xref: /openbmc/linux/net/mptcp/protocol.c (revision 19758688)
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 enum {
40 	MPTCP_CMSG_TS = BIT(0),
41 	MPTCP_CMSG_INQ = BIT(1),
42 };
43 
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
45 
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void mptcp_check_send_data_fin(struct sock *sk);
48 
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
51 
52 /* Returns end sequence number of the receiver's advertised window */
53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
54 {
55 	return READ_ONCE(msk->wnd_end);
56 }
57 
58 static bool mptcp_is_tcpsk(struct sock *sk)
59 {
60 	struct socket *sock = sk->sk_socket;
61 
62 	if (unlikely(sk->sk_prot == &tcp_prot)) {
63 		/* we are being invoked after mptcp_accept() has
64 		 * accepted a non-mp-capable flow: sk is a tcp_sk,
65 		 * not an mptcp one.
66 		 *
67 		 * Hand the socket over to tcp so all further socket ops
68 		 * bypass mptcp.
69 		 */
70 		WRITE_ONCE(sock->ops, &inet_stream_ops);
71 		return true;
72 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
73 	} else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
74 		WRITE_ONCE(sock->ops, &inet6_stream_ops);
75 		return true;
76 #endif
77 	}
78 
79 	return false;
80 }
81 
82 static int __mptcp_socket_create(struct mptcp_sock *msk)
83 {
84 	struct mptcp_subflow_context *subflow;
85 	struct sock *sk = (struct sock *)msk;
86 	struct socket *ssock;
87 	int err;
88 
89 	err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
90 	if (err)
91 		return err;
92 
93 	msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
94 	WRITE_ONCE(msk->first, ssock->sk);
95 	subflow = mptcp_subflow_ctx(ssock->sk);
96 	list_add(&subflow->node, &msk->conn_list);
97 	sock_hold(ssock->sk);
98 	subflow->request_mptcp = 1;
99 	subflow->subflow_id = msk->subflow_id++;
100 
101 	/* This is the first subflow, always with id 0 */
102 	subflow->local_id_valid = 1;
103 	mptcp_sock_graft(msk->first, sk->sk_socket);
104 	iput(SOCK_INODE(ssock));
105 
106 	return 0;
107 }
108 
109 /* If the MPC handshake is not started, returns the first subflow,
110  * eventually allocating it.
111  */
112 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
113 {
114 	struct sock *sk = (struct sock *)msk;
115 	int ret;
116 
117 	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
118 		return ERR_PTR(-EINVAL);
119 
120 	if (!msk->first) {
121 		ret = __mptcp_socket_create(msk);
122 		if (ret)
123 			return ERR_PTR(ret);
124 
125 		mptcp_sockopt_sync(msk, msk->first);
126 	}
127 
128 	return msk->first;
129 }
130 
131 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
132 {
133 	sk_drops_add(sk, skb);
134 	__kfree_skb(skb);
135 }
136 
137 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
138 {
139 	WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
140 		   mptcp_sk(sk)->rmem_fwd_alloc + size);
141 }
142 
143 static void mptcp_rmem_charge(struct sock *sk, int size)
144 {
145 	mptcp_rmem_fwd_alloc_add(sk, -size);
146 }
147 
148 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
149 			       struct sk_buff *from)
150 {
151 	bool fragstolen;
152 	int delta;
153 
154 	if (MPTCP_SKB_CB(from)->offset ||
155 	    !skb_try_coalesce(to, from, &fragstolen, &delta))
156 		return false;
157 
158 	pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
159 		 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
160 		 to->len, MPTCP_SKB_CB(from)->end_seq);
161 	MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
162 
163 	/* note the fwd memory can reach a negative value after accounting
164 	 * for the delta, but the later skb free will restore a non
165 	 * negative one
166 	 */
167 	atomic_add(delta, &sk->sk_rmem_alloc);
168 	mptcp_rmem_charge(sk, delta);
169 	kfree_skb_partial(from, fragstolen);
170 
171 	return true;
172 }
173 
174 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
175 				   struct sk_buff *from)
176 {
177 	if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
178 		return false;
179 
180 	return mptcp_try_coalesce((struct sock *)msk, to, from);
181 }
182 
183 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
184 {
185 	amount >>= PAGE_SHIFT;
186 	mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
187 	__sk_mem_reduce_allocated(sk, amount);
188 }
189 
190 static void mptcp_rmem_uncharge(struct sock *sk, int size)
191 {
192 	struct mptcp_sock *msk = mptcp_sk(sk);
193 	int reclaimable;
194 
195 	mptcp_rmem_fwd_alloc_add(sk, size);
196 	reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
197 
198 	/* see sk_mem_uncharge() for the rationale behind the following schema */
199 	if (unlikely(reclaimable >= PAGE_SIZE))
200 		__mptcp_rmem_reclaim(sk, reclaimable);
201 }
202 
203 static void mptcp_rfree(struct sk_buff *skb)
204 {
205 	unsigned int len = skb->truesize;
206 	struct sock *sk = skb->sk;
207 
208 	atomic_sub(len, &sk->sk_rmem_alloc);
209 	mptcp_rmem_uncharge(sk, len);
210 }
211 
212 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
213 {
214 	skb_orphan(skb);
215 	skb->sk = sk;
216 	skb->destructor = mptcp_rfree;
217 	atomic_add(skb->truesize, &sk->sk_rmem_alloc);
218 	mptcp_rmem_charge(sk, skb->truesize);
219 }
220 
221 /* "inspired" by tcp_data_queue_ofo(), main differences:
222  * - use mptcp seqs
223  * - don't cope with sacks
224  */
225 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
226 {
227 	struct sock *sk = (struct sock *)msk;
228 	struct rb_node **p, *parent;
229 	u64 seq, end_seq, max_seq;
230 	struct sk_buff *skb1;
231 
232 	seq = MPTCP_SKB_CB(skb)->map_seq;
233 	end_seq = MPTCP_SKB_CB(skb)->end_seq;
234 	max_seq = atomic64_read(&msk->rcv_wnd_sent);
235 
236 	pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
237 		 RB_EMPTY_ROOT(&msk->out_of_order_queue));
238 	if (after64(end_seq, max_seq)) {
239 		/* out of window */
240 		mptcp_drop(sk, skb);
241 		pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
242 			 (unsigned long long)end_seq - (unsigned long)max_seq,
243 			 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
244 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
245 		return;
246 	}
247 
248 	p = &msk->out_of_order_queue.rb_node;
249 	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
250 	if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
251 		rb_link_node(&skb->rbnode, NULL, p);
252 		rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
253 		msk->ooo_last_skb = skb;
254 		goto end;
255 	}
256 
257 	/* with 2 subflows, adding at end of ooo queue is quite likely
258 	 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
259 	 */
260 	if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
261 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
262 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
263 		return;
264 	}
265 
266 	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
267 	if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
268 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
269 		parent = &msk->ooo_last_skb->rbnode;
270 		p = &parent->rb_right;
271 		goto insert;
272 	}
273 
274 	/* Find place to insert this segment. Handle overlaps on the way. */
275 	parent = NULL;
276 	while (*p) {
277 		parent = *p;
278 		skb1 = rb_to_skb(parent);
279 		if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
280 			p = &parent->rb_left;
281 			continue;
282 		}
283 		if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
284 			if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
285 				/* All the bits are present. Drop. */
286 				mptcp_drop(sk, skb);
287 				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
288 				return;
289 			}
290 			if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
291 				/* partial overlap:
292 				 *     |     skb      |
293 				 *  |     skb1    |
294 				 * continue traversing
295 				 */
296 			} else {
297 				/* skb's seq == skb1's seq and skb covers skb1.
298 				 * Replace skb1 with skb.
299 				 */
300 				rb_replace_node(&skb1->rbnode, &skb->rbnode,
301 						&msk->out_of_order_queue);
302 				mptcp_drop(sk, skb1);
303 				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
304 				goto merge_right;
305 			}
306 		} else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
307 			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
308 			return;
309 		}
310 		p = &parent->rb_right;
311 	}
312 
313 insert:
314 	/* Insert segment into RB tree. */
315 	rb_link_node(&skb->rbnode, parent, p);
316 	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
317 
318 merge_right:
319 	/* Remove other segments covered by skb. */
320 	while ((skb1 = skb_rb_next(skb)) != NULL) {
321 		if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
322 			break;
323 		rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
324 		mptcp_drop(sk, skb1);
325 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
326 	}
327 	/* If there is no skb after us, we are the last_skb ! */
328 	if (!skb1)
329 		msk->ooo_last_skb = skb;
330 
331 end:
332 	skb_condense(skb);
333 	mptcp_set_owner_r(skb, sk);
334 }
335 
336 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
337 {
338 	struct mptcp_sock *msk = mptcp_sk(sk);
339 	int amt, amount;
340 
341 	if (size <= msk->rmem_fwd_alloc)
342 		return true;
343 
344 	size -= msk->rmem_fwd_alloc;
345 	amt = sk_mem_pages(size);
346 	amount = amt << PAGE_SHIFT;
347 	if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
348 		return false;
349 
350 	mptcp_rmem_fwd_alloc_add(sk, amount);
351 	return true;
352 }
353 
354 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
355 			     struct sk_buff *skb, unsigned int offset,
356 			     size_t copy_len)
357 {
358 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
359 	struct sock *sk = (struct sock *)msk;
360 	struct sk_buff *tail;
361 	bool has_rxtstamp;
362 
363 	__skb_unlink(skb, &ssk->sk_receive_queue);
364 
365 	skb_ext_reset(skb);
366 	skb_orphan(skb);
367 
368 	/* try to fetch required memory from subflow */
369 	if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
370 		goto drop;
371 
372 	has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
373 
374 	/* the skb map_seq accounts for the skb offset:
375 	 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
376 	 * value
377 	 */
378 	MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
379 	MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
380 	MPTCP_SKB_CB(skb)->offset = offset;
381 	MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
382 
383 	if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
384 		/* in sequence */
385 		msk->bytes_received += copy_len;
386 		WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
387 		tail = skb_peek_tail(&sk->sk_receive_queue);
388 		if (tail && mptcp_try_coalesce(sk, tail, skb))
389 			return true;
390 
391 		mptcp_set_owner_r(skb, sk);
392 		__skb_queue_tail(&sk->sk_receive_queue, skb);
393 		return true;
394 	} else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
395 		mptcp_data_queue_ofo(msk, skb);
396 		return false;
397 	}
398 
399 	/* old data, keep it simple and drop the whole pkt, sender
400 	 * will retransmit as needed, if needed.
401 	 */
402 	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
403 drop:
404 	mptcp_drop(sk, skb);
405 	return false;
406 }
407 
408 static void mptcp_stop_rtx_timer(struct sock *sk)
409 {
410 	struct inet_connection_sock *icsk = inet_csk(sk);
411 
412 	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
413 	mptcp_sk(sk)->timer_ival = 0;
414 }
415 
416 static void mptcp_close_wake_up(struct sock *sk)
417 {
418 	if (sock_flag(sk, SOCK_DEAD))
419 		return;
420 
421 	sk->sk_state_change(sk);
422 	if (sk->sk_shutdown == SHUTDOWN_MASK ||
423 	    sk->sk_state == TCP_CLOSE)
424 		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
425 	else
426 		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
427 }
428 
429 static bool mptcp_pending_data_fin_ack(struct sock *sk)
430 {
431 	struct mptcp_sock *msk = mptcp_sk(sk);
432 
433 	return ((1 << sk->sk_state) &
434 		(TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
435 	       msk->write_seq == READ_ONCE(msk->snd_una);
436 }
437 
438 static void mptcp_check_data_fin_ack(struct sock *sk)
439 {
440 	struct mptcp_sock *msk = mptcp_sk(sk);
441 
442 	/* Look for an acknowledged DATA_FIN */
443 	if (mptcp_pending_data_fin_ack(sk)) {
444 		WRITE_ONCE(msk->snd_data_fin_enable, 0);
445 
446 		switch (sk->sk_state) {
447 		case TCP_FIN_WAIT1:
448 			inet_sk_state_store(sk, TCP_FIN_WAIT2);
449 			break;
450 		case TCP_CLOSING:
451 		case TCP_LAST_ACK:
452 			inet_sk_state_store(sk, TCP_CLOSE);
453 			break;
454 		}
455 
456 		mptcp_close_wake_up(sk);
457 	}
458 }
459 
460 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
461 {
462 	struct mptcp_sock *msk = mptcp_sk(sk);
463 
464 	if (READ_ONCE(msk->rcv_data_fin) &&
465 	    ((1 << sk->sk_state) &
466 	     (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
467 		u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
468 
469 		if (msk->ack_seq == rcv_data_fin_seq) {
470 			if (seq)
471 				*seq = rcv_data_fin_seq;
472 
473 			return true;
474 		}
475 	}
476 
477 	return false;
478 }
479 
480 static void mptcp_set_datafin_timeout(struct sock *sk)
481 {
482 	struct inet_connection_sock *icsk = inet_csk(sk);
483 	u32 retransmits;
484 
485 	retransmits = min_t(u32, icsk->icsk_retransmits,
486 			    ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
487 
488 	mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
489 }
490 
491 static void __mptcp_set_timeout(struct sock *sk, long tout)
492 {
493 	mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
494 }
495 
496 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
497 {
498 	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
499 
500 	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
501 	       inet_csk(ssk)->icsk_timeout - jiffies : 0;
502 }
503 
504 static void mptcp_set_timeout(struct sock *sk)
505 {
506 	struct mptcp_subflow_context *subflow;
507 	long tout = 0;
508 
509 	mptcp_for_each_subflow(mptcp_sk(sk), subflow)
510 		tout = max(tout, mptcp_timeout_from_subflow(subflow));
511 	__mptcp_set_timeout(sk, tout);
512 }
513 
514 static inline bool tcp_can_send_ack(const struct sock *ssk)
515 {
516 	return !((1 << inet_sk_state_load(ssk)) &
517 	       (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
518 }
519 
520 void __mptcp_subflow_send_ack(struct sock *ssk)
521 {
522 	if (tcp_can_send_ack(ssk))
523 		tcp_send_ack(ssk);
524 }
525 
526 static void mptcp_subflow_send_ack(struct sock *ssk)
527 {
528 	bool slow;
529 
530 	slow = lock_sock_fast(ssk);
531 	__mptcp_subflow_send_ack(ssk);
532 	unlock_sock_fast(ssk, slow);
533 }
534 
535 static void mptcp_send_ack(struct mptcp_sock *msk)
536 {
537 	struct mptcp_subflow_context *subflow;
538 
539 	mptcp_for_each_subflow(msk, subflow)
540 		mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
541 }
542 
543 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
544 {
545 	bool slow;
546 
547 	slow = lock_sock_fast(ssk);
548 	if (tcp_can_send_ack(ssk))
549 		tcp_cleanup_rbuf(ssk, 1);
550 	unlock_sock_fast(ssk, slow);
551 }
552 
553 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
554 {
555 	const struct inet_connection_sock *icsk = inet_csk(ssk);
556 	u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
557 	const struct tcp_sock *tp = tcp_sk(ssk);
558 
559 	return (ack_pending & ICSK_ACK_SCHED) &&
560 		((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
561 		  READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
562 		 (rx_empty && ack_pending &
563 			      (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
564 }
565 
566 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
567 {
568 	int old_space = READ_ONCE(msk->old_wspace);
569 	struct mptcp_subflow_context *subflow;
570 	struct sock *sk = (struct sock *)msk;
571 	int space =  __mptcp_space(sk);
572 	bool cleanup, rx_empty;
573 
574 	cleanup = (space > 0) && (space >= (old_space << 1));
575 	rx_empty = !__mptcp_rmem(sk);
576 
577 	mptcp_for_each_subflow(msk, subflow) {
578 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
579 
580 		if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
581 			mptcp_subflow_cleanup_rbuf(ssk);
582 	}
583 }
584 
585 static bool mptcp_check_data_fin(struct sock *sk)
586 {
587 	struct mptcp_sock *msk = mptcp_sk(sk);
588 	u64 rcv_data_fin_seq;
589 	bool ret = false;
590 
591 	/* Need to ack a DATA_FIN received from a peer while this side
592 	 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
593 	 * msk->rcv_data_fin was set when parsing the incoming options
594 	 * at the subflow level and the msk lock was not held, so this
595 	 * is the first opportunity to act on the DATA_FIN and change
596 	 * the msk state.
597 	 *
598 	 * If we are caught up to the sequence number of the incoming
599 	 * DATA_FIN, send the DATA_ACK now and do state transition.  If
600 	 * not caught up, do nothing and let the recv code send DATA_ACK
601 	 * when catching up.
602 	 */
603 
604 	if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
605 		WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
606 		WRITE_ONCE(msk->rcv_data_fin, 0);
607 
608 		WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
609 		smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
610 
611 		switch (sk->sk_state) {
612 		case TCP_ESTABLISHED:
613 			inet_sk_state_store(sk, TCP_CLOSE_WAIT);
614 			break;
615 		case TCP_FIN_WAIT1:
616 			inet_sk_state_store(sk, TCP_CLOSING);
617 			break;
618 		case TCP_FIN_WAIT2:
619 			inet_sk_state_store(sk, TCP_CLOSE);
620 			break;
621 		default:
622 			/* Other states not expected */
623 			WARN_ON_ONCE(1);
624 			break;
625 		}
626 
627 		ret = true;
628 		if (!__mptcp_check_fallback(msk))
629 			mptcp_send_ack(msk);
630 		mptcp_close_wake_up(sk);
631 	}
632 	return ret;
633 }
634 
635 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
636 					   struct sock *ssk,
637 					   unsigned int *bytes)
638 {
639 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
640 	struct sock *sk = (struct sock *)msk;
641 	unsigned int moved = 0;
642 	bool more_data_avail;
643 	struct tcp_sock *tp;
644 	bool done = false;
645 	int sk_rbuf;
646 
647 	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
648 
649 	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
650 		int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
651 
652 		if (unlikely(ssk_rbuf > sk_rbuf)) {
653 			WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
654 			sk_rbuf = ssk_rbuf;
655 		}
656 	}
657 
658 	pr_debug("msk=%p ssk=%p", msk, ssk);
659 	tp = tcp_sk(ssk);
660 	do {
661 		u32 map_remaining, offset;
662 		u32 seq = tp->copied_seq;
663 		struct sk_buff *skb;
664 		bool fin;
665 
666 		/* try to move as much data as available */
667 		map_remaining = subflow->map_data_len -
668 				mptcp_subflow_get_map_offset(subflow);
669 
670 		skb = skb_peek(&ssk->sk_receive_queue);
671 		if (!skb) {
672 			/* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
673 			 * a different CPU can have already processed the pending
674 			 * data, stop here or we can enter an infinite loop
675 			 */
676 			if (!moved)
677 				done = true;
678 			break;
679 		}
680 
681 		if (__mptcp_check_fallback(msk)) {
682 			/* Under fallback skbs have no MPTCP extension and TCP could
683 			 * collapse them between the dummy map creation and the
684 			 * current dequeue. Be sure to adjust the map size.
685 			 */
686 			map_remaining = skb->len;
687 			subflow->map_data_len = skb->len;
688 		}
689 
690 		offset = seq - TCP_SKB_CB(skb)->seq;
691 		fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
692 		if (fin) {
693 			done = true;
694 			seq++;
695 		}
696 
697 		if (offset < skb->len) {
698 			size_t len = skb->len - offset;
699 
700 			if (tp->urg_data)
701 				done = true;
702 
703 			if (__mptcp_move_skb(msk, ssk, skb, offset, len))
704 				moved += len;
705 			seq += len;
706 
707 			if (WARN_ON_ONCE(map_remaining < len))
708 				break;
709 		} else {
710 			WARN_ON_ONCE(!fin);
711 			sk_eat_skb(ssk, skb);
712 			done = true;
713 		}
714 
715 		WRITE_ONCE(tp->copied_seq, seq);
716 		more_data_avail = mptcp_subflow_data_available(ssk);
717 
718 		if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
719 			done = true;
720 			break;
721 		}
722 	} while (more_data_avail);
723 
724 	*bytes += moved;
725 	return done;
726 }
727 
728 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
729 {
730 	struct sock *sk = (struct sock *)msk;
731 	struct sk_buff *skb, *tail;
732 	bool moved = false;
733 	struct rb_node *p;
734 	u64 end_seq;
735 
736 	p = rb_first(&msk->out_of_order_queue);
737 	pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
738 	while (p) {
739 		skb = rb_to_skb(p);
740 		if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
741 			break;
742 
743 		p = rb_next(p);
744 		rb_erase(&skb->rbnode, &msk->out_of_order_queue);
745 
746 		if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
747 				      msk->ack_seq))) {
748 			mptcp_drop(sk, skb);
749 			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
750 			continue;
751 		}
752 
753 		end_seq = MPTCP_SKB_CB(skb)->end_seq;
754 		tail = skb_peek_tail(&sk->sk_receive_queue);
755 		if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
756 			int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
757 
758 			/* skip overlapping data, if any */
759 			pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
760 				 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
761 				 delta);
762 			MPTCP_SKB_CB(skb)->offset += delta;
763 			MPTCP_SKB_CB(skb)->map_seq += delta;
764 			__skb_queue_tail(&sk->sk_receive_queue, skb);
765 		}
766 		msk->bytes_received += end_seq - msk->ack_seq;
767 		msk->ack_seq = end_seq;
768 		moved = true;
769 	}
770 	return moved;
771 }
772 
773 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
774 {
775 	int err = sock_error(ssk);
776 	int ssk_state;
777 
778 	if (!err)
779 		return false;
780 
781 	/* only propagate errors on fallen-back sockets or
782 	 * on MPC connect
783 	 */
784 	if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
785 		return false;
786 
787 	/* We need to propagate only transition to CLOSE state.
788 	 * Orphaned socket will see such state change via
789 	 * subflow_sched_work_if_closed() and that path will properly
790 	 * destroy the msk as needed.
791 	 */
792 	ssk_state = inet_sk_state_load(ssk);
793 	if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
794 		inet_sk_state_store(sk, ssk_state);
795 	WRITE_ONCE(sk->sk_err, -err);
796 
797 	/* This barrier is coupled with smp_rmb() in mptcp_poll() */
798 	smp_wmb();
799 	sk_error_report(sk);
800 	return true;
801 }
802 
803 void __mptcp_error_report(struct sock *sk)
804 {
805 	struct mptcp_subflow_context *subflow;
806 	struct mptcp_sock *msk = mptcp_sk(sk);
807 
808 	mptcp_for_each_subflow(msk, subflow)
809 		if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
810 			break;
811 }
812 
813 /* In most cases we will be able to lock the mptcp socket.  If its already
814  * owned, we need to defer to the work queue to avoid ABBA deadlock.
815  */
816 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
817 {
818 	struct sock *sk = (struct sock *)msk;
819 	unsigned int moved = 0;
820 
821 	__mptcp_move_skbs_from_subflow(msk, ssk, &moved);
822 	__mptcp_ofo_queue(msk);
823 	if (unlikely(ssk->sk_err)) {
824 		if (!sock_owned_by_user(sk))
825 			__mptcp_error_report(sk);
826 		else
827 			__set_bit(MPTCP_ERROR_REPORT,  &msk->cb_flags);
828 	}
829 
830 	/* If the moves have caught up with the DATA_FIN sequence number
831 	 * it's time to ack the DATA_FIN and change socket state, but
832 	 * this is not a good place to change state. Let the workqueue
833 	 * do it.
834 	 */
835 	if (mptcp_pending_data_fin(sk, NULL))
836 		mptcp_schedule_work(sk);
837 	return moved > 0;
838 }
839 
840 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
841 {
842 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
843 	struct mptcp_sock *msk = mptcp_sk(sk);
844 	int sk_rbuf, ssk_rbuf;
845 
846 	/* The peer can send data while we are shutting down this
847 	 * subflow at msk destruction time, but we must avoid enqueuing
848 	 * more data to the msk receive queue
849 	 */
850 	if (unlikely(subflow->disposable))
851 		return;
852 
853 	ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
854 	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
855 	if (unlikely(ssk_rbuf > sk_rbuf))
856 		sk_rbuf = ssk_rbuf;
857 
858 	/* over limit? can't append more skbs to msk, Also, no need to wake-up*/
859 	if (__mptcp_rmem(sk) > sk_rbuf) {
860 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
861 		return;
862 	}
863 
864 	/* Wake-up the reader only for in-sequence data */
865 	mptcp_data_lock(sk);
866 	if (move_skbs_to_msk(msk, ssk))
867 		sk->sk_data_ready(sk);
868 
869 	mptcp_data_unlock(sk);
870 }
871 
872 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
873 {
874 	mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
875 	WRITE_ONCE(msk->allow_infinite_fallback, false);
876 	mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
877 }
878 
879 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
880 {
881 	struct sock *sk = (struct sock *)msk;
882 
883 	if (sk->sk_state != TCP_ESTABLISHED)
884 		return false;
885 
886 	/* attach to msk socket only after we are sure we will deal with it
887 	 * at close time
888 	 */
889 	if (sk->sk_socket && !ssk->sk_socket)
890 		mptcp_sock_graft(ssk, sk->sk_socket);
891 
892 	mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
893 	mptcp_sockopt_sync_locked(msk, ssk);
894 	mptcp_subflow_joined(msk, ssk);
895 	mptcp_stop_tout_timer(sk);
896 	__mptcp_propagate_sndbuf(sk, ssk);
897 	return true;
898 }
899 
900 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
901 {
902 	struct mptcp_subflow_context *tmp, *subflow;
903 	struct mptcp_sock *msk = mptcp_sk(sk);
904 
905 	list_for_each_entry_safe(subflow, tmp, join_list, node) {
906 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
907 		bool slow = lock_sock_fast(ssk);
908 
909 		list_move_tail(&subflow->node, &msk->conn_list);
910 		if (!__mptcp_finish_join(msk, ssk))
911 			mptcp_subflow_reset(ssk);
912 		unlock_sock_fast(ssk, slow);
913 	}
914 }
915 
916 static bool mptcp_rtx_timer_pending(struct sock *sk)
917 {
918 	return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
919 }
920 
921 static void mptcp_reset_rtx_timer(struct sock *sk)
922 {
923 	struct inet_connection_sock *icsk = inet_csk(sk);
924 	unsigned long tout;
925 
926 	/* prevent rescheduling on close */
927 	if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
928 		return;
929 
930 	tout = mptcp_sk(sk)->timer_ival;
931 	sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
932 }
933 
934 bool mptcp_schedule_work(struct sock *sk)
935 {
936 	if (inet_sk_state_load(sk) != TCP_CLOSE &&
937 	    schedule_work(&mptcp_sk(sk)->work)) {
938 		/* each subflow already holds a reference to the sk, and the
939 		 * workqueue is invoked by a subflow, so sk can't go away here.
940 		 */
941 		sock_hold(sk);
942 		return true;
943 	}
944 	return false;
945 }
946 
947 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
948 {
949 	struct mptcp_subflow_context *subflow;
950 
951 	msk_owned_by_me(msk);
952 
953 	mptcp_for_each_subflow(msk, subflow) {
954 		if (READ_ONCE(subflow->data_avail))
955 			return mptcp_subflow_tcp_sock(subflow);
956 	}
957 
958 	return NULL;
959 }
960 
961 static bool mptcp_skb_can_collapse_to(u64 write_seq,
962 				      const struct sk_buff *skb,
963 				      const struct mptcp_ext *mpext)
964 {
965 	if (!tcp_skb_can_collapse_to(skb))
966 		return false;
967 
968 	/* can collapse only if MPTCP level sequence is in order and this
969 	 * mapping has not been xmitted yet
970 	 */
971 	return mpext && mpext->data_seq + mpext->data_len == write_seq &&
972 	       !mpext->frozen;
973 }
974 
975 /* we can append data to the given data frag if:
976  * - there is space available in the backing page_frag
977  * - the data frag tail matches the current page_frag free offset
978  * - the data frag end sequence number matches the current write seq
979  */
980 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
981 				       const struct page_frag *pfrag,
982 				       const struct mptcp_data_frag *df)
983 {
984 	return df && pfrag->page == df->page &&
985 		pfrag->size - pfrag->offset > 0 &&
986 		pfrag->offset == (df->offset + df->data_len) &&
987 		df->data_seq + df->data_len == msk->write_seq;
988 }
989 
990 static void dfrag_uncharge(struct sock *sk, int len)
991 {
992 	sk_mem_uncharge(sk, len);
993 	sk_wmem_queued_add(sk, -len);
994 }
995 
996 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
997 {
998 	int len = dfrag->data_len + dfrag->overhead;
999 
1000 	list_del(&dfrag->list);
1001 	dfrag_uncharge(sk, len);
1002 	put_page(dfrag->page);
1003 }
1004 
1005 static void __mptcp_clean_una(struct sock *sk)
1006 {
1007 	struct mptcp_sock *msk = mptcp_sk(sk);
1008 	struct mptcp_data_frag *dtmp, *dfrag;
1009 	u64 snd_una;
1010 
1011 	snd_una = msk->snd_una;
1012 	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1013 		if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1014 			break;
1015 
1016 		if (unlikely(dfrag == msk->first_pending)) {
1017 			/* in recovery mode can see ack after the current snd head */
1018 			if (WARN_ON_ONCE(!msk->recovery))
1019 				break;
1020 
1021 			WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1022 		}
1023 
1024 		dfrag_clear(sk, dfrag);
1025 	}
1026 
1027 	dfrag = mptcp_rtx_head(sk);
1028 	if (dfrag && after64(snd_una, dfrag->data_seq)) {
1029 		u64 delta = snd_una - dfrag->data_seq;
1030 
1031 		/* prevent wrap around in recovery mode */
1032 		if (unlikely(delta > dfrag->already_sent)) {
1033 			if (WARN_ON_ONCE(!msk->recovery))
1034 				goto out;
1035 			if (WARN_ON_ONCE(delta > dfrag->data_len))
1036 				goto out;
1037 			dfrag->already_sent += delta - dfrag->already_sent;
1038 		}
1039 
1040 		dfrag->data_seq += delta;
1041 		dfrag->offset += delta;
1042 		dfrag->data_len -= delta;
1043 		dfrag->already_sent -= delta;
1044 
1045 		dfrag_uncharge(sk, delta);
1046 	}
1047 
1048 	/* all retransmitted data acked, recovery completed */
1049 	if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1050 		msk->recovery = false;
1051 
1052 out:
1053 	if (snd_una == READ_ONCE(msk->snd_nxt) &&
1054 	    snd_una == READ_ONCE(msk->write_seq)) {
1055 		if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1056 			mptcp_stop_rtx_timer(sk);
1057 	} else {
1058 		mptcp_reset_rtx_timer(sk);
1059 	}
1060 }
1061 
1062 static void __mptcp_clean_una_wakeup(struct sock *sk)
1063 {
1064 	lockdep_assert_held_once(&sk->sk_lock.slock);
1065 
1066 	__mptcp_clean_una(sk);
1067 	mptcp_write_space(sk);
1068 }
1069 
1070 static void mptcp_clean_una_wakeup(struct sock *sk)
1071 {
1072 	mptcp_data_lock(sk);
1073 	__mptcp_clean_una_wakeup(sk);
1074 	mptcp_data_unlock(sk);
1075 }
1076 
1077 static void mptcp_enter_memory_pressure(struct sock *sk)
1078 {
1079 	struct mptcp_subflow_context *subflow;
1080 	struct mptcp_sock *msk = mptcp_sk(sk);
1081 	bool first = true;
1082 
1083 	mptcp_for_each_subflow(msk, subflow) {
1084 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1085 
1086 		if (first)
1087 			tcp_enter_memory_pressure(ssk);
1088 		sk_stream_moderate_sndbuf(ssk);
1089 
1090 		first = false;
1091 	}
1092 	__mptcp_sync_sndbuf(sk);
1093 }
1094 
1095 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1096  * data
1097  */
1098 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1099 {
1100 	if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1101 					pfrag, sk->sk_allocation)))
1102 		return true;
1103 
1104 	mptcp_enter_memory_pressure(sk);
1105 	return false;
1106 }
1107 
1108 static struct mptcp_data_frag *
1109 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1110 		      int orig_offset)
1111 {
1112 	int offset = ALIGN(orig_offset, sizeof(long));
1113 	struct mptcp_data_frag *dfrag;
1114 
1115 	dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1116 	dfrag->data_len = 0;
1117 	dfrag->data_seq = msk->write_seq;
1118 	dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1119 	dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1120 	dfrag->already_sent = 0;
1121 	dfrag->page = pfrag->page;
1122 
1123 	return dfrag;
1124 }
1125 
1126 struct mptcp_sendmsg_info {
1127 	int mss_now;
1128 	int size_goal;
1129 	u16 limit;
1130 	u16 sent;
1131 	unsigned int flags;
1132 	bool data_lock_held;
1133 };
1134 
1135 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1136 				    u64 data_seq, int avail_size)
1137 {
1138 	u64 window_end = mptcp_wnd_end(msk);
1139 	u64 mptcp_snd_wnd;
1140 
1141 	if (__mptcp_check_fallback(msk))
1142 		return avail_size;
1143 
1144 	mptcp_snd_wnd = window_end - data_seq;
1145 	avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1146 
1147 	if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1148 		tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1149 		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1150 	}
1151 
1152 	return avail_size;
1153 }
1154 
1155 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1156 {
1157 	struct skb_ext *mpext = __skb_ext_alloc(gfp);
1158 
1159 	if (!mpext)
1160 		return false;
1161 	__skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1162 	return true;
1163 }
1164 
1165 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1166 {
1167 	struct sk_buff *skb;
1168 
1169 	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1170 	if (likely(skb)) {
1171 		if (likely(__mptcp_add_ext(skb, gfp))) {
1172 			skb_reserve(skb, MAX_TCP_HEADER);
1173 			skb->ip_summed = CHECKSUM_PARTIAL;
1174 			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1175 			return skb;
1176 		}
1177 		__kfree_skb(skb);
1178 	} else {
1179 		mptcp_enter_memory_pressure(sk);
1180 	}
1181 	return NULL;
1182 }
1183 
1184 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1185 {
1186 	struct sk_buff *skb;
1187 
1188 	skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1189 	if (!skb)
1190 		return NULL;
1191 
1192 	if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1193 		tcp_skb_entail(ssk, skb);
1194 		return skb;
1195 	}
1196 	tcp_skb_tsorted_anchor_cleanup(skb);
1197 	kfree_skb(skb);
1198 	return NULL;
1199 }
1200 
1201 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1202 {
1203 	gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1204 
1205 	return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1206 }
1207 
1208 /* note: this always recompute the csum on the whole skb, even
1209  * if we just appended a single frag. More status info needed
1210  */
1211 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1212 {
1213 	struct mptcp_ext *mpext = mptcp_get_ext(skb);
1214 	__wsum csum = ~csum_unfold(mpext->csum);
1215 	int offset = skb->len - added;
1216 
1217 	mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1218 }
1219 
1220 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1221 				      struct sock *ssk,
1222 				      struct mptcp_ext *mpext)
1223 {
1224 	if (!mpext)
1225 		return;
1226 
1227 	mpext->infinite_map = 1;
1228 	mpext->data_len = 0;
1229 
1230 	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1231 	mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1232 	pr_fallback(msk);
1233 	mptcp_do_fallback(ssk);
1234 }
1235 
1236 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1237 
1238 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1239 			      struct mptcp_data_frag *dfrag,
1240 			      struct mptcp_sendmsg_info *info)
1241 {
1242 	u64 data_seq = dfrag->data_seq + info->sent;
1243 	int offset = dfrag->offset + info->sent;
1244 	struct mptcp_sock *msk = mptcp_sk(sk);
1245 	bool zero_window_probe = false;
1246 	struct mptcp_ext *mpext = NULL;
1247 	bool can_coalesce = false;
1248 	bool reuse_skb = true;
1249 	struct sk_buff *skb;
1250 	size_t copy;
1251 	int i;
1252 
1253 	pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1254 		 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1255 
1256 	if (WARN_ON_ONCE(info->sent > info->limit ||
1257 			 info->limit > dfrag->data_len))
1258 		return 0;
1259 
1260 	if (unlikely(!__tcp_can_send(ssk)))
1261 		return -EAGAIN;
1262 
1263 	/* compute send limit */
1264 	if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1265 		ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1266 	info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1267 	copy = info->size_goal;
1268 
1269 	skb = tcp_write_queue_tail(ssk);
1270 	if (skb && copy > skb->len) {
1271 		/* Limit the write to the size available in the
1272 		 * current skb, if any, so that we create at most a new skb.
1273 		 * Explicitly tells TCP internals to avoid collapsing on later
1274 		 * queue management operation, to avoid breaking the ext <->
1275 		 * SSN association set here
1276 		 */
1277 		mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1278 		if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1279 			TCP_SKB_CB(skb)->eor = 1;
1280 			goto alloc_skb;
1281 		}
1282 
1283 		i = skb_shinfo(skb)->nr_frags;
1284 		can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1285 		if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1286 			tcp_mark_push(tcp_sk(ssk), skb);
1287 			goto alloc_skb;
1288 		}
1289 
1290 		copy -= skb->len;
1291 	} else {
1292 alloc_skb:
1293 		skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1294 		if (!skb)
1295 			return -ENOMEM;
1296 
1297 		i = skb_shinfo(skb)->nr_frags;
1298 		reuse_skb = false;
1299 		mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1300 	}
1301 
1302 	/* Zero window and all data acked? Probe. */
1303 	copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1304 	if (copy == 0) {
1305 		u64 snd_una = READ_ONCE(msk->snd_una);
1306 
1307 		if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1308 			tcp_remove_empty_skb(ssk);
1309 			return 0;
1310 		}
1311 
1312 		zero_window_probe = true;
1313 		data_seq = snd_una - 1;
1314 		copy = 1;
1315 	}
1316 
1317 	copy = min_t(size_t, copy, info->limit - info->sent);
1318 	if (!sk_wmem_schedule(ssk, copy)) {
1319 		tcp_remove_empty_skb(ssk);
1320 		return -ENOMEM;
1321 	}
1322 
1323 	if (can_coalesce) {
1324 		skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1325 	} else {
1326 		get_page(dfrag->page);
1327 		skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1328 	}
1329 
1330 	skb->len += copy;
1331 	skb->data_len += copy;
1332 	skb->truesize += copy;
1333 	sk_wmem_queued_add(ssk, copy);
1334 	sk_mem_charge(ssk, copy);
1335 	WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1336 	TCP_SKB_CB(skb)->end_seq += copy;
1337 	tcp_skb_pcount_set(skb, 0);
1338 
1339 	/* on skb reuse we just need to update the DSS len */
1340 	if (reuse_skb) {
1341 		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1342 		mpext->data_len += copy;
1343 		goto out;
1344 	}
1345 
1346 	memset(mpext, 0, sizeof(*mpext));
1347 	mpext->data_seq = data_seq;
1348 	mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1349 	mpext->data_len = copy;
1350 	mpext->use_map = 1;
1351 	mpext->dsn64 = 1;
1352 
1353 	pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1354 		 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1355 		 mpext->dsn64);
1356 
1357 	if (zero_window_probe) {
1358 		mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1359 		mpext->frozen = 1;
1360 		if (READ_ONCE(msk->csum_enabled))
1361 			mptcp_update_data_checksum(skb, copy);
1362 		tcp_push_pending_frames(ssk);
1363 		return 0;
1364 	}
1365 out:
1366 	if (READ_ONCE(msk->csum_enabled))
1367 		mptcp_update_data_checksum(skb, copy);
1368 	if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1369 		mptcp_update_infinite_map(msk, ssk, mpext);
1370 	trace_mptcp_sendmsg_frag(mpext);
1371 	mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1372 	return copy;
1373 }
1374 
1375 #define MPTCP_SEND_BURST_SIZE		((1 << 16) - \
1376 					 sizeof(struct tcphdr) - \
1377 					 MAX_TCP_OPTION_SPACE - \
1378 					 sizeof(struct ipv6hdr) - \
1379 					 sizeof(struct frag_hdr))
1380 
1381 struct subflow_send_info {
1382 	struct sock *ssk;
1383 	u64 linger_time;
1384 };
1385 
1386 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1387 {
1388 	if (!subflow->stale)
1389 		return;
1390 
1391 	subflow->stale = 0;
1392 	MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1393 }
1394 
1395 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1396 {
1397 	if (unlikely(subflow->stale)) {
1398 		u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1399 
1400 		if (subflow->stale_rcv_tstamp == rcv_tstamp)
1401 			return false;
1402 
1403 		mptcp_subflow_set_active(subflow);
1404 	}
1405 	return __mptcp_subflow_active(subflow);
1406 }
1407 
1408 #define SSK_MODE_ACTIVE	0
1409 #define SSK_MODE_BACKUP	1
1410 #define SSK_MODE_MAX	2
1411 
1412 /* implement the mptcp packet scheduler;
1413  * returns the subflow that will transmit the next DSS
1414  * additionally updates the rtx timeout
1415  */
1416 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1417 {
1418 	struct subflow_send_info send_info[SSK_MODE_MAX];
1419 	struct mptcp_subflow_context *subflow;
1420 	struct sock *sk = (struct sock *)msk;
1421 	u32 pace, burst, wmem;
1422 	int i, nr_active = 0;
1423 	struct sock *ssk;
1424 	u64 linger_time;
1425 	long tout = 0;
1426 
1427 	/* pick the subflow with the lower wmem/wspace ratio */
1428 	for (i = 0; i < SSK_MODE_MAX; ++i) {
1429 		send_info[i].ssk = NULL;
1430 		send_info[i].linger_time = -1;
1431 	}
1432 
1433 	mptcp_for_each_subflow(msk, subflow) {
1434 		trace_mptcp_subflow_get_send(subflow);
1435 		ssk =  mptcp_subflow_tcp_sock(subflow);
1436 		if (!mptcp_subflow_active(subflow))
1437 			continue;
1438 
1439 		tout = max(tout, mptcp_timeout_from_subflow(subflow));
1440 		nr_active += !subflow->backup;
1441 		pace = subflow->avg_pacing_rate;
1442 		if (unlikely(!pace)) {
1443 			/* init pacing rate from socket */
1444 			subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1445 			pace = subflow->avg_pacing_rate;
1446 			if (!pace)
1447 				continue;
1448 		}
1449 
1450 		linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1451 		if (linger_time < send_info[subflow->backup].linger_time) {
1452 			send_info[subflow->backup].ssk = ssk;
1453 			send_info[subflow->backup].linger_time = linger_time;
1454 		}
1455 	}
1456 	__mptcp_set_timeout(sk, tout);
1457 
1458 	/* pick the best backup if no other subflow is active */
1459 	if (!nr_active)
1460 		send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1461 
1462 	/* According to the blest algorithm, to avoid HoL blocking for the
1463 	 * faster flow, we need to:
1464 	 * - estimate the faster flow linger time
1465 	 * - use the above to estimate the amount of byte transferred
1466 	 *   by the faster flow
1467 	 * - check that the amount of queued data is greter than the above,
1468 	 *   otherwise do not use the picked, slower, subflow
1469 	 * We select the subflow with the shorter estimated time to flush
1470 	 * the queued mem, which basically ensure the above. We just need
1471 	 * to check that subflow has a non empty cwin.
1472 	 */
1473 	ssk = send_info[SSK_MODE_ACTIVE].ssk;
1474 	if (!ssk || !sk_stream_memory_free(ssk))
1475 		return NULL;
1476 
1477 	burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1478 	wmem = READ_ONCE(ssk->sk_wmem_queued);
1479 	if (!burst)
1480 		return ssk;
1481 
1482 	subflow = mptcp_subflow_ctx(ssk);
1483 	subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1484 					   READ_ONCE(ssk->sk_pacing_rate) * burst,
1485 					   burst + wmem);
1486 	msk->snd_burst = burst;
1487 	return ssk;
1488 }
1489 
1490 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1491 {
1492 	tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1493 	release_sock(ssk);
1494 }
1495 
1496 static void mptcp_update_post_push(struct mptcp_sock *msk,
1497 				   struct mptcp_data_frag *dfrag,
1498 				   u32 sent)
1499 {
1500 	u64 snd_nxt_new = dfrag->data_seq;
1501 
1502 	dfrag->already_sent += sent;
1503 
1504 	msk->snd_burst -= sent;
1505 
1506 	snd_nxt_new += dfrag->already_sent;
1507 
1508 	/* snd_nxt_new can be smaller than snd_nxt in case mptcp
1509 	 * is recovering after a failover. In that event, this re-sends
1510 	 * old segments.
1511 	 *
1512 	 * Thus compute snd_nxt_new candidate based on
1513 	 * the dfrag->data_seq that was sent and the data
1514 	 * that has been handed to the subflow for transmission
1515 	 * and skip update in case it was old dfrag.
1516 	 */
1517 	if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1518 		msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1519 		msk->snd_nxt = snd_nxt_new;
1520 	}
1521 }
1522 
1523 void mptcp_check_and_set_pending(struct sock *sk)
1524 {
1525 	if (mptcp_send_head(sk)) {
1526 		mptcp_data_lock(sk);
1527 		mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1528 		mptcp_data_unlock(sk);
1529 	}
1530 }
1531 
1532 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1533 				  struct mptcp_sendmsg_info *info)
1534 {
1535 	struct mptcp_sock *msk = mptcp_sk(sk);
1536 	struct mptcp_data_frag *dfrag;
1537 	int len, copied = 0, err = 0;
1538 
1539 	while ((dfrag = mptcp_send_head(sk))) {
1540 		info->sent = dfrag->already_sent;
1541 		info->limit = dfrag->data_len;
1542 		len = dfrag->data_len - dfrag->already_sent;
1543 		while (len > 0) {
1544 			int ret = 0;
1545 
1546 			ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1547 			if (ret <= 0) {
1548 				err = copied ? : ret;
1549 				goto out;
1550 			}
1551 
1552 			info->sent += ret;
1553 			copied += ret;
1554 			len -= ret;
1555 
1556 			mptcp_update_post_push(msk, dfrag, ret);
1557 		}
1558 		WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1559 
1560 		if (msk->snd_burst <= 0 ||
1561 		    !sk_stream_memory_free(ssk) ||
1562 		    !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1563 			err = copied;
1564 			goto out;
1565 		}
1566 		mptcp_set_timeout(sk);
1567 	}
1568 	err = copied;
1569 
1570 out:
1571 	return err;
1572 }
1573 
1574 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1575 {
1576 	struct sock *prev_ssk = NULL, *ssk = NULL;
1577 	struct mptcp_sock *msk = mptcp_sk(sk);
1578 	struct mptcp_sendmsg_info info = {
1579 				.flags = flags,
1580 	};
1581 	bool do_check_data_fin = false;
1582 	int push_count = 1;
1583 
1584 	while (mptcp_send_head(sk) && (push_count > 0)) {
1585 		struct mptcp_subflow_context *subflow;
1586 		int ret = 0;
1587 
1588 		if (mptcp_sched_get_send(msk))
1589 			break;
1590 
1591 		push_count = 0;
1592 
1593 		mptcp_for_each_subflow(msk, subflow) {
1594 			if (READ_ONCE(subflow->scheduled)) {
1595 				mptcp_subflow_set_scheduled(subflow, false);
1596 
1597 				prev_ssk = ssk;
1598 				ssk = mptcp_subflow_tcp_sock(subflow);
1599 				if (ssk != prev_ssk) {
1600 					/* First check. If the ssk has changed since
1601 					 * the last round, release prev_ssk
1602 					 */
1603 					if (prev_ssk)
1604 						mptcp_push_release(prev_ssk, &info);
1605 
1606 					/* Need to lock the new subflow only if different
1607 					 * from the previous one, otherwise we are still
1608 					 * helding the relevant lock
1609 					 */
1610 					lock_sock(ssk);
1611 				}
1612 
1613 				push_count++;
1614 
1615 				ret = __subflow_push_pending(sk, ssk, &info);
1616 				if (ret <= 0) {
1617 					if (ret != -EAGAIN ||
1618 					    (1 << ssk->sk_state) &
1619 					     (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1620 						push_count--;
1621 					continue;
1622 				}
1623 				do_check_data_fin = true;
1624 			}
1625 		}
1626 	}
1627 
1628 	/* at this point we held the socket lock for the last subflow we used */
1629 	if (ssk)
1630 		mptcp_push_release(ssk, &info);
1631 
1632 	/* ensure the rtx timer is running */
1633 	if (!mptcp_rtx_timer_pending(sk))
1634 		mptcp_reset_rtx_timer(sk);
1635 	if (do_check_data_fin)
1636 		mptcp_check_send_data_fin(sk);
1637 }
1638 
1639 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1640 {
1641 	struct mptcp_sock *msk = mptcp_sk(sk);
1642 	struct mptcp_sendmsg_info info = {
1643 		.data_lock_held = true,
1644 	};
1645 	bool keep_pushing = true;
1646 	struct sock *xmit_ssk;
1647 	int copied = 0;
1648 
1649 	info.flags = 0;
1650 	while (mptcp_send_head(sk) && keep_pushing) {
1651 		struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1652 		int ret = 0;
1653 
1654 		/* check for a different subflow usage only after
1655 		 * spooling the first chunk of data
1656 		 */
1657 		if (first) {
1658 			mptcp_subflow_set_scheduled(subflow, false);
1659 			ret = __subflow_push_pending(sk, ssk, &info);
1660 			first = false;
1661 			if (ret <= 0)
1662 				break;
1663 			copied += ret;
1664 			continue;
1665 		}
1666 
1667 		if (mptcp_sched_get_send(msk))
1668 			goto out;
1669 
1670 		if (READ_ONCE(subflow->scheduled)) {
1671 			mptcp_subflow_set_scheduled(subflow, false);
1672 			ret = __subflow_push_pending(sk, ssk, &info);
1673 			if (ret <= 0)
1674 				keep_pushing = false;
1675 			copied += ret;
1676 		}
1677 
1678 		mptcp_for_each_subflow(msk, subflow) {
1679 			if (READ_ONCE(subflow->scheduled)) {
1680 				xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1681 				if (xmit_ssk != ssk) {
1682 					mptcp_subflow_delegate(subflow,
1683 							       MPTCP_DELEGATE_SEND);
1684 					keep_pushing = false;
1685 				}
1686 			}
1687 		}
1688 	}
1689 
1690 out:
1691 	/* __mptcp_alloc_tx_skb could have released some wmem and we are
1692 	 * not going to flush it via release_sock()
1693 	 */
1694 	if (copied) {
1695 		tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1696 			 info.size_goal);
1697 		if (!mptcp_rtx_timer_pending(sk))
1698 			mptcp_reset_rtx_timer(sk);
1699 
1700 		if (msk->snd_data_fin_enable &&
1701 		    msk->snd_nxt + 1 == msk->write_seq)
1702 			mptcp_schedule_work(sk);
1703 	}
1704 }
1705 
1706 static void mptcp_set_nospace(struct sock *sk)
1707 {
1708 	/* enable autotune */
1709 	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1710 
1711 	/* will be cleared on avail space */
1712 	set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1713 }
1714 
1715 static int mptcp_disconnect(struct sock *sk, int flags);
1716 
1717 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1718 				  size_t len, int *copied_syn)
1719 {
1720 	unsigned int saved_flags = msg->msg_flags;
1721 	struct mptcp_sock *msk = mptcp_sk(sk);
1722 	struct sock *ssk;
1723 	int ret;
1724 
1725 	/* on flags based fastopen the mptcp is supposed to create the
1726 	 * first subflow right now. Otherwise we are in the defer_connect
1727 	 * path, and the first subflow must be already present.
1728 	 * Since the defer_connect flag is cleared after the first succsful
1729 	 * fastopen attempt, no need to check for additional subflow status.
1730 	 */
1731 	if (msg->msg_flags & MSG_FASTOPEN) {
1732 		ssk = __mptcp_nmpc_sk(msk);
1733 		if (IS_ERR(ssk))
1734 			return PTR_ERR(ssk);
1735 	}
1736 	if (!msk->first)
1737 		return -EINVAL;
1738 
1739 	ssk = msk->first;
1740 
1741 	lock_sock(ssk);
1742 	msg->msg_flags |= MSG_DONTWAIT;
1743 	msk->fastopening = 1;
1744 	ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1745 	msk->fastopening = 0;
1746 	msg->msg_flags = saved_flags;
1747 	release_sock(ssk);
1748 
1749 	/* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1750 	if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1751 		ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1752 					    msg->msg_namelen, msg->msg_flags, 1);
1753 
1754 		/* Keep the same behaviour of plain TCP: zero the copied bytes in
1755 		 * case of any error, except timeout or signal
1756 		 */
1757 		if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1758 			*copied_syn = 0;
1759 	} else if (ret && ret != -EINPROGRESS) {
1760 		/* The disconnect() op called by tcp_sendmsg_fastopen()/
1761 		 * __inet_stream_connect() can fail, due to looking check,
1762 		 * see mptcp_disconnect().
1763 		 * Attempt it again outside the problematic scope.
1764 		 */
1765 		if (!mptcp_disconnect(sk, 0))
1766 			sk->sk_socket->state = SS_UNCONNECTED;
1767 	}
1768 	inet_clear_bit(DEFER_CONNECT, sk);
1769 
1770 	return ret;
1771 }
1772 
1773 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1774 {
1775 	struct mptcp_sock *msk = mptcp_sk(sk);
1776 	struct page_frag *pfrag;
1777 	size_t copied = 0;
1778 	int ret = 0;
1779 	long timeo;
1780 
1781 	/* silently ignore everything else */
1782 	msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1783 
1784 	lock_sock(sk);
1785 
1786 	if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1787 		     msg->msg_flags & MSG_FASTOPEN)) {
1788 		int copied_syn = 0;
1789 
1790 		ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1791 		copied += copied_syn;
1792 		if (ret == -EINPROGRESS && copied_syn > 0)
1793 			goto out;
1794 		else if (ret)
1795 			goto do_error;
1796 	}
1797 
1798 	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1799 
1800 	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1801 		ret = sk_stream_wait_connect(sk, &timeo);
1802 		if (ret)
1803 			goto do_error;
1804 	}
1805 
1806 	ret = -EPIPE;
1807 	if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1808 		goto do_error;
1809 
1810 	pfrag = sk_page_frag(sk);
1811 
1812 	while (msg_data_left(msg)) {
1813 		int total_ts, frag_truesize = 0;
1814 		struct mptcp_data_frag *dfrag;
1815 		bool dfrag_collapsed;
1816 		size_t psize, offset;
1817 
1818 		/* reuse tail pfrag, if possible, or carve a new one from the
1819 		 * page allocator
1820 		 */
1821 		dfrag = mptcp_pending_tail(sk);
1822 		dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1823 		if (!dfrag_collapsed) {
1824 			if (!sk_stream_memory_free(sk))
1825 				goto wait_for_memory;
1826 
1827 			if (!mptcp_page_frag_refill(sk, pfrag))
1828 				goto wait_for_memory;
1829 
1830 			dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1831 			frag_truesize = dfrag->overhead;
1832 		}
1833 
1834 		/* we do not bound vs wspace, to allow a single packet.
1835 		 * memory accounting will prevent execessive memory usage
1836 		 * anyway
1837 		 */
1838 		offset = dfrag->offset + dfrag->data_len;
1839 		psize = pfrag->size - offset;
1840 		psize = min_t(size_t, psize, msg_data_left(msg));
1841 		total_ts = psize + frag_truesize;
1842 
1843 		if (!sk_wmem_schedule(sk, total_ts))
1844 			goto wait_for_memory;
1845 
1846 		if (copy_page_from_iter(dfrag->page, offset, psize,
1847 					&msg->msg_iter) != psize) {
1848 			ret = -EFAULT;
1849 			goto do_error;
1850 		}
1851 
1852 		/* data successfully copied into the write queue */
1853 		sk_forward_alloc_add(sk, -total_ts);
1854 		copied += psize;
1855 		dfrag->data_len += psize;
1856 		frag_truesize += psize;
1857 		pfrag->offset += frag_truesize;
1858 		WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1859 
1860 		/* charge data on mptcp pending queue to the msk socket
1861 		 * Note: we charge such data both to sk and ssk
1862 		 */
1863 		sk_wmem_queued_add(sk, frag_truesize);
1864 		if (!dfrag_collapsed) {
1865 			get_page(dfrag->page);
1866 			list_add_tail(&dfrag->list, &msk->rtx_queue);
1867 			if (!msk->first_pending)
1868 				WRITE_ONCE(msk->first_pending, dfrag);
1869 		}
1870 		pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1871 			 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1872 			 !dfrag_collapsed);
1873 
1874 		continue;
1875 
1876 wait_for_memory:
1877 		mptcp_set_nospace(sk);
1878 		__mptcp_push_pending(sk, msg->msg_flags);
1879 		ret = sk_stream_wait_memory(sk, &timeo);
1880 		if (ret)
1881 			goto do_error;
1882 	}
1883 
1884 	if (copied)
1885 		__mptcp_push_pending(sk, msg->msg_flags);
1886 
1887 out:
1888 	release_sock(sk);
1889 	return copied;
1890 
1891 do_error:
1892 	if (copied)
1893 		goto out;
1894 
1895 	copied = sk_stream_error(sk, msg->msg_flags, ret);
1896 	goto out;
1897 }
1898 
1899 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1900 				struct msghdr *msg,
1901 				size_t len, int flags,
1902 				struct scm_timestamping_internal *tss,
1903 				int *cmsg_flags)
1904 {
1905 	struct sk_buff *skb, *tmp;
1906 	int copied = 0;
1907 
1908 	skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1909 		u32 offset = MPTCP_SKB_CB(skb)->offset;
1910 		u32 data_len = skb->len - offset;
1911 		u32 count = min_t(size_t, len - copied, data_len);
1912 		int err;
1913 
1914 		if (!(flags & MSG_TRUNC)) {
1915 			err = skb_copy_datagram_msg(skb, offset, msg, count);
1916 			if (unlikely(err < 0)) {
1917 				if (!copied)
1918 					return err;
1919 				break;
1920 			}
1921 		}
1922 
1923 		if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1924 			tcp_update_recv_tstamps(skb, tss);
1925 			*cmsg_flags |= MPTCP_CMSG_TS;
1926 		}
1927 
1928 		copied += count;
1929 
1930 		if (count < data_len) {
1931 			if (!(flags & MSG_PEEK)) {
1932 				MPTCP_SKB_CB(skb)->offset += count;
1933 				MPTCP_SKB_CB(skb)->map_seq += count;
1934 			}
1935 			break;
1936 		}
1937 
1938 		if (!(flags & MSG_PEEK)) {
1939 			/* we will bulk release the skb memory later */
1940 			skb->destructor = NULL;
1941 			WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1942 			__skb_unlink(skb, &msk->receive_queue);
1943 			__kfree_skb(skb);
1944 		}
1945 
1946 		if (copied >= len)
1947 			break;
1948 	}
1949 
1950 	return copied;
1951 }
1952 
1953 /* receive buffer autotuning.  See tcp_rcv_space_adjust for more information.
1954  *
1955  * Only difference: Use highest rtt estimate of the subflows in use.
1956  */
1957 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1958 {
1959 	struct mptcp_subflow_context *subflow;
1960 	struct sock *sk = (struct sock *)msk;
1961 	u8 scaling_ratio = U8_MAX;
1962 	u32 time, advmss = 1;
1963 	u64 rtt_us, mstamp;
1964 
1965 	msk_owned_by_me(msk);
1966 
1967 	if (copied <= 0)
1968 		return;
1969 
1970 	if (!msk->rcvspace_init)
1971 		mptcp_rcv_space_init(msk, msk->first);
1972 
1973 	msk->rcvq_space.copied += copied;
1974 
1975 	mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1976 	time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1977 
1978 	rtt_us = msk->rcvq_space.rtt_us;
1979 	if (rtt_us && time < (rtt_us >> 3))
1980 		return;
1981 
1982 	rtt_us = 0;
1983 	mptcp_for_each_subflow(msk, subflow) {
1984 		const struct tcp_sock *tp;
1985 		u64 sf_rtt_us;
1986 		u32 sf_advmss;
1987 
1988 		tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1989 
1990 		sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1991 		sf_advmss = READ_ONCE(tp->advmss);
1992 
1993 		rtt_us = max(sf_rtt_us, rtt_us);
1994 		advmss = max(sf_advmss, advmss);
1995 		scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1996 	}
1997 
1998 	msk->rcvq_space.rtt_us = rtt_us;
1999 	msk->scaling_ratio = scaling_ratio;
2000 	if (time < (rtt_us >> 3) || rtt_us == 0)
2001 		return;
2002 
2003 	if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2004 		goto new_measure;
2005 
2006 	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2007 	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2008 		u64 rcvwin, grow;
2009 		int rcvbuf;
2010 
2011 		rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2012 
2013 		grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2014 
2015 		do_div(grow, msk->rcvq_space.space);
2016 		rcvwin += (grow << 1);
2017 
2018 		rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2019 			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2020 
2021 		if (rcvbuf > sk->sk_rcvbuf) {
2022 			u32 window_clamp;
2023 
2024 			window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2025 			WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2026 
2027 			/* Make subflows follow along.  If we do not do this, we
2028 			 * get drops at subflow level if skbs can't be moved to
2029 			 * the mptcp rx queue fast enough (announced rcv_win can
2030 			 * exceed ssk->sk_rcvbuf).
2031 			 */
2032 			mptcp_for_each_subflow(msk, subflow) {
2033 				struct sock *ssk;
2034 				bool slow;
2035 
2036 				ssk = mptcp_subflow_tcp_sock(subflow);
2037 				slow = lock_sock_fast(ssk);
2038 				WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2039 				tcp_sk(ssk)->window_clamp = window_clamp;
2040 				tcp_cleanup_rbuf(ssk, 1);
2041 				unlock_sock_fast(ssk, slow);
2042 			}
2043 		}
2044 	}
2045 
2046 	msk->rcvq_space.space = msk->rcvq_space.copied;
2047 new_measure:
2048 	msk->rcvq_space.copied = 0;
2049 	msk->rcvq_space.time = mstamp;
2050 }
2051 
2052 static void __mptcp_update_rmem(struct sock *sk)
2053 {
2054 	struct mptcp_sock *msk = mptcp_sk(sk);
2055 
2056 	if (!msk->rmem_released)
2057 		return;
2058 
2059 	atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2060 	mptcp_rmem_uncharge(sk, msk->rmem_released);
2061 	WRITE_ONCE(msk->rmem_released, 0);
2062 }
2063 
2064 static void __mptcp_splice_receive_queue(struct sock *sk)
2065 {
2066 	struct mptcp_sock *msk = mptcp_sk(sk);
2067 
2068 	skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2069 }
2070 
2071 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2072 {
2073 	struct sock *sk = (struct sock *)msk;
2074 	unsigned int moved = 0;
2075 	bool ret, done;
2076 
2077 	do {
2078 		struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2079 		bool slowpath;
2080 
2081 		/* we can have data pending in the subflows only if the msk
2082 		 * receive buffer was full at subflow_data_ready() time,
2083 		 * that is an unlikely slow path.
2084 		 */
2085 		if (likely(!ssk))
2086 			break;
2087 
2088 		slowpath = lock_sock_fast(ssk);
2089 		mptcp_data_lock(sk);
2090 		__mptcp_update_rmem(sk);
2091 		done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2092 		mptcp_data_unlock(sk);
2093 
2094 		if (unlikely(ssk->sk_err))
2095 			__mptcp_error_report(sk);
2096 		unlock_sock_fast(ssk, slowpath);
2097 	} while (!done);
2098 
2099 	/* acquire the data lock only if some input data is pending */
2100 	ret = moved > 0;
2101 	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2102 	    !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2103 		mptcp_data_lock(sk);
2104 		__mptcp_update_rmem(sk);
2105 		ret |= __mptcp_ofo_queue(msk);
2106 		__mptcp_splice_receive_queue(sk);
2107 		mptcp_data_unlock(sk);
2108 	}
2109 	if (ret)
2110 		mptcp_check_data_fin((struct sock *)msk);
2111 	return !skb_queue_empty(&msk->receive_queue);
2112 }
2113 
2114 static unsigned int mptcp_inq_hint(const struct sock *sk)
2115 {
2116 	const struct mptcp_sock *msk = mptcp_sk(sk);
2117 	const struct sk_buff *skb;
2118 
2119 	skb = skb_peek(&msk->receive_queue);
2120 	if (skb) {
2121 		u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2122 
2123 		if (hint_val >= INT_MAX)
2124 			return INT_MAX;
2125 
2126 		return (unsigned int)hint_val;
2127 	}
2128 
2129 	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2130 		return 1;
2131 
2132 	return 0;
2133 }
2134 
2135 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2136 			 int flags, int *addr_len)
2137 {
2138 	struct mptcp_sock *msk = mptcp_sk(sk);
2139 	struct scm_timestamping_internal tss;
2140 	int copied = 0, cmsg_flags = 0;
2141 	int target;
2142 	long timeo;
2143 
2144 	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2145 	if (unlikely(flags & MSG_ERRQUEUE))
2146 		return inet_recv_error(sk, msg, len, addr_len);
2147 
2148 	lock_sock(sk);
2149 	if (unlikely(sk->sk_state == TCP_LISTEN)) {
2150 		copied = -ENOTCONN;
2151 		goto out_err;
2152 	}
2153 
2154 	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2155 
2156 	len = min_t(size_t, len, INT_MAX);
2157 	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2158 
2159 	if (unlikely(msk->recvmsg_inq))
2160 		cmsg_flags = MPTCP_CMSG_INQ;
2161 
2162 	while (copied < len) {
2163 		int bytes_read;
2164 
2165 		bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2166 		if (unlikely(bytes_read < 0)) {
2167 			if (!copied)
2168 				copied = bytes_read;
2169 			goto out_err;
2170 		}
2171 
2172 		copied += bytes_read;
2173 
2174 		/* be sure to advertise window change */
2175 		mptcp_cleanup_rbuf(msk);
2176 
2177 		if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2178 			continue;
2179 
2180 		/* only the master socket status is relevant here. The exit
2181 		 * conditions mirror closely tcp_recvmsg()
2182 		 */
2183 		if (copied >= target)
2184 			break;
2185 
2186 		if (copied) {
2187 			if (sk->sk_err ||
2188 			    sk->sk_state == TCP_CLOSE ||
2189 			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2190 			    !timeo ||
2191 			    signal_pending(current))
2192 				break;
2193 		} else {
2194 			if (sk->sk_err) {
2195 				copied = sock_error(sk);
2196 				break;
2197 			}
2198 
2199 			if (sk->sk_shutdown & RCV_SHUTDOWN) {
2200 				/* race breaker: the shutdown could be after the
2201 				 * previous receive queue check
2202 				 */
2203 				if (__mptcp_move_skbs(msk))
2204 					continue;
2205 				break;
2206 			}
2207 
2208 			if (sk->sk_state == TCP_CLOSE) {
2209 				copied = -ENOTCONN;
2210 				break;
2211 			}
2212 
2213 			if (!timeo) {
2214 				copied = -EAGAIN;
2215 				break;
2216 			}
2217 
2218 			if (signal_pending(current)) {
2219 				copied = sock_intr_errno(timeo);
2220 				break;
2221 			}
2222 		}
2223 
2224 		pr_debug("block timeout %ld", timeo);
2225 		sk_wait_data(sk, &timeo, NULL);
2226 	}
2227 
2228 out_err:
2229 	if (cmsg_flags && copied >= 0) {
2230 		if (cmsg_flags & MPTCP_CMSG_TS)
2231 			tcp_recv_timestamp(msg, sk, &tss);
2232 
2233 		if (cmsg_flags & MPTCP_CMSG_INQ) {
2234 			unsigned int inq = mptcp_inq_hint(sk);
2235 
2236 			put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2237 		}
2238 	}
2239 
2240 	pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2241 		 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2242 		 skb_queue_empty(&msk->receive_queue), copied);
2243 	if (!(flags & MSG_PEEK))
2244 		mptcp_rcv_space_adjust(msk, copied);
2245 
2246 	release_sock(sk);
2247 	return copied;
2248 }
2249 
2250 static void mptcp_retransmit_timer(struct timer_list *t)
2251 {
2252 	struct inet_connection_sock *icsk = from_timer(icsk, t,
2253 						       icsk_retransmit_timer);
2254 	struct sock *sk = &icsk->icsk_inet.sk;
2255 	struct mptcp_sock *msk = mptcp_sk(sk);
2256 
2257 	bh_lock_sock(sk);
2258 	if (!sock_owned_by_user(sk)) {
2259 		/* we need a process context to retransmit */
2260 		if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2261 			mptcp_schedule_work(sk);
2262 	} else {
2263 		/* delegate our work to tcp_release_cb() */
2264 		__set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2265 	}
2266 	bh_unlock_sock(sk);
2267 	sock_put(sk);
2268 }
2269 
2270 static void mptcp_tout_timer(struct timer_list *t)
2271 {
2272 	struct sock *sk = from_timer(sk, t, sk_timer);
2273 
2274 	mptcp_schedule_work(sk);
2275 	sock_put(sk);
2276 }
2277 
2278 /* Find an idle subflow.  Return NULL if there is unacked data at tcp
2279  * level.
2280  *
2281  * A backup subflow is returned only if that is the only kind available.
2282  */
2283 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2284 {
2285 	struct sock *backup = NULL, *pick = NULL;
2286 	struct mptcp_subflow_context *subflow;
2287 	int min_stale_count = INT_MAX;
2288 
2289 	mptcp_for_each_subflow(msk, subflow) {
2290 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2291 
2292 		if (!__mptcp_subflow_active(subflow))
2293 			continue;
2294 
2295 		/* still data outstanding at TCP level? skip this */
2296 		if (!tcp_rtx_and_write_queues_empty(ssk)) {
2297 			mptcp_pm_subflow_chk_stale(msk, ssk);
2298 			min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2299 			continue;
2300 		}
2301 
2302 		if (subflow->backup) {
2303 			if (!backup)
2304 				backup = ssk;
2305 			continue;
2306 		}
2307 
2308 		if (!pick)
2309 			pick = ssk;
2310 	}
2311 
2312 	if (pick)
2313 		return pick;
2314 
2315 	/* use backup only if there are no progresses anywhere */
2316 	return min_stale_count > 1 ? backup : NULL;
2317 }
2318 
2319 bool __mptcp_retransmit_pending_data(struct sock *sk)
2320 {
2321 	struct mptcp_data_frag *cur, *rtx_head;
2322 	struct mptcp_sock *msk = mptcp_sk(sk);
2323 
2324 	if (__mptcp_check_fallback(msk))
2325 		return false;
2326 
2327 	/* the closing socket has some data untransmitted and/or unacked:
2328 	 * some data in the mptcp rtx queue has not really xmitted yet.
2329 	 * keep it simple and re-inject the whole mptcp level rtx queue
2330 	 */
2331 	mptcp_data_lock(sk);
2332 	__mptcp_clean_una_wakeup(sk);
2333 	rtx_head = mptcp_rtx_head(sk);
2334 	if (!rtx_head) {
2335 		mptcp_data_unlock(sk);
2336 		return false;
2337 	}
2338 
2339 	msk->recovery_snd_nxt = msk->snd_nxt;
2340 	msk->recovery = true;
2341 	mptcp_data_unlock(sk);
2342 
2343 	msk->first_pending = rtx_head;
2344 	msk->snd_burst = 0;
2345 
2346 	/* be sure to clear the "sent status" on all re-injected fragments */
2347 	list_for_each_entry(cur, &msk->rtx_queue, list) {
2348 		if (!cur->already_sent)
2349 			break;
2350 		cur->already_sent = 0;
2351 	}
2352 
2353 	return true;
2354 }
2355 
2356 /* flags for __mptcp_close_ssk() */
2357 #define MPTCP_CF_PUSH		BIT(1)
2358 #define MPTCP_CF_FASTCLOSE	BIT(2)
2359 
2360 /* be sure to send a reset only if the caller asked for it, also
2361  * clean completely the subflow status when the subflow reaches
2362  * TCP_CLOSE state
2363  */
2364 static void __mptcp_subflow_disconnect(struct sock *ssk,
2365 				       struct mptcp_subflow_context *subflow,
2366 				       unsigned int flags)
2367 {
2368 	if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2369 	    (flags & MPTCP_CF_FASTCLOSE)) {
2370 		/* The MPTCP code never wait on the subflow sockets, TCP-level
2371 		 * disconnect should never fail
2372 		 */
2373 		WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2374 		mptcp_subflow_ctx_reset(subflow);
2375 	} else {
2376 		tcp_shutdown(ssk, SEND_SHUTDOWN);
2377 	}
2378 }
2379 
2380 /* subflow sockets can be either outgoing (connect) or incoming
2381  * (accept).
2382  *
2383  * Outgoing subflows use in-kernel sockets.
2384  * Incoming subflows do not have their own 'struct socket' allocated,
2385  * so we need to use tcp_close() after detaching them from the mptcp
2386  * parent socket.
2387  */
2388 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2389 			      struct mptcp_subflow_context *subflow,
2390 			      unsigned int flags)
2391 {
2392 	struct mptcp_sock *msk = mptcp_sk(sk);
2393 	bool dispose_it, need_push = false;
2394 
2395 	/* If the first subflow moved to a close state before accept, e.g. due
2396 	 * to an incoming reset or listener shutdown, the subflow socket is
2397 	 * already deleted by inet_child_forget() and the mptcp socket can't
2398 	 * survive too.
2399 	 */
2400 	if (msk->in_accept_queue && msk->first == ssk &&
2401 	    (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2402 		/* ensure later check in mptcp_worker() will dispose the msk */
2403 		mptcp_set_close_tout(sk, tcp_jiffies32 - (TCP_TIMEWAIT_LEN + 1));
2404 		sock_set_flag(sk, SOCK_DEAD);
2405 		lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2406 		mptcp_subflow_drop_ctx(ssk);
2407 		goto out_release;
2408 	}
2409 
2410 	dispose_it = msk->free_first || ssk != msk->first;
2411 	if (dispose_it)
2412 		list_del(&subflow->node);
2413 
2414 	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2415 
2416 	if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2417 		/* be sure to force the tcp_close path
2418 		 * to generate the egress reset
2419 		 */
2420 		ssk->sk_lingertime = 0;
2421 		sock_set_flag(ssk, SOCK_LINGER);
2422 		subflow->send_fastclose = 1;
2423 	}
2424 
2425 	need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2426 	if (!dispose_it) {
2427 		__mptcp_subflow_disconnect(ssk, subflow, flags);
2428 		release_sock(ssk);
2429 
2430 		goto out;
2431 	}
2432 
2433 	subflow->disposable = 1;
2434 
2435 	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2436 	 * the ssk has been already destroyed, we just need to release the
2437 	 * reference owned by msk;
2438 	 */
2439 	if (!inet_csk(ssk)->icsk_ulp_ops) {
2440 		WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2441 		kfree_rcu(subflow, rcu);
2442 	} else {
2443 		/* otherwise tcp will dispose of the ssk and subflow ctx */
2444 		__tcp_close(ssk, 0);
2445 
2446 		/* close acquired an extra ref */
2447 		__sock_put(ssk);
2448 	}
2449 
2450 out_release:
2451 	__mptcp_subflow_error_report(sk, ssk);
2452 	release_sock(ssk);
2453 
2454 	sock_put(ssk);
2455 
2456 	if (ssk == msk->first)
2457 		WRITE_ONCE(msk->first, NULL);
2458 
2459 out:
2460 	__mptcp_sync_sndbuf(sk);
2461 	if (need_push)
2462 		__mptcp_push_pending(sk, 0);
2463 
2464 	/* Catch every 'all subflows closed' scenario, including peers silently
2465 	 * closing them, e.g. due to timeout.
2466 	 * For established sockets, allow an additional timeout before closing,
2467 	 * as the protocol can still create more subflows.
2468 	 */
2469 	if (list_is_singular(&msk->conn_list) && msk->first &&
2470 	    inet_sk_state_load(msk->first) == TCP_CLOSE) {
2471 		if (sk->sk_state != TCP_ESTABLISHED ||
2472 		    msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2473 			inet_sk_state_store(sk, TCP_CLOSE);
2474 			mptcp_close_wake_up(sk);
2475 		} else {
2476 			mptcp_start_tout_timer(sk);
2477 		}
2478 	}
2479 }
2480 
2481 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2482 		     struct mptcp_subflow_context *subflow)
2483 {
2484 	if (sk->sk_state == TCP_ESTABLISHED)
2485 		mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2486 
2487 	/* subflow aborted before reaching the fully_established status
2488 	 * attempt the creation of the next subflow
2489 	 */
2490 	mptcp_pm_subflow_check_next(mptcp_sk(sk), ssk, subflow);
2491 
2492 	__mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2493 }
2494 
2495 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2496 {
2497 	return 0;
2498 }
2499 
2500 static void __mptcp_close_subflow(struct sock *sk)
2501 {
2502 	struct mptcp_subflow_context *subflow, *tmp;
2503 	struct mptcp_sock *msk = mptcp_sk(sk);
2504 
2505 	might_sleep();
2506 
2507 	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2508 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2509 
2510 		if (inet_sk_state_load(ssk) != TCP_CLOSE)
2511 			continue;
2512 
2513 		/* 'subflow_data_ready' will re-sched once rx queue is empty */
2514 		if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2515 			continue;
2516 
2517 		mptcp_close_ssk(sk, ssk, subflow);
2518 	}
2519 
2520 }
2521 
2522 static bool mptcp_close_tout_expired(const struct sock *sk)
2523 {
2524 	if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2525 	    sk->sk_state == TCP_CLOSE)
2526 		return false;
2527 
2528 	return time_after32(tcp_jiffies32,
2529 		  inet_csk(sk)->icsk_mtup.probe_timestamp + TCP_TIMEWAIT_LEN);
2530 }
2531 
2532 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2533 {
2534 	struct mptcp_subflow_context *subflow, *tmp;
2535 	struct sock *sk = (struct sock *)msk;
2536 
2537 	if (likely(!READ_ONCE(msk->rcv_fastclose)))
2538 		return;
2539 
2540 	mptcp_token_destroy(msk);
2541 
2542 	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2543 		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2544 		bool slow;
2545 
2546 		slow = lock_sock_fast(tcp_sk);
2547 		if (tcp_sk->sk_state != TCP_CLOSE) {
2548 			tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2549 			tcp_set_state(tcp_sk, TCP_CLOSE);
2550 		}
2551 		unlock_sock_fast(tcp_sk, slow);
2552 	}
2553 
2554 	/* Mirror the tcp_reset() error propagation */
2555 	switch (sk->sk_state) {
2556 	case TCP_SYN_SENT:
2557 		WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2558 		break;
2559 	case TCP_CLOSE_WAIT:
2560 		WRITE_ONCE(sk->sk_err, EPIPE);
2561 		break;
2562 	case TCP_CLOSE:
2563 		return;
2564 	default:
2565 		WRITE_ONCE(sk->sk_err, ECONNRESET);
2566 	}
2567 
2568 	inet_sk_state_store(sk, TCP_CLOSE);
2569 	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2570 	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2571 	set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2572 
2573 	/* the calling mptcp_worker will properly destroy the socket */
2574 	if (sock_flag(sk, SOCK_DEAD))
2575 		return;
2576 
2577 	sk->sk_state_change(sk);
2578 	sk_error_report(sk);
2579 }
2580 
2581 static void __mptcp_retrans(struct sock *sk)
2582 {
2583 	struct mptcp_sock *msk = mptcp_sk(sk);
2584 	struct mptcp_subflow_context *subflow;
2585 	struct mptcp_sendmsg_info info = {};
2586 	struct mptcp_data_frag *dfrag;
2587 	struct sock *ssk;
2588 	int ret, err;
2589 	u16 len = 0;
2590 
2591 	mptcp_clean_una_wakeup(sk);
2592 
2593 	/* first check ssk: need to kick "stale" logic */
2594 	err = mptcp_sched_get_retrans(msk);
2595 	dfrag = mptcp_rtx_head(sk);
2596 	if (!dfrag) {
2597 		if (mptcp_data_fin_enabled(msk)) {
2598 			struct inet_connection_sock *icsk = inet_csk(sk);
2599 
2600 			icsk->icsk_retransmits++;
2601 			mptcp_set_datafin_timeout(sk);
2602 			mptcp_send_ack(msk);
2603 
2604 			goto reset_timer;
2605 		}
2606 
2607 		if (!mptcp_send_head(sk))
2608 			return;
2609 
2610 		goto reset_timer;
2611 	}
2612 
2613 	if (err)
2614 		goto reset_timer;
2615 
2616 	mptcp_for_each_subflow(msk, subflow) {
2617 		if (READ_ONCE(subflow->scheduled)) {
2618 			u16 copied = 0;
2619 
2620 			mptcp_subflow_set_scheduled(subflow, false);
2621 
2622 			ssk = mptcp_subflow_tcp_sock(subflow);
2623 
2624 			lock_sock(ssk);
2625 
2626 			/* limit retransmission to the bytes already sent on some subflows */
2627 			info.sent = 0;
2628 			info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2629 								    dfrag->already_sent;
2630 			while (info.sent < info.limit) {
2631 				ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2632 				if (ret <= 0)
2633 					break;
2634 
2635 				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2636 				copied += ret;
2637 				info.sent += ret;
2638 			}
2639 			if (copied) {
2640 				len = max(copied, len);
2641 				tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2642 					 info.size_goal);
2643 				WRITE_ONCE(msk->allow_infinite_fallback, false);
2644 			}
2645 
2646 			release_sock(ssk);
2647 		}
2648 	}
2649 
2650 	msk->bytes_retrans += len;
2651 	dfrag->already_sent = max(dfrag->already_sent, len);
2652 
2653 reset_timer:
2654 	mptcp_check_and_set_pending(sk);
2655 
2656 	if (!mptcp_rtx_timer_pending(sk))
2657 		mptcp_reset_rtx_timer(sk);
2658 }
2659 
2660 /* schedule the timeout timer for the relevant event: either close timeout
2661  * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2662  */
2663 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2664 {
2665 	struct sock *sk = (struct sock *)msk;
2666 	unsigned long timeout, close_timeout;
2667 
2668 	if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2669 		return;
2670 
2671 	close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2672 			TCP_TIMEWAIT_LEN;
2673 
2674 	/* the close timeout takes precedence on the fail one, and here at least one of
2675 	 * them is active
2676 	 */
2677 	timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2678 
2679 	sk_reset_timer(sk, &sk->sk_timer, timeout);
2680 }
2681 
2682 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2683 {
2684 	struct sock *ssk = msk->first;
2685 	bool slow;
2686 
2687 	if (!ssk)
2688 		return;
2689 
2690 	pr_debug("MP_FAIL doesn't respond, reset the subflow");
2691 
2692 	slow = lock_sock_fast(ssk);
2693 	mptcp_subflow_reset(ssk);
2694 	WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2695 	unlock_sock_fast(ssk, slow);
2696 }
2697 
2698 static void mptcp_do_fastclose(struct sock *sk)
2699 {
2700 	struct mptcp_subflow_context *subflow, *tmp;
2701 	struct mptcp_sock *msk = mptcp_sk(sk);
2702 
2703 	inet_sk_state_store(sk, TCP_CLOSE);
2704 	mptcp_for_each_subflow_safe(msk, subflow, tmp)
2705 		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2706 				  subflow, MPTCP_CF_FASTCLOSE);
2707 }
2708 
2709 static void mptcp_worker(struct work_struct *work)
2710 {
2711 	struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2712 	struct sock *sk = (struct sock *)msk;
2713 	unsigned long fail_tout;
2714 	int state;
2715 
2716 	lock_sock(sk);
2717 	state = sk->sk_state;
2718 	if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2719 		goto unlock;
2720 
2721 	mptcp_check_fastclose(msk);
2722 
2723 	mptcp_pm_nl_work(msk);
2724 
2725 	mptcp_check_send_data_fin(sk);
2726 	mptcp_check_data_fin_ack(sk);
2727 	mptcp_check_data_fin(sk);
2728 
2729 	if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2730 		__mptcp_close_subflow(sk);
2731 
2732 	if (mptcp_close_tout_expired(sk)) {
2733 		mptcp_do_fastclose(sk);
2734 		mptcp_close_wake_up(sk);
2735 	}
2736 
2737 	if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2738 		__mptcp_destroy_sock(sk);
2739 		goto unlock;
2740 	}
2741 
2742 	if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2743 		__mptcp_retrans(sk);
2744 
2745 	fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2746 	if (fail_tout && time_after(jiffies, fail_tout))
2747 		mptcp_mp_fail_no_response(msk);
2748 
2749 unlock:
2750 	release_sock(sk);
2751 	sock_put(sk);
2752 }
2753 
2754 static void __mptcp_init_sock(struct sock *sk)
2755 {
2756 	struct mptcp_sock *msk = mptcp_sk(sk);
2757 
2758 	INIT_LIST_HEAD(&msk->conn_list);
2759 	INIT_LIST_HEAD(&msk->join_list);
2760 	INIT_LIST_HEAD(&msk->rtx_queue);
2761 	INIT_WORK(&msk->work, mptcp_worker);
2762 	__skb_queue_head_init(&msk->receive_queue);
2763 	msk->out_of_order_queue = RB_ROOT;
2764 	msk->first_pending = NULL;
2765 	msk->rmem_fwd_alloc = 0;
2766 	WRITE_ONCE(msk->rmem_released, 0);
2767 	msk->timer_ival = TCP_RTO_MIN;
2768 
2769 	WRITE_ONCE(msk->first, NULL);
2770 	inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2771 	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2772 	WRITE_ONCE(msk->allow_infinite_fallback, true);
2773 	msk->recovery = false;
2774 	msk->subflow_id = 1;
2775 
2776 	mptcp_pm_data_init(msk);
2777 
2778 	/* re-use the csk retrans timer for MPTCP-level retrans */
2779 	timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2780 	timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2781 }
2782 
2783 static void mptcp_ca_reset(struct sock *sk)
2784 {
2785 	struct inet_connection_sock *icsk = inet_csk(sk);
2786 
2787 	tcp_assign_congestion_control(sk);
2788 	strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2789 
2790 	/* no need to keep a reference to the ops, the name will suffice */
2791 	tcp_cleanup_congestion_control(sk);
2792 	icsk->icsk_ca_ops = NULL;
2793 }
2794 
2795 static int mptcp_init_sock(struct sock *sk)
2796 {
2797 	struct net *net = sock_net(sk);
2798 	int ret;
2799 
2800 	__mptcp_init_sock(sk);
2801 
2802 	if (!mptcp_is_enabled(net))
2803 		return -ENOPROTOOPT;
2804 
2805 	if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2806 		return -ENOMEM;
2807 
2808 	ret = mptcp_init_sched(mptcp_sk(sk),
2809 			       mptcp_sched_find(mptcp_get_scheduler(net)));
2810 	if (ret)
2811 		return ret;
2812 
2813 	set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2814 
2815 	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2816 	 * propagate the correct value
2817 	 */
2818 	mptcp_ca_reset(sk);
2819 
2820 	sk_sockets_allocated_inc(sk);
2821 	sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2822 	sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2823 
2824 	return 0;
2825 }
2826 
2827 static void __mptcp_clear_xmit(struct sock *sk)
2828 {
2829 	struct mptcp_sock *msk = mptcp_sk(sk);
2830 	struct mptcp_data_frag *dtmp, *dfrag;
2831 
2832 	WRITE_ONCE(msk->first_pending, NULL);
2833 	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2834 		dfrag_clear(sk, dfrag);
2835 }
2836 
2837 void mptcp_cancel_work(struct sock *sk)
2838 {
2839 	struct mptcp_sock *msk = mptcp_sk(sk);
2840 
2841 	if (cancel_work_sync(&msk->work))
2842 		__sock_put(sk);
2843 }
2844 
2845 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2846 {
2847 	lock_sock(ssk);
2848 
2849 	switch (ssk->sk_state) {
2850 	case TCP_LISTEN:
2851 		if (!(how & RCV_SHUTDOWN))
2852 			break;
2853 		fallthrough;
2854 	case TCP_SYN_SENT:
2855 		WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2856 		break;
2857 	default:
2858 		if (__mptcp_check_fallback(mptcp_sk(sk))) {
2859 			pr_debug("Fallback");
2860 			ssk->sk_shutdown |= how;
2861 			tcp_shutdown(ssk, how);
2862 
2863 			/* simulate the data_fin ack reception to let the state
2864 			 * machine move forward
2865 			 */
2866 			WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2867 			mptcp_schedule_work(sk);
2868 		} else {
2869 			pr_debug("Sending DATA_FIN on subflow %p", ssk);
2870 			tcp_send_ack(ssk);
2871 			if (!mptcp_rtx_timer_pending(sk))
2872 				mptcp_reset_rtx_timer(sk);
2873 		}
2874 		break;
2875 	}
2876 
2877 	release_sock(ssk);
2878 }
2879 
2880 static const unsigned char new_state[16] = {
2881 	/* current state:     new state:      action:	*/
2882 	[0 /* (Invalid) */] = TCP_CLOSE,
2883 	[TCP_ESTABLISHED]   = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2884 	[TCP_SYN_SENT]      = TCP_CLOSE,
2885 	[TCP_SYN_RECV]      = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2886 	[TCP_FIN_WAIT1]     = TCP_FIN_WAIT1,
2887 	[TCP_FIN_WAIT2]     = TCP_FIN_WAIT2,
2888 	[TCP_TIME_WAIT]     = TCP_CLOSE,	/* should not happen ! */
2889 	[TCP_CLOSE]         = TCP_CLOSE,
2890 	[TCP_CLOSE_WAIT]    = TCP_LAST_ACK  | TCP_ACTION_FIN,
2891 	[TCP_LAST_ACK]      = TCP_LAST_ACK,
2892 	[TCP_LISTEN]        = TCP_CLOSE,
2893 	[TCP_CLOSING]       = TCP_CLOSING,
2894 	[TCP_NEW_SYN_RECV]  = TCP_CLOSE,	/* should not happen ! */
2895 };
2896 
2897 static int mptcp_close_state(struct sock *sk)
2898 {
2899 	int next = (int)new_state[sk->sk_state];
2900 	int ns = next & TCP_STATE_MASK;
2901 
2902 	inet_sk_state_store(sk, ns);
2903 
2904 	return next & TCP_ACTION_FIN;
2905 }
2906 
2907 static void mptcp_check_send_data_fin(struct sock *sk)
2908 {
2909 	struct mptcp_subflow_context *subflow;
2910 	struct mptcp_sock *msk = mptcp_sk(sk);
2911 
2912 	pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2913 		 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2914 		 msk->snd_nxt, msk->write_seq);
2915 
2916 	/* we still need to enqueue subflows or not really shutting down,
2917 	 * skip this
2918 	 */
2919 	if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2920 	    mptcp_send_head(sk))
2921 		return;
2922 
2923 	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2924 
2925 	mptcp_for_each_subflow(msk, subflow) {
2926 		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2927 
2928 		mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2929 	}
2930 }
2931 
2932 static void __mptcp_wr_shutdown(struct sock *sk)
2933 {
2934 	struct mptcp_sock *msk = mptcp_sk(sk);
2935 
2936 	pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2937 		 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2938 		 !!mptcp_send_head(sk));
2939 
2940 	/* will be ignored by fallback sockets */
2941 	WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2942 	WRITE_ONCE(msk->snd_data_fin_enable, 1);
2943 
2944 	mptcp_check_send_data_fin(sk);
2945 }
2946 
2947 static void __mptcp_destroy_sock(struct sock *sk)
2948 {
2949 	struct mptcp_sock *msk = mptcp_sk(sk);
2950 
2951 	pr_debug("msk=%p", msk);
2952 
2953 	might_sleep();
2954 
2955 	mptcp_stop_rtx_timer(sk);
2956 	sk_stop_timer(sk, &sk->sk_timer);
2957 	msk->pm.status = 0;
2958 	mptcp_release_sched(msk);
2959 
2960 	sk->sk_prot->destroy(sk);
2961 
2962 	WARN_ON_ONCE(msk->rmem_fwd_alloc);
2963 	WARN_ON_ONCE(msk->rmem_released);
2964 	sk_stream_kill_queues(sk);
2965 	xfrm_sk_free_policy(sk);
2966 
2967 	sock_put(sk);
2968 }
2969 
2970 void __mptcp_unaccepted_force_close(struct sock *sk)
2971 {
2972 	sock_set_flag(sk, SOCK_DEAD);
2973 	mptcp_do_fastclose(sk);
2974 	__mptcp_destroy_sock(sk);
2975 }
2976 
2977 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2978 {
2979 	/* Concurrent splices from sk_receive_queue into receive_queue will
2980 	 * always show at least one non-empty queue when checked in this order.
2981 	 */
2982 	if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
2983 	    skb_queue_empty_lockless(&msk->receive_queue))
2984 		return 0;
2985 
2986 	return EPOLLIN | EPOLLRDNORM;
2987 }
2988 
2989 static void mptcp_check_listen_stop(struct sock *sk)
2990 {
2991 	struct sock *ssk;
2992 
2993 	if (inet_sk_state_load(sk) != TCP_LISTEN)
2994 		return;
2995 
2996 	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
2997 	ssk = mptcp_sk(sk)->first;
2998 	if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
2999 		return;
3000 
3001 	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3002 	tcp_set_state(ssk, TCP_CLOSE);
3003 	mptcp_subflow_queue_clean(sk, ssk);
3004 	inet_csk_listen_stop(ssk);
3005 	mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3006 	release_sock(ssk);
3007 }
3008 
3009 bool __mptcp_close(struct sock *sk, long timeout)
3010 {
3011 	struct mptcp_subflow_context *subflow;
3012 	struct mptcp_sock *msk = mptcp_sk(sk);
3013 	bool do_cancel_work = false;
3014 	int subflows_alive = 0;
3015 
3016 	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3017 
3018 	if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3019 		mptcp_check_listen_stop(sk);
3020 		inet_sk_state_store(sk, TCP_CLOSE);
3021 		goto cleanup;
3022 	}
3023 
3024 	if (mptcp_check_readable(msk) || timeout < 0) {
3025 		/* If the msk has read data, or the caller explicitly ask it,
3026 		 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3027 		 */
3028 		mptcp_do_fastclose(sk);
3029 		timeout = 0;
3030 	} else if (mptcp_close_state(sk)) {
3031 		__mptcp_wr_shutdown(sk);
3032 	}
3033 
3034 	sk_stream_wait_close(sk, timeout);
3035 
3036 cleanup:
3037 	/* orphan all the subflows */
3038 	mptcp_for_each_subflow(msk, subflow) {
3039 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3040 		bool slow = lock_sock_fast_nested(ssk);
3041 
3042 		subflows_alive += ssk->sk_state != TCP_CLOSE;
3043 
3044 		/* since the close timeout takes precedence on the fail one,
3045 		 * cancel the latter
3046 		 */
3047 		if (ssk == msk->first)
3048 			subflow->fail_tout = 0;
3049 
3050 		/* detach from the parent socket, but allow data_ready to
3051 		 * push incoming data into the mptcp stack, to properly ack it
3052 		 */
3053 		ssk->sk_socket = NULL;
3054 		ssk->sk_wq = NULL;
3055 		unlock_sock_fast(ssk, slow);
3056 	}
3057 	sock_orphan(sk);
3058 
3059 	/* all the subflows are closed, only timeout can change the msk
3060 	 * state, let's not keep resources busy for no reasons
3061 	 */
3062 	if (subflows_alive == 0)
3063 		inet_sk_state_store(sk, TCP_CLOSE);
3064 
3065 	sock_hold(sk);
3066 	pr_debug("msk=%p state=%d", sk, sk->sk_state);
3067 	if (msk->token)
3068 		mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3069 
3070 	if (sk->sk_state == TCP_CLOSE) {
3071 		__mptcp_destroy_sock(sk);
3072 		do_cancel_work = true;
3073 	} else {
3074 		mptcp_start_tout_timer(sk);
3075 	}
3076 
3077 	return do_cancel_work;
3078 }
3079 
3080 static void mptcp_close(struct sock *sk, long timeout)
3081 {
3082 	bool do_cancel_work;
3083 
3084 	lock_sock(sk);
3085 
3086 	do_cancel_work = __mptcp_close(sk, timeout);
3087 	release_sock(sk);
3088 	if (do_cancel_work)
3089 		mptcp_cancel_work(sk);
3090 
3091 	sock_put(sk);
3092 }
3093 
3094 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3095 {
3096 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3097 	const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3098 	struct ipv6_pinfo *msk6 = inet6_sk(msk);
3099 
3100 	msk->sk_v6_daddr = ssk->sk_v6_daddr;
3101 	msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3102 
3103 	if (msk6 && ssk6) {
3104 		msk6->saddr = ssk6->saddr;
3105 		msk6->flow_label = ssk6->flow_label;
3106 	}
3107 #endif
3108 
3109 	inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3110 	inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3111 	inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3112 	inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3113 	inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3114 	inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3115 }
3116 
3117 static int mptcp_disconnect(struct sock *sk, int flags)
3118 {
3119 	struct mptcp_sock *msk = mptcp_sk(sk);
3120 
3121 	/* We are on the fastopen error path. We can't call straight into the
3122 	 * subflows cleanup code due to lock nesting (we are already under
3123 	 * msk->firstsocket lock).
3124 	 */
3125 	if (msk->fastopening)
3126 		return -EBUSY;
3127 
3128 	mptcp_check_listen_stop(sk);
3129 	inet_sk_state_store(sk, TCP_CLOSE);
3130 
3131 	mptcp_stop_rtx_timer(sk);
3132 	mptcp_stop_tout_timer(sk);
3133 
3134 	if (msk->token)
3135 		mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3136 
3137 	/* msk->subflow is still intact, the following will not free the first
3138 	 * subflow
3139 	 */
3140 	mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3141 	WRITE_ONCE(msk->flags, 0);
3142 	msk->cb_flags = 0;
3143 	msk->recovery = false;
3144 	msk->can_ack = false;
3145 	msk->fully_established = false;
3146 	msk->rcv_data_fin = false;
3147 	msk->snd_data_fin_enable = false;
3148 	msk->rcv_fastclose = false;
3149 	msk->use_64bit_ack = false;
3150 	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3151 	mptcp_pm_data_reset(msk);
3152 	mptcp_ca_reset(sk);
3153 	msk->bytes_acked = 0;
3154 	msk->bytes_received = 0;
3155 	msk->bytes_sent = 0;
3156 	msk->bytes_retrans = 0;
3157 	msk->rcvspace_init = 0;
3158 
3159 	WRITE_ONCE(sk->sk_shutdown, 0);
3160 	sk_error_report(sk);
3161 	return 0;
3162 }
3163 
3164 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3165 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3166 {
3167 	unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3168 
3169 	return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3170 }
3171 #endif
3172 
3173 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3174 				 const struct mptcp_options_received *mp_opt,
3175 				 struct sock *ssk,
3176 				 struct request_sock *req)
3177 {
3178 	struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3179 	struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3180 	struct mptcp_sock *msk;
3181 
3182 	if (!nsk)
3183 		return NULL;
3184 
3185 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3186 	if (nsk->sk_family == AF_INET6)
3187 		inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3188 #endif
3189 
3190 	__mptcp_init_sock(nsk);
3191 
3192 	msk = mptcp_sk(nsk);
3193 	msk->local_key = subflow_req->local_key;
3194 	msk->token = subflow_req->token;
3195 	msk->in_accept_queue = 1;
3196 	WRITE_ONCE(msk->fully_established, false);
3197 	if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3198 		WRITE_ONCE(msk->csum_enabled, true);
3199 
3200 	msk->write_seq = subflow_req->idsn + 1;
3201 	msk->snd_nxt = msk->write_seq;
3202 	msk->snd_una = msk->write_seq;
3203 	msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3204 	msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3205 	mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3206 
3207 	/* passive msk is created after the first/MPC subflow */
3208 	msk->subflow_id = 2;
3209 
3210 	sock_reset_flag(nsk, SOCK_RCU_FREE);
3211 	security_inet_csk_clone(nsk, req);
3212 
3213 	/* this can't race with mptcp_close(), as the msk is
3214 	 * not yet exposted to user-space
3215 	 */
3216 	inet_sk_state_store(nsk, TCP_ESTABLISHED);
3217 
3218 	/* The msk maintain a ref to each subflow in the connections list */
3219 	WRITE_ONCE(msk->first, ssk);
3220 	list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3221 	sock_hold(ssk);
3222 
3223 	/* new mpc subflow takes ownership of the newly
3224 	 * created mptcp socket
3225 	 */
3226 	mptcp_token_accept(subflow_req, msk);
3227 
3228 	/* set msk addresses early to ensure mptcp_pm_get_local_id()
3229 	 * uses the correct data
3230 	 */
3231 	mptcp_copy_inaddrs(nsk, ssk);
3232 	__mptcp_propagate_sndbuf(nsk, ssk);
3233 
3234 	mptcp_rcv_space_init(msk, ssk);
3235 	bh_unlock_sock(nsk);
3236 
3237 	/* note: the newly allocated socket refcount is 2 now */
3238 	return nsk;
3239 }
3240 
3241 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3242 {
3243 	const struct tcp_sock *tp = tcp_sk(ssk);
3244 
3245 	msk->rcvspace_init = 1;
3246 	msk->rcvq_space.copied = 0;
3247 	msk->rcvq_space.rtt_us = 0;
3248 
3249 	msk->rcvq_space.time = tp->tcp_mstamp;
3250 
3251 	/* initial rcv_space offering made to peer */
3252 	msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3253 				      TCP_INIT_CWND * tp->advmss);
3254 	if (msk->rcvq_space.space == 0)
3255 		msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3256 }
3257 
3258 static struct sock *mptcp_accept(struct sock *ssk, int flags, int *err,
3259 				 bool kern)
3260 {
3261 	struct sock *newsk;
3262 
3263 	pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3264 	newsk = inet_csk_accept(ssk, flags, err, kern);
3265 	if (!newsk)
3266 		return NULL;
3267 
3268 	pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3269 	if (sk_is_mptcp(newsk)) {
3270 		struct mptcp_subflow_context *subflow;
3271 		struct sock *new_mptcp_sock;
3272 
3273 		subflow = mptcp_subflow_ctx(newsk);
3274 		new_mptcp_sock = subflow->conn;
3275 
3276 		/* is_mptcp should be false if subflow->conn is missing, see
3277 		 * subflow_syn_recv_sock()
3278 		 */
3279 		if (WARN_ON_ONCE(!new_mptcp_sock)) {
3280 			tcp_sk(newsk)->is_mptcp = 0;
3281 			goto out;
3282 		}
3283 
3284 		newsk = new_mptcp_sock;
3285 		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3286 	} else {
3287 		MPTCP_INC_STATS(sock_net(ssk),
3288 				MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3289 	}
3290 
3291 out:
3292 	newsk->sk_kern_sock = kern;
3293 	return newsk;
3294 }
3295 
3296 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3297 {
3298 	struct mptcp_subflow_context *subflow, *tmp;
3299 	struct sock *sk = (struct sock *)msk;
3300 
3301 	__mptcp_clear_xmit(sk);
3302 
3303 	/* join list will be eventually flushed (with rst) at sock lock release time */
3304 	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3305 		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3306 
3307 	/* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3308 	mptcp_data_lock(sk);
3309 	skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3310 	__skb_queue_purge(&sk->sk_receive_queue);
3311 	skb_rbtree_purge(&msk->out_of_order_queue);
3312 	mptcp_data_unlock(sk);
3313 
3314 	/* move all the rx fwd alloc into the sk_mem_reclaim_final in
3315 	 * inet_sock_destruct() will dispose it
3316 	 */
3317 	sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3318 	WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3319 	mptcp_token_destroy(msk);
3320 	mptcp_pm_free_anno_list(msk);
3321 	mptcp_free_local_addr_list(msk);
3322 }
3323 
3324 static void mptcp_destroy(struct sock *sk)
3325 {
3326 	struct mptcp_sock *msk = mptcp_sk(sk);
3327 
3328 	/* allow the following to close even the initial subflow */
3329 	msk->free_first = 1;
3330 	mptcp_destroy_common(msk, 0);
3331 	sk_sockets_allocated_dec(sk);
3332 }
3333 
3334 void __mptcp_data_acked(struct sock *sk)
3335 {
3336 	if (!sock_owned_by_user(sk))
3337 		__mptcp_clean_una(sk);
3338 	else
3339 		__set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3340 
3341 	if (mptcp_pending_data_fin_ack(sk))
3342 		mptcp_schedule_work(sk);
3343 }
3344 
3345 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3346 {
3347 	if (!mptcp_send_head(sk))
3348 		return;
3349 
3350 	if (!sock_owned_by_user(sk))
3351 		__mptcp_subflow_push_pending(sk, ssk, false);
3352 	else
3353 		__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3354 }
3355 
3356 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3357 				      BIT(MPTCP_RETRANSMIT) | \
3358 				      BIT(MPTCP_FLUSH_JOIN_LIST))
3359 
3360 /* processes deferred events and flush wmem */
3361 static void mptcp_release_cb(struct sock *sk)
3362 	__must_hold(&sk->sk_lock.slock)
3363 {
3364 	struct mptcp_sock *msk = mptcp_sk(sk);
3365 
3366 	for (;;) {
3367 		unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3368 		struct list_head join_list;
3369 
3370 		if (!flags)
3371 			break;
3372 
3373 		INIT_LIST_HEAD(&join_list);
3374 		list_splice_init(&msk->join_list, &join_list);
3375 
3376 		/* the following actions acquire the subflow socket lock
3377 		 *
3378 		 * 1) can't be invoked in atomic scope
3379 		 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3380 		 *    datapath acquires the msk socket spinlock while helding
3381 		 *    the subflow socket lock
3382 		 */
3383 		msk->cb_flags &= ~flags;
3384 		spin_unlock_bh(&sk->sk_lock.slock);
3385 
3386 		if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3387 			__mptcp_flush_join_list(sk, &join_list);
3388 		if (flags & BIT(MPTCP_PUSH_PENDING))
3389 			__mptcp_push_pending(sk, 0);
3390 		if (flags & BIT(MPTCP_RETRANSMIT))
3391 			__mptcp_retrans(sk);
3392 
3393 		cond_resched();
3394 		spin_lock_bh(&sk->sk_lock.slock);
3395 	}
3396 
3397 	if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3398 		__mptcp_clean_una_wakeup(sk);
3399 	if (unlikely(msk->cb_flags)) {
3400 		/* be sure to sync the msk state before taking actions
3401 		 * depending on sk_state (MPTCP_ERROR_REPORT)
3402 		 * On sk release avoid actions depending on the first subflow
3403 		 */
3404 		if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3405 			__mptcp_sync_state(sk, msk->pending_state);
3406 		if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3407 			__mptcp_error_report(sk);
3408 		if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3409 			__mptcp_sync_sndbuf(sk);
3410 	}
3411 
3412 	__mptcp_update_rmem(sk);
3413 }
3414 
3415 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3416  * TCP can't schedule delack timer before the subflow is fully established.
3417  * MPTCP uses the delack timer to do 3rd ack retransmissions
3418  */
3419 static void schedule_3rdack_retransmission(struct sock *ssk)
3420 {
3421 	struct inet_connection_sock *icsk = inet_csk(ssk);
3422 	struct tcp_sock *tp = tcp_sk(ssk);
3423 	unsigned long timeout;
3424 
3425 	if (mptcp_subflow_ctx(ssk)->fully_established)
3426 		return;
3427 
3428 	/* reschedule with a timeout above RTT, as we must look only for drop */
3429 	if (tp->srtt_us)
3430 		timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3431 	else
3432 		timeout = TCP_TIMEOUT_INIT;
3433 	timeout += jiffies;
3434 
3435 	WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3436 	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3437 	icsk->icsk_ack.timeout = timeout;
3438 	sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3439 }
3440 
3441 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3442 {
3443 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3444 	struct sock *sk = subflow->conn;
3445 
3446 	if (status & BIT(MPTCP_DELEGATE_SEND)) {
3447 		mptcp_data_lock(sk);
3448 		if (!sock_owned_by_user(sk))
3449 			__mptcp_subflow_push_pending(sk, ssk, true);
3450 		else
3451 			__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3452 		mptcp_data_unlock(sk);
3453 	}
3454 	if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3455 		mptcp_data_lock(sk);
3456 		if (!sock_owned_by_user(sk))
3457 			__mptcp_sync_sndbuf(sk);
3458 		else
3459 			__set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3460 		mptcp_data_unlock(sk);
3461 	}
3462 	if (status & BIT(MPTCP_DELEGATE_ACK))
3463 		schedule_3rdack_retransmission(ssk);
3464 }
3465 
3466 static int mptcp_hash(struct sock *sk)
3467 {
3468 	/* should never be called,
3469 	 * we hash the TCP subflows not the master socket
3470 	 */
3471 	WARN_ON_ONCE(1);
3472 	return 0;
3473 }
3474 
3475 static void mptcp_unhash(struct sock *sk)
3476 {
3477 	/* called from sk_common_release(), but nothing to do here */
3478 }
3479 
3480 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3481 {
3482 	struct mptcp_sock *msk = mptcp_sk(sk);
3483 
3484 	pr_debug("msk=%p, ssk=%p", msk, msk->first);
3485 	if (WARN_ON_ONCE(!msk->first))
3486 		return -EINVAL;
3487 
3488 	return inet_csk_get_port(msk->first, snum);
3489 }
3490 
3491 void mptcp_finish_connect(struct sock *ssk)
3492 {
3493 	struct mptcp_subflow_context *subflow;
3494 	struct mptcp_sock *msk;
3495 	struct sock *sk;
3496 
3497 	subflow = mptcp_subflow_ctx(ssk);
3498 	sk = subflow->conn;
3499 	msk = mptcp_sk(sk);
3500 
3501 	pr_debug("msk=%p, token=%u", sk, subflow->token);
3502 
3503 	subflow->map_seq = subflow->iasn;
3504 	subflow->map_subflow_seq = 1;
3505 
3506 	/* the socket is not connected yet, no msk/subflow ops can access/race
3507 	 * accessing the field below
3508 	 */
3509 	WRITE_ONCE(msk->local_key, subflow->local_key);
3510 	WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3511 	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3512 	WRITE_ONCE(msk->snd_una, msk->write_seq);
3513 	WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3514 
3515 	mptcp_pm_new_connection(msk, ssk, 0);
3516 }
3517 
3518 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3519 {
3520 	write_lock_bh(&sk->sk_callback_lock);
3521 	rcu_assign_pointer(sk->sk_wq, &parent->wq);
3522 	sk_set_socket(sk, parent);
3523 	sk->sk_uid = SOCK_INODE(parent)->i_uid;
3524 	write_unlock_bh(&sk->sk_callback_lock);
3525 }
3526 
3527 bool mptcp_finish_join(struct sock *ssk)
3528 {
3529 	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3530 	struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3531 	struct sock *parent = (void *)msk;
3532 	bool ret = true;
3533 
3534 	pr_debug("msk=%p, subflow=%p", msk, subflow);
3535 
3536 	/* mptcp socket already closing? */
3537 	if (!mptcp_is_fully_established(parent)) {
3538 		subflow->reset_reason = MPTCP_RST_EMPTCP;
3539 		return false;
3540 	}
3541 
3542 	/* active subflow, already present inside the conn_list */
3543 	if (!list_empty(&subflow->node)) {
3544 		mptcp_subflow_joined(msk, ssk);
3545 		mptcp_propagate_sndbuf(parent, ssk);
3546 		return true;
3547 	}
3548 
3549 	if (!mptcp_pm_allow_new_subflow(msk))
3550 		goto err_prohibited;
3551 
3552 	/* If we can't acquire msk socket lock here, let the release callback
3553 	 * handle it
3554 	 */
3555 	mptcp_data_lock(parent);
3556 	if (!sock_owned_by_user(parent)) {
3557 		ret = __mptcp_finish_join(msk, ssk);
3558 		if (ret) {
3559 			sock_hold(ssk);
3560 			list_add_tail(&subflow->node, &msk->conn_list);
3561 		}
3562 	} else {
3563 		sock_hold(ssk);
3564 		list_add_tail(&subflow->node, &msk->join_list);
3565 		__set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3566 	}
3567 	mptcp_data_unlock(parent);
3568 
3569 	if (!ret) {
3570 err_prohibited:
3571 		subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3572 		return false;
3573 	}
3574 
3575 	return true;
3576 }
3577 
3578 static void mptcp_shutdown(struct sock *sk, int how)
3579 {
3580 	pr_debug("sk=%p, how=%d", sk, how);
3581 
3582 	if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3583 		__mptcp_wr_shutdown(sk);
3584 }
3585 
3586 static int mptcp_forward_alloc_get(const struct sock *sk)
3587 {
3588 	return READ_ONCE(sk->sk_forward_alloc) +
3589 	       READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3590 }
3591 
3592 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3593 {
3594 	const struct sock *sk = (void *)msk;
3595 	u64 delta;
3596 
3597 	if (sk->sk_state == TCP_LISTEN)
3598 		return -EINVAL;
3599 
3600 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3601 		return 0;
3602 
3603 	delta = msk->write_seq - v;
3604 	if (__mptcp_check_fallback(msk) && msk->first) {
3605 		struct tcp_sock *tp = tcp_sk(msk->first);
3606 
3607 		/* the first subflow is disconnected after close - see
3608 		 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3609 		 * so ignore that status, too.
3610 		 */
3611 		if (!((1 << msk->first->sk_state) &
3612 		      (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3613 			delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3614 	}
3615 	if (delta > INT_MAX)
3616 		delta = INT_MAX;
3617 
3618 	return (int)delta;
3619 }
3620 
3621 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3622 {
3623 	struct mptcp_sock *msk = mptcp_sk(sk);
3624 	bool slow;
3625 
3626 	switch (cmd) {
3627 	case SIOCINQ:
3628 		if (sk->sk_state == TCP_LISTEN)
3629 			return -EINVAL;
3630 
3631 		lock_sock(sk);
3632 		__mptcp_move_skbs(msk);
3633 		*karg = mptcp_inq_hint(sk);
3634 		release_sock(sk);
3635 		break;
3636 	case SIOCOUTQ:
3637 		slow = lock_sock_fast(sk);
3638 		*karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3639 		unlock_sock_fast(sk, slow);
3640 		break;
3641 	case SIOCOUTQNSD:
3642 		slow = lock_sock_fast(sk);
3643 		*karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3644 		unlock_sock_fast(sk, slow);
3645 		break;
3646 	default:
3647 		return -ENOIOCTLCMD;
3648 	}
3649 
3650 	return 0;
3651 }
3652 
3653 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3654 					 struct mptcp_subflow_context *subflow)
3655 {
3656 	subflow->request_mptcp = 0;
3657 	__mptcp_do_fallback(msk);
3658 }
3659 
3660 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3661 {
3662 	struct mptcp_subflow_context *subflow;
3663 	struct mptcp_sock *msk = mptcp_sk(sk);
3664 	int err = -EINVAL;
3665 	struct sock *ssk;
3666 
3667 	ssk = __mptcp_nmpc_sk(msk);
3668 	if (IS_ERR(ssk))
3669 		return PTR_ERR(ssk);
3670 
3671 	inet_sk_state_store(sk, TCP_SYN_SENT);
3672 	subflow = mptcp_subflow_ctx(ssk);
3673 #ifdef CONFIG_TCP_MD5SIG
3674 	/* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3675 	 * TCP option space.
3676 	 */
3677 	if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3678 		mptcp_subflow_early_fallback(msk, subflow);
3679 #endif
3680 	if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3681 		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3682 		mptcp_subflow_early_fallback(msk, subflow);
3683 	}
3684 	if (likely(!__mptcp_check_fallback(msk)))
3685 		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3686 
3687 	/* if reaching here via the fastopen/sendmsg path, the caller already
3688 	 * acquired the subflow socket lock, too.
3689 	 */
3690 	if (!msk->fastopening)
3691 		lock_sock(ssk);
3692 
3693 	/* the following mirrors closely a very small chunk of code from
3694 	 * __inet_stream_connect()
3695 	 */
3696 	if (ssk->sk_state != TCP_CLOSE)
3697 		goto out;
3698 
3699 	if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3700 		err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3701 		if (err)
3702 			goto out;
3703 	}
3704 
3705 	err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3706 	if (err < 0)
3707 		goto out;
3708 
3709 	inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3710 
3711 out:
3712 	if (!msk->fastopening)
3713 		release_sock(ssk);
3714 
3715 	/* on successful connect, the msk state will be moved to established by
3716 	 * subflow_finish_connect()
3717 	 */
3718 	if (unlikely(err)) {
3719 		/* avoid leaving a dangling token in an unconnected socket */
3720 		mptcp_token_destroy(msk);
3721 		inet_sk_state_store(sk, TCP_CLOSE);
3722 		return err;
3723 	}
3724 
3725 	mptcp_copy_inaddrs(sk, ssk);
3726 	return 0;
3727 }
3728 
3729 static struct proto mptcp_prot = {
3730 	.name		= "MPTCP",
3731 	.owner		= THIS_MODULE,
3732 	.init		= mptcp_init_sock,
3733 	.connect	= mptcp_connect,
3734 	.disconnect	= mptcp_disconnect,
3735 	.close		= mptcp_close,
3736 	.accept		= mptcp_accept,
3737 	.setsockopt	= mptcp_setsockopt,
3738 	.getsockopt	= mptcp_getsockopt,
3739 	.shutdown	= mptcp_shutdown,
3740 	.destroy	= mptcp_destroy,
3741 	.sendmsg	= mptcp_sendmsg,
3742 	.ioctl		= mptcp_ioctl,
3743 	.recvmsg	= mptcp_recvmsg,
3744 	.release_cb	= mptcp_release_cb,
3745 	.hash		= mptcp_hash,
3746 	.unhash		= mptcp_unhash,
3747 	.get_port	= mptcp_get_port,
3748 	.forward_alloc_get	= mptcp_forward_alloc_get,
3749 	.sockets_allocated	= &mptcp_sockets_allocated,
3750 
3751 	.memory_allocated	= &tcp_memory_allocated,
3752 	.per_cpu_fw_alloc	= &tcp_memory_per_cpu_fw_alloc,
3753 
3754 	.memory_pressure	= &tcp_memory_pressure,
3755 	.sysctl_wmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_wmem),
3756 	.sysctl_rmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_rmem),
3757 	.sysctl_mem	= sysctl_tcp_mem,
3758 	.obj_size	= sizeof(struct mptcp_sock),
3759 	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
3760 	.no_autobind	= true,
3761 };
3762 
3763 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3764 {
3765 	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3766 	struct sock *ssk, *sk = sock->sk;
3767 	int err = -EINVAL;
3768 
3769 	lock_sock(sk);
3770 	ssk = __mptcp_nmpc_sk(msk);
3771 	if (IS_ERR(ssk)) {
3772 		err = PTR_ERR(ssk);
3773 		goto unlock;
3774 	}
3775 
3776 	if (sk->sk_family == AF_INET)
3777 		err = inet_bind_sk(ssk, uaddr, addr_len);
3778 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3779 	else if (sk->sk_family == AF_INET6)
3780 		err = inet6_bind_sk(ssk, uaddr, addr_len);
3781 #endif
3782 	if (!err)
3783 		mptcp_copy_inaddrs(sk, ssk);
3784 
3785 unlock:
3786 	release_sock(sk);
3787 	return err;
3788 }
3789 
3790 static int mptcp_listen(struct socket *sock, int backlog)
3791 {
3792 	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3793 	struct sock *sk = sock->sk;
3794 	struct sock *ssk;
3795 	int err;
3796 
3797 	pr_debug("msk=%p", msk);
3798 
3799 	lock_sock(sk);
3800 
3801 	err = -EINVAL;
3802 	if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3803 		goto unlock;
3804 
3805 	ssk = __mptcp_nmpc_sk(msk);
3806 	if (IS_ERR(ssk)) {
3807 		err = PTR_ERR(ssk);
3808 		goto unlock;
3809 	}
3810 
3811 	inet_sk_state_store(sk, TCP_LISTEN);
3812 	sock_set_flag(sk, SOCK_RCU_FREE);
3813 
3814 	lock_sock(ssk);
3815 	err = __inet_listen_sk(ssk, backlog);
3816 	release_sock(ssk);
3817 	inet_sk_state_store(sk, inet_sk_state_load(ssk));
3818 
3819 	if (!err) {
3820 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3821 		mptcp_copy_inaddrs(sk, ssk);
3822 		mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3823 	}
3824 
3825 unlock:
3826 	release_sock(sk);
3827 	return err;
3828 }
3829 
3830 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3831 			       int flags, bool kern)
3832 {
3833 	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3834 	struct sock *ssk, *newsk;
3835 	int err;
3836 
3837 	pr_debug("msk=%p", msk);
3838 
3839 	/* Buggy applications can call accept on socket states other then LISTEN
3840 	 * but no need to allocate the first subflow just to error out.
3841 	 */
3842 	ssk = READ_ONCE(msk->first);
3843 	if (!ssk)
3844 		return -EINVAL;
3845 
3846 	newsk = mptcp_accept(ssk, flags, &err, kern);
3847 	if (!newsk)
3848 		return err;
3849 
3850 	lock_sock(newsk);
3851 
3852 	__inet_accept(sock, newsock, newsk);
3853 	if (!mptcp_is_tcpsk(newsock->sk)) {
3854 		struct mptcp_sock *msk = mptcp_sk(newsk);
3855 		struct mptcp_subflow_context *subflow;
3856 
3857 		set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3858 		msk->in_accept_queue = 0;
3859 
3860 		/* set ssk->sk_socket of accept()ed flows to mptcp socket.
3861 		 * This is needed so NOSPACE flag can be set from tcp stack.
3862 		 */
3863 		mptcp_for_each_subflow(msk, subflow) {
3864 			struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3865 
3866 			if (!ssk->sk_socket)
3867 				mptcp_sock_graft(ssk, newsock);
3868 		}
3869 
3870 		/* Do late cleanup for the first subflow as necessary. Also
3871 		 * deal with bad peers not doing a complete shutdown.
3872 		 */
3873 		if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3874 			__mptcp_close_ssk(newsk, msk->first,
3875 					  mptcp_subflow_ctx(msk->first), 0);
3876 			if (unlikely(list_is_singular(&msk->conn_list)))
3877 				inet_sk_state_store(newsk, TCP_CLOSE);
3878 		}
3879 	}
3880 	release_sock(newsk);
3881 
3882 	return 0;
3883 }
3884 
3885 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3886 {
3887 	struct sock *sk = (struct sock *)msk;
3888 
3889 	if (sk_stream_is_writeable(sk))
3890 		return EPOLLOUT | EPOLLWRNORM;
3891 
3892 	mptcp_set_nospace(sk);
3893 	smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3894 	if (sk_stream_is_writeable(sk))
3895 		return EPOLLOUT | EPOLLWRNORM;
3896 
3897 	return 0;
3898 }
3899 
3900 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3901 			   struct poll_table_struct *wait)
3902 {
3903 	struct sock *sk = sock->sk;
3904 	struct mptcp_sock *msk;
3905 	__poll_t mask = 0;
3906 	u8 shutdown;
3907 	int state;
3908 
3909 	msk = mptcp_sk(sk);
3910 	sock_poll_wait(file, sock, wait);
3911 
3912 	state = inet_sk_state_load(sk);
3913 	pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3914 	if (state == TCP_LISTEN) {
3915 		struct sock *ssk = READ_ONCE(msk->first);
3916 
3917 		if (WARN_ON_ONCE(!ssk))
3918 			return 0;
3919 
3920 		return inet_csk_listen_poll(ssk);
3921 	}
3922 
3923 	shutdown = READ_ONCE(sk->sk_shutdown);
3924 	if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3925 		mask |= EPOLLHUP;
3926 	if (shutdown & RCV_SHUTDOWN)
3927 		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3928 
3929 	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3930 		mask |= mptcp_check_readable(msk);
3931 		if (shutdown & SEND_SHUTDOWN)
3932 			mask |= EPOLLOUT | EPOLLWRNORM;
3933 		else
3934 			mask |= mptcp_check_writeable(msk);
3935 	} else if (state == TCP_SYN_SENT &&
3936 		   inet_test_bit(DEFER_CONNECT, sk)) {
3937 		/* cf tcp_poll() note about TFO */
3938 		mask |= EPOLLOUT | EPOLLWRNORM;
3939 	}
3940 
3941 	/* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3942 	smp_rmb();
3943 	if (READ_ONCE(sk->sk_err))
3944 		mask |= EPOLLERR;
3945 
3946 	return mask;
3947 }
3948 
3949 static const struct proto_ops mptcp_stream_ops = {
3950 	.family		   = PF_INET,
3951 	.owner		   = THIS_MODULE,
3952 	.release	   = inet_release,
3953 	.bind		   = mptcp_bind,
3954 	.connect	   = inet_stream_connect,
3955 	.socketpair	   = sock_no_socketpair,
3956 	.accept		   = mptcp_stream_accept,
3957 	.getname	   = inet_getname,
3958 	.poll		   = mptcp_poll,
3959 	.ioctl		   = inet_ioctl,
3960 	.gettstamp	   = sock_gettstamp,
3961 	.listen		   = mptcp_listen,
3962 	.shutdown	   = inet_shutdown,
3963 	.setsockopt	   = sock_common_setsockopt,
3964 	.getsockopt	   = sock_common_getsockopt,
3965 	.sendmsg	   = inet_sendmsg,
3966 	.recvmsg	   = inet_recvmsg,
3967 	.mmap		   = sock_no_mmap,
3968 };
3969 
3970 static struct inet_protosw mptcp_protosw = {
3971 	.type		= SOCK_STREAM,
3972 	.protocol	= IPPROTO_MPTCP,
3973 	.prot		= &mptcp_prot,
3974 	.ops		= &mptcp_stream_ops,
3975 	.flags		= INET_PROTOSW_ICSK,
3976 };
3977 
3978 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3979 {
3980 	struct mptcp_delegated_action *delegated;
3981 	struct mptcp_subflow_context *subflow;
3982 	int work_done = 0;
3983 
3984 	delegated = container_of(napi, struct mptcp_delegated_action, napi);
3985 	while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3986 		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3987 
3988 		bh_lock_sock_nested(ssk);
3989 		if (!sock_owned_by_user(ssk)) {
3990 			mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
3991 		} else {
3992 			/* tcp_release_cb_override already processed
3993 			 * the action or will do at next release_sock().
3994 			 * In both case must dequeue the subflow here - on the same
3995 			 * CPU that scheduled it.
3996 			 */
3997 			smp_wmb();
3998 			clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
3999 		}
4000 		bh_unlock_sock(ssk);
4001 		sock_put(ssk);
4002 
4003 		if (++work_done == budget)
4004 			return budget;
4005 	}
4006 
4007 	/* always provide a 0 'work_done' argument, so that napi_complete_done
4008 	 * will not try accessing the NULL napi->dev ptr
4009 	 */
4010 	napi_complete_done(napi, 0);
4011 	return work_done;
4012 }
4013 
4014 void __init mptcp_proto_init(void)
4015 {
4016 	struct mptcp_delegated_action *delegated;
4017 	int cpu;
4018 
4019 	mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4020 
4021 	if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4022 		panic("Failed to allocate MPTCP pcpu counter\n");
4023 
4024 	init_dummy_netdev(&mptcp_napi_dev);
4025 	for_each_possible_cpu(cpu) {
4026 		delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4027 		INIT_LIST_HEAD(&delegated->head);
4028 		netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4029 				  mptcp_napi_poll);
4030 		napi_enable(&delegated->napi);
4031 	}
4032 
4033 	mptcp_subflow_init();
4034 	mptcp_pm_init();
4035 	mptcp_sched_init();
4036 	mptcp_token_init();
4037 
4038 	if (proto_register(&mptcp_prot, 1) != 0)
4039 		panic("Failed to register MPTCP proto.\n");
4040 
4041 	inet_register_protosw(&mptcp_protosw);
4042 
4043 	BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4044 }
4045 
4046 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4047 static const struct proto_ops mptcp_v6_stream_ops = {
4048 	.family		   = PF_INET6,
4049 	.owner		   = THIS_MODULE,
4050 	.release	   = inet6_release,
4051 	.bind		   = mptcp_bind,
4052 	.connect	   = inet_stream_connect,
4053 	.socketpair	   = sock_no_socketpair,
4054 	.accept		   = mptcp_stream_accept,
4055 	.getname	   = inet6_getname,
4056 	.poll		   = mptcp_poll,
4057 	.ioctl		   = inet6_ioctl,
4058 	.gettstamp	   = sock_gettstamp,
4059 	.listen		   = mptcp_listen,
4060 	.shutdown	   = inet_shutdown,
4061 	.setsockopt	   = sock_common_setsockopt,
4062 	.getsockopt	   = sock_common_getsockopt,
4063 	.sendmsg	   = inet6_sendmsg,
4064 	.recvmsg	   = inet6_recvmsg,
4065 	.mmap		   = sock_no_mmap,
4066 #ifdef CONFIG_COMPAT
4067 	.compat_ioctl	   = inet6_compat_ioctl,
4068 #endif
4069 };
4070 
4071 static struct proto mptcp_v6_prot;
4072 
4073 static struct inet_protosw mptcp_v6_protosw = {
4074 	.type		= SOCK_STREAM,
4075 	.protocol	= IPPROTO_MPTCP,
4076 	.prot		= &mptcp_v6_prot,
4077 	.ops		= &mptcp_v6_stream_ops,
4078 	.flags		= INET_PROTOSW_ICSK,
4079 };
4080 
4081 int __init mptcp_proto_v6_init(void)
4082 {
4083 	int err;
4084 
4085 	mptcp_v6_prot = mptcp_prot;
4086 	strcpy(mptcp_v6_prot.name, "MPTCPv6");
4087 	mptcp_v6_prot.slab = NULL;
4088 	mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4089 	mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4090 
4091 	err = proto_register(&mptcp_v6_prot, 1);
4092 	if (err)
4093 		return err;
4094 
4095 	err = inet6_register_protosw(&mptcp_v6_protosw);
4096 	if (err)
4097 		proto_unregister(&mptcp_v6_prot);
4098 
4099 	return err;
4100 }
4101 #endif
4102