Loading...
Note: File does not exist in v5.4.
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)
30struct mptcp6_sock {
31 struct mptcp_sock msk;
32 struct ipv6_pinfo np;
33};
34#endif
35
36struct mptcp_skb_cb {
37 u32 offset;
38};
39
40#define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
41
42static 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 */
48static 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
56static 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
80static 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
90static 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
115static 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
146static 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 */
155static 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
175static 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
210static 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
230static 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
240static 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
307static 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 */
398static 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
418void 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 }
442wake:
443 sk->sk_data_ready(sk);
444}
445
446static 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
456static bool mptcp_timer_pending(struct sock *sk)
457{
458 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
459}
460
461static 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
473void 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
483void 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
492static 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
513static 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
523static 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
538static 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
549static 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
557static void dfrag_uncharge(struct sock *sk, int len)
558{
559 sk_mem_uncharge(sk, len);
560 sk_wmem_queued_add(sk, -len);
561}
562
563static 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
572static 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
609out:
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 */
622static 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
633static struct mptcp_data_frag *
634mptcp_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
650static 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
797out:
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
806static 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
815static 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
850static 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
862static 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);
886restart:
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);
1000out:
1001 release_sock(sk);
1002 return copied ? : ret;
1003}
1004
1005static 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
1020static 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 */
1062static 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;
1147new_measure:
1148 msk->rcvq_space.copied = 0;
1149 msk->rcvq_space.time = mstamp;
1150}
1151
1152static 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
1171static 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 }
1263out_err:
1264 mptcp_rcv_space_adjust(msk, copied);
1265
1266 release_sock(sk);
1267 return copied;
1268}
1269
1270static 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
1283static 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 */
1307static 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 */
1341static 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
1358static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
1359{
1360 return 0;
1361}
1362
1363static 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
1386static void mptcp_worker(struct work_struct *work)
1387{
1388 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
1389 struct sock *ssk, *sk = &msk->sk.icsk_inet.sk;
1390 int orig_len, orig_offset, mss_now = 0, size_goal = 0;
1391 struct mptcp_data_frag *dfrag;
1392 u64 orig_write_seq;
1393 size_t copied = 0;
1394 struct msghdr msg = {
1395 .msg_flags = MSG_DONTWAIT,
1396 };
1397 long timeo = 0;
1398
1399 lock_sock(sk);
1400 mptcp_clean_una(sk);
1401 mptcp_check_data_fin_ack(sk);
1402 __mptcp_flush_join_list(msk);
1403 __mptcp_move_skbs(msk);
1404
1405 if (msk->pm.status)
1406 pm_work(msk);
1407
1408 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1409 mptcp_check_for_eof(msk);
1410
1411 mptcp_check_data_fin(sk);
1412
1413 if (!test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
1414 goto unlock;
1415
1416 dfrag = mptcp_rtx_head(sk);
1417 if (!dfrag)
1418 goto unlock;
1419
1420 if (!mptcp_ext_cache_refill(msk))
1421 goto reset_unlock;
1422
1423 ssk = mptcp_subflow_get_retrans(msk);
1424 if (!ssk)
1425 goto reset_unlock;
1426
1427 lock_sock(ssk);
1428
1429 orig_len = dfrag->data_len;
1430 orig_offset = dfrag->offset;
1431 orig_write_seq = dfrag->data_seq;
1432 while (dfrag->data_len > 0) {
1433 int ret = mptcp_sendmsg_frag(sk, ssk, &msg, dfrag, &timeo,
1434 &mss_now, &size_goal);
1435 if (ret < 0)
1436 break;
1437
1438 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
1439 copied += ret;
1440 dfrag->data_len -= ret;
1441 dfrag->offset += ret;
1442
1443 if (!mptcp_ext_cache_refill(msk))
1444 break;
1445 }
1446 if (copied)
1447 tcp_push(ssk, msg.msg_flags, mss_now, tcp_sk(ssk)->nonagle,
1448 size_goal);
1449
1450 dfrag->data_seq = orig_write_seq;
1451 dfrag->offset = orig_offset;
1452 dfrag->data_len = orig_len;
1453
1454 mptcp_set_timeout(sk, ssk);
1455 release_sock(ssk);
1456
1457reset_unlock:
1458 if (!mptcp_timer_pending(sk))
1459 mptcp_reset_timer(sk);
1460
1461unlock:
1462 release_sock(sk);
1463 sock_put(sk);
1464}
1465
1466static int __mptcp_init_sock(struct sock *sk)
1467{
1468 struct mptcp_sock *msk = mptcp_sk(sk);
1469
1470 spin_lock_init(&msk->join_list_lock);
1471
1472 INIT_LIST_HEAD(&msk->conn_list);
1473 INIT_LIST_HEAD(&msk->join_list);
1474 INIT_LIST_HEAD(&msk->rtx_queue);
1475 __set_bit(MPTCP_SEND_SPACE, &msk->flags);
1476 INIT_WORK(&msk->work, mptcp_worker);
1477
1478 msk->first = NULL;
1479 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
1480
1481 mptcp_pm_data_init(msk);
1482
1483 /* re-use the csk retrans timer for MPTCP-level retrans */
1484 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
1485
1486 return 0;
1487}
1488
1489static int mptcp_init_sock(struct sock *sk)
1490{
1491 struct net *net = sock_net(sk);
1492 int ret;
1493
1494 if (!mptcp_is_enabled(net))
1495 return -ENOPROTOOPT;
1496
1497 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
1498 return -ENOMEM;
1499
1500 ret = __mptcp_init_sock(sk);
1501 if (ret)
1502 return ret;
1503
1504 ret = __mptcp_socket_create(mptcp_sk(sk));
1505 if (ret)
1506 return ret;
1507
1508 sk_sockets_allocated_inc(sk);
1509 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
1510 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[2];
1511
1512 return 0;
1513}
1514
1515static void __mptcp_clear_xmit(struct sock *sk)
1516{
1517 struct mptcp_sock *msk = mptcp_sk(sk);
1518 struct mptcp_data_frag *dtmp, *dfrag;
1519
1520 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
1521
1522 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
1523 dfrag_clear(sk, dfrag);
1524}
1525
1526static void mptcp_cancel_work(struct sock *sk)
1527{
1528 struct mptcp_sock *msk = mptcp_sk(sk);
1529
1530 if (cancel_work_sync(&msk->work))
1531 sock_put(sk);
1532}
1533
1534static void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
1535{
1536 lock_sock(ssk);
1537
1538 switch (ssk->sk_state) {
1539 case TCP_LISTEN:
1540 if (!(how & RCV_SHUTDOWN))
1541 break;
1542 fallthrough;
1543 case TCP_SYN_SENT:
1544 tcp_disconnect(ssk, O_NONBLOCK);
1545 break;
1546 default:
1547 if (__mptcp_check_fallback(mptcp_sk(sk))) {
1548 pr_debug("Fallback");
1549 ssk->sk_shutdown |= how;
1550 tcp_shutdown(ssk, how);
1551 } else {
1552 pr_debug("Sending DATA_FIN on subflow %p", ssk);
1553 mptcp_set_timeout(sk, ssk);
1554 tcp_send_ack(ssk);
1555 }
1556 break;
1557 }
1558
1559 release_sock(ssk);
1560}
1561
1562static const unsigned char new_state[16] = {
1563 /* current state: new state: action: */
1564 [0 /* (Invalid) */] = TCP_CLOSE,
1565 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1566 [TCP_SYN_SENT] = TCP_CLOSE,
1567 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
1568 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
1569 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
1570 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
1571 [TCP_CLOSE] = TCP_CLOSE,
1572 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
1573 [TCP_LAST_ACK] = TCP_LAST_ACK,
1574 [TCP_LISTEN] = TCP_CLOSE,
1575 [TCP_CLOSING] = TCP_CLOSING,
1576 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
1577};
1578
1579static int mptcp_close_state(struct sock *sk)
1580{
1581 int next = (int)new_state[sk->sk_state];
1582 int ns = next & TCP_STATE_MASK;
1583
1584 inet_sk_state_store(sk, ns);
1585
1586 return next & TCP_ACTION_FIN;
1587}
1588
1589static void mptcp_close(struct sock *sk, long timeout)
1590{
1591 struct mptcp_subflow_context *subflow, *tmp;
1592 struct mptcp_sock *msk = mptcp_sk(sk);
1593 LIST_HEAD(conn_list);
1594
1595 lock_sock(sk);
1596 sk->sk_shutdown = SHUTDOWN_MASK;
1597
1598 if (sk->sk_state == TCP_LISTEN) {
1599 inet_sk_state_store(sk, TCP_CLOSE);
1600 goto cleanup;
1601 } else if (sk->sk_state == TCP_CLOSE) {
1602 goto cleanup;
1603 }
1604
1605 if (__mptcp_check_fallback(msk)) {
1606 goto update_state;
1607 } else if (mptcp_close_state(sk)) {
1608 pr_debug("Sending DATA_FIN sk=%p", sk);
1609 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
1610 WRITE_ONCE(msk->snd_data_fin_enable, 1);
1611
1612 mptcp_for_each_subflow(msk, subflow) {
1613 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
1614
1615 mptcp_subflow_shutdown(sk, tcp_sk, SHUTDOWN_MASK);
1616 }
1617 }
1618
1619 sk_stream_wait_close(sk, timeout);
1620
1621update_state:
1622 inet_sk_state_store(sk, TCP_CLOSE);
1623
1624cleanup:
1625 /* be sure to always acquire the join list lock, to sync vs
1626 * mptcp_finish_join().
1627 */
1628 spin_lock_bh(&msk->join_list_lock);
1629 list_splice_tail_init(&msk->join_list, &msk->conn_list);
1630 spin_unlock_bh(&msk->join_list_lock);
1631 list_splice_init(&msk->conn_list, &conn_list);
1632
1633 __mptcp_clear_xmit(sk);
1634
1635 release_sock(sk);
1636
1637 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
1638 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1639 __mptcp_close_ssk(sk, ssk, subflow, timeout);
1640 }
1641
1642 mptcp_cancel_work(sk);
1643
1644 __skb_queue_purge(&sk->sk_receive_queue);
1645
1646 sk_common_release(sk);
1647}
1648
1649static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
1650{
1651#if IS_ENABLED(CONFIG_MPTCP_IPV6)
1652 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
1653 struct ipv6_pinfo *msk6 = inet6_sk(msk);
1654
1655 msk->sk_v6_daddr = ssk->sk_v6_daddr;
1656 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
1657
1658 if (msk6 && ssk6) {
1659 msk6->saddr = ssk6->saddr;
1660 msk6->flow_label = ssk6->flow_label;
1661 }
1662#endif
1663
1664 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
1665 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
1666 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
1667 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
1668 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
1669 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
1670}
1671
1672static int mptcp_disconnect(struct sock *sk, int flags)
1673{
1674 /* Should never be called.
1675 * inet_stream_connect() calls ->disconnect, but that
1676 * refers to the subflow socket, not the mptcp one.
1677 */
1678 WARN_ON_ONCE(1);
1679 return 0;
1680}
1681
1682#if IS_ENABLED(CONFIG_MPTCP_IPV6)
1683static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
1684{
1685 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
1686
1687 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
1688}
1689#endif
1690
1691struct sock *mptcp_sk_clone(const struct sock *sk,
1692 const struct mptcp_options_received *mp_opt,
1693 struct request_sock *req)
1694{
1695 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
1696 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
1697 struct mptcp_sock *msk;
1698 u64 ack_seq;
1699
1700 if (!nsk)
1701 return NULL;
1702
1703#if IS_ENABLED(CONFIG_MPTCP_IPV6)
1704 if (nsk->sk_family == AF_INET6)
1705 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
1706#endif
1707
1708 __mptcp_init_sock(nsk);
1709
1710 msk = mptcp_sk(nsk);
1711 msk->local_key = subflow_req->local_key;
1712 msk->token = subflow_req->token;
1713 msk->subflow = NULL;
1714 WRITE_ONCE(msk->fully_established, false);
1715
1716 msk->write_seq = subflow_req->idsn + 1;
1717 atomic64_set(&msk->snd_una, msk->write_seq);
1718 if (mp_opt->mp_capable) {
1719 msk->can_ack = true;
1720 msk->remote_key = mp_opt->sndr_key;
1721 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
1722 ack_seq++;
1723 WRITE_ONCE(msk->ack_seq, ack_seq);
1724 }
1725
1726 sock_reset_flag(nsk, SOCK_RCU_FREE);
1727 /* will be fully established after successful MPC subflow creation */
1728 inet_sk_state_store(nsk, TCP_SYN_RECV);
1729 bh_unlock_sock(nsk);
1730
1731 /* keep a single reference */
1732 __sock_put(nsk);
1733 return nsk;
1734}
1735
1736void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
1737{
1738 const struct tcp_sock *tp = tcp_sk(ssk);
1739
1740 msk->rcvq_space.copied = 0;
1741 msk->rcvq_space.rtt_us = 0;
1742
1743 msk->rcvq_space.time = tp->tcp_mstamp;
1744
1745 /* initial rcv_space offering made to peer */
1746 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
1747 TCP_INIT_CWND * tp->advmss);
1748 if (msk->rcvq_space.space == 0)
1749 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
1750}
1751
1752static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
1753 bool kern)
1754{
1755 struct mptcp_sock *msk = mptcp_sk(sk);
1756 struct socket *listener;
1757 struct sock *newsk;
1758
1759 listener = __mptcp_nmpc_socket(msk);
1760 if (WARN_ON_ONCE(!listener)) {
1761 *err = -EINVAL;
1762 return NULL;
1763 }
1764
1765 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
1766 newsk = inet_csk_accept(listener->sk, flags, err, kern);
1767 if (!newsk)
1768 return NULL;
1769
1770 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
1771 if (sk_is_mptcp(newsk)) {
1772 struct mptcp_subflow_context *subflow;
1773 struct sock *new_mptcp_sock;
1774 struct sock *ssk = newsk;
1775
1776 subflow = mptcp_subflow_ctx(newsk);
1777 new_mptcp_sock = subflow->conn;
1778
1779 /* is_mptcp should be false if subflow->conn is missing, see
1780 * subflow_syn_recv_sock()
1781 */
1782 if (WARN_ON_ONCE(!new_mptcp_sock)) {
1783 tcp_sk(newsk)->is_mptcp = 0;
1784 return newsk;
1785 }
1786
1787 /* acquire the 2nd reference for the owning socket */
1788 sock_hold(new_mptcp_sock);
1789
1790 local_bh_disable();
1791 bh_lock_sock(new_mptcp_sock);
1792 msk = mptcp_sk(new_mptcp_sock);
1793 msk->first = newsk;
1794
1795 newsk = new_mptcp_sock;
1796 mptcp_copy_inaddrs(newsk, ssk);
1797 list_add(&subflow->node, &msk->conn_list);
1798
1799 mptcp_rcv_space_init(msk, ssk);
1800 bh_unlock_sock(new_mptcp_sock);
1801
1802 __MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
1803 local_bh_enable();
1804 } else {
1805 MPTCP_INC_STATS(sock_net(sk),
1806 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
1807 }
1808
1809 return newsk;
1810}
1811
1812static void mptcp_destroy(struct sock *sk)
1813{
1814 struct mptcp_sock *msk = mptcp_sk(sk);
1815
1816 mptcp_token_destroy(msk);
1817 if (msk->cached_ext)
1818 __skb_ext_put(msk->cached_ext);
1819
1820 sk_sockets_allocated_dec(sk);
1821}
1822
1823static int mptcp_setsockopt_sol_socket(struct mptcp_sock *msk, int optname,
1824 sockptr_t optval, unsigned int optlen)
1825{
1826 struct sock *sk = (struct sock *)msk;
1827 struct socket *ssock;
1828 int ret;
1829
1830 switch (optname) {
1831 case SO_REUSEPORT:
1832 case SO_REUSEADDR:
1833 lock_sock(sk);
1834 ssock = __mptcp_nmpc_socket(msk);
1835 if (!ssock) {
1836 release_sock(sk);
1837 return -EINVAL;
1838 }
1839
1840 ret = sock_setsockopt(ssock, SOL_SOCKET, optname, optval, optlen);
1841 if (ret == 0) {
1842 if (optname == SO_REUSEPORT)
1843 sk->sk_reuseport = ssock->sk->sk_reuseport;
1844 else if (optname == SO_REUSEADDR)
1845 sk->sk_reuse = ssock->sk->sk_reuse;
1846 }
1847 release_sock(sk);
1848 return ret;
1849 }
1850
1851 return sock_setsockopt(sk->sk_socket, SOL_SOCKET, optname, optval, optlen);
1852}
1853
1854static int mptcp_setsockopt_v6(struct mptcp_sock *msk, int optname,
1855 sockptr_t optval, unsigned int optlen)
1856{
1857 struct sock *sk = (struct sock *)msk;
1858 int ret = -EOPNOTSUPP;
1859 struct socket *ssock;
1860
1861 switch (optname) {
1862 case IPV6_V6ONLY:
1863 lock_sock(sk);
1864 ssock = __mptcp_nmpc_socket(msk);
1865 if (!ssock) {
1866 release_sock(sk);
1867 return -EINVAL;
1868 }
1869
1870 ret = tcp_setsockopt(ssock->sk, SOL_IPV6, optname, optval, optlen);
1871 if (ret == 0)
1872 sk->sk_ipv6only = ssock->sk->sk_ipv6only;
1873
1874 release_sock(sk);
1875 break;
1876 }
1877
1878 return ret;
1879}
1880
1881static int mptcp_setsockopt(struct sock *sk, int level, int optname,
1882 sockptr_t optval, unsigned int optlen)
1883{
1884 struct mptcp_sock *msk = mptcp_sk(sk);
1885 struct sock *ssk;
1886
1887 pr_debug("msk=%p", msk);
1888
1889 if (level == SOL_SOCKET)
1890 return mptcp_setsockopt_sol_socket(msk, optname, optval, optlen);
1891
1892 /* @@ the meaning of setsockopt() when the socket is connected and
1893 * there are multiple subflows is not yet defined. It is up to the
1894 * MPTCP-level socket to configure the subflows until the subflow
1895 * is in TCP fallback, when TCP socket options are passed through
1896 * to the one remaining subflow.
1897 */
1898 lock_sock(sk);
1899 ssk = __mptcp_tcp_fallback(msk);
1900 release_sock(sk);
1901 if (ssk)
1902 return tcp_setsockopt(ssk, level, optname, optval, optlen);
1903
1904 if (level == SOL_IPV6)
1905 return mptcp_setsockopt_v6(msk, optname, optval, optlen);
1906
1907 return -EOPNOTSUPP;
1908}
1909
1910static int mptcp_getsockopt(struct sock *sk, int level, int optname,
1911 char __user *optval, int __user *option)
1912{
1913 struct mptcp_sock *msk = mptcp_sk(sk);
1914 struct sock *ssk;
1915
1916 pr_debug("msk=%p", msk);
1917
1918 /* @@ the meaning of setsockopt() when the socket is connected and
1919 * there are multiple subflows is not yet defined. It is up to the
1920 * MPTCP-level socket to configure the subflows until the subflow
1921 * is in TCP fallback, when socket options are passed through
1922 * to the one remaining subflow.
1923 */
1924 lock_sock(sk);
1925 ssk = __mptcp_tcp_fallback(msk);
1926 release_sock(sk);
1927 if (ssk)
1928 return tcp_getsockopt(ssk, level, optname, optval, option);
1929
1930 return -EOPNOTSUPP;
1931}
1932
1933#define MPTCP_DEFERRED_ALL (TCPF_DELACK_TIMER_DEFERRED | \
1934 TCPF_WRITE_TIMER_DEFERRED)
1935
1936/* this is very alike tcp_release_cb() but we must handle differently a
1937 * different set of events
1938 */
1939static void mptcp_release_cb(struct sock *sk)
1940{
1941 unsigned long flags, nflags;
1942
1943 do {
1944 flags = sk->sk_tsq_flags;
1945 if (!(flags & MPTCP_DEFERRED_ALL))
1946 return;
1947 nflags = flags & ~MPTCP_DEFERRED_ALL;
1948 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1949
1950 sock_release_ownership(sk);
1951
1952 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1953 struct mptcp_sock *msk = mptcp_sk(sk);
1954 struct sock *ssk;
1955
1956 ssk = mptcp_subflow_recv_lookup(msk);
1957 if (!ssk || !schedule_work(&msk->work))
1958 __sock_put(sk);
1959 }
1960
1961 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1962 mptcp_retransmit_handler(sk);
1963 __sock_put(sk);
1964 }
1965}
1966
1967static int mptcp_hash(struct sock *sk)
1968{
1969 /* should never be called,
1970 * we hash the TCP subflows not the master socket
1971 */
1972 WARN_ON_ONCE(1);
1973 return 0;
1974}
1975
1976static void mptcp_unhash(struct sock *sk)
1977{
1978 /* called from sk_common_release(), but nothing to do here */
1979}
1980
1981static int mptcp_get_port(struct sock *sk, unsigned short snum)
1982{
1983 struct mptcp_sock *msk = mptcp_sk(sk);
1984 struct socket *ssock;
1985
1986 ssock = __mptcp_nmpc_socket(msk);
1987 pr_debug("msk=%p, subflow=%p", msk, ssock);
1988 if (WARN_ON_ONCE(!ssock))
1989 return -EINVAL;
1990
1991 return inet_csk_get_port(ssock->sk, snum);
1992}
1993
1994void mptcp_finish_connect(struct sock *ssk)
1995{
1996 struct mptcp_subflow_context *subflow;
1997 struct mptcp_sock *msk;
1998 struct sock *sk;
1999 u64 ack_seq;
2000
2001 subflow = mptcp_subflow_ctx(ssk);
2002 sk = subflow->conn;
2003 msk = mptcp_sk(sk);
2004
2005 pr_debug("msk=%p, token=%u", sk, subflow->token);
2006
2007 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2008 ack_seq++;
2009 subflow->map_seq = ack_seq;
2010 subflow->map_subflow_seq = 1;
2011
2012 /* the socket is not connected yet, no msk/subflow ops can access/race
2013 * accessing the field below
2014 */
2015 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2016 WRITE_ONCE(msk->local_key, subflow->local_key);
2017 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2018 WRITE_ONCE(msk->ack_seq, ack_seq);
2019 WRITE_ONCE(msk->can_ack, 1);
2020 atomic64_set(&msk->snd_una, msk->write_seq);
2021
2022 mptcp_pm_new_connection(msk, 0);
2023
2024 mptcp_rcv_space_init(msk, ssk);
2025}
2026
2027static void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2028{
2029 write_lock_bh(&sk->sk_callback_lock);
2030 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2031 sk_set_socket(sk, parent);
2032 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2033 write_unlock_bh(&sk->sk_callback_lock);
2034}
2035
2036bool mptcp_finish_join(struct sock *sk)
2037{
2038 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
2039 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2040 struct sock *parent = (void *)msk;
2041 struct socket *parent_sock;
2042 bool ret;
2043
2044 pr_debug("msk=%p, subflow=%p", msk, subflow);
2045
2046 /* mptcp socket already closing? */
2047 if (!mptcp_is_fully_established(parent))
2048 return false;
2049
2050 if (!msk->pm.server_side)
2051 return true;
2052
2053 if (!mptcp_pm_allow_new_subflow(msk))
2054 return false;
2055
2056 /* active connections are already on conn_list, and we can't acquire
2057 * msk lock here.
2058 * use the join list lock as synchronization point and double-check
2059 * msk status to avoid racing with mptcp_close()
2060 */
2061 spin_lock_bh(&msk->join_list_lock);
2062 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
2063 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node)))
2064 list_add_tail(&subflow->node, &msk->join_list);
2065 spin_unlock_bh(&msk->join_list_lock);
2066 if (!ret)
2067 return false;
2068
2069 /* attach to msk socket only after we are sure he will deal with us
2070 * at close time
2071 */
2072 parent_sock = READ_ONCE(parent->sk_socket);
2073 if (parent_sock && !sk->sk_socket)
2074 mptcp_sock_graft(sk, parent_sock);
2075 subflow->map_seq = READ_ONCE(msk->ack_seq);
2076 return true;
2077}
2078
2079static bool mptcp_memory_free(const struct sock *sk, int wake)
2080{
2081 struct mptcp_sock *msk = mptcp_sk(sk);
2082
2083 return wake ? test_bit(MPTCP_SEND_SPACE, &msk->flags) : true;
2084}
2085
2086static struct proto mptcp_prot = {
2087 .name = "MPTCP",
2088 .owner = THIS_MODULE,
2089 .init = mptcp_init_sock,
2090 .disconnect = mptcp_disconnect,
2091 .close = mptcp_close,
2092 .accept = mptcp_accept,
2093 .setsockopt = mptcp_setsockopt,
2094 .getsockopt = mptcp_getsockopt,
2095 .shutdown = tcp_shutdown,
2096 .destroy = mptcp_destroy,
2097 .sendmsg = mptcp_sendmsg,
2098 .recvmsg = mptcp_recvmsg,
2099 .release_cb = mptcp_release_cb,
2100 .hash = mptcp_hash,
2101 .unhash = mptcp_unhash,
2102 .get_port = mptcp_get_port,
2103 .sockets_allocated = &mptcp_sockets_allocated,
2104 .memory_allocated = &tcp_memory_allocated,
2105 .memory_pressure = &tcp_memory_pressure,
2106 .stream_memory_free = mptcp_memory_free,
2107 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2108 .sysctl_mem = sysctl_tcp_mem,
2109 .obj_size = sizeof(struct mptcp_sock),
2110 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2111 .no_autobind = true,
2112};
2113
2114static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
2115{
2116 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2117 struct socket *ssock;
2118 int err;
2119
2120 lock_sock(sock->sk);
2121 ssock = __mptcp_nmpc_socket(msk);
2122 if (!ssock) {
2123 err = -EINVAL;
2124 goto unlock;
2125 }
2126
2127 err = ssock->ops->bind(ssock, uaddr, addr_len);
2128 if (!err)
2129 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2130
2131unlock:
2132 release_sock(sock->sk);
2133 return err;
2134}
2135
2136static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
2137 struct mptcp_subflow_context *subflow)
2138{
2139 subflow->request_mptcp = 0;
2140 __mptcp_do_fallback(msk);
2141}
2142
2143static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
2144 int addr_len, int flags)
2145{
2146 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2147 struct mptcp_subflow_context *subflow;
2148 struct socket *ssock;
2149 int err;
2150
2151 lock_sock(sock->sk);
2152 if (sock->state != SS_UNCONNECTED && msk->subflow) {
2153 /* pending connection or invalid state, let existing subflow
2154 * cope with that
2155 */
2156 ssock = msk->subflow;
2157 goto do_connect;
2158 }
2159
2160 ssock = __mptcp_nmpc_socket(msk);
2161 if (!ssock) {
2162 err = -EINVAL;
2163 goto unlock;
2164 }
2165
2166 mptcp_token_destroy(msk);
2167 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
2168 subflow = mptcp_subflow_ctx(ssock->sk);
2169#ifdef CONFIG_TCP_MD5SIG
2170 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
2171 * TCP option space.
2172 */
2173 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
2174 mptcp_subflow_early_fallback(msk, subflow);
2175#endif
2176 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk))
2177 mptcp_subflow_early_fallback(msk, subflow);
2178
2179do_connect:
2180 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
2181 sock->state = ssock->state;
2182
2183 /* on successful connect, the msk state will be moved to established by
2184 * subflow_finish_connect()
2185 */
2186 if (!err || err == -EINPROGRESS)
2187 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2188 else
2189 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2190
2191unlock:
2192 release_sock(sock->sk);
2193 return err;
2194}
2195
2196static int mptcp_listen(struct socket *sock, int backlog)
2197{
2198 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2199 struct socket *ssock;
2200 int err;
2201
2202 pr_debug("msk=%p", msk);
2203
2204 lock_sock(sock->sk);
2205 ssock = __mptcp_nmpc_socket(msk);
2206 if (!ssock) {
2207 err = -EINVAL;
2208 goto unlock;
2209 }
2210
2211 mptcp_token_destroy(msk);
2212 inet_sk_state_store(sock->sk, TCP_LISTEN);
2213 sock_set_flag(sock->sk, SOCK_RCU_FREE);
2214
2215 err = ssock->ops->listen(ssock, backlog);
2216 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
2217 if (!err)
2218 mptcp_copy_inaddrs(sock->sk, ssock->sk);
2219
2220unlock:
2221 release_sock(sock->sk);
2222 return err;
2223}
2224
2225static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
2226 int flags, bool kern)
2227{
2228 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2229 struct socket *ssock;
2230 int err;
2231
2232 pr_debug("msk=%p", msk);
2233
2234 lock_sock(sock->sk);
2235 if (sock->sk->sk_state != TCP_LISTEN)
2236 goto unlock_fail;
2237
2238 ssock = __mptcp_nmpc_socket(msk);
2239 if (!ssock)
2240 goto unlock_fail;
2241
2242 clear_bit(MPTCP_DATA_READY, &msk->flags);
2243 sock_hold(ssock->sk);
2244 release_sock(sock->sk);
2245
2246 err = ssock->ops->accept(sock, newsock, flags, kern);
2247 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
2248 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
2249 struct mptcp_subflow_context *subflow;
2250
2251 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
2252 * This is needed so NOSPACE flag can be set from tcp stack.
2253 */
2254 __mptcp_flush_join_list(msk);
2255 mptcp_for_each_subflow(msk, subflow) {
2256 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2257
2258 if (!ssk->sk_socket)
2259 mptcp_sock_graft(ssk, newsock);
2260 }
2261 }
2262
2263 if (inet_csk_listen_poll(ssock->sk))
2264 set_bit(MPTCP_DATA_READY, &msk->flags);
2265 sock_put(ssock->sk);
2266 return err;
2267
2268unlock_fail:
2269 release_sock(sock->sk);
2270 return -EINVAL;
2271}
2272
2273static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
2274{
2275 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM :
2276 0;
2277}
2278
2279static __poll_t mptcp_poll(struct file *file, struct socket *sock,
2280 struct poll_table_struct *wait)
2281{
2282 struct sock *sk = sock->sk;
2283 struct mptcp_sock *msk;
2284 __poll_t mask = 0;
2285 int state;
2286
2287 msk = mptcp_sk(sk);
2288 sock_poll_wait(file, sock, wait);
2289
2290 state = inet_sk_state_load(sk);
2291 if (state == TCP_LISTEN)
2292 return mptcp_check_readable(msk);
2293
2294 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
2295 mask |= mptcp_check_readable(msk);
2296 if (sk_stream_is_writeable(sk) &&
2297 test_bit(MPTCP_SEND_SPACE, &msk->flags))
2298 mask |= EPOLLOUT | EPOLLWRNORM;
2299 }
2300 if (sk->sk_shutdown & RCV_SHUTDOWN)
2301 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2302
2303 return mask;
2304}
2305
2306static int mptcp_shutdown(struct socket *sock, int how)
2307{
2308 struct mptcp_sock *msk = mptcp_sk(sock->sk);
2309 struct mptcp_subflow_context *subflow;
2310 int ret = 0;
2311
2312 pr_debug("sk=%p, how=%d", msk, how);
2313
2314 lock_sock(sock->sk);
2315
2316 how++;
2317 if ((how & ~SHUTDOWN_MASK) || !how) {
2318 ret = -EINVAL;
2319 goto out_unlock;
2320 }
2321
2322 if (sock->state == SS_CONNECTING) {
2323 if ((1 << sock->sk->sk_state) &
2324 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE))
2325 sock->state = SS_DISCONNECTING;
2326 else
2327 sock->state = SS_CONNECTED;
2328 }
2329
2330 /* If we've already sent a FIN, or it's a closed state, skip this. */
2331 if (__mptcp_check_fallback(msk)) {
2332 if (how == SHUT_WR || how == SHUT_RDWR)
2333 inet_sk_state_store(sock->sk, TCP_FIN_WAIT1);
2334
2335 mptcp_for_each_subflow(msk, subflow) {
2336 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2337
2338 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2339 }
2340 } else if ((how & SEND_SHUTDOWN) &&
2341 ((1 << sock->sk->sk_state) &
2342 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2343 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) &&
2344 mptcp_close_state(sock->sk)) {
2345 __mptcp_flush_join_list(msk);
2346
2347 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2348 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2349
2350 mptcp_for_each_subflow(msk, subflow) {
2351 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2352
2353 mptcp_subflow_shutdown(sock->sk, tcp_sk, how);
2354 }
2355 }
2356
2357 /* Wake up anyone sleeping in poll. */
2358 sock->sk->sk_state_change(sock->sk);
2359
2360out_unlock:
2361 release_sock(sock->sk);
2362
2363 return ret;
2364}
2365
2366static const struct proto_ops mptcp_stream_ops = {
2367 .family = PF_INET,
2368 .owner = THIS_MODULE,
2369 .release = inet_release,
2370 .bind = mptcp_bind,
2371 .connect = mptcp_stream_connect,
2372 .socketpair = sock_no_socketpair,
2373 .accept = mptcp_stream_accept,
2374 .getname = inet_getname,
2375 .poll = mptcp_poll,
2376 .ioctl = inet_ioctl,
2377 .gettstamp = sock_gettstamp,
2378 .listen = mptcp_listen,
2379 .shutdown = mptcp_shutdown,
2380 .setsockopt = sock_common_setsockopt,
2381 .getsockopt = sock_common_getsockopt,
2382 .sendmsg = inet_sendmsg,
2383 .recvmsg = inet_recvmsg,
2384 .mmap = sock_no_mmap,
2385 .sendpage = inet_sendpage,
2386};
2387
2388static struct inet_protosw mptcp_protosw = {
2389 .type = SOCK_STREAM,
2390 .protocol = IPPROTO_MPTCP,
2391 .prot = &mptcp_prot,
2392 .ops = &mptcp_stream_ops,
2393 .flags = INET_PROTOSW_ICSK,
2394};
2395
2396void __init mptcp_proto_init(void)
2397{
2398 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
2399
2400 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
2401 panic("Failed to allocate MPTCP pcpu counter\n");
2402
2403 mptcp_subflow_init();
2404 mptcp_pm_init();
2405 mptcp_token_init();
2406
2407 if (proto_register(&mptcp_prot, 1) != 0)
2408 panic("Failed to register MPTCP proto.\n");
2409
2410 inet_register_protosw(&mptcp_protosw);
2411
2412 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
2413}
2414
2415#if IS_ENABLED(CONFIG_MPTCP_IPV6)
2416static const struct proto_ops mptcp_v6_stream_ops = {
2417 .family = PF_INET6,
2418 .owner = THIS_MODULE,
2419 .release = inet6_release,
2420 .bind = mptcp_bind,
2421 .connect = mptcp_stream_connect,
2422 .socketpair = sock_no_socketpair,
2423 .accept = mptcp_stream_accept,
2424 .getname = inet6_getname,
2425 .poll = mptcp_poll,
2426 .ioctl = inet6_ioctl,
2427 .gettstamp = sock_gettstamp,
2428 .listen = mptcp_listen,
2429 .shutdown = mptcp_shutdown,
2430 .setsockopt = sock_common_setsockopt,
2431 .getsockopt = sock_common_getsockopt,
2432 .sendmsg = inet6_sendmsg,
2433 .recvmsg = inet6_recvmsg,
2434 .mmap = sock_no_mmap,
2435 .sendpage = inet_sendpage,
2436#ifdef CONFIG_COMPAT
2437 .compat_ioctl = inet6_compat_ioctl,
2438#endif
2439};
2440
2441static struct proto mptcp_v6_prot;
2442
2443static void mptcp_v6_destroy(struct sock *sk)
2444{
2445 mptcp_destroy(sk);
2446 inet6_destroy_sock(sk);
2447}
2448
2449static struct inet_protosw mptcp_v6_protosw = {
2450 .type = SOCK_STREAM,
2451 .protocol = IPPROTO_MPTCP,
2452 .prot = &mptcp_v6_prot,
2453 .ops = &mptcp_v6_stream_ops,
2454 .flags = INET_PROTOSW_ICSK,
2455};
2456
2457int __init mptcp_proto_v6_init(void)
2458{
2459 int err;
2460
2461 mptcp_v6_prot = mptcp_prot;
2462 strcpy(mptcp_v6_prot.name, "MPTCPv6");
2463 mptcp_v6_prot.slab = NULL;
2464 mptcp_v6_prot.destroy = mptcp_v6_destroy;
2465 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
2466
2467 err = proto_register(&mptcp_v6_prot, 1);
2468 if (err)
2469 return err;
2470
2471 err = inet6_register_protosw(&mptcp_v6_protosw);
2472 if (err)
2473 proto_unregister(&mptcp_v6_prot);
2474
2475 return err;
2476}
2477#endif