Loading...
Note: File does not exist in v4.17.
1// SPDX-License-Identifier: GPL-2.0
2/* Multipath TCP
3 *
4 * Copyright (c) 2017 - 2019, Intel Corporation.
5 */
6
7#define pr_fmt(fmt) "MPTCP: " fmt
8
9#include <linux/kernel.h>
10#include <linux/module.h>
11#include <linux/netdevice.h>
12#include <linux/sched/signal.h>
13#include <linux/atomic.h>
14#include <net/sock.h>
15#include <net/inet_common.h>
16#include <net/inet_hashtables.h>
17#include <net/protocol.h>
18#include <net/tcp.h>
19#include <net/tcp_states.h>
20#if IS_ENABLED(CONFIG_MPTCP_IPV6)
21#include <net/transp_v6.h>
22#endif
23#include <net/mptcp.h>
24#include <net/xfrm.h>
25#include "protocol.h"
26#include "mib.h"
27
28#define CREATE_TRACE_POINTS
29#include <trace/events/mptcp.h>
30
31#if IS_ENABLED(CONFIG_MPTCP_IPV6)
32struct mptcp6_sock {
33 struct mptcp_sock msk;
34 struct ipv6_pinfo np;
35};
36#endif
37
38struct mptcp_skb_cb {
39 u64 map_seq;
40 u64 end_seq;
41 u32 offset;
42 u8 has_rxtstamp:1;
43};
44
45#define MPTCP_SKB_CB(__skb) ((struct mptcp_skb_cb *)&((__skb)->cb[0]))
46
47enum {
48 MPTCP_CMSG_TS = BIT(0),
49};
50
51static struct percpu_counter mptcp_sockets_allocated;
52
53static void __mptcp_destroy_sock(struct sock *sk);
54static void __mptcp_check_send_data_fin(struct sock *sk);
55
56DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
57static struct net_device mptcp_napi_dev;
58
59/* If msk has an initial subflow socket, and the MP_CAPABLE handshake has not
60 * completed yet or has failed, return the subflow socket.
61 * Otherwise return NULL.
62 */
63struct socket *__mptcp_nmpc_socket(const struct mptcp_sock *msk)
64{
65 if (!msk->subflow || READ_ONCE(msk->can_ack))
66 return NULL;
67
68 return msk->subflow;
69}
70
71/* Returns end sequence number of the receiver's advertised window */
72static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
73{
74 return READ_ONCE(msk->wnd_end);
75}
76
77static bool mptcp_is_tcpsk(struct sock *sk)
78{
79 struct socket *sock = sk->sk_socket;
80
81 if (unlikely(sk->sk_prot == &tcp_prot)) {
82 /* we are being invoked after mptcp_accept() has
83 * accepted a non-mp-capable flow: sk is a tcp_sk,
84 * not an mptcp one.
85 *
86 * Hand the socket over to tcp so all further socket ops
87 * bypass mptcp.
88 */
89 sock->ops = &inet_stream_ops;
90 return true;
91#if IS_ENABLED(CONFIG_MPTCP_IPV6)
92 } else if (unlikely(sk->sk_prot == &tcpv6_prot)) {
93 sock->ops = &inet6_stream_ops;
94 return true;
95#endif
96 }
97
98 return false;
99}
100
101static int __mptcp_socket_create(struct mptcp_sock *msk)
102{
103 struct mptcp_subflow_context *subflow;
104 struct sock *sk = (struct sock *)msk;
105 struct socket *ssock;
106 int err;
107
108 err = mptcp_subflow_create_socket(sk, &ssock);
109 if (err)
110 return err;
111
112 msk->first = ssock->sk;
113 msk->subflow = ssock;
114 subflow = mptcp_subflow_ctx(ssock->sk);
115 list_add(&subflow->node, &msk->conn_list);
116 sock_hold(ssock->sk);
117 subflow->request_mptcp = 1;
118 mptcp_sock_graft(msk->first, sk->sk_socket);
119
120 return 0;
121}
122
123static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
124{
125 sk_drops_add(sk, skb);
126 __kfree_skb(skb);
127}
128
129static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
130 struct sk_buff *from)
131{
132 bool fragstolen;
133 int delta;
134
135 if (MPTCP_SKB_CB(from)->offset ||
136 !skb_try_coalesce(to, from, &fragstolen, &delta))
137 return false;
138
139 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
140 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
141 to->len, MPTCP_SKB_CB(from)->end_seq);
142 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
143 kfree_skb_partial(from, fragstolen);
144 atomic_add(delta, &sk->sk_rmem_alloc);
145 sk_mem_charge(sk, delta);
146 return true;
147}
148
149static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
150 struct sk_buff *from)
151{
152 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
153 return false;
154
155 return mptcp_try_coalesce((struct sock *)msk, to, from);
156}
157
158/* "inspired" by tcp_data_queue_ofo(), main differences:
159 * - use mptcp seqs
160 * - don't cope with sacks
161 */
162static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
163{
164 struct sock *sk = (struct sock *)msk;
165 struct rb_node **p, *parent;
166 u64 seq, end_seq, max_seq;
167 struct sk_buff *skb1;
168
169 seq = MPTCP_SKB_CB(skb)->map_seq;
170 end_seq = MPTCP_SKB_CB(skb)->end_seq;
171 max_seq = READ_ONCE(msk->rcv_wnd_sent);
172
173 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
174 RB_EMPTY_ROOT(&msk->out_of_order_queue));
175 if (after64(end_seq, max_seq)) {
176 /* out of window */
177 mptcp_drop(sk, skb);
178 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
179 (unsigned long long)end_seq - (unsigned long)max_seq,
180 (unsigned long long)msk->rcv_wnd_sent);
181 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
182 return;
183 }
184
185 p = &msk->out_of_order_queue.rb_node;
186 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
187 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
188 rb_link_node(&skb->rbnode, NULL, p);
189 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
190 msk->ooo_last_skb = skb;
191 goto end;
192 }
193
194 /* with 2 subflows, adding at end of ooo queue is quite likely
195 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
196 */
197 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
198 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
199 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
200 return;
201 }
202
203 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
204 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
205 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
206 parent = &msk->ooo_last_skb->rbnode;
207 p = &parent->rb_right;
208 goto insert;
209 }
210
211 /* Find place to insert this segment. Handle overlaps on the way. */
212 parent = NULL;
213 while (*p) {
214 parent = *p;
215 skb1 = rb_to_skb(parent);
216 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
217 p = &parent->rb_left;
218 continue;
219 }
220 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
221 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
222 /* All the bits are present. Drop. */
223 mptcp_drop(sk, skb);
224 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
225 return;
226 }
227 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
228 /* partial overlap:
229 * | skb |
230 * | skb1 |
231 * continue traversing
232 */
233 } else {
234 /* skb's seq == skb1's seq and skb covers skb1.
235 * Replace skb1 with skb.
236 */
237 rb_replace_node(&skb1->rbnode, &skb->rbnode,
238 &msk->out_of_order_queue);
239 mptcp_drop(sk, skb1);
240 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
241 goto merge_right;
242 }
243 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
244 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
245 return;
246 }
247 p = &parent->rb_right;
248 }
249
250insert:
251 /* Insert segment into RB tree. */
252 rb_link_node(&skb->rbnode, parent, p);
253 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
254
255merge_right:
256 /* Remove other segments covered by skb. */
257 while ((skb1 = skb_rb_next(skb)) != NULL) {
258 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
259 break;
260 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
261 mptcp_drop(sk, skb1);
262 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
263 }
264 /* If there is no skb after us, we are the last_skb ! */
265 if (!skb1)
266 msk->ooo_last_skb = skb;
267
268end:
269 skb_condense(skb);
270 skb_set_owner_r(skb, sk);
271}
272
273static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
274 struct sk_buff *skb, unsigned int offset,
275 size_t copy_len)
276{
277 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
278 struct sock *sk = (struct sock *)msk;
279 struct sk_buff *tail;
280 bool has_rxtstamp;
281
282 __skb_unlink(skb, &ssk->sk_receive_queue);
283
284 skb_ext_reset(skb);
285 skb_orphan(skb);
286
287 /* try to fetch required memory from subflow */
288 if (!sk_rmem_schedule(sk, skb, skb->truesize)) {
289 int amount = sk_mem_pages(skb->truesize) << SK_MEM_QUANTUM_SHIFT;
290
291 if (ssk->sk_forward_alloc < amount)
292 goto drop;
293
294 ssk->sk_forward_alloc -= amount;
295 sk->sk_forward_alloc += amount;
296 }
297
298 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
299
300 /* the skb map_seq accounts for the skb offset:
301 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
302 * value
303 */
304 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
305 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
306 MPTCP_SKB_CB(skb)->offset = offset;
307 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
308
309 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
310 /* in sequence */
311 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
312 tail = skb_peek_tail(&sk->sk_receive_queue);
313 if (tail && mptcp_try_coalesce(sk, tail, skb))
314 return true;
315
316 skb_set_owner_r(skb, sk);
317 __skb_queue_tail(&sk->sk_receive_queue, skb);
318 return true;
319 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
320 mptcp_data_queue_ofo(msk, skb);
321 return false;
322 }
323
324 /* old data, keep it simple and drop the whole pkt, sender
325 * will retransmit as needed, if needed.
326 */
327 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
328drop:
329 mptcp_drop(sk, skb);
330 return false;
331}
332
333static void mptcp_stop_timer(struct sock *sk)
334{
335 struct inet_connection_sock *icsk = inet_csk(sk);
336
337 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
338 mptcp_sk(sk)->timer_ival = 0;
339}
340
341static void mptcp_close_wake_up(struct sock *sk)
342{
343 if (sock_flag(sk, SOCK_DEAD))
344 return;
345
346 sk->sk_state_change(sk);
347 if (sk->sk_shutdown == SHUTDOWN_MASK ||
348 sk->sk_state == TCP_CLOSE)
349 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
350 else
351 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
352}
353
354static bool mptcp_pending_data_fin_ack(struct sock *sk)
355{
356 struct mptcp_sock *msk = mptcp_sk(sk);
357
358 return !__mptcp_check_fallback(msk) &&
359 ((1 << sk->sk_state) &
360 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
361 msk->write_seq == READ_ONCE(msk->snd_una);
362}
363
364static void mptcp_check_data_fin_ack(struct sock *sk)
365{
366 struct mptcp_sock *msk = mptcp_sk(sk);
367
368 /* Look for an acknowledged DATA_FIN */
369 if (mptcp_pending_data_fin_ack(sk)) {
370 WRITE_ONCE(msk->snd_data_fin_enable, 0);
371
372 switch (sk->sk_state) {
373 case TCP_FIN_WAIT1:
374 inet_sk_state_store(sk, TCP_FIN_WAIT2);
375 break;
376 case TCP_CLOSING:
377 case TCP_LAST_ACK:
378 inet_sk_state_store(sk, TCP_CLOSE);
379 break;
380 }
381
382 mptcp_close_wake_up(sk);
383 }
384}
385
386static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
387{
388 struct mptcp_sock *msk = mptcp_sk(sk);
389
390 if (READ_ONCE(msk->rcv_data_fin) &&
391 ((1 << sk->sk_state) &
392 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
393 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
394
395 if (msk->ack_seq == rcv_data_fin_seq) {
396 if (seq)
397 *seq = rcv_data_fin_seq;
398
399 return true;
400 }
401 }
402
403 return false;
404}
405
406static void mptcp_set_datafin_timeout(const struct sock *sk)
407{
408 struct inet_connection_sock *icsk = inet_csk(sk);
409
410 mptcp_sk(sk)->timer_ival = min(TCP_RTO_MAX,
411 TCP_RTO_MIN << icsk->icsk_retransmits);
412}
413
414static void mptcp_set_timeout(const struct sock *sk, const struct sock *ssk)
415{
416 long tout = ssk && inet_csk(ssk)->icsk_pending ?
417 inet_csk(ssk)->icsk_timeout - jiffies : 0;
418
419 if (tout <= 0)
420 tout = mptcp_sk(sk)->timer_ival;
421 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
422}
423
424static bool tcp_can_send_ack(const struct sock *ssk)
425{
426 return !((1 << inet_sk_state_load(ssk)) &
427 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
428}
429
430void mptcp_subflow_send_ack(struct sock *ssk)
431{
432 bool slow;
433
434 slow = lock_sock_fast(ssk);
435 if (tcp_can_send_ack(ssk))
436 tcp_send_ack(ssk);
437 unlock_sock_fast(ssk, slow);
438}
439
440static void mptcp_send_ack(struct mptcp_sock *msk)
441{
442 struct mptcp_subflow_context *subflow;
443
444 mptcp_for_each_subflow(msk, subflow)
445 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
446}
447
448static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
449{
450 bool slow;
451
452 slow = lock_sock_fast(ssk);
453 if (tcp_can_send_ack(ssk))
454 tcp_cleanup_rbuf(ssk, 1);
455 unlock_sock_fast(ssk, slow);
456}
457
458static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
459{
460 const struct inet_connection_sock *icsk = inet_csk(ssk);
461 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
462 const struct tcp_sock *tp = tcp_sk(ssk);
463
464 return (ack_pending & ICSK_ACK_SCHED) &&
465 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
466 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
467 (rx_empty && ack_pending &
468 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
469}
470
471static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
472{
473 int old_space = READ_ONCE(msk->old_wspace);
474 struct mptcp_subflow_context *subflow;
475 struct sock *sk = (struct sock *)msk;
476 int space = __mptcp_space(sk);
477 bool cleanup, rx_empty;
478
479 cleanup = (space > 0) && (space >= (old_space << 1));
480 rx_empty = !__mptcp_rmem(sk);
481
482 mptcp_for_each_subflow(msk, subflow) {
483 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
484
485 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
486 mptcp_subflow_cleanup_rbuf(ssk);
487 }
488}
489
490static bool mptcp_check_data_fin(struct sock *sk)
491{
492 struct mptcp_sock *msk = mptcp_sk(sk);
493 u64 rcv_data_fin_seq;
494 bool ret = false;
495
496 if (__mptcp_check_fallback(msk))
497 return ret;
498
499 /* Need to ack a DATA_FIN received from a peer while this side
500 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
501 * msk->rcv_data_fin was set when parsing the incoming options
502 * at the subflow level and the msk lock was not held, so this
503 * is the first opportunity to act on the DATA_FIN and change
504 * the msk state.
505 *
506 * If we are caught up to the sequence number of the incoming
507 * DATA_FIN, send the DATA_ACK now and do state transition. If
508 * not caught up, do nothing and let the recv code send DATA_ACK
509 * when catching up.
510 */
511
512 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
513 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
514 WRITE_ONCE(msk->rcv_data_fin, 0);
515
516 sk->sk_shutdown |= RCV_SHUTDOWN;
517 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
518
519 switch (sk->sk_state) {
520 case TCP_ESTABLISHED:
521 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
522 break;
523 case TCP_FIN_WAIT1:
524 inet_sk_state_store(sk, TCP_CLOSING);
525 break;
526 case TCP_FIN_WAIT2:
527 inet_sk_state_store(sk, TCP_CLOSE);
528 break;
529 default:
530 /* Other states not expected */
531 WARN_ON_ONCE(1);
532 break;
533 }
534
535 ret = true;
536 mptcp_set_timeout(sk, NULL);
537 mptcp_send_ack(msk);
538 mptcp_close_wake_up(sk);
539 }
540 return ret;
541}
542
543static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
544 struct sock *ssk,
545 unsigned int *bytes)
546{
547 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
548 struct sock *sk = (struct sock *)msk;
549 unsigned int moved = 0;
550 bool more_data_avail;
551 struct tcp_sock *tp;
552 bool done = false;
553 int sk_rbuf;
554
555 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
556
557 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
558 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
559
560 if (unlikely(ssk_rbuf > sk_rbuf)) {
561 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
562 sk_rbuf = ssk_rbuf;
563 }
564 }
565
566 pr_debug("msk=%p ssk=%p", msk, ssk);
567 tp = tcp_sk(ssk);
568 do {
569 u32 map_remaining, offset;
570 u32 seq = tp->copied_seq;
571 struct sk_buff *skb;
572 bool fin;
573
574 /* try to move as much data as available */
575 map_remaining = subflow->map_data_len -
576 mptcp_subflow_get_map_offset(subflow);
577
578 skb = skb_peek(&ssk->sk_receive_queue);
579 if (!skb) {
580 /* if no data is found, a racing workqueue/recvmsg
581 * already processed the new data, stop here or we
582 * can enter an infinite loop
583 */
584 if (!moved)
585 done = true;
586 break;
587 }
588
589 if (__mptcp_check_fallback(msk)) {
590 /* if we are running under the workqueue, TCP could have
591 * collapsed skbs between dummy map creation and now
592 * be sure to adjust the size
593 */
594 map_remaining = skb->len;
595 subflow->map_data_len = skb->len;
596 }
597
598 offset = seq - TCP_SKB_CB(skb)->seq;
599 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
600 if (fin) {
601 done = true;
602 seq++;
603 }
604
605 if (offset < skb->len) {
606 size_t len = skb->len - offset;
607
608 if (tp->urg_data)
609 done = true;
610
611 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
612 moved += len;
613 seq += len;
614
615 if (WARN_ON_ONCE(map_remaining < len))
616 break;
617 } else {
618 WARN_ON_ONCE(!fin);
619 sk_eat_skb(ssk, skb);
620 done = true;
621 }
622
623 WRITE_ONCE(tp->copied_seq, seq);
624 more_data_avail = mptcp_subflow_data_available(ssk);
625
626 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
627 done = true;
628 break;
629 }
630 } while (more_data_avail);
631
632 *bytes += moved;
633 return done;
634}
635
636static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
637{
638 struct sock *sk = (struct sock *)msk;
639 struct sk_buff *skb, *tail;
640 bool moved = false;
641 struct rb_node *p;
642 u64 end_seq;
643
644 p = rb_first(&msk->out_of_order_queue);
645 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
646 while (p) {
647 skb = rb_to_skb(p);
648 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
649 break;
650
651 p = rb_next(p);
652 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
653
654 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
655 msk->ack_seq))) {
656 mptcp_drop(sk, skb);
657 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
658 continue;
659 }
660
661 end_seq = MPTCP_SKB_CB(skb)->end_seq;
662 tail = skb_peek_tail(&sk->sk_receive_queue);
663 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
664 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
665
666 /* skip overlapping data, if any */
667 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
668 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
669 delta);
670 MPTCP_SKB_CB(skb)->offset += delta;
671 __skb_queue_tail(&sk->sk_receive_queue, skb);
672 }
673 msk->ack_seq = end_seq;
674 moved = true;
675 }
676 return moved;
677}
678
679/* In most cases we will be able to lock the mptcp socket. If its already
680 * owned, we need to defer to the work queue to avoid ABBA deadlock.
681 */
682static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
683{
684 struct sock *sk = (struct sock *)msk;
685 unsigned int moved = 0;
686
687 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
688 __mptcp_ofo_queue(msk);
689 if (unlikely(ssk->sk_err)) {
690 if (!sock_owned_by_user(sk))
691 __mptcp_error_report(sk);
692 else
693 set_bit(MPTCP_ERROR_REPORT, &msk->flags);
694 }
695
696 /* If the moves have caught up with the DATA_FIN sequence number
697 * it's time to ack the DATA_FIN and change socket state, but
698 * this is not a good place to change state. Let the workqueue
699 * do it.
700 */
701 if (mptcp_pending_data_fin(sk, NULL))
702 mptcp_schedule_work(sk);
703 return moved > 0;
704}
705
706void mptcp_data_ready(struct sock *sk, struct sock *ssk)
707{
708 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
709 struct mptcp_sock *msk = mptcp_sk(sk);
710 int sk_rbuf, ssk_rbuf;
711
712 /* The peer can send data while we are shutting down this
713 * subflow at msk destruction time, but we must avoid enqueuing
714 * more data to the msk receive queue
715 */
716 if (unlikely(subflow->disposable))
717 return;
718
719 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
720 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
721 if (unlikely(ssk_rbuf > sk_rbuf))
722 sk_rbuf = ssk_rbuf;
723
724 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
725 if (__mptcp_rmem(sk) > sk_rbuf) {
726 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
727 return;
728 }
729
730 /* Wake-up the reader only for in-sequence data */
731 mptcp_data_lock(sk);
732 if (move_skbs_to_msk(msk, ssk))
733 sk->sk_data_ready(sk);
734
735 mptcp_data_unlock(sk);
736}
737
738static bool mptcp_do_flush_join_list(struct mptcp_sock *msk)
739{
740 struct mptcp_subflow_context *subflow;
741 bool ret = false;
742
743 if (likely(list_empty(&msk->join_list)))
744 return false;
745
746 spin_lock_bh(&msk->join_list_lock);
747 list_for_each_entry(subflow, &msk->join_list, node) {
748 u32 sseq = READ_ONCE(subflow->setsockopt_seq);
749
750 mptcp_propagate_sndbuf((struct sock *)msk, mptcp_subflow_tcp_sock(subflow));
751 if (READ_ONCE(msk->setsockopt_seq) != sseq)
752 ret = true;
753 }
754 list_splice_tail_init(&msk->join_list, &msk->conn_list);
755 spin_unlock_bh(&msk->join_list_lock);
756
757 return ret;
758}
759
760void __mptcp_flush_join_list(struct mptcp_sock *msk)
761{
762 if (likely(!mptcp_do_flush_join_list(msk)))
763 return;
764
765 if (!test_and_set_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags))
766 mptcp_schedule_work((struct sock *)msk);
767}
768
769static void mptcp_flush_join_list(struct mptcp_sock *msk)
770{
771 bool sync_needed = test_and_clear_bit(MPTCP_WORK_SYNC_SETSOCKOPT, &msk->flags);
772
773 might_sleep();
774
775 if (!mptcp_do_flush_join_list(msk) && !sync_needed)
776 return;
777
778 mptcp_sockopt_sync_all(msk);
779}
780
781static bool mptcp_timer_pending(struct sock *sk)
782{
783 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
784}
785
786static void mptcp_reset_timer(struct sock *sk)
787{
788 struct inet_connection_sock *icsk = inet_csk(sk);
789 unsigned long tout;
790
791 /* prevent rescheduling on close */
792 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
793 return;
794
795 /* should never be called with mptcp level timer cleared */
796 tout = READ_ONCE(mptcp_sk(sk)->timer_ival);
797 if (WARN_ON_ONCE(!tout))
798 tout = TCP_RTO_MIN;
799 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
800}
801
802bool mptcp_schedule_work(struct sock *sk)
803{
804 if (inet_sk_state_load(sk) != TCP_CLOSE &&
805 schedule_work(&mptcp_sk(sk)->work)) {
806 /* each subflow already holds a reference to the sk, and the
807 * workqueue is invoked by a subflow, so sk can't go away here.
808 */
809 sock_hold(sk);
810 return true;
811 }
812 return false;
813}
814
815void mptcp_subflow_eof(struct sock *sk)
816{
817 if (!test_and_set_bit(MPTCP_WORK_EOF, &mptcp_sk(sk)->flags))
818 mptcp_schedule_work(sk);
819}
820
821static void mptcp_check_for_eof(struct mptcp_sock *msk)
822{
823 struct mptcp_subflow_context *subflow;
824 struct sock *sk = (struct sock *)msk;
825 int receivers = 0;
826
827 mptcp_for_each_subflow(msk, subflow)
828 receivers += !subflow->rx_eof;
829 if (receivers)
830 return;
831
832 if (!(sk->sk_shutdown & RCV_SHUTDOWN)) {
833 /* hopefully temporary hack: propagate shutdown status
834 * to msk, when all subflows agree on it
835 */
836 sk->sk_shutdown |= RCV_SHUTDOWN;
837
838 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
839 sk->sk_data_ready(sk);
840 }
841
842 switch (sk->sk_state) {
843 case TCP_ESTABLISHED:
844 inet_sk_state_store(sk, TCP_CLOSE_WAIT);
845 break;
846 case TCP_FIN_WAIT1:
847 inet_sk_state_store(sk, TCP_CLOSING);
848 break;
849 case TCP_FIN_WAIT2:
850 inet_sk_state_store(sk, TCP_CLOSE);
851 break;
852 default:
853 return;
854 }
855 mptcp_close_wake_up(sk);
856}
857
858static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
859{
860 struct mptcp_subflow_context *subflow;
861 struct sock *sk = (struct sock *)msk;
862
863 sock_owned_by_me(sk);
864
865 mptcp_for_each_subflow(msk, subflow) {
866 if (READ_ONCE(subflow->data_avail))
867 return mptcp_subflow_tcp_sock(subflow);
868 }
869
870 return NULL;
871}
872
873static bool mptcp_skb_can_collapse_to(u64 write_seq,
874 const struct sk_buff *skb,
875 const struct mptcp_ext *mpext)
876{
877 if (!tcp_skb_can_collapse_to(skb))
878 return false;
879
880 /* can collapse only if MPTCP level sequence is in order and this
881 * mapping has not been xmitted yet
882 */
883 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
884 !mpext->frozen;
885}
886
887/* we can append data to the given data frag if:
888 * - there is space available in the backing page_frag
889 * - the data frag tail matches the current page_frag free offset
890 * - the data frag end sequence number matches the current write seq
891 */
892static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
893 const struct page_frag *pfrag,
894 const struct mptcp_data_frag *df)
895{
896 return df && pfrag->page == df->page &&
897 pfrag->size - pfrag->offset > 0 &&
898 pfrag->offset == (df->offset + df->data_len) &&
899 df->data_seq + df->data_len == msk->write_seq;
900}
901
902static int mptcp_wmem_with_overhead(int size)
903{
904 return size + ((sizeof(struct mptcp_data_frag) * size) >> PAGE_SHIFT);
905}
906
907static void __mptcp_wmem_reserve(struct sock *sk, int size)
908{
909 int amount = mptcp_wmem_with_overhead(size);
910 struct mptcp_sock *msk = mptcp_sk(sk);
911
912 WARN_ON_ONCE(msk->wmem_reserved);
913 if (WARN_ON_ONCE(amount < 0))
914 amount = 0;
915
916 if (amount <= sk->sk_forward_alloc)
917 goto reserve;
918
919 /* under memory pressure try to reserve at most a single page
920 * otherwise try to reserve the full estimate and fallback
921 * to a single page before entering the error path
922 */
923 if ((tcp_under_memory_pressure(sk) && amount > PAGE_SIZE) ||
924 !sk_wmem_schedule(sk, amount)) {
925 if (amount <= PAGE_SIZE)
926 goto nomem;
927
928 amount = PAGE_SIZE;
929 if (!sk_wmem_schedule(sk, amount))
930 goto nomem;
931 }
932
933reserve:
934 msk->wmem_reserved = amount;
935 sk->sk_forward_alloc -= amount;
936 return;
937
938nomem:
939 /* we will wait for memory on next allocation */
940 msk->wmem_reserved = -1;
941}
942
943static void __mptcp_update_wmem(struct sock *sk)
944{
945 struct mptcp_sock *msk = mptcp_sk(sk);
946
947#ifdef CONFIG_LOCKDEP
948 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
949#endif
950
951 if (!msk->wmem_reserved)
952 return;
953
954 if (msk->wmem_reserved < 0)
955 msk->wmem_reserved = 0;
956 if (msk->wmem_reserved > 0) {
957 sk->sk_forward_alloc += msk->wmem_reserved;
958 msk->wmem_reserved = 0;
959 }
960}
961
962static bool mptcp_wmem_alloc(struct sock *sk, int size)
963{
964 struct mptcp_sock *msk = mptcp_sk(sk);
965
966 /* check for pre-existing error condition */
967 if (msk->wmem_reserved < 0)
968 return false;
969
970 if (msk->wmem_reserved >= size)
971 goto account;
972
973 mptcp_data_lock(sk);
974 if (!sk_wmem_schedule(sk, size)) {
975 mptcp_data_unlock(sk);
976 return false;
977 }
978
979 sk->sk_forward_alloc -= size;
980 msk->wmem_reserved += size;
981 mptcp_data_unlock(sk);
982
983account:
984 msk->wmem_reserved -= size;
985 return true;
986}
987
988static void mptcp_wmem_uncharge(struct sock *sk, int size)
989{
990 struct mptcp_sock *msk = mptcp_sk(sk);
991
992 if (msk->wmem_reserved < 0)
993 msk->wmem_reserved = 0;
994 msk->wmem_reserved += size;
995}
996
997static void __mptcp_mem_reclaim_partial(struct sock *sk)
998{
999 lockdep_assert_held_once(&sk->sk_lock.slock);
1000 __mptcp_update_wmem(sk);
1001 sk_mem_reclaim_partial(sk);
1002}
1003
1004static void mptcp_mem_reclaim_partial(struct sock *sk)
1005{
1006 struct mptcp_sock *msk = mptcp_sk(sk);
1007
1008 /* if we are experiencing a transint allocation error,
1009 * the forward allocation memory has been already
1010 * released
1011 */
1012 if (msk->wmem_reserved < 0)
1013 return;
1014
1015 mptcp_data_lock(sk);
1016 sk->sk_forward_alloc += msk->wmem_reserved;
1017 sk_mem_reclaim_partial(sk);
1018 msk->wmem_reserved = sk->sk_forward_alloc;
1019 sk->sk_forward_alloc = 0;
1020 mptcp_data_unlock(sk);
1021}
1022
1023static void dfrag_uncharge(struct sock *sk, int len)
1024{
1025 sk_mem_uncharge(sk, len);
1026 sk_wmem_queued_add(sk, -len);
1027}
1028
1029static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1030{
1031 int len = dfrag->data_len + dfrag->overhead;
1032
1033 list_del(&dfrag->list);
1034 dfrag_uncharge(sk, len);
1035 put_page(dfrag->page);
1036}
1037
1038static void __mptcp_clean_una(struct sock *sk)
1039{
1040 struct mptcp_sock *msk = mptcp_sk(sk);
1041 struct mptcp_data_frag *dtmp, *dfrag;
1042 bool cleaned = false;
1043 u64 snd_una;
1044
1045 /* on fallback we just need to ignore snd_una, as this is really
1046 * plain TCP
1047 */
1048 if (__mptcp_check_fallback(msk))
1049 msk->snd_una = READ_ONCE(msk->snd_nxt);
1050
1051 snd_una = msk->snd_una;
1052 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1053 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1054 break;
1055
1056 if (WARN_ON_ONCE(dfrag == msk->first_pending))
1057 break;
1058 dfrag_clear(sk, dfrag);
1059 cleaned = true;
1060 }
1061
1062 dfrag = mptcp_rtx_head(sk);
1063 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1064 u64 delta = snd_una - dfrag->data_seq;
1065
1066 if (WARN_ON_ONCE(delta > dfrag->already_sent))
1067 goto out;
1068
1069 dfrag->data_seq += delta;
1070 dfrag->offset += delta;
1071 dfrag->data_len -= delta;
1072 dfrag->already_sent -= delta;
1073
1074 dfrag_uncharge(sk, delta);
1075 cleaned = true;
1076 }
1077
1078out:
1079 if (cleaned && tcp_under_memory_pressure(sk))
1080 __mptcp_mem_reclaim_partial(sk);
1081
1082 if (snd_una == READ_ONCE(msk->snd_nxt)) {
1083 if (msk->timer_ival && !mptcp_data_fin_enabled(msk))
1084 mptcp_stop_timer(sk);
1085 } else {
1086 mptcp_reset_timer(sk);
1087 }
1088}
1089
1090static void __mptcp_clean_una_wakeup(struct sock *sk)
1091{
1092#ifdef CONFIG_LOCKDEP
1093 WARN_ON_ONCE(!lockdep_is_held(&sk->sk_lock.slock));
1094#endif
1095 __mptcp_clean_una(sk);
1096 mptcp_write_space(sk);
1097}
1098
1099static void mptcp_clean_una_wakeup(struct sock *sk)
1100{
1101 mptcp_data_lock(sk);
1102 __mptcp_clean_una_wakeup(sk);
1103 mptcp_data_unlock(sk);
1104}
1105
1106static void mptcp_enter_memory_pressure(struct sock *sk)
1107{
1108 struct mptcp_subflow_context *subflow;
1109 struct mptcp_sock *msk = mptcp_sk(sk);
1110 bool first = true;
1111
1112 sk_stream_moderate_sndbuf(sk);
1113 mptcp_for_each_subflow(msk, subflow) {
1114 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1115
1116 if (first)
1117 tcp_enter_memory_pressure(ssk);
1118 sk_stream_moderate_sndbuf(ssk);
1119 first = false;
1120 }
1121}
1122
1123/* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1124 * data
1125 */
1126static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1127{
1128 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1129 pfrag, sk->sk_allocation)))
1130 return true;
1131
1132 mptcp_enter_memory_pressure(sk);
1133 return false;
1134}
1135
1136static struct mptcp_data_frag *
1137mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1138 int orig_offset)
1139{
1140 int offset = ALIGN(orig_offset, sizeof(long));
1141 struct mptcp_data_frag *dfrag;
1142
1143 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1144 dfrag->data_len = 0;
1145 dfrag->data_seq = msk->write_seq;
1146 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1147 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1148 dfrag->already_sent = 0;
1149 dfrag->page = pfrag->page;
1150
1151 return dfrag;
1152}
1153
1154struct mptcp_sendmsg_info {
1155 int mss_now;
1156 int size_goal;
1157 u16 limit;
1158 u16 sent;
1159 unsigned int flags;
1160 bool data_lock_held;
1161};
1162
1163static int mptcp_check_allowed_size(struct mptcp_sock *msk, u64 data_seq,
1164 int avail_size)
1165{
1166 u64 window_end = mptcp_wnd_end(msk);
1167
1168 if (__mptcp_check_fallback(msk))
1169 return avail_size;
1170
1171 if (!before64(data_seq + avail_size, window_end)) {
1172 u64 allowed_size = window_end - data_seq;
1173
1174 return min_t(unsigned int, allowed_size, avail_size);
1175 }
1176
1177 return avail_size;
1178}
1179
1180static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1181{
1182 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1183
1184 if (!mpext)
1185 return false;
1186 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1187 return true;
1188}
1189
1190static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1191{
1192 struct sk_buff *skb;
1193
1194 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1195 if (likely(skb)) {
1196 if (likely(__mptcp_add_ext(skb, gfp))) {
1197 skb_reserve(skb, MAX_TCP_HEADER);
1198 skb->reserved_tailroom = skb->end - skb->tail;
1199 return skb;
1200 }
1201 __kfree_skb(skb);
1202 } else {
1203 mptcp_enter_memory_pressure(sk);
1204 }
1205 return NULL;
1206}
1207
1208static bool __mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1209{
1210 struct sk_buff *skb;
1211
1212 if (ssk->sk_tx_skb_cache) {
1213 skb = ssk->sk_tx_skb_cache;
1214 if (unlikely(!skb_ext_find(skb, SKB_EXT_MPTCP) &&
1215 !__mptcp_add_ext(skb, gfp)))
1216 return false;
1217 return true;
1218 }
1219
1220 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1221 if (!skb)
1222 return false;
1223
1224 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1225 ssk->sk_tx_skb_cache = skb;
1226 return true;
1227 }
1228 kfree_skb(skb);
1229 return false;
1230}
1231
1232static bool mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1233{
1234 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1235
1236 if (unlikely(tcp_under_memory_pressure(sk))) {
1237 if (data_lock_held)
1238 __mptcp_mem_reclaim_partial(sk);
1239 else
1240 mptcp_mem_reclaim_partial(sk);
1241 }
1242 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1243}
1244
1245/* note: this always recompute the csum on the whole skb, even
1246 * if we just appended a single frag. More status info needed
1247 */
1248static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1249{
1250 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1251 __wsum csum = ~csum_unfold(mpext->csum);
1252 int offset = skb->len - added;
1253
1254 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1255}
1256
1257static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1258 struct mptcp_data_frag *dfrag,
1259 struct mptcp_sendmsg_info *info)
1260{
1261 u64 data_seq = dfrag->data_seq + info->sent;
1262 struct mptcp_sock *msk = mptcp_sk(sk);
1263 bool zero_window_probe = false;
1264 struct mptcp_ext *mpext = NULL;
1265 struct sk_buff *skb, *tail;
1266 bool must_collapse = false;
1267 int size_bias = 0;
1268 int avail_size;
1269 size_t ret = 0;
1270
1271 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1272 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1273
1274 /* compute send limit */
1275 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1276 avail_size = info->size_goal;
1277 skb = tcp_write_queue_tail(ssk);
1278 if (skb) {
1279 /* Limit the write to the size available in the
1280 * current skb, if any, so that we create at most a new skb.
1281 * Explicitly tells TCP internals to avoid collapsing on later
1282 * queue management operation, to avoid breaking the ext <->
1283 * SSN association set here
1284 */
1285 mpext = skb_ext_find(skb, SKB_EXT_MPTCP);
1286 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1287 TCP_SKB_CB(skb)->eor = 1;
1288 goto alloc_skb;
1289 }
1290
1291 must_collapse = (info->size_goal > skb->len) &&
1292 (skb_shinfo(skb)->nr_frags < sysctl_max_skb_frags);
1293 if (must_collapse) {
1294 size_bias = skb->len;
1295 avail_size = info->size_goal - skb->len;
1296 }
1297 }
1298
1299alloc_skb:
1300 if (!must_collapse &&
1301 !mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held))
1302 return 0;
1303
1304 /* Zero window and all data acked? Probe. */
1305 avail_size = mptcp_check_allowed_size(msk, data_seq, avail_size);
1306 if (avail_size == 0) {
1307 u64 snd_una = READ_ONCE(msk->snd_una);
1308
1309 if (skb || snd_una != msk->snd_nxt)
1310 return 0;
1311 zero_window_probe = true;
1312 data_seq = snd_una - 1;
1313 avail_size = 1;
1314 }
1315
1316 if (WARN_ON_ONCE(info->sent > info->limit ||
1317 info->limit > dfrag->data_len))
1318 return 0;
1319
1320 ret = info->limit - info->sent;
1321 tail = tcp_build_frag(ssk, avail_size + size_bias, info->flags,
1322 dfrag->page, dfrag->offset + info->sent, &ret);
1323 if (!tail) {
1324 tcp_remove_empty_skb(sk, tcp_write_queue_tail(ssk));
1325 return -ENOMEM;
1326 }
1327
1328 /* if the tail skb is still the cached one, collapsing really happened.
1329 */
1330 if (skb == tail) {
1331 TCP_SKB_CB(tail)->tcp_flags &= ~TCPHDR_PSH;
1332 mpext->data_len += ret;
1333 WARN_ON_ONCE(zero_window_probe);
1334 goto out;
1335 }
1336
1337 mpext = skb_ext_find(tail, SKB_EXT_MPTCP);
1338 if (WARN_ON_ONCE(!mpext)) {
1339 /* should never reach here, stream corrupted */
1340 return -EINVAL;
1341 }
1342
1343 memset(mpext, 0, sizeof(*mpext));
1344 mpext->data_seq = data_seq;
1345 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1346 mpext->data_len = ret;
1347 mpext->use_map = 1;
1348 mpext->dsn64 = 1;
1349
1350 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1351 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1352 mpext->dsn64);
1353
1354 if (zero_window_probe) {
1355 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1356 mpext->frozen = 1;
1357 if (READ_ONCE(msk->csum_enabled))
1358 mptcp_update_data_checksum(tail, ret);
1359 tcp_push_pending_frames(ssk);
1360 return 0;
1361 }
1362out:
1363 if (READ_ONCE(msk->csum_enabled))
1364 mptcp_update_data_checksum(tail, ret);
1365 mptcp_subflow_ctx(ssk)->rel_write_seq += ret;
1366 return ret;
1367}
1368
1369#define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1370 sizeof(struct tcphdr) - \
1371 MAX_TCP_OPTION_SPACE - \
1372 sizeof(struct ipv6hdr) - \
1373 sizeof(struct frag_hdr))
1374
1375struct subflow_send_info {
1376 struct sock *ssk;
1377 u64 ratio;
1378};
1379
1380static struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1381{
1382 struct subflow_send_info send_info[2];
1383 struct mptcp_subflow_context *subflow;
1384 int i, nr_active = 0;
1385 struct sock *ssk;
1386 u64 ratio;
1387 u32 pace;
1388
1389 sock_owned_by_me((struct sock *)msk);
1390
1391 if (__mptcp_check_fallback(msk)) {
1392 if (!msk->first)
1393 return NULL;
1394 return sk_stream_memory_free(msk->first) ? msk->first : NULL;
1395 }
1396
1397 /* re-use last subflow, if the burst allow that */
1398 if (msk->last_snd && msk->snd_burst > 0 &&
1399 sk_stream_memory_free(msk->last_snd) &&
1400 mptcp_subflow_active(mptcp_subflow_ctx(msk->last_snd)))
1401 return msk->last_snd;
1402
1403 /* pick the subflow with the lower wmem/wspace ratio */
1404 for (i = 0; i < 2; ++i) {
1405 send_info[i].ssk = NULL;
1406 send_info[i].ratio = -1;
1407 }
1408 mptcp_for_each_subflow(msk, subflow) {
1409 trace_mptcp_subflow_get_send(subflow);
1410 ssk = mptcp_subflow_tcp_sock(subflow);
1411 if (!mptcp_subflow_active(subflow))
1412 continue;
1413
1414 nr_active += !subflow->backup;
1415 if (!sk_stream_memory_free(subflow->tcp_sock) || !tcp_sk(ssk)->snd_wnd)
1416 continue;
1417
1418 pace = READ_ONCE(ssk->sk_pacing_rate);
1419 if (!pace)
1420 continue;
1421
1422 ratio = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32,
1423 pace);
1424 if (ratio < send_info[subflow->backup].ratio) {
1425 send_info[subflow->backup].ssk = ssk;
1426 send_info[subflow->backup].ratio = ratio;
1427 }
1428 }
1429
1430 /* pick the best backup if no other subflow is active */
1431 if (!nr_active)
1432 send_info[0].ssk = send_info[1].ssk;
1433
1434 if (send_info[0].ssk) {
1435 msk->last_snd = send_info[0].ssk;
1436 msk->snd_burst = min_t(int, MPTCP_SEND_BURST_SIZE,
1437 tcp_sk(msk->last_snd)->snd_wnd);
1438 return msk->last_snd;
1439 }
1440
1441 return NULL;
1442}
1443
1444static void mptcp_push_release(struct sock *sk, struct sock *ssk,
1445 struct mptcp_sendmsg_info *info)
1446{
1447 mptcp_set_timeout(sk, ssk);
1448 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1449 release_sock(ssk);
1450}
1451
1452static void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1453{
1454 struct sock *prev_ssk = NULL, *ssk = NULL;
1455 struct mptcp_sock *msk = mptcp_sk(sk);
1456 struct mptcp_sendmsg_info info = {
1457 .flags = flags,
1458 };
1459 struct mptcp_data_frag *dfrag;
1460 int len, copied = 0;
1461
1462 while ((dfrag = mptcp_send_head(sk))) {
1463 info.sent = dfrag->already_sent;
1464 info.limit = dfrag->data_len;
1465 len = dfrag->data_len - dfrag->already_sent;
1466 while (len > 0) {
1467 int ret = 0;
1468
1469 prev_ssk = ssk;
1470 mptcp_flush_join_list(msk);
1471 ssk = mptcp_subflow_get_send(msk);
1472
1473 /* try to keep the subflow socket lock across
1474 * consecutive xmit on the same socket
1475 */
1476 if (ssk != prev_ssk && prev_ssk)
1477 mptcp_push_release(sk, prev_ssk, &info);
1478 if (!ssk)
1479 goto out;
1480
1481 if (ssk != prev_ssk || !prev_ssk)
1482 lock_sock(ssk);
1483
1484 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1485 if (ret <= 0) {
1486 mptcp_push_release(sk, ssk, &info);
1487 goto out;
1488 }
1489
1490 info.sent += ret;
1491 dfrag->already_sent += ret;
1492 msk->snd_nxt += ret;
1493 msk->snd_burst -= ret;
1494 msk->tx_pending_data -= ret;
1495 copied += ret;
1496 len -= ret;
1497 }
1498 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1499 }
1500
1501 /* at this point we held the socket lock for the last subflow we used */
1502 if (ssk)
1503 mptcp_push_release(sk, ssk, &info);
1504
1505out:
1506 if (copied) {
1507 /* start the timer, if it's not pending */
1508 if (!mptcp_timer_pending(sk))
1509 mptcp_reset_timer(sk);
1510 __mptcp_check_send_data_fin(sk);
1511 }
1512}
1513
1514static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk)
1515{
1516 struct mptcp_sock *msk = mptcp_sk(sk);
1517 struct mptcp_sendmsg_info info = {
1518 .data_lock_held = true,
1519 };
1520 struct mptcp_data_frag *dfrag;
1521 struct sock *xmit_ssk;
1522 int len, copied = 0;
1523 bool first = true;
1524
1525 info.flags = 0;
1526 while ((dfrag = mptcp_send_head(sk))) {
1527 info.sent = dfrag->already_sent;
1528 info.limit = dfrag->data_len;
1529 len = dfrag->data_len - dfrag->already_sent;
1530 while (len > 0) {
1531 int ret = 0;
1532
1533 /* the caller already invoked the packet scheduler,
1534 * check for a different subflow usage only after
1535 * spooling the first chunk of data
1536 */
1537 xmit_ssk = first ? ssk : mptcp_subflow_get_send(mptcp_sk(sk));
1538 if (!xmit_ssk)
1539 goto out;
1540 if (xmit_ssk != ssk) {
1541 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
1542 goto out;
1543 }
1544
1545 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
1546 if (ret <= 0)
1547 goto out;
1548
1549 info.sent += ret;
1550 dfrag->already_sent += ret;
1551 msk->snd_nxt += ret;
1552 msk->snd_burst -= ret;
1553 msk->tx_pending_data -= ret;
1554 copied += ret;
1555 len -= ret;
1556 first = false;
1557 }
1558 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1559 }
1560
1561out:
1562 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1563 * not going to flush it via release_sock()
1564 */
1565 __mptcp_update_wmem(sk);
1566 if (copied) {
1567 mptcp_set_timeout(sk, ssk);
1568 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1569 info.size_goal);
1570 if (!mptcp_timer_pending(sk))
1571 mptcp_reset_timer(sk);
1572
1573 if (msk->snd_data_fin_enable &&
1574 msk->snd_nxt + 1 == msk->write_seq)
1575 mptcp_schedule_work(sk);
1576 }
1577}
1578
1579static void mptcp_set_nospace(struct sock *sk)
1580{
1581 /* enable autotune */
1582 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1583
1584 /* will be cleared on avail space */
1585 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1586}
1587
1588static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1589{
1590 struct mptcp_sock *msk = mptcp_sk(sk);
1591 struct page_frag *pfrag;
1592 size_t copied = 0;
1593 int ret = 0;
1594 long timeo;
1595
1596 /* we don't support FASTOPEN yet */
1597 if (msg->msg_flags & MSG_FASTOPEN)
1598 return -EOPNOTSUPP;
1599
1600 /* silently ignore everything else */
1601 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL;
1602
1603 mptcp_lock_sock(sk, __mptcp_wmem_reserve(sk, min_t(size_t, 1 << 20, len)));
1604
1605 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1606
1607 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1608 ret = sk_stream_wait_connect(sk, &timeo);
1609 if (ret)
1610 goto out;
1611 }
1612
1613 pfrag = sk_page_frag(sk);
1614
1615 while (msg_data_left(msg)) {
1616 int total_ts, frag_truesize = 0;
1617 struct mptcp_data_frag *dfrag;
1618 bool dfrag_collapsed;
1619 size_t psize, offset;
1620
1621 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
1622 ret = -EPIPE;
1623 goto out;
1624 }
1625
1626 /* reuse tail pfrag, if possible, or carve a new one from the
1627 * page allocator
1628 */
1629 dfrag = mptcp_pending_tail(sk);
1630 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1631 if (!dfrag_collapsed) {
1632 if (!sk_stream_memory_free(sk))
1633 goto wait_for_memory;
1634
1635 if (!mptcp_page_frag_refill(sk, pfrag))
1636 goto wait_for_memory;
1637
1638 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1639 frag_truesize = dfrag->overhead;
1640 }
1641
1642 /* we do not bound vs wspace, to allow a single packet.
1643 * memory accounting will prevent execessive memory usage
1644 * anyway
1645 */
1646 offset = dfrag->offset + dfrag->data_len;
1647 psize = pfrag->size - offset;
1648 psize = min_t(size_t, psize, msg_data_left(msg));
1649 total_ts = psize + frag_truesize;
1650
1651 if (!mptcp_wmem_alloc(sk, total_ts))
1652 goto wait_for_memory;
1653
1654 if (copy_page_from_iter(dfrag->page, offset, psize,
1655 &msg->msg_iter) != psize) {
1656 mptcp_wmem_uncharge(sk, psize + frag_truesize);
1657 ret = -EFAULT;
1658 goto out;
1659 }
1660
1661 /* data successfully copied into the write queue */
1662 copied += psize;
1663 dfrag->data_len += psize;
1664 frag_truesize += psize;
1665 pfrag->offset += frag_truesize;
1666 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1667 msk->tx_pending_data += psize;
1668
1669 /* charge data on mptcp pending queue to the msk socket
1670 * Note: we charge such data both to sk and ssk
1671 */
1672 sk_wmem_queued_add(sk, frag_truesize);
1673 if (!dfrag_collapsed) {
1674 get_page(dfrag->page);
1675 list_add_tail(&dfrag->list, &msk->rtx_queue);
1676 if (!msk->first_pending)
1677 WRITE_ONCE(msk->first_pending, dfrag);
1678 }
1679 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1680 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1681 !dfrag_collapsed);
1682
1683 continue;
1684
1685wait_for_memory:
1686 mptcp_set_nospace(sk);
1687 __mptcp_push_pending(sk, msg->msg_flags);
1688 ret = sk_stream_wait_memory(sk, &timeo);
1689 if (ret)
1690 goto out;
1691 }
1692
1693 if (copied)
1694 __mptcp_push_pending(sk, msg->msg_flags);
1695
1696out:
1697 release_sock(sk);
1698 return copied ? : ret;
1699}
1700
1701static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1702 struct msghdr *msg,
1703 size_t len, int flags,
1704 struct scm_timestamping_internal *tss,
1705 int *cmsg_flags)
1706{
1707 struct sk_buff *skb, *tmp;
1708 int copied = 0;
1709
1710 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1711 u32 offset = MPTCP_SKB_CB(skb)->offset;
1712 u32 data_len = skb->len - offset;
1713 u32 count = min_t(size_t, len - copied, data_len);
1714 int err;
1715
1716 if (!(flags & MSG_TRUNC)) {
1717 err = skb_copy_datagram_msg(skb, offset, msg, count);
1718 if (unlikely(err < 0)) {
1719 if (!copied)
1720 return err;
1721 break;
1722 }
1723 }
1724
1725 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1726 tcp_update_recv_tstamps(skb, tss);
1727 *cmsg_flags |= MPTCP_CMSG_TS;
1728 }
1729
1730 copied += count;
1731
1732 if (count < data_len) {
1733 if (!(flags & MSG_PEEK))
1734 MPTCP_SKB_CB(skb)->offset += count;
1735 break;
1736 }
1737
1738 if (!(flags & MSG_PEEK)) {
1739 /* we will bulk release the skb memory later */
1740 skb->destructor = NULL;
1741 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1742 __skb_unlink(skb, &msk->receive_queue);
1743 __kfree_skb(skb);
1744 }
1745
1746 if (copied >= len)
1747 break;
1748 }
1749
1750 return copied;
1751}
1752
1753/* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1754 *
1755 * Only difference: Use highest rtt estimate of the subflows in use.
1756 */
1757static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1758{
1759 struct mptcp_subflow_context *subflow;
1760 struct sock *sk = (struct sock *)msk;
1761 u32 time, advmss = 1;
1762 u64 rtt_us, mstamp;
1763
1764 sock_owned_by_me(sk);
1765
1766 if (copied <= 0)
1767 return;
1768
1769 msk->rcvq_space.copied += copied;
1770
1771 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1772 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1773
1774 rtt_us = msk->rcvq_space.rtt_us;
1775 if (rtt_us && time < (rtt_us >> 3))
1776 return;
1777
1778 rtt_us = 0;
1779 mptcp_for_each_subflow(msk, subflow) {
1780 const struct tcp_sock *tp;
1781 u64 sf_rtt_us;
1782 u32 sf_advmss;
1783
1784 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1785
1786 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1787 sf_advmss = READ_ONCE(tp->advmss);
1788
1789 rtt_us = max(sf_rtt_us, rtt_us);
1790 advmss = max(sf_advmss, advmss);
1791 }
1792
1793 msk->rcvq_space.rtt_us = rtt_us;
1794 if (time < (rtt_us >> 3) || rtt_us == 0)
1795 return;
1796
1797 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1798 goto new_measure;
1799
1800 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
1801 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
1802 int rcvmem, rcvbuf;
1803 u64 rcvwin, grow;
1804
1805 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
1806
1807 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
1808
1809 do_div(grow, msk->rcvq_space.space);
1810 rcvwin += (grow << 1);
1811
1812 rcvmem = SKB_TRUESIZE(advmss + MAX_TCP_HEADER);
1813 while (tcp_win_from_space(sk, rcvmem) < advmss)
1814 rcvmem += 128;
1815
1816 do_div(rcvwin, advmss);
1817 rcvbuf = min_t(u64, rcvwin * rcvmem,
1818 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
1819
1820 if (rcvbuf > sk->sk_rcvbuf) {
1821 u32 window_clamp;
1822
1823 window_clamp = tcp_win_from_space(sk, rcvbuf);
1824 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
1825
1826 /* Make subflows follow along. If we do not do this, we
1827 * get drops at subflow level if skbs can't be moved to
1828 * the mptcp rx queue fast enough (announced rcv_win can
1829 * exceed ssk->sk_rcvbuf).
1830 */
1831 mptcp_for_each_subflow(msk, subflow) {
1832 struct sock *ssk;
1833 bool slow;
1834
1835 ssk = mptcp_subflow_tcp_sock(subflow);
1836 slow = lock_sock_fast(ssk);
1837 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
1838 tcp_sk(ssk)->window_clamp = window_clamp;
1839 tcp_cleanup_rbuf(ssk, 1);
1840 unlock_sock_fast(ssk, slow);
1841 }
1842 }
1843 }
1844
1845 msk->rcvq_space.space = msk->rcvq_space.copied;
1846new_measure:
1847 msk->rcvq_space.copied = 0;
1848 msk->rcvq_space.time = mstamp;
1849}
1850
1851static void __mptcp_update_rmem(struct sock *sk)
1852{
1853 struct mptcp_sock *msk = mptcp_sk(sk);
1854
1855 if (!msk->rmem_released)
1856 return;
1857
1858 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
1859 sk_mem_uncharge(sk, msk->rmem_released);
1860 WRITE_ONCE(msk->rmem_released, 0);
1861}
1862
1863static void __mptcp_splice_receive_queue(struct sock *sk)
1864{
1865 struct mptcp_sock *msk = mptcp_sk(sk);
1866
1867 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
1868}
1869
1870static bool __mptcp_move_skbs(struct mptcp_sock *msk)
1871{
1872 struct sock *sk = (struct sock *)msk;
1873 unsigned int moved = 0;
1874 bool ret, done;
1875
1876 mptcp_flush_join_list(msk);
1877 do {
1878 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
1879 bool slowpath;
1880
1881 /* we can have data pending in the subflows only if the msk
1882 * receive buffer was full at subflow_data_ready() time,
1883 * that is an unlikely slow path.
1884 */
1885 if (likely(!ssk))
1886 break;
1887
1888 slowpath = lock_sock_fast(ssk);
1889 mptcp_data_lock(sk);
1890 __mptcp_update_rmem(sk);
1891 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
1892 mptcp_data_unlock(sk);
1893
1894 if (unlikely(ssk->sk_err))
1895 __mptcp_error_report(sk);
1896 unlock_sock_fast(ssk, slowpath);
1897 } while (!done);
1898
1899 /* acquire the data lock only if some input data is pending */
1900 ret = moved > 0;
1901 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
1902 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
1903 mptcp_data_lock(sk);
1904 __mptcp_update_rmem(sk);
1905 ret |= __mptcp_ofo_queue(msk);
1906 __mptcp_splice_receive_queue(sk);
1907 mptcp_data_unlock(sk);
1908 }
1909 if (ret)
1910 mptcp_check_data_fin((struct sock *)msk);
1911 return !skb_queue_empty(&msk->receive_queue);
1912}
1913
1914static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
1915 int nonblock, int flags, int *addr_len)
1916{
1917 struct mptcp_sock *msk = mptcp_sk(sk);
1918 struct scm_timestamping_internal tss;
1919 int copied = 0, cmsg_flags = 0;
1920 int target;
1921 long timeo;
1922
1923 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
1924 if (unlikely(flags & MSG_ERRQUEUE))
1925 return inet_recv_error(sk, msg, len, addr_len);
1926
1927 mptcp_lock_sock(sk, __mptcp_splice_receive_queue(sk));
1928 if (unlikely(sk->sk_state == TCP_LISTEN)) {
1929 copied = -ENOTCONN;
1930 goto out_err;
1931 }
1932
1933 timeo = sock_rcvtimeo(sk, nonblock);
1934
1935 len = min_t(size_t, len, INT_MAX);
1936 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1937
1938 while (copied < len) {
1939 int bytes_read;
1940
1941 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
1942 if (unlikely(bytes_read < 0)) {
1943 if (!copied)
1944 copied = bytes_read;
1945 goto out_err;
1946 }
1947
1948 copied += bytes_read;
1949
1950 /* be sure to advertise window change */
1951 mptcp_cleanup_rbuf(msk);
1952
1953 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
1954 continue;
1955
1956 /* only the master socket status is relevant here. The exit
1957 * conditions mirror closely tcp_recvmsg()
1958 */
1959 if (copied >= target)
1960 break;
1961
1962 if (copied) {
1963 if (sk->sk_err ||
1964 sk->sk_state == TCP_CLOSE ||
1965 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1966 !timeo ||
1967 signal_pending(current))
1968 break;
1969 } else {
1970 if (sk->sk_err) {
1971 copied = sock_error(sk);
1972 break;
1973 }
1974
1975 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
1976 mptcp_check_for_eof(msk);
1977
1978 if (sk->sk_shutdown & RCV_SHUTDOWN) {
1979 /* race breaker: the shutdown could be after the
1980 * previous receive queue check
1981 */
1982 if (__mptcp_move_skbs(msk))
1983 continue;
1984 break;
1985 }
1986
1987 if (sk->sk_state == TCP_CLOSE) {
1988 copied = -ENOTCONN;
1989 break;
1990 }
1991
1992 if (!timeo) {
1993 copied = -EAGAIN;
1994 break;
1995 }
1996
1997 if (signal_pending(current)) {
1998 copied = sock_intr_errno(timeo);
1999 break;
2000 }
2001 }
2002
2003 pr_debug("block timeout %ld", timeo);
2004 sk_wait_data(sk, &timeo, NULL);
2005 }
2006
2007out_err:
2008 if (cmsg_flags && copied >= 0) {
2009 if (cmsg_flags & MPTCP_CMSG_TS)
2010 tcp_recv_timestamp(msg, sk, &tss);
2011 }
2012
2013 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2014 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2015 skb_queue_empty(&msk->receive_queue), copied);
2016 if (!(flags & MSG_PEEK))
2017 mptcp_rcv_space_adjust(msk, copied);
2018
2019 release_sock(sk);
2020 return copied;
2021}
2022
2023static void mptcp_retransmit_timer(struct timer_list *t)
2024{
2025 struct inet_connection_sock *icsk = from_timer(icsk, t,
2026 icsk_retransmit_timer);
2027 struct sock *sk = &icsk->icsk_inet.sk;
2028 struct mptcp_sock *msk = mptcp_sk(sk);
2029
2030 bh_lock_sock(sk);
2031 if (!sock_owned_by_user(sk)) {
2032 /* we need a process context to retransmit */
2033 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2034 mptcp_schedule_work(sk);
2035 } else {
2036 /* delegate our work to tcp_release_cb() */
2037 set_bit(MPTCP_RETRANSMIT, &msk->flags);
2038 }
2039 bh_unlock_sock(sk);
2040 sock_put(sk);
2041}
2042
2043static void mptcp_timeout_timer(struct timer_list *t)
2044{
2045 struct sock *sk = from_timer(sk, t, sk_timer);
2046
2047 mptcp_schedule_work(sk);
2048 sock_put(sk);
2049}
2050
2051/* Find an idle subflow. Return NULL if there is unacked data at tcp
2052 * level.
2053 *
2054 * A backup subflow is returned only if that is the only kind available.
2055 */
2056static struct sock *mptcp_subflow_get_retrans(const struct mptcp_sock *msk)
2057{
2058 struct mptcp_subflow_context *subflow;
2059 struct sock *backup = NULL;
2060
2061 sock_owned_by_me((const struct sock *)msk);
2062
2063 if (__mptcp_check_fallback(msk))
2064 return NULL;
2065
2066 mptcp_for_each_subflow(msk, subflow) {
2067 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2068
2069 if (!mptcp_subflow_active(subflow))
2070 continue;
2071
2072 /* still data outstanding at TCP level? Don't retransmit. */
2073 if (!tcp_write_queue_empty(ssk)) {
2074 if (inet_csk(ssk)->icsk_ca_state >= TCP_CA_Loss)
2075 continue;
2076 return NULL;
2077 }
2078
2079 if (subflow->backup) {
2080 if (!backup)
2081 backup = ssk;
2082 continue;
2083 }
2084
2085 return ssk;
2086 }
2087
2088 return backup;
2089}
2090
2091static void mptcp_dispose_initial_subflow(struct mptcp_sock *msk)
2092{
2093 if (msk->subflow) {
2094 iput(SOCK_INODE(msk->subflow));
2095 msk->subflow = NULL;
2096 }
2097}
2098
2099/* subflow sockets can be either outgoing (connect) or incoming
2100 * (accept).
2101 *
2102 * Outgoing subflows use in-kernel sockets.
2103 * Incoming subflows do not have their own 'struct socket' allocated,
2104 * so we need to use tcp_close() after detaching them from the mptcp
2105 * parent socket.
2106 */
2107static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2108 struct mptcp_subflow_context *subflow)
2109{
2110 struct mptcp_sock *msk = mptcp_sk(sk);
2111
2112 list_del(&subflow->node);
2113
2114 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2115
2116 /* if we are invoked by the msk cleanup code, the subflow is
2117 * already orphaned
2118 */
2119 if (ssk->sk_socket)
2120 sock_orphan(ssk);
2121
2122 subflow->disposable = 1;
2123
2124 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2125 * the ssk has been already destroyed, we just need to release the
2126 * reference owned by msk;
2127 */
2128 if (!inet_csk(ssk)->icsk_ulp_ops) {
2129 kfree_rcu(subflow, rcu);
2130 } else {
2131 /* otherwise tcp will dispose of the ssk and subflow ctx */
2132 __tcp_close(ssk, 0);
2133
2134 /* close acquired an extra ref */
2135 __sock_put(ssk);
2136 }
2137 release_sock(ssk);
2138
2139 sock_put(ssk);
2140
2141 if (ssk == msk->last_snd)
2142 msk->last_snd = NULL;
2143
2144 if (ssk == msk->first)
2145 msk->first = NULL;
2146
2147 if (msk->subflow && ssk == msk->subflow->sk)
2148 mptcp_dispose_initial_subflow(msk);
2149}
2150
2151void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2152 struct mptcp_subflow_context *subflow)
2153{
2154 if (sk->sk_state == TCP_ESTABLISHED)
2155 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2156 __mptcp_close_ssk(sk, ssk, subflow);
2157}
2158
2159static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2160{
2161 return 0;
2162}
2163
2164static void __mptcp_close_subflow(struct mptcp_sock *msk)
2165{
2166 struct mptcp_subflow_context *subflow, *tmp;
2167
2168 might_sleep();
2169
2170 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2171 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2172
2173 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2174 continue;
2175
2176 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2177 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2178 continue;
2179
2180 mptcp_close_ssk((struct sock *)msk, ssk, subflow);
2181 }
2182}
2183
2184static bool mptcp_check_close_timeout(const struct sock *sk)
2185{
2186 s32 delta = tcp_jiffies32 - inet_csk(sk)->icsk_mtup.probe_timestamp;
2187 struct mptcp_subflow_context *subflow;
2188
2189 if (delta >= TCP_TIMEWAIT_LEN)
2190 return true;
2191
2192 /* if all subflows are in closed status don't bother with additional
2193 * timeout
2194 */
2195 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2196 if (inet_sk_state_load(mptcp_subflow_tcp_sock(subflow)) !=
2197 TCP_CLOSE)
2198 return false;
2199 }
2200 return true;
2201}
2202
2203static void mptcp_check_fastclose(struct mptcp_sock *msk)
2204{
2205 struct mptcp_subflow_context *subflow, *tmp;
2206 struct sock *sk = &msk->sk.icsk_inet.sk;
2207
2208 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2209 return;
2210
2211 mptcp_token_destroy(msk);
2212
2213 list_for_each_entry_safe(subflow, tmp, &msk->conn_list, node) {
2214 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2215 bool slow;
2216
2217 slow = lock_sock_fast(tcp_sk);
2218 if (tcp_sk->sk_state != TCP_CLOSE) {
2219 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2220 tcp_set_state(tcp_sk, TCP_CLOSE);
2221 }
2222 unlock_sock_fast(tcp_sk, slow);
2223 }
2224
2225 inet_sk_state_store(sk, TCP_CLOSE);
2226 sk->sk_shutdown = SHUTDOWN_MASK;
2227 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2228 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2229
2230 mptcp_close_wake_up(sk);
2231}
2232
2233static void __mptcp_retrans(struct sock *sk)
2234{
2235 struct mptcp_sock *msk = mptcp_sk(sk);
2236 struct mptcp_sendmsg_info info = {};
2237 struct mptcp_data_frag *dfrag;
2238 size_t copied = 0;
2239 struct sock *ssk;
2240 int ret;
2241
2242 mptcp_clean_una_wakeup(sk);
2243 dfrag = mptcp_rtx_head(sk);
2244 if (!dfrag) {
2245 if (mptcp_data_fin_enabled(msk)) {
2246 struct inet_connection_sock *icsk = inet_csk(sk);
2247
2248 icsk->icsk_retransmits++;
2249 mptcp_set_datafin_timeout(sk);
2250 mptcp_send_ack(msk);
2251
2252 goto reset_timer;
2253 }
2254
2255 return;
2256 }
2257
2258 ssk = mptcp_subflow_get_retrans(msk);
2259 if (!ssk)
2260 goto reset_timer;
2261
2262 lock_sock(ssk);
2263
2264 /* limit retransmission to the bytes already sent on some subflows */
2265 info.sent = 0;
2266 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len : dfrag->already_sent;
2267 while (info.sent < info.limit) {
2268 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2269 if (ret <= 0)
2270 break;
2271
2272 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2273 copied += ret;
2274 info.sent += ret;
2275 }
2276 if (copied) {
2277 dfrag->already_sent = max(dfrag->already_sent, info.sent);
2278 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2279 info.size_goal);
2280 }
2281
2282 mptcp_set_timeout(sk, ssk);
2283 release_sock(ssk);
2284
2285reset_timer:
2286 if (!mptcp_timer_pending(sk))
2287 mptcp_reset_timer(sk);
2288}
2289
2290static void mptcp_worker(struct work_struct *work)
2291{
2292 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2293 struct sock *sk = &msk->sk.icsk_inet.sk;
2294 int state;
2295
2296 lock_sock(sk);
2297 state = sk->sk_state;
2298 if (unlikely(state == TCP_CLOSE))
2299 goto unlock;
2300
2301 mptcp_check_data_fin_ack(sk);
2302 mptcp_flush_join_list(msk);
2303
2304 mptcp_check_fastclose(msk);
2305
2306 if (msk->pm.status)
2307 mptcp_pm_nl_work(msk);
2308
2309 if (test_and_clear_bit(MPTCP_WORK_EOF, &msk->flags))
2310 mptcp_check_for_eof(msk);
2311
2312 __mptcp_check_send_data_fin(sk);
2313 mptcp_check_data_fin(sk);
2314
2315 /* There is no point in keeping around an orphaned sk timedout or
2316 * closed, but we need the msk around to reply to incoming DATA_FIN,
2317 * even if it is orphaned and in FIN_WAIT2 state
2318 */
2319 if (sock_flag(sk, SOCK_DEAD) &&
2320 (mptcp_check_close_timeout(sk) || sk->sk_state == TCP_CLOSE)) {
2321 inet_sk_state_store(sk, TCP_CLOSE);
2322 __mptcp_destroy_sock(sk);
2323 goto unlock;
2324 }
2325
2326 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2327 __mptcp_close_subflow(msk);
2328
2329 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2330 __mptcp_retrans(sk);
2331
2332unlock:
2333 release_sock(sk);
2334 sock_put(sk);
2335}
2336
2337static int __mptcp_init_sock(struct sock *sk)
2338{
2339 struct mptcp_sock *msk = mptcp_sk(sk);
2340
2341 spin_lock_init(&msk->join_list_lock);
2342
2343 INIT_LIST_HEAD(&msk->conn_list);
2344 INIT_LIST_HEAD(&msk->join_list);
2345 INIT_LIST_HEAD(&msk->rtx_queue);
2346 INIT_WORK(&msk->work, mptcp_worker);
2347 __skb_queue_head_init(&msk->receive_queue);
2348 msk->out_of_order_queue = RB_ROOT;
2349 msk->first_pending = NULL;
2350 msk->wmem_reserved = 0;
2351 WRITE_ONCE(msk->rmem_released, 0);
2352 msk->tx_pending_data = 0;
2353
2354 msk->first = NULL;
2355 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2356 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2357
2358 mptcp_pm_data_init(msk);
2359
2360 /* re-use the csk retrans timer for MPTCP-level retrans */
2361 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2362 timer_setup(&sk->sk_timer, mptcp_timeout_timer, 0);
2363
2364 return 0;
2365}
2366
2367static int mptcp_init_sock(struct sock *sk)
2368{
2369 struct inet_connection_sock *icsk = inet_csk(sk);
2370 struct net *net = sock_net(sk);
2371 int ret;
2372
2373 ret = __mptcp_init_sock(sk);
2374 if (ret)
2375 return ret;
2376
2377 if (!mptcp_is_enabled(net))
2378 return -ENOPROTOOPT;
2379
2380 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2381 return -ENOMEM;
2382
2383 ret = __mptcp_socket_create(mptcp_sk(sk));
2384 if (ret)
2385 return ret;
2386
2387 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2388 * propagate the correct value
2389 */
2390 tcp_assign_congestion_control(sk);
2391 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2392
2393 /* no need to keep a reference to the ops, the name will suffice */
2394 tcp_cleanup_congestion_control(sk);
2395 icsk->icsk_ca_ops = NULL;
2396
2397 sk_sockets_allocated_inc(sk);
2398 sk->sk_rcvbuf = sock_net(sk)->ipv4.sysctl_tcp_rmem[1];
2399 sk->sk_sndbuf = sock_net(sk)->ipv4.sysctl_tcp_wmem[1];
2400
2401 return 0;
2402}
2403
2404static void __mptcp_clear_xmit(struct sock *sk)
2405{
2406 struct mptcp_sock *msk = mptcp_sk(sk);
2407 struct mptcp_data_frag *dtmp, *dfrag;
2408
2409 WRITE_ONCE(msk->first_pending, NULL);
2410 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2411 dfrag_clear(sk, dfrag);
2412}
2413
2414static void mptcp_cancel_work(struct sock *sk)
2415{
2416 struct mptcp_sock *msk = mptcp_sk(sk);
2417
2418 if (cancel_work_sync(&msk->work))
2419 __sock_put(sk);
2420}
2421
2422void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2423{
2424 lock_sock(ssk);
2425
2426 switch (ssk->sk_state) {
2427 case TCP_LISTEN:
2428 if (!(how & RCV_SHUTDOWN))
2429 break;
2430 fallthrough;
2431 case TCP_SYN_SENT:
2432 tcp_disconnect(ssk, O_NONBLOCK);
2433 break;
2434 default:
2435 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2436 pr_debug("Fallback");
2437 ssk->sk_shutdown |= how;
2438 tcp_shutdown(ssk, how);
2439 } else {
2440 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2441 mptcp_set_timeout(sk, ssk);
2442 tcp_send_ack(ssk);
2443 if (!mptcp_timer_pending(sk))
2444 mptcp_reset_timer(sk);
2445 }
2446 break;
2447 }
2448
2449 release_sock(ssk);
2450}
2451
2452static const unsigned char new_state[16] = {
2453 /* current state: new state: action: */
2454 [0 /* (Invalid) */] = TCP_CLOSE,
2455 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2456 [TCP_SYN_SENT] = TCP_CLOSE,
2457 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2458 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2459 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2460 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2461 [TCP_CLOSE] = TCP_CLOSE,
2462 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2463 [TCP_LAST_ACK] = TCP_LAST_ACK,
2464 [TCP_LISTEN] = TCP_CLOSE,
2465 [TCP_CLOSING] = TCP_CLOSING,
2466 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2467};
2468
2469static int mptcp_close_state(struct sock *sk)
2470{
2471 int next = (int)new_state[sk->sk_state];
2472 int ns = next & TCP_STATE_MASK;
2473
2474 inet_sk_state_store(sk, ns);
2475
2476 return next & TCP_ACTION_FIN;
2477}
2478
2479static void __mptcp_check_send_data_fin(struct sock *sk)
2480{
2481 struct mptcp_subflow_context *subflow;
2482 struct mptcp_sock *msk = mptcp_sk(sk);
2483
2484 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2485 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2486 msk->snd_nxt, msk->write_seq);
2487
2488 /* we still need to enqueue subflows or not really shutting down,
2489 * skip this
2490 */
2491 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2492 mptcp_send_head(sk))
2493 return;
2494
2495 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2496
2497 /* fallback socket will not get data_fin/ack, can move to the next
2498 * state now
2499 */
2500 if (__mptcp_check_fallback(msk)) {
2501 if ((1 << sk->sk_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
2502 inet_sk_state_store(sk, TCP_CLOSE);
2503 mptcp_close_wake_up(sk);
2504 } else if (sk->sk_state == TCP_FIN_WAIT1) {
2505 inet_sk_state_store(sk, TCP_FIN_WAIT2);
2506 }
2507 }
2508
2509 mptcp_flush_join_list(msk);
2510 mptcp_for_each_subflow(msk, subflow) {
2511 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2512
2513 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2514 }
2515}
2516
2517static void __mptcp_wr_shutdown(struct sock *sk)
2518{
2519 struct mptcp_sock *msk = mptcp_sk(sk);
2520
2521 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2522 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2523 !!mptcp_send_head(sk));
2524
2525 /* will be ignored by fallback sockets */
2526 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2527 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2528
2529 __mptcp_check_send_data_fin(sk);
2530}
2531
2532static void __mptcp_destroy_sock(struct sock *sk)
2533{
2534 struct mptcp_subflow_context *subflow, *tmp;
2535 struct mptcp_sock *msk = mptcp_sk(sk);
2536 LIST_HEAD(conn_list);
2537
2538 pr_debug("msk=%p", msk);
2539
2540 might_sleep();
2541
2542 /* be sure to always acquire the join list lock, to sync vs
2543 * mptcp_finish_join().
2544 */
2545 spin_lock_bh(&msk->join_list_lock);
2546 list_splice_tail_init(&msk->join_list, &msk->conn_list);
2547 spin_unlock_bh(&msk->join_list_lock);
2548 list_splice_init(&msk->conn_list, &conn_list);
2549
2550 sk_stop_timer(sk, &msk->sk.icsk_retransmit_timer);
2551 sk_stop_timer(sk, &sk->sk_timer);
2552 msk->pm.status = 0;
2553
2554 list_for_each_entry_safe(subflow, tmp, &conn_list, node) {
2555 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2556 __mptcp_close_ssk(sk, ssk, subflow);
2557 }
2558
2559 sk->sk_prot->destroy(sk);
2560
2561 WARN_ON_ONCE(msk->wmem_reserved);
2562 WARN_ON_ONCE(msk->rmem_released);
2563 sk_stream_kill_queues(sk);
2564 xfrm_sk_free_policy(sk);
2565
2566 sk_refcnt_debug_release(sk);
2567 mptcp_dispose_initial_subflow(msk);
2568 sock_put(sk);
2569}
2570
2571static void mptcp_close(struct sock *sk, long timeout)
2572{
2573 struct mptcp_subflow_context *subflow;
2574 bool do_cancel_work = false;
2575
2576 lock_sock(sk);
2577 sk->sk_shutdown = SHUTDOWN_MASK;
2578
2579 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
2580 inet_sk_state_store(sk, TCP_CLOSE);
2581 goto cleanup;
2582 }
2583
2584 if (mptcp_close_state(sk))
2585 __mptcp_wr_shutdown(sk);
2586
2587 sk_stream_wait_close(sk, timeout);
2588
2589cleanup:
2590 /* orphan all the subflows */
2591 inet_csk(sk)->icsk_mtup.probe_timestamp = tcp_jiffies32;
2592 mptcp_for_each_subflow(mptcp_sk(sk), subflow) {
2593 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2594 bool slow = lock_sock_fast(ssk);
2595
2596 sock_orphan(ssk);
2597 unlock_sock_fast(ssk, slow);
2598 }
2599 sock_orphan(sk);
2600
2601 sock_hold(sk);
2602 pr_debug("msk=%p state=%d", sk, sk->sk_state);
2603 if (sk->sk_state == TCP_CLOSE) {
2604 __mptcp_destroy_sock(sk);
2605 do_cancel_work = true;
2606 } else {
2607 sk_reset_timer(sk, &sk->sk_timer, jiffies + TCP_TIMEWAIT_LEN);
2608 }
2609 release_sock(sk);
2610 if (do_cancel_work)
2611 mptcp_cancel_work(sk);
2612
2613 if (mptcp_sk(sk)->token)
2614 mptcp_event(MPTCP_EVENT_CLOSED, mptcp_sk(sk), NULL, GFP_KERNEL);
2615
2616 sock_put(sk);
2617}
2618
2619static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
2620{
2621#if IS_ENABLED(CONFIG_MPTCP_IPV6)
2622 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
2623 struct ipv6_pinfo *msk6 = inet6_sk(msk);
2624
2625 msk->sk_v6_daddr = ssk->sk_v6_daddr;
2626 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
2627
2628 if (msk6 && ssk6) {
2629 msk6->saddr = ssk6->saddr;
2630 msk6->flow_label = ssk6->flow_label;
2631 }
2632#endif
2633
2634 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
2635 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
2636 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
2637 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
2638 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
2639 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
2640}
2641
2642static int mptcp_disconnect(struct sock *sk, int flags)
2643{
2644 struct mptcp_subflow_context *subflow;
2645 struct mptcp_sock *msk = mptcp_sk(sk);
2646
2647 mptcp_do_flush_join_list(msk);
2648
2649 mptcp_for_each_subflow(msk, subflow) {
2650 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2651
2652 lock_sock(ssk);
2653 tcp_disconnect(ssk, flags);
2654 release_sock(ssk);
2655 }
2656 return 0;
2657}
2658
2659#if IS_ENABLED(CONFIG_MPTCP_IPV6)
2660static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
2661{
2662 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
2663
2664 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
2665}
2666#endif
2667
2668struct sock *mptcp_sk_clone(const struct sock *sk,
2669 const struct mptcp_options_received *mp_opt,
2670 struct request_sock *req)
2671{
2672 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
2673 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
2674 struct mptcp_sock *msk;
2675 u64 ack_seq;
2676
2677 if (!nsk)
2678 return NULL;
2679
2680#if IS_ENABLED(CONFIG_MPTCP_IPV6)
2681 if (nsk->sk_family == AF_INET6)
2682 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
2683#endif
2684
2685 __mptcp_init_sock(nsk);
2686
2687 msk = mptcp_sk(nsk);
2688 msk->local_key = subflow_req->local_key;
2689 msk->token = subflow_req->token;
2690 msk->subflow = NULL;
2691 WRITE_ONCE(msk->fully_established, false);
2692 if (mp_opt->csum_reqd)
2693 WRITE_ONCE(msk->csum_enabled, true);
2694
2695 msk->write_seq = subflow_req->idsn + 1;
2696 msk->snd_nxt = msk->write_seq;
2697 msk->snd_una = msk->write_seq;
2698 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
2699 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
2700
2701 if (mp_opt->mp_capable) {
2702 msk->can_ack = true;
2703 msk->remote_key = mp_opt->sndr_key;
2704 mptcp_crypto_key_sha(msk->remote_key, NULL, &ack_seq);
2705 ack_seq++;
2706 WRITE_ONCE(msk->ack_seq, ack_seq);
2707 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2708 }
2709
2710 sock_reset_flag(nsk, SOCK_RCU_FREE);
2711 /* will be fully established after successful MPC subflow creation */
2712 inet_sk_state_store(nsk, TCP_SYN_RECV);
2713
2714 security_inet_csk_clone(nsk, req);
2715 bh_unlock_sock(nsk);
2716
2717 /* keep a single reference */
2718 __sock_put(nsk);
2719 return nsk;
2720}
2721
2722void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
2723{
2724 const struct tcp_sock *tp = tcp_sk(ssk);
2725
2726 msk->rcvq_space.copied = 0;
2727 msk->rcvq_space.rtt_us = 0;
2728
2729 msk->rcvq_space.time = tp->tcp_mstamp;
2730
2731 /* initial rcv_space offering made to peer */
2732 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
2733 TCP_INIT_CWND * tp->advmss);
2734 if (msk->rcvq_space.space == 0)
2735 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
2736
2737 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
2738}
2739
2740static struct sock *mptcp_accept(struct sock *sk, int flags, int *err,
2741 bool kern)
2742{
2743 struct mptcp_sock *msk = mptcp_sk(sk);
2744 struct socket *listener;
2745 struct sock *newsk;
2746
2747 listener = __mptcp_nmpc_socket(msk);
2748 if (WARN_ON_ONCE(!listener)) {
2749 *err = -EINVAL;
2750 return NULL;
2751 }
2752
2753 pr_debug("msk=%p, listener=%p", msk, mptcp_subflow_ctx(listener->sk));
2754 newsk = inet_csk_accept(listener->sk, flags, err, kern);
2755 if (!newsk)
2756 return NULL;
2757
2758 pr_debug("msk=%p, subflow is mptcp=%d", msk, sk_is_mptcp(newsk));
2759 if (sk_is_mptcp(newsk)) {
2760 struct mptcp_subflow_context *subflow;
2761 struct sock *new_mptcp_sock;
2762
2763 subflow = mptcp_subflow_ctx(newsk);
2764 new_mptcp_sock = subflow->conn;
2765
2766 /* is_mptcp should be false if subflow->conn is missing, see
2767 * subflow_syn_recv_sock()
2768 */
2769 if (WARN_ON_ONCE(!new_mptcp_sock)) {
2770 tcp_sk(newsk)->is_mptcp = 0;
2771 return newsk;
2772 }
2773
2774 /* acquire the 2nd reference for the owning socket */
2775 sock_hold(new_mptcp_sock);
2776 newsk = new_mptcp_sock;
2777 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
2778 } else {
2779 MPTCP_INC_STATS(sock_net(sk),
2780 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
2781 }
2782
2783 return newsk;
2784}
2785
2786void mptcp_destroy_common(struct mptcp_sock *msk)
2787{
2788 struct sock *sk = (struct sock *)msk;
2789
2790 __mptcp_clear_xmit(sk);
2791
2792 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
2793 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
2794
2795 skb_rbtree_purge(&msk->out_of_order_queue);
2796 mptcp_token_destroy(msk);
2797 mptcp_pm_free_anno_list(msk);
2798}
2799
2800static void mptcp_destroy(struct sock *sk)
2801{
2802 struct mptcp_sock *msk = mptcp_sk(sk);
2803
2804 mptcp_destroy_common(msk);
2805 sk_sockets_allocated_dec(sk);
2806}
2807
2808void __mptcp_data_acked(struct sock *sk)
2809{
2810 if (!sock_owned_by_user(sk))
2811 __mptcp_clean_una(sk);
2812 else
2813 set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags);
2814
2815 if (mptcp_pending_data_fin_ack(sk))
2816 mptcp_schedule_work(sk);
2817}
2818
2819void __mptcp_check_push(struct sock *sk, struct sock *ssk)
2820{
2821 if (!mptcp_send_head(sk))
2822 return;
2823
2824 if (!sock_owned_by_user(sk)) {
2825 struct sock *xmit_ssk = mptcp_subflow_get_send(mptcp_sk(sk));
2826
2827 if (xmit_ssk == ssk)
2828 __mptcp_subflow_push_pending(sk, ssk);
2829 else if (xmit_ssk)
2830 mptcp_subflow_delegate(mptcp_subflow_ctx(xmit_ssk));
2831 } else {
2832 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2833 }
2834}
2835
2836/* processes deferred events and flush wmem */
2837static void mptcp_release_cb(struct sock *sk)
2838{
2839 for (;;) {
2840 unsigned long flags = 0;
2841
2842 if (test_and_clear_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags))
2843 flags |= BIT(MPTCP_PUSH_PENDING);
2844 if (test_and_clear_bit(MPTCP_RETRANSMIT, &mptcp_sk(sk)->flags))
2845 flags |= BIT(MPTCP_RETRANSMIT);
2846 if (!flags)
2847 break;
2848
2849 /* the following actions acquire the subflow socket lock
2850 *
2851 * 1) can't be invoked in atomic scope
2852 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
2853 * datapath acquires the msk socket spinlock while helding
2854 * the subflow socket lock
2855 */
2856
2857 spin_unlock_bh(&sk->sk_lock.slock);
2858 if (flags & BIT(MPTCP_PUSH_PENDING))
2859 __mptcp_push_pending(sk, 0);
2860 if (flags & BIT(MPTCP_RETRANSMIT))
2861 __mptcp_retrans(sk);
2862
2863 cond_resched();
2864 spin_lock_bh(&sk->sk_lock.slock);
2865 }
2866
2867 /* be sure to set the current sk state before tacking actions
2868 * depending on sk_state
2869 */
2870 if (test_and_clear_bit(MPTCP_CONNECTED, &mptcp_sk(sk)->flags))
2871 __mptcp_set_connected(sk);
2872 if (test_and_clear_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->flags))
2873 __mptcp_clean_una_wakeup(sk);
2874 if (test_and_clear_bit(MPTCP_ERROR_REPORT, &mptcp_sk(sk)->flags))
2875 __mptcp_error_report(sk);
2876
2877 /* push_pending may touch wmem_reserved, ensure we do the cleanup
2878 * later
2879 */
2880 __mptcp_update_wmem(sk);
2881 __mptcp_update_rmem(sk);
2882}
2883
2884void mptcp_subflow_process_delegated(struct sock *ssk)
2885{
2886 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2887 struct sock *sk = subflow->conn;
2888
2889 mptcp_data_lock(sk);
2890 if (!sock_owned_by_user(sk))
2891 __mptcp_subflow_push_pending(sk, ssk);
2892 else
2893 set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->flags);
2894 mptcp_data_unlock(sk);
2895 mptcp_subflow_delegated_done(subflow);
2896}
2897
2898static int mptcp_hash(struct sock *sk)
2899{
2900 /* should never be called,
2901 * we hash the TCP subflows not the master socket
2902 */
2903 WARN_ON_ONCE(1);
2904 return 0;
2905}
2906
2907static void mptcp_unhash(struct sock *sk)
2908{
2909 /* called from sk_common_release(), but nothing to do here */
2910}
2911
2912static int mptcp_get_port(struct sock *sk, unsigned short snum)
2913{
2914 struct mptcp_sock *msk = mptcp_sk(sk);
2915 struct socket *ssock;
2916
2917 ssock = __mptcp_nmpc_socket(msk);
2918 pr_debug("msk=%p, subflow=%p", msk, ssock);
2919 if (WARN_ON_ONCE(!ssock))
2920 return -EINVAL;
2921
2922 return inet_csk_get_port(ssock->sk, snum);
2923}
2924
2925void mptcp_finish_connect(struct sock *ssk)
2926{
2927 struct mptcp_subflow_context *subflow;
2928 struct mptcp_sock *msk;
2929 struct sock *sk;
2930 u64 ack_seq;
2931
2932 subflow = mptcp_subflow_ctx(ssk);
2933 sk = subflow->conn;
2934 msk = mptcp_sk(sk);
2935
2936 pr_debug("msk=%p, token=%u", sk, subflow->token);
2937
2938 mptcp_crypto_key_sha(subflow->remote_key, NULL, &ack_seq);
2939 ack_seq++;
2940 subflow->map_seq = ack_seq;
2941 subflow->map_subflow_seq = 1;
2942
2943 /* the socket is not connected yet, no msk/subflow ops can access/race
2944 * accessing the field below
2945 */
2946 WRITE_ONCE(msk->remote_key, subflow->remote_key);
2947 WRITE_ONCE(msk->local_key, subflow->local_key);
2948 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
2949 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2950 WRITE_ONCE(msk->ack_seq, ack_seq);
2951 WRITE_ONCE(msk->rcv_wnd_sent, ack_seq);
2952 WRITE_ONCE(msk->can_ack, 1);
2953 WRITE_ONCE(msk->snd_una, msk->write_seq);
2954
2955 mptcp_pm_new_connection(msk, ssk, 0);
2956
2957 mptcp_rcv_space_init(msk, ssk);
2958}
2959
2960void mptcp_sock_graft(struct sock *sk, struct socket *parent)
2961{
2962 write_lock_bh(&sk->sk_callback_lock);
2963 rcu_assign_pointer(sk->sk_wq, &parent->wq);
2964 sk_set_socket(sk, parent);
2965 sk->sk_uid = SOCK_INODE(parent)->i_uid;
2966 write_unlock_bh(&sk->sk_callback_lock);
2967}
2968
2969bool mptcp_finish_join(struct sock *ssk)
2970{
2971 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
2972 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
2973 struct sock *parent = (void *)msk;
2974 struct socket *parent_sock;
2975 bool ret;
2976
2977 pr_debug("msk=%p, subflow=%p", msk, subflow);
2978
2979 /* mptcp socket already closing? */
2980 if (!mptcp_is_fully_established(parent)) {
2981 subflow->reset_reason = MPTCP_RST_EMPTCP;
2982 return false;
2983 }
2984
2985 if (!msk->pm.server_side)
2986 goto out;
2987
2988 if (!mptcp_pm_allow_new_subflow(msk)) {
2989 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
2990 return false;
2991 }
2992
2993 /* active connections are already on conn_list, and we can't acquire
2994 * msk lock here.
2995 * use the join list lock as synchronization point and double-check
2996 * msk status to avoid racing with __mptcp_destroy_sock()
2997 */
2998 spin_lock_bh(&msk->join_list_lock);
2999 ret = inet_sk_state_load(parent) == TCP_ESTABLISHED;
3000 if (ret && !WARN_ON_ONCE(!list_empty(&subflow->node))) {
3001 list_add_tail(&subflow->node, &msk->join_list);
3002 sock_hold(ssk);
3003 }
3004 spin_unlock_bh(&msk->join_list_lock);
3005 if (!ret) {
3006 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3007 return false;
3008 }
3009
3010 /* attach to msk socket only after we are sure he will deal with us
3011 * at close time
3012 */
3013 parent_sock = READ_ONCE(parent->sk_socket);
3014 if (parent_sock && !ssk->sk_socket)
3015 mptcp_sock_graft(ssk, parent_sock);
3016 subflow->map_seq = READ_ONCE(msk->ack_seq);
3017out:
3018 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
3019 return true;
3020}
3021
3022static void mptcp_shutdown(struct sock *sk, int how)
3023{
3024 pr_debug("sk=%p, how=%d", sk, how);
3025
3026 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3027 __mptcp_wr_shutdown(sk);
3028}
3029
3030static struct proto mptcp_prot = {
3031 .name = "MPTCP",
3032 .owner = THIS_MODULE,
3033 .init = mptcp_init_sock,
3034 .disconnect = mptcp_disconnect,
3035 .close = mptcp_close,
3036 .accept = mptcp_accept,
3037 .setsockopt = mptcp_setsockopt,
3038 .getsockopt = mptcp_getsockopt,
3039 .shutdown = mptcp_shutdown,
3040 .destroy = mptcp_destroy,
3041 .sendmsg = mptcp_sendmsg,
3042 .recvmsg = mptcp_recvmsg,
3043 .release_cb = mptcp_release_cb,
3044 .hash = mptcp_hash,
3045 .unhash = mptcp_unhash,
3046 .get_port = mptcp_get_port,
3047 .sockets_allocated = &mptcp_sockets_allocated,
3048 .memory_allocated = &tcp_memory_allocated,
3049 .memory_pressure = &tcp_memory_pressure,
3050 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3051 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3052 .sysctl_mem = sysctl_tcp_mem,
3053 .obj_size = sizeof(struct mptcp_sock),
3054 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3055 .no_autobind = true,
3056};
3057
3058static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3059{
3060 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3061 struct socket *ssock;
3062 int err;
3063
3064 lock_sock(sock->sk);
3065 ssock = __mptcp_nmpc_socket(msk);
3066 if (!ssock) {
3067 err = -EINVAL;
3068 goto unlock;
3069 }
3070
3071 err = ssock->ops->bind(ssock, uaddr, addr_len);
3072 if (!err)
3073 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3074
3075unlock:
3076 release_sock(sock->sk);
3077 return err;
3078}
3079
3080static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3081 struct mptcp_subflow_context *subflow)
3082{
3083 subflow->request_mptcp = 0;
3084 __mptcp_do_fallback(msk);
3085}
3086
3087static int mptcp_stream_connect(struct socket *sock, struct sockaddr *uaddr,
3088 int addr_len, int flags)
3089{
3090 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3091 struct mptcp_subflow_context *subflow;
3092 struct socket *ssock;
3093 int err;
3094
3095 lock_sock(sock->sk);
3096 if (sock->state != SS_UNCONNECTED && msk->subflow) {
3097 /* pending connection or invalid state, let existing subflow
3098 * cope with that
3099 */
3100 ssock = msk->subflow;
3101 goto do_connect;
3102 }
3103
3104 ssock = __mptcp_nmpc_socket(msk);
3105 if (!ssock) {
3106 err = -EINVAL;
3107 goto unlock;
3108 }
3109
3110 mptcp_token_destroy(msk);
3111 inet_sk_state_store(sock->sk, TCP_SYN_SENT);
3112 subflow = mptcp_subflow_ctx(ssock->sk);
3113#ifdef CONFIG_TCP_MD5SIG
3114 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3115 * TCP option space.
3116 */
3117 if (rcu_access_pointer(tcp_sk(ssock->sk)->md5sig_info))
3118 mptcp_subflow_early_fallback(msk, subflow);
3119#endif
3120 if (subflow->request_mptcp && mptcp_token_new_connect(ssock->sk)) {
3121 MPTCP_INC_STATS(sock_net(ssock->sk), MPTCP_MIB_TOKENFALLBACKINIT);
3122 mptcp_subflow_early_fallback(msk, subflow);
3123 }
3124 if (likely(!__mptcp_check_fallback(msk)))
3125 MPTCP_INC_STATS(sock_net(sock->sk), MPTCP_MIB_MPCAPABLEACTIVE);
3126
3127do_connect:
3128 err = ssock->ops->connect(ssock, uaddr, addr_len, flags);
3129 sock->state = ssock->state;
3130
3131 /* on successful connect, the msk state will be moved to established by
3132 * subflow_finish_connect()
3133 */
3134 if (!err || err == -EINPROGRESS)
3135 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3136 else
3137 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3138
3139unlock:
3140 release_sock(sock->sk);
3141 return err;
3142}
3143
3144static int mptcp_listen(struct socket *sock, int backlog)
3145{
3146 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3147 struct socket *ssock;
3148 int err;
3149
3150 pr_debug("msk=%p", msk);
3151
3152 lock_sock(sock->sk);
3153 ssock = __mptcp_nmpc_socket(msk);
3154 if (!ssock) {
3155 err = -EINVAL;
3156 goto unlock;
3157 }
3158
3159 mptcp_token_destroy(msk);
3160 inet_sk_state_store(sock->sk, TCP_LISTEN);
3161 sock_set_flag(sock->sk, SOCK_RCU_FREE);
3162
3163 err = ssock->ops->listen(ssock, backlog);
3164 inet_sk_state_store(sock->sk, inet_sk_state_load(ssock->sk));
3165 if (!err)
3166 mptcp_copy_inaddrs(sock->sk, ssock->sk);
3167
3168unlock:
3169 release_sock(sock->sk);
3170 return err;
3171}
3172
3173static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3174 int flags, bool kern)
3175{
3176 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3177 struct socket *ssock;
3178 int err;
3179
3180 pr_debug("msk=%p", msk);
3181
3182 lock_sock(sock->sk);
3183 if (sock->sk->sk_state != TCP_LISTEN)
3184 goto unlock_fail;
3185
3186 ssock = __mptcp_nmpc_socket(msk);
3187 if (!ssock)
3188 goto unlock_fail;
3189
3190 clear_bit(MPTCP_DATA_READY, &msk->flags);
3191 sock_hold(ssock->sk);
3192 release_sock(sock->sk);
3193
3194 err = ssock->ops->accept(sock, newsock, flags, kern);
3195 if (err == 0 && !mptcp_is_tcpsk(newsock->sk)) {
3196 struct mptcp_sock *msk = mptcp_sk(newsock->sk);
3197 struct mptcp_subflow_context *subflow;
3198 struct sock *newsk = newsock->sk;
3199
3200 lock_sock(newsk);
3201
3202 /* PM/worker can now acquire the first subflow socket
3203 * lock without racing with listener queue cleanup,
3204 * we can notify it, if needed.
3205 *
3206 * Even if remote has reset the initial subflow by now
3207 * the refcnt is still at least one.
3208 */
3209 subflow = mptcp_subflow_ctx(msk->first);
3210 list_add(&subflow->node, &msk->conn_list);
3211 sock_hold(msk->first);
3212 if (mptcp_is_fully_established(newsk))
3213 mptcp_pm_fully_established(msk, msk->first, GFP_KERNEL);
3214
3215 mptcp_copy_inaddrs(newsk, msk->first);
3216 mptcp_rcv_space_init(msk, msk->first);
3217 mptcp_propagate_sndbuf(newsk, msk->first);
3218
3219 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3220 * This is needed so NOSPACE flag can be set from tcp stack.
3221 */
3222 mptcp_flush_join_list(msk);
3223 mptcp_for_each_subflow(msk, subflow) {
3224 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3225
3226 if (!ssk->sk_socket)
3227 mptcp_sock_graft(ssk, newsock);
3228 }
3229 release_sock(newsk);
3230 }
3231
3232 if (inet_csk_listen_poll(ssock->sk))
3233 set_bit(MPTCP_DATA_READY, &msk->flags);
3234 sock_put(ssock->sk);
3235 return err;
3236
3237unlock_fail:
3238 release_sock(sock->sk);
3239 return -EINVAL;
3240}
3241
3242static __poll_t mptcp_check_readable(struct mptcp_sock *msk)
3243{
3244 /* Concurrent splices from sk_receive_queue into receive_queue will
3245 * always show at least one non-empty queue when checked in this order.
3246 */
3247 if (skb_queue_empty_lockless(&((struct sock *)msk)->sk_receive_queue) &&
3248 skb_queue_empty_lockless(&msk->receive_queue))
3249 return 0;
3250
3251 return EPOLLIN | EPOLLRDNORM;
3252}
3253
3254static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3255{
3256 struct sock *sk = (struct sock *)msk;
3257
3258 if (unlikely(sk->sk_shutdown & SEND_SHUTDOWN))
3259 return EPOLLOUT | EPOLLWRNORM;
3260
3261 if (sk_stream_is_writeable(sk))
3262 return EPOLLOUT | EPOLLWRNORM;
3263
3264 mptcp_set_nospace(sk);
3265 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3266 if (sk_stream_is_writeable(sk))
3267 return EPOLLOUT | EPOLLWRNORM;
3268
3269 return 0;
3270}
3271
3272static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3273 struct poll_table_struct *wait)
3274{
3275 struct sock *sk = sock->sk;
3276 struct mptcp_sock *msk;
3277 __poll_t mask = 0;
3278 int state;
3279
3280 msk = mptcp_sk(sk);
3281 sock_poll_wait(file, sock, wait);
3282
3283 state = inet_sk_state_load(sk);
3284 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3285 if (state == TCP_LISTEN)
3286 return test_bit(MPTCP_DATA_READY, &msk->flags) ? EPOLLIN | EPOLLRDNORM : 0;
3287
3288 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3289 mask |= mptcp_check_readable(msk);
3290 mask |= mptcp_check_writeable(msk);
3291 }
3292 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3293 mask |= EPOLLHUP;
3294 if (sk->sk_shutdown & RCV_SHUTDOWN)
3295 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3296
3297 /* This barrier is coupled with smp_wmb() in tcp_reset() */
3298 smp_rmb();
3299 if (sk->sk_err)
3300 mask |= EPOLLERR;
3301
3302 return mask;
3303}
3304
3305static const struct proto_ops mptcp_stream_ops = {
3306 .family = PF_INET,
3307 .owner = THIS_MODULE,
3308 .release = inet_release,
3309 .bind = mptcp_bind,
3310 .connect = mptcp_stream_connect,
3311 .socketpair = sock_no_socketpair,
3312 .accept = mptcp_stream_accept,
3313 .getname = inet_getname,
3314 .poll = mptcp_poll,
3315 .ioctl = inet_ioctl,
3316 .gettstamp = sock_gettstamp,
3317 .listen = mptcp_listen,
3318 .shutdown = inet_shutdown,
3319 .setsockopt = sock_common_setsockopt,
3320 .getsockopt = sock_common_getsockopt,
3321 .sendmsg = inet_sendmsg,
3322 .recvmsg = inet_recvmsg,
3323 .mmap = sock_no_mmap,
3324 .sendpage = inet_sendpage,
3325};
3326
3327static struct inet_protosw mptcp_protosw = {
3328 .type = SOCK_STREAM,
3329 .protocol = IPPROTO_MPTCP,
3330 .prot = &mptcp_prot,
3331 .ops = &mptcp_stream_ops,
3332 .flags = INET_PROTOSW_ICSK,
3333};
3334
3335static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3336{
3337 struct mptcp_delegated_action *delegated;
3338 struct mptcp_subflow_context *subflow;
3339 int work_done = 0;
3340
3341 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3342 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3343 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3344
3345 bh_lock_sock_nested(ssk);
3346 if (!sock_owned_by_user(ssk) &&
3347 mptcp_subflow_has_delegated_action(subflow))
3348 mptcp_subflow_process_delegated(ssk);
3349 /* ... elsewhere tcp_release_cb_override already processed
3350 * the action or will do at next release_sock().
3351 * In both case must dequeue the subflow here - on the same
3352 * CPU that scheduled it.
3353 */
3354 bh_unlock_sock(ssk);
3355 sock_put(ssk);
3356
3357 if (++work_done == budget)
3358 return budget;
3359 }
3360
3361 /* always provide a 0 'work_done' argument, so that napi_complete_done
3362 * will not try accessing the NULL napi->dev ptr
3363 */
3364 napi_complete_done(napi, 0);
3365 return work_done;
3366}
3367
3368void __init mptcp_proto_init(void)
3369{
3370 struct mptcp_delegated_action *delegated;
3371 int cpu;
3372
3373 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
3374
3375 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
3376 panic("Failed to allocate MPTCP pcpu counter\n");
3377
3378 init_dummy_netdev(&mptcp_napi_dev);
3379 for_each_possible_cpu(cpu) {
3380 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
3381 INIT_LIST_HEAD(&delegated->head);
3382 netif_tx_napi_add(&mptcp_napi_dev, &delegated->napi, mptcp_napi_poll,
3383 NAPI_POLL_WEIGHT);
3384 napi_enable(&delegated->napi);
3385 }
3386
3387 mptcp_subflow_init();
3388 mptcp_pm_init();
3389 mptcp_token_init();
3390
3391 if (proto_register(&mptcp_prot, 1) != 0)
3392 panic("Failed to register MPTCP proto.\n");
3393
3394 inet_register_protosw(&mptcp_protosw);
3395
3396 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
3397}
3398
3399#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3400static const struct proto_ops mptcp_v6_stream_ops = {
3401 .family = PF_INET6,
3402 .owner = THIS_MODULE,
3403 .release = inet6_release,
3404 .bind = mptcp_bind,
3405 .connect = mptcp_stream_connect,
3406 .socketpair = sock_no_socketpair,
3407 .accept = mptcp_stream_accept,
3408 .getname = inet6_getname,
3409 .poll = mptcp_poll,
3410 .ioctl = inet6_ioctl,
3411 .gettstamp = sock_gettstamp,
3412 .listen = mptcp_listen,
3413 .shutdown = inet_shutdown,
3414 .setsockopt = sock_common_setsockopt,
3415 .getsockopt = sock_common_getsockopt,
3416 .sendmsg = inet6_sendmsg,
3417 .recvmsg = inet6_recvmsg,
3418 .mmap = sock_no_mmap,
3419 .sendpage = inet_sendpage,
3420#ifdef CONFIG_COMPAT
3421 .compat_ioctl = inet6_compat_ioctl,
3422#endif
3423};
3424
3425static struct proto mptcp_v6_prot;
3426
3427static void mptcp_v6_destroy(struct sock *sk)
3428{
3429 mptcp_destroy(sk);
3430 inet6_destroy_sock(sk);
3431}
3432
3433static struct inet_protosw mptcp_v6_protosw = {
3434 .type = SOCK_STREAM,
3435 .protocol = IPPROTO_MPTCP,
3436 .prot = &mptcp_v6_prot,
3437 .ops = &mptcp_v6_stream_ops,
3438 .flags = INET_PROTOSW_ICSK,
3439};
3440
3441int __init mptcp_proto_v6_init(void)
3442{
3443 int err;
3444
3445 mptcp_v6_prot = mptcp_prot;
3446 strcpy(mptcp_v6_prot.name, "MPTCPv6");
3447 mptcp_v6_prot.slab = NULL;
3448 mptcp_v6_prot.destroy = mptcp_v6_destroy;
3449 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
3450
3451 err = proto_register(&mptcp_v6_prot, 1);
3452 if (err)
3453 return err;
3454
3455 err = inet6_register_protosw(&mptcp_v6_protosw);
3456 if (err)
3457 proto_unregister(&mptcp_v6_prot);
3458
3459 return err;
3460}
3461#endif