]> git.ipfire.org Git - thirdparty/kernel/stable.git/blob - net/mptcp/protocol.c
projects / thirdparty / kernel / stable.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
mptcp: drop the push_pending field
[thirdparty/kernel/stable.git] / net / mptcp / protocol.c
1 // SPDX-License-Identifier: GPL-2.0
2 /* Multipath TCP
3 *
4 * Copyright (c) 2017 - 2019, Intel Corporation.
5 */
6
7 #define pr_fmt(fmt) "MPTCP: " fmt
8
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/netdevice.h>
12 #include <linux/sched/signal.h>
13 #include <linux/atomic.h>
14 #include <net/sock.h>
15 #include <net/inet_common.h>
16 #include <net/inet_hashtables.h>
17 #include <net/protocol.h>
18 #include <net/tcp.h>
19 #include <net/tcp_states.h>
20 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
21 #include <net/transp_v6.h>
22 #endif
23 #include <net/mptcp.h>
24 #include <net/xfrm.h>
25 #include <asm/ioctls.h>
26 #include "protocol.h"
27 #include "mib.h"
28
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/mptcp.h>
31
32 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
33 struct mptcp6_sock {
34 struct mptcp_sock msk;
35 struct ipv6_pinfo np;
36 };
37 #endif
38
39 enum {
40 MPTCP_CMSG_TS = BIT(0),
41 MPTCP_CMSG_INQ = BIT(1),
42 };
43
44 static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
45
46 static void __mptcp_destroy_sock(struct sock *sk);
47 static void mptcp_check_send_data_fin(struct sock *sk);
48
49 DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
50 static struct net_device mptcp_napi_dev;
51
52 /* Returns end sequence number of the receiver's advertised window */
53 static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
54 {
55 return READ_ONCE(msk->wnd_end);
56 }
57
58 static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk)
59 {
60 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
61 if (sk->sk_prot == &tcpv6_prot)
62 return &inet6_stream_ops;
63 #endif
64 WARN_ON_ONCE(sk->sk_prot != &tcp_prot);
65 return &inet_stream_ops;
66 }
67
68 static int __mptcp_socket_create(struct mptcp_sock *msk)
69 {
70 struct mptcp_subflow_context *subflow;
71 struct sock *sk = (struct sock *)msk;
72 struct socket *ssock;
73 int err;
74
75 err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
76 if (err)
77 return err;
78
79 msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
80 WRITE_ONCE(msk->first, ssock->sk);
81 subflow = mptcp_subflow_ctx(ssock->sk);
82 list_add(&subflow->node, &msk->conn_list);
83 sock_hold(ssock->sk);
84 subflow->request_mptcp = 1;
85 subflow->subflow_id = msk->subflow_id++;
86
87 /* This is the first subflow, always with id 0 */
88 subflow->local_id_valid = 1;
89 mptcp_sock_graft(msk->first, sk->sk_socket);
90 iput(SOCK_INODE(ssock));
91
92 return 0;
93 }
94
95 /* If the MPC handshake is not started, returns the first subflow,
96 * eventually allocating it.
97 */
98 struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
99 {
100 struct sock *sk = (struct sock *)msk;
101 int ret;
102
103 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
104 return ERR_PTR(-EINVAL);
105
106 if (!msk->first) {
107 ret = __mptcp_socket_create(msk);
108 if (ret)
109 return ERR_PTR(ret);
110 }
111
112 return msk->first;
113 }
114
115 static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
116 {
117 sk_drops_add(sk, skb);
118 __kfree_skb(skb);
119 }
120
121 static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
122 {
123 WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
124 mptcp_sk(sk)->rmem_fwd_alloc + size);
125 }
126
127 static void mptcp_rmem_charge(struct sock *sk, int size)
128 {
129 mptcp_rmem_fwd_alloc_add(sk, -size);
130 }
131
132 static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
133 struct sk_buff *from)
134 {
135 bool fragstolen;
136 int delta;
137
138 if (MPTCP_SKB_CB(from)->offset ||
139 !skb_try_coalesce(to, from, &fragstolen, &delta))
140 return false;
141
142 pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
143 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
144 to->len, MPTCP_SKB_CB(from)->end_seq);
145 MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
146
147 /* note the fwd memory can reach a negative value after accounting
148 * for the delta, but the later skb free will restore a non
149 * negative one
150 */
151 atomic_add(delta, &sk->sk_rmem_alloc);
152 mptcp_rmem_charge(sk, delta);
153 kfree_skb_partial(from, fragstolen);
154
155 return true;
156 }
157
158 static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
159 struct sk_buff *from)
160 {
161 if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
162 return false;
163
164 return mptcp_try_coalesce((struct sock *)msk, to, from);
165 }
166
167 static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
168 {
169 amount >>= PAGE_SHIFT;
170 mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
171 __sk_mem_reduce_allocated(sk, amount);
172 }
173
174 static void mptcp_rmem_uncharge(struct sock *sk, int size)
175 {
176 struct mptcp_sock *msk = mptcp_sk(sk);
177 int reclaimable;
178
179 mptcp_rmem_fwd_alloc_add(sk, size);
180 reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
181
182 /* see sk_mem_uncharge() for the rationale behind the following schema */
183 if (unlikely(reclaimable >= PAGE_SIZE))
184 __mptcp_rmem_reclaim(sk, reclaimable);
185 }
186
187 static void mptcp_rfree(struct sk_buff *skb)
188 {
189 unsigned int len = skb->truesize;
190 struct sock *sk = skb->sk;
191
192 atomic_sub(len, &sk->sk_rmem_alloc);
193 mptcp_rmem_uncharge(sk, len);
194 }
195
196 void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
197 {
198 skb_orphan(skb);
199 skb->sk = sk;
200 skb->destructor = mptcp_rfree;
201 atomic_add(skb->truesize, &sk->sk_rmem_alloc);
202 mptcp_rmem_charge(sk, skb->truesize);
203 }
204
205 /* "inspired" by tcp_data_queue_ofo(), main differences:
206 * - use mptcp seqs
207 * - don't cope with sacks
208 */
209 static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
210 {
211 struct sock *sk = (struct sock *)msk;
212 struct rb_node **p, *parent;
213 u64 seq, end_seq, max_seq;
214 struct sk_buff *skb1;
215
216 seq = MPTCP_SKB_CB(skb)->map_seq;
217 end_seq = MPTCP_SKB_CB(skb)->end_seq;
218 max_seq = atomic64_read(&msk->rcv_wnd_sent);
219
220 pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
221 RB_EMPTY_ROOT(&msk->out_of_order_queue));
222 if (after64(end_seq, max_seq)) {
223 /* out of window */
224 mptcp_drop(sk, skb);
225 pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
226 (unsigned long long)end_seq - (unsigned long)max_seq,
227 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
228 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
229 return;
230 }
231
232 p = &msk->out_of_order_queue.rb_node;
233 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
234 if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
235 rb_link_node(&skb->rbnode, NULL, p);
236 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
237 msk->ooo_last_skb = skb;
238 goto end;
239 }
240
241 /* with 2 subflows, adding at end of ooo queue is quite likely
242 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
243 */
244 if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
245 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
246 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
247 return;
248 }
249
250 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
251 if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
252 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
253 parent = &msk->ooo_last_skb->rbnode;
254 p = &parent->rb_right;
255 goto insert;
256 }
257
258 /* Find place to insert this segment. Handle overlaps on the way. */
259 parent = NULL;
260 while (*p) {
261 parent = *p;
262 skb1 = rb_to_skb(parent);
263 if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
264 p = &parent->rb_left;
265 continue;
266 }
267 if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
268 if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
269 /* All the bits are present. Drop. */
270 mptcp_drop(sk, skb);
271 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
272 return;
273 }
274 if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
275 /* partial overlap:
276 * | skb |
277 * | skb1 |
278 * continue traversing
279 */
280 } else {
281 /* skb's seq == skb1's seq and skb covers skb1.
282 * Replace skb1 with skb.
283 */
284 rb_replace_node(&skb1->rbnode, &skb->rbnode,
285 &msk->out_of_order_queue);
286 mptcp_drop(sk, skb1);
287 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
288 goto merge_right;
289 }
290 } else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
291 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
292 return;
293 }
294 p = &parent->rb_right;
295 }
296
297 insert:
298 /* Insert segment into RB tree. */
299 rb_link_node(&skb->rbnode, parent, p);
300 rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
301
302 merge_right:
303 /* Remove other segments covered by skb. */
304 while ((skb1 = skb_rb_next(skb)) != NULL) {
305 if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
306 break;
307 rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
308 mptcp_drop(sk, skb1);
309 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
310 }
311 /* If there is no skb after us, we are the last_skb ! */
312 if (!skb1)
313 msk->ooo_last_skb = skb;
314
315 end:
316 skb_condense(skb);
317 mptcp_set_owner_r(skb, sk);
318 }
319
320 static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
321 {
322 struct mptcp_sock *msk = mptcp_sk(sk);
323 int amt, amount;
324
325 if (size <= msk->rmem_fwd_alloc)
326 return true;
327
328 size -= msk->rmem_fwd_alloc;
329 amt = sk_mem_pages(size);
330 amount = amt << PAGE_SHIFT;
331 if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
332 return false;
333
334 mptcp_rmem_fwd_alloc_add(sk, amount);
335 return true;
336 }
337
338 static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
339 struct sk_buff *skb, unsigned int offset,
340 size_t copy_len)
341 {
342 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
343 struct sock *sk = (struct sock *)msk;
344 struct sk_buff *tail;
345 bool has_rxtstamp;
346
347 __skb_unlink(skb, &ssk->sk_receive_queue);
348
349 skb_ext_reset(skb);
350 skb_orphan(skb);
351
352 /* try to fetch required memory from subflow */
353 if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
354 goto drop;
355
356 has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
357
358 /* the skb map_seq accounts for the skb offset:
359 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
360 * value
361 */
362 MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
363 MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
364 MPTCP_SKB_CB(skb)->offset = offset;
365 MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
366
367 if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
368 /* in sequence */
369 msk->bytes_received += copy_len;
370 WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
371 tail = skb_peek_tail(&sk->sk_receive_queue);
372 if (tail && mptcp_try_coalesce(sk, tail, skb))
373 return true;
374
375 mptcp_set_owner_r(skb, sk);
376 __skb_queue_tail(&sk->sk_receive_queue, skb);
377 return true;
378 } else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
379 mptcp_data_queue_ofo(msk, skb);
380 return false;
381 }
382
383 /* old data, keep it simple and drop the whole pkt, sender
384 * will retransmit as needed, if needed.
385 */
386 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
387 drop:
388 mptcp_drop(sk, skb);
389 return false;
390 }
391
392 static void mptcp_stop_rtx_timer(struct sock *sk)
393 {
394 struct inet_connection_sock *icsk = inet_csk(sk);
395
396 sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
397 mptcp_sk(sk)->timer_ival = 0;
398 }
399
400 static void mptcp_close_wake_up(struct sock *sk)
401 {
402 if (sock_flag(sk, SOCK_DEAD))
403 return;
404
405 sk->sk_state_change(sk);
406 if (sk->sk_shutdown == SHUTDOWN_MASK ||
407 sk->sk_state == TCP_CLOSE)
408 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
409 else
410 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
411 }
412
413 static bool mptcp_pending_data_fin_ack(struct sock *sk)
414 {
415 struct mptcp_sock *msk = mptcp_sk(sk);
416
417 return ((1 << sk->sk_state) &
418 (TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
419 msk->write_seq == READ_ONCE(msk->snd_una);
420 }
421
422 static void mptcp_check_data_fin_ack(struct sock *sk)
423 {
424 struct mptcp_sock *msk = mptcp_sk(sk);
425
426 /* Look for an acknowledged DATA_FIN */
427 if (mptcp_pending_data_fin_ack(sk)) {
428 WRITE_ONCE(msk->snd_data_fin_enable, 0);
429
430 switch (sk->sk_state) {
431 case TCP_FIN_WAIT1:
432 mptcp_set_state(sk, TCP_FIN_WAIT2);
433 break;
434 case TCP_CLOSING:
435 case TCP_LAST_ACK:
436 mptcp_set_state(sk, TCP_CLOSE);
437 break;
438 }
439
440 mptcp_close_wake_up(sk);
441 }
442 }
443
444 static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
445 {
446 struct mptcp_sock *msk = mptcp_sk(sk);
447
448 if (READ_ONCE(msk->rcv_data_fin) &&
449 ((1 << sk->sk_state) &
450 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
451 u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
452
453 if (msk->ack_seq == rcv_data_fin_seq) {
454 if (seq)
455 *seq = rcv_data_fin_seq;
456
457 return true;
458 }
459 }
460
461 return false;
462 }
463
464 static void mptcp_set_datafin_timeout(struct sock *sk)
465 {
466 struct inet_connection_sock *icsk = inet_csk(sk);
467 u32 retransmits;
468
469 retransmits = min_t(u32, icsk->icsk_retransmits,
470 ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
471
472 mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
473 }
474
475 static void __mptcp_set_timeout(struct sock *sk, long tout)
476 {
477 mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
478 }
479
480 static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
481 {
482 const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
483
484 return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
485 inet_csk(ssk)->icsk_timeout - jiffies : 0;
486 }
487
488 static void mptcp_set_timeout(struct sock *sk)
489 {
490 struct mptcp_subflow_context *subflow;
491 long tout = 0;
492
493 mptcp_for_each_subflow(mptcp_sk(sk), subflow)
494 tout = max(tout, mptcp_timeout_from_subflow(subflow));
495 __mptcp_set_timeout(sk, tout);
496 }
497
498 static inline bool tcp_can_send_ack(const struct sock *ssk)
499 {
500 return !((1 << inet_sk_state_load(ssk)) &
501 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
502 }
503
504 void __mptcp_subflow_send_ack(struct sock *ssk)
505 {
506 if (tcp_can_send_ack(ssk))
507 tcp_send_ack(ssk);
508 }
509
510 static void mptcp_subflow_send_ack(struct sock *ssk)
511 {
512 bool slow;
513
514 slow = lock_sock_fast(ssk);
515 __mptcp_subflow_send_ack(ssk);
516 unlock_sock_fast(ssk, slow);
517 }
518
519 static void mptcp_send_ack(struct mptcp_sock *msk)
520 {
521 struct mptcp_subflow_context *subflow;
522
523 mptcp_for_each_subflow(msk, subflow)
524 mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
525 }
526
527 static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
528 {
529 bool slow;
530
531 slow = lock_sock_fast(ssk);
532 if (tcp_can_send_ack(ssk))
533 tcp_cleanup_rbuf(ssk, 1);
534 unlock_sock_fast(ssk, slow);
535 }
536
537 static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
538 {
539 const struct inet_connection_sock *icsk = inet_csk(ssk);
540 u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
541 const struct tcp_sock *tp = tcp_sk(ssk);
542
543 return (ack_pending & ICSK_ACK_SCHED) &&
544 ((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
545 READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
546 (rx_empty && ack_pending &
547 (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
548 }
549
550 static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
551 {
552 int old_space = READ_ONCE(msk->old_wspace);
553 struct mptcp_subflow_context *subflow;
554 struct sock *sk = (struct sock *)msk;
555 int space = __mptcp_space(sk);
556 bool cleanup, rx_empty;
557
558 cleanup = (space > 0) && (space >= (old_space << 1));
559 rx_empty = !__mptcp_rmem(sk);
560
561 mptcp_for_each_subflow(msk, subflow) {
562 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
563
564 if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
565 mptcp_subflow_cleanup_rbuf(ssk);
566 }
567 }
568
569 static bool mptcp_check_data_fin(struct sock *sk)
570 {
571 struct mptcp_sock *msk = mptcp_sk(sk);
572 u64 rcv_data_fin_seq;
573 bool ret = false;
574
575 /* Need to ack a DATA_FIN received from a peer while this side
576 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
577 * msk->rcv_data_fin was set when parsing the incoming options
578 * at the subflow level and the msk lock was not held, so this
579 * is the first opportunity to act on the DATA_FIN and change
580 * the msk state.
581 *
582 * If we are caught up to the sequence number of the incoming
583 * DATA_FIN, send the DATA_ACK now and do state transition. If
584 * not caught up, do nothing and let the recv code send DATA_ACK
585 * when catching up.
586 */
587
588 if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
589 WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
590 WRITE_ONCE(msk->rcv_data_fin, 0);
591
592 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
593 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
594
595 switch (sk->sk_state) {
596 case TCP_ESTABLISHED:
597 mptcp_set_state(sk, TCP_CLOSE_WAIT);
598 break;
599 case TCP_FIN_WAIT1:
600 mptcp_set_state(sk, TCP_CLOSING);
601 break;
602 case TCP_FIN_WAIT2:
603 mptcp_set_state(sk, TCP_CLOSE);
604 break;
605 default:
606 /* Other states not expected */
607 WARN_ON_ONCE(1);
608 break;
609 }
610
611 ret = true;
612 if (!__mptcp_check_fallback(msk))
613 mptcp_send_ack(msk);
614 mptcp_close_wake_up(sk);
615 }
616 return ret;
617 }
618
619 static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
620 struct sock *ssk,
621 unsigned int *bytes)
622 {
623 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
624 struct sock *sk = (struct sock *)msk;
625 unsigned int moved = 0;
626 bool more_data_avail;
627 struct tcp_sock *tp;
628 bool done = false;
629 int sk_rbuf;
630
631 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
632
633 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
634 int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
635
636 if (unlikely(ssk_rbuf > sk_rbuf)) {
637 WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
638 sk_rbuf = ssk_rbuf;
639 }
640 }
641
642 pr_debug("msk=%p ssk=%p", msk, ssk);
643 tp = tcp_sk(ssk);
644 do {
645 u32 map_remaining, offset;
646 u32 seq = tp->copied_seq;
647 struct sk_buff *skb;
648 bool fin;
649
650 /* try to move as much data as available */
651 map_remaining = subflow->map_data_len -
652 mptcp_subflow_get_map_offset(subflow);
653
654 skb = skb_peek(&ssk->sk_receive_queue);
655 if (!skb) {
656 /* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
657 * a different CPU can have already processed the pending
658 * data, stop here or we can enter an infinite loop
659 */
660 if (!moved)
661 done = true;
662 break;
663 }
664
665 if (__mptcp_check_fallback(msk)) {
666 /* Under fallback skbs have no MPTCP extension and TCP could
667 * collapse them between the dummy map creation and the
668 * current dequeue. Be sure to adjust the map size.
669 */
670 map_remaining = skb->len;
671 subflow->map_data_len = skb->len;
672 }
673
674 offset = seq - TCP_SKB_CB(skb)->seq;
675 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
676 if (fin) {
677 done = true;
678 seq++;
679 }
680
681 if (offset < skb->len) {
682 size_t len = skb->len - offset;
683
684 if (tp->urg_data)
685 done = true;
686
687 if (__mptcp_move_skb(msk, ssk, skb, offset, len))
688 moved += len;
689 seq += len;
690
691 if (WARN_ON_ONCE(map_remaining < len))
692 break;
693 } else {
694 WARN_ON_ONCE(!fin);
695 sk_eat_skb(ssk, skb);
696 done = true;
697 }
698
699 WRITE_ONCE(tp->copied_seq, seq);
700 more_data_avail = mptcp_subflow_data_available(ssk);
701
702 if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
703 done = true;
704 break;
705 }
706 } while (more_data_avail);
707
708 *bytes += moved;
709 return done;
710 }
711
712 static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
713 {
714 struct sock *sk = (struct sock *)msk;
715 struct sk_buff *skb, *tail;
716 bool moved = false;
717 struct rb_node *p;
718 u64 end_seq;
719
720 p = rb_first(&msk->out_of_order_queue);
721 pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
722 while (p) {
723 skb = rb_to_skb(p);
724 if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
725 break;
726
727 p = rb_next(p);
728 rb_erase(&skb->rbnode, &msk->out_of_order_queue);
729
730 if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
731 msk->ack_seq))) {
732 mptcp_drop(sk, skb);
733 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
734 continue;
735 }
736
737 end_seq = MPTCP_SKB_CB(skb)->end_seq;
738 tail = skb_peek_tail(&sk->sk_receive_queue);
739 if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
740 int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
741
742 /* skip overlapping data, if any */
743 pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
744 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
745 delta);
746 MPTCP_SKB_CB(skb)->offset += delta;
747 MPTCP_SKB_CB(skb)->map_seq += delta;
748 __skb_queue_tail(&sk->sk_receive_queue, skb);
749 }
750 msk->bytes_received += end_seq - msk->ack_seq;
751 msk->ack_seq = end_seq;
752 moved = true;
753 }
754 return moved;
755 }
756
757 static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
758 {
759 int err = sock_error(ssk);
760 int ssk_state;
761
762 if (!err)
763 return false;
764
765 /* only propagate errors on fallen-back sockets or
766 * on MPC connect
767 */
768 if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
769 return false;
770
771 /* We need to propagate only transition to CLOSE state.
772 * Orphaned socket will see such state change via
773 * subflow_sched_work_if_closed() and that path will properly
774 * destroy the msk as needed.
775 */
776 ssk_state = inet_sk_state_load(ssk);
777 if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
778 mptcp_set_state(sk, ssk_state);
779 WRITE_ONCE(sk->sk_err, -err);
780
781 /* This barrier is coupled with smp_rmb() in mptcp_poll() */
782 smp_wmb();
783 sk_error_report(sk);
784 return true;
785 }
786
787 void __mptcp_error_report(struct sock *sk)
788 {
789 struct mptcp_subflow_context *subflow;
790 struct mptcp_sock *msk = mptcp_sk(sk);
791
792 mptcp_for_each_subflow(msk, subflow)
793 if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
794 break;
795 }
796
797 /* In most cases we will be able to lock the mptcp socket. If its already
798 * owned, we need to defer to the work queue to avoid ABBA deadlock.
799 */
800 static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
801 {
802 struct sock *sk = (struct sock *)msk;
803 unsigned int moved = 0;
804
805 __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
806 __mptcp_ofo_queue(msk);
807 if (unlikely(ssk->sk_err)) {
808 if (!sock_owned_by_user(sk))
809 __mptcp_error_report(sk);
810 else
811 __set_bit(MPTCP_ERROR_REPORT, &msk->cb_flags);
812 }
813
814 /* If the moves have caught up with the DATA_FIN sequence number
815 * it's time to ack the DATA_FIN and change socket state, but
816 * this is not a good place to change state. Let the workqueue
817 * do it.
818 */
819 if (mptcp_pending_data_fin(sk, NULL))
820 mptcp_schedule_work(sk);
821 return moved > 0;
822 }
823
824 void mptcp_data_ready(struct sock *sk, struct sock *ssk)
825 {
826 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
827 struct mptcp_sock *msk = mptcp_sk(sk);
828 int sk_rbuf, ssk_rbuf;
829
830 /* The peer can send data while we are shutting down this
831 * subflow at msk destruction time, but we must avoid enqueuing
832 * more data to the msk receive queue
833 */
834 if (unlikely(subflow->disposable))
835 return;
836
837 ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
838 sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
839 if (unlikely(ssk_rbuf > sk_rbuf))
840 sk_rbuf = ssk_rbuf;
841
842 /* over limit? can't append more skbs to msk, Also, no need to wake-up*/
843 if (__mptcp_rmem(sk) > sk_rbuf) {
844 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
845 return;
846 }
847
848 /* Wake-up the reader only for in-sequence data */
849 mptcp_data_lock(sk);
850 if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
851 sk->sk_data_ready(sk);
852 mptcp_data_unlock(sk);
853 }
854
855 static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
856 {
857 mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
858 WRITE_ONCE(msk->allow_infinite_fallback, false);
859 mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
860 }
861
862 static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
863 {
864 struct sock *sk = (struct sock *)msk;
865
866 if (sk->sk_state != TCP_ESTABLISHED)
867 return false;
868
869 /* attach to msk socket only after we are sure we will deal with it
870 * at close time
871 */
872 if (sk->sk_socket && !ssk->sk_socket)
873 mptcp_sock_graft(ssk, sk->sk_socket);
874
875 mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
876 mptcp_sockopt_sync_locked(msk, ssk);
877 mptcp_subflow_joined(msk, ssk);
878 mptcp_stop_tout_timer(sk);
879 __mptcp_propagate_sndbuf(sk, ssk);
880 return true;
881 }
882
883 static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
884 {
885 struct mptcp_subflow_context *tmp, *subflow;
886 struct mptcp_sock *msk = mptcp_sk(sk);
887
888 list_for_each_entry_safe(subflow, tmp, join_list, node) {
889 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
890 bool slow = lock_sock_fast(ssk);
891
892 list_move_tail(&subflow->node, &msk->conn_list);
893 if (!__mptcp_finish_join(msk, ssk))
894 mptcp_subflow_reset(ssk);
895 unlock_sock_fast(ssk, slow);
896 }
897 }
898
899 static bool mptcp_rtx_timer_pending(struct sock *sk)
900 {
901 return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
902 }
903
904 static void mptcp_reset_rtx_timer(struct sock *sk)
905 {
906 struct inet_connection_sock *icsk = inet_csk(sk);
907 unsigned long tout;
908
909 /* prevent rescheduling on close */
910 if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
911 return;
912
913 tout = mptcp_sk(sk)->timer_ival;
914 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
915 }
916
917 bool mptcp_schedule_work(struct sock *sk)
918 {
919 if (inet_sk_state_load(sk) != TCP_CLOSE &&
920 schedule_work(&mptcp_sk(sk)->work)) {
921 /* each subflow already holds a reference to the sk, and the
922 * workqueue is invoked by a subflow, so sk can't go away here.
923 */
924 sock_hold(sk);
925 return true;
926 }
927 return false;
928 }
929
930 static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
931 {
932 struct mptcp_subflow_context *subflow;
933
934 msk_owned_by_me(msk);
935
936 mptcp_for_each_subflow(msk, subflow) {
937 if (READ_ONCE(subflow->data_avail))
938 return mptcp_subflow_tcp_sock(subflow);
939 }
940
941 return NULL;
942 }
943
944 static bool mptcp_skb_can_collapse_to(u64 write_seq,
945 const struct sk_buff *skb,
946 const struct mptcp_ext *mpext)
947 {
948 if (!tcp_skb_can_collapse_to(skb))
949 return false;
950
951 /* can collapse only if MPTCP level sequence is in order and this
952 * mapping has not been xmitted yet
953 */
954 return mpext && mpext->data_seq + mpext->data_len == write_seq &&
955 !mpext->frozen;
956 }
957
958 /* we can append data to the given data frag if:
959 * - there is space available in the backing page_frag
960 * - the data frag tail matches the current page_frag free offset
961 * - the data frag end sequence number matches the current write seq
962 */
963 static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
964 const struct page_frag *pfrag,
965 const struct mptcp_data_frag *df)
966 {
967 return df && pfrag->page == df->page &&
968 pfrag->size - pfrag->offset > 0 &&
969 pfrag->offset == (df->offset + df->data_len) &&
970 df->data_seq + df->data_len == msk->write_seq;
971 }
972
973 static void dfrag_uncharge(struct sock *sk, int len)
974 {
975 sk_mem_uncharge(sk, len);
976 sk_wmem_queued_add(sk, -len);
977 }
978
979 static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
980 {
981 int len = dfrag->data_len + dfrag->overhead;
982
983 list_del(&dfrag->list);
984 dfrag_uncharge(sk, len);
985 put_page(dfrag->page);
986 }
987
988 static void __mptcp_clean_una(struct sock *sk)
989 {
990 struct mptcp_sock *msk = mptcp_sk(sk);
991 struct mptcp_data_frag *dtmp, *dfrag;
992 u64 snd_una;
993
994 snd_una = msk->snd_una;
995 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
996 if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
997 break;
998
999 if (unlikely(dfrag == msk->first_pending)) {
1000 /* in recovery mode can see ack after the current snd head */
1001 if (WARN_ON_ONCE(!msk->recovery))
1002 break;
1003
1004 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1005 }
1006
1007 dfrag_clear(sk, dfrag);
1008 }
1009
1010 dfrag = mptcp_rtx_head(sk);
1011 if (dfrag && after64(snd_una, dfrag->data_seq)) {
1012 u64 delta = snd_una - dfrag->data_seq;
1013
1014 /* prevent wrap around in recovery mode */
1015 if (unlikely(delta > dfrag->already_sent)) {
1016 if (WARN_ON_ONCE(!msk->recovery))
1017 goto out;
1018 if (WARN_ON_ONCE(delta > dfrag->data_len))
1019 goto out;
1020 dfrag->already_sent += delta - dfrag->already_sent;
1021 }
1022
1023 dfrag->data_seq += delta;
1024 dfrag->offset += delta;
1025 dfrag->data_len -= delta;
1026 dfrag->already_sent -= delta;
1027
1028 dfrag_uncharge(sk, delta);
1029 }
1030
1031 /* all retransmitted data acked, recovery completed */
1032 if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1033 msk->recovery = false;
1034
1035 out:
1036 if (snd_una == READ_ONCE(msk->snd_nxt) &&
1037 snd_una == READ_ONCE(msk->write_seq)) {
1038 if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1039 mptcp_stop_rtx_timer(sk);
1040 } else {
1041 mptcp_reset_rtx_timer(sk);
1042 }
1043 }
1044
1045 static void __mptcp_clean_una_wakeup(struct sock *sk)
1046 {
1047 lockdep_assert_held_once(&sk->sk_lock.slock);
1048
1049 __mptcp_clean_una(sk);
1050 mptcp_write_space(sk);
1051 }
1052
1053 static void mptcp_clean_una_wakeup(struct sock *sk)
1054 {
1055 mptcp_data_lock(sk);
1056 __mptcp_clean_una_wakeup(sk);
1057 mptcp_data_unlock(sk);
1058 }
1059
1060 static void mptcp_enter_memory_pressure(struct sock *sk)
1061 {
1062 struct mptcp_subflow_context *subflow;
1063 struct mptcp_sock *msk = mptcp_sk(sk);
1064 bool first = true;
1065
1066 mptcp_for_each_subflow(msk, subflow) {
1067 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1068
1069 if (first)
1070 tcp_enter_memory_pressure(ssk);
1071 sk_stream_moderate_sndbuf(ssk);
1072
1073 first = false;
1074 }
1075 __mptcp_sync_sndbuf(sk);
1076 }
1077
1078 /* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1079 * data
1080 */
1081 static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1082 {
1083 if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1084 pfrag, sk->sk_allocation)))
1085 return true;
1086
1087 mptcp_enter_memory_pressure(sk);
1088 return false;
1089 }
1090
1091 static struct mptcp_data_frag *
1092 mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1093 int orig_offset)
1094 {
1095 int offset = ALIGN(orig_offset, sizeof(long));
1096 struct mptcp_data_frag *dfrag;
1097
1098 dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1099 dfrag->data_len = 0;
1100 dfrag->data_seq = msk->write_seq;
1101 dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1102 dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1103 dfrag->already_sent = 0;
1104 dfrag->page = pfrag->page;
1105
1106 return dfrag;
1107 }
1108
1109 struct mptcp_sendmsg_info {
1110 int mss_now;
1111 int size_goal;
1112 u16 limit;
1113 u16 sent;
1114 unsigned int flags;
1115 bool data_lock_held;
1116 };
1117
1118 static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1119 u64 data_seq, int avail_size)
1120 {
1121 u64 window_end = mptcp_wnd_end(msk);
1122 u64 mptcp_snd_wnd;
1123
1124 if (__mptcp_check_fallback(msk))
1125 return avail_size;
1126
1127 mptcp_snd_wnd = window_end - data_seq;
1128 avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1129
1130 if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1131 tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1132 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1133 }
1134
1135 return avail_size;
1136 }
1137
1138 static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1139 {
1140 struct skb_ext *mpext = __skb_ext_alloc(gfp);
1141
1142 if (!mpext)
1143 return false;
1144 __skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1145 return true;
1146 }
1147
1148 static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1149 {
1150 struct sk_buff *skb;
1151
1152 skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1153 if (likely(skb)) {
1154 if (likely(__mptcp_add_ext(skb, gfp))) {
1155 skb_reserve(skb, MAX_TCP_HEADER);
1156 skb->ip_summed = CHECKSUM_PARTIAL;
1157 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1158 return skb;
1159 }
1160 __kfree_skb(skb);
1161 } else {
1162 mptcp_enter_memory_pressure(sk);
1163 }
1164 return NULL;
1165 }
1166
1167 static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1168 {
1169 struct sk_buff *skb;
1170
1171 skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1172 if (!skb)
1173 return NULL;
1174
1175 if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1176 tcp_skb_entail(ssk, skb);
1177 return skb;
1178 }
1179 tcp_skb_tsorted_anchor_cleanup(skb);
1180 kfree_skb(skb);
1181 return NULL;
1182 }
1183
1184 static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1185 {
1186 gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1187
1188 return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1189 }
1190
1191 /* note: this always recompute the csum on the whole skb, even
1192 * if we just appended a single frag. More status info needed
1193 */
1194 static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1195 {
1196 struct mptcp_ext *mpext = mptcp_get_ext(skb);
1197 __wsum csum = ~csum_unfold(mpext->csum);
1198 int offset = skb->len - added;
1199
1200 mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1201 }
1202
1203 static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1204 struct sock *ssk,
1205 struct mptcp_ext *mpext)
1206 {
1207 if (!mpext)
1208 return;
1209
1210 mpext->infinite_map = 1;
1211 mpext->data_len = 0;
1212
1213 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1214 mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1215 pr_fallback(msk);
1216 mptcp_do_fallback(ssk);
1217 }
1218
1219 #define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1220
1221 static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1222 struct mptcp_data_frag *dfrag,
1223 struct mptcp_sendmsg_info *info)
1224 {
1225 u64 data_seq = dfrag->data_seq + info->sent;
1226 int offset = dfrag->offset + info->sent;
1227 struct mptcp_sock *msk = mptcp_sk(sk);
1228 bool zero_window_probe = false;
1229 struct mptcp_ext *mpext = NULL;
1230 bool can_coalesce = false;
1231 bool reuse_skb = true;
1232 struct sk_buff *skb;
1233 size_t copy;
1234 int i;
1235
1236 pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1237 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1238
1239 if (WARN_ON_ONCE(info->sent > info->limit ||
1240 info->limit > dfrag->data_len))
1241 return 0;
1242
1243 if (unlikely(!__tcp_can_send(ssk)))
1244 return -EAGAIN;
1245
1246 /* compute send limit */
1247 if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1248 ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1249 info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1250 copy = info->size_goal;
1251
1252 skb = tcp_write_queue_tail(ssk);
1253 if (skb && copy > skb->len) {
1254 /* Limit the write to the size available in the
1255 * current skb, if any, so that we create at most a new skb.
1256 * Explicitly tells TCP internals to avoid collapsing on later
1257 * queue management operation, to avoid breaking the ext <->
1258 * SSN association set here
1259 */
1260 mpext = mptcp_get_ext(skb);
1261 if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1262 TCP_SKB_CB(skb)->eor = 1;
1263 goto alloc_skb;
1264 }
1265
1266 i = skb_shinfo(skb)->nr_frags;
1267 can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1268 if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1269 tcp_mark_push(tcp_sk(ssk), skb);
1270 goto alloc_skb;
1271 }
1272
1273 copy -= skb->len;
1274 } else {
1275 alloc_skb:
1276 skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1277 if (!skb)
1278 return -ENOMEM;
1279
1280 i = skb_shinfo(skb)->nr_frags;
1281 reuse_skb = false;
1282 mpext = mptcp_get_ext(skb);
1283 }
1284
1285 /* Zero window and all data acked? Probe. */
1286 copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1287 if (copy == 0) {
1288 u64 snd_una = READ_ONCE(msk->snd_una);
1289
1290 if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1291 tcp_remove_empty_skb(ssk);
1292 return 0;
1293 }
1294
1295 zero_window_probe = true;
1296 data_seq = snd_una - 1;
1297 copy = 1;
1298 }
1299
1300 copy = min_t(size_t, copy, info->limit - info->sent);
1301 if (!sk_wmem_schedule(ssk, copy)) {
1302 tcp_remove_empty_skb(ssk);
1303 return -ENOMEM;
1304 }
1305
1306 if (can_coalesce) {
1307 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1308 } else {
1309 get_page(dfrag->page);
1310 skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1311 }
1312
1313 skb->len += copy;
1314 skb->data_len += copy;
1315 skb->truesize += copy;
1316 sk_wmem_queued_add(ssk, copy);
1317 sk_mem_charge(ssk, copy);
1318 WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1319 TCP_SKB_CB(skb)->end_seq += copy;
1320 tcp_skb_pcount_set(skb, 0);
1321
1322 /* on skb reuse we just need to update the DSS len */
1323 if (reuse_skb) {
1324 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1325 mpext->data_len += copy;
1326 goto out;
1327 }
1328
1329 memset(mpext, 0, sizeof(*mpext));
1330 mpext->data_seq = data_seq;
1331 mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1332 mpext->data_len = copy;
1333 mpext->use_map = 1;
1334 mpext->dsn64 = 1;
1335
1336 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1337 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1338 mpext->dsn64);
1339
1340 if (zero_window_probe) {
1341 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1342 mpext->frozen = 1;
1343 if (READ_ONCE(msk->csum_enabled))
1344 mptcp_update_data_checksum(skb, copy);
1345 tcp_push_pending_frames(ssk);
1346 return 0;
1347 }
1348 out:
1349 if (READ_ONCE(msk->csum_enabled))
1350 mptcp_update_data_checksum(skb, copy);
1351 if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1352 mptcp_update_infinite_map(msk, ssk, mpext);
1353 trace_mptcp_sendmsg_frag(mpext);
1354 mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1355 return copy;
1356 }
1357
1358 #define MPTCP_SEND_BURST_SIZE ((1 << 16) - \
1359 sizeof(struct tcphdr) - \
1360 MAX_TCP_OPTION_SPACE - \
1361 sizeof(struct ipv6hdr) - \
1362 sizeof(struct frag_hdr))
1363
1364 struct subflow_send_info {
1365 struct sock *ssk;
1366 u64 linger_time;
1367 };
1368
1369 void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1370 {
1371 if (!subflow->stale)
1372 return;
1373
1374 subflow->stale = 0;
1375 MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1376 }
1377
1378 bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1379 {
1380 if (unlikely(subflow->stale)) {
1381 u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1382
1383 if (subflow->stale_rcv_tstamp == rcv_tstamp)
1384 return false;
1385
1386 mptcp_subflow_set_active(subflow);
1387 }
1388 return __mptcp_subflow_active(subflow);
1389 }
1390
1391 #define SSK_MODE_ACTIVE 0
1392 #define SSK_MODE_BACKUP 1
1393 #define SSK_MODE_MAX 2
1394
1395 /* implement the mptcp packet scheduler;
1396 * returns the subflow that will transmit the next DSS
1397 * additionally updates the rtx timeout
1398 */
1399 struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1400 {
1401 struct subflow_send_info send_info[SSK_MODE_MAX];
1402 struct mptcp_subflow_context *subflow;
1403 struct sock *sk = (struct sock *)msk;
1404 u32 pace, burst, wmem;
1405 int i, nr_active = 0;
1406 struct sock *ssk;
1407 u64 linger_time;
1408 long tout = 0;
1409
1410 /* pick the subflow with the lower wmem/wspace ratio */
1411 for (i = 0; i < SSK_MODE_MAX; ++i) {
1412 send_info[i].ssk = NULL;
1413 send_info[i].linger_time = -1;
1414 }
1415
1416 mptcp_for_each_subflow(msk, subflow) {
1417 trace_mptcp_subflow_get_send(subflow);
1418 ssk = mptcp_subflow_tcp_sock(subflow);
1419 if (!mptcp_subflow_active(subflow))
1420 continue;
1421
1422 tout = max(tout, mptcp_timeout_from_subflow(subflow));
1423 nr_active += !subflow->backup;
1424 pace = subflow->avg_pacing_rate;
1425 if (unlikely(!pace)) {
1426 /* init pacing rate from socket */
1427 subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1428 pace = subflow->avg_pacing_rate;
1429 if (!pace)
1430 continue;
1431 }
1432
1433 linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1434 if (linger_time < send_info[subflow->backup].linger_time) {
1435 send_info[subflow->backup].ssk = ssk;
1436 send_info[subflow->backup].linger_time = linger_time;
1437 }
1438 }
1439 __mptcp_set_timeout(sk, tout);
1440
1441 /* pick the best backup if no other subflow is active */
1442 if (!nr_active)
1443 send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1444
1445 /* According to the blest algorithm, to avoid HoL blocking for the
1446 * faster flow, we need to:
1447 * - estimate the faster flow linger time
1448 * - use the above to estimate the amount of byte transferred
1449 * by the faster flow
1450 * - check that the amount of queued data is greter than the above,
1451 * otherwise do not use the picked, slower, subflow
1452 * We select the subflow with the shorter estimated time to flush
1453 * the queued mem, which basically ensure the above. We just need
1454 * to check that subflow has a non empty cwin.
1455 */
1456 ssk = send_info[SSK_MODE_ACTIVE].ssk;
1457 if (!ssk || !sk_stream_memory_free(ssk))
1458 return NULL;
1459
1460 burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1461 wmem = READ_ONCE(ssk->sk_wmem_queued);
1462 if (!burst)
1463 return ssk;
1464
1465 subflow = mptcp_subflow_ctx(ssk);
1466 subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1467 READ_ONCE(ssk->sk_pacing_rate) * burst,
1468 burst + wmem);
1469 msk->snd_burst = burst;
1470 return ssk;
1471 }
1472
1473 static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1474 {
1475 tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1476 release_sock(ssk);
1477 }
1478
1479 static void mptcp_update_post_push(struct mptcp_sock *msk,
1480 struct mptcp_data_frag *dfrag,
1481 u32 sent)
1482 {
1483 u64 snd_nxt_new = dfrag->data_seq;
1484
1485 dfrag->already_sent += sent;
1486
1487 msk->snd_burst -= sent;
1488
1489 snd_nxt_new += dfrag->already_sent;
1490
1491 /* snd_nxt_new can be smaller than snd_nxt in case mptcp
1492 * is recovering after a failover. In that event, this re-sends
1493 * old segments.
1494 *
1495 * Thus compute snd_nxt_new candidate based on
1496 * the dfrag->data_seq that was sent and the data
1497 * that has been handed to the subflow for transmission
1498 * and skip update in case it was old dfrag.
1499 */
1500 if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1501 msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1502 msk->snd_nxt = snd_nxt_new;
1503 }
1504 }
1505
1506 void mptcp_check_and_set_pending(struct sock *sk)
1507 {
1508 if (mptcp_send_head(sk)) {
1509 mptcp_data_lock(sk);
1510 mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1511 mptcp_data_unlock(sk);
1512 }
1513 }
1514
1515 static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1516 struct mptcp_sendmsg_info *info)
1517 {
1518 struct mptcp_sock *msk = mptcp_sk(sk);
1519 struct mptcp_data_frag *dfrag;
1520 int len, copied = 0, err = 0;
1521
1522 while ((dfrag = mptcp_send_head(sk))) {
1523 info->sent = dfrag->already_sent;
1524 info->limit = dfrag->data_len;
1525 len = dfrag->data_len - dfrag->already_sent;
1526 while (len > 0) {
1527 int ret = 0;
1528
1529 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1530 if (ret <= 0) {
1531 err = copied ? : ret;
1532 goto out;
1533 }
1534
1535 info->sent += ret;
1536 copied += ret;
1537 len -= ret;
1538
1539 mptcp_update_post_push(msk, dfrag, ret);
1540 }
1541 WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1542
1543 if (msk->snd_burst <= 0 ||
1544 !sk_stream_memory_free(ssk) ||
1545 !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1546 err = copied;
1547 goto out;
1548 }
1549 mptcp_set_timeout(sk);
1550 }
1551 err = copied;
1552
1553 out:
1554 return err;
1555 }
1556
1557 void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1558 {
1559 struct sock *prev_ssk = NULL, *ssk = NULL;
1560 struct mptcp_sock *msk = mptcp_sk(sk);
1561 struct mptcp_sendmsg_info info = {
1562 .flags = flags,
1563 };
1564 bool do_check_data_fin = false;
1565 int push_count = 1;
1566
1567 while (mptcp_send_head(sk) && (push_count > 0)) {
1568 struct mptcp_subflow_context *subflow;
1569 int ret = 0;
1570
1571 if (mptcp_sched_get_send(msk))
1572 break;
1573
1574 push_count = 0;
1575
1576 mptcp_for_each_subflow(msk, subflow) {
1577 if (READ_ONCE(subflow->scheduled)) {
1578 mptcp_subflow_set_scheduled(subflow, false);
1579
1580 prev_ssk = ssk;
1581 ssk = mptcp_subflow_tcp_sock(subflow);
1582 if (ssk != prev_ssk) {
1583 /* First check. If the ssk has changed since
1584 * the last round, release prev_ssk
1585 */
1586 if (prev_ssk)
1587 mptcp_push_release(prev_ssk, &info);
1588
1589 /* Need to lock the new subflow only if different
1590 * from the previous one, otherwise we are still
1591 * helding the relevant lock
1592 */
1593 lock_sock(ssk);
1594 }
1595
1596 push_count++;
1597
1598 ret = __subflow_push_pending(sk, ssk, &info);
1599 if (ret <= 0) {
1600 if (ret != -EAGAIN ||
1601 (1 << ssk->sk_state) &
1602 (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1603 push_count--;
1604 continue;
1605 }
1606 do_check_data_fin = true;
1607 }
1608 }
1609 }
1610
1611 /* at this point we held the socket lock for the last subflow we used */
1612 if (ssk)
1613 mptcp_push_release(ssk, &info);
1614
1615 /* ensure the rtx timer is running */
1616 if (!mptcp_rtx_timer_pending(sk))
1617 mptcp_reset_rtx_timer(sk);
1618 if (do_check_data_fin)
1619 mptcp_check_send_data_fin(sk);
1620 }
1621
1622 static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1623 {
1624 struct mptcp_sock *msk = mptcp_sk(sk);
1625 struct mptcp_sendmsg_info info = {
1626 .data_lock_held = true,
1627 };
1628 bool keep_pushing = true;
1629 struct sock *xmit_ssk;
1630 int copied = 0;
1631
1632 info.flags = 0;
1633 while (mptcp_send_head(sk) && keep_pushing) {
1634 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1635 int ret = 0;
1636
1637 /* check for a different subflow usage only after
1638 * spooling the first chunk of data
1639 */
1640 if (first) {
1641 mptcp_subflow_set_scheduled(subflow, false);
1642 ret = __subflow_push_pending(sk, ssk, &info);
1643 first = false;
1644 if (ret <= 0)
1645 break;
1646 copied += ret;
1647 continue;
1648 }
1649
1650 if (mptcp_sched_get_send(msk))
1651 goto out;
1652
1653 if (READ_ONCE(subflow->scheduled)) {
1654 mptcp_subflow_set_scheduled(subflow, false);
1655 ret = __subflow_push_pending(sk, ssk, &info);
1656 if (ret <= 0)
1657 keep_pushing = false;
1658 copied += ret;
1659 }
1660
1661 mptcp_for_each_subflow(msk, subflow) {
1662 if (READ_ONCE(subflow->scheduled)) {
1663 xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1664 if (xmit_ssk != ssk) {
1665 mptcp_subflow_delegate(subflow,
1666 MPTCP_DELEGATE_SEND);
1667 keep_pushing = false;
1668 }
1669 }
1670 }
1671 }
1672
1673 out:
1674 /* __mptcp_alloc_tx_skb could have released some wmem and we are
1675 * not going to flush it via release_sock()
1676 */
1677 if (copied) {
1678 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1679 info.size_goal);
1680 if (!mptcp_rtx_timer_pending(sk))
1681 mptcp_reset_rtx_timer(sk);
1682
1683 if (msk->snd_data_fin_enable &&
1684 msk->snd_nxt + 1 == msk->write_seq)
1685 mptcp_schedule_work(sk);
1686 }
1687 }
1688
1689 static void mptcp_set_nospace(struct sock *sk)
1690 {
1691 /* enable autotune */
1692 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1693
1694 /* will be cleared on avail space */
1695 set_bit(MPTCP_NOSPACE, &mptcp_sk(sk)->flags);
1696 }
1697
1698 static int mptcp_disconnect(struct sock *sk, int flags);
1699
1700 static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1701 size_t len, int *copied_syn)
1702 {
1703 unsigned int saved_flags = msg->msg_flags;
1704 struct mptcp_sock *msk = mptcp_sk(sk);
1705 struct sock *ssk;
1706 int ret;
1707
1708 /* on flags based fastopen the mptcp is supposed to create the
1709 * first subflow right now. Otherwise we are in the defer_connect
1710 * path, and the first subflow must be already present.
1711 * Since the defer_connect flag is cleared after the first succsful
1712 * fastopen attempt, no need to check for additional subflow status.
1713 */
1714 if (msg->msg_flags & MSG_FASTOPEN) {
1715 ssk = __mptcp_nmpc_sk(msk);
1716 if (IS_ERR(ssk))
1717 return PTR_ERR(ssk);
1718 }
1719 if (!msk->first)
1720 return -EINVAL;
1721
1722 ssk = msk->first;
1723
1724 lock_sock(ssk);
1725 msg->msg_flags |= MSG_DONTWAIT;
1726 msk->fastopening = 1;
1727 ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1728 msk->fastopening = 0;
1729 msg->msg_flags = saved_flags;
1730 release_sock(ssk);
1731
1732 /* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1733 if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1734 ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1735 msg->msg_namelen, msg->msg_flags, 1);
1736
1737 /* Keep the same behaviour of plain TCP: zero the copied bytes in
1738 * case of any error, except timeout or signal
1739 */
1740 if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1741 *copied_syn = 0;
1742 } else if (ret && ret != -EINPROGRESS) {
1743 /* The disconnect() op called by tcp_sendmsg_fastopen()/
1744 * __inet_stream_connect() can fail, due to looking check,
1745 * see mptcp_disconnect().
1746 * Attempt it again outside the problematic scope.
1747 */
1748 if (!mptcp_disconnect(sk, 0))
1749 sk->sk_socket->state = SS_UNCONNECTED;
1750 }
1751 inet_clear_bit(DEFER_CONNECT, sk);
1752
1753 return ret;
1754 }
1755
1756 static int do_copy_data_nocache(struct sock *sk, int copy,
1757 struct iov_iter *from, char *to)
1758 {
1759 if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1760 if (!copy_from_iter_full_nocache(to, copy, from))
1761 return -EFAULT;
1762 } else if (!copy_from_iter_full(to, copy, from)) {
1763 return -EFAULT;
1764 }
1765 return 0;
1766 }
1767
1768 static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1769 {
1770 struct mptcp_sock *msk = mptcp_sk(sk);
1771 struct page_frag *pfrag;
1772 size_t copied = 0;
1773 int ret = 0;
1774 long timeo;
1775
1776 /* silently ignore everything else */
1777 msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1778
1779 lock_sock(sk);
1780
1781 if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1782 msg->msg_flags & MSG_FASTOPEN)) {
1783 int copied_syn = 0;
1784
1785 ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1786 copied += copied_syn;
1787 if (ret == -EINPROGRESS && copied_syn > 0)
1788 goto out;
1789 else if (ret)
1790 goto do_error;
1791 }
1792
1793 timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1794
1795 if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1796 ret = sk_stream_wait_connect(sk, &timeo);
1797 if (ret)
1798 goto do_error;
1799 }
1800
1801 ret = -EPIPE;
1802 if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1803 goto do_error;
1804
1805 pfrag = sk_page_frag(sk);
1806
1807 while (msg_data_left(msg)) {
1808 int total_ts, frag_truesize = 0;
1809 struct mptcp_data_frag *dfrag;
1810 bool dfrag_collapsed;
1811 size_t psize, offset;
1812
1813 /* reuse tail pfrag, if possible, or carve a new one from the
1814 * page allocator
1815 */
1816 dfrag = mptcp_pending_tail(sk);
1817 dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1818 if (!dfrag_collapsed) {
1819 if (!sk_stream_memory_free(sk))
1820 goto wait_for_memory;
1821
1822 if (!mptcp_page_frag_refill(sk, pfrag))
1823 goto wait_for_memory;
1824
1825 dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1826 frag_truesize = dfrag->overhead;
1827 }
1828
1829 /* we do not bound vs wspace, to allow a single packet.
1830 * memory accounting will prevent execessive memory usage
1831 * anyway
1832 */
1833 offset = dfrag->offset + dfrag->data_len;
1834 psize = pfrag->size - offset;
1835 psize = min_t(size_t, psize, msg_data_left(msg));
1836 total_ts = psize + frag_truesize;
1837
1838 if (!sk_wmem_schedule(sk, total_ts))
1839 goto wait_for_memory;
1840
1841 ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1842 page_address(dfrag->page) + offset);
1843 if (ret)
1844 goto do_error;
1845
1846 /* data successfully copied into the write queue */
1847 sk_forward_alloc_add(sk, -total_ts);
1848 copied += psize;
1849 dfrag->data_len += psize;
1850 frag_truesize += psize;
1851 pfrag->offset += frag_truesize;
1852 WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1853
1854 /* charge data on mptcp pending queue to the msk socket
1855 * Note: we charge such data both to sk and ssk
1856 */
1857 sk_wmem_queued_add(sk, frag_truesize);
1858 if (!dfrag_collapsed) {
1859 get_page(dfrag->page);
1860 list_add_tail(&dfrag->list, &msk->rtx_queue);
1861 if (!msk->first_pending)
1862 WRITE_ONCE(msk->first_pending, dfrag);
1863 }
1864 pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1865 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1866 !dfrag_collapsed);
1867
1868 continue;
1869
1870 wait_for_memory:
1871 mptcp_set_nospace(sk);
1872 __mptcp_push_pending(sk, msg->msg_flags);
1873 ret = sk_stream_wait_memory(sk, &timeo);
1874 if (ret)
1875 goto do_error;
1876 }
1877
1878 if (copied)
1879 __mptcp_push_pending(sk, msg->msg_flags);
1880
1881 out:
1882 release_sock(sk);
1883 return copied;
1884
1885 do_error:
1886 if (copied)
1887 goto out;
1888
1889 copied = sk_stream_error(sk, msg->msg_flags, ret);
1890 goto out;
1891 }
1892
1893 static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1894 struct msghdr *msg,
1895 size_t len, int flags,
1896 struct scm_timestamping_internal *tss,
1897 int *cmsg_flags)
1898 {
1899 struct sk_buff *skb, *tmp;
1900 int copied = 0;
1901
1902 skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1903 u32 offset = MPTCP_SKB_CB(skb)->offset;
1904 u32 data_len = skb->len - offset;
1905 u32 count = min_t(size_t, len - copied, data_len);
1906 int err;
1907
1908 if (!(flags & MSG_TRUNC)) {
1909 err = skb_copy_datagram_msg(skb, offset, msg, count);
1910 if (unlikely(err < 0)) {
1911 if (!copied)
1912 return err;
1913 break;
1914 }
1915 }
1916
1917 if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1918 tcp_update_recv_tstamps(skb, tss);
1919 *cmsg_flags |= MPTCP_CMSG_TS;
1920 }
1921
1922 copied += count;
1923
1924 if (count < data_len) {
1925 if (!(flags & MSG_PEEK)) {
1926 MPTCP_SKB_CB(skb)->offset += count;
1927 MPTCP_SKB_CB(skb)->map_seq += count;
1928 msk->bytes_consumed += count;
1929 }
1930 break;
1931 }
1932
1933 if (!(flags & MSG_PEEK)) {
1934 /* we will bulk release the skb memory later */
1935 skb->destructor = NULL;
1936 WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1937 __skb_unlink(skb, &msk->receive_queue);
1938 __kfree_skb(skb);
1939 msk->bytes_consumed += count;
1940 }
1941
1942 if (copied >= len)
1943 break;
1944 }
1945
1946 return copied;
1947 }
1948
1949 /* receive buffer autotuning. See tcp_rcv_space_adjust for more information.
1950 *
1951 * Only difference: Use highest rtt estimate of the subflows in use.
1952 */
1953 static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1954 {
1955 struct mptcp_subflow_context *subflow;
1956 struct sock *sk = (struct sock *)msk;
1957 u8 scaling_ratio = U8_MAX;
1958 u32 time, advmss = 1;
1959 u64 rtt_us, mstamp;
1960
1961 msk_owned_by_me(msk);
1962
1963 if (copied <= 0)
1964 return;
1965
1966 msk->rcvq_space.copied += copied;
1967
1968 mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1969 time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1970
1971 rtt_us = msk->rcvq_space.rtt_us;
1972 if (rtt_us && time < (rtt_us >> 3))
1973 return;
1974
1975 rtt_us = 0;
1976 mptcp_for_each_subflow(msk, subflow) {
1977 const struct tcp_sock *tp;
1978 u64 sf_rtt_us;
1979 u32 sf_advmss;
1980
1981 tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
1982
1983 sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
1984 sf_advmss = READ_ONCE(tp->advmss);
1985
1986 rtt_us = max(sf_rtt_us, rtt_us);
1987 advmss = max(sf_advmss, advmss);
1988 scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
1989 }
1990
1991 msk->rcvq_space.rtt_us = rtt_us;
1992 msk->scaling_ratio = scaling_ratio;
1993 if (time < (rtt_us >> 3) || rtt_us == 0)
1994 return;
1995
1996 if (msk->rcvq_space.copied <= msk->rcvq_space.space)
1997 goto new_measure;
1998
1999 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2000 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2001 u64 rcvwin, grow;
2002 int rcvbuf;
2003
2004 rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2005
2006 grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2007
2008 do_div(grow, msk->rcvq_space.space);
2009 rcvwin += (grow << 1);
2010
2011 rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2012 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2013
2014 if (rcvbuf > sk->sk_rcvbuf) {
2015 u32 window_clamp;
2016
2017 window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2018 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2019
2020 /* Make subflows follow along. If we do not do this, we
2021 * get drops at subflow level if skbs can't be moved to
2022 * the mptcp rx queue fast enough (announced rcv_win can
2023 * exceed ssk->sk_rcvbuf).
2024 */
2025 mptcp_for_each_subflow(msk, subflow) {
2026 struct sock *ssk;
2027 bool slow;
2028
2029 ssk = mptcp_subflow_tcp_sock(subflow);
2030 slow = lock_sock_fast(ssk);
2031 WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2032 tcp_sk(ssk)->window_clamp = window_clamp;
2033 tcp_cleanup_rbuf(ssk, 1);
2034 unlock_sock_fast(ssk, slow);
2035 }
2036 }
2037 }
2038
2039 msk->rcvq_space.space = msk->rcvq_space.copied;
2040 new_measure:
2041 msk->rcvq_space.copied = 0;
2042 msk->rcvq_space.time = mstamp;
2043 }
2044
2045 static void __mptcp_update_rmem(struct sock *sk)
2046 {
2047 struct mptcp_sock *msk = mptcp_sk(sk);
2048
2049 if (!msk->rmem_released)
2050 return;
2051
2052 atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2053 mptcp_rmem_uncharge(sk, msk->rmem_released);
2054 WRITE_ONCE(msk->rmem_released, 0);
2055 }
2056
2057 static void __mptcp_splice_receive_queue(struct sock *sk)
2058 {
2059 struct mptcp_sock *msk = mptcp_sk(sk);
2060
2061 skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2062 }
2063
2064 static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2065 {
2066 struct sock *sk = (struct sock *)msk;
2067 unsigned int moved = 0;
2068 bool ret, done;
2069
2070 do {
2071 struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2072 bool slowpath;
2073
2074 /* we can have data pending in the subflows only if the msk
2075 * receive buffer was full at subflow_data_ready() time,
2076 * that is an unlikely slow path.
2077 */
2078 if (likely(!ssk))
2079 break;
2080
2081 slowpath = lock_sock_fast(ssk);
2082 mptcp_data_lock(sk);
2083 __mptcp_update_rmem(sk);
2084 done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2085 mptcp_data_unlock(sk);
2086
2087 if (unlikely(ssk->sk_err))
2088 __mptcp_error_report(sk);
2089 unlock_sock_fast(ssk, slowpath);
2090 } while (!done);
2091
2092 /* acquire the data lock only if some input data is pending */
2093 ret = moved > 0;
2094 if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2095 !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2096 mptcp_data_lock(sk);
2097 __mptcp_update_rmem(sk);
2098 ret |= __mptcp_ofo_queue(msk);
2099 __mptcp_splice_receive_queue(sk);
2100 mptcp_data_unlock(sk);
2101 }
2102 if (ret)
2103 mptcp_check_data_fin((struct sock *)msk);
2104 return !skb_queue_empty(&msk->receive_queue);
2105 }
2106
2107 static unsigned int mptcp_inq_hint(const struct sock *sk)
2108 {
2109 const struct mptcp_sock *msk = mptcp_sk(sk);
2110 const struct sk_buff *skb;
2111
2112 skb = skb_peek(&msk->receive_queue);
2113 if (skb) {
2114 u64 hint_val = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
2115
2116 if (hint_val >= INT_MAX)
2117 return INT_MAX;
2118
2119 return (unsigned int)hint_val;
2120 }
2121
2122 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2123 return 1;
2124
2125 return 0;
2126 }
2127
2128 static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2129 int flags, int *addr_len)
2130 {
2131 struct mptcp_sock *msk = mptcp_sk(sk);
2132 struct scm_timestamping_internal tss;
2133 int copied = 0, cmsg_flags = 0;
2134 int target;
2135 long timeo;
2136
2137 /* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2138 if (unlikely(flags & MSG_ERRQUEUE))
2139 return inet_recv_error(sk, msg, len, addr_len);
2140
2141 lock_sock(sk);
2142 if (unlikely(sk->sk_state == TCP_LISTEN)) {
2143 copied = -ENOTCONN;
2144 goto out_err;
2145 }
2146
2147 timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2148
2149 len = min_t(size_t, len, INT_MAX);
2150 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2151
2152 if (unlikely(msk->recvmsg_inq))
2153 cmsg_flags = MPTCP_CMSG_INQ;
2154
2155 while (copied < len) {
2156 int bytes_read;
2157
2158 bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2159 if (unlikely(bytes_read < 0)) {
2160 if (!copied)
2161 copied = bytes_read;
2162 goto out_err;
2163 }
2164
2165 copied += bytes_read;
2166
2167 /* be sure to advertise window change */
2168 mptcp_cleanup_rbuf(msk);
2169
2170 if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2171 continue;
2172
2173 /* only the master socket status is relevant here. The exit
2174 * conditions mirror closely tcp_recvmsg()
2175 */
2176 if (copied >= target)
2177 break;
2178
2179 if (copied) {
2180 if (sk->sk_err ||
2181 sk->sk_state == TCP_CLOSE ||
2182 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2183 !timeo ||
2184 signal_pending(current))
2185 break;
2186 } else {
2187 if (sk->sk_err) {
2188 copied = sock_error(sk);
2189 break;
2190 }
2191
2192 if (sk->sk_shutdown & RCV_SHUTDOWN) {
2193 /* race breaker: the shutdown could be after the
2194 * previous receive queue check
2195 */
2196 if (__mptcp_move_skbs(msk))
2197 continue;
2198 break;
2199 }
2200
2201 if (sk->sk_state == TCP_CLOSE) {
2202 copied = -ENOTCONN;
2203 break;
2204 }
2205
2206 if (!timeo) {
2207 copied = -EAGAIN;
2208 break;
2209 }
2210
2211 if (signal_pending(current)) {
2212 copied = sock_intr_errno(timeo);
2213 break;
2214 }
2215 }
2216
2217 pr_debug("block timeout %ld", timeo);
2218 sk_wait_data(sk, &timeo, NULL);
2219 }
2220
2221 out_err:
2222 if (cmsg_flags && copied >= 0) {
2223 if (cmsg_flags & MPTCP_CMSG_TS)
2224 tcp_recv_timestamp(msg, sk, &tss);
2225
2226 if (cmsg_flags & MPTCP_CMSG_INQ) {
2227 unsigned int inq = mptcp_inq_hint(sk);
2228
2229 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2230 }
2231 }
2232
2233 pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2234 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2235 skb_queue_empty(&msk->receive_queue), copied);
2236 if (!(flags & MSG_PEEK))
2237 mptcp_rcv_space_adjust(msk, copied);
2238
2239 release_sock(sk);
2240 return copied;
2241 }
2242
2243 static void mptcp_retransmit_timer(struct timer_list *t)
2244 {
2245 struct inet_connection_sock *icsk = from_timer(icsk, t,
2246 icsk_retransmit_timer);
2247 struct sock *sk = &icsk->icsk_inet.sk;
2248 struct mptcp_sock *msk = mptcp_sk(sk);
2249
2250 bh_lock_sock(sk);
2251 if (!sock_owned_by_user(sk)) {
2252 /* we need a process context to retransmit */
2253 if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2254 mptcp_schedule_work(sk);
2255 } else {
2256 /* delegate our work to tcp_release_cb() */
2257 __set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2258 }
2259 bh_unlock_sock(sk);
2260 sock_put(sk);
2261 }
2262
2263 static void mptcp_tout_timer(struct timer_list *t)
2264 {
2265 struct sock *sk = from_timer(sk, t, sk_timer);
2266
2267 mptcp_schedule_work(sk);
2268 sock_put(sk);
2269 }
2270
2271 /* Find an idle subflow. Return NULL if there is unacked data at tcp
2272 * level.
2273 *
2274 * A backup subflow is returned only if that is the only kind available.
2275 */
2276 struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2277 {
2278 struct sock *backup = NULL, *pick = NULL;
2279 struct mptcp_subflow_context *subflow;
2280 int min_stale_count = INT_MAX;
2281
2282 mptcp_for_each_subflow(msk, subflow) {
2283 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2284
2285 if (!__mptcp_subflow_active(subflow))
2286 continue;
2287
2288 /* still data outstanding at TCP level? skip this */
2289 if (!tcp_rtx_and_write_queues_empty(ssk)) {
2290 mptcp_pm_subflow_chk_stale(msk, ssk);
2291 min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2292 continue;
2293 }
2294
2295 if (subflow->backup) {
2296 if (!backup)
2297 backup = ssk;
2298 continue;
2299 }
2300
2301 if (!pick)
2302 pick = ssk;
2303 }
2304
2305 if (pick)
2306 return pick;
2307
2308 /* use backup only if there are no progresses anywhere */
2309 return min_stale_count > 1 ? backup : NULL;
2310 }
2311
2312 bool __mptcp_retransmit_pending_data(struct sock *sk)
2313 {
2314 struct mptcp_data_frag *cur, *rtx_head;
2315 struct mptcp_sock *msk = mptcp_sk(sk);
2316
2317 if (__mptcp_check_fallback(msk))
2318 return false;
2319
2320 /* the closing socket has some data untransmitted and/or unacked:
2321 * some data in the mptcp rtx queue has not really xmitted yet.
2322 * keep it simple and re-inject the whole mptcp level rtx queue
2323 */
2324 mptcp_data_lock(sk);
2325 __mptcp_clean_una_wakeup(sk);
2326 rtx_head = mptcp_rtx_head(sk);
2327 if (!rtx_head) {
2328 mptcp_data_unlock(sk);
2329 return false;
2330 }
2331
2332 msk->recovery_snd_nxt = msk->snd_nxt;
2333 msk->recovery = true;
2334 mptcp_data_unlock(sk);
2335
2336 msk->first_pending = rtx_head;
2337 msk->snd_burst = 0;
2338
2339 /* be sure to clear the "sent status" on all re-injected fragments */
2340 list_for_each_entry(cur, &msk->rtx_queue, list) {
2341 if (!cur->already_sent)
2342 break;
2343 cur->already_sent = 0;
2344 }
2345
2346 return true;
2347 }
2348
2349 /* flags for __mptcp_close_ssk() */
2350 #define MPTCP_CF_PUSH BIT(1)
2351 #define MPTCP_CF_FASTCLOSE BIT(2)
2352
2353 /* be sure to send a reset only if the caller asked for it, also
2354 * clean completely the subflow status when the subflow reaches
2355 * TCP_CLOSE state
2356 */
2357 static void __mptcp_subflow_disconnect(struct sock *ssk,
2358 struct mptcp_subflow_context *subflow,
2359 unsigned int flags)
2360 {
2361 if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2362 (flags & MPTCP_CF_FASTCLOSE)) {
2363 /* The MPTCP code never wait on the subflow sockets, TCP-level
2364 * disconnect should never fail
2365 */
2366 WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2367 mptcp_subflow_ctx_reset(subflow);
2368 } else {
2369 tcp_shutdown(ssk, SEND_SHUTDOWN);
2370 }
2371 }
2372
2373 /* subflow sockets can be either outgoing (connect) or incoming
2374 * (accept).
2375 *
2376 * Outgoing subflows use in-kernel sockets.
2377 * Incoming subflows do not have their own 'struct socket' allocated,
2378 * so we need to use tcp_close() after detaching them from the mptcp
2379 * parent socket.
2380 */
2381 static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2382 struct mptcp_subflow_context *subflow,
2383 unsigned int flags)
2384 {
2385 struct mptcp_sock *msk = mptcp_sk(sk);
2386 bool dispose_it, need_push = false;
2387
2388 /* If the first subflow moved to a close state before accept, e.g. due
2389 * to an incoming reset or listener shutdown, the subflow socket is
2390 * already deleted by inet_child_forget() and the mptcp socket can't
2391 * survive too.
2392 */
2393 if (msk->in_accept_queue && msk->first == ssk &&
2394 (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2395 /* ensure later check in mptcp_worker() will dispose the msk */
2396 sock_set_flag(sk, SOCK_DEAD);
2397 mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2398 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2399 mptcp_subflow_drop_ctx(ssk);
2400 goto out_release;
2401 }
2402
2403 dispose_it = msk->free_first || ssk != msk->first;
2404 if (dispose_it)
2405 list_del(&subflow->node);
2406
2407 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2408
2409 if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2410 /* be sure to force the tcp_close path
2411 * to generate the egress reset
2412 */
2413 ssk->sk_lingertime = 0;
2414 sock_set_flag(ssk, SOCK_LINGER);
2415 subflow->send_fastclose = 1;
2416 }
2417
2418 need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2419 if (!dispose_it) {
2420 __mptcp_subflow_disconnect(ssk, subflow, flags);
2421 release_sock(ssk);
2422
2423 goto out;
2424 }
2425
2426 subflow->disposable = 1;
2427
2428 /* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2429 * the ssk has been already destroyed, we just need to release the
2430 * reference owned by msk;
2431 */
2432 if (!inet_csk(ssk)->icsk_ulp_ops) {
2433 WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2434 kfree_rcu(subflow, rcu);
2435 } else {
2436 /* otherwise tcp will dispose of the ssk and subflow ctx */
2437 __tcp_close(ssk, 0);
2438
2439 /* close acquired an extra ref */
2440 __sock_put(ssk);
2441 }
2442
2443 out_release:
2444 __mptcp_subflow_error_report(sk, ssk);
2445 release_sock(ssk);
2446
2447 sock_put(ssk);
2448
2449 if (ssk == msk->first)
2450 WRITE_ONCE(msk->first, NULL);
2451
2452 out:
2453 __mptcp_sync_sndbuf(sk);
2454 if (need_push)
2455 __mptcp_push_pending(sk, 0);
2456
2457 /* Catch every 'all subflows closed' scenario, including peers silently
2458 * closing them, e.g. due to timeout.
2459 * For established sockets, allow an additional timeout before closing,
2460 * as the protocol can still create more subflows.
2461 */
2462 if (list_is_singular(&msk->conn_list) && msk->first &&
2463 inet_sk_state_load(msk->first) == TCP_CLOSE) {
2464 if (sk->sk_state != TCP_ESTABLISHED ||
2465 msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2466 mptcp_set_state(sk, TCP_CLOSE);
2467 mptcp_close_wake_up(sk);
2468 } else {
2469 mptcp_start_tout_timer(sk);
2470 }
2471 }
2472 }
2473
2474 void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2475 struct mptcp_subflow_context *subflow)
2476 {
2477 if (sk->sk_state == TCP_ESTABLISHED)
2478 mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2479
2480 /* subflow aborted before reaching the fully_established status
2481 * attempt the creation of the next subflow
2482 */
2483 mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2484
2485 __mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2486 }
2487
2488 static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2489 {
2490 return 0;
2491 }
2492
2493 static void __mptcp_close_subflow(struct sock *sk)
2494 {
2495 struct mptcp_subflow_context *subflow, *tmp;
2496 struct mptcp_sock *msk = mptcp_sk(sk);
2497
2498 might_sleep();
2499
2500 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2501 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2502
2503 if (inet_sk_state_load(ssk) != TCP_CLOSE)
2504 continue;
2505
2506 /* 'subflow_data_ready' will re-sched once rx queue is empty */
2507 if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2508 continue;
2509
2510 mptcp_close_ssk(sk, ssk, subflow);
2511 }
2512
2513 }
2514
2515 static bool mptcp_close_tout_expired(const struct sock *sk)
2516 {
2517 if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2518 sk->sk_state == TCP_CLOSE)
2519 return false;
2520
2521 return time_after32(tcp_jiffies32,
2522 inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2523 }
2524
2525 static void mptcp_check_fastclose(struct mptcp_sock *msk)
2526 {
2527 struct mptcp_subflow_context *subflow, *tmp;
2528 struct sock *sk = (struct sock *)msk;
2529
2530 if (likely(!READ_ONCE(msk->rcv_fastclose)))
2531 return;
2532
2533 mptcp_token_destroy(msk);
2534
2535 mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2536 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2537 bool slow;
2538
2539 slow = lock_sock_fast(tcp_sk);
2540 if (tcp_sk->sk_state != TCP_CLOSE) {
2541 tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2542 tcp_set_state(tcp_sk, TCP_CLOSE);
2543 }
2544 unlock_sock_fast(tcp_sk, slow);
2545 }
2546
2547 /* Mirror the tcp_reset() error propagation */
2548 switch (sk->sk_state) {
2549 case TCP_SYN_SENT:
2550 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2551 break;
2552 case TCP_CLOSE_WAIT:
2553 WRITE_ONCE(sk->sk_err, EPIPE);
2554 break;
2555 case TCP_CLOSE:
2556 return;
2557 default:
2558 WRITE_ONCE(sk->sk_err, ECONNRESET);
2559 }
2560
2561 mptcp_set_state(sk, TCP_CLOSE);
2562 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2563 smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2564 set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2565
2566 /* the calling mptcp_worker will properly destroy the socket */
2567 if (sock_flag(sk, SOCK_DEAD))
2568 return;
2569
2570 sk->sk_state_change(sk);
2571 sk_error_report(sk);
2572 }
2573
2574 static void __mptcp_retrans(struct sock *sk)
2575 {
2576 struct mptcp_sock *msk = mptcp_sk(sk);
2577 struct mptcp_subflow_context *subflow;
2578 struct mptcp_sendmsg_info info = {};
2579 struct mptcp_data_frag *dfrag;
2580 struct sock *ssk;
2581 int ret, err;
2582 u16 len = 0;
2583
2584 mptcp_clean_una_wakeup(sk);
2585
2586 /* first check ssk: need to kick "stale" logic */
2587 err = mptcp_sched_get_retrans(msk);
2588 dfrag = mptcp_rtx_head(sk);
2589 if (!dfrag) {
2590 if (mptcp_data_fin_enabled(msk)) {
2591 struct inet_connection_sock *icsk = inet_csk(sk);
2592
2593 icsk->icsk_retransmits++;
2594 mptcp_set_datafin_timeout(sk);
2595 mptcp_send_ack(msk);
2596
2597 goto reset_timer;
2598 }
2599
2600 if (!mptcp_send_head(sk))
2601 return;
2602
2603 goto reset_timer;
2604 }
2605
2606 if (err)
2607 goto reset_timer;
2608
2609 mptcp_for_each_subflow(msk, subflow) {
2610 if (READ_ONCE(subflow->scheduled)) {
2611 u16 copied = 0;
2612
2613 mptcp_subflow_set_scheduled(subflow, false);
2614
2615 ssk = mptcp_subflow_tcp_sock(subflow);
2616
2617 lock_sock(ssk);
2618
2619 /* limit retransmission to the bytes already sent on some subflows */
2620 info.sent = 0;
2621 info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2622 dfrag->already_sent;
2623 while (info.sent < info.limit) {
2624 ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2625 if (ret <= 0)
2626 break;
2627
2628 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2629 copied += ret;
2630 info.sent += ret;
2631 }
2632 if (copied) {
2633 len = max(copied, len);
2634 tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2635 info.size_goal);
2636 WRITE_ONCE(msk->allow_infinite_fallback, false);
2637 }
2638
2639 release_sock(ssk);
2640 }
2641 }
2642
2643 msk->bytes_retrans += len;
2644 dfrag->already_sent = max(dfrag->already_sent, len);
2645
2646 reset_timer:
2647 mptcp_check_and_set_pending(sk);
2648
2649 if (!mptcp_rtx_timer_pending(sk))
2650 mptcp_reset_rtx_timer(sk);
2651 }
2652
2653 /* schedule the timeout timer for the relevant event: either close timeout
2654 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2655 */
2656 void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2657 {
2658 struct sock *sk = (struct sock *)msk;
2659 unsigned long timeout, close_timeout;
2660
2661 if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2662 return;
2663
2664 close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2665 mptcp_close_timeout(sk);
2666
2667 /* the close timeout takes precedence on the fail one, and here at least one of
2668 * them is active
2669 */
2670 timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2671
2672 sk_reset_timer(sk, &sk->sk_timer, timeout);
2673 }
2674
2675 static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2676 {
2677 struct sock *ssk = msk->first;
2678 bool slow;
2679
2680 if (!ssk)
2681 return;
2682
2683 pr_debug("MP_FAIL doesn't respond, reset the subflow");
2684
2685 slow = lock_sock_fast(ssk);
2686 mptcp_subflow_reset(ssk);
2687 WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2688 unlock_sock_fast(ssk, slow);
2689 }
2690
2691 static void mptcp_do_fastclose(struct sock *sk)
2692 {
2693 struct mptcp_subflow_context *subflow, *tmp;
2694 struct mptcp_sock *msk = mptcp_sk(sk);
2695
2696 mptcp_set_state(sk, TCP_CLOSE);
2697 mptcp_for_each_subflow_safe(msk, subflow, tmp)
2698 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2699 subflow, MPTCP_CF_FASTCLOSE);
2700 }
2701
2702 static void mptcp_worker(struct work_struct *work)
2703 {
2704 struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2705 struct sock *sk = (struct sock *)msk;
2706 unsigned long fail_tout;
2707 int state;
2708
2709 lock_sock(sk);
2710 state = sk->sk_state;
2711 if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2712 goto unlock;
2713
2714 mptcp_check_fastclose(msk);
2715
2716 mptcp_pm_nl_work(msk);
2717
2718 mptcp_check_send_data_fin(sk);
2719 mptcp_check_data_fin_ack(sk);
2720 mptcp_check_data_fin(sk);
2721
2722 if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2723 __mptcp_close_subflow(sk);
2724
2725 if (mptcp_close_tout_expired(sk)) {
2726 mptcp_do_fastclose(sk);
2727 mptcp_close_wake_up(sk);
2728 }
2729
2730 if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2731 __mptcp_destroy_sock(sk);
2732 goto unlock;
2733 }
2734
2735 if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2736 __mptcp_retrans(sk);
2737
2738 fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2739 if (fail_tout && time_after(jiffies, fail_tout))
2740 mptcp_mp_fail_no_response(msk);
2741
2742 unlock:
2743 release_sock(sk);
2744 sock_put(sk);
2745 }
2746
2747 static void __mptcp_init_sock(struct sock *sk)
2748 {
2749 struct mptcp_sock *msk = mptcp_sk(sk);
2750
2751 INIT_LIST_HEAD(&msk->conn_list);
2752 INIT_LIST_HEAD(&msk->join_list);
2753 INIT_LIST_HEAD(&msk->rtx_queue);
2754 INIT_WORK(&msk->work, mptcp_worker);
2755 __skb_queue_head_init(&msk->receive_queue);
2756 msk->out_of_order_queue = RB_ROOT;
2757 msk->first_pending = NULL;
2758 msk->rmem_fwd_alloc = 0;
2759 WRITE_ONCE(msk->rmem_released, 0);
2760 msk->timer_ival = TCP_RTO_MIN;
2761 msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2762
2763 WRITE_ONCE(msk->first, NULL);
2764 inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2765 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2766 WRITE_ONCE(msk->allow_infinite_fallback, true);
2767 msk->recovery = false;
2768 msk->subflow_id = 1;
2769
2770 mptcp_pm_data_init(msk);
2771
2772 /* re-use the csk retrans timer for MPTCP-level retrans */
2773 timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2774 timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2775 }
2776
2777 static void mptcp_ca_reset(struct sock *sk)
2778 {
2779 struct inet_connection_sock *icsk = inet_csk(sk);
2780
2781 tcp_assign_congestion_control(sk);
2782 strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2783
2784 /* no need to keep a reference to the ops, the name will suffice */
2785 tcp_cleanup_congestion_control(sk);
2786 icsk->icsk_ca_ops = NULL;
2787 }
2788
2789 static int mptcp_init_sock(struct sock *sk)
2790 {
2791 struct net *net = sock_net(sk);
2792 int ret;
2793
2794 __mptcp_init_sock(sk);
2795
2796 if (!mptcp_is_enabled(net))
2797 return -ENOPROTOOPT;
2798
2799 if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2800 return -ENOMEM;
2801
2802 ret = mptcp_init_sched(mptcp_sk(sk),
2803 mptcp_sched_find(mptcp_get_scheduler(net)));
2804 if (ret)
2805 return ret;
2806
2807 set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2808
2809 /* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2810 * propagate the correct value
2811 */
2812 mptcp_ca_reset(sk);
2813
2814 sk_sockets_allocated_inc(sk);
2815 sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2816 sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2817
2818 return 0;
2819 }
2820
2821 static void __mptcp_clear_xmit(struct sock *sk)
2822 {
2823 struct mptcp_sock *msk = mptcp_sk(sk);
2824 struct mptcp_data_frag *dtmp, *dfrag;
2825
2826 WRITE_ONCE(msk->first_pending, NULL);
2827 list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2828 dfrag_clear(sk, dfrag);
2829 }
2830
2831 void mptcp_cancel_work(struct sock *sk)
2832 {
2833 struct mptcp_sock *msk = mptcp_sk(sk);
2834
2835 if (cancel_work_sync(&msk->work))
2836 __sock_put(sk);
2837 }
2838
2839 void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2840 {
2841 lock_sock(ssk);
2842
2843 switch (ssk->sk_state) {
2844 case TCP_LISTEN:
2845 if (!(how & RCV_SHUTDOWN))
2846 break;
2847 fallthrough;
2848 case TCP_SYN_SENT:
2849 WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2850 break;
2851 default:
2852 if (__mptcp_check_fallback(mptcp_sk(sk))) {
2853 pr_debug("Fallback");
2854 ssk->sk_shutdown |= how;
2855 tcp_shutdown(ssk, how);
2856
2857 /* simulate the data_fin ack reception to let the state
2858 * machine move forward
2859 */
2860 WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2861 mptcp_schedule_work(sk);
2862 } else {
2863 pr_debug("Sending DATA_FIN on subflow %p", ssk);
2864 tcp_send_ack(ssk);
2865 if (!mptcp_rtx_timer_pending(sk))
2866 mptcp_reset_rtx_timer(sk);
2867 }
2868 break;
2869 }
2870
2871 release_sock(ssk);
2872 }
2873
2874 void mptcp_set_state(struct sock *sk, int state)
2875 {
2876 int oldstate = sk->sk_state;
2877
2878 switch (state) {
2879 case TCP_ESTABLISHED:
2880 if (oldstate != TCP_ESTABLISHED)
2881 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2882 break;
2883
2884 default:
2885 if (oldstate == TCP_ESTABLISHED)
2886 MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2887 }
2888
2889 inet_sk_state_store(sk, state);
2890 }
2891
2892 static const unsigned char new_state[16] = {
2893 /* current state: new state: action: */
2894 [0 /* (Invalid) */] = TCP_CLOSE,
2895 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2896 [TCP_SYN_SENT] = TCP_CLOSE,
2897 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2898 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2899 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2900 [TCP_TIME_WAIT] = TCP_CLOSE, /* should not happen ! */
2901 [TCP_CLOSE] = TCP_CLOSE,
2902 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2903 [TCP_LAST_ACK] = TCP_LAST_ACK,
2904 [TCP_LISTEN] = TCP_CLOSE,
2905 [TCP_CLOSING] = TCP_CLOSING,
2906 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2907 };
2908
2909 static int mptcp_close_state(struct sock *sk)
2910 {
2911 int next = (int)new_state[sk->sk_state];
2912 int ns = next & TCP_STATE_MASK;
2913
2914 mptcp_set_state(sk, ns);
2915
2916 return next & TCP_ACTION_FIN;
2917 }
2918
2919 static void mptcp_check_send_data_fin(struct sock *sk)
2920 {
2921 struct mptcp_subflow_context *subflow;
2922 struct mptcp_sock *msk = mptcp_sk(sk);
2923
2924 pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2925 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2926 msk->snd_nxt, msk->write_seq);
2927
2928 /* we still need to enqueue subflows or not really shutting down,
2929 * skip this
2930 */
2931 if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2932 mptcp_send_head(sk))
2933 return;
2934
2935 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2936
2937 mptcp_for_each_subflow(msk, subflow) {
2938 struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2939
2940 mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2941 }
2942 }
2943
2944 static void __mptcp_wr_shutdown(struct sock *sk)
2945 {
2946 struct mptcp_sock *msk = mptcp_sk(sk);
2947
2948 pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2949 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2950 !!mptcp_send_head(sk));
2951
2952 /* will be ignored by fallback sockets */
2953 WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2954 WRITE_ONCE(msk->snd_data_fin_enable, 1);
2955
2956 mptcp_check_send_data_fin(sk);
2957 }
2958
2959 static void __mptcp_destroy_sock(struct sock *sk)
2960 {
2961 struct mptcp_sock *msk = mptcp_sk(sk);
2962
2963 pr_debug("msk=%p", msk);
2964
2965 might_sleep();
2966
2967 mptcp_stop_rtx_timer(sk);
2968 sk_stop_timer(sk, &sk->sk_timer);
2969 msk->pm.status = 0;
2970 mptcp_release_sched(msk);
2971
2972 sk->sk_prot->destroy(sk);
2973
2974 WARN_ON_ONCE(msk->rmem_fwd_alloc);
2975 WARN_ON_ONCE(msk->rmem_released);
2976 sk_stream_kill_queues(sk);
2977 xfrm_sk_free_policy(sk);
2978
2979 sock_put(sk);
2980 }
2981
2982 void __mptcp_unaccepted_force_close(struct sock *sk)
2983 {
2984 sock_set_flag(sk, SOCK_DEAD);
2985 mptcp_do_fastclose(sk);
2986 __mptcp_destroy_sock(sk);
2987 }
2988
2989 static __poll_t mptcp_check_readable(struct sock *sk)
2990 {
2991 return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
2992 }
2993
2994 static void mptcp_check_listen_stop(struct sock *sk)
2995 {
2996 struct sock *ssk;
2997
2998 if (inet_sk_state_load(sk) != TCP_LISTEN)
2999 return;
3000
3001 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3002 ssk = mptcp_sk(sk)->first;
3003 if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3004 return;
3005
3006 lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3007 tcp_set_state(ssk, TCP_CLOSE);
3008 mptcp_subflow_queue_clean(sk, ssk);
3009 inet_csk_listen_stop(ssk);
3010 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3011 release_sock(ssk);
3012 }
3013
3014 bool __mptcp_close(struct sock *sk, long timeout)
3015 {
3016 struct mptcp_subflow_context *subflow;
3017 struct mptcp_sock *msk = mptcp_sk(sk);
3018 bool do_cancel_work = false;
3019 int subflows_alive = 0;
3020
3021 WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3022
3023 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3024 mptcp_check_listen_stop(sk);
3025 mptcp_set_state(sk, TCP_CLOSE);
3026 goto cleanup;
3027 }
3028
3029 if (mptcp_data_avail(msk) || timeout < 0) {
3030 /* If the msk has read data, or the caller explicitly ask it,
3031 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3032 */
3033 mptcp_do_fastclose(sk);
3034 timeout = 0;
3035 } else if (mptcp_close_state(sk)) {
3036 __mptcp_wr_shutdown(sk);
3037 }
3038
3039 sk_stream_wait_close(sk, timeout);
3040
3041 cleanup:
3042 /* orphan all the subflows */
3043 mptcp_for_each_subflow(msk, subflow) {
3044 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3045 bool slow = lock_sock_fast_nested(ssk);
3046
3047 subflows_alive += ssk->sk_state != TCP_CLOSE;
3048
3049 /* since the close timeout takes precedence on the fail one,
3050 * cancel the latter
3051 */
3052 if (ssk == msk->first)
3053 subflow->fail_tout = 0;
3054
3055 /* detach from the parent socket, but allow data_ready to
3056 * push incoming data into the mptcp stack, to properly ack it
3057 */
3058 ssk->sk_socket = NULL;
3059 ssk->sk_wq = NULL;
3060 unlock_sock_fast(ssk, slow);
3061 }
3062 sock_orphan(sk);
3063
3064 /* all the subflows are closed, only timeout can change the msk
3065 * state, let's not keep resources busy for no reasons
3066 */
3067 if (subflows_alive == 0)
3068 mptcp_set_state(sk, TCP_CLOSE);
3069
3070 sock_hold(sk);
3071 pr_debug("msk=%p state=%d", sk, sk->sk_state);
3072 if (msk->token)
3073 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3074
3075 if (sk->sk_state == TCP_CLOSE) {
3076 __mptcp_destroy_sock(sk);
3077 do_cancel_work = true;
3078 } else {
3079 mptcp_start_tout_timer(sk);
3080 }
3081
3082 return do_cancel_work;
3083 }
3084
3085 static void mptcp_close(struct sock *sk, long timeout)
3086 {
3087 bool do_cancel_work;
3088
3089 lock_sock(sk);
3090
3091 do_cancel_work = __mptcp_close(sk, timeout);
3092 release_sock(sk);
3093 if (do_cancel_work)
3094 mptcp_cancel_work(sk);
3095
3096 sock_put(sk);
3097 }
3098
3099 static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3100 {
3101 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3102 const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3103 struct ipv6_pinfo *msk6 = inet6_sk(msk);
3104
3105 msk->sk_v6_daddr = ssk->sk_v6_daddr;
3106 msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3107
3108 if (msk6 && ssk6) {
3109 msk6->saddr = ssk6->saddr;
3110 msk6->flow_label = ssk6->flow_label;
3111 }
3112 #endif
3113
3114 inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3115 inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3116 inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3117 inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3118 inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3119 inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3120 }
3121
3122 static int mptcp_disconnect(struct sock *sk, int flags)
3123 {
3124 struct mptcp_sock *msk = mptcp_sk(sk);
3125
3126 /* We are on the fastopen error path. We can't call straight into the
3127 * subflows cleanup code due to lock nesting (we are already under
3128 * msk->firstsocket lock).
3129 */
3130 if (msk->fastopening)
3131 return -EBUSY;
3132
3133 mptcp_check_listen_stop(sk);
3134 mptcp_set_state(sk, TCP_CLOSE);
3135
3136 mptcp_stop_rtx_timer(sk);
3137 mptcp_stop_tout_timer(sk);
3138
3139 if (msk->token)
3140 mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3141
3142 /* msk->subflow is still intact, the following will not free the first
3143 * subflow
3144 */
3145 mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3146 WRITE_ONCE(msk->flags, 0);
3147 msk->cb_flags = 0;
3148 msk->recovery = false;
3149 msk->can_ack = false;
3150 msk->fully_established = false;
3151 msk->rcv_data_fin = false;
3152 msk->snd_data_fin_enable = false;
3153 msk->rcv_fastclose = false;
3154 msk->use_64bit_ack = false;
3155 msk->bytes_consumed = 0;
3156 WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3157 mptcp_pm_data_reset(msk);
3158 mptcp_ca_reset(sk);
3159 msk->bytes_acked = 0;
3160 msk->bytes_received = 0;
3161 msk->bytes_sent = 0;
3162 msk->bytes_retrans = 0;
3163
3164 WRITE_ONCE(sk->sk_shutdown, 0);
3165 sk_error_report(sk);
3166 return 0;
3167 }
3168
3169 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3170 static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3171 {
3172 unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3173
3174 return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3175 }
3176 #endif
3177
3178 struct sock *mptcp_sk_clone_init(const struct sock *sk,
3179 const struct mptcp_options_received *mp_opt,
3180 struct sock *ssk,
3181 struct request_sock *req)
3182 {
3183 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3184 struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3185 struct mptcp_sock *msk;
3186
3187 if (!nsk)
3188 return NULL;
3189
3190 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3191 if (nsk->sk_family == AF_INET6)
3192 inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3193 #endif
3194
3195 __mptcp_init_sock(nsk);
3196
3197 msk = mptcp_sk(nsk);
3198 msk->local_key = subflow_req->local_key;
3199 msk->token = subflow_req->token;
3200 msk->in_accept_queue = 1;
3201 WRITE_ONCE(msk->fully_established, false);
3202 if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3203 WRITE_ONCE(msk->csum_enabled, true);
3204
3205 msk->write_seq = subflow_req->idsn + 1;
3206 msk->snd_nxt = msk->write_seq;
3207 msk->snd_una = msk->write_seq;
3208 msk->wnd_end = msk->snd_nxt + req->rsk_rcv_wnd;
3209 msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3210 mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3211
3212 /* passive msk is created after the first/MPC subflow */
3213 msk->subflow_id = 2;
3214
3215 sock_reset_flag(nsk, SOCK_RCU_FREE);
3216 security_inet_csk_clone(nsk, req);
3217
3218 /* this can't race with mptcp_close(), as the msk is
3219 * not yet exposted to user-space
3220 */
3221 mptcp_set_state(nsk, TCP_ESTABLISHED);
3222
3223 /* The msk maintain a ref to each subflow in the connections list */
3224 WRITE_ONCE(msk->first, ssk);
3225 list_add(&mptcp_subflow_ctx(ssk)->node, &msk->conn_list);
3226 sock_hold(ssk);
3227
3228 /* new mpc subflow takes ownership of the newly
3229 * created mptcp socket
3230 */
3231 mptcp_token_accept(subflow_req, msk);
3232
3233 /* set msk addresses early to ensure mptcp_pm_get_local_id()
3234 * uses the correct data
3235 */
3236 mptcp_copy_inaddrs(nsk, ssk);
3237 __mptcp_propagate_sndbuf(nsk, ssk);
3238
3239 mptcp_rcv_space_init(msk, ssk);
3240 bh_unlock_sock(nsk);
3241
3242 /* note: the newly allocated socket refcount is 2 now */
3243 return nsk;
3244 }
3245
3246 void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3247 {
3248 const struct tcp_sock *tp = tcp_sk(ssk);
3249
3250 msk->rcvq_space.copied = 0;
3251 msk->rcvq_space.rtt_us = 0;
3252
3253 msk->rcvq_space.time = tp->tcp_mstamp;
3254
3255 /* initial rcv_space offering made to peer */
3256 msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3257 TCP_INIT_CWND * tp->advmss);
3258 if (msk->rcvq_space.space == 0)
3259 msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3260
3261 WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3262 }
3263
3264 void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3265 {
3266 struct mptcp_subflow_context *subflow, *tmp;
3267 struct sock *sk = (struct sock *)msk;
3268
3269 __mptcp_clear_xmit(sk);
3270
3271 /* join list will be eventually flushed (with rst) at sock lock release time */
3272 mptcp_for_each_subflow_safe(msk, subflow, tmp)
3273 __mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3274
3275 /* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3276 mptcp_data_lock(sk);
3277 skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3278 __skb_queue_purge(&sk->sk_receive_queue);
3279 skb_rbtree_purge(&msk->out_of_order_queue);
3280 mptcp_data_unlock(sk);
3281
3282 /* move all the rx fwd alloc into the sk_mem_reclaim_final in
3283 * inet_sock_destruct() will dispose it
3284 */
3285 sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3286 WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3287 mptcp_token_destroy(msk);
3288 mptcp_pm_free_anno_list(msk);
3289 mptcp_free_local_addr_list(msk);
3290 }
3291
3292 static void mptcp_destroy(struct sock *sk)
3293 {
3294 struct mptcp_sock *msk = mptcp_sk(sk);
3295
3296 /* allow the following to close even the initial subflow */
3297 msk->free_first = 1;
3298 mptcp_destroy_common(msk, 0);
3299 sk_sockets_allocated_dec(sk);
3300 }
3301
3302 void __mptcp_data_acked(struct sock *sk)
3303 {
3304 if (!sock_owned_by_user(sk))
3305 __mptcp_clean_una(sk);
3306 else
3307 __set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3308
3309 if (mptcp_pending_data_fin_ack(sk))
3310 mptcp_schedule_work(sk);
3311 }
3312
3313 void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3314 {
3315 if (!mptcp_send_head(sk))
3316 return;
3317
3318 if (!sock_owned_by_user(sk))
3319 __mptcp_subflow_push_pending(sk, ssk, false);
3320 else
3321 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3322 }
3323
3324 #define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3325 BIT(MPTCP_RETRANSMIT) | \
3326 BIT(MPTCP_FLUSH_JOIN_LIST))
3327
3328 /* processes deferred events and flush wmem */
3329 static void mptcp_release_cb(struct sock *sk)
3330 __must_hold(&sk->sk_lock.slock)
3331 {
3332 struct mptcp_sock *msk = mptcp_sk(sk);
3333
3334 for (;;) {
3335 unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3336 struct list_head join_list;
3337
3338 if (!flags)
3339 break;
3340
3341 INIT_LIST_HEAD(&join_list);
3342 list_splice_init(&msk->join_list, &join_list);
3343
3344 /* the following actions acquire the subflow socket lock
3345 *
3346 * 1) can't be invoked in atomic scope
3347 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3348 * datapath acquires the msk socket spinlock while helding
3349 * the subflow socket lock
3350 */
3351 msk->cb_flags &= ~flags;
3352 spin_unlock_bh(&sk->sk_lock.slock);
3353
3354 if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3355 __mptcp_flush_join_list(sk, &join_list);
3356 if (flags & BIT(MPTCP_PUSH_PENDING))
3357 __mptcp_push_pending(sk, 0);
3358 if (flags & BIT(MPTCP_RETRANSMIT))
3359 __mptcp_retrans(sk);
3360
3361 cond_resched();
3362 spin_lock_bh(&sk->sk_lock.slock);
3363 }
3364
3365 if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3366 __mptcp_clean_una_wakeup(sk);
3367 if (unlikely(msk->cb_flags)) {
3368 /* be sure to sync the msk state before taking actions
3369 * depending on sk_state (MPTCP_ERROR_REPORT)
3370 * On sk release avoid actions depending on the first subflow
3371 */
3372 if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3373 __mptcp_sync_state(sk, msk->pending_state);
3374 if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3375 __mptcp_error_report(sk);
3376 if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3377 __mptcp_sync_sndbuf(sk);
3378 }
3379
3380 __mptcp_update_rmem(sk);
3381 }
3382
3383 /* MP_JOIN client subflow must wait for 4th ack before sending any data:
3384 * TCP can't schedule delack timer before the subflow is fully established.
3385 * MPTCP uses the delack timer to do 3rd ack retransmissions
3386 */
3387 static void schedule_3rdack_retransmission(struct sock *ssk)
3388 {
3389 struct inet_connection_sock *icsk = inet_csk(ssk);
3390 struct tcp_sock *tp = tcp_sk(ssk);
3391 unsigned long timeout;
3392
3393 if (mptcp_subflow_ctx(ssk)->fully_established)
3394 return;
3395
3396 /* reschedule with a timeout above RTT, as we must look only for drop */
3397 if (tp->srtt_us)
3398 timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3399 else
3400 timeout = TCP_TIMEOUT_INIT;
3401 timeout += jiffies;
3402
3403 WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3404 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3405 icsk->icsk_ack.timeout = timeout;
3406 sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3407 }
3408
3409 void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3410 {
3411 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3412 struct sock *sk = subflow->conn;
3413
3414 if (status & BIT(MPTCP_DELEGATE_SEND)) {
3415 mptcp_data_lock(sk);
3416 if (!sock_owned_by_user(sk))
3417 __mptcp_subflow_push_pending(sk, ssk, true);
3418 else
3419 __set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3420 mptcp_data_unlock(sk);
3421 }
3422 if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3423 mptcp_data_lock(sk);
3424 if (!sock_owned_by_user(sk))
3425 __mptcp_sync_sndbuf(sk);
3426 else
3427 __set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3428 mptcp_data_unlock(sk);
3429 }
3430 if (status & BIT(MPTCP_DELEGATE_ACK))
3431 schedule_3rdack_retransmission(ssk);
3432 }
3433
3434 static int mptcp_hash(struct sock *sk)
3435 {
3436 /* should never be called,
3437 * we hash the TCP subflows not the master socket
3438 */
3439 WARN_ON_ONCE(1);
3440 return 0;
3441 }
3442
3443 static void mptcp_unhash(struct sock *sk)
3444 {
3445 /* called from sk_common_release(), but nothing to do here */
3446 }
3447
3448 static int mptcp_get_port(struct sock *sk, unsigned short snum)
3449 {
3450 struct mptcp_sock *msk = mptcp_sk(sk);
3451
3452 pr_debug("msk=%p, ssk=%p", msk, msk->first);
3453 if (WARN_ON_ONCE(!msk->first))
3454 return -EINVAL;
3455
3456 return inet_csk_get_port(msk->first, snum);
3457 }
3458
3459 void mptcp_finish_connect(struct sock *ssk)
3460 {
3461 struct mptcp_subflow_context *subflow;
3462 struct mptcp_sock *msk;
3463 struct sock *sk;
3464
3465 subflow = mptcp_subflow_ctx(ssk);
3466 sk = subflow->conn;
3467 msk = mptcp_sk(sk);
3468
3469 pr_debug("msk=%p, token=%u", sk, subflow->token);
3470
3471 subflow->map_seq = subflow->iasn;
3472 subflow->map_subflow_seq = 1;
3473
3474 /* the socket is not connected yet, no msk/subflow ops can access/race
3475 * accessing the field below
3476 */
3477 WRITE_ONCE(msk->local_key, subflow->local_key);
3478 WRITE_ONCE(msk->write_seq, subflow->idsn + 1);
3479 WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3480 WRITE_ONCE(msk->snd_una, msk->write_seq);
3481
3482 mptcp_pm_new_connection(msk, ssk, 0);
3483
3484 mptcp_rcv_space_init(msk, ssk);
3485 }
3486
3487 void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3488 {
3489 write_lock_bh(&sk->sk_callback_lock);
3490 rcu_assign_pointer(sk->sk_wq, &parent->wq);
3491 sk_set_socket(sk, parent);
3492 sk->sk_uid = SOCK_INODE(parent)->i_uid;
3493 write_unlock_bh(&sk->sk_callback_lock);
3494 }
3495
3496 bool mptcp_finish_join(struct sock *ssk)
3497 {
3498 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3499 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3500 struct sock *parent = (void *)msk;
3501 bool ret = true;
3502
3503 pr_debug("msk=%p, subflow=%p", msk, subflow);
3504
3505 /* mptcp socket already closing? */
3506 if (!mptcp_is_fully_established(parent)) {
3507 subflow->reset_reason = MPTCP_RST_EMPTCP;
3508 return false;
3509 }
3510
3511 /* active subflow, already present inside the conn_list */
3512 if (!list_empty(&subflow->node)) {
3513 mptcp_subflow_joined(msk, ssk);
3514 mptcp_propagate_sndbuf(parent, ssk);
3515 return true;
3516 }
3517
3518 if (!mptcp_pm_allow_new_subflow(msk))
3519 goto err_prohibited;
3520
3521 /* If we can't acquire msk socket lock here, let the release callback
3522 * handle it
3523 */
3524 mptcp_data_lock(parent);
3525 if (!sock_owned_by_user(parent)) {
3526 ret = __mptcp_finish_join(msk, ssk);
3527 if (ret) {
3528 sock_hold(ssk);
3529 list_add_tail(&subflow->node, &msk->conn_list);
3530 }
3531 } else {
3532 sock_hold(ssk);
3533 list_add_tail(&subflow->node, &msk->join_list);
3534 __set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3535 }
3536 mptcp_data_unlock(parent);
3537
3538 if (!ret) {
3539 err_prohibited:
3540 subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3541 return false;
3542 }
3543
3544 return true;
3545 }
3546
3547 static void mptcp_shutdown(struct sock *sk, int how)
3548 {
3549 pr_debug("sk=%p, how=%d", sk, how);
3550
3551 if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3552 __mptcp_wr_shutdown(sk);
3553 }
3554
3555 static int mptcp_forward_alloc_get(const struct sock *sk)
3556 {
3557 return READ_ONCE(sk->sk_forward_alloc) +
3558 READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3559 }
3560
3561 static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3562 {
3563 const struct sock *sk = (void *)msk;
3564 u64 delta;
3565
3566 if (sk->sk_state == TCP_LISTEN)
3567 return -EINVAL;
3568
3569 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3570 return 0;
3571
3572 delta = msk->write_seq - v;
3573 if (__mptcp_check_fallback(msk) && msk->first) {
3574 struct tcp_sock *tp = tcp_sk(msk->first);
3575
3576 /* the first subflow is disconnected after close - see
3577 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3578 * so ignore that status, too.
3579 */
3580 if (!((1 << msk->first->sk_state) &
3581 (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3582 delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3583 }
3584 if (delta > INT_MAX)
3585 delta = INT_MAX;
3586
3587 return (int)delta;
3588 }
3589
3590 static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3591 {
3592 struct mptcp_sock *msk = mptcp_sk(sk);
3593 bool slow;
3594
3595 switch (cmd) {
3596 case SIOCINQ:
3597 if (sk->sk_state == TCP_LISTEN)
3598 return -EINVAL;
3599
3600 lock_sock(sk);
3601 __mptcp_move_skbs(msk);
3602 *karg = mptcp_inq_hint(sk);
3603 release_sock(sk);
3604 break;
3605 case SIOCOUTQ:
3606 slow = lock_sock_fast(sk);
3607 *karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3608 unlock_sock_fast(sk, slow);
3609 break;
3610 case SIOCOUTQNSD:
3611 slow = lock_sock_fast(sk);
3612 *karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3613 unlock_sock_fast(sk, slow);
3614 break;
3615 default:
3616 return -ENOIOCTLCMD;
3617 }
3618
3619 return 0;
3620 }
3621
3622 static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3623 struct mptcp_subflow_context *subflow)
3624 {
3625 subflow->request_mptcp = 0;
3626 __mptcp_do_fallback(msk);
3627 }
3628
3629 static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3630 {
3631 struct mptcp_subflow_context *subflow;
3632 struct mptcp_sock *msk = mptcp_sk(sk);
3633 int err = -EINVAL;
3634 struct sock *ssk;
3635
3636 ssk = __mptcp_nmpc_sk(msk);
3637 if (IS_ERR(ssk))
3638 return PTR_ERR(ssk);
3639
3640 mptcp_set_state(sk, TCP_SYN_SENT);
3641 subflow = mptcp_subflow_ctx(ssk);
3642 #ifdef CONFIG_TCP_MD5SIG
3643 /* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3644 * TCP option space.
3645 */
3646 if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3647 mptcp_subflow_early_fallback(msk, subflow);
3648 #endif
3649 if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3650 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3651 mptcp_subflow_early_fallback(msk, subflow);
3652 }
3653 if (likely(!__mptcp_check_fallback(msk)))
3654 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3655
3656 /* if reaching here via the fastopen/sendmsg path, the caller already
3657 * acquired the subflow socket lock, too.
3658 */
3659 if (!msk->fastopening)
3660 lock_sock(ssk);
3661
3662 /* the following mirrors closely a very small chunk of code from
3663 * __inet_stream_connect()
3664 */
3665 if (ssk->sk_state != TCP_CLOSE)
3666 goto out;
3667
3668 if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3669 err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3670 if (err)
3671 goto out;
3672 }
3673
3674 err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3675 if (err < 0)
3676 goto out;
3677
3678 inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3679
3680 out:
3681 if (!msk->fastopening)
3682 release_sock(ssk);
3683
3684 /* on successful connect, the msk state will be moved to established by
3685 * subflow_finish_connect()
3686 */
3687 if (unlikely(err)) {
3688 /* avoid leaving a dangling token in an unconnected socket */
3689 mptcp_token_destroy(msk);
3690 mptcp_set_state(sk, TCP_CLOSE);
3691 return err;
3692 }
3693
3694 mptcp_copy_inaddrs(sk, ssk);
3695 return 0;
3696 }
3697
3698 static struct proto mptcp_prot = {
3699 .name = "MPTCP",
3700 .owner = THIS_MODULE,
3701 .init = mptcp_init_sock,
3702 .connect = mptcp_connect,
3703 .disconnect = mptcp_disconnect,
3704 .close = mptcp_close,
3705 .setsockopt = mptcp_setsockopt,
3706 .getsockopt = mptcp_getsockopt,
3707 .shutdown = mptcp_shutdown,
3708 .destroy = mptcp_destroy,
3709 .sendmsg = mptcp_sendmsg,
3710 .ioctl = mptcp_ioctl,
3711 .recvmsg = mptcp_recvmsg,
3712 .release_cb = mptcp_release_cb,
3713 .hash = mptcp_hash,
3714 .unhash = mptcp_unhash,
3715 .get_port = mptcp_get_port,
3716 .forward_alloc_get = mptcp_forward_alloc_get,
3717 .sockets_allocated = &mptcp_sockets_allocated,
3718
3719 .memory_allocated = &tcp_memory_allocated,
3720 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3721
3722 .memory_pressure = &tcp_memory_pressure,
3723 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3724 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3725 .sysctl_mem = sysctl_tcp_mem,
3726 .obj_size = sizeof(struct mptcp_sock),
3727 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3728 .no_autobind = true,
3729 };
3730
3731 static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3732 {
3733 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3734 struct sock *ssk, *sk = sock->sk;
3735 int err = -EINVAL;
3736
3737 lock_sock(sk);
3738 ssk = __mptcp_nmpc_sk(msk);
3739 if (IS_ERR(ssk)) {
3740 err = PTR_ERR(ssk);
3741 goto unlock;
3742 }
3743
3744 if (sk->sk_family == AF_INET)
3745 err = inet_bind_sk(ssk, uaddr, addr_len);
3746 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
3747 else if (sk->sk_family == AF_INET6)
3748 err = inet6_bind_sk(ssk, uaddr, addr_len);
3749 #endif
3750 if (!err)
3751 mptcp_copy_inaddrs(sk, ssk);
3752
3753 unlock:
3754 release_sock(sk);
3755 return err;
3756 }
3757
3758 static int mptcp_listen(struct socket *sock, int backlog)
3759 {
3760 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3761 struct sock *sk = sock->sk;
3762 struct sock *ssk;
3763 int err;
3764
3765 pr_debug("msk=%p", msk);
3766
3767 lock_sock(sk);
3768
3769 err = -EINVAL;
3770 if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3771 goto unlock;
3772
3773 ssk = __mptcp_nmpc_sk(msk);
3774 if (IS_ERR(ssk)) {
3775 err = PTR_ERR(ssk);
3776 goto unlock;
3777 }
3778
3779 mptcp_set_state(sk, TCP_LISTEN);
3780 sock_set_flag(sk, SOCK_RCU_FREE);
3781
3782 lock_sock(ssk);
3783 err = __inet_listen_sk(ssk, backlog);
3784 release_sock(ssk);
3785 mptcp_set_state(sk, inet_sk_state_load(ssk));
3786
3787 if (!err) {
3788 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3789 mptcp_copy_inaddrs(sk, ssk);
3790 mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3791 }
3792
3793 unlock:
3794 release_sock(sk);
3795 return err;
3796 }
3797
3798 static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3799 int flags, bool kern)
3800 {
3801 struct mptcp_sock *msk = mptcp_sk(sock->sk);
3802 struct sock *ssk, *newsk;
3803 int err;
3804
3805 pr_debug("msk=%p", msk);
3806
3807 /* Buggy applications can call accept on socket states other then LISTEN
3808 * but no need to allocate the first subflow just to error out.
3809 */
3810 ssk = READ_ONCE(msk->first);
3811 if (!ssk)
3812 return -EINVAL;
3813
3814 pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3815 newsk = inet_csk_accept(ssk, flags, &err, kern);
3816 if (!newsk)
3817 return err;
3818
3819 pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3820 if (sk_is_mptcp(newsk)) {
3821 struct mptcp_subflow_context *subflow;
3822 struct sock *new_mptcp_sock;
3823
3824 subflow = mptcp_subflow_ctx(newsk);
3825 new_mptcp_sock = subflow->conn;
3826
3827 /* is_mptcp should be false if subflow->conn is missing, see
3828 * subflow_syn_recv_sock()
3829 */
3830 if (WARN_ON_ONCE(!new_mptcp_sock)) {
3831 tcp_sk(newsk)->is_mptcp = 0;
3832 goto tcpfallback;
3833 }
3834
3835 newsk = new_mptcp_sock;
3836 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3837
3838 newsk->sk_kern_sock = kern;
3839 lock_sock(newsk);
3840 __inet_accept(sock, newsock, newsk);
3841
3842 set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3843 msk = mptcp_sk(newsk);
3844 msk->in_accept_queue = 0;
3845
3846 /* set ssk->sk_socket of accept()ed flows to mptcp socket.
3847 * This is needed so NOSPACE flag can be set from tcp stack.
3848 */
3849 mptcp_for_each_subflow(msk, subflow) {
3850 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3851
3852 if (!ssk->sk_socket)
3853 mptcp_sock_graft(ssk, newsock);
3854 }
3855
3856 /* Do late cleanup for the first subflow as necessary. Also
3857 * deal with bad peers not doing a complete shutdown.
3858 */
3859 if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3860 __mptcp_close_ssk(newsk, msk->first,
3861 mptcp_subflow_ctx(msk->first), 0);
3862 if (unlikely(list_is_singular(&msk->conn_list)))
3863 mptcp_set_state(newsk, TCP_CLOSE);
3864 }
3865 } else {
3866 MPTCP_INC_STATS(sock_net(ssk),
3867 MPTCP_MIB_MPCAPABLEPASSIVEFALLBACK);
3868 tcpfallback:
3869 newsk->sk_kern_sock = kern;
3870 lock_sock(newsk);
3871 __inet_accept(sock, newsock, newsk);
3872 /* we are being invoked after accepting a non-mp-capable
3873 * flow: sk is a tcp_sk, not an mptcp one.
3874 *
3875 * Hand the socket over to tcp so all further socket ops
3876 * bypass mptcp.
3877 */
3878 WRITE_ONCE(newsock->sk->sk_socket->ops,
3879 mptcp_fallback_tcp_ops(newsock->sk));
3880 }
3881 release_sock(newsk);
3882
3883 return 0;
3884 }
3885
3886 static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3887 {
3888 struct sock *sk = (struct sock *)msk;
3889
3890 if (sk_stream_is_writeable(sk))
3891 return EPOLLOUT | EPOLLWRNORM;
3892
3893 mptcp_set_nospace(sk);
3894 smp_mb__after_atomic(); /* msk->flags is changed by write_space cb */
3895 if (sk_stream_is_writeable(sk))
3896 return EPOLLOUT | EPOLLWRNORM;
3897
3898 return 0;
3899 }
3900
3901 static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3902 struct poll_table_struct *wait)
3903 {
3904 struct sock *sk = sock->sk;
3905 struct mptcp_sock *msk;
3906 __poll_t mask = 0;
3907 u8 shutdown;
3908 int state;
3909
3910 msk = mptcp_sk(sk);
3911 sock_poll_wait(file, sock, wait);
3912
3913 state = inet_sk_state_load(sk);
3914 pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3915 if (state == TCP_LISTEN) {
3916 struct sock *ssk = READ_ONCE(msk->first);
3917
3918 if (WARN_ON_ONCE(!ssk))
3919 return 0;
3920
3921 return inet_csk_listen_poll(ssk);
3922 }
3923
3924 shutdown = READ_ONCE(sk->sk_shutdown);
3925 if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
3926 mask |= EPOLLHUP;
3927 if (shutdown & RCV_SHUTDOWN)
3928 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
3929
3930 if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
3931 mask |= mptcp_check_readable(sk);
3932 if (shutdown & SEND_SHUTDOWN)
3933 mask |= EPOLLOUT | EPOLLWRNORM;
3934 else
3935 mask |= mptcp_check_writeable(msk);
3936 } else if (state == TCP_SYN_SENT &&
3937 inet_test_bit(DEFER_CONNECT, sk)) {
3938 /* cf tcp_poll() note about TFO */
3939 mask |= EPOLLOUT | EPOLLWRNORM;
3940 }
3941
3942 /* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
3943 smp_rmb();
3944 if (READ_ONCE(sk->sk_err))
3945 mask |= EPOLLERR;
3946
3947 return mask;
3948 }
3949
3950 static const struct proto_ops mptcp_stream_ops = {
3951 .family = PF_INET,
3952 .owner = THIS_MODULE,
3953 .release = inet_release,
3954 .bind = mptcp_bind,
3955 .connect = inet_stream_connect,
3956 .socketpair = sock_no_socketpair,
3957 .accept = mptcp_stream_accept,
3958 .getname = inet_getname,
3959 .poll = mptcp_poll,
3960 .ioctl = inet_ioctl,
3961 .gettstamp = sock_gettstamp,
3962 .listen = mptcp_listen,
3963 .shutdown = inet_shutdown,
3964 .setsockopt = sock_common_setsockopt,
3965 .getsockopt = sock_common_getsockopt,
3966 .sendmsg = inet_sendmsg,
3967 .recvmsg = inet_recvmsg,
3968 .mmap = sock_no_mmap,
3969 .set_rcvlowat = mptcp_set_rcvlowat,
3970 };
3971
3972 static struct inet_protosw mptcp_protosw = {
3973 .type = SOCK_STREAM,
3974 .protocol = IPPROTO_MPTCP,
3975 .prot = &mptcp_prot,
3976 .ops = &mptcp_stream_ops,
3977 .flags = INET_PROTOSW_ICSK,
3978 };
3979
3980 static int mptcp_napi_poll(struct napi_struct *napi, int budget)
3981 {
3982 struct mptcp_delegated_action *delegated;
3983 struct mptcp_subflow_context *subflow;
3984 int work_done = 0;
3985
3986 delegated = container_of(napi, struct mptcp_delegated_action, napi);
3987 while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
3988 struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3989
3990 bh_lock_sock_nested(ssk);
3991 if (!sock_owned_by_user(ssk)) {
3992 mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
3993 } else {
3994 /* tcp_release_cb_override already processed
3995 * the action or will do at next release_sock().
3996 * In both case must dequeue the subflow here - on the same
3997 * CPU that scheduled it.
3998 */
3999 smp_wmb();
4000 clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4001 }
4002 bh_unlock_sock(ssk);
4003 sock_put(ssk);
4004
4005 if (++work_done == budget)
4006 return budget;
4007 }
4008
4009 /* always provide a 0 'work_done' argument, so that napi_complete_done
4010 * will not try accessing the NULL napi->dev ptr
4011 */
4012 napi_complete_done(napi, 0);
4013 return work_done;
4014 }
4015
4016 void __init mptcp_proto_init(void)
4017 {
4018 struct mptcp_delegated_action *delegated;
4019 int cpu;
4020
4021 mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4022
4023 if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4024 panic("Failed to allocate MPTCP pcpu counter\n");
4025
4026 init_dummy_netdev(&mptcp_napi_dev);
4027 for_each_possible_cpu(cpu) {
4028 delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4029 INIT_LIST_HEAD(&delegated->head);
4030 netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4031 mptcp_napi_poll);
4032 napi_enable(&delegated->napi);
4033 }
4034
4035 mptcp_subflow_init();
4036 mptcp_pm_init();
4037 mptcp_sched_init();
4038 mptcp_token_init();
4039
4040 if (proto_register(&mptcp_prot, 1) != 0)
4041 panic("Failed to register MPTCP proto.\n");
4042
4043 inet_register_protosw(&mptcp_protosw);
4044
4045 BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4046 }
4047
4048 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
4049 static const struct proto_ops mptcp_v6_stream_ops = {
4050 .family = PF_INET6,
4051 .owner = THIS_MODULE,
4052 .release = inet6_release,
4053 .bind = mptcp_bind,
4054 .connect = inet_stream_connect,
4055 .socketpair = sock_no_socketpair,
4056 .accept = mptcp_stream_accept,
4057 .getname = inet6_getname,
4058 .poll = mptcp_poll,
4059 .ioctl = inet6_ioctl,
4060 .gettstamp = sock_gettstamp,
4061 .listen = mptcp_listen,
4062 .shutdown = inet_shutdown,
4063 .setsockopt = sock_common_setsockopt,
4064 .getsockopt = sock_common_getsockopt,
4065 .sendmsg = inet6_sendmsg,
4066 .recvmsg = inet6_recvmsg,
4067 .mmap = sock_no_mmap,
4068 #ifdef CONFIG_COMPAT
4069 .compat_ioctl = inet6_compat_ioctl,
4070 #endif
4071 .set_rcvlowat = mptcp_set_rcvlowat,
4072 };
4073
4074 static struct proto mptcp_v6_prot;
4075
4076 static struct inet_protosw mptcp_v6_protosw = {
4077 .type = SOCK_STREAM,
4078 .protocol = IPPROTO_MPTCP,
4079 .prot = &mptcp_v6_prot,
4080 .ops = &mptcp_v6_stream_ops,
4081 .flags = INET_PROTOSW_ICSK,
4082 };
4083
4084 int __init mptcp_proto_v6_init(void)
4085 {
4086 int err;
4087
4088 mptcp_v6_prot = mptcp_prot;
4089 strcpy(mptcp_v6_prot.name, "MPTCPv6");
4090 mptcp_v6_prot.slab = NULL;
4091 mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4092 mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4093
4094 err = proto_register(&mptcp_v6_prot, 1);
4095 if (err)
4096 return err;
4097
4098 err = inet6_register_protosw(&mptcp_v6_protosw);
4099 if (err)
4100 proto_unregister(&mptcp_v6_prot);
4101
4102 return err;
4103 }
4104 #endif
Mirror https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git