]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - net/mptcp/protocol.c
projects / mirror_ubuntu-hirsute-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
mptcp: fix race in release_cb
[mirror_ubuntu-hirsute-kernel.git] / net / mptcp / protocol.c
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 <linux/igmp.h>
15 #include <net/sock.h>
16 #include <net/inet_common.h>
17 #include <net/inet_hashtables.h>
18 #include <net/protocol.h>
19 #include <net/tcp.h>
20 #include <net/tcp_states.h>
21 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
22 #include <net/transp_v6.h>
23 #include <net/addrconf.h>
24 #endif
25 #include <net/mptcp.h>
26 #include <net/xfrm.h>
27 #include "protocol.h"
28 #include "mib.h"
29
30 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
31 struct mptcp6_sock {
32 struct mptcp_sock msk;
33 struct ipv6_pinfo np;
34 };
35 #endif
36
37 struct mptcp_skb_cb {
38 u64 map_seq;
39 u64 end_seq;
40 u32 offset;
41 };
42
43 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
44
45 static struct percpu_counter mptcp_sockets_allocated;
46
47 static void __mptcp_destroy_sock(struct sock *sk);
48 static void __mptcp_check_send_data_fin(struct sock *sk);
49
50 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
51 * completed yet or has failed, return the subflow socket.
52 * Otherwise return NULL.
53 */
54 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
55 {
56 if (!msk->subflow || READ_ONCE(msk->can_ack))
57 return NULL;
58
59 return msk->subflow;
60 }
61
62 /* Returns end sequence number of the receiver's advertised window */
63 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
64 {
65 return READ_ONCE(msk->wnd_end);
66 }
67
68 static bool mptcp_is_tcpsk(struct sock *sk)
69 {
70 struct socket *sock = sk->sk_socket;
71
72 if (unlikely(sk->sk_prot == &tcp_prot)) {
73 /* we are being invoked after mptcp_accept() has
74 * accepted a non-mp-capable flow: sk is a tcp_sk,
75 * not an mptcp one.
76 *
77 * Hand the socket over to tcp so all further socket ops
78 * bypass mptcp.
79 */
80 sock->ops = &inet_stream_ops;
81 return true;
82 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
83 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
84 sock->ops = &inet6_stream_ops;
85 return true;
86 #endif
87 }
88
89 return false;
90 }
91
92 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
93 {
94 sock_owned_by_me((const struct sock *)msk);
95
96 if (likely(!__mptcp_check_fallback(msk)))
97 return NULL;
98
99 return msk->first;
100 }
101
102 static int __mptcp_socket_create(struct mptcp_sock *msk)
103 {
104 struct mptcp_subflow_context *subflow;
105 struct sock *sk = (struct sock *)msk;
106 struct socket *ssock;
107 int err;
108
109 err = mptcp_subflow_create_socket(sk, &ssock);
110 if (err)
111 return err;
112
113 msk->first = ssock->sk;
114 msk->subflow = ssock;
115 subflow = mptcp_subflow_ctx(ssock->sk);
116 list_add(&subflow->node, &msk->conn_list);
117 sock_hold(ssock->sk);
118 subflow->request_mptcp = 1;
119 mptcp_sock_graft(msk->first, sk->sk_socket);
120
121 return 0;
122 }
123
124 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
125 {
126 sk_drops_add(sk, skb);
127 __kfree_skb(skb);
128 }
129
130 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
131 struct sk_buff *from)
132 {
133 bool fragstolen;
134 int delta;
135
136 if (MPTCP_SKB_CB(from)->offset ||
137 !skb_try_coalesce(to, from, &fragstolen, &delta))
138 return false;
139
140 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
141 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
142 to->len, MPTCP_SKB_CB(from)->end_seq);
143 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
144 kfree_skb_partial(from, fragstolen);
145 atomic_add(delta, &sk->sk_rmem_alloc);
146 sk_mem_charge(sk, delta);
147 return true;
148 }
149
150 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
151 struct sk_buff *from)
152 {
153 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
154 return false;
155
156 return mptcp_try_coalesce((struct sock *)msk, to, from);
157 }
158
159 /* "inspired" by tcp_data_queue_ofo(), main differences:
160 * - use mptcp seqs
161 * - don't cope with sacks
162 */
163 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
164 {
165 struct sock *sk = (struct sock *)msk;
166 struct rb_node **p, *parent;
167 u64 seq, end_seq, max_seq;
168 struct sk_buff *skb1;
169
170 seq = MPTCP_SKB_CB(skb)->map_seq;
171 end_seq = MPTCP_SKB_CB(skb)->end_seq;
172 max_seq = READ_ONCE(msk->rcv_wnd_sent);
173
174 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
175 RB_EMPTY_ROOT(&msk->out_of_order_queue));
176 if (after64(end_seq, max_seq)) {
177 /* out of window */
178 mptcp_drop(sk, skb);
179 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
180 (unsigned long long)end_seq - (unsigned long)max_seq,
181 (unsigned long long)msk->rcv_wnd_sent);
182 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
183 return;
184 }
185
186 p = &msk->out_of_order_queue.rb_node;
187 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
188 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
189 rb_link_node(&skb->rbnode, NULL, p);
190 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
191 msk->ooo_last_skb = skb;
192 goto end;
193 }
194
195 /* with 2 subflows, adding at end of ooo queue is quite likely
196 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
197 */
198 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
199 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
200 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
201 return;
202 }
203
204 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
205 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
206 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
207 parent = &msk->ooo_last_skb->rbnode;
208 p = &parent->rb_right;
209 goto insert;
210 }
211
212 /* Find place to insert this segment. Handle overlaps on the way. */
213 parent = NULL;
214 while (*p) {
215 parent = *p;
216 skb1 = rb_to_skb(parent);
217 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
218 p = &parent->rb_left;
219 continue;
220 }
221 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
222 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
223 /* All the bits are present. Drop. */
224 mptcp_drop(sk, skb);
225 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
226 return;
227 }
228 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
229 /* partial overlap:
230 * | skb |
231 * | skb1 |
232 * continue traversing
233 */
234 } else {
235 /* skb's seq == skb1's seq and skb covers skb1.
236 * Replace skb1 with skb.
237 */
238 rb_replace_node(&skb1->rbnode, &skb->rbnode,
239 &msk->out_of_order_queue);
240 mptcp_drop(sk, skb1);
241 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
242 goto merge_right;
243 }
244 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
245 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
246 return;
247 }
248 p = &parent->rb_right;
249 }
250
251 insert:
252 /* Insert segment into RB tree. */
253 rb_link_node(&skb->rbnode, parent, p);
254 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
255
256 merge_right:
257 /* Remove other segments covered by skb. */
258 while ((skb1 = skb_rb_next(skb)) != NULL) {
259 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
260 break;
261 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
262 mptcp_drop(sk, skb1);
263 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
264 }
265 /* If there is no skb after us, we are the last_skb ! */
266 if (!skb1)
267 msk->ooo_last_skb = skb;
268
269 end:
270 skb_condense(skb);
271 skb_set_owner_r(skb, sk);
272 }
273
274 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
275 struct sk_buff *skb, unsigned int offset,
276 size_t copy_len)
277 {
278 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
279 struct sock *sk = (struct sock *)msk;
280 struct sk_buff *tail;
281
282 __skb_unlink(skb, &ssk->sk_receive_queue);
283
284 skb_ext_reset(skb);
285 skb_orphan(skb);
286
287 /* try to fetch required memory from subflow */
288 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
289 if (ssk->sk_forward_alloc < skb->truesize)
290 goto drop;
291 __sk_mem_reclaim(ssk, skb->truesize);
292 if (!sk_rmem_schedule(sk, skb, skb->truesize))
293 goto drop;
294 }
295
296 /* the skb map_seq accounts for the skb offset:
297 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
298 * value
299 */
300 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
301 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
302 MPTCP_SKB_CB(skb)->offset = offset;
303
304 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
305 /* in sequence */
306 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
307 tail = skb_peek_tail(&sk->sk_receive_queue);
308 if (tail && mptcp_try_coalesce(sk, tail, skb))
309 return true;
310
311 skb_set_owner_r(skb, sk);
312 __skb_queue_tail(&sk->sk_receive_queue, skb);
313 return true;
314 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
315 mptcp_data_queue_ofo(msk, skb);
316 return false;
317 }
318
319 /* old data, keep it simple and drop the whole pkt, sender
320 * will retransmit as needed, if needed.
321 */
322 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
323 drop:
324 mptcp_drop(sk, skb);
325 return false;
326 }
327
328 static void mptcp_stop_timer(struct sock *sk)
329 {
330 struct inet_connection_sock *icsk = inet_csk(sk);
331
332 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
333 mptcp_sk(sk)->timer_ival = 0;
334 }
335
336 static void mptcp_close_wake_up(struct sock *sk)
337 {
338 if (sock_flag(sk, SOCK_DEAD))
339 return;
340
341 sk->sk_state_change(sk);
342 if (sk->sk_shutdown == SHUTDOWN_MASK ||
343 sk->sk_state == TCP_CLOSE)
344 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
345 else
346 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
347 }
348
349 static bool mptcp_pending_data_fin_ack(struct sock *sk)
350 {
351 struct mptcp_sock *msk = mptcp_sk(sk);
352
353 return !__mptcp_check_fallback(msk) &&
354 ((1 << sk->sk_state) &
355 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
356 msk->write_seq == READ_ONCE(msk->snd_una);
357 }
358
359 static void mptcp_check_data_fin_ack(struct sock *sk)
360 {
361 struct mptcp_sock *msk = mptcp_sk(sk);
362
363 /* Look for an acknowledged DATA_FIN */
364 if (mptcp_pending_data_fin_ack(sk)) {
365 WRITE_ONCE(msk->snd_data_fin_enable, 0);
366
367 switch (sk->sk_state) {
368 case TCP_FIN_WAIT1:
369 inet_sk_state_store(sk, TCP_FIN_WAIT2);
370 break;
371 case TCP_CLOSING:
372 case TCP_LAST_ACK:
373 inet_sk_state_store(sk, TCP_CLOSE);
374 break;
375 }
376
377 mptcp_close_wake_up(sk);
378 }
379 }
380
381 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
382 {
383 struct mptcp_sock *msk = mptcp_sk(sk);
384
385 if (READ_ONCE(msk->rcv_data_fin) &&
386 ((1 << sk->sk_state) &
387 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
388 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
389
390 if (msk->ack_seq == rcv_data_fin_seq) {
391 if (seq)
392 *seq = rcv_data_fin_seq;
393
394 return true;
395 }
396 }
397
398 return false;
399 }
400
401 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
402 {
403 long tout = ssk && inet_csk(ssk)->icsk_pending ?
404 inet_csk(ssk)->icsk_timeout - jiffies : 0;
405
406 if (tout <= 0)
407 tout = mptcp_sk(sk)->timer_ival;
408 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
409 }
410
411 static bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
412 {
413 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
414
415 /* can't send if JOIN hasn't completed yet (i.e. is usable for mptcp) */
416 if (subflow->request_join && !subflow->fully_established)
417 return false;
418
419 /* only send if our side has not closed yet */
420 return ((1 << ssk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT));
421 }
422
423 static bool tcp_can_send_ack(const struct sock *ssk)
424 {
425 return !((1 << inet_sk_state_load(ssk)) &
426 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
427 }
428
429 static void mptcp_send_ack(struct mptcp_sock *msk)
430 {
431 struct mptcp_subflow_context *subflow;
432
433 mptcp_for_each_subflow(msk, subflow) {
434 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
435
436 lock_sock(ssk);
437 if (tcp_can_send_ack(ssk))
438 tcp_send_ack(ssk);
439 release_sock(ssk);
440 }
441 }
442
443 static bool mptcp_subflow_cleanup_rbuf(struct sock *ssk)
444 {
445 int ret;
446
447 lock_sock(ssk);
448 ret = tcp_can_send_ack(ssk);
449 if (ret)
450 tcp_cleanup_rbuf(ssk, 1);
451 release_sock(ssk);
452 return ret;
453 }
454
455 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
456 {
457 struct sock *ack_hint = READ_ONCE(msk->ack_hint);
458 struct mptcp_subflow_context *subflow;
459
460 /* if the hinted ssk is still active, try to use it */
461 if (likely(ack_hint)) {
462 mptcp_for_each_subflow(msk, subflow) {
463 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
464
465 if (ack_hint == ssk && mptcp_subflow_cleanup_rbuf(ssk))
466 return;
467 }
468 }
469
470 /* otherwise pick the first active subflow */
471 mptcp_for_each_subflow(msk, subflow)
472 if (mptcp_subflow_cleanup_rbuf(mptcp_subflow_tcp_sock(subflow)))
473 return;
474 }
475
476 static bool mptcp_check_data_fin(struct sock *sk)
477 {
478 struct mptcp_sock *msk = mptcp_sk(sk);
479 u64 rcv_data_fin_seq;
480 bool ret = false;
481
482 if (__mptcp_check_fallback(msk) || !msk->first)
483 return ret;
484
485 /* Need to ack a DATA_FIN received from a peer while this side
486 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
487 * msk->rcv_data_fin was set when parsing the incoming options
488 * at the subflow level and the msk lock was not held, so this
489 * is the first opportunity to act on the DATA_FIN and change
490 * the msk state.
491 *
492 * If we are caught up to the sequence number of the incoming
493 * DATA_FIN, send the DATA_ACK now and do state transition. If
494 * not caught up, do nothing and let the recv code send DATA_ACK
495 * when catching up.
496 */
497
498 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
499 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
500 WRITE_ONCE(msk->rcv_data_fin, 0);
501
502 sk->sk_shutdown |= RCV_SHUTDOWN;
503 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
504 set_bit(MPTCP_DATA_READY, &msk->flags);
505
506 switch (sk->sk_state) {
507 case TCP_ESTABLISHED:
508 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
509 break;
510 case TCP_FIN_WAIT1:
511 inet_sk_state_store(sk, TCP_CLOSING);
512 break;
513 case TCP_FIN_WAIT2:
514 inet_sk_state_store(sk, TCP_CLOSE);
515 break;
516 default:
517 /* Other states not expected */
518 WARN_ON_ONCE(1);
519 break;
520 }
521
522 ret = true;
523 mptcp_set_timeout(sk, NULL);
524 mptcp_send_ack(msk);
525 mptcp_close_wake_up(sk);
526 }
527 return ret;
528 }
529
530 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
531 struct sock *ssk,
532 unsigned int *bytes)
533 {
534 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
535 struct sock *sk = (struct sock *)msk;
536 unsigned int moved = 0;
537 bool more_data_avail;
538 struct tcp_sock *tp;
539 bool done = false;
540 int sk_rbuf;
541
542 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
543
544 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
545 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
546
547 if (unlikely(ssk_rbuf > sk_rbuf)) {
548 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
549 sk_rbuf = ssk_rbuf;
550 }
551 }
552
553 pr_debug("msk=%p ssk=%p", msk, ssk);
554 tp = tcp_sk(ssk);
555 do {
556 u32 map_remaining, offset;
557 u32 seq = tp->copied_seq;
558 struct sk_buff *skb;
559 bool fin;
560
561 /* try to move as much data as available */
562 map_remaining = subflow->map_data_len -
563 mptcp_subflow_get_map_offset(subflow);
564
565 skb = skb_peek(&ssk->sk_receive_queue);
566 if (!skb) {
567 /* if no data is found, a racing workqueue/recvmsg
568 * already processed the new data, stop here or we
569 * can enter an infinite loop
570 */
571 if (!moved)
572 done = true;
573 break;
574 }
575
576 if (__mptcp_check_fallback(msk)) {
577 /* if we are running under the workqueue, TCP could have
578 * collapsed skbs between dummy map creation and now
579 * be sure to adjust the size
580 */
581 map_remaining = skb->len;
582 subflow->map_data_len = skb->len;
583 }
584
585 offset = seq - TCP_SKB_CB(skb)->seq;
586 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
587 if (fin) {
588 done = true;
589 seq++;
590 }
591
592 if (offset < skb->len) {
593 size_t len = skb->len - offset;
594
595 if (tp->urg_data)
596 done = true;
597
598 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
599 moved += len;
600 seq += len;
601
602 if (WARN_ON_ONCE(map_remaining < len))
603 break;
604 } else {
605 WARN_ON_ONCE(!fin);
606 sk_eat_skb(ssk, skb);
607 done = true;
608 }
609
610 WRITE_ONCE(tp->copied_seq, seq);
611 more_data_avail = mptcp_subflow_data_available(ssk);
612
613 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
614 done = true;
615 break;
616 }
617 } while (more_data_avail);
618 WRITE_ONCE(msk->ack_hint, ssk);
619
620 *bytes += moved;
621 return done;
622 }
623
624 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
625 {
626 struct sock *sk = (struct sock *)msk;
627 struct sk_buff *skb, *tail;
628 bool moved = false;
629 struct rb_node *p;
630 u64 end_seq;
631
632 p = rb_first(&msk->out_of_order_queue);
633 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
634 while (p) {
635 skb = rb_to_skb(p);
636 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
637 break;
638
639 p = rb_next(p);
640 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
641
642 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
643 msk->ack_seq))) {
644 mptcp_drop(sk, skb);
645 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
646 continue;
647 }
648
649 end_seq = MPTCP_SKB_CB(skb)->end_seq;
650 tail = skb_peek_tail(&sk->sk_receive_queue);
651 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
652 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
653
654 /* skip overlapping data, if any */
655 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
656 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
657 delta);
658 MPTCP_SKB_CB(skb)->offset += delta;
659 __skb_queue_tail(&sk->sk_receive_queue, skb);
660 }
661 msk->ack_seq = end_seq;
662 moved = true;
663 }
664 return moved;
665 }
666
667 /* In most cases we will be able to lock the mptcp socket. If its already
668 * owned, we need to defer to the work queue to avoid ABBA deadlock.
669 */
670 static void move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
671 {
672 struct sock *sk = (struct sock *)msk;
673 unsigned int moved = 0;
674
675 if (inet_sk_state_load(sk) == TCP_CLOSE)
676 return;
677
678 mptcp_data_lock(sk);
679
680 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
681 __mptcp_ofo_queue(msk);
682
683 /* If the moves have caught up with the DATA_FIN sequence number
684 * it's time to ack the DATA_FIN and change socket state, but
685 * this is not a good place to change state. Let the workqueue
686 * do it.
687 */
688 if (mptcp_pending_data_fin(sk, NULL))
689 mptcp_schedule_work(sk);
690 mptcp_data_unlock(sk);
691 }
692
693 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
694 {
695 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
696 struct mptcp_sock *msk = mptcp_sk(sk);
697 int sk_rbuf, ssk_rbuf;
698 bool wake;
699
700 /* The peer can send data while we are shutting down this
701 * subflow at msk destruction time, but we must avoid enqueuing
702 * more data to the msk receive queue
703 */
704 if (unlikely(subflow->disposable))
705 return;
706
707 /* move_skbs_to_msk below can legitly clear the data_avail flag,
708 * but we will need later to properly woke the reader, cache its
709 * value
710 */
711 wake = subflow->data_avail == MPTCP_SUBFLOW_DATA_AVAIL;
712 if (wake)
713 set_bit(MPTCP_DATA_READY, &msk->flags);
714
715 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
716 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
717 if (unlikely(ssk_rbuf > sk_rbuf))
718 sk_rbuf = ssk_rbuf;
719
720 /* over limit? can't append more skbs to msk */
721 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf)
722 goto wake;
723
724 move_skbs_to_msk(msk, ssk);
725
726 wake:
727 if (wake)
728 sk->sk_data_ready(sk);
729 }
730
731 void __mptcp_flush_join_list(struct mptcp_sock *msk)
732 {
733 if (likely(list_empty(&msk->join_list)))
734 return;
735
736 spin_lock_bh(&msk->join_list_lock);
737 list_splice_tail_init(&msk->join_list, &msk->conn_list);
738 spin_unlock_bh(&msk->join_list_lock);
739 }
740
741 static bool mptcp_timer_pending(struct sock *sk)
742 {
743 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
744 }
745
746 static void mptcp_reset_timer(struct sock *sk)
747 {
748 struct inet_connection_sock *icsk = inet_csk(sk);
749 unsigned long tout;
750
751 /* prevent rescheduling on close */
752 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
753 return;
754
755 /* should never be called with mptcp level timer cleared */
756 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
757 if (WARN_ON_ONCE(!tout))
758 tout = TCP_RTO_MIN;
759 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
760 }
761
762 bool mptcp_schedule_work(struct sock *sk)
763 {
764 if (inet_sk_state_load(sk) != TCP_CLOSE &&
765 schedule_work(&mptcp_sk(sk)->work)) {
766 /* each subflow already holds a reference to the sk, and the
767 * workqueue is invoked by a subflow, so sk can't go away here.
768 */
769 sock_hold(sk);
770 return true;
771 }
772 return false;
773 }
774
775 void mptcp_subflow_eof(struct sock *sk)
776 {
777 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
778 mptcp_schedule_work(sk);
779 }
780
781 static void mptcp_check_for_eof(struct mptcp_sock *msk)
782 {
783 struct mptcp_subflow_context *subflow;
784 struct sock *sk = (struct sock *)msk;
785 int receivers = 0;
786
787 mptcp_for_each_subflow(msk, subflow)
788 receivers += !subflow->rx_eof;
789 if (receivers)
790 return;
791
792 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
793 /* hopefully temporary hack: propagate shutdown status
794 * to msk, when all subflows agree on it
795 */
796 sk->sk_shutdown |= RCV_SHUTDOWN;
797
798 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
799 set_bit(MPTCP_DATA_READY, &msk->flags);
800 sk->sk_data_ready(sk);
801 }
802
803 switch (sk->sk_state) {
804 case TCP_ESTABLISHED:
805 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
806 break;
807 case TCP_FIN_WAIT1:
808 inet_sk_state_store(sk, TCP_CLOSING);
809 break;
810 case TCP_FIN_WAIT2:
811 inet_sk_state_store(sk, TCP_CLOSE);
812 break;
813 default:
814 return;
815 }
816 mptcp_close_wake_up(sk);
817 }
818
819 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
820 {
821 struct mptcp_subflow_context *subflow;
822 struct sock *sk = (struct sock *)msk;
823
824 sock_owned_by_me(sk);
825
826 mptcp_for_each_subflow(msk, subflow) {
827 if (subflow->data_avail)
828 return mptcp_subflow_tcp_sock(subflow);
829 }
830
831 return NULL;
832 }
833
834 static bool mptcp_skb_can_collapse_to(u64 write_seq,
835 const struct sk_buff *skb,
836 const struct mptcp_ext *mpext)
837 {
838 if (!tcp_skb_can_collapse_to(skb))
839 return false;
840
841 /* can collapse only if MPTCP level sequence is in order and this
842 * mapping has not been xmitted yet
843 */
844 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
845 !mpext->frozen;
846 }
847
848 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
849 const struct page_frag *pfrag,
850 const struct mptcp_data_frag *df)
851 {
852 return df && pfrag->page == df->page &&
853 pfrag->size - pfrag->offset > 0 &&
854 df->data_seq + df->data_len == msk->write_seq;
855 }
856
857 static int mptcp_wmem_with_overhead(struct sock *sk, int size)
858 {
859 struct mptcp_sock *msk = mptcp_sk(sk);
860 int ret, skbs;
861
862 ret = size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
863 skbs = (msk->tx_pending_data + size) / msk->size_goal_cache;
864 if (skbs < msk->skb_tx_cache.qlen)
865 return ret;
866
867 return ret + (skbs - msk->skb_tx_cache.qlen) * SKB_TRUESIZE(MAX_TCP_HEADER);
868 }
869
870 static void __mptcp_wmem_reserve(struct sock *sk, int size)
871 {
872 int amount = mptcp_wmem_with_overhead(sk, size);
873 struct mptcp_sock *msk = mptcp_sk(sk);
874
875 WARN_ON_ONCE(msk->wmem_reserved);
876 if (WARN_ON_ONCE(amount < 0))
877 amount = 0;
878
879 if (amount <= sk->sk_forward_alloc)
880 goto reserve;
881
882 /* under memory pressure try to reserve at most a single page
883 * otherwise try to reserve the full estimate and fallback
884 * to a single page before entering the error path
885 */
886 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
887 !sk_wmem_schedule(sk, amount)) {
888 if (amount <= PAGE_SIZE)
889 goto nomem;
890
891 amount = PAGE_SIZE;
892 if (!sk_wmem_schedule(sk, amount))
893 goto nomem;
894 }
895
896 reserve:
897 msk->wmem_reserved = amount;
898 sk->sk_forward_alloc -= amount;
899 return;
900
901 nomem:
902 /* we will wait for memory on next allocation */
903 msk->wmem_reserved = -1;
904 }
905
906 static void __mptcp_update_wmem(struct sock *sk)
907 {
908 struct mptcp_sock *msk = mptcp_sk(sk);
909
910 if (!msk->wmem_reserved)
911 return;
912
913 if (msk->wmem_reserved < 0)
914 msk->wmem_reserved = 0;
915 if (msk->wmem_reserved > 0) {
916 sk->sk_forward_alloc += msk->wmem_reserved;
917 msk->wmem_reserved = 0;
918 }
919 }
920
921 static bool mptcp_wmem_alloc(struct sock *sk, int size)
922 {
923 struct mptcp_sock *msk = mptcp_sk(sk);
924
925 /* check for pre-existing error condition */
926 if (msk->wmem_reserved < 0)
927 return false;
928
929 if (msk->wmem_reserved >= size)
930 goto account;
931
932 mptcp_data_lock(sk);
933 if (!sk_wmem_schedule(sk, size)) {
934 mptcp_data_unlock(sk);
935 return false;
936 }
937
938 sk->sk_forward_alloc -= size;
939 msk->wmem_reserved += size;
940 mptcp_data_unlock(sk);
941
942 account:
943 msk->wmem_reserved -= size;
944 return true;
945 }
946
947 static void mptcp_wmem_uncharge(struct sock *sk, int size)
948 {
949 struct mptcp_sock *msk = mptcp_sk(sk);
950
951 if (msk->wmem_reserved < 0)
952 msk->wmem_reserved = 0;
953 msk->wmem_reserved += size;
954 }
955
956 static void mptcp_mem_reclaim_partial(struct sock *sk)
957 {
958 struct mptcp_sock *msk = mptcp_sk(sk);
959
960 /* if we are experiencing a transint allocation error,
961 * the forward allocation memory has been already
962 * released
963 */
964 if (msk->wmem_reserved < 0)
965 return;
966
967 mptcp_data_lock(sk);
968 sk->sk_forward_alloc += msk->wmem_reserved;
969 sk_mem_reclaim_partial(sk);
970 msk->wmem_reserved = sk->sk_forward_alloc;
971 sk->sk_forward_alloc = 0;
972 mptcp_data_unlock(sk);
973 }
974
975 static void dfrag_uncharge(struct sock *sk, int len)
976 {
977 sk_mem_uncharge(sk, len);
978 sk_wmem_queued_add(sk, -len);
979 }
980
981 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
982 {
983 int len = dfrag->data_len + dfrag->overhead;
984
985 list_del(&dfrag->list);
986 dfrag_uncharge(sk, len);
987 put_page(dfrag->page);
988 }
989
990 static void __mptcp_clean_una(struct sock *sk)
991 {
992 struct mptcp_sock *msk = mptcp_sk(sk);
993 struct mptcp_data_frag *dtmp, *dfrag;
994 bool cleaned = false;
995 u64 snd_una;
996
997 /* on fallback we just need to ignore snd_una, as this is really
998 * plain TCP
999 */
1000 if (__mptcp_check_fallback(msk))
1001 msk->snd_una = READ_ONCE(msk->snd_nxt);
1002
1003 snd_una = msk->snd_una;
1004 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1005 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1006 break;
1007
1008 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1009 break;
1010 dfrag_clear(sk, dfrag);
1011 cleaned = true;
1012 }
1013
1014 dfrag = mptcp_rtx_head(sk);
1015 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1016 u64 delta = snd_una - dfrag->data_seq;
1017
1018 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1019 goto out;
1020
1021 dfrag->data_seq += delta;
1022 dfrag->offset += delta;
1023 dfrag->data_len -= delta;
1024 dfrag->already_sent -= delta;
1025
1026 dfrag_uncharge(sk, delta);
1027 cleaned = true;
1028 }
1029
1030 out:
1031 if (cleaned) {
1032 if (tcp_under_memory_pressure(sk)) {
1033 __mptcp_update_wmem(sk);
1034 sk_mem_reclaim_partial(sk);
1035 }
1036
1037 if (sk_stream_is_writeable(sk)) {
1038 /* pairs with memory barrier in mptcp_poll */
1039 smp_mb();
1040 if (test_and_clear_bit(MPTCP_NOSPACE, &msk->flags))
1041 sk_stream_write_space(sk);
1042 }
1043 }
1044
1045 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1046 if (msk->timer_ival)
1047 mptcp_stop_timer(sk);
1048 } else {
1049 mptcp_reset_timer(sk);
1050 }
1051 }
1052
1053 static void mptcp_enter_memory_pressure(struct sock *sk)
1054 {
1055 struct mptcp_subflow_context *subflow;
1056 struct mptcp_sock *msk = mptcp_sk(sk);
1057 bool first = true;
1058
1059 sk_stream_moderate_sndbuf(sk);
1060 mptcp_for_each_subflow(msk, subflow) {
1061 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1062
1063 if (first)
1064 tcp_enter_memory_pressure(ssk);
1065 sk_stream_moderate_sndbuf(ssk);
1066 first = false;
1067 }
1068 }
1069
1070 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1071 * data
1072 */
1073 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1074 {
1075 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1076 pfrag, sk->sk_allocation)))
1077 return true;
1078
1079 mptcp_enter_memory_pressure(sk);
1080 return false;
1081 }
1082
1083 static struct mptcp_data_frag *
1084 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1085 int orig_offset)
1086 {
1087 int offset = ALIGN(orig_offset, sizeof(long));
1088 struct mptcp_data_frag *dfrag;
1089
1090 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1091 dfrag->data_len = 0;
1092 dfrag->data_seq = msk->write_seq;
1093 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1094 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1095 dfrag->already_sent = 0;
1096 dfrag->page = pfrag->page;
1097
1098 return dfrag;
1099 }
1100
1101 struct mptcp_sendmsg_info {
1102 int mss_now;
1103 int size_goal;
1104 u16 limit;
1105 u16 sent;
1106 unsigned int flags;
1107 };
1108
1109 static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1110 int avail_size)
1111 {
1112 u64 window_end = mptcp_wnd_end(msk);
1113
1114 if (__mptcp_check_fallback(msk))
1115 return avail_size;
1116
1117 if (!before64(data_seq + avail_size, window_end)) {
1118 u64 allowed_size = window_end - data_seq;
1119
1120 return min_t(unsigned int, allowed_size, avail_size);
1121 }
1122
1123 return avail_size;
1124 }
1125
1126 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1127 {
1128 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1129
1130 if (!mpext)
1131 return false;
1132 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1133 return true;
1134 }
1135
1136 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1137 {
1138 struct sk_buff *skb;
1139
1140 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1141 if (likely(skb)) {
1142 if (likely(__mptcp_add_ext(skb, gfp))) {
1143 skb_reserve(skb, MAX_TCP_HEADER);
1144 skb->reserved_tailroom = skb->end - skb->tail;
1145 return skb;
1146 }
1147 __kfree_skb(skb);
1148 } else {
1149 mptcp_enter_memory_pressure(sk);
1150 }
1151 return NULL;
1152 }
1153
1154 static bool mptcp_tx_cache_refill(struct sock *sk, int size,
1155 struct sk_buff_head *skbs, int *total_ts)
1156 {
1157 struct mptcp_sock *msk = mptcp_sk(sk);
1158 struct sk_buff *skb;
1159 int space_needed;
1160
1161 if (unlikely(tcp_under_memory_pressure(sk))) {
1162 mptcp_mem_reclaim_partial(sk);
1163
1164 /* under pressure pre-allocate at most a single skb */
1165 if (msk->skb_tx_cache.qlen)
1166 return true;
1167 space_needed = msk->size_goal_cache;
1168 } else {
1169 space_needed = msk->tx_pending_data + size -
1170 msk->skb_tx_cache.qlen * msk->size_goal_cache;
1171 }
1172
1173 while (space_needed > 0) {
1174 skb = __mptcp_do_alloc_tx_skb(sk, sk->sk_allocation);
1175 if (unlikely(!skb)) {
1176 /* under memory pressure, try to pass the caller a
1177 * single skb to allow forward progress
1178 */
1179 while (skbs->qlen > 1) {
1180 skb = __skb_dequeue_tail(skbs);
1181 *total_ts -= skb->truesize;
1182 __kfree_skb(skb);
1183 }
1184 return skbs->qlen > 0;
1185 }
1186
1187 *total_ts += skb->truesize;
1188 __skb_queue_tail(skbs, skb);
1189 space_needed -= msk->size_goal_cache;
1190 }
1191 return true;
1192 }
1193
1194 static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1195 {
1196 struct mptcp_sock *msk = mptcp_sk(sk);
1197 struct sk_buff *skb;
1198
1199 if (ssk->sk_tx_skb_cache) {
1200 skb = ssk->sk_tx_skb_cache;
1201 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1202 !__mptcp_add_ext(skb, gfp)))
1203 return false;
1204 return true;
1205 }
1206
1207 skb = skb_peek(&msk->skb_tx_cache);
1208 if (skb) {
1209 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1210 skb = __skb_dequeue(&msk->skb_tx_cache);
1211 if (WARN_ON_ONCE(!skb))
1212 return false;
1213
1214 mptcp_wmem_uncharge(sk, skb->truesize);
1215 ssk->sk_tx_skb_cache = skb;
1216 return true;
1217 }
1218
1219 /* over memory limit, no point to try to allocate a new skb */
1220 return false;
1221 }
1222
1223 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1224 if (!skb)
1225 return false;
1226
1227 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1228 ssk->sk_tx_skb_cache = skb;
1229 return true;
1230 }
1231 kfree_skb(skb);
1232 return false;
1233 }
1234
1235 static bool mptcp_must_reclaim_memory(struct sock *sk, struct sock *ssk)
1236 {
1237 return !ssk->sk_tx_skb_cache &&
1238 !skb_peek(&mptcp_sk(sk)->skb_tx_cache) &&
1239 tcp_under_memory_pressure(sk);
1240 }
1241
1242 static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk)
1243 {
1244 if (unlikely(mptcp_must_reclaim_memory(sk, ssk)))
1245 mptcp_mem_reclaim_partial(sk);
1246 return __mptcp_alloc_tx_skb(sk, ssk, sk->sk_allocation);
1247 }
1248
1249 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1250 struct mptcp_data_frag *dfrag,
1251 struct mptcp_sendmsg_info *info)
1252 {
1253 u64 data_seq = dfrag->data_seq + info->sent;
1254 struct mptcp_sock *msk = mptcp_sk(sk);
1255 bool zero_window_probe = false;
1256 struct mptcp_ext *mpext = NULL;
1257 struct sk_buff *skb, *tail;
1258 bool can_collapse = false;
1259 int size_bias = 0;
1260 int avail_size;
1261 size_t ret = 0;
1262
1263 pr_debug("msk=%p ssk=%p sending dfrag at seq=%lld len=%d already sent=%d",
1264 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1265
1266 /* compute send limit */
1267 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1268 avail_size = info->size_goal;
1269 msk->size_goal_cache = info->size_goal;
1270 skb = tcp_write_queue_tail(ssk);
1271 if (skb) {
1272 /* Limit the write to the size available in the
1273 * current skb, if any, so that we create at most a new skb.
1274 * Explicitly tells TCP internals to avoid collapsing on later
1275 * queue management operation, to avoid breaking the ext <->
1276 * SSN association set here
1277 */
1278 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1279 can_collapse = (info->size_goal - skb->len > 0) &&
1280 mptcp_skb_can_collapse_to(data_seq, skb, mpext);
1281 if (!can_collapse) {
1282 TCP_SKB_CB(skb)->eor = 1;
1283 } else {
1284 size_bias = skb->len;
1285 avail_size = info->size_goal - skb->len;
1286 }
1287 }
1288
1289 /* Zero window and all data acked? Probe. */
1290 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1291 if (avail_size == 0) {
1292 u64 snd_una = READ_ONCE(msk->snd_una);
1293
1294 if (skb || snd_una != msk->snd_nxt)
1295 return 0;
1296 zero_window_probe = true;
1297 data_seq = snd_una - 1;
1298 avail_size = 1;
1299 }
1300
1301 if (WARN_ON_ONCE(info->sent > info->limit ||
1302 info->limit > dfrag->data_len))
1303 return 0;
1304
1305 ret = info->limit - info->sent;
1306 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1307 dfrag->page, dfrag->offset + info->sent, &ret);
1308 if (!tail) {
1309 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1310 return -ENOMEM;
1311 }
1312
1313 /* if the tail skb is still the cached one, collapsing really happened.
1314 */
1315 if (skb == tail) {
1316 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1317 mpext->data_len += ret;
1318 WARN_ON_ONCE(!can_collapse);
1319 WARN_ON_ONCE(zero_window_probe);
1320 goto out;
1321 }
1322
1323 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1324 if (WARN_ON_ONCE(!mpext)) {
1325 /* should never reach here, stream corrupted */
1326 return -EINVAL;
1327 }
1328
1329 memset(mpext, 0, sizeof(*mpext));
1330 mpext->data_seq = data_seq;
1331 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1332 mpext->data_len = ret;
1333 mpext->use_map = 1;
1334 mpext->dsn64 = 1;
1335
1336 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1337 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1338 mpext->dsn64);
1339
1340 if (zero_window_probe) {
1341 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1342 mpext->frozen = 1;
1343 ret = 0;
1344 tcp_push_pending_frames(ssk);
1345 }
1346 out:
1347 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1348 return ret;
1349 }
1350
1351 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1352 sizeof(struct tcphdr) - \
1353 MAX_TCP_OPTION_SPACE - \
1354 sizeof(struct ipv6hdr) - \
1355 sizeof(struct frag_hdr))
1356
1357 struct subflow_send_info {
1358 struct sock *ssk;
1359 u64 ratio;
1360 };
1361
1362 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk,
1363 u32 *sndbuf)
1364 {
1365 struct subflow_send_info send_info[2];
1366 struct mptcp_subflow_context *subflow;
1367 int i, nr_active = 0;
1368 struct sock *ssk;
1369 u64 ratio;
1370 u32 pace;
1371
1372 sock_owned_by_me((struct sock *)msk);
1373
1374 *sndbuf = 0;
1375 if (__mptcp_check_fallback(msk)) {
1376 if (!msk->first)
1377 return NULL;
1378 *sndbuf = msk->first->sk_sndbuf;
1379 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1380 }
1381
1382 /* re-use last subflow, if the burst allow that */
1383 if (msk->last_snd && msk->snd_burst > 0 &&
1384 sk_stream_memory_free(msk->last_snd) &&
1385 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd))) {
1386 mptcp_for_each_subflow(msk, subflow) {
1387 ssk = mptcp_subflow_tcp_sock(subflow);
1388 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1389 }
1390 return msk->last_snd;
1391 }
1392
1393 /* pick the subflow with the lower wmem/wspace ratio */
1394 for (i = 0; i < 2; ++i) {
1395 send_info[i].ssk = NULL;
1396 send_info[i].ratio = -1;
1397 }
1398 mptcp_for_each_subflow(msk, subflow) {
1399 ssk = mptcp_subflow_tcp_sock(subflow);
1400 if (!mptcp_subflow_active(subflow))
1401 continue;
1402
1403 nr_active += !subflow->backup;
1404 *sndbuf = max(tcp_sk(ssk)->snd_wnd, *sndbuf);
1405 if (!sk_stream_memory_free(subflow->tcp_sock))
1406 continue;
1407
1408 pace = READ_ONCE(ssk->sk_pacing_rate);
1409 if (!pace)
1410 continue;
1411
1412 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1413 pace);
1414 if (ratio < send_info[subflow->backup].ratio) {
1415 send_info[subflow->backup].ssk = ssk;
1416 send_info[subflow->backup].ratio = ratio;
1417 }
1418 }
1419
1420 pr_debug("msk=%p nr_active=%d ssk=%p:%lld backup=%p:%lld",
1421 msk, nr_active, send_info[0].ssk, send_info[0].ratio,
1422 send_info[1].ssk, send_info[1].ratio);
1423
1424 /* pick the best backup if no other subflow is active */
1425 if (!nr_active)
1426 send_info[0].ssk = send_info[1].ssk;
1427
1428 if (send_info[0].ssk) {
1429 msk->last_snd = send_info[0].ssk;
1430 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1431 sk_stream_wspace(msk->last_snd));
1432 return msk->last_snd;
1433 }
1434 return NULL;
1435 }
1436
1437 static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1438 struct mptcp_sendmsg_info *info)
1439 {
1440 mptcp_set_timeout(sk, ssk);
1441 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1442 release_sock(ssk);
1443 }
1444
1445 static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1446 {
1447 struct sock *prev_ssk = NULL, *ssk = NULL;
1448 struct mptcp_sock *msk = mptcp_sk(sk);
1449 struct mptcp_sendmsg_info info = {
1450 .flags = flags,
1451 };
1452 struct mptcp_data_frag *dfrag;
1453 int len, copied = 0;
1454 u32 sndbuf;
1455
1456 while ((dfrag = mptcp_send_head(sk))) {
1457 info.sent = dfrag->already_sent;
1458 info.limit = dfrag->data_len;
1459 len = dfrag->data_len - dfrag->already_sent;
1460 while (len > 0) {
1461 int ret = 0;
1462
1463 prev_ssk = ssk;
1464 __mptcp_flush_join_list(msk);
1465 ssk = mptcp_subflow_get_send(msk, &sndbuf);
1466
1467 /* do auto tuning */
1468 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1469 sndbuf > READ_ONCE(sk->sk_sndbuf))
1470 WRITE_ONCE(sk->sk_sndbuf, sndbuf);
1471
1472 /* try to keep the subflow socket lock across
1473 * consecutive xmit on the same socket
1474 */
1475 if (ssk != prev_ssk && prev_ssk)
1476 mptcp_push_release(sk, prev_ssk, &info);
1477 if (!ssk)
1478 goto out;
1479
1480 if (ssk != prev_ssk || !prev_ssk)
1481 lock_sock(ssk);
1482
1483 /* keep it simple and always provide a new skb for the
1484 * subflow, even if we will not use it when collapsing
1485 * on the pending one
1486 */
1487 if (!mptcp_alloc_tx_skb(sk, ssk)) {
1488 mptcp_push_release(sk, ssk, &info);
1489 goto out;
1490 }
1491
1492 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1493 if (ret <= 0) {
1494 mptcp_push_release(sk, ssk, &info);
1495 goto out;
1496 }
1497
1498 info.sent += ret;
1499 dfrag->already_sent += ret;
1500 msk->snd_nxt += ret;
1501 msk->snd_burst -= ret;
1502 msk->tx_pending_data -= ret;
1503 copied += ret;
1504 len -= ret;
1505 }
1506 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1507 }
1508
1509 /* at this point we held the socket lock for the last subflow we used */
1510 if (ssk)
1511 mptcp_push_release(sk, ssk, &info);
1512
1513 out:
1514 if (copied) {
1515 /* start the timer, if it's not pending */
1516 if (!mptcp_timer_pending(sk))
1517 mptcp_reset_timer(sk);
1518 __mptcp_check_send_data_fin(sk);
1519 }
1520 }
1521
1522 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1523 {
1524 struct mptcp_sock *msk = mptcp_sk(sk);
1525 struct mptcp_sendmsg_info info;
1526 struct mptcp_data_frag *dfrag;
1527 int len, copied = 0;
1528
1529 info.flags = 0;
1530 while ((dfrag = mptcp_send_head(sk))) {
1531 info.sent = dfrag->already_sent;
1532 info.limit = dfrag->data_len;
1533 len = dfrag->data_len - dfrag->already_sent;
1534 while (len > 0) {
1535 int ret = 0;
1536
1537 /* do auto tuning */
1538 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
1539 ssk->sk_sndbuf > READ_ONCE(sk->sk_sndbuf))
1540 WRITE_ONCE(sk->sk_sndbuf, ssk->sk_sndbuf);
1541
1542 if (unlikely(mptcp_must_reclaim_memory(sk, ssk))) {
1543 __mptcp_update_wmem(sk);
1544 sk_mem_reclaim_partial(sk);
1545 }
1546 if (!__mptcp_alloc_tx_skb(sk, ssk, GFP_ATOMIC))
1547 goto out;
1548
1549 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1550 if (ret <= 0)
1551 goto out;
1552
1553 info.sent += ret;
1554 dfrag->already_sent += ret;
1555 msk->snd_nxt += ret;
1556 msk->snd_burst -= ret;
1557 msk->tx_pending_data -= ret;
1558 copied += ret;
1559 len -= ret;
1560 }
1561 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1562 }
1563
1564 out:
1565 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1566 * not going to flush it via release_sock()
1567 */
1568 __mptcp_update_wmem(sk);
1569 if (copied) {
1570 mptcp_set_timeout(sk, ssk);
1571 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1572 info.size_goal);
1573 if (!mptcp_timer_pending(sk))
1574 mptcp_reset_timer(sk);
1575
1576 if (msk->snd_data_fin_enable &&
1577 msk->snd_nxt + 1 == msk->write_seq)
1578 mptcp_schedule_work(sk);
1579 }
1580 }
1581
1582 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1583 {
1584 struct mptcp_sock *msk = mptcp_sk(sk);
1585 struct page_frag *pfrag;
1586 size_t copied = 0;
1587 int ret = 0;
1588 long timeo;
1589
1590 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
1591 return -EOPNOTSUPP;
1592
1593 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1594
1595 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1596
1597 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1598 ret = sk_stream_wait_connect(sk, &timeo);
1599 if (ret)
1600 goto out;
1601 }
1602
1603 pfrag = sk_page_frag(sk);
1604
1605 while (msg_data_left(msg)) {
1606 int total_ts, frag_truesize = 0;
1607 struct mptcp_data_frag *dfrag;
1608 struct sk_buff_head skbs;
1609 bool dfrag_collapsed;
1610 size_t psize, offset;
1611
1612 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1613 ret = -EPIPE;
1614 goto out;
1615 }
1616
1617 /* reuse tail pfrag, if possible, or carve a new one from the
1618 * page allocator
1619 */
1620 dfrag = mptcp_pending_tail(sk);
1621 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1622 if (!dfrag_collapsed) {
1623 if (!sk_stream_memory_free(sk))
1624 goto wait_for_memory;
1625
1626 if (!mptcp_page_frag_refill(sk, pfrag))
1627 goto wait_for_memory;
1628
1629 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1630 frag_truesize = dfrag->overhead;
1631 }
1632
1633 /* we do not bound vs wspace, to allow a single packet.
1634 * memory accounting will prevent execessive memory usage
1635 * anyway
1636 */
1637 offset = dfrag->offset + dfrag->data_len;
1638 psize = pfrag->size - offset;
1639 psize = min_t(size_t, psize, msg_data_left(msg));
1640 total_ts = psize + frag_truesize;
1641 __skb_queue_head_init(&skbs);
1642 if (!mptcp_tx_cache_refill(sk, psize, &skbs, &total_ts))
1643 goto wait_for_memory;
1644
1645 if (!mptcp_wmem_alloc(sk, total_ts)) {
1646 __skb_queue_purge(&skbs);
1647 goto wait_for_memory;
1648 }
1649
1650 skb_queue_splice_tail(&skbs, &msk->skb_tx_cache);
1651 if (copy_page_from_iter(dfrag->page, offset, psize,
1652 &msg->msg_iter) != psize) {
1653 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1654 ret = -EFAULT;
1655 goto out;
1656 }
1657
1658 /* data successfully copied into the write queue */
1659 copied += psize;
1660 dfrag->data_len += psize;
1661 frag_truesize += psize;
1662 pfrag->offset += frag_truesize;
1663 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1664 msk->tx_pending_data += psize;
1665
1666 /* charge data on mptcp pending queue to the msk socket
1667 * Note: we charge such data both to sk and ssk
1668 */
1669 sk_wmem_queued_add(sk, frag_truesize);
1670 if (!dfrag_collapsed) {
1671 get_page(dfrag->page);
1672 list_add_tail(&dfrag->list, &msk->rtx_queue);
1673 if (!msk->first_pending)
1674 WRITE_ONCE(msk->first_pending, dfrag);
1675 }
1676 pr_debug("msk=%p dfrag at seq=%lld len=%d sent=%d new=%d", msk,
1677 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1678 !dfrag_collapsed);
1679
1680 continue;
1681
1682 wait_for_memory:
1683 set_bit(MPTCP_NOSPACE, &msk->flags);
1684 __mptcp_push_pending(sk, msg->msg_flags);
1685 ret = sk_stream_wait_memory(sk, &timeo);
1686 if (ret)
1687 goto out;
1688 }
1689
1690 if (copied)
1691 __mptcp_push_pending(sk, msg->msg_flags);
1692
1693 out:
1694 release_sock(sk);
1695 return copied ? : ret;
1696 }
1697
1698 static void mptcp_wait_data(struct sock *sk, long *timeo)
1699 {
1700 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1701 struct mptcp_sock *msk = mptcp_sk(sk);
1702
1703 add_wait_queue(sk_sleep(sk), &wait);
1704 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1705
1706 sk_wait_event(sk, timeo,
1707 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1708
1709 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1710 remove_wait_queue(sk_sleep(sk), &wait);
1711 }
1712
1713 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1714 struct msghdr *msg,
1715 size_t len)
1716 {
1717 struct sk_buff *skb;
1718 int copied = 0;
1719
1720 while ((skb = skb_peek(&msk->receive_queue)) != NULL) {
1721 u32 offset = MPTCP_SKB_CB(skb)->offset;
1722 u32 data_len = skb->len - offset;
1723 u32 count = min_t(size_t, len - copied, data_len);
1724 int err;
1725
1726 err = skb_copy_datagram_msg(skb, offset, msg, count);
1727 if (unlikely(err < 0)) {
1728 if (!copied)
1729 return err;
1730 break;
1731 }
1732
1733 copied += count;
1734
1735 if (count < data_len) {
1736 MPTCP_SKB_CB(skb)->offset += count;
1737 break;
1738 }
1739
1740 /* we will bulk release the skb memory later */
1741 skb->destructor = NULL;
1742 msk->rmem_released += skb->truesize;
1743 __skb_unlink(skb, &msk->receive_queue);
1744 __kfree_skb(skb);
1745
1746 if (copied >= len)
1747 break;
1748 }
1749
1750 return copied;
1751 }
1752
1753 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1754 *
1755 * Only difference: Use highest rtt estimate of the subflows in use.
1756 */
1757 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1758 {
1759 struct mptcp_subflow_context *subflow;
1760 struct sock *sk = (struct sock *)msk;
1761 u32 time, advmss = 1;
1762 u64 rtt_us, mstamp;
1763
1764 sock_owned_by_me(sk);
1765
1766 if (copied <= 0)
1767 return;
1768
1769 msk->rcvq_space.copied += copied;
1770
1771 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1772 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1773
1774 rtt_us = msk->rcvq_space.rtt_us;
1775 if (rtt_us && time < (rtt_us >> 3))
1776 return;
1777
1778 rtt_us = 0;
1779 mptcp_for_each_subflow(msk, subflow) {
1780 const struct tcp_sock *tp;
1781 u64 sf_rtt_us;
1782 u32 sf_advmss;
1783
1784 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1785
1786 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1787 sf_advmss = READ_ONCE(tp->advmss);
1788
1789 rtt_us = max(sf_rtt_us, rtt_us);
1790 advmss = max(sf_advmss, advmss);
1791 }
1792
1793 msk->rcvq_space.rtt_us = rtt_us;
1794 if (time < (rtt_us >> 3) || rtt_us == 0)
1795 return;
1796
1797 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1798 goto new_measure;
1799
1800 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1801 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1802 int rcvmem, rcvbuf;
1803 u64 rcvwin, grow;
1804
1805 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1806
1807 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1808
1809 do_div(grow, msk->rcvq_space.space);
1810 rcvwin += (grow << 1);
1811
1812 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1813 while (tcp_win_from_space(sk, rcvmem) < advmss)
1814 rcvmem += 128;
1815
1816 do_div(rcvwin, advmss);
1817 rcvbuf = min_t(u64, rcvwin * rcvmem,
1818 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1819
1820 if (rcvbuf > sk->sk_rcvbuf) {
1821 u32 window_clamp;
1822
1823 window_clamp = tcp_win_from_space(sk, rcvbuf);
1824 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1825
1826 /* Make subflows follow along. If we do not do this, we
1827 * get drops at subflow level if skbs can't be moved to
1828 * the mptcp rx queue fast enough (announced rcv_win can
1829 * exceed ssk->sk_rcvbuf).
1830 */
1831 mptcp_for_each_subflow(msk, subflow) {
1832 struct sock *ssk;
1833 bool slow;
1834
1835 ssk = mptcp_subflow_tcp_sock(subflow);
1836 slow = lock_sock_fast(ssk);
1837 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1838 tcp_sk(ssk)->window_clamp = window_clamp;
1839 tcp_cleanup_rbuf(ssk, 1);
1840 unlock_sock_fast(ssk, slow);
1841 }
1842 }
1843 }
1844
1845 msk->rcvq_space.space = msk->rcvq_space.copied;
1846 new_measure:
1847 msk->rcvq_space.copied = 0;
1848 msk->rcvq_space.time = mstamp;
1849 }
1850
1851 static void __mptcp_update_rmem(struct sock *sk)
1852 {
1853 struct mptcp_sock *msk = mptcp_sk(sk);
1854
1855 if (!msk->rmem_released)
1856 return;
1857
1858 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1859 sk_mem_uncharge(sk, msk->rmem_released);
1860 msk->rmem_released = 0;
1861 }
1862
1863 static void __mptcp_splice_receive_queue(struct sock *sk)
1864 {
1865 struct mptcp_sock *msk = mptcp_sk(sk);
1866
1867 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1868 }
1869
1870 static bool __mptcp_move_skbs(struct mptcp_sock *msk, unsigned int rcv)
1871 {
1872 struct sock *sk = (struct sock *)msk;
1873 unsigned int moved = 0;
1874 bool ret, done;
1875
1876 __mptcp_flush_join_list(msk);
1877 do {
1878 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1879 bool slowpath;
1880
1881 /* we can have data pending in the subflows only if the msk
1882 * receive buffer was full at subflow_data_ready() time,
1883 * that is an unlikely slow path.
1884 */
1885 if (likely(!ssk))
1886 break;
1887
1888 slowpath = lock_sock_fast(ssk);
1889 mptcp_data_lock(sk);
1890 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1891 mptcp_data_unlock(sk);
1892 if (moved && rcv) {
1893 WRITE_ONCE(msk->rmem_pending, min(rcv, moved));
1894 tcp_cleanup_rbuf(ssk, 1);
1895 WRITE_ONCE(msk->rmem_pending, 0);
1896 }
1897 unlock_sock_fast(ssk, slowpath);
1898 } while (!done);
1899
1900 /* acquire the data lock only if some input data is pending */
1901 ret = moved > 0;
1902 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1903 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1904 mptcp_data_lock(sk);
1905 __mptcp_update_rmem(sk);
1906 ret |= __mptcp_ofo_queue(msk);
1907 __mptcp_splice_receive_queue(sk);
1908 mptcp_data_unlock(sk);
1909 }
1910 if (ret)
1911 mptcp_check_data_fin((struct sock *)msk);
1912 return !skb_queue_empty(&msk->receive_queue);
1913 }
1914
1915 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1916 int nonblock, int flags, int *addr_len)
1917 {
1918 struct mptcp_sock *msk = mptcp_sk(sk);
1919 int copied = 0;
1920 int target;
1921 long timeo;
1922
1923 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1924 return -EOPNOTSUPP;
1925
1926 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1927 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1928 copied = -ENOTCONN;
1929 goto out_err;
1930 }
1931
1932 timeo = sock_rcvtimeo(sk, nonblock);
1933
1934 len = min_t(size_t, len, INT_MAX);
1935 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1936
1937 while (copied < len) {
1938 int bytes_read, old_space;
1939
1940 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1941 if (unlikely(bytes_read < 0)) {
1942 if (!copied)
1943 copied = bytes_read;
1944 goto out_err;
1945 }
1946
1947 copied += bytes_read;
1948
1949 if (skb_queue_empty(&msk->receive_queue) &&
1950 __mptcp_move_skbs(msk, len - copied))
1951 continue;
1952
1953 /* be sure to advertise window change */
1954 old_space = READ_ONCE(msk->old_wspace);
1955 if ((tcp_space(sk) - old_space) >= old_space)
1956 mptcp_cleanup_rbuf(msk);
1957
1958 /* only the master socket status is relevant here. The exit
1959 * conditions mirror closely tcp_recvmsg()
1960 */
1961 if (copied >= target)
1962 break;
1963
1964 if (copied) {
1965 if (sk->sk_err ||
1966 sk->sk_state == TCP_CLOSE ||
1967 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1968 !timeo ||
1969 signal_pending(current))
1970 break;
1971 } else {
1972 if (sk->sk_err) {
1973 copied = sock_error(sk);
1974 break;
1975 }
1976
1977 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1978 mptcp_check_for_eof(msk);
1979
1980 if (sk->sk_shutdown & RCV_SHUTDOWN) {
1981 /* race breaker: the shutdown could be after the
1982 * previous receive queue check
1983 */
1984 if (__mptcp_move_skbs(msk, len - copied))
1985 continue;
1986 break;
1987 }
1988
1989 if (sk->sk_state == TCP_CLOSE) {
1990 copied = -ENOTCONN;
1991 break;
1992 }
1993
1994 if (!timeo) {
1995 copied = -EAGAIN;
1996 break;
1997 }
1998
1999 if (signal_pending(current)) {
2000 copied = sock_intr_errno(timeo);
2001 break;
2002 }
2003 }
2004
2005 pr_debug("block timeout %ld", timeo);
2006 mptcp_wait_data(sk, &timeo);
2007 }
2008
2009 if (skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2010 skb_queue_empty(&msk->receive_queue)) {
2011 /* entire backlog drained, clear DATA_READY. */
2012 clear_bit(MPTCP_DATA_READY, &msk->flags);
2013
2014 /* .. race-breaker: ssk might have gotten new data
2015 * after last __mptcp_move_skbs() returned false.
2016 */
2017 if (unlikely(__mptcp_move_skbs(msk, 0)))
2018 set_bit(MPTCP_DATA_READY, &msk->flags);
2019 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
2020 /* data to read but mptcp_wait_data() cleared DATA_READY */
2021 set_bit(MPTCP_DATA_READY, &msk->flags);
2022 }
2023 out_err:
2024 pr_debug("msk=%p data_ready=%d rx queue empty=%d copied=%d",
2025 msk, test_bit(MPTCP_DATA_READY, &msk->flags),
2026 skb_queue_empty_lockless(&sk->sk_receive_queue), copied);
2027 mptcp_rcv_space_adjust(msk, copied);
2028
2029 release_sock(sk);
2030 return copied;
2031 }
2032
2033 static void mptcp_retransmit_handler(struct sock *sk)
2034 {
2035 struct mptcp_sock *msk = mptcp_sk(sk);
2036
2037 set_bit(MPTCP_WORK_RTX, &msk->flags);
2038 mptcp_schedule_work(sk);
2039 }
2040
2041 static void mptcp_retransmit_timer(struct timer_list *t)
2042 {
2043 struct inet_connection_sock *icsk = from_timer(icsk, t,
2044 icsk_retransmit_timer);
2045 struct sock *sk = &icsk->icsk_inet.sk;
2046
2047 bh_lock_sock(sk);
2048 if (!sock_owned_by_user(sk)) {
2049 mptcp_retransmit_handler(sk);
2050 } else {
2051 /* delegate our work to tcp_release_cb() */
2052 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
2053 &sk->sk_tsq_flags))
2054 sock_hold(sk);
2055 }
2056 bh_unlock_sock(sk);
2057 sock_put(sk);
2058 }
2059
2060 static void mptcp_timeout_timer(struct timer_list *t)
2061 {
2062 struct sock *sk = from_timer(sk, t, sk_timer);
2063
2064 mptcp_schedule_work(sk);
2065 sock_put(sk);
2066 }
2067
2068 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2069 * level.
2070 *
2071 * A backup subflow is returned only if that is the only kind available.
2072 */
2073 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2074 {
2075 struct mptcp_subflow_context *subflow;
2076 struct sock *backup = NULL;
2077
2078 sock_owned_by_me((const struct sock *)msk);
2079
2080 if (__mptcp_check_fallback(msk))
2081 return NULL;
2082
2083 mptcp_for_each_subflow(msk, subflow) {
2084 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2085
2086 if (!mptcp_subflow_active(subflow))
2087 continue;
2088
2089 /* still data outstanding at TCP level? Don't retransmit. */
2090 if (!tcp_write_queue_empty(ssk)) {
2091 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2092 continue;
2093 return NULL;
2094 }
2095
2096 if (subflow->backup) {
2097 if (!backup)
2098 backup = ssk;
2099 continue;
2100 }
2101
2102 return ssk;
2103 }
2104
2105 return backup;
2106 }
2107
2108 static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2109 {
2110 if (msk->subflow) {
2111 iput(SOCK_INODE(msk->subflow));
2112 msk->subflow = NULL;
2113 }
2114 }
2115
2116 /* subflow sockets can be either outgoing (connect) or incoming
2117 * (accept).
2118 *
2119 * Outgoing subflows use in-kernel sockets.
2120 * Incoming subflows do not have their own 'struct socket' allocated,
2121 * so we need to use tcp_close() after detaching them from the mptcp
2122 * parent socket.
2123 */
2124 void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2125 struct mptcp_subflow_context *subflow)
2126 {
2127 struct mptcp_sock *msk = mptcp_sk(sk);
2128
2129 list_del(&subflow->node);
2130
2131 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2132
2133 /* if we are invoked by the msk cleanup code, the subflow is
2134 * already orphaned
2135 */
2136 if (ssk->sk_socket)
2137 sock_orphan(ssk);
2138
2139 subflow->disposable = 1;
2140
2141 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2142 * the ssk has been already destroyed, we just need to release the
2143 * reference owned by msk;
2144 */
2145 if (!inet_csk(ssk)->icsk_ulp_ops) {
2146 kfree_rcu(subflow, rcu);
2147 } else {
2148 /* otherwise tcp will dispose of the ssk and subflow ctx */
2149 __tcp_close(ssk, 0);
2150
2151 /* close acquired an extra ref */
2152 __sock_put(ssk);
2153 }
2154 release_sock(ssk);
2155
2156 sock_put(ssk);
2157
2158 if (ssk == msk->last_snd)
2159 msk->last_snd = NULL;
2160
2161 if (msk->subflow && ssk == msk->subflow->sk)
2162 mptcp_dispose_initial_subflow(msk);
2163 }
2164
2165 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2166 {
2167 return 0;
2168 }
2169
2170 static void pm_work(struct mptcp_sock *msk)
2171 {
2172 struct mptcp_pm_data *pm = &msk->pm;
2173
2174 spin_lock_bh(&msk->pm.lock);
2175
2176 pr_debug("msk=%p status=%x", msk, pm->status);
2177 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
2178 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
2179 mptcp_pm_nl_add_addr_received(msk);
2180 }
2181 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_SEND_ACK)) {
2182 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_SEND_ACK);
2183 mptcp_pm_nl_add_addr_send_ack(msk);
2184 }
2185 if (pm->status & BIT(MPTCP_PM_RM_ADDR_RECEIVED)) {
2186 pm->status &= ~BIT(MPTCP_PM_RM_ADDR_RECEIVED);
2187 mptcp_pm_nl_rm_addr_received(msk);
2188 }
2189 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
2190 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
2191 mptcp_pm_nl_fully_established(msk);
2192 }
2193 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
2194 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
2195 mptcp_pm_nl_subflow_established(msk);
2196 }
2197
2198 spin_unlock_bh(&msk->pm.lock);
2199 }
2200
2201 static void __mptcp_close_subflow(struct mptcp_sock *msk)
2202 {
2203 struct mptcp_subflow_context *subflow, *tmp;
2204
2205 might_sleep();
2206
2207 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2208 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2209
2210 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2211 continue;
2212
2213 __mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2214 }
2215 }
2216
2217 static bool mptcp_check_close_timeout(const struct sock *sk)
2218 {
2219 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2220 struct mptcp_subflow_context *subflow;
2221
2222 if (delta >= TCP_TIMEWAIT_LEN)
2223 return true;
2224
2225 /* if all subflows are in closed status don't bother with additional
2226 * timeout
2227 */
2228 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2229 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2230 TCP_CLOSE)
2231 return false;
2232 }
2233 return true;
2234 }
2235
2236 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2237 {
2238 struct mptcp_subflow_context *subflow, *tmp;
2239 struct sock *sk = &msk->sk.icsk_inet.sk;
2240
2241 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2242 return;
2243
2244 mptcp_token_destroy(msk);
2245
2246 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2247 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2248
2249 lock_sock(tcp_sk);
2250 if (tcp_sk->sk_state != TCP_CLOSE) {
2251 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2252 tcp_set_state(tcp_sk, TCP_CLOSE);
2253 }
2254 release_sock(tcp_sk);
2255 }
2256
2257 inet_sk_state_store(sk, TCP_CLOSE);
2258 sk->sk_shutdown = SHUTDOWN_MASK;
2259 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2260 set_bit(MPTCP_DATA_READY, &msk->flags);
2261 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2262
2263 mptcp_close_wake_up(sk);
2264 }
2265
2266 static void mptcp_worker(struct work_struct *work)
2267 {
2268 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2269 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
2270 struct mptcp_sendmsg_info info = {};
2271 struct mptcp_data_frag *dfrag;
2272 size_t copied = 0;
2273 int state, ret;
2274
2275 lock_sock(sk);
2276 state = sk->sk_state;
2277 if (unlikely(state == TCP_CLOSE))
2278 goto unlock;
2279
2280 mptcp_check_data_fin_ack(sk);
2281 __mptcp_flush_join_list(msk);
2282
2283 mptcp_check_fastclose(msk);
2284
2285 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2286 __mptcp_close_subflow(msk);
2287
2288 if (msk->pm.status)
2289 pm_work(msk);
2290
2291 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2292 mptcp_check_for_eof(msk);
2293
2294 __mptcp_check_send_data_fin(sk);
2295 mptcp_check_data_fin(sk);
2296
2297 /* There is no point in keeping around an orphaned sk timedout or
2298 * closed, but we need the msk around to reply to incoming DATA_FIN,
2299 * even if it is orphaned and in FIN_WAIT2 state
2300 */
2301 if (sock_flag(sk, SOCK_DEAD) &&
2302 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2303 inet_sk_state_store(sk, TCP_CLOSE);
2304 __mptcp_destroy_sock(sk);
2305 goto unlock;
2306 }
2307
2308 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2309 goto unlock;
2310
2311 __mptcp_clean_una(sk);
2312 dfrag = mptcp_rtx_head(sk);
2313 if (!dfrag)
2314 goto unlock;
2315
2316 ssk = mptcp_subflow_get_retrans(msk);
2317 if (!ssk)
2318 goto reset_unlock;
2319
2320 lock_sock(ssk);
2321
2322 /* limit retransmission to the bytes already sent on some subflows */
2323 info.sent = 0;
2324 info.limit = dfrag->already_sent;
2325 while (info.sent < dfrag->already_sent) {
2326 if (!mptcp_alloc_tx_skb(sk, ssk))
2327 break;
2328
2329 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2330 if (ret <= 0)
2331 break;
2332
2333 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2334 copied += ret;
2335 info.sent += ret;
2336 }
2337 if (copied)
2338 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2339 info.size_goal);
2340
2341 mptcp_set_timeout(sk, ssk);
2342 release_sock(ssk);
2343
2344 reset_unlock:
2345 if (!mptcp_timer_pending(sk))
2346 mptcp_reset_timer(sk);
2347
2348 unlock:
2349 release_sock(sk);
2350 sock_put(sk);
2351 }
2352
2353 static int __mptcp_init_sock(struct sock *sk)
2354 {
2355 struct mptcp_sock *msk = mptcp_sk(sk);
2356
2357 spin_lock_init(&msk->join_list_lock);
2358
2359 INIT_LIST_HEAD(&msk->conn_list);
2360 INIT_LIST_HEAD(&msk->join_list);
2361 INIT_LIST_HEAD(&msk->rtx_queue);
2362 INIT_WORK(&msk->work, mptcp_worker);
2363 __skb_queue_head_init(&msk->receive_queue);
2364 __skb_queue_head_init(&msk->skb_tx_cache);
2365 msk->out_of_order_queue = RB_ROOT;
2366 msk->first_pending = NULL;
2367 msk->wmem_reserved = 0;
2368 msk->rmem_released = 0;
2369 msk->tx_pending_data = 0;
2370 msk->size_goal_cache = TCP_BASE_MSS;
2371
2372 msk->ack_hint = NULL;
2373 msk->first = NULL;
2374 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2375
2376 mptcp_pm_data_init(msk);
2377
2378 /* re-use the csk retrans timer for MPTCP-level retrans */
2379 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2380 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2381 return 0;
2382 }
2383
2384 static int mptcp_init_sock(struct sock *sk)
2385 {
2386 struct net *net = sock_net(sk);
2387 int ret;
2388
2389 ret = __mptcp_init_sock(sk);
2390 if (ret)
2391 return ret;
2392
2393 if (!mptcp_is_enabled(net))
2394 return -ENOPROTOOPT;
2395
2396 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2397 return -ENOMEM;
2398
2399 ret = __mptcp_socket_create(mptcp_sk(sk));
2400 if (ret)
2401 return ret;
2402
2403 sk_sockets_allocated_inc(sk);
2404 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2405 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2406
2407 return 0;
2408 }
2409
2410 static void __mptcp_clear_xmit(struct sock *sk)
2411 {
2412 struct mptcp_sock *msk = mptcp_sk(sk);
2413 struct mptcp_data_frag *dtmp, *dfrag;
2414 struct sk_buff *skb;
2415
2416 WRITE_ONCE(msk->first_pending, NULL);
2417 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2418 dfrag_clear(sk, dfrag);
2419 while ((skb = __skb_dequeue(&msk->skb_tx_cache)) != NULL) {
2420 sk->sk_forward_alloc += skb->truesize;
2421 kfree_skb(skb);
2422 }
2423 }
2424
2425 static void mptcp_cancel_work(struct sock *sk)
2426 {
2427 struct mptcp_sock *msk = mptcp_sk(sk);
2428
2429 if (cancel_work_sync(&msk->work))
2430 __sock_put(sk);
2431 }
2432
2433 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2434 {
2435 lock_sock(ssk);
2436
2437 switch (ssk->sk_state) {
2438 case TCP_LISTEN:
2439 if (!(how & RCV_SHUTDOWN))
2440 break;
2441 fallthrough;
2442 case TCP_SYN_SENT:
2443 tcp_disconnect(ssk, O_NONBLOCK);
2444 break;
2445 default:
2446 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2447 pr_debug("Fallback");
2448 ssk->sk_shutdown |= how;
2449 tcp_shutdown(ssk, how);
2450 } else {
2451 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2452 mptcp_set_timeout(sk, ssk);
2453 tcp_send_ack(ssk);
2454 }
2455 break;
2456 }
2457
2458 release_sock(ssk);
2459 }
2460
2461 static const unsigned char new_state[16] = {
2462 /* current state: new state: action: */
2463 [0 /* (Invalid) */] = TCP_CLOSE,
2464 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2465 [TCP_SYN_SENT] = TCP_CLOSE,
2466 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2467 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2468 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2469 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2470 [TCP_CLOSE] = TCP_CLOSE,
2471 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2472 [TCP_LAST_ACK] = TCP_LAST_ACK,
2473 [TCP_LISTEN] = TCP_CLOSE,
2474 [TCP_CLOSING] = TCP_CLOSING,
2475 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2476 };
2477
2478 static int mptcp_close_state(struct sock *sk)
2479 {
2480 int next = (int)new_state[sk->sk_state];
2481 int ns = next & TCP_STATE_MASK;
2482
2483 inet_sk_state_store(sk, ns);
2484
2485 return next & TCP_ACTION_FIN;
2486 }
2487
2488 static void __mptcp_check_send_data_fin(struct sock *sk)
2489 {
2490 struct mptcp_subflow_context *subflow;
2491 struct mptcp_sock *msk = mptcp_sk(sk);
2492
2493 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2494 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2495 msk->snd_nxt, msk->write_seq);
2496
2497 /* we still need to enqueue subflows or not really shutting down,
2498 * skip this
2499 */
2500 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2501 mptcp_send_head(sk))
2502 return;
2503
2504 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2505
2506 /* fallback socket will not get data_fin/ack, can move to the next
2507 * state now
2508 */
2509 if (__mptcp_check_fallback(msk)) {
2510 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2511 inet_sk_state_store(sk, TCP_CLOSE);
2512 mptcp_close_wake_up(sk);
2513 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2514 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2515 }
2516 }
2517
2518 __mptcp_flush_join_list(msk);
2519 mptcp_for_each_subflow(msk, subflow) {
2520 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2521
2522 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2523 }
2524 }
2525
2526 static void __mptcp_wr_shutdown(struct sock *sk)
2527 {
2528 struct mptcp_sock *msk = mptcp_sk(sk);
2529
2530 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2531 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2532 !!mptcp_send_head(sk));
2533
2534 /* will be ignored by fallback sockets */
2535 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2536 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2537
2538 __mptcp_check_send_data_fin(sk);
2539 }
2540
2541 static void __mptcp_destroy_sock(struct sock *sk)
2542 {
2543 struct mptcp_subflow_context *subflow, *tmp;
2544 struct mptcp_sock *msk = mptcp_sk(sk);
2545 LIST_HEAD(conn_list);
2546
2547 pr_debug("msk=%p", msk);
2548
2549 might_sleep();
2550
2551 /* be sure to always acquire the join list lock, to sync vs
2552 * mptcp_finish_join().
2553 */
2554 spin_lock_bh(&msk->join_list_lock);
2555 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2556 spin_unlock_bh(&msk->join_list_lock);
2557 list_splice_init(&msk->conn_list, &conn_list);
2558
2559 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2560 sk_stop_timer(sk, &sk->sk_timer);
2561 msk->pm.status = 0;
2562
2563 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2564 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2565 __mptcp_close_ssk(sk, ssk, subflow);
2566 }
2567
2568 sk->sk_prot->destroy(sk);
2569
2570 WARN_ON_ONCE(msk->wmem_reserved);
2571 WARN_ON_ONCE(msk->rmem_released);
2572 sk_stream_kill_queues(sk);
2573 xfrm_sk_free_policy(sk);
2574 sk_refcnt_debug_release(sk);
2575 mptcp_dispose_initial_subflow(msk);
2576 sock_put(sk);
2577 }
2578
2579 static void mptcp_close(struct sock *sk, long timeout)
2580 {
2581 struct mptcp_subflow_context *subflow;
2582 bool do_cancel_work = false;
2583
2584 lock_sock(sk);
2585 sk->sk_shutdown = SHUTDOWN_MASK;
2586
2587 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2588 inet_sk_state_store(sk, TCP_CLOSE);
2589 goto cleanup;
2590 }
2591
2592 if (mptcp_close_state(sk))
2593 __mptcp_wr_shutdown(sk);
2594
2595 sk_stream_wait_close(sk, timeout);
2596
2597 cleanup:
2598 /* orphan all the subflows */
2599 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2600 list_for_each_entry(subflow, &mptcp_sk(sk)->conn_list, node) {
2601 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2602 bool slow = lock_sock_fast(ssk);
2603
2604 sock_orphan(ssk);
2605 unlock_sock_fast(ssk, slow);
2606 }
2607 sock_orphan(sk);
2608
2609 sock_hold(sk);
2610 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2611 if (sk->sk_state == TCP_CLOSE) {
2612 __mptcp_destroy_sock(sk);
2613 do_cancel_work = true;
2614 } else {
2615 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2616 }
2617 release_sock(sk);
2618 if (do_cancel_work)
2619 mptcp_cancel_work(sk);
2620 sock_put(sk);
2621 }
2622
2623 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2624 {
2625 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2626 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2627 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2628
2629 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2630 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2631
2632 if (msk6 && ssk6) {
2633 msk6->saddr = ssk6->saddr;
2634 msk6->flow_label = ssk6->flow_label;
2635 }
2636 #endif
2637
2638 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2639 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2640 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2641 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2642 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2643 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2644 }
2645
2646 static int mptcp_disconnect(struct sock *sk, int flags)
2647 {
2648 struct mptcp_subflow_context *subflow;
2649 struct mptcp_sock *msk = mptcp_sk(sk);
2650
2651 __mptcp_flush_join_list(msk);
2652 mptcp_for_each_subflow(msk, subflow) {
2653 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2654
2655 lock_sock(ssk);
2656 tcp_disconnect(ssk, flags);
2657 release_sock(ssk);
2658 }
2659 return 0;
2660 }
2661
2662 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2663 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2664 {
2665 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2666
2667 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2668 }
2669 #endif
2670
2671 struct sock *mptcp_sk_clone(const struct sock *sk,
2672 const struct mptcp_options_received *mp_opt,
2673 struct request_sock *req)
2674 {
2675 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2676 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2677 struct mptcp_sock *msk;
2678 u64 ack_seq;
2679
2680 if (!nsk)
2681 return NULL;
2682
2683 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2684 if (nsk->sk_family == AF_INET6)
2685 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2686 #endif
2687
2688 __mptcp_init_sock(nsk);
2689
2690 msk = mptcp_sk(nsk);
2691 msk->local_key = subflow_req->local_key;
2692 msk->token = subflow_req->token;
2693 msk->subflow = NULL;
2694 WRITE_ONCE(msk->fully_established, false);
2695
2696 msk->write_seq = subflow_req->idsn + 1;
2697 msk->snd_nxt = msk->write_seq;
2698 msk->snd_una = msk->write_seq;
2699 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2700
2701 if (mp_opt->mp_capable) {
2702 msk->can_ack = true;
2703 msk->remote_key = mp_opt->sndr_key;
2704 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2705 ack_seq++;
2706 WRITE_ONCE(msk->ack_seq, ack_seq);
2707 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2708 }
2709
2710 sock_reset_flag(nsk, SOCK_RCU_FREE);
2711 /* will be fully established after successful MPC subflow creation */
2712 inet_sk_state_store(nsk, TCP_SYN_RECV);
2713
2714 security_inet_csk_clone(nsk, req);
2715 bh_unlock_sock(nsk);
2716
2717 /* keep a single reference */
2718 __sock_put(nsk);
2719 return nsk;
2720 }
2721
2722 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2723 {
2724 const struct tcp_sock *tp = tcp_sk(ssk);
2725
2726 msk->rcvq_space.copied = 0;
2727 msk->rcvq_space.rtt_us = 0;
2728
2729 msk->rcvq_space.time = tp->tcp_mstamp;
2730
2731 /* initial rcv_space offering made to peer */
2732 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2733 TCP_INIT_CWND * tp->advmss);
2734 if (msk->rcvq_space.space == 0)
2735 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2736
2737 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2738 }
2739
2740 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2741 bool kern)
2742 {
2743 struct mptcp_sock *msk = mptcp_sk(sk);
2744 struct socket *listener;
2745 struct sock *newsk;
2746
2747 listener = __mptcp_nmpc_socket(msk);
2748 if (WARN_ON_ONCE(!listener)) {
2749 *err = -EINVAL;
2750 return NULL;
2751 }
2752
2753 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2754 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2755 if (!newsk)
2756 return NULL;
2757
2758 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2759 if (sk_is_mptcp(newsk)) {
2760 struct mptcp_subflow_context *subflow;
2761 struct sock *new_mptcp_sock;
2762
2763 subflow = mptcp_subflow_ctx(newsk);
2764 new_mptcp_sock = subflow->conn;
2765
2766 /* is_mptcp should be false if subflow->conn is missing, see
2767 * subflow_syn_recv_sock()
2768 */
2769 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2770 tcp_sk(newsk)->is_mptcp = 0;
2771 return newsk;
2772 }
2773
2774 /* acquire the 2nd reference for the owning socket */
2775 sock_hold(new_mptcp_sock);
2776 newsk = new_mptcp_sock;
2777 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2778 } else {
2779 MPTCP_INC_STATS(sock_net(sk),
2780 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2781 }
2782
2783 return newsk;
2784 }
2785
2786 void mptcp_destroy_common(struct mptcp_sock *msk)
2787 {
2788 struct sock *sk = (struct sock *)msk;
2789
2790 __mptcp_clear_xmit(sk);
2791
2792 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2793 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2794
2795 skb_rbtree_purge(&msk->out_of_order_queue);
2796 mptcp_token_destroy(msk);
2797 mptcp_pm_free_anno_list(msk);
2798 }
2799
2800 static void mptcp_destroy(struct sock *sk)
2801 {
2802 struct mptcp_sock *msk = mptcp_sk(sk);
2803
2804 mptcp_destroy_common(msk);
2805 sk_sockets_allocated_dec(sk);
2806 }
2807
2808 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
2809 sockptr_t optval, unsigned int optlen)
2810 {
2811 struct sock *sk = (struct sock *)msk;
2812 struct socket *ssock;
2813 int ret;
2814
2815 switch (optname) {
2816 case SO_REUSEPORT:
2817 case SO_REUSEADDR:
2818 lock_sock(sk);
2819 ssock = __mptcp_nmpc_socket(msk);
2820 if (!ssock) {
2821 release_sock(sk);
2822 return -EINVAL;
2823 }
2824
2825 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
2826 if (ret == 0) {
2827 if (optname == SO_REUSEPORT)
2828 sk->sk_reuseport = ssock->sk->sk_reuseport;
2829 else if (optname == SO_REUSEADDR)
2830 sk->sk_reuse = ssock->sk->sk_reuse;
2831 }
2832 release_sock(sk);
2833 return ret;
2834 }
2835
2836 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
2837 }
2838
2839 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
2840 sockptr_t optval, unsigned int optlen)
2841 {
2842 struct sock *sk = (struct sock *)msk;
2843 int ret = -EOPNOTSUPP;
2844 struct socket *ssock;
2845
2846 switch (optname) {
2847 case IPV6_V6ONLY:
2848 lock_sock(sk);
2849 ssock = __mptcp_nmpc_socket(msk);
2850 if (!ssock) {
2851 release_sock(sk);
2852 return -EINVAL;
2853 }
2854
2855 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
2856 if (ret == 0)
2857 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
2858
2859 release_sock(sk);
2860 break;
2861 }
2862
2863 return ret;
2864 }
2865
2866 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
2867 sockptr_t optval, unsigned int optlen)
2868 {
2869 struct mptcp_sock *msk = mptcp_sk(sk);
2870 struct sock *ssk;
2871
2872 pr_debug("msk=%p", msk);
2873
2874 if (level == SOL_SOCKET)
2875 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
2876
2877 /* @@ the meaning of setsockopt() when the socket is connected and
2878 * there are multiple subflows is not yet defined. It is up to the
2879 * MPTCP-level socket to configure the subflows until the subflow
2880 * is in TCP fallback, when TCP socket options are passed through
2881 * to the one remaining subflow.
2882 */
2883 lock_sock(sk);
2884 ssk = __mptcp_tcp_fallback(msk);
2885 release_sock(sk);
2886 if (ssk)
2887 return tcp_setsockopt(ssk, level, optname, optval, optlen);
2888
2889 if (level == SOL_IPV6)
2890 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
2891
2892 return -EOPNOTSUPP;
2893 }
2894
2895 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
2896 char __user *optval, int __user *option)
2897 {
2898 struct mptcp_sock *msk = mptcp_sk(sk);
2899 struct sock *ssk;
2900
2901 pr_debug("msk=%p", msk);
2902
2903 /* @@ the meaning of setsockopt() when the socket is connected and
2904 * there are multiple subflows is not yet defined. It is up to the
2905 * MPTCP-level socket to configure the subflows until the subflow
2906 * is in TCP fallback, when socket options are passed through
2907 * to the one remaining subflow.
2908 */
2909 lock_sock(sk);
2910 ssk = __mptcp_tcp_fallback(msk);
2911 release_sock(sk);
2912 if (ssk)
2913 return tcp_getsockopt(ssk, level, optname, optval, option);
2914
2915 return -EOPNOTSUPP;
2916 }
2917
2918 void __mptcp_data_acked(struct sock *sk)
2919 {
2920 if (!sock_owned_by_user(sk))
2921 __mptcp_clean_una(sk);
2922 else
2923 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2924
2925 if (mptcp_pending_data_fin_ack(sk))
2926 mptcp_schedule_work(sk);
2927 }
2928
2929 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2930 {
2931 if (!mptcp_send_head(sk))
2932 return;
2933
2934 if (!sock_owned_by_user(sk))
2935 __mptcp_subflow_push_pending(sk, ssk);
2936 else
2937 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2938 }
2939
2940 #define MPTCP_DEFERRED_ALL (TCPF_WRITE_TIMER_DEFERRED)
2941
2942 /* processes deferred events and flush wmem */
2943 static void mptcp_release_cb(struct sock *sk)
2944 {
2945 unsigned long flags, nflags;
2946
2947 for (;;) {
2948 flags = 0;
2949 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2950 flags |= MPTCP_PUSH_PENDING;
2951 if (!flags)
2952 break;
2953
2954 /* the following actions acquire the subflow socket lock
2955 *
2956 * 1) can't be invoked in atomic scope
2957 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2958 * datapath acquires the msk socket spinlock while helding
2959 * the subflow socket lock
2960 */
2961
2962 spin_unlock_bh(&sk->sk_lock.slock);
2963 if (flags & MPTCP_PUSH_PENDING)
2964 __mptcp_push_pending(sk, 0);
2965
2966 cond_resched();
2967 spin_lock_bh(&sk->sk_lock.slock);
2968 }
2969
2970 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2971 __mptcp_clean_una(sk);
2972 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2973 __mptcp_error_report(sk);
2974
2975 /* push_pending may touch wmem_reserved, ensure we do the cleanup
2976 * later
2977 */
2978 __mptcp_update_wmem(sk);
2979 __mptcp_update_rmem(sk);
2980
2981 do {
2982 flags = sk->sk_tsq_flags;
2983 if (!(flags & MPTCP_DEFERRED_ALL))
2984 return;
2985 nflags = flags & ~MPTCP_DEFERRED_ALL;
2986 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
2987
2988 sock_release_ownership(sk);
2989
2990 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
2991 mptcp_retransmit_handler(sk);
2992 __sock_put(sk);
2993 }
2994 }
2995
2996 static int mptcp_hash(struct sock *sk)
2997 {
2998 /* should never be called,
2999 * we hash the TCP subflows not the master socket
3000 */
3001 WARN_ON_ONCE(1);
3002 return 0;
3003 }
3004
3005 static void mptcp_unhash(struct sock *sk)
3006 {
3007 /* called from sk_common_release(), but nothing to do here */
3008 }
3009
3010 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3011 {
3012 struct mptcp_sock *msk = mptcp_sk(sk);
3013 struct socket *ssock;
3014
3015 ssock = __mptcp_nmpc_socket(msk);
3016 pr_debug("msk=%p, subflow=%p", msk, ssock);
3017 if (WARN_ON_ONCE(!ssock))
3018 return -EINVAL;
3019
3020 return inet_csk_get_port(ssock->sk, snum);
3021 }
3022
3023 void mptcp_finish_connect(struct sock *ssk)
3024 {
3025 struct mptcp_subflow_context *subflow;
3026 struct mptcp_sock *msk;
3027 struct sock *sk;
3028 u64 ack_seq;
3029
3030 subflow = mptcp_subflow_ctx(ssk);
3031 sk = subflow->conn;
3032 msk = mptcp_sk(sk);
3033
3034 pr_debug("msk=%p, token=%u", sk, subflow->token);
3035
3036 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
3037 ack_seq++;
3038 subflow->map_seq = ack_seq;
3039 subflow->map_subflow_seq = 1;
3040
3041 /* the socket is not connected yet, no msk/subflow ops can access/race
3042 * accessing the field below
3043 */
3044 WRITE_ONCE(msk->remote_key, subflow->remote_key);
3045 WRITE_ONCE(msk->local_key, subflow->local_key);
3046 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3047 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3048 WRITE_ONCE(msk->ack_seq, ack_seq);
3049 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
3050 WRITE_ONCE(msk->can_ack, 1);
3051 WRITE_ONCE(msk->snd_una, msk->write_seq);
3052
3053 mptcp_pm_new_connection(msk, 0);
3054
3055 mptcp_rcv_space_init(msk, ssk);
3056 }
3057
3058 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3059 {
3060 write_lock_bh(&sk->sk_callback_lock);
3061 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3062 sk_set_socket(sk, parent);
3063 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3064 write_unlock_bh(&sk->sk_callback_lock);
3065 }
3066
3067 bool mptcp_finish_join(struct sock *ssk)
3068 {
3069 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3070 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3071 struct sock *parent = (void *)msk;
3072 struct socket *parent_sock;
3073 bool ret;
3074
3075 pr_debug("msk=%p, subflow=%p", msk, subflow);
3076
3077 /* mptcp socket already closing? */
3078 if (!mptcp_is_fully_established(parent))
3079 return false;
3080
3081 if (!msk->pm.server_side)
3082 return true;
3083
3084 if (!mptcp_pm_allow_new_subflow(msk))
3085 return false;
3086
3087 /* active connections are already on conn_list, and we can't acquire
3088 * msk lock here.
3089 * use the join list lock as synchronization point and double-check
3090 * msk status to avoid racing with __mptcp_destroy_sock()
3091 */
3092 spin_lock_bh(&msk->join_list_lock);
3093 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3094 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3095 list_add_tail(&subflow->node, &msk->join_list);
3096 sock_hold(ssk);
3097 }
3098 spin_unlock_bh(&msk->join_list_lock);
3099 if (!ret)
3100 return false;
3101
3102 /* attach to msk socket only after we are sure he will deal with us
3103 * at close time
3104 */
3105 parent_sock = READ_ONCE(parent->sk_socket);
3106 if (parent_sock && !ssk->sk_socket)
3107 mptcp_sock_graft(ssk, parent_sock);
3108 subflow->map_seq = READ_ONCE(msk->ack_seq);
3109 return true;
3110 }
3111
3112 static void mptcp_shutdown(struct sock *sk, int how)
3113 {
3114 pr_debug("sk=%p, how=%d", sk, how);
3115
3116 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3117 __mptcp_wr_shutdown(sk);
3118 }
3119
3120 static struct proto mptcp_prot = {
3121 .name = "MPTCP",
3122 .owner = THIS_MODULE,
3123 .init = mptcp_init_sock,
3124 .disconnect = mptcp_disconnect,
3125 .close = mptcp_close,
3126 .accept = mptcp_accept,
3127 .setsockopt = mptcp_setsockopt,
3128 .getsockopt = mptcp_getsockopt,
3129 .shutdown = mptcp_shutdown,
3130 .destroy = mptcp_destroy,
3131 .sendmsg = mptcp_sendmsg,
3132 .recvmsg = mptcp_recvmsg,
3133 .release_cb = mptcp_release_cb,
3134 .hash = mptcp_hash,
3135 .unhash = mptcp_unhash,
3136 .get_port = mptcp_get_port,
3137 .sockets_allocated = &mptcp_sockets_allocated,
3138 .memory_allocated = &tcp_memory_allocated,
3139 .memory_pressure = &tcp_memory_pressure,
3140 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3141 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3142 .sysctl_mem = sysctl_tcp_mem,
3143 .obj_size = sizeof(struct mptcp_sock),
3144 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3145 .no_autobind = true,
3146 };
3147
3148 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3149 {
3150 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3151 struct socket *ssock;
3152 int err;
3153
3154 lock_sock(sock->sk);
3155 ssock = __mptcp_nmpc_socket(msk);
3156 if (!ssock) {
3157 err = -EINVAL;
3158 goto unlock;
3159 }
3160
3161 err = ssock->ops->bind(ssock, uaddr, addr_len);
3162 if (!err)
3163 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3164
3165 unlock:
3166 release_sock(sock->sk);
3167 return err;
3168 }
3169
3170 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3171 struct mptcp_subflow_context *subflow)
3172 {
3173 subflow->request_mptcp = 0;
3174 __mptcp_do_fallback(msk);
3175 }
3176
3177 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3178 int addr_len, int flags)
3179 {
3180 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3181 struct mptcp_subflow_context *subflow;
3182 struct socket *ssock;
3183 int err;
3184
3185 lock_sock(sock->sk);
3186 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3187 /* pending connection or invalid state, let existing subflow
3188 * cope with that
3189 */
3190 ssock = msk->subflow;
3191 goto do_connect;
3192 }
3193
3194 ssock = __mptcp_nmpc_socket(msk);
3195 if (!ssock) {
3196 err = -EINVAL;
3197 goto unlock;
3198 }
3199
3200 mptcp_token_destroy(msk);
3201 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3202 subflow = mptcp_subflow_ctx(ssock->sk);
3203 #ifdef CONFIG_TCP_MD5SIG
3204 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3205 * TCP option space.
3206 */
3207 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3208 mptcp_subflow_early_fallback(msk, subflow);
3209 #endif
3210 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
3211 mptcp_subflow_early_fallback(msk, subflow);
3212
3213 do_connect:
3214 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3215 sock->state = ssock->state;
3216
3217 /* on successful connect, the msk state will be moved to established by
3218 * subflow_finish_connect()
3219 */
3220 if (!err || err == -EINPROGRESS)
3221 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3222 else
3223 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3224
3225 unlock:
3226 release_sock(sock->sk);
3227 return err;
3228 }
3229
3230 static int mptcp_listen(struct socket *sock, int backlog)
3231 {
3232 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3233 struct socket *ssock;
3234 int err;
3235
3236 pr_debug("msk=%p", msk);
3237
3238 lock_sock(sock->sk);
3239 ssock = __mptcp_nmpc_socket(msk);
3240 if (!ssock) {
3241 err = -EINVAL;
3242 goto unlock;
3243 }
3244
3245 mptcp_token_destroy(msk);
3246 inet_sk_state_store(sock->sk, TCP_LISTEN);
3247 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3248
3249 err = ssock->ops->listen(ssock, backlog);
3250 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3251 if (!err)
3252 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3253
3254 unlock:
3255 release_sock(sock->sk);
3256 return err;
3257 }
3258
3259 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3260 int flags, bool kern)
3261 {
3262 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3263 struct socket *ssock;
3264 int err;
3265
3266 pr_debug("msk=%p", msk);
3267
3268 lock_sock(sock->sk);
3269 if (sock->sk->sk_state != TCP_LISTEN)
3270 goto unlock_fail;
3271
3272 ssock = __mptcp_nmpc_socket(msk);
3273 if (!ssock)
3274 goto unlock_fail;
3275
3276 clear_bit(MPTCP_DATA_READY, &msk->flags);
3277 sock_hold(ssock->sk);
3278 release_sock(sock->sk);
3279
3280 err = ssock->ops->accept(sock, newsock, flags, kern);
3281 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3282 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3283 struct mptcp_subflow_context *subflow;
3284 struct sock *newsk = newsock->sk;
3285 bool slowpath;
3286
3287 slowpath = lock_sock_fast(newsk);
3288
3289 /* PM/worker can now acquire the first subflow socket
3290 * lock without racing with listener queue cleanup,
3291 * we can notify it, if needed.
3292 */
3293 subflow = mptcp_subflow_ctx(msk->first);
3294 list_add(&subflow->node, &msk->conn_list);
3295 sock_hold(msk->first);
3296 if (mptcp_is_fully_established(newsk))
3297 mptcp_pm_fully_established(msk);
3298
3299 mptcp_copy_inaddrs(newsk, msk->first);
3300 mptcp_rcv_space_init(msk, msk->first);
3301
3302 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3303 * This is needed so NOSPACE flag can be set from tcp stack.
3304 */
3305 __mptcp_flush_join_list(msk);
3306 mptcp_for_each_subflow(msk, subflow) {
3307 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3308
3309 if (!ssk->sk_socket)
3310 mptcp_sock_graft(ssk, newsock);
3311 }
3312 unlock_sock_fast(newsk, slowpath);
3313 }
3314
3315 if (inet_csk_listen_poll(ssock->sk))
3316 set_bit(MPTCP_DATA_READY, &msk->flags);
3317 sock_put(ssock->sk);
3318 return err;
3319
3320 unlock_fail:
3321 release_sock(sock->sk);
3322 return -EINVAL;
3323 }
3324
3325 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3326 {
3327 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
3328 0;
3329 }
3330
3331 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3332 {
3333 struct sock *sk = (struct sock *)msk;
3334
3335 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3336 return EPOLLOUT | EPOLLWRNORM;
3337
3338 if (sk_stream_is_writeable(sk))
3339 return EPOLLOUT | EPOLLWRNORM;
3340
3341 set_bit(MPTCP_NOSPACE, &msk->flags);
3342 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3343 if (sk_stream_is_writeable(sk))
3344 return EPOLLOUT | EPOLLWRNORM;
3345
3346 return 0;
3347 }
3348
3349 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3350 struct poll_table_struct *wait)
3351 {
3352 struct sock *sk = sock->sk;
3353 struct mptcp_sock *msk;
3354 __poll_t mask = 0;
3355 int state;
3356
3357 msk = mptcp_sk(sk);
3358 sock_poll_wait(file, sock, wait);
3359
3360 state = inet_sk_state_load(sk);
3361 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3362 if (state == TCP_LISTEN)
3363 return mptcp_check_readable(msk);
3364
3365 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3366 mask |= mptcp_check_readable(msk);
3367 mask |= mptcp_check_writeable(msk);
3368 }
3369 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3370 mask |= EPOLLHUP;
3371 if (sk->sk_shutdown & RCV_SHUTDOWN)
3372 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3373
3374 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3375 smp_rmb();
3376 if (sk->sk_err)
3377 mask |= EPOLLERR;
3378
3379 return mask;
3380 }
3381
3382 static int mptcp_release(struct socket *sock)
3383 {
3384 struct mptcp_subflow_context *subflow;
3385 struct sock *sk = sock->sk;
3386 struct mptcp_sock *msk;
3387
3388 if (!sk)
3389 return 0;
3390
3391 lock_sock(sk);
3392
3393 msk = mptcp_sk(sk);
3394
3395 mptcp_for_each_subflow(msk, subflow) {
3396 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3397
3398 ip_mc_drop_socket(ssk);
3399 }
3400
3401 release_sock(sk);
3402
3403 return inet_release(sock);
3404 }
3405
3406 static const struct proto_ops mptcp_stream_ops = {
3407 .family = PF_INET,
3408 .owner = THIS_MODULE,
3409 .release = mptcp_release,
3410 .bind = mptcp_bind,
3411 .connect = mptcp_stream_connect,
3412 .socketpair = sock_no_socketpair,
3413 .accept = mptcp_stream_accept,
3414 .getname = inet_getname,
3415 .poll = mptcp_poll,
3416 .ioctl = inet_ioctl,
3417 .gettstamp = sock_gettstamp,
3418 .listen = mptcp_listen,
3419 .shutdown = inet_shutdown,
3420 .setsockopt = sock_common_setsockopt,
3421 .getsockopt = sock_common_getsockopt,
3422 .sendmsg = inet_sendmsg,
3423 .recvmsg = inet_recvmsg,
3424 .mmap = sock_no_mmap,
3425 .sendpage = inet_sendpage,
3426 };
3427
3428 static struct inet_protosw mptcp_protosw = {
3429 .type = SOCK_STREAM,
3430 .protocol = IPPROTO_MPTCP,
3431 .prot = &mptcp_prot,
3432 .ops = &mptcp_stream_ops,
3433 .flags = INET_PROTOSW_ICSK,
3434 };
3435
3436 void __init mptcp_proto_init(void)
3437 {
3438 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3439
3440 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3441 panic("Failed to allocate MPTCP pcpu counter\n");
3442
3443 mptcp_subflow_init();
3444 mptcp_pm_init();
3445 mptcp_token_init();
3446
3447 if (proto_register(&mptcp_prot, 1) != 0)
3448 panic("Failed to register MPTCP proto.\n");
3449
3450 inet_register_protosw(&mptcp_protosw);
3451
3452 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3453 }
3454
3455 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3456 static int mptcp6_release(struct socket *sock)
3457 {
3458 struct mptcp_subflow_context *subflow;
3459 struct mptcp_sock *msk;
3460 struct sock *sk = sock->sk;
3461
3462 if (!sk)
3463 return 0;
3464
3465 lock_sock(sk);
3466
3467 msk = mptcp_sk(sk);
3468
3469 mptcp_for_each_subflow(msk, subflow) {
3470 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3471
3472 ip_mc_drop_socket(ssk);
3473 ipv6_sock_mc_close(ssk);
3474 ipv6_sock_ac_close(ssk);
3475 }
3476
3477 release_sock(sk);
3478 return inet6_release(sock);
3479 }
3480
3481 static const struct proto_ops mptcp_v6_stream_ops = {
3482 .family = PF_INET6,
3483 .owner = THIS_MODULE,
3484 .release = mptcp6_release,
3485 .bind = mptcp_bind,
3486 .connect = mptcp_stream_connect,
3487 .socketpair = sock_no_socketpair,
3488 .accept = mptcp_stream_accept,
3489 .getname = inet6_getname,
3490 .poll = mptcp_poll,
3491 .ioctl = inet6_ioctl,
3492 .gettstamp = sock_gettstamp,
3493 .listen = mptcp_listen,
3494 .shutdown = inet_shutdown,
3495 .setsockopt = sock_common_setsockopt,
3496 .getsockopt = sock_common_getsockopt,
3497 .sendmsg = inet6_sendmsg,
3498 .recvmsg = inet6_recvmsg,
3499 .mmap = sock_no_mmap,
3500 .sendpage = inet_sendpage,
3501 #ifdef CONFIG_COMPAT
3502 .compat_ioctl = inet6_compat_ioctl,
3503 #endif
3504 };
3505
3506 static struct proto mptcp_v6_prot;
3507
3508 static void mptcp_v6_destroy(struct sock *sk)
3509 {
3510 mptcp_destroy(sk);
3511 inet6_destroy_sock(sk);
3512 }
3513
3514 static struct inet_protosw mptcp_v6_protosw = {
3515 .type = SOCK_STREAM,
3516 .protocol = IPPROTO_MPTCP,
3517 .prot = &mptcp_v6_prot,
3518 .ops = &mptcp_v6_stream_ops,
3519 .flags = INET_PROTOSW_ICSK,
3520 };
3521
3522 int __init mptcp_proto_v6_init(void)
3523 {
3524 int err;
3525
3526 mptcp_v6_prot = mptcp_prot;
3527 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3528 mptcp_v6_prot.slab = NULL;
3529 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3530 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3531
3532 err = proto_register(&mptcp_v6_prot, 1);
3533 if (err)
3534 return err;
3535
3536 err = inet6_register_protosw(&mptcp_v6_protosw);
3537 if (err)
3538 proto_unregister(&mptcp_v6_prot);
3539
3540 return err;
3541 }
3542 #endif
Ubuntu Hirsute Linux kernel mirror