return avail_size;
1138 }
1139 
1140 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1141 {
1142 	struct skb_ext *mpext = __skb_ext_alloc(gfp);
1143 
1144 	if (!mpext)
1145 		return false;
1146 	__skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1147 	return true;
1148 }
1149 
1150 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1151 {
1152 	struct sk_buff *skb;
1153 
1154 	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1155 	if (likely(skb)) {
1156 		if (likely(__mptcp_add_ext(skb, gfp))) {
1157 			skb_reserve(skb, MAX_TCP_HEADER);
1158 			skb->ip_summed = CHECKSUM_PARTIAL;
1159 			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1160 			return skb;
1161 		}
1162 		__kfree_skb(skb);
1163 	} else {
1164 		mptcp_enter_memory_pressure(sk);
1165 	}
1166 	return NULL;
1167 }
1168 
1169 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1170 {
1171 	struct sk_buff *skb;
1172 
1173 	skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1174 	if (!skb)
1175 		return NULL;
1176 
1177 	if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1178 		tcp_skb_entail(ssk, skb);
1179 		return skb;
1180 	}
1181 	tcp_skb_tsorted_anchor_cleanup(skb);
1182 	kfree_skb(skb);
1183 	return NULL;
1184 }
1185 
1186 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1187 {
1188 	gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1189 
1190 	return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1191 }
1192 
1193 /* note: this always recompute the csum on the whole skb, even
1194  * if we just appended a single frag. More status info needed
1195  */
1196 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1197 {
1198 	struct mptcp_ext *mpext = mptcp_get_ext(skb);
1199 	__wsum csum = ~csum_unfold(mpext->csum);
1200 	int offset = skb->len - added;
1201 
1202 	mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1203 }
1204 
1205 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1206 				      struct sock *ssk,
1207 				      struct mptcp_ext *mpext)
1208 {
1209 	if (!mpext)
1210 		return;
1211 
1212 	mpext->infinite_map = 1;
1213 	mpext->data_len = 0;
1214 
1215 	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1216 	mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1217 	pr_fallback(msk);
1218 	mptcp_do_fallback(ssk);
1219 }
1220 
1221 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1222 			      struct mptcp_data_frag *dfrag,
1223 			      struct mptcp_sendmsg_info *info)
1224 {
1225 	u64 data_seq = dfrag->data_seq + info->sent;
1226 	int offset = dfrag->offset + info->sent;
1227 	struct mptcp_sock *msk = mptcp_sk(sk);
1228 	bool zero_window_probe = false;
1229 	struct mptcp_ext *mpext = NULL;
1230 	bool can_coalesce = false;
1231 	bool reuse_skb = true;
1232 	struct sk_buff *skb;
1233 	size_t copy;
1234 	int i;
1235 
1236 	pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1237 		 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1238 
1239 	if (WARN_ON_ONCE(info->sent > info->limit ||
1240 			 info->limit > dfrag->data_len))
1241 		return 0;
1242 
1243 	if (unlikely(!__tcp_can_send(ssk)))
1244 		return -EAGAIN;
1245 
1246 	/* compute send limit */
1247 	info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1248 	copy = info->size_goal;
1249 
1250 	skb = tcp_write_queue_tail(ssk);
1251 	if (skb && copy > skb->len) {
1252 		/* Limit the write to the size available in the
1253 		 * current skb, if any, so that we create at most a new skb.
1254 		 * Explicitly tells TCP internals to avoid collapsing on later
1255 		 * queue management operation, to avoid breaking the ext <->
1256 		 * SSN association set here
1257 		 */
1258 		mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1259 		if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1260 			TCP_SKB_CB(skb)->eor = 1;
1261 			goto alloc_skb;
1262 		}
1263 
1264 		i = skb_shinfo(skb)->nr_frags;
1265 		can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1266 		if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1267 			tcp_mark_push(tcp_sk(ssk), skb);
1268 			goto alloc_skb;
1269 		}
1270 
1271 		copy -= skb->len;
1272 	} else {
1273 alloc_skb:
1274 		skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1275 		if (!skb)
1276 			return -ENOMEM;
1277 
1278 		i = skb_shinfo(skb)->nr_frags;
1279 		reuse_skb = false;
1280 		mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1281 	}
1282 
1283 	/* Zero window and all data acked? Probe. */
1284 	copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1285 	if (copy == 0) {
1286 		u64 snd_una = READ_ONCE(msk->snd_una);
1287 
1288 		if (snd_una != msk->snd_nxt) {
1289 			tcp_remove_empty_skb(ssk);
1290 			return 0;
1291 		}
1292 
1293 		zero_window_probe = true;
1294 		data_seq = snd_una - 1;
1295 		copy = 1;
1296 
1297 		/* all mptcp-level data is acked, no skbs should be present into the
1298 		 * ssk write queue
1299 		 */
1300 		WARN_ON_ONCE(reuse_skb);
1301 	}
1302 
1303 	copy = min_t(size_t, copy, info->limit - info->sent);
1304 	if (!sk_wmem_schedule(ssk, copy)) {
1305 		tcp_remove_empty_skb(ssk);
1306 		return -ENOMEM;
1307 	}
1308 
1309 	if (can_coalesce) {
1310 		skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1311 	} else {
1312 		get_page(dfrag->page);
1313 		skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1314 	}
1315 
1316 	skb->len += copy;
1317 	skb->data_len += copy;
1318 	skb->truesize += copy;
1319 	sk_wmem_queued_add(ssk, copy);
1320 	sk_mem_charge(ssk, copy);
1321 	WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1322 	TCP_SKB_CB(skb)->end_seq += copy;
1323 	tcp_skb_pcount_set(skb, 0);
1324 
1325 	/* on skb reuse we just need to update the DSS len */
1326 	if (reuse_skb) {
1327 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1328 		mpext->data_len += copy;
1329 		WARN_ON_ONCE(zero_window_probe);
1330 		goto out;
1331 	}
1332 
1333 	memset(mpext, 0, sizeof(*mpext));
1334 	mpext->data_seq = data_seq;
1335 	mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1336 	mpext->data_len = copy;
1337 	mpext->use_map = 1;
1338 	mpext->dsn64 = 1;
1339 
1340 	pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1341 		 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1342 		 mpext->dsn64);
1343 
1344 	if (zero_window_probe) {
1345 		mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1346 		mpext->frozen = 1;
1347 		if (READ_ONCE(msk->csum_enabled))
1348 			mptcp_update_data_checksum(skb, copy);
1349 		tcp_push_pending_frames(ssk);
1350 		return 0;
1351 	}
1352 out:
1353 	if (READ_ONCE(msk->csum_enabled))
1354 		mptcp_update_data_checksum(skb, copy);
1355 	if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1356 		mptcp_update_infinite_map(msk, ssk, mpext);
1357 	trace_mptcp_sendmsg_frag(mpext);
1358 	mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1359 	return copy;
1360 }
1361 
1362 #define MPTCP_SEND_BURST_SIZE		((1 << 16) - \
1363 					 sizeof(struct tcphdr) - \
1364 					 MAX_TCP_OPTION_SPACE - \
1365 					 sizeof(struct ipv6hdr) - \
1366 					 sizeof(struct frag_hdr))
1367 
1368 struct subflow_send_info {
1369 	struct sock *ssk;
1370 	u64 linger_time;
1371 };
1372 
1373 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1374 {
1375 	if (!subflow->stale)
1376 		return;
1377 
1378 	subflow->stale = 0;
1379 	MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1380 }
1381 
1382 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1383 {
1384 	if (unlikely(subflow->stale)) {
1385 		u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1386 
1387 		if (subflow->stale_rcv_tstamp == rcv_tstamp)
1388 			return false;
1389 
1390 		mptcp_subflow_set_active(subflow);
1391 	}
1392 	return __mptcp_subflow_active(subflow);
1393 }
1394 
1395 #define SSK_MODE_ACTIVE	0
1396 #define SSK_MODE_BACKUP	1
1397 #define SSK_MODE_MAX	2
1398 
1399 /* implement the mptcp packet scheduler;
1400  * returns the subflow that will transmit the next DSS
1401  * additionally updates the rtx timeout
1402  */
1403 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1404 {
1405 	struct subflow_send_info send_info[SSK_MODE_MAX];
1406 	struct mptcp_subflow_context *subflow;
1407 	struct sock *sk = (struct sock *)msk;
1408 	u32 pace, burst, wmem;
1409 	int i, nr_active = 0;
1410 	struct sock *ssk;
1411 	u64 linger_time;
1412 	long tout = 0;
1413 
1414 	sock_owned_by_me(sk);
1415 
1416 	if (__mptcp_check_fallback(msk)) {
1417 		if (!msk->first)
1418 			return NULL;
1419 		return __tcp_can_send(msk->first) &&
1420 		       sk_stream_memory_free(msk->first) ? msk->first : NULL;
1421 	}
1422 
1423 	/* re-use last subflow, if the burst allow that */
1424 	if (msk->last_snd && msk->snd_burst > 0 &&
1425 	    sk_stream_memory_free(msk->last_snd) &&
1426 	    mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1427 		mptcp_set_timeout(sk);
1428 		return msk->last_snd;
1429 	}
1430 
1431 	/* pick the subflow with the lower wmem/wspace ratio */
1432 	for (i = 0; i < SSK_MODE_MAX; ++i) {
1433 		send_info[i].ssk = NULL;
1434 		send_info[i].linger_time = -1;
1435 	}
1436 
1437 	mptcp_for_each_subflow(msk, subflow) {
1438 		trace_mptcp_subflow_get_send(subflow);
1439 		ssk =  mptcp_subflow_tcp_sock(subflow);
1440 		if (!mptcp_subflow_active(subflow))
1441 			continue;
1442 
1443 		tout = max(tout, mptcp_timeout_from_subflow(subflow));
1444 		nr_active += !subflow->backup;
1445 		pace = subflow->avg_pacing_rate;
1446 		if (unlikely(!pace)) {
1447 			/* init pacing rate from socket */
1448 			subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1449 			pace = subflow->avg_pacing_rate;
1450 			if (!pace)
1451 				continue;
1452 		}
1453 
1454 		linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1455 		if (linger_time < send_info[subflow->backup].linger_time) {
1456 			send_info[subflow->backup].ssk = ssk;
1457 			send_info[subflow->backup].linger_time = linger_time;
1458 		}
1459 	}
1460 	__mptcp_set_timeout(sk, tout);
1461 
1462 	/* pick the best backup if no other subflow is active */
1463 	if (!nr_active)
1464 		send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1465 
1466 	/* According to the blest algorithm, to avoid HoL blocking for the
1467 	 * faster flow, we need to:
1468 	 * - estimate the faster flow linger time
1469 	 * - use the above to estimate the amount of byte transferred
1470 	 *   by the faster flow
1471 	 * - check that the amount of queued data is greter than the above,
1472 	 *   otherwise do not use the picked, slower, subflow
1473 	 * We select the subflow with the shorter estimated time to flush
1474 	 * the queued mem, which basically ensure the above. We just need
1475 	 * to check that subflow has a non empty cwin.
1476 	 */
1477 	ssk = send_info[SSK_MODE_ACTIVE].ssk;
1478 	if (!ssk || !sk_stream_memory_free(ssk))
1479 		return NULL;
1480 
1481 	burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1482 	wmem = READ_ONCE(ssk->sk_wmem_queued);
1483 	if (!burst) {
1484 		msk->last_snd = NULL;
1485 		return ssk;
1486 	}
1487 
1488 	subflow = mptcp_subflow_ctx(ssk);
1489 	subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1490 					   READ_ONCE(ssk->sk_pacing_rate) * burst,
1491 					   burst + wmem);
1492 	msk->last_snd = ssk;
1493 	msk->snd_burst = burst;
1494 	return ssk;
1495 }
1496 
1497 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1498 {
1499 	tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1500 	release_sock(ssk);
1501 }
1502 
1503 static void mptcp_update_post_push(struct mptcp_sock *msk,
1504 				   struct mptcp_data_frag *dfrag,
1505 				   u32 sent)
1506 {
1507 	u64 snd_nxt_new = dfrag->data_seq;
1508 
1509 	dfrag->already_sent += sent;
1510 
1511 	msk->snd_burst -= sent;
1512 
1513 	snd_nxt_new += dfrag->already_sent;
1514 
1515 	/* snd_nxt_new can be smaller than snd_nxt in case mptcp
1516 	 * is recovering after a failover. In that event, this re-sends
1517 	 * old segments.
1518 	 *
1519 	 * Thus compute snd_nxt_new candidate based on
1520 	 * the dfrag->data_seq that was sent and the data
1521 	 * that has been handed to the subflow for transmission
1522 	 * and skip update in case it was old dfrag.
1523 	 */
1524 	if (likely(after64(snd_nxt_new, msk->snd_nxt)))
1525 		msk->snd_nxt = snd_nxt_new;
1526 }
1527 
1528 void mptcp_check_and_set_pending(struct sock *sk)
1529 {
1530 	if (mptcp_send_head(sk))
1531 		mptcp_sk(sk)->push_pending |= BIT(MPTCP_PUSH_PENDING);
1532 }
1533 
1534 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1535 {
1536 	struct sock *prev_ssk = NULL, *ssk = NULL;
1537 	struct mptcp_sock *msk = mptcp_sk(sk);
1538 	struct mptcp_sendmsg_info info = {
1539 				.flags = flags,
1540 	};
1541 	bool do_check_data_fin = false;
1542 	struct mptcp_data_frag *dfrag;
1543 	int len;
1544 
1545 	while ((dfrag = mptcp_send_head(sk))) {
1546 		info.sent = dfrag->already_sent;
1547 		info.limit = dfrag->data_len;
1548 		len = dfrag->data_len - dfrag->already_sent;
1549 		while (len > 0) {
1550 			int ret = 0;
1551 
1552 			prev_ssk = ssk;
1553 			ssk = mptcp_subflow_get_send(msk);
1554 
1555 			/* First check. If the ssk has changed since
1556 			 * the last round, release prev_ssk
1557 			 */
1558 			if (ssk != prev_ssk && prev_ssk)
1559 				mptcp_push_release(prev_ssk, &info);
1560 			if (!ssk)
1561 				goto out;
1562 
1563 			/* Need to lock the new subflow only if different
1564 			 * from the previous one, otherwise we are still
1565 			 * helding the relevant lock
1566 			 */
1567 			if (ssk != prev_ssk)
1568 				lock_sock(ssk);
1569 
1570 			ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1571 			if (ret <= 0) {
1572 				if (ret == -EAGAIN)
1573 					continue;
1574 				mptcp_push_release(ssk, &info);
1575 				goto out;
1576 			}
1577 
1578 			do_check_data_fin = true;
1579 			info.sent += ret;
1580 			len -= ret;
1581 
1582 			mptcp_update_post_push(msk, dfrag, ret);
1583 		}
1584 		WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1585 	}
1586 
1587 	/* at this point we held the socket lock for the last subflow we used */
1588 	if (ssk)
1589 		mptcp_push_release(ssk, &info);
1590 
1591 out:
1592 	/* ensure the rtx timer is running */
1593 	if (!mptcp_timer_pending(sk))
1594 		mptcp_reset_timer(sk);
1595 	if (do_check_data_fin)
1596 		__mptcp_check_send_data_fin(sk);
1597 }
1598 
1599 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1600 {
1601 	struct mptcp_sock *msk = mptcp_sk(sk);
1602 	struct mptcp_sendmsg_info info = {
1603 		.data_lock_held = true,
1604 	};
1605 	struct mptcp_data_frag *dfrag;
1606 	struct sock *xmit_ssk;
1607 	int len, copied = 0;
1608 	bool first = true;
1609 
1610 	info.flags = 0;
1611 	while ((dfrag = mptcp_send_head(sk))) {
1612 		info.sent = dfrag->already_sent;
1613 		info.limit = dfrag->data_len;
1614 		len = dfrag->data_len - dfrag->already_sent;
1615 		while (len > 0) {
1616 			int ret = 0;
1617 
1618 			/* the caller already invoked the packet scheduler,
1619 			 * check for a different subflow usage only after
1620 			 * spooling the first chunk of data
1621 			 */
1622 			xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1623 			if (!xmit_ssk)
1624 				goto out;
1625 			if (xmit_ssk != ssk) {
1626 				mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk),
1627 						       MPTCP_DELEGATE_SEND);
1628 				goto out;
1629 			}
1630 
1631 			ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1632 			if (ret <= 0)
1633 				goto out;
1634 
1635 			info.sent += ret;
1636 			copied += ret;
1637 			len -= ret;
1638 			first = false;
1639 
1640 			mptcp_update_post_push(msk, dfrag, ret);
1641 		}
1642 		WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1643 	}
1644 
1645 out:
1646 	/* __mptcp_alloc_tx_skb could have released some wmem and we are
1647 	 * not going to flush it via release_sock()
1648 	 */
1649 	if (copied) {
1650 		tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1651 			 info.size_goal);
1652 		if (!mptcp_timer_pending(sk))
1653 			mptcp_reset_timer(sk);
1654 
1655 		if (msk->snd_data_fin_enable &&
1656 		    msk->snd_nxt + 1 == msk->write_seq)
1657 			mptcp_schedule_work(sk);
1658 	}
1659 }
1660 
1661 static void mptcp_set_nospace(struct sock *sk)
1662 {
1663 	/* enable autotune */
1664 	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1665 
1666 	/* will be cleared on avail space */
1667 	set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1668 }
1669 
1670 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1671 {
1672 	struct mptcp_sock *msk = mptcp_sk(sk);
1673 	struct page_frag *pfrag;
1674 	size_t copied = 0;
1675 	int ret = 0;
1676 	long timeo;
1677 
1678 	/* we don't support FASTOPEN yet */
1679 	if (msg->msg_flags & MSG_FASTOPEN)
1680 		return -EOPNOTSUPP;
1681 
1682 	/* silently ignore everything else */
1683 	msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1684 
1685 	lock_sock(sk);
1686 
1687 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1688 
1689 	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1690 		ret = sk_stream_wait_connect(sk, &timeo);
1691 		if (ret)
1692 			goto out;
1693 	}
1694 
1695 	pfrag = sk_page_frag(sk);
1696 
1697 	while (msg_data_left(msg)) {
1698 		int total_ts, frag_truesize = 0;
1699 		struct mptcp_data_frag *dfrag;
1700 		bool dfrag_collapsed;
1701 		size_t psize, offset;
1702 
1703 		if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1704 			ret = -EPIPE;
1705 			goto out;
1706 		}
1707 
1708 		/* reuse tail pfrag, if possible, or carve a new one from the
1709 		 * page allocator
1710 		 */
1711 		dfrag = mptcp_pending_tail(sk);
1712 		dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1713 		if (!dfrag_collapsed) {
1714 			if (!sk_stream_memory_free(sk))
1715 				goto wait_for_memory;
1716 
1717 			if (!mptcp_page_frag_refill(sk, pfrag))
1718 				goto wait_for_memory;
1719 
1720 			dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1721 			frag_truesize = dfrag->overhead;
1722 		}
1723 
1724 		/* we do not bound vs wspace, to allow a single packet.
1725 		 * memory accounting will prevent execessive memory usage
1726 		 * anyway
1727 		 */
1728 		offset = dfrag->offset + dfrag->data_len;
1729 		psize = pfrag->size - offset;
1730 		psize = min_t(size_t, psize, msg_data_left(msg));
1731 		total_ts = psize + frag_truesize;
1732 
1733 		if (!sk_wmem_schedule(sk, total_ts))
1734 			goto wait_for_memory;
1735 
1736 		if (copy_page_from_iter(dfrag->page, offset, psize,
1737 					&msg->msg_iter) != psize) {
1738 			ret = -EFAULT;
1739 			goto out;
1740 		}
1741 
1742 		/* data successfully copied into the write queue */
1743 		sk->sk_forward_alloc -= total_ts;
1744 		copied += psize;
1745 		dfrag->data_len += psize;
1746 		frag_truesize += psize;
1747 		pfrag->offset += frag_truesize;
1748 		WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1749 
1750 		/* charge data on mptcp pending queue to the msk socket
1751 		 * Note: we charge such data both to sk and ssk
1752 		 */
1753 		sk_wmem_queued_add(sk, frag_truesize);
1754 		if (!dfrag_collapsed) {
1755 			get_page(dfrag->page);
1756 			list_add_tail(&dfrag->list, &msk->rtx_queue);
1757 			if (!msk->first_pending)
1758 				WRITE_ONCE(msk->first_pending, dfrag);
1759 		}
1760 		pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1761 			 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1762 			 !dfrag_collapsed);
1763 
1764 		continue;
1765 
1766 wait_for_memory:
1767 		mptcp_set_nospace(sk);
1768 		__mptcp_push_pending(sk, msg->msg_flags);
1769 		ret = sk_stream_wait_memory(sk, &timeo);
1770 		if (ret)
1771 			goto out;
1772 	}
1773 
1774 	if (copied)
1775 		__mptcp_push_pending(sk, msg->msg_flags);
1776 
1777 out:
1778 	release_sock(sk);
1779 	return copied ? : ret;
1780 }
1781 
1782 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1783 				struct msghdr *msg,
1784 				size_t len, int flags,
1785 				struct scm_timestamping_internal *tss,
1786 				int *cmsg_flags)
1787 {
1788 	struct sk_buff *skb, *tmp;
1789 	int copied = 0;
1790 
1791 	skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1792 		u32 offset = MPTCP_SKB_CB(skb)->offset;
1793 		u32 data_len = skb->len - offset;
1794 		u32 count = min_t(size_t, len - copied, data_len);
1795 		int err;
1796 
1797 		if (!(flags & MSG_TRUNC)) {
1798 			err = skb_copy_datagram_msg(skb, offset, msg, count);
1799 			if (unlikely(err < 0)) {
1800 				if (!copied)
1801 					return err;
1802 				break;
1803 			}
1804 		}
1805 
1806 		if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1807 			tcp_update_recv_tstamps(skb, tss);
1808 			*cmsg_flags |= MPTCP_CMSG_TS;
1809 		}
1810 
1811 		copied += count;
1812 
1813 		if (count < data_len) {
1814 			if (!(flags & MSG_PEEK)) {
1815 				MPTCP_SKB_CB(skb)->offset += count;
1816 				MPTCP_SKB_CB(skb)->map_seq += count;
1817 			}
1818 			break;
1819 		}
1820 
1821 		if (!(flags & MSG_PEEK)) {
1822 			/* we will bulk release the skb memory later */
1823 			skb->destructor = NULL;
1824 			WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1825 			__skb_unlink(skb, &msk->receive_queue);
1826 			__kfree_skb(skb);
1827 		}
1828 
1829 		if (copied >= len)
1830 			break;
1831 	}
1832 
1833 	return copied;
1834 }
1835 
1836 /* receive buffer autotuning.  See tcp_rcv_space_adjust for more information.
1837  *
1838  * Only difference: Use highest rtt estimate of the subflows in use.
1839  */
1840 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1841 {
1842 	struct mptcp_subflow_context *subflow;
1843 	struct sock *sk = (struct sock *)msk;
1844 	u32 time, advmss = 1;
1845 	u64 rtt_us, mstamp;
1846 
1847 	sock_owned_by_me(sk);
1848 
1849 	if (copied <= 0)
1850 		return;
1851 
1852 	msk->rcvq_space.copied += copied;
1853 
1854 	mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1855 	time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1856 
1857 	rtt_us = msk->rcvq_space.rtt_us;
1858 	if (rtt_us && time < (rtt_us >> 3))
1859 		return;
1860 
1861 	rtt_us = 0;
1862 	mptcp_for_each_subflow(msk, subflow) {
1863 		const struct tcp_sock *tp;
1864 		u64 sf_rtt_us;
1865 		u32 sf_advmss;
1866 
1867 		tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1868 
1869 		sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1870 		sf_advmss = READ_ONCE(tp->advmss);
1871 
1872 		rtt_us = max(sf_rtt_us, rtt_us);
1873 		advmss = max(sf_advmss, advmss);
1874 	}
1875 
1876 	msk->rcvq_space.rtt_us = rtt_us;
1877 	if (time < (rtt_us >> 3) || rtt_us == 0)
1878 		return;
1879 
1880 	if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1881 		goto new_measure;
1882 
1883 	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
1884 	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1885 		int rcvmem, rcvbuf;
1886 		u64 rcvwin, grow;
1887 
1888 		rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1889 
1890 		grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1891 
1892 		do_div(grow, msk->rcvq_space.space);
1893 		rcvwin += (grow << 1);
1894 
1895 		rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1896 		while (tcp_win_from_space(sk, rcvmem) < advmss)
1897 			rcvmem += 128;
1898 
1899 		do_div(rcvwin, advmss);
1900 		rcvbuf = min_t(u64, rcvwin * rcvmem,
1901 			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
1902 
1903 		if (rcvbuf > sk->sk_rcvbuf) {
1904 			u32 window_clamp;
1905 
1906 			window_clamp = tcp_win_from_space(sk, rcvbuf);
1907 			WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1908 
1909 			/* Make subflows follow along.  If we do not do this, we
1910 			 * get drops at subflow level if skbs can't be moved to
1911 			 * the mptcp rx queue fast enough (announced rcv_win can
1912 			 * exceed ssk->sk_rcvbuf).
1913 			 */
1914 			mptcp_for_each_subflow(msk, subflow) {
1915 				struct sock *ssk;
1916 				bool slow;
1917 
1918 				ssk = mptcp_subflow_tcp_sock(subflow);
1919 				slow = lock_sock_fast(ssk);
1920 				WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1921 				tcp_sk(ssk)->window_clamp = window_clamp;
1922 				tcp_cleanup_rbuf(ssk, 1);
1923 				unlock_sock_fast(ssk, slow);
1924 			}
1925 		}
1926 	}
1927 
1928 	msk->rcvq_space.space = msk->rcvq_space.copied;
1929 new_measure:
1930 	msk->rcvq_space.copied = 0;
1931 	msk->rcvq_space.time = mstamp;
1932 }
1933 
1934 static void __mptcp_update_rmem(struct sock *sk)
1935 {
1936 	struct mptcp_sock *msk = mptcp_sk(sk);
1937 
1938 	if (!msk->rmem_released)
1939 		return;
1940 
1941 	atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1942 	mptcp_rmem_uncharge(sk, msk->rmem_released);
1943 	WRITE_ONCE(msk->rmem_released, 0);
1944 }
1945 
1946 static void __mptcp_splice_receive_queue(struct sock *sk)
1947 {
1948 	struct mptcp_sock *msk = mptcp_sk(sk);
1949 
1950 	skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1951 }
1952 
1953 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1954 {
1955 	struct sock *sk = (struct sock *)msk;
1956 	unsigned int moved = 0;
1957 	bool ret, done;
1958 
1959 	do {
1960 		struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1961 		bool slowpath;
1962 
1963 		/* we can have data pending in the subflows only if the msk
1964 		 * receive buffer was full at subflow_data_ready() time,
1965 		 * that is an unlikely slow path.
1966 		 */
1967 		if (likely(!ssk))
1968 			break;
1969 
1970 		slowpath = lock_sock_fast(ssk);
1971 		mptcp_data_lock(sk);
1972 		__mptcp_update_rmem(sk);
1973 		done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1974 		mptcp_data_unlock(sk);
1975 
1976 		if (unlikely(ssk->sk_err))
1977 			__mptcp_error_report(sk);
1978 		unlock_sock_fast(ssk, slowpath);
1979 	} while (!done);
1980 
1981 	/* acquire the data lock only if some input data is pending */
1982 	ret = moved > 0;
1983 	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1984 	    !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1985 		mptcp_data_lock(sk);
1986 		__mptcp_update_rmem(sk);
1987 		ret |= __mptcp_ofo_queue(msk);
1988 		__mptcp_splice_receive_queue(sk);
1989 		mptcp_data_unlock(sk);
1990 	}
1991 	if (ret)
1992 		mptcp_check_data_fin((struct sock *)msk);
1993 	return !skb_queue_empty(&msk->receive_queue);
1994 }
1995 
1996 static unsigned int mptcp_inq_hint(const struct sock *sk)
1997 {
1998 	const struct mptcp_sock *msk = mptcp_sk(sk);
1999 	const struct sk_buff *skb;
2000 
2001 	skb = skb_peek(&msk->receive_queue);
2002 	if (skb) {
2003 		u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2004 
2005 		if (hint_val >= INT_MAX)
2006 			return INT_MAX;
2007 
2008 		return (unsigned int)hint_val;
2009 	}
2010 
2011 	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2012 		return 1;
2013 
2014 	return 0;
2015 }
2016 
2017 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2018 			 int flags, int *addr_len)
2019 {
2020 	struct mptcp_sock *msk = mptcp_sk(sk);
2021 	struct scm_timestamping_internal tss;
2022 	int copied = 0, cmsg_flags = 0;
2023 	int target;
2024 	long timeo;
2025 
2026 	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2027 	if (unlikely(flags & MSG_ERRQUEUE))
2028 		return inet_recv_error(sk, msg, len, addr_len);
2029 
2030 	lock_sock(sk);
2031 	if (unlikely(sk->sk_state == TCP_LISTEN)) {
2032 		copied = -ENOTCONN;
2033 		goto out_err;
2034 	}
2035 
2036 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2037 
2038 	len = min_t(size_t, len, INT_MAX);
2039 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2040 
2041 	if (unlikely(msk->recvmsg_inq))
2042 		cmsg_flags = MPTCP_CMSG_INQ;
2043 
2044 	while (copied < len) {
2045 		int bytes_read;
2046 
2047 		bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2048 		if (unlikely(bytes_read < 0)) {
2049 			if (!copied)
2050 				copied = bytes_read;
2051 			goto out_err;
2052 		}
2053 
2054 		copied += bytes_read;
2055 
2056 		/* be sure to advertise window change */
2057 		mptcp_cleanup_rbuf(msk);
2058 
2059 		if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2060 			continue;
2061 
2062 		/* only the master socket status is relevant here. The exit
2063 		 * conditions mirror closely tcp_recvmsg()
2064 		 */
2065 		if (copied >= target)
2066 			break;
2067 
2068 		if (copied) {
2069 			if (sk->sk_err ||
2070 			    sk->sk_state == TCP_CLOSE ||
2071 			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2072 			    !timeo ||
2073 			    signal_pending(current))
2074 				break;
2075 		} else {
2076 			if (sk->sk_err) {
2077 				copied = sock_error(sk);
2078 				break;
2079 			}
2080 
2081 			if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2082 				mptcp_check_for_eof(msk);
2083 
2084 			if (sk->sk_shutdown & RCV_SHUTDOWN) {
2085 				/* race breaker: the shutdown could be after the
2086 				 * previous receive queue check
2087 				 */
2088 				if (__mptcp_move_skbs(msk))
2089 					continue;
2090 				break;
2091 			}
2092 
2093 			if (sk->sk_state == TCP_CLOSE) {
2094 				copied = -ENOTCONN;
2095 				break;
2096 			}
2097 
2098 			if (!timeo) {
2099 				copied = -EAGAIN;
2100 				break;
2101 			}
2102 
2103 			if (signal_pending(current)) {
2104 				copied = sock_intr_errno(timeo);
2105 				break;
2106 			}
2107 		}
2108 
2109 		pr_debug("block timeout %ld", timeo);
2110 		sk_wait_data(sk, &timeo, NULL);
2111 	}
2112 
2113 out_err:
2114 	if (cmsg_flags && copied >= 0) {
2115 		if (cmsg_flags & MPTCP_CMSG_TS)
2116 			tcp_recv_timestamp(msg, sk, &tss);
2117 
2118 		if (cmsg_flags & MPTCP_CMSG_INQ) {
2119 			unsigned int inq = mptcp_inq_hint(sk);
2120 
2121 			put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2122 		}
2123 	}
2124 
2125 	pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2126 		 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2127 		 skb_queue_empty(&msk->receive_queue), copied);
2128 	if (!(flags & MSG_PEEK))
2129 		mptcp_rcv_space_adjust(msk, copied);
2130 
2131 	release_sock(sk);
2132 	return copied;
2133 }
2134 
2135 static void mptcp_retransmit_timer(struct timer_list *t)
2136 {
2137 	struct inet_connection_sock *icsk = from_timer(icsk, t,
2138 						       icsk_retransmit_timer);
2139 	struct sock *sk = &icsk->icsk_inet.sk;
2140 	struct mptcp_sock *msk = mptcp_sk(sk);
2141 
2142 	bh_lock_sock(sk);
2143 	if (!sock_owned_by_user(sk)) {
2144 		/* we need a process context to retransmit */
2145 		if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2146 			mptcp_schedule_work(sk);
2147 	} else {
2148 		/* delegate our work to tcp_release_cb() */
2149 		__set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2150 	}
2151 	bh_unlock_sock(sk);
2152 	sock_put(sk);
2153 }
2154 
2155 static void mptcp_timeout_timer(struct timer_list *t)
2156 {
2157 	struct sock *sk = from_timer(sk, t, sk_timer);
2158 
2159 	mptcp_schedule_work(sk);
2160 	sock_put(sk);
2161 }
2162 
2163 /* Find an idle subflow.  Return NULL if there is unacked data at tcp
2164  * level.
2165  *
2166  * A backup subflow is returned only if that is the only kind available.
2167  */
2168 static struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2169 {
2170 	struct sock *backup = NULL, *pick = NULL;
2171 	struct mptcp_subflow_context *subflow;
2172 	int min_stale_count = INT_MAX;
2173 
2174 	sock_owned_by_me((const struct sock *)msk);
2175 
2176 	if (__mptcp_check_fallback(msk))
2177 		return NULL;
2178 
2179 	mptcp_for_each_subflow(msk, subflow) {
2180 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2181 
2182 		if (!__mptcp_subflow_active(subflow))
2183 			continue;
2184 
2185 		/* still data outstanding at TCP level? skip this */
2186 		if (!tcp_rtx_and_write_queues_empty(ssk)) {
2187 			mptcp_pm_subflow_chk_stale(msk, ssk);
2188 			min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2189 			continue;
2190 		}
2191 
2192 		if (subflow->backup) {
2193 			if (!backup)
2194 				backup = ssk;
2195 			continue;
2196 		}
2197 
2198 		if (!pick)
2199 			pick = ssk;
2200 	}
2201 
2202 	if (pick)
2203 		return pick;
2204 
2205 	/* use backup only if there are no progresses anywhere */
2206 	return min_stale_count > 1 ? backup : NULL;
2207 }
2208 
2209 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2210 {
2211 	if (msk->subflow) {
2212 		iput(SOCK_INODE(msk->subflow));
2213 		msk->subflow = NULL;
2214 	}
2215 }
2216 
2217 bool __mptcp_retransmit_pending_data(struct sock *sk)
2218 {
2219 	struct mptcp_data_frag *cur, *rtx_head;
2220 	struct mptcp_sock *msk = mptcp_sk(sk);
2221 
2222 	if (__mptcp_check_fallback(mptcp_sk(sk)))
2223 		return false;
2224 
2225 	if (tcp_rtx_and_write_queues_empty(sk))
2226 		return false;
2227 
2228 	/* the closing socket has some data untransmitted and/or unacked:
2229 	 * some data in the mptcp rtx queue has not really xmitted yet.
2230 	 * keep it simple and re-inject the whole mptcp level rtx queue
2231 	 */
2232 	mptcp_data_lock(sk);
2233 	__mptcp_clean_una_wakeup(sk);
2234 	rtx_head = mptcp_rtx_head(sk);
2235 	if (!rtx_head) {
2236 		mptcp_data_unlock(sk);
2237 		return false;
2238 	}
2239 
2240 	msk->recovery_snd_nxt = msk->snd_nxt;
2241 	msk->recovery = true;
2242 	mptcp_data_unlock(sk);
2243 
2244 	msk->first_pending = rtx_head;
2245 	msk->snd_burst = 0;
2246 
2247 	/* be sure to clear the "sent status" on all re-injected fragments */
2248 	list_for_each_entry(cur, &msk->rtx_queue, list) {
2249 		if (!cur->already_sent)
2250 			break;
2251 		cur->already_sent = 0;
2252 	}
2253 
2254 	return true;
2255 }
2256 
2257 /* flags for __mptcp_close_ssk() */
2258 #define MPTCP_CF_PUSH		BIT(1)
2259 #define MPTCP_CF_FASTCLOSE	BIT(2)
2260 
2261 /* subflow sockets can be either outgoing (connect) or incoming
2262  * (accept).
2263  *
2264  * Outgoing subflows use in-kernel sockets.
2265  * Incoming subflows do not have their own 'struct socket' allocated,
2266  * so we need to use tcp_close() after detaching them from the mptcp
2267  * parent socket.
2268  */
2269 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2270 			      struct mptcp_subflow_context *subflow,
2271 			      unsigned int flags)
2272 {
2273 	struct mptcp_sock *msk = mptcp_sk(sk);
2274 	bool need_push, dispose_it;
2275 
2276 	dispose_it = !msk->subflow || ssk != msk->subflow->sk;
2277 	if (dispose_it)
2278 		list_del(&subflow->node);
2279 
2280 	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2281 
2282 	if (flags & MPTCP_CF_FASTCLOSE)
2283 		subflow->send_fastclose = 1;
2284 
2285 	need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2286 	if (!dispose_it) {
2287 		tcp_disconnect(ssk, 0);
2288 		msk->subflow->state = SS_UNCONNECTED;
2289 		mptcp_subflow_ctx_reset(subflow);
2290 		release_sock(ssk);
2291 
2292 		goto out;
2293 	}
2294 
2295 	/* if we are invoked by the msk cleanup code, the subflow is
2296 	 * already orphaned
2297 	 */
2298 	if (ssk->sk_socket)
2299 		sock_orphan(ssk);
2300 
2301 	subflow->disposable = 1;
2302 
2303 	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2304 	 * the ssk has been already destroyed, we just need to release the
2305 	 * reference owned by msk;
2306 	 */
2307 	if (!inet_csk(ssk)->icsk_ulp_ops) {
2308 		kfree_rcu(subflow, rcu);
2309 	} else {
2310 		/* otherwise tcp will dispose of the ssk and subflow ctx */
2311 		if (ssk->sk_state == TCP_LISTEN) {
2312 			tcp_set_state(ssk, TCP_CLOSE);
2313 			mptcp_subflow_queue_clean(ssk);
2314 			inet_csk_listen_stop(ssk);
2315 		}
2316 		__tcp_close(ssk, 0);
2317 
2318 		/* close acquired an extra ref */
2319 		__sock_put(ssk);
2320 	}
2321 	release_sock(ssk);
2322 
2323 	sock_put(ssk);
2324 
2325 	if (ssk == msk->first)
2326 		msk->first = NULL;
2327 
2328 out:
2329 	if (ssk == msk->last_snd)
2330 		msk->last_snd = NULL;
2331 
2332 	if (need_push)
2333 		__mptcp_push_pending(sk, 0);
2334 }
2335 
2336 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2337 		     struct mptcp_subflow_context *subflow)
2338 {
2339 	if (sk->sk_state == TCP_ESTABLISHED)
2340 		mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2341 
2342 	/* subflow aborted before reaching the fully_established status
2343 	 * attempt the creation of the next subflow
2344 	 */
2345 	mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2346 
2347 	__mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2348 }
2349 
2350 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2351 {
2352 	return 0;
2353 }
2354 
2355 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2356 {
2357 	struct mptcp_subflow_context *subflow, *tmp;
2358 
2359 	might_sleep();
2360 
2361 	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2362 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2363 
2364 		if (inet_sk_state_load(ssk) != TCP_CLOSE)
2365 			continue;
2366 
2367 		/* 'subflow_data_ready' will re-sched once rx queue is empty */
2368 		if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2369 			continue;
2370 
2371 		mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2372 	}
2373 }
2374 
2375 static bool mptcp_check_close_timeout(const struct sock *sk)
2376 {
2377 	s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2378 	struct mptcp_subflow_context *subflow;
2379 
2380 	if (delta >= TCP_TIMEWAIT_LEN)
2381 		return true;
2382 
2383 	/* if all subflows are in closed status don't bother with additional
2384 	 * timeout
2385 	 */
2386 	mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2387 		if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2388 		    TCP_CLOSE)
2389 			return false;
2390 	}
2391 	return true;
2392 }
2393 
2394 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2395 {
2396 	struct mptcp_subflow_context *subflow, *tmp;
2397 	struct sock *sk = &msk->sk.icsk_inet.sk;
2398 
2399 	if (likely(!READ_ONCE(msk->rcv_fastclose)))
2400 		return;
2401 
2402 	mptcp_token_destroy(msk);
2403 
2404 	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2405 		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2406 		bool slow;
2407 
2408 		slow = lock_sock_fast(tcp_sk);
2409 		if (tcp_sk->sk_state != TCP_CLOSE) {
2410 			tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2411 			tcp_set_state(tcp_sk, TCP_CLOSE);
2412 		}
2413 		unlock_sock_fast(tcp_sk, slow);
2414 	}
2415 
2416 	inet_sk_state_store(sk, TCP_CLOSE);
2417 	sk->sk_shutdown = SHUTDOWN_MASK;
2418 	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2419 	set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2420 
2421 	mptcp_close_wake_up(sk);
2422 }
2423 
2424 static void __mptcp_retrans(struct sock *sk)
2425 {
2426 	struct mptcp_sock *msk = mptcp_sk(sk);
2427 	struct mptcp_sendmsg_info info = {};
2428 	struct mptcp_data_frag *dfrag;
2429 	size_t copied = 0;
2430 	struct sock *ssk;
2431 	int ret;
2432 
2433 	mptcp_clean_una_wakeup(sk);
2434 
2435 	/* first check ssk: need to kick "stale" logic */
2436 	ssk = mptcp_subflow_get_retrans(msk);
2437 	dfrag = mptcp_rtx_head(sk);
2438 	if (!dfrag) {
2439 		if (mptcp_data_fin_enabled(msk)) {
2440 			struct inet_connection_sock *icsk = inet_csk(sk);
2441 
2442 			icsk->icsk_retransmits++;
2443 			mptcp_set_datafin_timeout(sk);
2444 			mptcp_send_ack(msk);
2445 
2446 			goto reset_timer;
2447 		}
2448 
2449 		if (!mptcp_send_head(sk))
2450 			return;
2451 
2452 		goto reset_timer;
2453 	}
2454 
2455 	if (!ssk)
2456 		goto reset_timer;
2457 
2458 	lock_sock(ssk);
2459 
2460 	/* limit retransmission to the bytes already sent on some subflows */
2461 	info.sent = 0;
2462 	info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2463 	while (info.sent < info.limit) {
2464 		ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2465 		if (ret <= 0)
2466 			break;
2467 
2468 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2469 		copied += ret;
2470 		info.sent += ret;
2471 	}
2472 	if (copied) {
2473 		dfrag->already_sent = max(dfrag->already_sent, info.sent);
2474 		tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2475 			 info.size_goal);
2476 		WRITE_ONCE(msk->allow_infinite_fallback, false);
2477 	}
2478 
2479 	release_sock(ssk);
2480 
2481 reset_timer:
2482 	mptcp_check_and_set_pending(sk);
2483 
2484 	if (!mptcp_timer_pending(sk))
2485 		mptcp_reset_timer(sk);
2486 }
2487 
2488 /* schedule the timeout timer for the relevant event: either close timeout
2489  * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2490  */
2491 void mptcp_reset_timeout(struct mptcp_sock *msk, unsigned long fail_tout)
2492 {
2493 	struct sock *sk = (struct sock *)msk;
2494 	unsigned long timeout, close_timeout;
2495 
2496 	if (!fail_tout && !sock_flag(sk, SOCK_DEAD))
2497 		return;
2498 
2499 	close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies + TCP_TIMEWAIT_LEN;
2500 
2501 	/* the close timeout takes precedence on the fail one, and here at least one of
2502 	 * them is active
2503 	 */
2504 	timeout = sock_flag(sk, SOCK_DEAD) ? close_timeout : fail_tout;
2505 
2506 	sk_reset_timer(sk, &sk->sk_timer, timeout);
2507 }
2508 
2509 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2510 {
2511 	struct sock *ssk = msk->first;
2512 	bool slow;
2513 
2514 	if (!ssk)
2515 		return;
2516 
2517 	pr_debug("MP_FAIL doesn't respond, reset the subflow");
2518 
2519 	slow = lock_sock_fast(ssk);
2520 	mptcp_subflow_reset(ssk);
2521 	WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2522 	unlock_sock_fast(ssk, slow);
2523 
2524 	mptcp_reset_timeout(msk, 0);
2525 }
2526 
2527 static void mptcp_worker(struct work_struct *work)
2528 {
2529 	struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2530 	struct sock *sk = &msk->sk.icsk_inet.sk;
2531 	unsigned long fail_tout;
2532 	int state;
2533 
2534 	lock_sock(sk);
2535 	state = sk->sk_state;
2536 	if (unlikely(state == TCP_CLOSE))
2537 		goto unlock;
2538 
2539 	mptcp_check_data_fin_ack(sk);
2540 
2541 	mptcp_check_fastclose(msk);
2542 
2543 	mptcp_pm_nl_work(msk);
2544 
2545 	if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2546 		mptcp_check_for_eof(msk);
2547 
2548 	__mptcp_check_send_data_fin(sk);
2549 	mptcp_check_data_fin(sk);
2550 
2551 	/* There is no point in keeping around an orphaned sk timedout or
2552 	 * closed, but we need the msk around to reply to incoming DATA_FIN,
2553 	 * even if it is orphaned and in FIN_WAIT2 state
2554 	 */
2555 	if (sock_flag(sk, SOCK_DEAD) &&
2556 	    (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2557 		inet_sk_state_store(sk, TCP_CLOSE);
2558 		__mptcp_destroy_sock(sk);
2559 		goto unlock;
2560 	}
2561 
2562 	if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2563 		__mptcp_close_subflow(msk);
2564 
2565 	if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2566 		__mptcp_retrans(sk);
2567 
2568 	fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2569 	if (fail_tout && time_after(jiffies, fail_tout))
2570 		mptcp_mp_fail_no_response(msk);
2571 
2572 unlock:
2573 	release_sock(sk);
2574 	sock_put(sk);
2575 }
2576 
2577 static int __mptcp_init_sock(struct sock *sk)
2578 {
2579 	struct mptcp_sock *msk = mptcp_sk(sk);
2580 
2581 	INIT_LIST_HEAD(&msk->conn_list);
2582 	INIT_LIST_HEAD(&msk->join_list);
2583 	INIT_LIST_HEAD(&msk->rtx_queue);
2584 	INIT_WORK(&msk->work, mptcp_worker);
2585 	__skb_queue_head_init(&msk->receive_queue);
2586 	msk->out_of_order_queue = RB_ROOT;
2587 	msk->first_pending = NULL;
2588 	msk->rmem_fwd_alloc = 0;
2589 	WRITE_ONCE(msk->rmem_released, 0);
2590 	msk->timer_ival = TCP_RTO_MIN;
2591 
2592 	msk->first = NULL;
2593 	inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2594 	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2595 	WRITE_ONCE(msk->allow_infinite_fallback, true);
2596 	msk->recovery = false;
2597 
2598 	mptcp_pm_data_init(msk);
2599 
2600 	/* re-use the csk retrans timer for MPTCP-level retrans */
2601 	timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2602 	timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2603 
2604 	return 0;
2605 }
2606 
2607 static void mptcp_ca_reset(struct sock *sk)
2608 {
2609 	struct inet_connection_sock *icsk = inet_csk(sk);
2610 
2611 	tcp_assign_congestion_control(sk);
2612 	strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2613 
2614 	/* no need to keep a reference to the ops, the name will suffice */
2615 	tcp_cleanup_congestion_control(sk);
2616 	icsk->icsk_ca_ops = NULL;
2617 }
2618 
2619 static int mptcp_init_sock(struct sock *sk)
2620 {
2621 	struct net *net = sock_net(sk);
2622 	int ret;
2623 
2624 	ret = __mptcp_init_sock(sk);
2625 	if (ret)
2626 		return ret;
2627 
2628 	if (!mptcp_is_enabled(net))
2629 		return -ENOPROTOOPT;
2630 
2631 	if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2632 		return -ENOMEM;
2633 
2634 	ret = __mptcp_socket_create(mptcp_sk(sk));
2635 	if (ret)
2636 		return ret;
2637 
2638 	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2639 	 * propagate the correct value
2640 	 */
2641 	mptcp_ca_reset(sk);
2642 
2643 	sk_sockets_allocated_inc(sk);
2644 	sk->sk_rcvbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
2645 	sk->sk_sndbuf = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
2646 
2647 	return 0;
2648 }
2649 
2650 static void __mptcp_clear_xmit(struct sock *sk)
2651 {
2652 	struct mptcp_sock *msk = mptcp_sk(sk);
2653 	struct mptcp_data_frag *dtmp, *dfrag;
2654 
2655 	WRITE_ONCE(msk->first_pending, NULL);
2656 	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2657 		dfrag_clear(sk, dfrag);
2658 }
2659 
2660 static void mptcp_cancel_work(struct sock *sk)
2661 {
2662 	struct mptcp_sock *msk = mptcp_sk(sk);
2663 
2664 	if (cancel_work_sync(&msk->work))
2665 		__sock_put(sk);
2666 }
2667 
2668 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2669 {
2670 	lock_sock(ssk);
2671 
2672 	switch (ssk->sk_state) {
2673 	case TCP_LISTEN:
2674 		if (!(how & RCV_SHUTDOWN))
2675 			break;
2676 		fallthrough;
2677 	case TCP_SYN_SENT:
2678 		tcp_disconnect(ssk, O_NONBLOCK);
2679 		break;
2680 	default:
2681 		if (__mptcp_check_fallback(mptcp_sk(sk))) {
2682 			pr_debug("Fallback");
2683 			ssk->sk_shutdown |= how;
2684 			tcp_shutdown(ssk, how);
2685 		} else {
2686 			pr_debug("Sending DATA_FIN on subflow %p", ssk);
2687 			tcp_send_ack(ssk);
2688 			if (!mptcp_timer_pending(sk))
2689 				mptcp_reset_timer(sk);
2690 		}
2691 		break;
2692 	}
2693 
2694 	release_sock(ssk);
2695 }
2696 
2697 static const unsigned char new_state[16] = {
2698 	/* current state:     new state:      action:	*/
2699 	[0 /* (Invalid) */] = TCP_CLOSE,
2700 	[TCP_ESTABLISHED]   = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2701 	[TCP_SYN_SENT]      = TCP_CLOSE,
2702 	[TCP_SYN_RECV]      = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2703 	[TCP_FIN_WAIT1]     = TCP_FIN_WAIT1,
2704 	[TCP_FIN_WAIT2]     = TCP_FIN_WAIT2,
2705 	[TCP_TIME_WAIT]     = TCP_CLOSE,	/* should not happen ! */
2706 	[TCP_CLOSE]         = TCP_CLOSE,
2707 	[TCP_CLOSE_WAIT]    = TCP_LAST_ACK  | TCP_ACTION_FIN,
2708 	[TCP_LAST_ACK]      = TCP_LAST_ACK,
2709 	[TCP_LISTEN]        = TCP_CLOSE,
2710 	[TCP_CLOSING]       = TCP_CLOSING,
2711 	[TCP_NEW_SYN_RECV]  = TCP_CLOSE,	/* should not happen ! */
2712 };
2713 
2714 static int mptcp_close_state(struct sock *sk)
2715 {
2716 	int next = (int)new_state[sk->sk_state];
2717 	int ns = next & TCP_STATE_MASK;
2718 
2719 	inet_sk_state_store(sk, ns);
2720 
2721 	return next & TCP_ACTION_FIN;
2722 }
2723 
2724 static void __mptcp_check_send_data_fin(struct sock *sk)
2725 {
2726 	struct mptcp_subflow_context *subflow;
2727 	struct mptcp_sock *msk = mptcp_sk(sk);
2728 
2729 	pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2730 		 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2731 		 msk->snd_nxt, msk->write_seq);
2732 
2733 	/* we still need to enqueue subflows or not really shutting down,
2734 	 * skip this
2735 	 */
2736 	if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2737 	    mptcp_send_head(sk))
2738 		return;
2739 
2740 	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2741 
2742 	/* fallback socket will not get data_fin/ack, can move to the next
2743 	 * state now
2744 	 */
2745 	if (__mptcp_check_fallback(msk)) {
2746 		WRITE_ONCE(msk->snd_una, msk->write_seq);
2747 		if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2748 			inet_sk_state_store(sk, TCP_CLOSE);
2749 			mptcp_close_wake_up(sk);
2750 		} else if (sk->sk_state == TCP_FIN_WAIT1) {
2751 			inet_sk_state_store(sk, TCP_FIN_WAIT2);
2752 		}
2753 	}
2754 
2755 	mptcp_for_each_subflow(msk, subflow) {
2756 		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2757 
2758 		mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2759 	}
2760 }
2761 
2762 static void __mptcp_wr_shutdown(struct sock *sk)
2763 {
2764 	struct mptcp_sock *msk = mptcp_sk(sk);
2765 
2766 	pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2767 		 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2768 		 !!mptcp_send_head(sk));
2769 
2770 	/* will be ignored by fallback sockets */
2771 	WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2772 	WRITE_ONCE(msk->snd_data_fin_enable, 1);
2773 
2774 	__mptcp_check_send_data_fin(sk);
2775 }
2776 
2777 static void __mptcp_destroy_sock(struct sock *sk)
2778 {
2779 	struct mptcp_sock *msk = mptcp_sk(sk);
2780 
2781 	pr_debug("msk=%p", msk);
2782 
2783 	might_sleep();
2784 
2785 	mptcp_stop_timer(sk);
2786 	sk_stop_timer(sk, &sk->sk_timer);
2787 	msk->pm.status = 0;
2788 
2789 	sk->sk_prot->destroy(sk);
2790 
2791 	WARN_ON_ONCE(msk->rmem_fwd_alloc);
2792 	WARN_ON_ONCE(msk->rmem_released);
2793 	sk_stream_kill_queues(sk);
2794 	xfrm_sk_free_policy(sk);
2795 
2796 	sk_refcnt_debug_release(sk);
2797 	sock_put(sk);
2798 }
2799 
2800 static void mptcp_close(struct sock *sk, long timeout)
2801 {
2802 	struct mptcp_subflow_context *subflow;
2803 	struct mptcp_sock *msk = mptcp_sk(sk);
2804 	bool do_cancel_work = false;
2805 
2806 	lock_sock(sk);
2807 	sk->sk_shutdown = SHUTDOWN_MASK;
2808 
2809 	if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2810 		inet_sk_state_store(sk, TCP_CLOSE);
2811 		goto cleanup;
2812 	}
2813 
2814 	if (mptcp_close_state(sk))
2815 		__mptcp_wr_shutdown(sk);
2816 
2817 	sk_stream_wait_close(sk, timeout);
2818 
2819 cleanup:
2820 	/* orphan all the subflows */
2821 	inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2822 	mptcp_for_each_subflow(msk, subflow) {
2823 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2824 		bool slow = lock_sock_fast_nested(ssk);
2825 
2826 		/* since the close timeout takes precedence on the fail one,
2827 		 * cancel the latter
2828 		 */
2829 		if (ssk == msk->first)
2830 			subflow->fail_tout = 0;
2831 
2832 		sock_orphan(ssk);
2833 		unlock_sock_fast(ssk, slow);
2834 	}
2835 	sock_orphan(sk);
2836 
2837 	sock_hold(sk);
2838 	pr_debug("msk=%p state=%d", sk, sk->sk_state);
2839 	if (mptcp_sk(sk)->token)
2840 		mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
2841 
2842 	if (sk->sk_state == TCP_CLOSE) {
2843 		__mptcp_destroy_sock(sk);
2844 		do_cancel_work = true;
2845 	} else {
2846 		mptcp_reset_timeout(msk, 0);
2847 	}
2848 	release_sock(sk);
2849 	if (do_cancel_work)
2850 		mptcp_cancel_work(sk);
2851 
2852 	sock_put(sk);
2853 }
2854 
2855 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2856 {
2857 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2858 	const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2859 	struct ipv6_pinfo *msk6 = inet6_sk(msk);
2860 
2861 	msk->sk_v6_daddr = ssk->sk_v6_daddr;
2862 	msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2863 
2864 	if (msk6 && ssk6) {
2865 		msk6->saddr = ssk6->saddr;
2866 		msk6->flow_label = ssk6->flow_label;
2867 	}
2868 #endif
2869 
2870 	inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2871 	inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2872 	inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2873 	inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2874 	inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2875 	inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2876 }
2877 
2878 static int mptcp_disconnect(struct sock *sk, int flags)
2879 {
2880 	struct mptcp_sock *msk = mptcp_sk(sk);
2881 
2882 	inet_sk_state_store(sk, TCP_CLOSE);
2883 
2884 	mptcp_stop_timer(sk);
2885 	sk_stop_timer(sk, &sk->sk_timer);
2886 
2887 	if (mptcp_sk(sk)->token)
2888 		mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2889 
2890 	/* msk->subflow is still intact, the following will not free the first
2891 	 * subflow
2892 	 */
2893 	mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
2894 	msk->last_snd = NULL;
2895 	WRITE_ONCE(msk->flags, 0);
2896 	msk->cb_flags = 0;
2897 	msk->push_pending = 0;
2898 	msk->recovery = false;
2899 	msk->can_ack = false;
2900 	msk->fully_established = false;
2901 	msk->rcv_data_fin = false;
2902 	msk->snd_data_fin_enable = false;
2903 	msk->rcv_fastclose = false;
2904 	msk->use_64bit_ack = false;
2905 	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2906 	mptcp_pm_data_reset(msk);
2907 	mptcp_ca_reset(sk);
2908 
2909 	sk->sk_shutdown = 0;
2910 	sk_error_report(sk);
2911 	return 0;
2912 }
2913 
2914 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2915 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2916 {
2917 	unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2918 
2919 	return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2920 }
2921 #endif
2922 
2923 struct sock *mptcp_sk_clone(const struct sock *sk,
2924 			    const struct mptcp_options_received *mp_opt,
2925 			    struct request_sock *req)
2926 {
2927 	struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2928 	struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2929 	struct mptcp_sock *msk;
2930 	u64 ack_seq;
2931 
2932 	if (!nsk)
2933 		return NULL;
2934 
2935 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2936 	if (nsk->sk_family == AF_INET6)
2937 		inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2938 #endif
2939 
2940 	__mptcp_init_sock(nsk);
2941 
2942 	msk = mptcp_sk(nsk);
2943 	msk->local_key = subflow_req->local_key;
2944 	msk->token = subflow_req->token;
2945 	msk->subflow = NULL;
2946 	WRITE_ONCE(msk->fully_established, false);
2947 	if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
2948 		WRITE_ONCE(msk->csum_enabled, true);
2949 
2950 	msk->write_seq = subflow_req->idsn + 1;
2951 	msk->snd_nxt = msk->write_seq;
2952 	msk->snd_una = msk->write_seq;
2953 	msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2954 	msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2955 
2956 	if (mp_opt->suboptions & OPTIONS_MPTCP_MPC) {
2957 		msk->can_ack = true;
2958 		msk->remote_key = mp_opt->sndr_key;
2959 		mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2960 		ack_seq++;
2961 		WRITE_ONCE(msk->ack_seq, ack_seq);
2962 		atomic64_set(&msk->rcv_wnd_sent, ack_seq);
2963 	}
2964 
2965 	sock_reset_flag(nsk, SOCK_RCU_FREE);
2966 	/* will be fully established after successful MPC subflow creation */
2967 	inet_sk_state_store(nsk, TCP_SYN_RECV);
2968 
2969 	security_inet_csk_clone(nsk, req);
2970 	bh_unlock_sock(nsk);
2971 
2972 	/* keep a single reference */
2973 	__sock_put(nsk);
2974 	return nsk;
2975 }
2976 
2977 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2978 {
2979 	const struct tcp_sock *tp = tcp_sk(ssk);
2980 
2981 	msk->rcvq_space.copied = 0;
2982 	msk->rcvq_space.rtt_us = 0;
2983 
2984 	msk->rcvq_space.time = tp->tcp_mstamp;
2985 
2986 	/* initial rcv_space offering made to peer */
2987 	msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2988 				      TCP_INIT_CWND * tp->advmss);
2989 	if (msk->rcvq_space.space == 0)
2990 		msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2991 
2992 	WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2993 }
2994 
2995 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2996 				 bool kern)
2997 {
2998 	struct mptcp_sock *msk = mptcp_sk(sk);
2999 	struct socket *listener;
3000 	struct sock *newsk;
3001 
3002 	listener = __mptcp_nmpc_socket(msk);
3003 	if (WARN_ON_ONCE(!listener)) {
3004 		*err = -EINVAL;
3005 		return NULL;
3006 	}
3007 
3008 	pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
3009 	newsk = inet_csk_accept(listener->sk, flags, err, kern);
3010 	if (!newsk)
3011 		return NULL;
3012 
3013 	pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
3014 	if (sk_is_mptcp(newsk)) {
3015 		struct mptcp_subflow_context *subflow;
3016 		struct sock *new_mptcp_sock;
3017 
3018 		subflow = mptcp_subflow_ctx(newsk);
3019 		new_mptcp_sock = subflow->conn;
3020 
3021 		/* is_mptcp should be false if subflow->conn is missing, see
3022 		 * subflow_syn_recv_sock()
3023 		 */
3024 		if (WARN_ON_ONCE(!new_mptcp_sock)) {
3025 			tcp_sk(newsk)->is_mptcp = 0;
3026 			goto out;
3027 		}
3028 
3029 		/* acquire the 2nd reference for the owning socket */
3030 		sock_hold(new_mptcp_sock);
3031 		newsk = new_mptcp_sock;
3032 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3033 	} else {
3034 		MPTCP_INC_STATS(sock_net(sk),
3035 				MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3036 	}
3037 
3038 out:
3039 	newsk->sk_kern_sock = kern;
3040 	return newsk;
3041 }
3042 
3043 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3044 {
3045 	struct mptcp_subflow_context *subflow, *tmp;
3046 	struct sock *sk = (struct sock *)msk;
3047 
3048 	__mptcp_clear_xmit(sk);
3049 
3050 	/* join list will be eventually flushed (with rst) at sock lock release time */
3051 	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3052 		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3053 
3054 	/* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3055 	mptcp_data_lock(sk);
3056 	skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3057 	__skb_queue_purge(&sk->sk_receive_queue);
3058 	skb_rbtree_purge(&msk->out_of_order_queue);
3059 	mptcp_data_unlock(sk);
3060 
3061 	/* move all the rx fwd alloc into the sk_mem_reclaim_final in
3062 	 * inet_sock_destruct() will dispose it
3063 	 */
3064 	sk->sk_forward_alloc += msk->rmem_fwd_alloc;
3065 	msk->rmem_fwd_alloc = 0;
3066 	mptcp_token_destroy(msk);
3067 	mptcp_pm_free_anno_list(msk);
3068 	mptcp_free_local_addr_list(msk);
3069 }
3070 
3071 static void mptcp_destroy(struct sock *sk)
3072 {
3073 	struct mptcp_sock *msk = mptcp_sk(sk);
3074 
3075 	/* clears msk->subflow, allowing the following to close
3076 	 * even the initial subflow
3077 	 */
3078 	mptcp_dispose_initial_subflow(msk);
3079 	mptcp_destroy_common(msk, 0);
3080 	sk_sockets_allocated_dec(sk);
3081 }
3082 
3083 void __mptcp_data_acked(struct sock *sk)
3084 {
3085 	if (!sock_owned_by_user(sk))
3086 		__mptcp_clean_una(sk);
3087 	else
3088 		__set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3089 
3090 	if (mptcp_pending_data_fin_ack(sk))
3091 		mptcp_schedule_work(sk);
3092 }
3093 
3094 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3095 {
3096 	if (!mptcp_send_head(sk))
3097 		return;
3098 
3099 	if (!sock_owned_by_user(sk)) {
3100 		struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
3101 
3102 		if (xmit_ssk == ssk)
3103 			__mptcp_subflow_push_pending(sk, ssk);
3104 		else if (xmit_ssk)
3105 			mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk), MPTCP_DELEGATE_SEND);
3106 	} else {
3107 		__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3108 	}
3109 }
3110 
3111 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3112 				      BIT(MPTCP_RETRANSMIT) | \
3113 				      BIT(MPTCP_FLUSH_JOIN_LIST))
3114 
3115 /* processes deferred events and flush wmem */
3116 static void mptcp_release_cb(struct sock *sk)
3117 	__must_hold(&sk->sk_lock.slock)
3118 {
3119 	struct mptcp_sock *msk = mptcp_sk(sk);
3120 
3121 	for (;;) {
3122 		unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED) |
3123 				      msk->push_pending;
3124 		if (!flags)
3125 			break;
3126 
3127 		/* the following actions acquire the subflow socket lock
3128 		 *
3129 		 * 1) can't be invoked in atomic scope
3130 		 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3131 		 *    datapath acquires the msk socket spinlock while helding
3132 		 *    the subflow socket lock
3133 		 */
3134 		msk->push_pending = 0;
3135 		msk->cb_flags &= ~flags;
3136 		spin_unlock_bh(&sk->sk_lock.slock);
3137 		if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3138 			__mptcp_flush_join_list(sk);
3139 		if (flags & BIT(MPTCP_PUSH_PENDING))
3140 			__mptcp_push_pending(sk, 0);
3141 		if (flags & BIT(MPTCP_RETRANSMIT))
3142 			__mptcp_retrans(sk);
3143 
3144 		cond_resched();
3145 		spin_lock_bh(&sk->sk_lock.slock);
3146 	}
3147 
3148 	if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3149 		__mptcp_clean_una_wakeup(sk);
3150 	if (unlikely(&msk->cb_flags)) {
3151 		/* be sure to set the current sk state before tacking actions
3152 		 * depending on sk_state, that is processing MPTCP_ERROR_REPORT
3153 		 */
3154 		if (__test_and_clear_bit(MPTCP_CONNECTED, &msk->cb_flags))
3155 			__mptcp_set_connected(sk);
3156 		if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3157 			__mptcp_error_report(sk);
3158 		if (__test_and_clear_bit(MPTCP_RESET_SCHEDULER, &msk->cb_flags))
3159 			msk->last_snd = NULL;
3160 	}
3161 
3162 	__mptcp_update_rmem(sk);
3163 }
3164 
3165 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3166  * TCP can't schedule delack timer before the subflow is fully established.
3167  * MPTCP uses the delack timer to do 3rd ack retransmissions
3168  */
3169 static void schedule_3rdack_retransmission(struct sock *ssk)
3170 {
3171 	struct inet_connection_sock *icsk = inet_csk(ssk);
3172 	struct tcp_sock *tp = tcp_sk(ssk);
3173 	unsigned long timeout;
3174 
3175 	if (mptcp_subflow_ctx(ssk)->fully_established)
3176 		return;
3177 
3178 	/* reschedule with a timeout above RTT, as we must look only for drop */
3179 	if (tp->srtt_us)
3180 		timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3181 	else
3182 		timeout = TCP_TIMEOUT_INIT;
3183 	timeout += jiffies;
3184 
3185 	WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3186 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3187 	icsk->icsk_ack.timeout = timeout;
3188 	sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3189 }
3190 
3191 void mptcp_subflow_process_delegated(struct sock *ssk)
3192 {
3193 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3194 	struct sock *sk = subflow->conn;
3195 
3196 	if (test_bit(MPTCP_DELEGATE_SEND, &subflow->delegated_status)) {
3197 		mptcp_data_lock(sk);
3198 		if (!sock_owned_by_user(sk))
3199 			__mptcp_subflow_push_pending(sk, ssk);
3200 		else
3201 			__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3202 		mptcp_data_unlock(sk);
3203 		mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_SEND);
3204 	}
3205 	if (test_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status)) {
3206 		schedule_3rdack_retransmission(ssk);
3207 		mptcp_subflow_delegated_done(subflow, MPTCP_DELEGATE_ACK);
3208 	}
3209 }
3210 
3211 static int mptcp_hash(struct sock *sk)
3212 {
3213 	/* should never be called,
3214 	 * we hash the TCP subflows not the master socket
3215 	 */
3216 	WARN_ON_ONCE(1);
3217 	return 0;
3218 }
3219 
3220 static void mptcp_unhash(struct sock *sk)
3221 {
3222 	/* called from sk_common_release(), but nothing to do here */
3223 }
3224 
3225 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3226 {
3227 	struct mptcp_sock *msk = mptcp_sk(sk);
3228 	struct socket *ssock;
3229 
3230 	ssock = __mptcp_nmpc_socket(msk);
3231 	pr_debug("msk=%p, subflow=%p", msk, ssock);
3232 	if (WARN_ON_ONCE(!ssock))
3233 		return -EINVAL;
3234 
3235 	return inet_csk_get_port(ssock->sk, snum);
3236 }
3237 
3238 void mptcp_finish_connect(struct sock *ssk)
3239 {
3240 	struct mptcp_subflow_context *subflow;
3241 	struct mptcp_sock *msk;
3242 	struct sock *sk;
3243 	u64 ack_seq;
3244 
3245 	subflow = mptcp_subflow_ctx(ssk);
3246 	sk = subflow->conn;
3247 	msk = mptcp_sk(sk);
3248 
3249 	pr_debug("msk=%p, token=%u", sk, subflow->token);
3250 
3251 	mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3252 	ack_seq++;
3253 	subflow->map_seq = ack_seq;
3254 	subflow->map_subflow_seq = 1;
3255 
3256 	/* the socket is not connected yet, no msk/subflow ops can access/race
3257 	 * accessing the field below
3258 	 */
3259 	WRITE_ONCE(msk->remote_key, subflow->remote_key);
3260 	WRITE_ONCE(msk->local_key, subflow->local_key);
3261 	WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3262 	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3263 	WRITE_ONCE(msk->ack_seq, ack_seq);
3264 	WRITE_ONCE(msk->can_ack, 1);
3265 	WRITE_ONCE(msk->snd_una, msk->write_seq);
3266 	atomic64_set(&msk->rcv_wnd_sent, ack_seq);
3267 
3268 	mptcp_pm_new_connection(msk, ssk, 0);
3269 
3270 	mptcp_rcv_space_init(msk, ssk);
3271 }
3272 
3273 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3274 {
3275 	write_lock_bh(&sk->sk_callback_lock);
3276 	rcu_assign_pointer(sk->sk_wq, &parent->wq);
3277 	sk_set_socket(sk, parent);
3278 	sk->sk_uid = SOCK_INODE(parent)->i_uid;
3279 	write_unlock_bh(&sk->sk_callback_lock);
3280 }
3281 
3282 bool mptcp_finish_join(struct sock *ssk)
3283 {
3284 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3285 	struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3286 	struct sock *parent = (void *)msk;
3287 	bool ret = true;
3288 
3289 	pr_debug("msk=%p, subflow=%p", msk, subflow);
3290 
3291 	/* mptcp socket already closing? */
3292 	if (!mptcp_is_fully_established(parent)) {
3293 		subflow->reset_reason = MPTCP_RST_EMPTCP;
3294 		return false;
3295 	}
3296 
3297 	if (!list_empty(&subflow->node))
3298 		goto out;
3299 
3300 	if (!mptcp_pm_allow_new_subflow(msk))
3301 		goto err_prohibited;
3302 
3303 	/* active connections are already on conn_list.
3304 	 * If we can't acquire msk socket lock here, let the release callback
3305 	 * handle it
3306 	 */
3307 	mptcp_data_lock(parent);
3308 	if (!sock_owned_by_user(parent)) {
3309 		ret = __mptcp_finish_join(msk, ssk);
3310 		if (ret) {
3311 			sock_hold(ssk);
3312 			list_add_tail(&subflow->node, &msk->conn_list);
3313 		}
3314 	} else {
3315 		sock_hold(ssk);
3316 		list_add_tail(&subflow->node, &msk->join_list);
3317 		__set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3318 	}
3319 	mptcp_data_unlock(parent);
3320 
3321 	if (!ret) {
3322 err_prohibited:
3323 		subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3324 		return false;
3325 	}
3326 
3327 	subflow->map_seq = READ_ONCE(msk->ack_seq);
3328 	WRITE_ONCE(msk->allow_infinite_fallback, false);
3329 
3330 out:
3331 	mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3332 	return true;
3333 }
3334 
3335 static void mptcp_shutdown(struct sock *sk, int how)
3336 {
3337 	pr_debug("sk=%p, how=%d", sk, how);
3338 
3339 	if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3340 		__mptcp_wr_shutdown(sk);
3341 }
3342 
3343 static int mptcp_forward_alloc_get(const struct sock *sk)
3344 {
3345 	return sk->sk_forward_alloc + mptcp_sk(sk)->rmem_fwd_alloc;
3346 }
3347 
3348 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3349 {
3350 	const struct sock *sk = (void *)msk;
3351 	u64 delta;
3352 
3353 	if (sk->sk_state == TCP_LISTEN)
3354 		return -EINVAL;
3355 
3356 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3357 		return 0;
3358 
3359 	delta = msk->write_seq - v;
3360 	if (__mptcp_check_fallback(msk) && msk->first) {
3361 		struct tcp_sock *tp = tcp_sk(msk->first);
3362 
3363 		/* the first subflow is disconnected after close - see
3364 		 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3365 		 * so ignore that status, too.
3366 		 */
3367 		if (!((1 << msk->first->sk_state) &
3368 		      (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3369 			delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3370 	}
3371 	if (delta > INT_MAX)
3372 		delta = INT_MAX;
3373 
3374 	return (int)delta;
3375 }
3376 
3377 static int mptcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
3378 {
3379 	struct mptcp_sock *msk = mptcp_sk(sk);
3380 	bool slow;
3381 	int answ;
3382 
3383 	switch (cmd) {
3384 	case SIOCINQ:
3385 		if (sk->sk_state == TCP_LISTEN)
3386 			return -EINVAL;
3387 
3388 		lock_sock(sk);
3389 		__mptcp_move_skbs(msk);
3390 		answ = mptcp_inq_hint(sk);
3391 		release_sock(sk);
3392 		break;
3393 	case SIOCOUTQ:
3394 		slow = lock_sock_fast(sk);
3395 		answ = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3396 		unlock_sock_fast(sk, slow);
3397 		break;
3398 	case SIOCOUTQNSD:
3399 		slow = lock_sock_fast(sk);
3400 		answ = mptcp_ioctl_outq(msk, msk->snd_nxt);
3401 		unlock_sock_fast(sk, slow);
3402 		break;
3403 	default:
3404 		return -ENOIOCTLCMD;
3405 	}
3406 
3407 	return put_user(answ, (int __user *)arg);
3408 }
3409 
3410 static struct proto mptcp_prot = {
3411 	.name		= "MPTCP",
3412 	.owner		= THIS_MODULE,
3413 	.init		= mptcp_init_sock,
3414 	.disconnect	= mptcp_disconnect,
3415 	.close		= mptcp_close,
3416 	.accept		= mptcp_accept,
3417 	.setsockopt	= mptcp_setsockopt,
3418 	.getsockopt	= mptcp_getsockopt,
3419 	.shutdown	= mptcp_shutdown,
3420 	.destroy	= mptcp_destroy,
3421 	.sendmsg	= mptcp_sendmsg,
3422 	.ioctl		= mptcp_ioctl,
3423 	.recvmsg	= mptcp_recvmsg,
3424 	.release_cb	= mptcp_release_cb,
3425 	.hash		= mptcp_hash,
3426 	.unhash		= mptcp_unhash,
3427 	.get_port	= mptcp_get_port,
3428 	.forward_alloc_get	= mptcp_forward_alloc_get,
3429 	.sockets_allocated	= &mptcp_sockets_allocated,
3430 
3431 	.memory_allocated	= &tcp_memory_allocated,
3432 	.per_cpu_fw_alloc	= &tcp_memory_per_cpu_fw_alloc,
3433 
3434 	.memory_pressure	= &tcp_memory_pressure,
3435 	.sysctl_wmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_wmem),
3436 	.sysctl_rmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_rmem),
3437 	.sysctl_mem	= sysctl_tcp_mem,
3438 	.obj_size	= sizeof(struct mptcp_sock),
3439 	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
3440 	.no_autobind	= true,
3441 };
3442 
3443 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3444 {
3445 	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3446 	struct socket *ssock;
3447 	int err;
3448 
3449 	lock_sock(sock->sk);
3450 	ssock = __mptcp_nmpc_socket(msk);
3451 	if (!ssock) {
3452 		err = -EINVAL;
3453 		goto unlock;
3454 	}
3455 
3456 	err = ssock->ops->bind(ssock, uaddr, addr_len);
3457 	if (!err)
3458 		mptcp_copy_inaddrs(sock->sk, ssock->sk);
3459 
3460 unlock:
3461 	release_sock(sock->sk);
3462 	return err;
3463 }
3464 
3465 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3466 					 struct mptcp_subflow_context *subflow)
3467 {
3468 	subflow->request_mptcp = 0;
3469 	__mptcp_do_fallback(msk);
3470 }
3471 
3472 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3473 				int addr_len, int flags)
3474 {
3475 	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3476 	struct mptcp_subflow_context *subflow;
3477 	struct socket *ssock;
3478 	int err = -EINVAL;
3479 
3480 	lock_sock(sock->sk);
3481 	if (uaddr) {
3482 		if (addr_len < sizeof(uaddr->sa_family))
3483 			goto unlock;
3484 
3485 		if (uaddr->sa_family == AF_UNSPEC) {
3486 			err = mptcp_disconnect(sock->sk, flags);
3487 			sock->state = err ? SS_DISCONNECTING : SS_UNCONNECTED;
3488 			goto unlock;
3489 		}
3490 	}
3491 
3492 	if (sock->state != SS_UNCONNECTED && msk->subflow) {
3493 		/* pending connection or invalid state, let existing subflow
3494 		 * cope with that
3495 		 */
3496 		ssock = msk->subflow;
3497 		goto do_connect;
3498 	}
3499 
3500 	ssock = __mptcp_nmpc_socket(msk);
3501 	if (!ssock)
3502 		goto unlock;
3503 
3504 	mptcp_token_destroy(msk);
3505 	inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3506 	subflow = mptcp_subflow_ctx(ssock->sk);
3507 #ifdef CONFIG_TCP_MD5SIG
3508 	/* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3509 	 * TCP option space.
3510 	 */
3511 	if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3512 		mptcp_subflow_early_fallback(msk, subflow);
3513 #endif
3514 	if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3515 		MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3516 		mptcp_subflow_early_fallback(msk, subflow);
3517 	}
3518 	if (likely(!__mptcp_check_fallback(msk)))
3519 		MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3520 
3521 do_connect:
3522 	err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3523 	sock->state = ssock->state;
3524 
3525 	/* on successful connect, the msk state will be moved to established by
3526 	 * subflow_finish_connect()
3527 	 */
3528 	if (!err || err == -EINPROGRESS)
3529 		mptcp_copy_inaddrs(sock->sk, ssock->sk);
3530 	else
3531 		inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3532 
3533 unlock:
3534 	release_sock(sock->sk);
3535 	return err;
3536 }
3537 
3538 static int mptcp_listen(struct socket *sock, int backlog)
3539 {
3540 	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3541 	struct socket *ssock;
3542 	int err;
3543 
3544 	pr_debug("msk=%p", msk);
3545 
3546 	lock_sock(sock->sk);
3547 	ssock = __mptcp_nmpc_socket(msk);
3548 	if (!ssock) {
3549 		err = -EINVAL;
3550 		goto unlock;
3551 	}
3552 
3553 	mptcp_token_destroy(msk);
3554 	inet_sk_state_store(sock->sk, TCP_LISTEN);
3555 	sock_set_flag(sock->sk, SOCK_RCU_FREE);
3556 
3557 	err = ssock->ops->listen(ssock, backlog);
3558 	inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3559 	if (!err)
3560 		mptcp_copy_inaddrs(sock->sk, ssock->sk);
3561 
3562 unlock:
3563 	release_sock(sock->sk);
3564 	return err;
3565 }
3566 
3567 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3568 			       int flags, bool kern)
3569 {
3570 	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3571 	struct socket *ssock;
3572 	int err;
3573 
3574 	pr_debug("msk=%p", msk);
3575 
3576 	ssock = __mptcp_nmpc_socket(msk);
3577 	if (!ssock)
3578 		return -EINVAL;
3579 
3580 	err = ssock->ops->accept(sock, newsock, flags, kern);
3581 	if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3582 		struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3583 		struct mptcp_subflow_context *subflow;
3584 		struct sock *newsk = newsock->sk;
3585 
3586 		lock_sock(newsk);
3587 
3588 		/* PM/worker can now acquire the first subflow socket
3589 		 * lock without racing with listener queue cleanup,
3590 		 * we can notify it, if needed.
3591 		 *
3592 		 * Even if remote has reset the initial subflow by now
3593 		 * the refcnt is still at least one.
3594 		 */
3595 		subflow = mptcp_subflow_ctx(msk->first);
3596 		list_add(&subflow->node, &msk->conn_list);
3597 		sock_hold(msk->first);
3598 		if (mptcp_is_fully_established(newsk))
3599 			mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3600 
3601 		mptcp_copy_inaddrs(newsk, msk->first);
3602 		mptcp_rcv_space_init(msk, msk->first);
3603 		mptcp_propagate_sndbuf(newsk, msk->first);
3604 
3605 		/* set ssk->sk_socket of accept()ed flows to mptcp socket.
3606 		 * This is needed so NOSPACE flag can be set from tcp stack.
3607 		 */
3608 		mptcp_for_each_subflow(msk, subflow) {
3609 			struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3610 
3611 			if (!ssk->sk_socket)
3612 				mptcp_sock_graft(ssk, newsock);
3613 		}
3614 		release_sock(newsk);
3615 	}
3616 
3617 	return err;
3618 }
3619 
3620 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3621 {
3622 	/* Concurrent splices from sk_receive_queue into receive_queue will
3623 	 * always show at least one non-empty queue when checked in this order.
3624 	 */
3625 	if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
3626 	    skb_queue_empty_lockless(&msk->receive_queue))
3627 		return 0;
3628 
3629 	return EPOLLIN | EPOLLRDNORM;
3630 }
3631 
3632 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3633 {
3634 	struct sock *sk = (struct sock *)msk;
3635 
3636 	if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3637 		return EPOLLOUT | EPOLLWRNORM;
3638 
3639 	if (sk_stream_is_writeable(sk))
3640 		return EPOLLOUT | EPOLLWRNORM;
3641 
3642 	mptcp_set_nospace(sk);
3643 	smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3644 	if (sk_stream_is_writeable(sk))
3645 		return EPOLLOUT | EPOLLWRNORM;
3646 
3647 	return 0;
3648 }
3649 
3650 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3651 			   struct poll_table_struct *wait)
3652 {
3653 	struct sock *sk = sock->sk;
3654 	struct mptcp_sock *msk;
3655 	__poll_t mask = 0;
3656 	int state;
3657 
3658 	msk = mptcp_sk(sk);
3659 	sock_poll_wait(file, sock, wait);
3660 
3661 	state = inet_sk_state_load(sk);
3662 	pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3663 	if (state == TCP_LISTEN) {
3664 		if (WARN_ON_ONCE(!msk->subflow || !msk->subflow->sk))
3665 			return 0;
3666 
3667 		return inet_csk_listen_poll(msk->subflow->sk);
3668 	}
3669 
3670 	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3671 		mask |= mptcp_check_readable(msk);
3672 		mask |= mptcp_check_writeable(msk);
3673 	}
3674 	if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3675 		mask |= EPOLLHUP;
3676 	if (sk->sk_shutdown & RCV_SHUTDOWN)
3677 		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3678 
3679 	/* This barrier is coupled with smp_wmb() in tcp_reset() */
3680 	smp_rmb();
3681 	if (sk->sk_err)
3682 		mask |= EPOLLERR;
3683 
3684 	return mask;
3685 }
3686 
3687 static const struct proto_ops mptcp_stream_ops = {
3688 	.family		   = PF_INET,
3689 	.owner		   = THIS_MODULE,
3690 	.release	   = inet_release,
3691 	.bind		   = mptcp_bind,
3692 	.connect	   = mptcp_stream_connect,
3693 	.socketpair	   = sock_no_socketpair,
3694 	.accept		   = mptcp_stream_accept,
3695 	.getname	   = inet_getname,
3696 	.poll		   = mptcp_poll,
3697 	.ioctl		   = inet_ioctl,
3698 	.gettstamp	   = sock_gettstamp,
3699 	.listen		   = mptcp_listen,
3700 	.shutdown	   = inet_shutdown,
3701 	.setsockopt	   = sock_common_setsockopt,
3702 	.getsockopt	   = sock_common_getsockopt,
3703 	.sendmsg	   = inet_sendmsg,
3704 	.recvmsg	   = inet_recvmsg,
3705 	.mmap		   = sock_no_mmap,
3706 	.sendpage	   = inet_sendpage,
3707 };
3708 
3709 static struct inet_protosw mptcp_protosw = {
3710 	.type		= SOCK_STREAM,
3711 	.protocol	= IPPROTO_MPTCP,
3712 	.prot		= &mptcp_prot,
3713 	.ops		= &mptcp_stream_ops,
3714 	.flags		= INET_PROTOSW_ICSK,
3715 };
3716 
3717 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3718 {
3719 	struct mptcp_delegated_action *delegated;
3720 	struct mptcp_subflow_context *subflow;
3721 	int work_done = 0;
3722 
3723 	delegated = container_of(napi, struct mptcp_delegated_action, napi);
3724 	while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3725 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3726 
3727 		bh_lock_sock_nested(ssk);
3728 		if (!sock_owned_by_user(ssk) &&
3729 		    mptcp_subflow_has_delegated_action(subflow))
3730 			mptcp_subflow_process_delegated(ssk);
3731 		/* ... elsewhere tcp_release_cb_override already processed
3732 		 * the action or will do at next release_sock().
3733 		 * In both case must dequeue the subflow here - on the same
3734 		 * CPU that scheduled it.
3735 		 */
3736 		bh_unlock_sock(ssk);
3737 		sock_put(ssk);
3738 
3739 		if (++work_done == budget)
3740 			return budget;
3741 	}
3742 
3743 	/* always provide a 0 'work_done' argument, so that napi_complete_done
3744 	 * will not try accessing the NULL napi->dev ptr
3745 	 */
3746 	napi_complete_done(napi, 0);
3747 	return work_done;
3748 }
3749 
3750 void __init mptcp_proto_init(void)
3751 {
3752 	struct mptcp_delegated_action *delegated;
3753 	int cpu;
3754 
3755 	mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3756 
3757 	if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3758 		panic("Failed to allocate MPTCP pcpu counter\n");
3759 
3760 	init_dummy_netdev(&mptcp_napi_dev);
3761 	for_each_possible_cpu(cpu) {
3762 		delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3763 		INIT_LIST_HEAD(&delegated->head);
3764 		netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
3765 				  mptcp_napi_poll);
3766 		napi_enable(&delegated->napi);
3767 	}
3768 
3769 	mptcp_subflow_init();
3770 	mptcp_pm_init();
3771 	mptcp_token_init();
3772 
3773 	if (proto_register(&mptcp_prot, 1) != 0)
3774 		panic("Failed to register MPTCP proto.\n");
3775 
3776 	inet_register_protosw(&mptcp_protosw);
3777 
3778 	BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3779 }
3780 
3781 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3782 static const struct proto_ops mptcp_v6_stream_ops = {
3783 	.family		   = PF_INET6,
3784 	.owner		   = THIS_MODULE,
3785 	.release	   = inet6_release,
3786 	.bind		   = mptcp_bind,
3787 	.connect	   = mptcp_stream_connect,
3788 	.socketpair	   = sock_no_socketpair,
3789 	.accept		   = mptcp_stream_accept,
3790 	.getname	   = inet6_getname,
3791 	.poll		   = mptcp_poll,
3792 	.ioctl		   = inet6_ioctl,
3793 	.gettstamp	   = sock_gettstamp,
3794 	.listen		   = mptcp_listen,
3795 	.shutdown	   = inet_shutdown,
3796 	.setsockopt	   = sock_common_setsockopt,
3797 	.getsockopt	   = sock_common_getsockopt,
3798 	.sendmsg	   = inet6_sendmsg,
3799 	.recvmsg	   = inet6_recvmsg,
3800 	.mmap		   = sock_no_mmap,
3801 	.sendpage	   = inet_sendpage,
3802 #ifdef CONFIG_COMPAT
3803 	.compat_ioctl	   = inet6_compat_ioctl,
3804 #endif
3805 };
3806 
3807 static struct proto mptcp_v6_prot;
3808 
3809 static void mptcp_v6_destroy(struct sock *sk)
3810 {
3811 	mptcp_destroy(sk);
3812 	inet6_destroy_sock(sk);
3813 }
3814 
3815 static struct inet_protosw mptcp_v6_protosw = {
3816 	.type		= SOCK_STREAM,
3817 	.protocol	= IPPROTO_MPTCP,
3818 	.prot		= &mptcp_v6_prot,
3819 	.ops		= &mptcp_v6_stream_ops,
3820 	.flags		= INET_PROTOSW_ICSK,
3821 };
3822 
3823 int __init mptcp_proto_v6_init(void)
3824 {
3825 	int err;
3826 
3827 	mptcp_v6_prot = mptcp_prot;
3828 	strcpy(mptcp_v6_prot.name, "MPTCPv6");
3829 	mptcp_v6_prot.slab = NULL;
3830 	mptcp_v6_prot.destroy = mptcp_v6_destroy;
3831 	mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3832 
3833 	err = proto_register(&mptcp_v6_prot, 1);
3834 	if (err)
3835 		return err;
3836 
3837 	err = inet6_register_protosw(&mptcp_v6_protosw);
3838 	if (err)
3839 		proto_unregister(&mptcp_v6_prot);
3840 
3841 	return err;
3842 }
3843 #endif
3844