]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - net/mptcp/protocol.c
projects / mirror_ubuntu-jammy-kernel.git / blob
? search:


b295eb6e9580b10b13bcc4948f22dae89be147e4
[mirror_ubuntu-jammy-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 <net/sock.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
18 #include <net/tcp.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
22 #endif
23 #include <net/mptcp.h>
24 #include "protocol.h"
25 #include "mib.h"
26
27 #define MPTCP_SAME_STATE TCP_MAX_STATES
28
29 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
30 struct mptcp6_sock {
31 struct mptcp_sock msk;
32 struct ipv6_pinfo np;
33 };
34 #endif
35
36 struct mptcp_skb_cb {
37 u32 offset;
38 };
39
40 #define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
41
42 static struct percpu_counter mptcp_sockets_allocated;
43
44 /* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
45 * completed yet or has failed, return the subflow socket.
46 * Otherwise return NULL.
47 */
48 static struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
49 {
50 if (!msk->subflow || READ_ONCE(msk->can_ack))
51 return NULL;
52
53 return msk->subflow;
54 }
55
56 static bool mptcp_is_tcpsk(struct sock *sk)
57 {
58 struct socket *sock = sk->sk_socket;
59
60 if (unlikely(sk->sk_prot == &tcp_prot)) {
61 /* we are being invoked after mptcp_accept() has
62 * accepted a non-mp-capable flow: sk is a tcp_sk,
63 * not an mptcp one.
64 *
65 * Hand the socket over to tcp so all further socket ops
66 * bypass mptcp.
67 */
68 sock->ops = &inet_stream_ops;
69 return true;
70 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
71 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
72 sock->ops = &inet6_stream_ops;
73 return true;
74 #endif
75 }
76
77 return false;
78 }
79
80 static struct sock *__mptcp_tcp_fallback(struct mptcp_sock *msk)
81 {
82 sock_owned_by_me((const struct sock *)msk);
83
84 if (likely(!__mptcp_check_fallback(msk)))
85 return NULL;
86
87 return msk->first;
88 }
89
90 static int __mptcp_socket_create(struct mptcp_sock *msk)
91 {
92 struct mptcp_subflow_context *subflow;
93 struct sock *sk = (struct sock *)msk;
94 struct socket *ssock;
95 int err;
96
97 err = mptcp_subflow_create_socket(sk, &ssock);
98 if (err)
99 return err;
100
101 msk->first = ssock->sk;
102 msk->subflow = ssock;
103 subflow = mptcp_subflow_ctx(ssock->sk);
104 list_add(&subflow->node, &msk->conn_list);
105 subflow->request_mptcp = 1;
106
107 /* accept() will wait on first subflow sk_wq, and we always wakes up
108 * via msk->sk_socket
109 */
110 RCU_INIT_POINTER(msk->first->sk_wq, &sk->sk_socket->wq);
111
112 return 0;
113 }
114
115 static void __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
116 struct sk_buff *skb,
117 unsigned int offset, size_t copy_len)
118 {
119 struct sock *sk = (struct sock *)msk;
120 struct sk_buff *tail;
121
122 __skb_unlink(skb, &ssk->sk_receive_queue);
123
124 skb_ext_reset(skb);
125 skb_orphan(skb);
126 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
127
128 tail = skb_peek_tail(&sk->sk_receive_queue);
129 if (offset == 0 && tail) {
130 bool fragstolen;
131 int delta;
132
133 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
134 kfree_skb_partial(skb, fragstolen);
135 atomic_add(delta, &sk->sk_rmem_alloc);
136 sk_mem_charge(sk, delta);
137 return;
138 }
139 }
140
141 skb_set_owner_r(skb, sk);
142 __skb_queue_tail(&sk->sk_receive_queue, skb);
143 MPTCP_SKB_CB(skb)->offset = offset;
144 }
145
146 static void mptcp_stop_timer(struct sock *sk)
147 {
148 struct inet_connection_sock *icsk = inet_csk(sk);
149
150 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
151 mptcp_sk(sk)->timer_ival = 0;
152 }
153
154 /* both sockets must be locked */
155 static bool mptcp_subflow_dsn_valid(const struct mptcp_sock *msk,
156 struct sock *ssk)
157 {
158 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
159 u64 dsn = mptcp_subflow_get_mapped_dsn(subflow);
160
161 /* revalidate data sequence number.
162 *
163 * mptcp_subflow_data_available() is usually called
164 * without msk lock. Its unlikely (but possible)
165 * that msk->ack_seq has been advanced since the last
166 * call found in-sequence data.
167 */
168 if (likely(dsn == msk->ack_seq))
169 return true;
170
171 subflow->data_avail = 0;
172 return mptcp_subflow_data_available(ssk);
173 }
174
175 static void mptcp_check_data_fin_ack(struct sock *sk)
176 {
177 struct mptcp_sock *msk = mptcp_sk(sk);
178
179 if (__mptcp_check_fallback(msk))
180 return;
181
182 /* Look for an acknowledged DATA_FIN */
183 if (((1 << sk->sk_state) &
184 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
185 msk->write_seq == atomic64_read(&msk->snd_una)) {
186 mptcp_stop_timer(sk);
187
188 WRITE_ONCE(msk->snd_data_fin_enable, 0);
189
190 switch (sk->sk_state) {
191 case TCP_FIN_WAIT1:
192 inet_sk_state_store(sk, TCP_FIN_WAIT2);
193 sk->sk_state_change(sk);
194 break;
195 case TCP_CLOSING:
196 case TCP_LAST_ACK:
197 inet_sk_state_store(sk, TCP_CLOSE);
198 sk->sk_state_change(sk);
199 break;
200 }
201
202 if (sk->sk_shutdown == SHUTDOWN_MASK ||
203 sk->sk_state == TCP_CLOSE)
204 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
205 else
206 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
207 }
208 }
209
210 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
211 {
212 struct mptcp_sock *msk = mptcp_sk(sk);
213
214 if (READ_ONCE(msk->rcv_data_fin) &&
215 ((1 << sk->sk_state) &
216 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
217 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
218
219 if (msk->ack_seq == rcv_data_fin_seq) {
220 if (seq)
221 *seq = rcv_data_fin_seq;
222
223 return true;
224 }
225 }
226
227 return false;
228 }
229
230 static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
231 {
232 long tout = ssk && inet_csk(ssk)->icsk_pending ?
233 inet_csk(ssk)->icsk_timeout - jiffies : 0;
234
235 if (tout <= 0)
236 tout = mptcp_sk(sk)->timer_ival;
237 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
238 }
239
240 static void mptcp_check_data_fin(struct sock *sk)
241 {
242 struct mptcp_sock *msk = mptcp_sk(sk);
243 u64 rcv_data_fin_seq;
244
245 if (__mptcp_check_fallback(msk) || !msk->first)
246 return;
247
248 /* Need to ack a DATA_FIN received from a peer while this side
249 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
250 * msk->rcv_data_fin was set when parsing the incoming options
251 * at the subflow level and the msk lock was not held, so this
252 * is the first opportunity to act on the DATA_FIN and change
253 * the msk state.
254 *
255 * If we are caught up to the sequence number of the incoming
256 * DATA_FIN, send the DATA_ACK now and do state transition. If
257 * not caught up, do nothing and let the recv code send DATA_ACK
258 * when catching up.
259 */
260
261 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
262 struct mptcp_subflow_context *subflow;
263
264 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
265 WRITE_ONCE(msk->rcv_data_fin, 0);
266
267 sk->sk_shutdown |= RCV_SHUTDOWN;
268 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
269 set_bit(MPTCP_DATA_READY, &msk->flags);
270
271 switch (sk->sk_state) {
272 case TCP_ESTABLISHED:
273 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
274 break;
275 case TCP_FIN_WAIT1:
276 inet_sk_state_store(sk, TCP_CLOSING);
277 break;
278 case TCP_FIN_WAIT2:
279 inet_sk_state_store(sk, TCP_CLOSE);
280 // @@ Close subflows now?
281 break;
282 default:
283 /* Other states not expected */
284 WARN_ON_ONCE(1);
285 break;
286 }
287
288 mptcp_set_timeout(sk, NULL);
289 mptcp_for_each_subflow(msk, subflow) {
290 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
291
292 lock_sock(ssk);
293 tcp_send_ack(ssk);
294 release_sock(ssk);
295 }
296
297 sk->sk_state_change(sk);
298
299 if (sk->sk_shutdown == SHUTDOWN_MASK ||
300 sk->sk_state == TCP_CLOSE)
301 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
302 else
303 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
304 }
305 }
306
307 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
308 struct sock *ssk,
309 unsigned int *bytes)
310 {
311 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
312 struct sock *sk = (struct sock *)msk;
313 unsigned int moved = 0;
314 bool more_data_avail;
315 struct tcp_sock *tp;
316 bool done = false;
317
318 if (!mptcp_subflow_dsn_valid(msk, ssk)) {
319 *bytes = 0;
320 return false;
321 }
322
323 tp = tcp_sk(ssk);
324 do {
325 u32 map_remaining, offset;
326 u32 seq = tp->copied_seq;
327 struct sk_buff *skb;
328 bool fin;
329
330 /* try to move as much data as available */
331 map_remaining = subflow->map_data_len -
332 mptcp_subflow_get_map_offset(subflow);
333
334 skb = skb_peek(&ssk->sk_receive_queue);
335 if (!skb)
336 break;
337
338 if (__mptcp_check_fallback(msk)) {
339 /* if we are running under the workqueue, TCP could have
340 * collapsed skbs between dummy map creation and now
341 * be sure to adjust the size
342 */
343 map_remaining = skb->len;
344 subflow->map_data_len = skb->len;
345 }
346
347 offset = seq - TCP_SKB_CB(skb)->seq;
348 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
349 if (fin) {
350 done = true;
351 seq++;
352 }
353
354 if (offset < skb->len) {
355 size_t len = skb->len - offset;
356
357 if (tp->urg_data)
358 done = true;
359
360 __mptcp_move_skb(msk, ssk, skb, offset, len);
361 seq += len;
362 moved += len;
363
364 if (WARN_ON_ONCE(map_remaining < len))
365 break;
366 } else {
367 WARN_ON_ONCE(!fin);
368 sk_eat_skb(ssk, skb);
369 done = true;
370 }
371
372 WRITE_ONCE(tp->copied_seq, seq);
373 more_data_avail = mptcp_subflow_data_available(ssk);
374
375 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf)) {
376 done = true;
377 break;
378 }
379 } while (more_data_avail);
380
381 *bytes = moved;
382
383 /* If the moves have caught up with the DATA_FIN sequence number
384 * it's time to ack the DATA_FIN and change socket state, but
385 * this is not a good place to change state. Let the workqueue
386 * do it.
387 */
388 if (mptcp_pending_data_fin(sk, NULL) &&
389 schedule_work(&msk->work))
390 sock_hold(sk);
391
392 return done;
393 }
394
395 /* In most cases we will be able to lock the mptcp socket. If its already
396 * owned, we need to defer to the work queue to avoid ABBA deadlock.
397 */
398 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
399 {
400 struct sock *sk = (struct sock *)msk;
401 unsigned int moved = 0;
402
403 if (READ_ONCE(sk->sk_lock.owned))
404 return false;
405
406 if (unlikely(!spin_trylock_bh(&sk->sk_lock.slock)))
407 return false;
408
409 /* must re-check after taking the lock */
410 if (!READ_ONCE(sk->sk_lock.owned))
411 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
412
413 spin_unlock_bh(&sk->sk_lock.slock);
414
415 return moved > 0;
416 }
417
418 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
419 {
420 struct mptcp_sock *msk = mptcp_sk(sk);
421
422 set_bit(MPTCP_DATA_READY, &msk->flags);
423
424 if (atomic_read(&sk->sk_rmem_alloc) < READ_ONCE(sk->sk_rcvbuf) &&
425 move_skbs_to_msk(msk, ssk))
426 goto wake;
427
428 /* don't schedule if mptcp sk is (still) over limit */
429 if (atomic_read(&sk->sk_rmem_alloc) > READ_ONCE(sk->sk_rcvbuf))
430 goto wake;
431
432 /* mptcp socket is owned, release_cb should retry */
433 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED,
434 &sk->sk_tsq_flags)) {
435 sock_hold(sk);
436
437 /* need to try again, its possible release_cb() has already
438 * been called after the test_and_set_bit() above.
439 */
440 move_skbs_to_msk(msk, ssk);
441 }
442 wake:
443 sk->sk_data_ready(sk);
444 }
445
446 static void __mptcp_flush_join_list(struct mptcp_sock *msk)
447 {
448 if (likely(list_empty(&msk->join_list)))
449 return;
450
451 spin_lock_bh(&msk->join_list_lock);
452 list_splice_tail_init(&msk->join_list, &msk->conn_list);
453 spin_unlock_bh(&msk->join_list_lock);
454 }
455
456 static bool mptcp_timer_pending(struct sock *sk)
457 {
458 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
459 }
460
461 static void mptcp_reset_timer(struct sock *sk)
462 {
463 struct inet_connection_sock *icsk = inet_csk(sk);
464 unsigned long tout;
465
466 /* should never be called with mptcp level timer cleared */
467 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
468 if (WARN_ON_ONCE(!tout))
469 tout = TCP_RTO_MIN;
470 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
471 }
472
473 void mptcp_data_acked(struct sock *sk)
474 {
475 mptcp_reset_timer(sk);
476
477 if ((!sk_stream_is_writeable(sk) ||
478 (inet_sk_state_load(sk) != TCP_ESTABLISHED)) &&
479 schedule_work(&mptcp_sk(sk)->work))
480 sock_hold(sk);
481 }
482
483 void mptcp_subflow_eof(struct sock *sk)
484 {
485 struct mptcp_sock *msk = mptcp_sk(sk);
486
487 if (!test_and_set_bit(MPTCP_WORK_EOF, &msk->flags) &&
488 schedule_work(&msk->work))
489 sock_hold(sk);
490 }
491
492 static void mptcp_check_for_eof(struct mptcp_sock *msk)
493 {
494 struct mptcp_subflow_context *subflow;
495 struct sock *sk = (struct sock *)msk;
496 int receivers = 0;
497
498 mptcp_for_each_subflow(msk, subflow)
499 receivers += !subflow->rx_eof;
500
501 if (!receivers && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
502 /* hopefully temporary hack: propagate shutdown status
503 * to msk, when all subflows agree on it
504 */
505 sk->sk_shutdown |= RCV_SHUTDOWN;
506
507 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
508 set_bit(MPTCP_DATA_READY, &msk->flags);
509 sk->sk_data_ready(sk);
510 }
511 }
512
513 static bool mptcp_ext_cache_refill(struct mptcp_sock *msk)
514 {
515 const struct sock *sk = (const struct sock *)msk;
516
517 if (!msk->cached_ext)
518 msk->cached_ext = __skb_ext_alloc(sk->sk_allocation);
519
520 return !!msk->cached_ext;
521 }
522
523 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
524 {
525 struct mptcp_subflow_context *subflow;
526 struct sock *sk = (struct sock *)msk;
527
528 sock_owned_by_me(sk);
529
530 mptcp_for_each_subflow(msk, subflow) {
531 if (subflow->data_avail)
532 return mptcp_subflow_tcp_sock(subflow);
533 }
534
535 return NULL;
536 }
537
538 static bool mptcp_skb_can_collapse_to(u64 write_seq,
539 const struct sk_buff *skb,
540 const struct mptcp_ext *mpext)
541 {
542 if (!tcp_skb_can_collapse_to(skb))
543 return false;
544
545 /* can collapse only if MPTCP level sequence is in order */
546 return mpext && mpext->data_seq + mpext->data_len == write_seq;
547 }
548
549 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
550 const struct page_frag *pfrag,
551 const struct mptcp_data_frag *df)
552 {
553 return df && pfrag->page == df->page &&
554 df->data_seq + df->data_len == msk->write_seq;
555 }
556
557 static void dfrag_uncharge(struct sock *sk, int len)
558 {
559 sk_mem_uncharge(sk, len);
560 sk_wmem_queued_add(sk, -len);
561 }
562
563 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
564 {
565 int len = dfrag->data_len + dfrag->overhead;
566
567 list_del(&dfrag->list);
568 dfrag_uncharge(sk, len);
569 put_page(dfrag->page);
570 }
571
572 static void mptcp_clean_una(struct sock *sk)
573 {
574 struct mptcp_sock *msk = mptcp_sk(sk);
575 struct mptcp_data_frag *dtmp, *dfrag;
576 bool cleaned = false;
577 u64 snd_una;
578
579 /* on fallback we just need to ignore snd_una, as this is really
580 * plain TCP
581 */
582 if (__mptcp_check_fallback(msk))
583 atomic64_set(&msk->snd_una, msk->write_seq);
584 snd_una = atomic64_read(&msk->snd_una);
585
586 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
587 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
588 break;
589
590 dfrag_clear(sk, dfrag);
591 cleaned = true;
592 }
593
594 dfrag = mptcp_rtx_head(sk);
595 if (dfrag && after64(snd_una, dfrag->data_seq)) {
596 u64 delta = snd_una - dfrag->data_seq;
597
598 if (WARN_ON_ONCE(delta > dfrag->data_len))
599 goto out;
600
601 dfrag->data_seq += delta;
602 dfrag->offset += delta;
603 dfrag->data_len -= delta;
604
605 dfrag_uncharge(sk, delta);
606 cleaned = true;
607 }
608
609 out:
610 if (cleaned) {
611 sk_mem_reclaim_partial(sk);
612
613 /* Only wake up writers if a subflow is ready */
614 if (test_bit(MPTCP_SEND_SPACE, &msk->flags))
615 sk_stream_write_space(sk);
616 }
617 }
618
619 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
620 * data
621 */
622 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
623 {
624 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
625 pfrag, sk->sk_allocation)))
626 return true;
627
628 sk->sk_prot->enter_memory_pressure(sk);
629 sk_stream_moderate_sndbuf(sk);
630 return false;
631 }
632
633 static struct mptcp_data_frag *
634 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
635 int orig_offset)
636 {
637 int offset = ALIGN(orig_offset, sizeof(long));
638 struct mptcp_data_frag *dfrag;
639
640 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
641 dfrag->data_len = 0;
642 dfrag->data_seq = msk->write_seq;
643 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
644 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
645 dfrag->page = pfrag->page;
646
647 return dfrag;
648 }
649
650 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
651 struct msghdr *msg, struct mptcp_data_frag *dfrag,
652 long *timeo, int *pmss_now,
653 int *ps_goal)
654 {
655 int mss_now, avail_size, size_goal, offset, ret, frag_truesize = 0;
656 bool dfrag_collapsed, can_collapse = false;
657 struct mptcp_sock *msk = mptcp_sk(sk);
658 struct mptcp_ext *mpext = NULL;
659 bool retransmission = !!dfrag;
660 struct sk_buff *skb, *tail;
661 struct page_frag *pfrag;
662 struct page *page;
663 u64 *write_seq;
664 size_t psize;
665
666 /* use the mptcp page cache so that we can easily move the data
667 * from one substream to another, but do per subflow memory accounting
668 * Note: pfrag is used only !retransmission, but the compiler if
669 * fooled into a warning if we don't init here
670 */
671 pfrag = sk_page_frag(sk);
672 if (!retransmission) {
673 write_seq = &msk->write_seq;
674 page = pfrag->page;
675 } else {
676 write_seq = &dfrag->data_seq;
677 page = dfrag->page;
678 }
679
680 /* compute copy limit */
681 mss_now = tcp_send_mss(ssk, &size_goal, msg->msg_flags);
682 *pmss_now = mss_now;
683 *ps_goal = size_goal;
684 avail_size = size_goal;
685 skb = tcp_write_queue_tail(ssk);
686 if (skb) {
687 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
688
689 /* Limit the write to the size available in the
690 * current skb, if any, so that we create at most a new skb.
691 * Explicitly tells TCP internals to avoid collapsing on later
692 * queue management operation, to avoid breaking the ext <->
693 * SSN association set here
694 */
695 can_collapse = (size_goal - skb->len > 0) &&
696 mptcp_skb_can_collapse_to(*write_seq, skb, mpext);
697 if (!can_collapse)
698 TCP_SKB_CB(skb)->eor = 1;
699 else
700 avail_size = size_goal - skb->len;
701 }
702
703 if (!retransmission) {
704 /* reuse tail pfrag, if possible, or carve a new one from the
705 * page allocator
706 */
707 dfrag = mptcp_rtx_tail(sk);
708 offset = pfrag->offset;
709 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
710 if (!dfrag_collapsed) {
711 dfrag = mptcp_carve_data_frag(msk, pfrag, offset);
712 offset = dfrag->offset;
713 frag_truesize = dfrag->overhead;
714 }
715 psize = min_t(size_t, pfrag->size - offset, avail_size);
716
717 /* Copy to page */
718 pr_debug("left=%zu", msg_data_left(msg));
719 psize = copy_page_from_iter(pfrag->page, offset,
720 min_t(size_t, msg_data_left(msg),
721 psize),
722 &msg->msg_iter);
723 pr_debug("left=%zu", msg_data_left(msg));
724 if (!psize)
725 return -EINVAL;
726
727 if (!sk_wmem_schedule(sk, psize + dfrag->overhead)) {
728 iov_iter_revert(&msg->msg_iter, psize);
729 return -ENOMEM;
730 }
731 } else {
732 offset = dfrag->offset;
733 psize = min_t(size_t, dfrag->data_len, avail_size);
734 }
735
736 /* tell the TCP stack to delay the push so that we can safely
737 * access the skb after the sendpages call
738 */
739 ret = do_tcp_sendpages(ssk, page, offset, psize,
740 msg->msg_flags | MSG_SENDPAGE_NOTLAST | MSG_DONTWAIT);
741 if (ret <= 0) {
742 if (!retransmission)
743 iov_iter_revert(&msg->msg_iter, psize);
744 return ret;
745 }
746
747 frag_truesize += ret;
748 if (!retransmission) {
749 if (unlikely(ret < psize))
750 iov_iter_revert(&msg->msg_iter, psize - ret);
751
752 /* send successful, keep track of sent data for mptcp-level
753 * retransmission
754 */
755 dfrag->data_len += ret;
756 if (!dfrag_collapsed) {
757 get_page(dfrag->page);
758 list_add_tail(&dfrag->list, &msk->rtx_queue);
759 sk_wmem_queued_add(sk, frag_truesize);
760 } else {
761 sk_wmem_queued_add(sk, ret);
762 }
763
764 /* charge data on mptcp rtx queue to the master socket
765 * Note: we charge such data both to sk and ssk
766 */
767 sk->sk_forward_alloc -= frag_truesize;
768 }
769
770 /* if the tail skb extension is still the cached one, collapsing
771 * really happened. Note: we can't check for 'same skb' as the sk_buff
772 * hdr on tail can be transmitted, freed and re-allocated by the
773 * do_tcp_sendpages() call
774 */
775 tail = tcp_write_queue_tail(ssk);
776 if (mpext && tail && mpext == skb_ext_find(tail, SKB_EXT_MPTCP)) {
777 WARN_ON_ONCE(!can_collapse);
778 mpext->data_len += ret;
779 goto out;
780 }
781
782 skb = tcp_write_queue_tail(ssk);
783 mpext = __skb_ext_set(skb, SKB_EXT_MPTCP, msk->cached_ext);
784 msk->cached_ext = NULL;
785
786 memset(mpext, 0, sizeof(*mpext));
787 mpext->data_seq = *write_seq;
788 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
789 mpext->data_len = ret;
790 mpext->use_map = 1;
791 mpext->dsn64 = 1;
792
793 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
794 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
795 mpext->dsn64);
796
797 out:
798 if (!retransmission)
799 pfrag->offset += frag_truesize;
800 WRITE_ONCE(*write_seq, *write_seq + ret);
801 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
802
803 return ret;
804 }
805
806 static void mptcp_nospace(struct mptcp_sock *msk, struct socket *sock)
807 {
808 clear_bit(MPTCP_SEND_SPACE, &msk->flags);
809 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
810
811 /* enables sk->write_space() callbacks */
812 set_bit(SOCK_NOSPACE, &sock->flags);
813 }
814
815 static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
816 {
817 struct mptcp_subflow_context *subflow;
818 struct sock *backup = NULL;
819
820 sock_owned_by_me((const struct sock *)msk);
821
822 if (!mptcp_ext_cache_refill(msk))
823 return NULL;
824
825 mptcp_for_each_subflow(msk, subflow) {
826 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
827
828 if (!sk_stream_memory_free(ssk)) {
829 struct socket *sock = ssk->sk_socket;
830
831 if (sock)
832 mptcp_nospace(msk, sock);
833
834 return NULL;
835 }
836
837 if (subflow->backup) {
838 if (!backup)
839 backup = ssk;
840
841 continue;
842 }
843
844 return ssk;
845 }
846
847 return backup;
848 }
849
850 static void ssk_check_wmem(struct mptcp_sock *msk, struct sock *ssk)
851 {
852 struct socket *sock;
853
854 if (likely(sk_stream_is_writeable(ssk)))
855 return;
856
857 sock = READ_ONCE(ssk->sk_socket);
858 if (sock)
859 mptcp_nospace(msk, sock);
860 }
861
862 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
863 {
864 int mss_now = 0, size_goal = 0, ret = 0;
865 struct mptcp_sock *msk = mptcp_sk(sk);
866 struct page_frag *pfrag;
867 size_t copied = 0;
868 struct sock *ssk;
869 bool tx_ok;
870 long timeo;
871
872 if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL))
873 return -EOPNOTSUPP;
874
875 lock_sock(sk);
876
877 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
878
879 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
880 ret = sk_stream_wait_connect(sk, &timeo);
881 if (ret)
882 goto out;
883 }
884
885 pfrag = sk_page_frag(sk);
886 restart:
887 mptcp_clean_una(sk);
888
889 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
890 ret = -EPIPE;
891 goto out;
892 }
893
894 __mptcp_flush_join_list(msk);
895 ssk = mptcp_subflow_get_send(msk);
896 while (!sk_stream_memory_free(sk) ||
897 !ssk ||
898 !mptcp_page_frag_refill(ssk, pfrag)) {
899 if (ssk) {
900 /* make sure retransmit timer is
901 * running before we wait for memory.
902 *
903 * The retransmit timer might be needed
904 * to make the peer send an up-to-date
905 * MPTCP Ack.
906 */
907 mptcp_set_timeout(sk, ssk);
908 if (!mptcp_timer_pending(sk))
909 mptcp_reset_timer(sk);
910 }
911
912 ret = sk_stream_wait_memory(sk, &timeo);
913 if (ret)
914 goto out;
915
916 mptcp_clean_una(sk);
917
918 ssk = mptcp_subflow_get_send(msk);
919 if (list_empty(&msk->conn_list)) {
920 ret = -ENOTCONN;
921 goto out;
922 }
923 }
924
925 pr_debug("conn_list->subflow=%p", ssk);
926
927 lock_sock(ssk);
928 tx_ok = msg_data_left(msg);
929 while (tx_ok) {
930 ret = mptcp_sendmsg_frag(sk, ssk, msg, NULL, &timeo, &mss_now,
931 &size_goal);
932 if (ret < 0) {
933 if (ret == -EAGAIN && timeo > 0) {
934 mptcp_set_timeout(sk, ssk);
935 release_sock(ssk);
936 goto restart;
937 }
938 break;
939 }
940
941 copied += ret;
942
943 tx_ok = msg_data_left(msg);
944 if (!tx_ok)
945 break;
946
947 if (!sk_stream_memory_free(ssk) ||
948 !mptcp_page_frag_refill(ssk, pfrag) ||
949 !mptcp_ext_cache_refill(msk)) {
950 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
951 tcp_push(ssk, msg->msg_flags, mss_now,
952 tcp_sk(ssk)->nonagle, size_goal);
953 mptcp_set_timeout(sk, ssk);
954 release_sock(ssk);
955 goto restart;
956 }
957
958 /* memory is charged to mptcp level socket as well, i.e.
959 * if msg is very large, mptcp socket may run out of buffer
960 * space. mptcp_clean_una() will release data that has
961 * been acked at mptcp level in the mean time, so there is
962 * a good chance we can continue sending data right away.
963 *
964 * Normally, when the tcp subflow can accept more data, then
965 * so can the MPTCP socket. However, we need to cope with
966 * peers that might lag behind in their MPTCP-level
967 * acknowledgements, i.e. data might have been acked at
968 * tcp level only. So, we must also check the MPTCP socket
969 * limits before we send more data.
970 */
971 if (unlikely(!sk_stream_memory_free(sk))) {
972 tcp_push(ssk, msg->msg_flags, mss_now,
973 tcp_sk(ssk)->nonagle, size_goal);
974 mptcp_clean_una(sk);
975 if (!sk_stream_memory_free(sk)) {
976 /* can't send more for now, need to wait for
977 * MPTCP-level ACKs from peer.
978 *
979 * Wakeup will happen via mptcp_clean_una().
980 */
981 mptcp_set_timeout(sk, ssk);
982 release_sock(ssk);
983 goto restart;
984 }
985 }
986 }
987
988 mptcp_set_timeout(sk, ssk);
989 if (copied) {
990 tcp_push(ssk, msg->msg_flags, mss_now, tcp_sk(ssk)->nonagle,
991 size_goal);
992
993 /* start the timer, if it's not pending */
994 if (!mptcp_timer_pending(sk))
995 mptcp_reset_timer(sk);
996 }
997
998 ssk_check_wmem(msk, ssk);
999 release_sock(ssk);
1000 out:
1001 release_sock(sk);
1002 return copied ? : ret;
1003 }
1004
1005 static void mptcp_wait_data(struct sock *sk, long *timeo)
1006 {
1007 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1008 struct mptcp_sock *msk = mptcp_sk(sk);
1009
1010 add_wait_queue(sk_sleep(sk), &wait);
1011 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1012
1013 sk_wait_event(sk, timeo,
1014 test_and_clear_bit(MPTCP_DATA_READY, &msk->flags), &wait);
1015
1016 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1017 remove_wait_queue(sk_sleep(sk), &wait);
1018 }
1019
1020 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1021 struct msghdr *msg,
1022 size_t len)
1023 {
1024 struct sock *sk = (struct sock *)msk;
1025 struct sk_buff *skb;
1026 int copied = 0;
1027
1028 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1029 u32 offset = MPTCP_SKB_CB(skb)->offset;
1030 u32 data_len = skb->len - offset;
1031 u32 count = min_t(size_t, len - copied, data_len);
1032 int err;
1033
1034 err = skb_copy_datagram_msg(skb, offset, msg, count);
1035 if (unlikely(err < 0)) {
1036 if (!copied)
1037 return err;
1038 break;
1039 }
1040
1041 copied += count;
1042
1043 if (count < data_len) {
1044 MPTCP_SKB_CB(skb)->offset += count;
1045 break;
1046 }
1047
1048 __skb_unlink(skb, &sk->sk_receive_queue);
1049 __kfree_skb(skb);
1050
1051 if (copied >= len)
1052 break;
1053 }
1054
1055 return copied;
1056 }
1057
1058 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1059 *
1060 * Only difference: Use highest rtt estimate of the subflows in use.
1061 */
1062 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1063 {
1064 struct mptcp_subflow_context *subflow;
1065 struct sock *sk = (struct sock *)msk;
1066 u32 time, advmss = 1;
1067 u64 rtt_us, mstamp;
1068
1069 sock_owned_by_me(sk);
1070
1071 if (copied <= 0)
1072 return;
1073
1074 msk->rcvq_space.copied += copied;
1075
1076 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1077 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1078
1079 rtt_us = msk->rcvq_space.rtt_us;
1080 if (rtt_us && time < (rtt_us >> 3))
1081 return;
1082
1083 rtt_us = 0;
1084 mptcp_for_each_subflow(msk, subflow) {
1085 const struct tcp_sock *tp;
1086 u64 sf_rtt_us;
1087 u32 sf_advmss;
1088
1089 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1090
1091 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1092 sf_advmss = READ_ONCE(tp->advmss);
1093
1094 rtt_us = max(sf_rtt_us, rtt_us);
1095 advmss = max(sf_advmss, advmss);
1096 }
1097
1098 msk->rcvq_space.rtt_us = rtt_us;
1099 if (time < (rtt_us >> 3) || rtt_us == 0)
1100 return;
1101
1102 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1103 goto new_measure;
1104
1105 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1106 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1107 int rcvmem, rcvbuf;
1108 u64 rcvwin, grow;
1109
1110 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1111
1112 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1113
1114 do_div(grow, msk->rcvq_space.space);
1115 rcvwin += (grow << 1);
1116
1117 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1118 while (tcp_win_from_space(sk, rcvmem) < advmss)
1119 rcvmem += 128;
1120
1121 do_div(rcvwin, advmss);
1122 rcvbuf = min_t(u64, rcvwin * rcvmem,
1123 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1124
1125 if (rcvbuf > sk->sk_rcvbuf) {
1126 u32 window_clamp;
1127
1128 window_clamp = tcp_win_from_space(sk, rcvbuf);
1129 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1130
1131 /* Make subflows follow along. If we do not do this, we
1132 * get drops at subflow level if skbs can't be moved to
1133 * the mptcp rx queue fast enough (announced rcv_win can
1134 * exceed ssk->sk_rcvbuf).
1135 */
1136 mptcp_for_each_subflow(msk, subflow) {
1137 struct sock *ssk;
1138
1139 ssk = mptcp_subflow_tcp_sock(subflow);
1140 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1141 tcp_sk(ssk)->window_clamp = window_clamp;
1142 }
1143 }
1144 }
1145
1146 msk->rcvq_space.space = msk->rcvq_space.copied;
1147 new_measure:
1148 msk->rcvq_space.copied = 0;
1149 msk->rcvq_space.time = mstamp;
1150 }
1151
1152 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1153 {
1154 unsigned int moved = 0;
1155 bool done;
1156
1157 do {
1158 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1159
1160 if (!ssk)
1161 break;
1162
1163 lock_sock(ssk);
1164 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1165 release_sock(ssk);
1166 } while (!done);
1167
1168 return moved > 0;
1169 }
1170
1171 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1172 int nonblock, int flags, int *addr_len)
1173 {
1174 struct mptcp_sock *msk = mptcp_sk(sk);
1175 int copied = 0;
1176 int target;
1177 long timeo;
1178
1179 if (msg->msg_flags & ~(MSG_WAITALL | MSG_DONTWAIT))
1180 return -EOPNOTSUPP;
1181
1182 lock_sock(sk);
1183 timeo = sock_rcvtimeo(sk, nonblock);
1184
1185 len = min_t(size_t, len, INT_MAX);
1186 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1187 __mptcp_flush_join_list(msk);
1188
1189 while (len > (size_t)copied) {
1190 int bytes_read;
1191
1192 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied);
1193 if (unlikely(bytes_read < 0)) {
1194 if (!copied)
1195 copied = bytes_read;
1196 goto out_err;
1197 }
1198
1199 copied += bytes_read;
1200
1201 if (skb_queue_empty(&sk->sk_receive_queue) &&
1202 __mptcp_move_skbs(msk))
1203 continue;
1204
1205 /* only the master socket status is relevant here. The exit
1206 * conditions mirror closely tcp_recvmsg()
1207 */
1208 if (copied >= target)
1209 break;
1210
1211 if (copied) {
1212 if (sk->sk_err ||
1213 sk->sk_state == TCP_CLOSE ||
1214 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1215 !timeo ||
1216 signal_pending(current))
1217 break;
1218 } else {
1219 if (sk->sk_err) {
1220 copied = sock_error(sk);
1221 break;
1222 }
1223
1224 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1225 mptcp_check_for_eof(msk);
1226
1227 if (sk->sk_shutdown & RCV_SHUTDOWN)
1228 break;
1229
1230 if (sk->sk_state == TCP_CLOSE) {
1231 copied = -ENOTCONN;
1232 break;
1233 }
1234
1235 if (!timeo) {
1236 copied = -EAGAIN;
1237 break;
1238 }
1239
1240 if (signal_pending(current)) {
1241 copied = sock_intr_errno(timeo);
1242 break;
1243 }
1244 }
1245
1246 pr_debug("block timeout %ld", timeo);
1247 mptcp_wait_data(sk, &timeo);
1248 }
1249
1250 if (skb_queue_empty(&sk->sk_receive_queue)) {
1251 /* entire backlog drained, clear DATA_READY. */
1252 clear_bit(MPTCP_DATA_READY, &msk->flags);
1253
1254 /* .. race-breaker: ssk might have gotten new data
1255 * after last __mptcp_move_skbs() returned false.
1256 */
1257 if (unlikely(__mptcp_move_skbs(msk)))
1258 set_bit(MPTCP_DATA_READY, &msk->flags);
1259 } else if (unlikely(!test_bit(MPTCP_DATA_READY, &msk->flags))) {
1260 /* data to read but mptcp_wait_data() cleared DATA_READY */
1261 set_bit(MPTCP_DATA_READY, &msk->flags);
1262 }
1263 out_err:
1264 mptcp_rcv_space_adjust(msk, copied);
1265
1266 release_sock(sk);
1267 return copied;
1268 }
1269
1270 static void mptcp_retransmit_handler(struct sock *sk)
1271 {
1272 struct mptcp_sock *msk = mptcp_sk(sk);
1273
1274 if (atomic64_read(&msk->snd_una) == READ_ONCE(msk->write_seq)) {
1275 mptcp_stop_timer(sk);
1276 } else {
1277 set_bit(MPTCP_WORK_RTX, &msk->flags);
1278 if (schedule_work(&msk->work))
1279 sock_hold(sk);
1280 }
1281 }
1282
1283 static void mptcp_retransmit_timer(struct timer_list *t)
1284 {
1285 struct inet_connection_sock *icsk = from_timer(icsk, t,
1286 icsk_retransmit_timer);
1287 struct sock *sk = &icsk->icsk_inet.sk;
1288
1289 bh_lock_sock(sk);
1290 if (!sock_owned_by_user(sk)) {
1291 mptcp_retransmit_handler(sk);
1292 } else {
1293 /* delegate our work to tcp_release_cb() */
1294 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED,
1295 &sk->sk_tsq_flags))
1296 sock_hold(sk);
1297 }
1298 bh_unlock_sock(sk);
1299 sock_put(sk);
1300 }
1301
1302 /* Find an idle subflow. Return NULL if there is unacked data at tcp
1303 * level.
1304 *
1305 * A backup subflow is returned only if that is the only kind available.
1306 */
1307 static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
1308 {
1309 struct mptcp_subflow_context *subflow;
1310 struct sock *backup = NULL;
1311
1312 sock_owned_by_me((const struct sock *)msk);
1313
1314 mptcp_for_each_subflow(msk, subflow) {
1315 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1316
1317 /* still data outstanding at TCP level? Don't retransmit. */
1318 if (!tcp_write_queue_empty(ssk))
1319 return NULL;
1320
1321 if (subflow->backup) {
1322 if (!backup)
1323 backup = ssk;
1324 continue;
1325 }
1326
1327 return ssk;
1328 }
1329
1330 return backup;
1331 }
1332
1333 /* subflow sockets can be either outgoing (connect) or incoming
1334 * (accept).
1335 *
1336 * Outgoing subflows use in-kernel sockets.
1337 * Incoming subflows do not have their own 'struct socket' allocated,
1338 * so we need to use tcp_close() after detaching them from the mptcp
1339 * parent socket.
1340 */
1341 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
1342 struct mptcp_subflow_context *subflow,
1343 long timeout)
1344 {
1345 struct socket *sock = READ_ONCE(ssk->sk_socket);
1346
1347 list_del(&subflow->node);
1348
1349 if (sock && sock != sk->sk_socket) {
1350 /* outgoing subflow */
1351 sock_release(sock);
1352 } else {
1353 /* incoming subflow */
1354 tcp_close(ssk, timeout);
1355 }
1356 }
1357
1358 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1359 {
1360 return 0;
1361 }
1362
1363 static void pm_work(struct mptcp_sock *msk)
1364 {
1365 struct mptcp_pm_data *pm = &msk->pm;
1366
1367 spin_lock_bh(&msk->pm.lock);
1368
1369 pr_debug("msk=%p status=%x", msk, pm->status);
1370 if (pm->status & BIT(MPTCP_PM_ADD_ADDR_RECEIVED)) {
1371 pm->status &= ~BIT(MPTCP_PM_ADD_ADDR_RECEIVED);
1372 mptcp_pm_nl_add_addr_received(msk);
1373 }
1374 if (pm->status & BIT(MPTCP_PM_ESTABLISHED)) {
1375 pm->status &= ~BIT(MPTCP_PM_ESTABLISHED);
1376 mptcp_pm_nl_fully_established(msk);
1377 }
1378 if (pm->status & BIT(MPTCP_PM_SUBFLOW_ESTABLISHED)) {
1379 pm->status &= ~BIT(MPTCP_PM_SUBFLOW_ESTABLISHED);
1380 mptcp_pm_nl_subflow_established(msk);
1381 }
1382
1383 spin_unlock_bh(&msk->pm.lock);
1384 }
1385
1386 static void __mptcp_close_subflow(struct mptcp_sock *msk)
1387 {
1388 struct mptcp_subflow_context *subflow, *tmp;
1389
1390 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
1391 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1392
1393 if (inet_sk_state_load(ssk) != TCP_CLOSE)
1394 continue;
1395
1396 __mptcp_close_ssk((struct sock *)msk, ssk, subflow, 0);
1397 }
1398 }
1399
1400 static void mptcp_worker(struct work_struct *work)
1401 {
1402 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1403 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1404 int orig_len, orig_offset, mss_now = 0, size_goal = 0;
1405 struct mptcp_data_frag *dfrag;
1406 u64 orig_write_seq;
1407 size_t copied = 0;
1408 struct msghdr msg = {
1409 .msg_flags = MSG_DONTWAIT,
1410 };
1411 long timeo = 0;
1412
1413 lock_sock(sk);
1414 mptcp_clean_una(sk);
1415 mptcp_check_data_fin_ack(sk);
1416 __mptcp_flush_join_list(msk);
1417 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
1418 __mptcp_close_subflow(msk);
1419
1420 __mptcp_move_skbs(msk);
1421
1422 if (msk->pm.status)
1423 pm_work(msk);
1424
1425 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1426 mptcp_check_for_eof(msk);
1427
1428 mptcp_check_data_fin(sk);
1429
1430 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1431 goto unlock;
1432
1433 dfrag = mptcp_rtx_head(sk);
1434 if (!dfrag)
1435 goto unlock;
1436
1437 if (!mptcp_ext_cache_refill(msk))
1438 goto reset_unlock;
1439
1440 ssk = mptcp_subflow_get_retrans(msk);
1441 if (!ssk)
1442 goto reset_unlock;
1443
1444 lock_sock(ssk);
1445
1446 orig_len = dfrag->data_len;
1447 orig_offset = dfrag->offset;
1448 orig_write_seq = dfrag->data_seq;
1449 while (dfrag->data_len > 0) {
1450 int ret = mptcp_sendmsg_frag(sk, ssk, &msg, dfrag, &timeo,
1451 &mss_now, &size_goal);
1452 if (ret < 0)
1453 break;
1454
1455 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1456 copied += ret;
1457 dfrag->data_len -= ret;
1458 dfrag->offset += ret;
1459
1460 if (!mptcp_ext_cache_refill(msk))
1461 break;
1462 }
1463 if (copied)
1464 tcp_push(ssk, msg.msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1465 size_goal);
1466
1467 dfrag->data_seq = orig_write_seq;
1468 dfrag->offset = orig_offset;
1469 dfrag->data_len = orig_len;
1470
1471 mptcp_set_timeout(sk, ssk);
1472 release_sock(ssk);
1473
1474 reset_unlock:
1475 if (!mptcp_timer_pending(sk))
1476 mptcp_reset_timer(sk);
1477
1478 unlock:
1479 release_sock(sk);
1480 sock_put(sk);
1481 }
1482
1483 static int __mptcp_init_sock(struct sock *sk)
1484 {
1485 struct mptcp_sock *msk = mptcp_sk(sk);
1486
1487 spin_lock_init(&msk->join_list_lock);
1488
1489 INIT_LIST_HEAD(&msk->conn_list);
1490 INIT_LIST_HEAD(&msk->join_list);
1491 INIT_LIST_HEAD(&msk->rtx_queue);
1492 __set_bit(MPTCP_SEND_SPACE, &msk->flags);
1493 INIT_WORK(&msk->work, mptcp_worker);
1494
1495 msk->first = NULL;
1496 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1497
1498 mptcp_pm_data_init(msk);
1499
1500 /* re-use the csk retrans timer for MPTCP-level retrans */
1501 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1502
1503 return 0;
1504 }
1505
1506 static int mptcp_init_sock(struct sock *sk)
1507 {
1508 struct net *net = sock_net(sk);
1509 int ret;
1510
1511 if (!mptcp_is_enabled(net))
1512 return -ENOPROTOOPT;
1513
1514 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1515 return -ENOMEM;
1516
1517 ret = __mptcp_init_sock(sk);
1518 if (ret)
1519 return ret;
1520
1521 ret = __mptcp_socket_create(mptcp_sk(sk));
1522 if (ret)
1523 return ret;
1524
1525 sk_sockets_allocated_inc(sk);
1526 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1527 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[2];
1528
1529 return 0;
1530 }
1531
1532 static void __mptcp_clear_xmit(struct sock *sk)
1533 {
1534 struct mptcp_sock *msk = mptcp_sk(sk);
1535 struct mptcp_data_frag *dtmp, *dfrag;
1536
1537 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
1538
1539 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
1540 dfrag_clear(sk, dfrag);
1541 }
1542
1543 static void mptcp_cancel_work(struct sock *sk)
1544 {
1545 struct mptcp_sock *msk = mptcp_sk(sk);
1546
1547 if (cancel_work_sync(&msk->work))
1548 sock_put(sk);
1549 }
1550
1551 static void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
1552 {
1553 lock_sock(ssk);
1554
1555 switch (ssk->sk_state) {
1556 case TCP_LISTEN:
1557 if (!(how & RCV_SHUTDOWN))
1558 break;
1559 fallthrough;
1560 case TCP_SYN_SENT:
1561 tcp_disconnect(ssk, O_NONBLOCK);
1562 break;
1563 default:
1564 if (__mptcp_check_fallback(mptcp_sk(sk))) {
1565 pr_debug("Fallback");
1566 ssk->sk_shutdown |= how;
1567 tcp_shutdown(ssk, how);
1568 } else {
1569 pr_debug("Sending DATA_FIN on subflow %p", ssk);
1570 mptcp_set_timeout(sk, ssk);
1571 tcp_send_ack(ssk);
1572 }
1573 break;
1574 }
1575
1576 release_sock(ssk);
1577 }
1578
1579 static const unsigned char new_state[16] = {
1580 /* current state: new state: action: */
1581 [0 /* (Invalid) */] = TCP_CLOSE,
1582 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1583 [TCP_SYN_SENT] = TCP_CLOSE,
1584 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1585 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1586 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1587 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
1588 [TCP_CLOSE] = TCP_CLOSE,
1589 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1590 [TCP_LAST_ACK] = TCP_LAST_ACK,
1591 [TCP_LISTEN] = TCP_CLOSE,
1592 [TCP_CLOSING] = TCP_CLOSING,
1593 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1594 };
1595
1596 static int mptcp_close_state(struct sock *sk)
1597 {
1598 int next = (int)new_state[sk->sk_state];
1599 int ns = next & TCP_STATE_MASK;
1600
1601 inet_sk_state_store(sk, ns);
1602
1603 return next & TCP_ACTION_FIN;
1604 }
1605
1606 static void mptcp_close(struct sock *sk, long timeout)
1607 {
1608 struct mptcp_subflow_context *subflow, *tmp;
1609 struct mptcp_sock *msk = mptcp_sk(sk);
1610 LIST_HEAD(conn_list);
1611
1612 lock_sock(sk);
1613 sk->sk_shutdown = SHUTDOWN_MASK;
1614
1615 if (sk->sk_state == TCP_LISTEN) {
1616 inet_sk_state_store(sk, TCP_CLOSE);
1617 goto cleanup;
1618 } else if (sk->sk_state == TCP_CLOSE) {
1619 goto cleanup;
1620 }
1621
1622 if (__mptcp_check_fallback(msk)) {
1623 goto update_state;
1624 } else if (mptcp_close_state(sk)) {
1625 pr_debug("Sending DATA_FIN sk=%p", sk);
1626 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
1627 WRITE_ONCE(msk->snd_data_fin_enable, 1);
1628
1629 mptcp_for_each_subflow(msk, subflow) {
1630 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
1631
1632 mptcp_subflow_shutdown(sk, tcp_sk, SHUTDOWN_MASK);
1633 }
1634 }
1635
1636 sk_stream_wait_close(sk, timeout);
1637
1638 update_state:
1639 inet_sk_state_store(sk, TCP_CLOSE);
1640
1641 cleanup:
1642 /* be sure to always acquire the join list lock, to sync vs
1643 * mptcp_finish_join().
1644 */
1645 spin_lock_bh(&msk->join_list_lock);
1646 list_splice_tail_init(&msk->join_list, &msk->conn_list);
1647 spin_unlock_bh(&msk->join_list_lock);
1648 list_splice_init(&msk->conn_list, &conn_list);
1649
1650 __mptcp_clear_xmit(sk);
1651
1652 release_sock(sk);
1653
1654 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
1655 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1656 __mptcp_close_ssk(sk, ssk, subflow, timeout);
1657 }
1658
1659 mptcp_cancel_work(sk);
1660
1661 __skb_queue_purge(&sk->sk_receive_queue);
1662
1663 sk_common_release(sk);
1664 }
1665
1666 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
1667 {
1668 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1669 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
1670 struct ipv6_pinfo *msk6 = inet6_sk(msk);
1671
1672 msk->sk_v6_daddr = ssk->sk_v6_daddr;
1673 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
1674
1675 if (msk6 && ssk6) {
1676 msk6->saddr = ssk6->saddr;
1677 msk6->flow_label = ssk6->flow_label;
1678 }
1679 #endif
1680
1681 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
1682 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
1683 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
1684 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
1685 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
1686 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
1687 }
1688
1689 static int mptcp_disconnect(struct sock *sk, int flags)
1690 {
1691 /* Should never be called.
1692 * inet_stream_connect() calls ->disconnect, but that
1693 * refers to the subflow socket, not the mptcp one.
1694 */
1695 WARN_ON_ONCE(1);
1696 return 0;
1697 }
1698
1699 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1700 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
1701 {
1702 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
1703
1704 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
1705 }
1706 #endif
1707
1708 struct sock *mptcp_sk_clone(const struct sock *sk,
1709 const struct mptcp_options_received *mp_opt,
1710 struct request_sock *req)
1711 {
1712 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
1713 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
1714 struct mptcp_sock *msk;
1715 u64 ack_seq;
1716
1717 if (!nsk)
1718 return NULL;
1719
1720 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1721 if (nsk->sk_family == AF_INET6)
1722 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
1723 #endif
1724
1725 __mptcp_init_sock(nsk);
1726
1727 msk = mptcp_sk(nsk);
1728 msk->local_key = subflow_req->local_key;
1729 msk->token = subflow_req->token;
1730 msk->subflow = NULL;
1731 WRITE_ONCE(msk->fully_established, false);
1732
1733 msk->write_seq = subflow_req->idsn + 1;
1734 atomic64_set(&msk->snd_una, msk->write_seq);
1735 if (mp_opt->mp_capable) {
1736 msk->can_ack = true;
1737 msk->remote_key = mp_opt->sndr_key;
1738 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
1739 ack_seq++;
1740 WRITE_ONCE(msk->ack_seq, ack_seq);
1741 }
1742
1743 sock_reset_flag(nsk, SOCK_RCU_FREE);
1744 /* will be fully established after successful MPC subflow creation */
1745 inet_sk_state_store(nsk, TCP_SYN_RECV);
1746 bh_unlock_sock(nsk);
1747
1748 /* keep a single reference */
1749 __sock_put(nsk);
1750 return nsk;
1751 }
1752
1753 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
1754 {
1755 const struct tcp_sock *tp = tcp_sk(ssk);
1756
1757 msk->rcvq_space.copied = 0;
1758 msk->rcvq_space.rtt_us = 0;
1759
1760 msk->rcvq_space.time = tp->tcp_mstamp;
1761
1762 /* initial rcv_space offering made to peer */
1763 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
1764 TCP_INIT_CWND * tp->advmss);
1765 if (msk->rcvq_space.space == 0)
1766 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
1767 }
1768
1769 static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
1770 bool kern)
1771 {
1772 struct mptcp_sock *msk = mptcp_sk(sk);
1773 struct socket *listener;
1774 struct sock *newsk;
1775
1776 listener = __mptcp_nmpc_socket(msk);
1777 if (WARN_ON_ONCE(!listener)) {
1778 *err = -EINVAL;
1779 return NULL;
1780 }
1781
1782 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
1783 newsk = inet_csk_accept(listener->sk, flags, err, kern);
1784 if (!newsk)
1785 return NULL;
1786
1787 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
1788 if (sk_is_mptcp(newsk)) {
1789 struct mptcp_subflow_context *subflow;
1790 struct sock *new_mptcp_sock;
1791 struct sock *ssk = newsk;
1792
1793 subflow = mptcp_subflow_ctx(newsk);
1794 new_mptcp_sock = subflow->conn;
1795
1796 /* is_mptcp should be false if subflow->conn is missing, see
1797 * subflow_syn_recv_sock()
1798 */
1799 if (WARN_ON_ONCE(!new_mptcp_sock)) {
1800 tcp_sk(newsk)->is_mptcp = 0;
1801 return newsk;
1802 }
1803
1804 /* acquire the 2nd reference for the owning socket */
1805 sock_hold(new_mptcp_sock);
1806
1807 local_bh_disable();
1808 bh_lock_sock(new_mptcp_sock);
1809 msk = mptcp_sk(new_mptcp_sock);
1810 msk->first = newsk;
1811
1812 newsk = new_mptcp_sock;
1813 mptcp_copy_inaddrs(newsk, ssk);
1814 list_add(&subflow->node, &msk->conn_list);
1815
1816 mptcp_rcv_space_init(msk, ssk);
1817 bh_unlock_sock(new_mptcp_sock);
1818
1819 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
1820 local_bh_enable();
1821 } else {
1822 MPTCP_INC_STATS(sock_net(sk),
1823 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
1824 }
1825
1826 return newsk;
1827 }
1828
1829 static void mptcp_destroy(struct sock *sk)
1830 {
1831 struct mptcp_sock *msk = mptcp_sk(sk);
1832
1833 mptcp_token_destroy(msk);
1834 if (msk->cached_ext)
1835 __skb_ext_put(msk->cached_ext);
1836
1837 sk_sockets_allocated_dec(sk);
1838 }
1839
1840 static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
1841 sockptr_t optval, unsigned int optlen)
1842 {
1843 struct sock *sk = (struct sock *)msk;
1844 struct socket *ssock;
1845 int ret;
1846
1847 switch (optname) {
1848 case SO_REUSEPORT:
1849 case SO_REUSEADDR:
1850 lock_sock(sk);
1851 ssock = __mptcp_nmpc_socket(msk);
1852 if (!ssock) {
1853 release_sock(sk);
1854 return -EINVAL;
1855 }
1856
1857 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
1858 if (ret == 0) {
1859 if (optname == SO_REUSEPORT)
1860 sk->sk_reuseport = ssock->sk->sk_reuseport;
1861 else if (optname == SO_REUSEADDR)
1862 sk->sk_reuse = ssock->sk->sk_reuse;
1863 }
1864 release_sock(sk);
1865 return ret;
1866 }
1867
1868 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
1869 }
1870
1871 static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
1872 sockptr_t optval, unsigned int optlen)
1873 {
1874 struct sock *sk = (struct sock *)msk;
1875 int ret = -EOPNOTSUPP;
1876 struct socket *ssock;
1877
1878 switch (optname) {
1879 case IPV6_V6ONLY:
1880 lock_sock(sk);
1881 ssock = __mptcp_nmpc_socket(msk);
1882 if (!ssock) {
1883 release_sock(sk);
1884 return -EINVAL;
1885 }
1886
1887 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
1888 if (ret == 0)
1889 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
1890
1891 release_sock(sk);
1892 break;
1893 }
1894
1895 return ret;
1896 }
1897
1898 static int mptcp_setsockopt(struct sock *sk, int level, int optname,
1899 sockptr_t optval, unsigned int optlen)
1900 {
1901 struct mptcp_sock *msk = mptcp_sk(sk);
1902 struct sock *ssk;
1903
1904 pr_debug("msk=%p", msk);
1905
1906 if (level == SOL_SOCKET)
1907 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
1908
1909 /* @@ the meaning of setsockopt() when the socket is connected and
1910 * there are multiple subflows is not yet defined. It is up to the
1911 * MPTCP-level socket to configure the subflows until the subflow
1912 * is in TCP fallback, when TCP socket options are passed through
1913 * to the one remaining subflow.
1914 */
1915 lock_sock(sk);
1916 ssk = __mptcp_tcp_fallback(msk);
1917 release_sock(sk);
1918 if (ssk)
1919 return tcp_setsockopt(ssk, level, optname, optval, optlen);
1920
1921 if (level == SOL_IPV6)
1922 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
1923
1924 return -EOPNOTSUPP;
1925 }
1926
1927 static int mptcp_getsockopt(struct sock *sk, int level, int optname,
1928 char __user *optval, int __user *option)
1929 {
1930 struct mptcp_sock *msk = mptcp_sk(sk);
1931 struct sock *ssk;
1932
1933 pr_debug("msk=%p", msk);
1934
1935 /* @@ the meaning of setsockopt() when the socket is connected and
1936 * there are multiple subflows is not yet defined. It is up to the
1937 * MPTCP-level socket to configure the subflows until the subflow
1938 * is in TCP fallback, when socket options are passed through
1939 * to the one remaining subflow.
1940 */
1941 lock_sock(sk);
1942 ssk = __mptcp_tcp_fallback(msk);
1943 release_sock(sk);
1944 if (ssk)
1945 return tcp_getsockopt(ssk, level, optname, optval, option);
1946
1947 return -EOPNOTSUPP;
1948 }
1949
1950 #define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
1951 TCPF_WRITE_TIMER_DEFERRED)
1952
1953 /* this is very alike tcp_release_cb() but we must handle differently a
1954 * different set of events
1955 */
1956 static void mptcp_release_cb(struct sock *sk)
1957 {
1958 unsigned long flags, nflags;
1959
1960 do {
1961 flags = sk->sk_tsq_flags;
1962 if (!(flags & MPTCP_DEFERRED_ALL))
1963 return;
1964 nflags = flags & ~MPTCP_DEFERRED_ALL;
1965 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1966
1967 sock_release_ownership(sk);
1968
1969 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1970 struct mptcp_sock *msk = mptcp_sk(sk);
1971 struct sock *ssk;
1972
1973 ssk = mptcp_subflow_recv_lookup(msk);
1974 if (!ssk || !schedule_work(&msk->work))
1975 __sock_put(sk);
1976 }
1977
1978 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1979 mptcp_retransmit_handler(sk);
1980 __sock_put(sk);
1981 }
1982 }
1983
1984 static int mptcp_hash(struct sock *sk)
1985 {
1986 /* should never be called,
1987 * we hash the TCP subflows not the master socket
1988 */
1989 WARN_ON_ONCE(1);
1990 return 0;
1991 }
1992
1993 static void mptcp_unhash(struct sock *sk)
1994 {
1995 /* called from sk_common_release(), but nothing to do here */
1996 }
1997
1998 static int mptcp_get_port(struct sock *sk, unsigned short snum)
1999 {
2000 struct mptcp_sock *msk = mptcp_sk(sk);
2001 struct socket *ssock;
2002
2003 ssock = __mptcp_nmpc_socket(msk);
2004 pr_debug("msk=%p, subflow=%p", msk, ssock);
2005 if (WARN_ON_ONCE(!ssock))
2006 return -EINVAL;
2007
2008 return inet_csk_get_port(ssock->sk, snum);
2009 }
2010
2011 void mptcp_finish_connect(struct sock *ssk)
2012 {
2013 struct mptcp_subflow_context *subflow;
2014 struct mptcp_sock *msk;
2015 struct sock *sk;
2016 u64 ack_seq;
2017
2018 subflow = mptcp_subflow_ctx(ssk);
2019 sk = subflow->conn;
2020 msk = mptcp_sk(sk);
2021
2022 pr_debug("msk=%p, token=%u", sk, subflow->token);
2023
2024 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2025 ack_seq++;
2026 subflow->map_seq = ack_seq;
2027 subflow->map_subflow_seq = 1;
2028
2029 /* the socket is not connected yet, no msk/subflow ops can access/race
2030 * accessing the field below
2031 */
2032 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2033 WRITE_ONCE(msk->local_key, subflow->local_key);
2034 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2035 WRITE_ONCE(msk->ack_seq, ack_seq);
2036 WRITE_ONCE(msk->can_ack, 1);
2037 atomic64_set(&msk->snd_una, msk->write_seq);
2038
2039 mptcp_pm_new_connection(msk, 0);
2040
2041 mptcp_rcv_space_init(msk, ssk);
2042 }
2043
2044 static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2045 {
2046 write_lock_bh(&sk->sk_callback_lock);
2047 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2048 sk_set_socket(sk, parent);
2049 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2050 write_unlock_bh(&sk->sk_callback_lock);
2051 }
2052
2053 bool mptcp_finish_join(struct sock *sk)
2054 {
2055 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
2056 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2057 struct sock *parent = (void *)msk;
2058 struct socket *parent_sock;
2059 bool ret;
2060
2061 pr_debug("msk=%p, subflow=%p", msk, subflow);
2062
2063 /* mptcp socket already closing? */
2064 if (!mptcp_is_fully_established(parent))
2065 return false;
2066
2067 if (!msk->pm.server_side)
2068 return true;
2069
2070 if (!mptcp_pm_allow_new_subflow(msk))
2071 return false;
2072
2073 /* active connections are already on conn_list, and we can't acquire
2074 * msk lock here.
2075 * use the join list lock as synchronization point and double-check
2076 * msk status to avoid racing with mptcp_close()
2077 */
2078 spin_lock_bh(&msk->join_list_lock);
2079 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2080 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node)))
2081 list_add_tail(&subflow->node, &msk->join_list);
2082 spin_unlock_bh(&msk->join_list_lock);
2083 if (!ret)
2084 return false;
2085
2086 /* attach to msk socket only after we are sure he will deal with us
2087 * at close time
2088 */
2089 parent_sock = READ_ONCE(parent->sk_socket);
2090 if (parent_sock && !sk->sk_socket)
2091 mptcp_sock_graft(sk, parent_sock);
2092 subflow->map_seq = READ_ONCE(msk->ack_seq);
2093 return true;
2094 }
2095
2096 static bool mptcp_memory_free(const struct sock *sk, int wake)
2097 {
2098 struct mptcp_sock *msk = mptcp_sk(sk);
2099
2100 return wake ? test_bit(MPTCP_SEND_SPACE, &msk->flags) : true;
2101 }
2102
2103 static struct proto mptcp_prot = {
2104 .name = "MPTCP",
2105 .owner = THIS_MODULE,
2106 .init = mptcp_init_sock,
2107 .disconnect = mptcp_disconnect,
2108 .close = mptcp_close,
2109 .accept = mptcp_accept,
2110 .setsockopt = mptcp_setsockopt,
2111 .getsockopt = mptcp_getsockopt,
2112 .shutdown = tcp_shutdown,
2113 .destroy = mptcp_destroy,
2114 .sendmsg = mptcp_sendmsg,
2115 .recvmsg = mptcp_recvmsg,
2116 .release_cb = mptcp_release_cb,
2117 .hash = mptcp_hash,
2118 .unhash = mptcp_unhash,
2119 .get_port = mptcp_get_port,
2120 .sockets_allocated = &mptcp_sockets_allocated,
2121 .memory_allocated = &tcp_memory_allocated,
2122 .memory_pressure = &tcp_memory_pressure,
2123 .stream_memory_free = mptcp_memory_free,
2124 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2125 .sysctl_mem = sysctl_tcp_mem,
2126 .obj_size = sizeof(struct mptcp_sock),
2127 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2128 .no_autobind = true,
2129 };
2130
2131 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2132 {
2133 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2134 struct socket *ssock;
2135 int err;
2136
2137 lock_sock(sock->sk);
2138 ssock = __mptcp_nmpc_socket(msk);
2139 if (!ssock) {
2140 err = -EINVAL;
2141 goto unlock;
2142 }
2143
2144 err = ssock->ops->bind(ssock, uaddr, addr_len);
2145 if (!err)
2146 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2147
2148 unlock:
2149 release_sock(sock->sk);
2150 return err;
2151 }
2152
2153 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2154 struct mptcp_subflow_context *subflow)
2155 {
2156 subflow->request_mptcp = 0;
2157 __mptcp_do_fallback(msk);
2158 }
2159
2160 static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2161 int addr_len, int flags)
2162 {
2163 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2164 struct mptcp_subflow_context *subflow;
2165 struct socket *ssock;
2166 int err;
2167
2168 lock_sock(sock->sk);
2169 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2170 /* pending connection or invalid state, let existing subflow
2171 * cope with that
2172 */
2173 ssock = msk->subflow;
2174 goto do_connect;
2175 }
2176
2177 ssock = __mptcp_nmpc_socket(msk);
2178 if (!ssock) {
2179 err = -EINVAL;
2180 goto unlock;
2181 }
2182
2183 mptcp_token_destroy(msk);
2184 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2185 subflow = mptcp_subflow_ctx(ssock->sk);
2186 #ifdef CONFIG_TCP_MD5SIG
2187 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2188 * TCP option space.
2189 */
2190 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2191 mptcp_subflow_early_fallback(msk, subflow);
2192 #endif
2193 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2194 mptcp_subflow_early_fallback(msk, subflow);
2195
2196 do_connect:
2197 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2198 sock->state = ssock->state;
2199
2200 /* on successful connect, the msk state will be moved to established by
2201 * subflow_finish_connect()
2202 */
2203 if (!err || err == -EINPROGRESS)
2204 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2205 else
2206 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2207
2208 unlock:
2209 release_sock(sock->sk);
2210 return err;
2211 }
2212
2213 static int mptcp_listen(struct socket *sock, int backlog)
2214 {
2215 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2216 struct socket *ssock;
2217 int err;
2218
2219 pr_debug("msk=%p", msk);
2220
2221 lock_sock(sock->sk);
2222 ssock = __mptcp_nmpc_socket(msk);
2223 if (!ssock) {
2224 err = -EINVAL;
2225 goto unlock;
2226 }
2227
2228 mptcp_token_destroy(msk);
2229 inet_sk_state_store(sock->sk, TCP_LISTEN);
2230 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2231
2232 err = ssock->ops->listen(ssock, backlog);
2233 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2234 if (!err)
2235 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2236
2237 unlock:
2238 release_sock(sock->sk);
2239 return err;
2240 }
2241
2242 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2243 int flags, bool kern)
2244 {
2245 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2246 struct socket *ssock;
2247 int err;
2248
2249 pr_debug("msk=%p", msk);
2250
2251 lock_sock(sock->sk);
2252 if (sock->sk->sk_state != TCP_LISTEN)
2253 goto unlock_fail;
2254
2255 ssock = __mptcp_nmpc_socket(msk);
2256 if (!ssock)
2257 goto unlock_fail;
2258
2259 clear_bit(MPTCP_DATA_READY, &msk->flags);
2260 sock_hold(ssock->sk);
2261 release_sock(sock->sk);
2262
2263 err = ssock->ops->accept(sock, newsock, flags, kern);
2264 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2265 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2266 struct mptcp_subflow_context *subflow;
2267
2268 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2269 * This is needed so NOSPACE flag can be set from tcp stack.
2270 */
2271 __mptcp_flush_join_list(msk);
2272 mptcp_for_each_subflow(msk, subflow) {
2273 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2274
2275 if (!ssk->sk_socket)
2276 mptcp_sock_graft(ssk, newsock);
2277 }
2278 }
2279
2280 if (inet_csk_listen_poll(ssock->sk))
2281 set_bit(MPTCP_DATA_READY, &msk->flags);
2282 sock_put(ssock->sk);
2283 return err;
2284
2285 unlock_fail:
2286 release_sock(sock->sk);
2287 return -EINVAL;
2288 }
2289
2290 static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2291 {
2292 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2293 0;
2294 }
2295
2296 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2297 struct poll_table_struct *wait)
2298 {
2299 struct sock *sk = sock->sk;
2300 struct mptcp_sock *msk;
2301 __poll_t mask = 0;
2302 int state;
2303
2304 msk = mptcp_sk(sk);
2305 sock_poll_wait(file, sock, wait);
2306
2307 state = inet_sk_state_load(sk);
2308 if (state == TCP_LISTEN)
2309 return mptcp_check_readable(msk);
2310
2311 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2312 mask |= mptcp_check_readable(msk);
2313 if (sk_stream_is_writeable(sk) &&
2314 test_bit(MPTCP_SEND_SPACE, &msk->flags))
2315 mask |= EPOLLOUT | EPOLLWRNORM;
2316 }
2317 if (sk->sk_shutdown & RCV_SHUTDOWN)
2318 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2319
2320 return mask;
2321 }
2322
2323 static int mptcp_shutdown(struct socket *sock, int how)
2324 {
2325 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2326 struct mptcp_subflow_context *subflow;
2327 int ret = 0;
2328
2329 pr_debug("sk=%p, how=%d", msk, how);
2330
2331 lock_sock(sock->sk);
2332
2333 how++;
2334 if ((how & ~SHUTDOWN_MASK) || !how) {
2335 ret = -EINVAL;
2336 goto out_unlock;
2337 }
2338
2339 if (sock->state == SS_CONNECTING) {
2340 if ((1 << sock->sk->sk_state) &
2341 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2342 sock->state = SS_DISCONNECTING;
2343 else
2344 sock->state = SS_CONNECTED;
2345 }
2346
2347 /* If we've already sent a FIN, or it's a closed state, skip this. */
2348 if (__mptcp_check_fallback(msk)) {
2349 if (how == SHUT_WR || how == SHUT_RDWR)
2350 inet_sk_state_store(sock->sk, TCP_FIN_WAIT1);
2351
2352 mptcp_for_each_subflow(msk, subflow) {
2353 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2354
2355 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2356 }
2357 } else if ((how & SEND_SHUTDOWN) &&
2358 ((1 << sock->sk->sk_state) &
2359 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2360 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) &&
2361 mptcp_close_state(sock->sk)) {
2362 __mptcp_flush_join_list(msk);
2363
2364 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2365 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2366
2367 mptcp_for_each_subflow(msk, subflow) {
2368 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2369
2370 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2371 }
2372 }
2373
2374 /* Wake up anyone sleeping in poll. */
2375 sock->sk->sk_state_change(sock->sk);
2376
2377 out_unlock:
2378 release_sock(sock->sk);
2379
2380 return ret;
2381 }
2382
2383 static const struct proto_ops mptcp_stream_ops = {
2384 .family = PF_INET,
2385 .owner = THIS_MODULE,
2386 .release = inet_release,
2387 .bind = mptcp_bind,
2388 .connect = mptcp_stream_connect,
2389 .socketpair = sock_no_socketpair,
2390 .accept = mptcp_stream_accept,
2391 .getname = inet_getname,
2392 .poll = mptcp_poll,
2393 .ioctl = inet_ioctl,
2394 .gettstamp = sock_gettstamp,
2395 .listen = mptcp_listen,
2396 .shutdown = mptcp_shutdown,
2397 .setsockopt = sock_common_setsockopt,
2398 .getsockopt = sock_common_getsockopt,
2399 .sendmsg = inet_sendmsg,
2400 .recvmsg = inet_recvmsg,
2401 .mmap = sock_no_mmap,
2402 .sendpage = inet_sendpage,
2403 };
2404
2405 static struct inet_protosw mptcp_protosw = {
2406 .type = SOCK_STREAM,
2407 .protocol = IPPROTO_MPTCP,
2408 .prot = &mptcp_prot,
2409 .ops = &mptcp_stream_ops,
2410 .flags = INET_PROTOSW_ICSK,
2411 };
2412
2413 void __init mptcp_proto_init(void)
2414 {
2415 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2416
2417 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2418 panic("Failed to allocate MPTCP pcpu counter\n");
2419
2420 mptcp_subflow_init();
2421 mptcp_pm_init();
2422 mptcp_token_init();
2423
2424 if (proto_register(&mptcp_prot, 1) != 0)
2425 panic("Failed to register MPTCP proto.\n");
2426
2427 inet_register_protosw(&mptcp_protosw);
2428
2429 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2430 }
2431
2432 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
2433 static const struct proto_ops mptcp_v6_stream_ops = {
2434 .family = PF_INET6,
2435 .owner = THIS_MODULE,
2436 .release = inet6_release,
2437 .bind = mptcp_bind,
2438 .connect = mptcp_stream_connect,
2439 .socketpair = sock_no_socketpair,
2440 .accept = mptcp_stream_accept,
2441 .getname = inet6_getname,
2442 .poll = mptcp_poll,
2443 .ioctl = inet6_ioctl,
2444 .gettstamp = sock_gettstamp,
2445 .listen = mptcp_listen,
2446 .shutdown = mptcp_shutdown,
2447 .setsockopt = sock_common_setsockopt,
2448 .getsockopt = sock_common_getsockopt,
2449 .sendmsg = inet6_sendmsg,
2450 .recvmsg = inet6_recvmsg,
2451 .mmap = sock_no_mmap,
2452 .sendpage = inet_sendpage,
2453 #ifdef CONFIG_COMPAT
2454 .compat_ioctl = inet6_compat_ioctl,
2455 #endif
2456 };
2457
2458 static struct proto mptcp_v6_prot;
2459
2460 static void mptcp_v6_destroy(struct sock *sk)
2461 {
2462 mptcp_destroy(sk);
2463 inet6_destroy_sock(sk);
2464 }
2465
2466 static struct inet_protosw mptcp_v6_protosw = {
2467 .type = SOCK_STREAM,
2468 .protocol = IPPROTO_MPTCP,
2469 .prot = &mptcp_v6_prot,
2470 .ops = &mptcp_v6_stream_ops,
2471 .flags = INET_PROTOSW_ICSK,
2472 };
2473
2474 int __init mptcp_proto_v6_init(void)
2475 {
2476 int err;
2477
2478 mptcp_v6_prot = mptcp_prot;
2479 strcpy(mptcp_v6_prot.name, "MPTCPv6");
2480 mptcp_v6_prot.slab = NULL;
2481 mptcp_v6_prot.destroy = mptcp_v6_destroy;
2482 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
2483
2484 err = proto_register(&mptcp_v6_prot, 1);
2485 if (err)
2486 return err;
2487
2488 err = inet6_register_protosw(&mptcp_v6_protosw);
2489 if (err)
2490 proto_unregister(&mptcp_v6_prot);
2491
2492 return err;
2493 }
2494 #endif
Ubuntu Jammy Linux kernel mirror